Datasets:

emozilla

/

dolma-v1_7-arxiv

like 1

\section{Introduction} \label{sec:1} Gliese 876 (GJ876) is an M-dwarf star ($M_\star \simeq 0.37 M_\odot$) 4.6 pc away that harbors a well-studied multi-planetary system \citep{1998ApJ...505L.147M,2007ASSL..350.....V,2014MNRAS.438.2413V}. Through radial velocity measurements, we know of at least four planets orbiting GJ876, the outer three of which are in a 4:2:1 (Laplace) mean-motion resonance (MMR) \citep{2010ApJ...719..890R}. Leveraging the strong resonant interactions, several studies have been able to break the mass and inclination degeneracy of the system, providing strong constraints on the full orbital parameters and planetary masses \citep{2010A&A...511A..21C,2010ApJ...719..890R,2016MNRAS.455.2484N,2018A&A...609A.117T,2018AJ....155..106M}. Of particular note is that the system is chaotic on potentially observable timescales, and that the mass ratios of GJ876-c ($M_c \simeq 265 M_\oplus$, $P_c = 30 d$, $e_c = 0.26$) and GJ876-b ($M_b \simeq 845 M_\oplus$, $P_b = 61 d$, $e_b = 0.036$) are large enough to potentially open a deep and wide gap in their natal disk -- possibly extending out to the outermost planet, GJ876-e ($M_e \simeq 16 M_\oplus$, $P_e = 124 d$, $e_e \simeq 0.03$) \citep{2001A&A...374.1092S,2010A&A...510A...4R}. Recently, \cite{2015AJ....149..167B} investigated the formation of GJ876 through \emph{stochastic} disk migration. They attributed the stochasticity to turbulence in the disk, and argued that it was essential for producing the observed chaotic state of the system. Historically, the magnetorotational instability (MRI) was thought to be the driver of turbulence in protoplanetary disks \citep{1991ApJ...376..214B}. However, recent simulations have shown that non-ideal MHD processes can suppress or completely shutoff the MRI in the midplanes of protoplanetary disks where planets are expected to form \citep[e.g.][]{2013ApJ...769...76B,2014A&A...566A..56L,2015MNRAS.454.1117S}. If these regions are laminar or weakly turbulent, then stochastic forcing of a smoothly migrating planet may be negligible. It then becomes important to know whether or not smooth migration alone can account for all of the observed characteristics of GJ876. Moreover, while previous studies have focused on constructing the 2:1 resonance of GJ876-c,b through smooth migration \citep[e.g.,][]{2002ApJ...567..596L}, there has yet to be an exhaustive analysis focused on constructing the chaotic Laplace resonance through disk migration. This paper is organized as follows. In \S\ref{sec:2} we outline our procedure for simulating the migration history of GJ876 and our method for efficiently exploring parameter space. In \S\ref{sec:3} we present the main results from our suite of simulations. In \S\ref{sec:4} we explore several different variations on the procedure outlined in \S\ref{sec:2}. Finally, we end in \S\ref{sec:5} by discussing improvements to our smooth migration model and the possibility of extending our procedure to other resonant chain exoplanet systems. \section{Numerical model} \label{sec:2} \subsection{Smooth migration} Smooth migration models attempt to parameterize the complex energy and angular momentum transfer processes between a planet and its surrounding disk by specifying the exponential damping timescales associated with semi-major axis evolution, $\tau_a \equiv | a/\dot{a}|$, and eccentricity evolution, $\tau_e \equiv |e/\dot{e}|$ \citep[for a review see e.g.][]{2012ARA&A..50..211K}. The damping timescales for a planet of mass $M_p$ orbiting a star of mass $M_\star$ at semi-major axis $a$ and orbital period $P$ are, \begin{align} \tau_a^{-1} = C_a \left(\frac{4 \pi a^2 \Sigma_p M_p}{M_\star^2}\right) \left(\frac{H}{a} \right)^{-2} P^{-1} , \label{eq:tau_a} \\ \tau_e^{-1} = C_e \left(\frac{4 \pi a^2 \Sigma_p M_p}{M_\star^2}\right) \left(\frac{H}{a} \right)^{-4} P^{-1} , \label{eq:tau_e} \end{align} where $\Sigma_p$ and $H/a$ are the surface density and aspect ratio of the disk at the location of the migrating planet \citep{1980ApJ...241..425G}. We adopt the proportionality constants $C_a = 2.175$ and $C_e = 0.39$ \citep{2002ApJ...565.1257T,2004ApJ...602..388T}, but note that they depend strongly on the local disk structure \citep[e.g.][]{1980ApJ...241..425G,1981ApJ...243.1062G,1993ApJ...419..155A,1993ApJ...419..166A,2003ApJ...585.1024G,2012ARA&A..50..211K}. Equations \eqref{eq:tau_a} and \eqref{eq:tau_e} correspond to the so-called Type I migration regime \citep{1997Icar..126..261W,2014AJ....147...32G}. The ratio of these timescales, $K$, depends on the local disk aspect ratio as, \begin{equation} \label{eq:K} K \equiv \frac{\tau_a}{\tau_e} \propto \left( \frac{H}{a} \right)^{-2}. \end{equation} In the N-body simulations presented in the next section, we only apply semi-major axis damping forces to GJ876-e as its migration rate should be relatively fast compared to the slow Type II migration rates of GJ876-c,b \citep[e.g.][]{2012ARA&A..50..211K,2015A&A...574A..52D}. However, since the eccentricity damping rates of the inner planets may be faster than their semi-major axis damping rates by a factor of $K$, and given that the equilibrium eccentricities of planets migrating in MMR depend on their eccentricity damping timescale \citep[e.g.][]{2002ApJ...567..596L,2014AJ....147...32G}, we apply eccentricity damping forces to the planets GJ876-c,b with a damping timescale denoted by $\tau_{e,1}$, in our N-body simulations. \subsection{Simulation setup} \label{sec:setup} Our goal is to map out the parameter space in $K$, $\tau_a$, and $\tau_{e,1}$ for the system properties of GJ876. Our simulation procedure is similar to that of \citet{2017ApJ...840L..19T}, who studied the formation history of the Trappist-1 resonant chain. For a given $K$, $\tau_a$, and $\tau_{e,1}$ we first initialize a coplanar, three planet system around a $M_\star = 1 M_\odot$ star. The masses of the three planets are chosen to reproduce the best fit mass ratios of planets GJ876-c,b and e (which we henceforth label as planets 1, 2, and 3). These are set to $M_1 = 2.15\times10^{-3} M_\star$, $M_2 = 6.85 \times 10^{-3} M_\star$, and $M_3 = 1.30 \times 10^{-4} M_\star$ \citep{2018AJ....155..106M}. The initial orbital periods are $P_1 = 1$ year, $P_2/P_1= 2.2$, and $P_3/P_2 = 10$. Once initialized, we integrate each planetary system for a time of $10 \tau_a$ while applying the aforementioned damping forces. After this initial damping phase, we remove all damping forces from the system over a time of $\tau_a$. If the system is in the Laplace resonance, we rescale the system to the observed properties of GJ876. This entails changing the mass of the central star to $M_\star = 0.37 M_\odot$ and the inner planet's period to the observed period of GJ876-c, all while keeping the planet-to-star mass ratios and period ratios the same as in the previous phase. Finally, we integrate the rescaled system without any damping forces for $10^5$ years. This rescaling of the simulation after the migration phase is similar to what was done in \citet{2017ApJ...840L..19T} for the Trappist-1 system, and allows us to avoid having to fine tune the initial conditions or adjust the duration of the damping phase so that the planets end up near their observed periods by the end of the damping period. All integrations are done with the REBOUND\footnote{\url{http://github.com/hannorein/rebound}, v3.5.2.} N-body code \citep{Rein:2012cd}. For the damping forces we use the REBOUND extension REBOUNDX\footnote{\url{http://github.com/dtamayo/reboundx}, v2.18.1.} to introduce orbit-averaged forces which give the proper damping timescales \citep{2000MNRAS.315..823P}. We use the WHFast integrator for all simulations with $40$ timesteps per initial orbital period of the inner planet \citep{Rein:2015cl}. We found this to be in good agreement with simulations run with the more accurate (but slower) IAS15 integrator \citep{Rein:2015ib}. We show the time evolution of the planetary periods, eccentricities, Laplace resonance angle, and two 2:1 resonance angles of a representative simulation in Figure \ref{fig:sim_demo}. The Laplace angle is defined in terms of the mean longitudes, $\lambda$, of the planets as $\Phi_L = \lambda_1 - 3 \lambda_2 + 2 \lambda_3$. The 2:1 MMR resonant angles are defined as $\Phi_{1,2} = 2\lambda_2 - \lambda_1 - \varpi_2$ and $\Phi_{2,3} = 2 \lambda_3 - \lambda_2 - \varpi_2$, where $\varpi$ denotes the the longitude of pericenter. In this example simulation, for times less than $10 \tau_a$ the outer planet experiences damping forces with $ \log_{10} K = 3.125$ and $\tau_a = 10^4 P_1$. The inner planets also experience eccentricity damping forces with $\tau_{e,1} = 10^{5.5} P_1$. During this phase the planets are brought into successive 2:1 MMRs, evidenced by $\Phi_{1,2}$ and $\Phi_{2,3}$ librating about $0^\circ$ with small amplitude. At nearly the same time the system also catches into the Laplace resonance. As the system continues to migrate while in the 4:2:1 resonance, the eccentricity of the inner planet grows to $e_1 \approx 0.25$, while $e_2 \approx e_3 \approx 0.03$. Between times of $10 \tau_a$ and $11 \tau_a$, which corresponds to the region between the two vertical dashed lines in Figure \ref{fig:sim_demo}, we gradually remove the damping forces. Then at a time of $11 \tau_a$ we completely remove all damping forces and rescale the system to an inner period of $30$ days and a central mass of $0.37 M_\odot$. Finally, we evolve the system for $10^5$ years where it remains in a stable, resonant configuration. \begin{figure} \centering \plotone{f1.pdf} \caption{Time evolution of a simulation with $(\log_{10}(K),\log_{10}(\tau_a/P_1),\log_{10}(\tau_{e,1}/P_1)) = (3.125,4,5.5)$ demonstrating our simulation procedure. The vertical lines mark the end of the initial damping phase and the start of the damping-free phase. During this time period the damping forces on the planets are reduced to zero over a time of $\tau_a$. At the beginning of the damping-free phase we rescale the system to the observed period of the inner planet. In descending order, the panels show the time evolution of the orbital periods, eccentricities, the Laplace resonant angle, and two, 2:1 MMR angles. } \label{fig:sim_demo} \end{figure} \begin{figure*} \centering \plotone{f2.pdf} \caption{The Laplace angle libration amplitude and our AMR grid in $\tau_a$ and $K=\tau_a/\tau_e$ space for fixed $\tau_{e,1}=10^5 P_1$, at increasing levels of refinement. The color of each cell is the average libration amplitude of $10$ simulations where the initial angles were randomized. The opacity of each cell indicates its instability fraction, with white cells indicating that all $10$ simulations went unstable. The observed libration amplitude of GJ876 is approximately $26^\circ \pm 5^\circ$ \citep{2018AJ....155..106M}. } \label{fig:grid} \end{figure*} \subsection{Efficiently exploring parameter space} \begin{figure*}[th] \centering \plotone{f3.pdf} \caption{Overview of the resonant angles (top row), eccentricities (middle row), inner and outer period ratios, and Lyapunov time (bottom row) for the simulation procedure described in \S\ref{sec:setup}. All quantities are averaged over $\tau_{e,1}$. An overbar denotes the time-averaged value, while angle brackets denote the RMS libration amplitude. When present, regions falling within the red contours signify $\le 3 \sigma$ agreement with the observed values from \citet{2018AJ....155..106M} (see \S\ref{sec:chi2}). In the Lyapunov time plot, simulations which did not catch into the Laplace resonance by the end of the damping phase are not shown (see \S\ref{sec:chaos}). } \label{fig:sum} \end{figure*} To efficiently sample the 3D parameter space of $K$, $\tau_a$, and $\tau_{e,1}$, we adopt the techniques of adaptive mesh refinement \citep[AMR;][]{1989JCoPh..82...64B} to focus most of our computational efforts on parts of parameter space where the system is in the Laplace resonance. The "mesh" used here is the discretized 3D parameter space of $(K,\tau_a,\tau_{e,1})$, where at each point on the mesh we run an N-body simulation. We apply the refinement procedure of \citet{Lohner:1987hq} \citep[see also e.g.][]{2000ApJS..131..273F} to the RMS libration amplitude of the Laplace angle. We define the RMS libration amplitude as, \begin{equation} \langle \Phi_L \rangle = \sqrt{\frac{2}{T} \int_{t_{\rm final}-T}^{t_{\rm final}} dt\, \Phi_L^2(t)} , \end{equation} where the averaging time $T=10^4 P_1$ for resonant systems\footnote{For simulations that do not catch into resonance the averaging time is $10\%$ of $\tau_a$.}. We focus our refinement on regions where $\langle \Phi_L \rangle < 100^\circ$ and where there is a boundary between an unstable\footnote{For simplicity we define "unstable" to mean an ejection of a planet from the system or when the Hill spheres of two planets overlap.} cell and a cell in resonance. Our adaptive grid uses up to $8$ refinement levels, with each refinement level doubling the number of points in $K$, $\tau_a$, and $\tau_{e,1}$, respectively. Our final grid has an effective uniform resolution of $256^3$ logarithmically spaced points in the domain $\tau_a \in [10^2, 10^6] P_1$ , $K \in [1,10^4]$, and $\tau_{e,1} \in [10^4,10^8] P_1$\footnote{Preliminary simulations ruled out regions where $\tau_a < 10^2 P_1$, $K>10^4$, and $\tau_{e,1} < 10^4 P_1$.}. We determine the AMR grid through one set of simulations, and then repeat each simulation with randomized initial phase angles for a total of $10$ simulations per $(K,\tau_a,\tau_{e,1})$. In Figure \ref{fig:grid} we show our AMR grid and $\langle \Phi_L \rangle$ for a slice at $\tau_{e,1} = 10^5 P_1$. The opacity of each cell shows the instability fraction, with white cells indicating that all $10$ initial conditions for that $(K,\tau_a,\tau_{e,1})$ point went unstable. In the end, we run a total of $\sim1.3$ million simulations, roughly two orders of magnitude less than the $\sim160$ million simulations required to have a uniform resolution of $256^3$ points over our range of $K$, $\tau_a$, and $\tau_{e,1}$. These savings allow us to explore several different variations on the setup described in \S\ref{sec:setup}, which we detail further in \S\ref{sec:ics}. \section{Results} \label{sec:3} Our main results are shown in Figure \ref{fig:sum}, where we plot the $(K,\tau_a)$ distributions, integrated over $\tau_{e,1}$, of several important quantities, and in Figures \ref{fig:slice_1}-\ref{fig:slice_3}, where we show the $\tau_{e,1}$ dependence of $\Phi_L$ and $e_1$ at several slices of constant $\tau_a$ and $K$. \subsection{Resonant angles and period ratios} The top row of Figure \ref{fig:sum} shows the libration amplitudes of the two-body, 2:1 resonance angles, $\Phi_{1,2} = 2 \lambda_2 - \lambda_1 - \varpi_2$, $\Phi_{2,3} = 2 \lambda_3 - \lambda_2 - \varpi_2$, and $\Phi_L$. We find that the majority of parameter space does not produce systems in the Laplace resonance. Rather, there is a bounded resonant region where the libration amplitudes of $\Phi_{1,2}$, $\Phi_{2,3}$, and $\Phi_L$ tend to be smallest in the center and quickly rise (or become unstable) at the edges. We argue in the next section that the instabilities at $K \lesssim 10$ are the result of the eccentricity of the outer planet increasing to the point of orbit-crossing with the middle planet, while the instabilities at $K \gtrsim 10$ are primarily due to high eccentricities of the inner planet. The observed libration amplitude of $\Phi_L$ is of moderate to large amplitude \citep[$\simeq 20^\circ-30^\circ$;][]{2016MNRAS.455.2484N,2018AJ....155..106M}, and so we are primarily interested in the transitional regions of parameter space. The bottom row of Figure \ref{fig:sum} shows the period ratios for the inner and outer pair of planets. Since we start the outer planet well outside the 3:1 MMR, the outer two planets catch into MMRs with period ratios greater than 2 for slow enough damping timescales. This is not the case for the inner pair of planets, as they are initially placed just wide of the 2:1 resonance. Moreover, once the outer pair is locked into resonance, the effective migration rate slows by a factor dependent on the mass ratios \citep{2002ApJ...567..596L}. This reduction of the effective migration rate makes it easier for the inner pair of planets to capture into the 2:1 resonance, rather than push through to the 3:2 MMR \citep{2014AJ....147...32G}. \begin{figure}[h] \centering \plotone{f4.pdf} \caption{Similar to Figure \ref{fig:sum}, but we focus on the Laplace angle libration amplitude for four constant values of $\tau_{e,1}$: $10^{4},10^{4.5}, 10^{5}$ and $10^7 P_1$. As $\tau_{e,1}$ decreases, the resonant region shrinks and moves to lower $\tau_a$ and $K$.} \label{fig:slice_1} \end{figure} \begin{figure}[h] \centering \plotone{f5.pdf} \caption{Similar to Figure \ref{fig:slice_1}, but for slices through parameter space for constant $\tau_a$ (left column) and constant $K$ (right column) for the Laplace angle libration amplitude. The $\tau_a$ slices are $10^{4} P_1$ and $10^{4.5} P_1$, while the $K$ slices are at $10^2$ and $10^3$.} \label{fig:slice_2} \end{figure} In Figure \ref{fig:slice_1} we show the libration amplitude of $\Phi_L$ for four different values of $\tau_{e,1}$: $10^{4} P_1$, $10^{4.5} P_1$, $10^5 P_1$, and $10^7 P_1$. As $\tau_{e,1}$ decreases the stable Laplace resonance region shrinks and shifts towards lower $\tau_a$ and lower $K$ which were unstable at larger values of $\tau_{e,1}$. The magnitude of $\langle \Phi_L \rangle$ is also typically larger for lower $\tau_{e,1}$. Note that the unstable region for $K \lesssim 10$ is relatively robust to the value of $\tau_{e,1}$, reinforcing the understanding that this is connected to the large eccentricity of the outer planet. On the other hand, the region of instability at the lower end of the resonant region for $K \gtrsim 10$ shrinks as $\tau_{e,1}$ decreases, as the equilibrium eccentricity of the inner planet is lower (see also Figure \ref{fig:slice_3}). Similar to Figure \ref{fig:slice_1}, Figure \ref{fig:slice_2} shows $\langle \Phi_L \rangle$ for cuts of constant $\tau_a$ and $K$. In the left panels, we show the $(K,\tau_{e,1})$ dependence of $\langle \Phi_L \rangle$ for $\tau_a = 10^4 P_1$ and $10^{4.5} P_1$, while in the right panels we show the $(\tau_a,\tau_{e,1})$ dependence for $K = 10^2$ and $10^3$. In all cases, $\langle \Phi_L \rangle$ is roughly independent of $\tau_{e,1}$ until $\tau_{e,1} \sim \tau_a$ at which point the system quickly escapes from the Laplace resonance. \subsection{Eccentricities} The middle row of Figure \ref{fig:sum} shows the final time-averaged eccentricities of each simulation. When in the three-body resonance, the inner planet's eccentricity is typically pumped to large values ($e_1 \sim0.4$ for large $\tau_{e,1}$), while $e_2$ remains low, and $e_3$ increases as $K$ decreases. Figure \ref{fig:slice_3} shows the $\tau_{e,1}$ dependence of $e_1$ for the same constant values of $\tau_a$ and $K$ as discussed previously in Figure \ref{fig:slice_2}. Similar to $\langle \Phi_L \rangle$, for fixed $\tau_a$ and $K$, the inner planet's eccentricity is roughly independent of $\tau_{e,1}$ when $\tau_{e,1}$ is large. When $\tau_{e,1}$ becomes comparable in magnitude to $\tau_a$ the inner planet reaches an equilibrium eccentricity lower than $e_1 \sim 0.4$. However, the transition to $e_1 < 0.2$ is relatively quick as $\tau_{e,1}$ decreases while the transition is more gradual for fixed $\tau_{e,1}$ and increasing either $\tau_a$ or $K$. In the next section, we explore more quantitatively how well our simulated systems compare to the observed GJ876. \begin{figure}[h] \centering \plotone{f6.pdf} \caption{Same as Figure \ref{fig:slice_2}, but for the eccentricity of the inner planet.} \label{fig:slice_3} \end{figure} \subsection{Comparison to the real GJ876} \label{sec:chi2} \begin{figure}[t] \centering \plotone{f7.pdf} \caption{The $\chi^2$ statistic defined in Equation \eqref{eq:chi2}. The top panel shows $\chi^2$ as a function of $K$ and $\tau_a$. The bottom panel shows $\chi^2$ as a function of disk thickness, $H/a$, and disk surface density, $\Sigma$ at the location of the outer planet. To convert from $\tau_a$ and $K$ to $\Sigma$ and $H/a$ we use Equations \eqref{eq:tau_a} and \eqref{eq:tau_e}.} \label{fig:chi2} \end{figure} The red contours in Figures \ref{fig:sum}-\ref{fig:slice_3} show the $3 \sigma$ contours of the observed values for the three resonant angles, $\Phi_{1,2} = 10.4^\circ \pm 1.8^\circ$, $\Phi_{2,3} = 25.0^\circ \pm 4.65^\circ$, $\Phi_L = 26.6^\circ \pm 5.0^\circ$, and three eccentricities, $e_1 = 0.257 \pm 0.002$, $e_2 = 0.033 \pm 0.002$, and $e_3 = 0.03 \pm 0.023$, taken from \citet{2018AJ....155..106M}. For simplicity, we have averaged any asymmetric errors. The resonant angles are within the $3 \sigma$ contours for a wide range of $K$, $\tau_a$, and $\tau_{e,1}$. The eccentricities of the inner planets, on the other hand, are near their observed values in a much narrower range of parameter space. This suggests that the inner planets' eccentricities provide a strong constraint on the region of parameter space consistent with the observed system. To quantify how well a system agrees with GJ876 we calculate a $\chi^2$ statistic for each simulation, which we define as, \begin{equation} \label{eq:chi2} \chi^2 = \frac{1}{6} \sum_{i=1}^6 \left( \frac{ y_i - y_\text{obs,i}}{\sigma_\text{obs,i}} \right)^2 , \end{equation} and where $y_{\rm obs}$ and $\sigma_{\rm obs}$ are the observed values and uncertainties of $\Phi_{1,2},\Phi_{2,3}, \Phi_L, e_1, e_2$, and $e_3$. For a given $(K,\tau_a)$ pair we show the minimum $\chi^2$ across all $\tau_{e,1}$ values in the top panel of Figure \ref{fig:chi2}. This combined statistic attains its minimum in the region of parameter space near $K\simeq 30-3,000$ and $\tau_a\simeq 10^3 -10^4 P_1$. In particular, there are two regions of best fit: one around $K\simeq30$ and $\tau_a \simeq 3 \times 10^3-10^4 P_1$ and another around $K\simeq 300-1,000$ and $\tau_a \simeq 10^4 P_1$. The region at larger $K$ and longer $\tau_a$ corresponds to systems with weak eccentricity damping on the inner planets (large $\tau_{e,1}$), while the region at lower $K$ and shorter $\tau_a$ corresponds to systems with short $\tau_{e,1}$. For simplicity we use equal weighting for the six parameters when computing $\chi^2$ in Equation \eqref{eq:chi2}. If the resonant angles are given more weight in the sum, the $\chi^2$ values generally come down since the relative uncertainties in the angles are larger, but the overall best-fit region does not change. This is clear from the top row of Figure \ref{fig:sum} which shows that the region enclosed by the $3 \sigma$ contours is the same for the three resonant angles and coincides with the overall low $\chi^2$ region in the top panel of Figure \ref{fig:chi2}. Similarly, if we give more weight to the eccentricities, the minimum $\chi^2$ region as shown in Figure \ref{fig:chi2} experiences little change. In the bottom panel of Figure \ref{fig:chi2} we use Equations \eqref{eq:tau_a} and \eqref{eq:tau_e} to derive a constraint on the surface density of gas and disk thickness near the outer Laplace planet. Again, there are two regions of parameter space where the simulations match the observed system particularly well. One has low surface density and thickness with $\Sigma \simeq \text{few} \times 10^2$ $\text{g/cm}^2$ and $H/a \simeq 0.01-0.02$, and the other has a higher surface density and thickness with $\Sigma \simeq \text{few} \times 10^3 - 10^5$ $\text{g/cm}^2$ and $H/a \simeq \text{few} \times 10^{-2}$. This latter region typically requires stronger eccentricity damping of the inner planets in order to drive the system into the observed configuration. In other words, if we did not damp the inner planets in addition to the outermost planet, we would infer a very thin, low surface density disk near GJ876-e at the formation time of the Laplace resonance. As previously mentioned, these estimates depend on the overall normalizations connecting the disk properties to the semi-major axis and eccentricity damping timescales, and are best determined by hydrodynamical models \citep{1980ApJ...241..425G,1981ApJ...243.1062G,1993ApJ...419..155A,1993ApJ...419..166A,2002ApJ...567..596L,2002ApJ...565.1257T,2012ARA&A..50..211K}. Nevertheless, it is encouraging that we obtain reasonable numbers \citep[cf.][]{2005ApJ...631.1134A} for the inner regions of GJ876's protoplanetary disk given our simplified migration model. \subsection{Chaos} \label{sec:chaos} \begin{figure} \centering \plotone{f8.pdf} \caption{\emph{Top}: Normalized histogram of $\tau_{\rm lyap}$ for simulations with $10^4 \le \tau_{e,1} < 10^5 P_1$ (blue line) and $10^5 \le \tau_{e,1} < 10^8 P_1$. Because of our non-uniform grid, each simulation is given a weight of $8^{8 - \ell}$, where $\ell$ is the level of refinement for that simulation. \emph{Bottom}: A 2D histogram showing the dependence of $\tau_{\rm lyap}$ on the inner planets' eccentricity damping rate. The color of a pixel centered on a given $(\tau_{\rm lyap},\tau_{e,1})$ shows the total number of times across all $(K,\tau_a)$ pairs that the Lyapunov time was that value. The counts are normalized to the total number.} \label{fig:lyap_2d} \end{figure} Dynamical fits to GJ876 strongly suggest that the system is chaotic with a Lyapunov timescale between tens to thousands of years \citep{2015AJ....149..167B,2016MNRAS.455.2484N,Marti:2016ep,2018AJ....155..106M}. The Lyapunov timescale, $\tau_{\rm lyap}$, characterizes the time it takes for nearly identical initial conditions to diverge, and hence is one measure of chaos \citep[e.g.][]{1999ssd..book.....M}. During the non-damping stage for systems deemed in the Laplace resonance, we track $\tau_{\rm lyap}$ by integrating the variational equations and monitoring the temporal evolution of the MEGNO number \citep[for a description of the algorithms used see e.g.;][]{2003PhyD..182..151C,Rein:2016hv}. The bottom right plot of Figure \ref{fig:sum} shows the 2D distribution of $\tau_{\rm lyap}$ for our standard set of simulations. In the resonant region, we typically find $\tau_{\rm lyap} \gtrsim 10^3$ years, with shorter timescales on the fringes of the region. In Figure \ref{fig:lyap_2d} we show how the Lyapunov time varies with $\tau_{e,1}$. In the top panel, we show the distribution of $\tau_{\rm lyap}$ for all of our simulations. Recall that we only track the chaos indicators in our simulations if they exit the damping stage with $\langle \Phi_L \rangle \le 100^\circ$. To get a sense of the dependence on $\tau_{e,1}$, we split the distribution into systems with strong ($\tau_{e,1} < 10^5 P_1$) and weak ($\tau_{e,1} \ge 10^5 P_1$) eccentricity damping on the inner planets. The distributions are mostly similar with peaks around a few thousand years, but systems with shorter $\tau_{e,1}$ have a higher chance of having $\tau_{\rm lyap} < 10^3$ years. We explore this further in the bottom panel of Figure \ref{fig:lyap_2d}, where we show the 2D distribution of $\tau_{\rm lyap}$ and $\tau_{e,1}$. For a given value of $\tau_{e,1}$ we compute the histogram of $\tau_{\rm lyap}$ across all values of $K$ and $\tau_a$. The color of each pixel in Figure \ref{fig:lyap_2d} shows the normalized, total number of times that particular value of $\tau_{\rm lyap}$ occurred. Again, the peak of the distribution for all $\tau_{e,1}$ occurs around Lyapunov times of a few thousand years. At the strongest damping rates, the distribution flattens out with more occurrences of $\tau_{\rm lyap} < 10^3$ years and $\tau_{\rm lyap} > 10^4$ years. \begin{figure} \centering \plotone{f9.pdf} \caption{Correlations of the libration amplitudes for the 4:1 resonant angle, $4 \lambda_3 - \lambda_1 - 2 \varpi_1 - \varpi_3$, with $\tau_{\rm lyap}$, for two ranges of $\tau_{e,1}$. The top panel is for stronger eccentricity damping of the inner planet ($\tau_{e,1} < 10^5 P_1$), while the bottom panel is for weaker eccentricity damping ($\tau_{e,1} \ge 10^5 P_1$). Only simulations with $\langle \Phi_L \rangle \le 50^\circ$ are shown and the color of each point corresponds to the value of $ \langle \Phi_L\rangle$. Unlike for the Laplace angle, the libration amplitude here corresponds to the standard deviation (as opposed to the RMS).} \label{fig:aqr} \end{figure} \subsubsection{The double apsidal corotation resonance} A longer $\tau_{\rm lyap}$ is consistent with \citet{2015AJ....149..167B}, who found that in addition to the decades long $\tau_{\rm lyap}$, there was also a region of phase space where $\tau_{\rm lyap} \gtrsim 10^3$ years. Moreover, \citet{2018AJ....155..106M} associate this with a low energy, double apsidal corotation resonance where the angles $\varpi_3-\varpi_2$, $\varpi_2-\varpi_1$, $2 \lambda_3 - \lambda_2 - \varpi_3$, and $4 \lambda_3 - \lambda_1 - 2 \varpi_1 - \varpi_3$ all librate. Figure \ref{fig:aqr} shows the correlation of one of these angles, $4 \lambda_3 - \lambda_1 - 2 \varpi_1 - \varpi_3$, with $\tau_{\rm lyap}$ and $\tau_{e,1}$ for simulations with $\langle \Phi_L \rangle < 50^\circ$\footnote{The other angles associated with the double apsidal corotation resonance show the same correlation with $\tau_{\rm lyap}$ and $\tau_{e,1}$.}. This angle is associated with the 4:1 MMR between the innermost and outermost planets of the resonant chain. The top panel shows the distribution for $\tau_{e,1} < 10^5 P_1$, while the bottom panel shows the distribution for $\tau_{e,1} \ge 10^5 P_1$. Systems with $\tau_{\rm lyap} \gtrsim 10^3$ years can either have these angles librate with small amplitude or circulate and have $\langle \Phi_L \rangle \lesssim 10^\circ$, while systems with $\tau_{\rm lyap} \lesssim 10^3$ years have predominantly larger libration amplitudes and $\langle \Phi_L \rangle \gtrsim 10^\circ$. This large libration amplitude suggests that the angles are switching between libration and circulation as pointed out in \citet{2018AJ....155..106M}. Moreover, for the shortest $\tau_{e,1}$ values, we do not find any simulations where these angles have low librations amplitudes. This suggests that if the true system lies in the region of parameter space where the angles associated with the double apsidal corotation resonance librate with small amplitude, then there was little or no eccentricity damping of the inner planets at the time of the Laplace resonance formation. \section{Discussion} \label{sec:4} \subsection{Alternative initial conditions} \label{sec:ics} \begin{figure} \centering \plotone{f10_a.pdf} \plotone{f10_b.pdf} \caption{Results from variations on our initial setup. \emph{Top}: The distribution of $\tau_{\rm lyap}$ for the different setups described in \S\ref{sec:ics}. \emph{Middle}: The $\tau_{e,1}$ integrated $\langle \Phi_L \rangle$ for the 3D setup. \emph{Bottom}: The $\tau_{e,1}$ integrated $\tau_{\rm lyap}$ for the setup which includes GJ876-d.} \label{fig:lyap_3} \end{figure} Given that the majority of our simulations sit in the longer Lyapunov timescale part of phase space, we wish to now examine possible alternatives to our standard setup described in \S\ref{sec:setup} which could potentially increase the number of initial conditions leading to $\tau_{\rm lyap} \lesssim 10^3$ years. In Figure \ref{fig:lyap_3} we summarize the results of relaxing three of our initial assumptions: the timescale of the damping removal, coplanarity, and neglecting the fourth planet of the system, GJ876-d. \paragraph{Abrupt damping removal} In our standard set of simulations, we slowly removed the damping forces on the outer planet over one damping timescale. We ran an additional set of simulations where we instead abruptly turned off all damping forces. The distribution of $\tau_{\rm lyap}$ as shown in the top panel of Figure \ref{fig:lyap_3} is relatively insensitive to how smoothly the damping forces are removed. The only significant change is slightly more systems to the left of the main peak at $\tau_{\rm lyap} \simeq 3000$ years. \paragraph{Mutual inclinations} Another major simplification of our study is in assuming that the Laplace planets are coplanar with each other. This is a relatively robust assumption, however, since the observed Laplace planets are nearly coplanar \citep{2016MNRAS.455.2484N}. Nevertheless, we ran an additional set of simulations where we gave each planet a random initial inclination of up to $20^\circ$. The distribution of $\tau_{\rm lyap}$ is relatively unaffected by allowing the planets to be mutually inclined. The middle panel of Figure \ref{fig:lyap_3} shows that allowing the planets to have non-zero inclinations narrows the Laplace resonance parameter space and increases the instability fractions. \paragraph{GJ876-d} The two-day inner planet, GJ876-d ($M_d \simeq 7 M_\oplus, P_d = 1.94 d, e_d \simeq 0.11$\footnote{We use the larger eccentricity from \citet{2016MNRAS.455.2484N} as opposed to the smaller eccentricity from \citet{2018AJ....155..106M}.}), has so far been neglected in the resonance capture and subsequent evolution of our models \citep{2016MNRAS.455.2484N}. While this is likely a good assumption for the process of resonance capture, it is less so when studying the overall system parameters (such as instability and chaos) during the long term evolution of the entire system. To this end, we ran our standard set of simulations again, but after removing the damping forces we additionally included GJ876-d at its observed period and eccentricity. When GJ876-d is included, we also include the precessional effects of GR via REBOUNDX \citep{1986IAUS..114..105N}. Including GJ876-d does not noticeably affect the distribution for $\tau_{\rm lyap} < 10^3$ years. Surprisingly, however, it does increase the frequency of systems with $\tau_{\rm lyap} > 10^4$ years. Additionally, other effects not present in our simplified smooth migration models, such as the stochastic forcing used in \citet{2015AJ....149..167B}, or the presence of an eccentric disk during the damping phase \citep[e.g.][]{2005A&A...437..727K}, may shorten or significantly affect the chaotic timescale of the system. \section{Conclusions and future work}\label{sec:5} Since the majority of parameter space where GJ876 is in resonance has $\langle \Phi_L \rangle \lesssim 10^\circ$ and $\tau_{\rm lyap} \gtrsim 10^{3}$ years, it is somewhat surprising that the observed system has such large libration amplitude and short Lyapunov time. Further observations of the system will yield better constraints on these two characteristics as well as on the eccentricities of the planets \citep{2018AJ....155..106M}. In particular, the eccentricity of GJ876-c provides the strongest constraint on the appropriate damping timescales driving the system into resonance. In addition to further observations, more detailed, hydrodynamical simulations of the three Laplace planets in GJ876 are required to determine whether or not the formation scenario we have presented here is truly viable. In particular, such simulations can help explain how an eccentric disk \citep{2005A&A...437..727K}, the presence of an inner disk \citep{2008A&A...483..325C}, how the disk disperses \citep{2005A&A...437..727K}, the level of disk viscosity \citep[e.g.][]{2017ApJ...839..100F}, and other effects determine the migration rates, final eccentricities, resonance angles, and Lyapunov timescale of GJ876. Hydrodynamical simulations will also improve upon the connection between damping timescales and disk properties (e.g. Equations \eqref{eq:tau_a} and \eqref{eq:tau_e}). On this topic, \citet{2018arXiv180804223C} have recently simulated the construction of the Laplace resonance in GJ876 for several different disk thicknesses, masses, and viscosities using hydrodynamical simulations of two-dimensional locally isothermal disks. Compared to our smooth migration model, the migration rate of GJ876-e in their simulations can vary wildly with time as the presence of an eccentric disk outside of GJ876-c,b induces large variations in the torque felt by the outer planet. In many of their cases this prevents the system from forming the 4:2:1 MMR as GJ876-e is caught at larger period ratios with GJ876-b. In order to form the Laplace resonance, they offer two main avenues. The first is for GJ876-e to open a partial gap in the disk, allowing it to gradually remove the disk eccentricity induced by GJ876-b,c and migrate into the 2:1 resonance with GJ876-b. This scenario requires the disk thickness and viscosity to be low enough for GJ876-e to open a (partial) gap. The second is for the resonance capture to take place later in the systems lifetime when the disk surface density is lower. Further simulations covering a wider range of thicknesses and disk masses would facilitate a more robust comparison to our results (e.g.\,to Figure \ref{fig:chi2}). A natural extension of our AMR method is to the study of other resonant chain systems. While GJ876 is unique in its diversity of masses, systems discovered by \emph{Kepler} typically have roughly equal mass planets orbiting their host star in a compact configuration. Our method can be readily applied to systems containing resonant chains of Earth mass, or Super-Earth mass planets, and can put meaningful constraints on the local disk properties present when the planets captured into resonance \citep[see also e.g.][]{2015A&A...579A.128D}. Examples of such systems include Kepler-60, Kepler-223, Trappist 1, and others \citep{2016Natur.533..509M,2016MNRAS.455L.104G,2017NatAs...1E.129L}. In particular, chains of non-gap opening planets may undergo Type I migration, and if they have measured masses, the parameter space to explore consists only of the density and temperature profile of the disk. This low-dimensional parameter space allows for our efficient AMR scheme to fully explore the relevant parameter space. \acknowledgments The mesh refinement code used is available at \url{http://github.com/adamdempsey90/NDTAMR}. We would like to thank the referee for a thorough and insightful report which greatly improved the manuscript. We would also like to thank Sam Hadden, Yoram Lithwick, Diego Mu{\~n}oz, Matt Payne, Hanno Rein, and Dan Tamayo for useful discussions, as well as Sarah Millholland and Greg Laughlin for sharing an early version of their paper with us. This research was supported in part through the computational resources and staff contributions provided for the Quest high performance computing facility at Northwestern University which is jointly supported by the Office of the Provost, the Office for Research, and Northwestern University Information Technology. B.E.N. acknowledges support from the Data Science Initiative at Northwestern University.

434bf830c45951129857276da302bf278015c23b

\section{INTRODUCTION} In machine learning, signal processing, and control, many problems are innately time-varying. In dynamical systems, the challenge is to deal with time-varying parameters, which has motivated a long standing research on online identification \cite{pra16, bra06, bra16, li11, bit94}. In signal processing, instead, the problem is to track time-varying signals (for example, moving targets). More in general, we talk about online optimization when the problem can be formulated as the minimization of a time-varying cost functional. In the last years, the literature on this topic has grown rapidly, with particular attention to the convex case. \emph{Online convex optimization} (OCO, \cite{sha12book,haz16book}) can be described as a game in which at each time step $t=1,\dots,T$ a player has to minimize a convex cost functional $f_{t}$ revealed by an adversary. In principle, the player might find the minimum by convex programming, but the need for a real time response typically prevents it, which leads to the development of suboptimal strategies able to track the desired time-varying target. In some cases, a known dynamic is envisaged in the model, while in others a completely adversarial approach is considered, with no prior information on the evolution. In the OCO literature, a benchmark to evaluate a strategy is provided by the regret \cite{sha12book,haz16book,zin03,haz07,mok16,hos16,sha18}, which basically measures the difference between the player's sequence of decisions and the best strategy in hindsight (\emph{i.e.,} the minimization of each $f_t$). A strategy is defined successful if its regret is sublinear in $T$. In the last decade, much attention has been devoted to the regret analysis in OCO: in \cite{zin03}, a gradient-based algorithm was proposed that achieves $\mathcal{O}(\sqrt{T})$ behavior; some years later, \cite{haz07} obtained $\mathcal{O}(\log{T})$ in the case of strongly convex functionals. This result was improved in \cite{mok16}, which obtained $\mathcal{O}(1+C_T)$, where $C_T$ is the sum of the distances between successive reference points. Very recently, regret has been investigated also in OCO distributed settings \cite{hos16,sha18}. In most of the mentioned works, $f_t$ is assumed to be differentiable, with bounded gradient. This does not envisage sparsity-promoting convex cost functionals, such as, Lasso \cite{tib96} and Elastic-net \cite{zou05}, which contain a (non-differentiable) $\ell_1$-norm. Such functionals are often used for model selection and data compression in estimation/identification problems \cite{hof13,pil14,yil18}. The key idea is that the $\ell_1$ regularization promotes sparse solutions (namely, solutions with many zeros), and can be used to build parsimonious models. This is of particular interest when models are built from large and noisy datasets. Moreover, in signal processing, sparsity-promoting functionals have gained new popularity with the advent of Compressed Sensing (CS, \cite{don06,fou13}), which states that sparse signals can be recovered from compressed linear measurements. The aim of this work is to tackle the problem of sparse OCO. Since the best results in terms of regret analysis have been obtained so far with strongly convex functionals, we consider the (strongly-convex) Elastic-net model. We then elaborate an online strategy, based on iterative soft thresholding (IST, \cite{dau04, for10}), and we analyze its regret. Finally, we present a practical application for the identification of time-varying systems. In the context of sparse signal processing and CS, the problem of tracking time-varying sparse signals has been tackled, but not in the regret analysis perspective. Moreover, in most works a known dynamics was assumed. In \cite{zin13}, an approximate message passing method was proposed to track time-varying sparse signals (acquired according to the CS paradigm), knowing that the dynamics is ruled by a Markov model; simulations showed good efficiency of the approximate message passing in different applications, but no theoretical analysis was provided. In \cite{cha16}, a dynamic filtering via $\ell_{1}$ minimization was proposed, again assuming a Markov evolution model. The related optimization problem was formulated, at time $t$, as a Lasso plus a term promoting the consistency with the prediction of previous time step. Two algorithms were proposed: BPDN-DF, based on basis pursuit and Kalman filtering optimization, has theoretical guarantees on the error boundedness \cite[Theorem III.1]{cha16}, but degrades when the inaccuracy in the dynamics model increases; RWL1-DF is a re-weighted $\ell_1$ filtering, which achieves better performance according to numerical tests. \cite{mot15} considered the same model as \cite{cha16} (except for an $\ell_1$-norm term that enforces consistency with the prediction), then proposed an iterative algorithm and conditions for exact recovery. A Kalman filtering approach was considered also in \cite{zac12}, where a dynamic iterative pursuit was proposed, based on a variant of orthogonal matching pursuit including prediction. In \cite{bal15}, IST was analyzed for tracking time-varying sparse signals, and a theoretical error bound was provided under the assumption of boundedness of the sparse signal and of its derivative. We finally mention the slightly different problem of streaming sets measurements, in which the unknown signal is static, but observations come sequentially, and one aims to refine the estimate online. Homotopy methods \cite{sal10,hof13} were proposed to tackle it. The approach is recursive: recovery is not performed from scratch each time, but past estimation is used to update the current one. The paper is organized as follows. In Section \ref{sec:ps}, we state the optimization problem in a rigorous way. In Section \ref{sec:pm}, we introduce the proposed algorithm, whose regret is analyzed in Section \ref{sec:ra}. We then show a few numerical results in an online identification example (Section \ref{sec:nr}), and we finally draw some conclusions. \section{PROBLEM STATEMENT}\label{sec:ps} In OCO, at each time $t$ a convex cost functional $f_t$ is revealed to a player, that plays its action $x_{t+1}$. A suitable performance metric to evaluate the strategy is the so-called dynamic regret, which is defined as follows \cite{mok16}: \begin{equation*} \mathrm{\mathbf{Reg}}^d_T (z_1,\dots,z_T):=\sum_{t=1}^{T}{f_{t}(x_{t})-f_{t}(z_t)} \end{equation*} where \begin{equation}\label{def:zt} z_t:=\argmin{x\in\mathcal{X}}f_{t}(x). \end{equation} where $\mathcal{X}\subseteq \mathbb{R}^n$ is a suitable state space. We specify that the notion of static regret is considered in some works, where the desired target is the minimum of $\sum_t f_t$ \cite{mok16}. Dynamic regret instead is focused on tracking, which is the case of our interest. We now formulate the problem using a sparse signal processing notation (the analogy with system identification will be clear in Section \ref{sec:nr}). We consider a time-varying sparse signal $v_t\in\mathbb{R}^n$, $t=1,\dots, T$, $T\in\mathbb{N}$, \emph{i.e.}, at each time step $t$, $v_t$ has $k_t\ll n$ non-zero components. We assume that linear measurements \begin{equation}\label{acqisition} y_t=A_t v_t+e_t,~~A_t\in\mathbb{R}^{m, n} \end{equation} are available, where $e_t$ is a possible measurement noise. Underdetermined problems with $m<n$ (as in the CS setting) are envisaged. It is well known that the problem of sparse signal recovery can be recast into a convex optimization problem, leveraging the fact that $\ell_1$ norm promotes sparsity \cite{for10}. Specifically, minimizing $\frac{1}{2}\left\|y_t-A_t x\right\|_2^2+\lambda\left\|x\right\|_1$, $\lambda>0$ produces sparse solutions that are consistent with measurements: this is the popular Lasso problem \cite{tib96}. In this paper, following the rationale of \cite{mok16}, we propose to use the (strongly convex) Elastic-net model \cite{uma10}, which adds a Tikhonov regularization term to the Lasso and reads as follows: \begin{equation}\label{eq:elasticnet} \begin{split} &f_t(x):=\frac{1}{2}\left\|y_t-A_t x\right\|_2^2+\lambda\left\|x\right\|_1+\frac{\mu}{2}\left\|x\right\|_2^2\\ &\lambda>0, \mu>0, t=1,\dots,T.\\ \end{split} \end{equation} A preliminary version of this model (with constant $A_t=A$) was introduced in \cite{fox17}. From a practical viewpoint, Tikhonov regularization promotes a grouping effect of correlated variables and is preferred in those sparse applications where it is undesired to discard correlated variables. In other terms, if two columns of $A_t$ are strongly correlated, Lasso would discard one of them, while Elastic-net would preserve both. For simplicity, we assume that the weight $\lambda$ is constant in $t$, which intrinsically assumes that the sparsity level $k_t$ is not significantly changing in time. Our goal is to solve the following optimization problem: \begin{equation}\label{eq:theproblem} \text{For any } t=1,\dots,T,~~~ \min_{x\in\mathbb{R}^n} f_t(x). \end{equation} Since each $f_t$ is strongly convex, in principle a convex optimization algorithm can be performed at each time step to get the desired minimum. Nevertheless, in the practice such algorithms are too slow to provide a solution in real time, in particular in case of large $n$. For this motivation, we propose to investigate a low-complexity iterative method that can online estimate $v_t$. Such method is expected to be suboptimal, but will turn out to be successful in terms of dynamic regret and efficient in numerical experiments. \section{PROPOSED METHOD}\label{sec:pm} Our idea is to adapt IST for problem \eqref{eq:theproblem} and analyze its dynamic regret. IST is a gradient-based algorithm which has been proved to converge to a minimum of Lasso \cite{for10} and to the minimum of the Elastic-net \cite{uma10}. Our goal is to prove that it can be used successfully also in the dynamic case. The choice of IST has different motivations. On the one hand, its simplicity allows a straightforward implementation and makes its theoretical analysis affordable. On the other hand, IST is prone to decentralization and parallel implementation \cite{rfm15,fox14,fox16,fia13}. In this work, we deal with a centralized setting, but the extension to a distributed setting (for example, a sensor network) might be investigated in the future. A drawback of IST is its convergence slowness, in terms of number of iterations, which makes other iterative algorithms more popular, firstly the alternating direction method of multipliers (ADMM, \cite{boy10}). The implementation and the analysis proposed in this paper for IST can be extended to ADMM, which will be object of future work. IST for (static) Elastic-net was derived in \cite{uma10}. Since our derivation is slightly different, we now describe it in detail. \subsection{Batch IST for Elastic-net} Let us define: \begin{equation*} \begin{split} &g(x):=\frac{1}{2}\left\| y-Ax\right\| _{2}^{2}+\frac{\mu}{2}\|x\|_{2}^{2};~~f(x):=g(x)+\lambda\|x\|_{1}\\ &s(x,b):=\frac{1}{2\tau}\left\| x-b\right\| _{2}^{2}-\frac{1}{2}\left\| Ax-Ab\right\| _{2}^{2}\\ &f(x,b):=f(x)+s(x,b)\\ \end{split} \end{equation*} where $x,b\in\mathbb{R}^{n}$, $A\in\mathbb{R}^{m, n}$, $y\in\mathbb{R}^m$, and $\tau>0$ is chosen so that $\tau\|A\|_{2}^{2}<1$; in this way, $s(x,b)\geq0$. As a consequence, following \cite{for10}, we can easily see that minimizing $f(x)$ is equivalent to minimizing the so-called surrogate functional $f(x,b).$ In particular, if $z$ is the minimum of $f(x)$, then $(z,z)$ is a minimum for $f(x,b)$. Adding $s(x,b)$, we cancel the term $\|Ax\|_{2}^{2}$, then, when $b$ is fixed the problem $\min_{x\in\mathbb{R}^{n}}f(x,b)$ is separable in the single components of $x$ and its solution can be written in closed form. In fact, the terms of $f(x,b)$ depending on $x$ can be grouped as follows: \begin{align*} & \frac{1}{2}\left(\mu+\frac{1}{\tau}\right)\|x\|_{2}^{2}-\langle x,\frac{b}{\tau}+A^{T}(y-Ab)\rangle+\lambda\|x\|_{1}=\\ & =\frac{1}{2}\left(\mu+\frac{1}{\tau}\right)\left\|x-\frac{b+\tau A^{T}(y-Ab)}{1+\mu\tau}\right\|_{2}^{2}+\lambda\|x\|_{1}+c \end{align*} where $c\in\mathbb{R}$ does not depend on $x$. We then conclude: \begin{equation} \argmin{x\in\mathbb{R}^{n}}f(x,b)=\mathbb{S}_{\frac{\lambda\tau}{\mu\tau+1}}\left[\frac{b+\tau A^{T}(y-Ab)}{1+\mu\tau}\right] \end{equation} where $\mathbb{S}_{\beta}:\mathbb{R}^{n}\to\mathbb{R}^{n}$, $\beta>0$, is the well-known component-wise soft thresholding operator \cite{for10}, defined as follows: for $z\in\mathbb{R}$, $\mathbb{S}_{\beta}[z]=z-\beta$ if $z>\beta$; $\mathbb{S}_{\beta}[z]=z+\beta$ if $z<-\beta$; $\mathbb{S}_{\beta}[z]=0$ otherwise. On the other hand, fixed $x$, it is easy to check that $x=\argmin{b\in\mathbb{R}^{n}}f(x,b)$. Alternating the minimizations in $x$ and $b$, we obtain Algorithm \ref{alg:IST_static}, which converges to the minimum of $f(x)$ (this can be easily proved exploiting the contractivity). Practically, the algorithm stops when a suitable numerical convergence criterion is met. Notice that in this batch procedure the time $t$ only refers to the algorithm's iterations. \begin{algorithm} \caption{Batch IST for Elastic-net} \label{alg:IST_static} \begin{algorithmic}[1] % \STATE $x_0=0$ % \FOR{$t=1,\dots,T_{stop}$} % \STATE $x_t=\mathbb{S}_{\frac{\lambda\tau}{\mu\tau+1}}\left[\frac{x_{t-1}+\tau A^{T}(y-Ax_{t-1})}{1+\mu\tau}\right]$ % \ENDFOR \end{algorithmic} \end{algorithm} \subsection{Online IST for dynamic Elastic-net} In order to tackle the dynamic problem \eqref{eq:elasticnet}-\eqref{eq:theproblem}, we propose an online version of Algorithm \ref{alg:IST_static}, which performs an IST step at each time $t$. This is summarized in Algorithm \ref{alg:IST_dynamic}, which will be theoretically analyzed in Section \ref{sec:ra}. If the computational resources allow it, one could perform $r>1$ IST steps at each $t$, as we will discuss in Section \ref{sec:nr}. \begin{algorithm} \caption{Online IST for dynamic Elastic-net} \label{alg:IST_dynamic} \begin{algorithmic}[1] % \STATE $x_0=0$ % \FOR{$t=1,\dots,T$} % \STATE $x_t=\mathbb{S}_{\frac{\lambda\tau}{\mu\tau+1}}\left[\frac{x_{t-1}+\tau A_t^{T}(y_t-A_tx_{t-1})}{1+\mu\tau}\right]$ % \ENDFOR % \end{algorithmic} \end{algorithm} \subsection{Related literature} We observe that online IST is an instance of Composite Objective Mirror Descent (COMID, \cite{duc10}), a popular technique in machine learning. Developed in the same years in different frameworks, both IST and COMID come from the forward-backward techniques. COMID was introduced to tackle online optimization with a regularizer term, \emph{i.e.}, the minimization of cost functionals of kind $\sum_{t}f_{t}(x)+r(x)$, where $r(x)$ is a (static) regularizer. $f_{t}$ and $r$ are both supposed to be convex. Its update is given by $x_{t+1}=\argmin x\langle f'_{t}(x_{t}),x\rangle+B_{\psi}(x,x_{t})+\eta r(x)$, where $B_{\psi}$ is a Bregman divergence. In \cite{duc10}, the static regret was proved to behave as $\mathcal{O}(\sqrt{T})$ and $\mathcal{O}(\log T)$, for convex and strongly convex functions, respectively. The assumptions to obtain this result, however, include the boundedness of $\|f'_{t}\|$ and $\eta$ vanishing as $\frac{1}{\sqrt{T}}.$ These assumptions are not required in this paper. In particular, vanishing parameters are generally not desirable when the time horizon $T$ is large or tends to infinity. \section{REGRET ANALYSIS}\label{sec:ra} In this section, we prove that $\mathrm{\mathbf{Reg}}^d_T$ for Online IST for dynamic Elastic-net is sublinear; in particular, it only depends on the evolution of $z_t$ \eqref{def:zt} (or similarly, on the evolution of $A_t$ and $v_t$ \eqref{acqisition}). We now provide some lemmas that build the proof of our main result. At the end of the section, we discuss some consequences of it. In the following, we assume that the evolution of the system is bounded. \begin{assumption} For $t=1,\dots,T$, $\tau \|A_t\|_2^2\leq 1$, and $\|v_t\|_2\leq v_M$ for some $v_M>0$. \end{assumption} As a consequence, also $\|z_t\|_2<z_M$ for some $z_M>0$. We remark that this an assumption on the system's evolution, while we do not force any boundedness on the algorithm's evolution $x_t$ or $f_t(x_t)$. This is an improvement with respect to \cite{mok16}, where the boundedness of $\nabla f_t$ was required \cite[Assumption 3]{mok16}, which for example is not satisfied for quadratic cost functions, unless a bounded state space is assumed \cite[Example in Section V]{mok16}. Given any $x\in\mathbb{R}^n$, let us define \begin{equation*} \Gamma_t(x):=\mathbb{S}_{\frac{\lambda\tau}{\mu\tau+1}}\left[\frac{x+\tau A_t^{T}(y_t-A_tx)}{1+\mu\tau}\right]. \end{equation*} \begin{lemma}\label{lem:contr}[Contractivity] Let $z_t=\argmin{x\in\mathbb{R}^n} f_t(x)$. Then, for any $x\in\mathbb{R}^n$, $$\|\Gamma_t(x)-z_t\|_2\leq \frac{1}{1+\mu\tau}\|x-z_t\|_2.$$ \end{lemma} \begin{proof} Since $z_t$ is the minimum, it is also a stationary point: $\Gamma_t(z_t)=z_t$. Since $\mathbb{S}_{\beta}$ is non-expansive, we have: \begin{align*} &\|\Gamma_t(x)-z_t\|_2 =\|\Gamma_t(x)-\Gamma_t(z_t)\|_2\\ & \leq\left\|\frac{x+\tau A_t^{T}(y_t-A_t x)}{1+\mu\tau}-\frac{z_t+\tau A_t^{T}(y_t-A_tz)}{1+\mu\tau}\right\|_2\\ & \leq\frac{\left\|I-\tau A_t^T A_t \right\|_2}{1+\mu\tau}\|x-z_t\|_2=\frac{1}{1+\mu\tau}\|x-z_t\|_2 \end{align*} where $I\in\mathbb{R}^{n, n}$ is the identity matrix. \end{proof} Using the contractivity, we can prove that, at a given $t$, the distance between the played action and the current minimum is controlled by the distance between successive minima. Let $$\Delta_t:=\|z_t-z_{t-1}\|_2.$$ \begin{lemma}\label{lem:somme} For any $t=1,\dots,T$, \begin{equation*} \begin{split} (a)~~&\sum_{t=2}^T \left\|x_t -z_t\right\|_2\leq c_1 + c_2 \sum_{t=2}^T \Delta_t\\ (b)~~&\sum_{t=2}^T \left\|x_t -z_t\right\|^2\leq c_3 + c_4 \sum_{t=2}^T \Delta_t^2+c_5 \sum_{t=2}^T \Delta_t\\ \end{split} \end{equation*} where $c_i>0$, $i=1,\dots,5$ are assessed in the proof. \end{lemma} \begin{proof} Using the triangle inequality and Lemma \ref{lem:contr}, for any $t=2,\dots,T$, \begin{equation*} \begin{split} \|x_t-z_t\|_2&\leq \|x_t-z_{t-1}\|_2 +\Delta_t\\ &\leq \frac{1}{1+\mu\tau}\|x_{t-1}-z_{t-1}\|_2 +\Delta_t.\\ \end{split} \end{equation*} Summing over $t=2,\dots,T$, \begin{equation*} \begin{split} &\frac{\mu\tau}{1+\mu\tau}\sum_{t=2}^T \|x_t-z_t\|_2\leq \frac{\|x_{1}-z_{1}\|_2 - \|x_{T}-z_{T}\|_2}{1+\mu\tau}+\\&~~~ + \sum_{t=2}^T \Delta_t. \end{split} \end{equation*} We have then proved $(a)$ with \begin{equation*} \begin{split} c_1&= \frac{1}{\mu\tau}\left(\|x_{1}-z_{1}\|_2 - \|x_{T}-z_{T}\|_2\right);\\ c_2&=\frac{1+\mu\tau}{\mu\tau}.\\ \end{split} \end{equation*} To prove $(b)$, we use: \begin{equation*} \begin{split} &\|x_t-z_t\|_2^2\leq \|x_t-z_{t-1}\|_2^2 +\Delta_t^2+2\|x_t-z_{t-1}\|_2\Delta_t \\ &\leq \frac{\|x_{t-1}-z_{t-1}\|_2^2}{(1+\mu\tau)^2} +\Delta_t^2+\frac{4 z_M}{1+\mu\tau}\|x_{t-1}-z_{t-1}\|_2.\\ \end{split} \end{equation*} Summing over $t=2,\dots,T$, we obtain $(b)$ with \begin{equation*} \begin{split} c_3&= \frac{\|x_{1}-z_{1}\|_2^2 - \|x_{T}-z_{T}\|_2^2}{\mu\tau(\mu\tau+2)}+\\&~~~+\frac{4 z_M}{1+\mu\tau}\left(\|x_{1}-z_{1}\|_2 - \|x_{T}-z_{T}\|_2\right);\\ c_4&=\frac{(1+\mu\tau^2)}{\mu\tau(\mu\tau+2)};~~~c_5=\frac{4 z_M}{1+\mu\tau}. \end{split} \end{equation*} \end{proof} \begin{lemma}\label{lem:surro} $$f_{t-1}(x_t)-f_{t-1}(z_{t-1})\leq \frac{1}{\tau}\left\| x_{t-1}-z_{t-1}\right\|_2^2.$$ \end{lemma} \begin{proof} Notice that $$f_{t-1}(x_t) \leq f_{t-1}(x_t,x_{t-1}) \leq f_{t-1}(z_{t-1},x_{t-1}),$$ using the fact that $x_t$ minimizes $f(\cdot, x_{t-1})$. Now, \begin{equation*} \begin{split} f_{t-1}(z_{t-1},x_{t-1})-f_{t-1}(z_{t-1})&=s(x_{t-1},z_{t-1})\\ &\leq\frac{\left\| x_{t-1}-z_{t-1}\right\|_2^2}{\tau}. \end{split} \end{equation*} \end{proof} \begin{lemma}\label{lem:d_t} Let \begin{equation*} \begin{split}d_t(x)&:=f_t(x)-f_{t-1}(x)\\&=\frac{1}{2}\left\|A_t x-y_t\right\|_2^2-\frac{1}{2}\left\|A_{t-1} x-y_{t-1}\right\|_2^2. \end{split} \end{equation*} Then, \begin{equation*} \begin{split} d_t(x_t)-d_t(z_t)\leq \gamma_1 \left\| x_t-z_t\right\|_2 +\gamma_2 \left\| x_t-z_t\right\|_2^2 \end{split} \end{equation*} where $\gamma_1>0, \gamma_2>0$ are assessed in the proof. \end{lemma} \begin{proof} Observing that $\left\| x_t+z_t\right\|_2\leq \left\| x_t-z_t\right\|_2+2\left\|z_t\right\|_2$, straightforward computation leads to the following bound: \begin{equation*} \begin{split} &d_t(x_t)-d_t(z_t)\leq \left\| x_t-z_t\right\|_2 \left\| A_t^T y_t -A_{t-1}^T y_{t-1}\right\|_2\\&~~+ \frac{1}{2} \left\| x_t-z_t\right\|_2 \left\| A_t^T A_t-A_{t-1}^T A_{t-1}\right\|_2 \left\| x_t+z_t\right\|_2\\ & \leq \frac{2 v_M + 2 z_M}{\tau}\left\| x_t-z_t\right\|_2 + \frac{1}{\tau} \left\| x_t-z_t\right\|_2^2. \end{split} \end{equation*} \end{proof} Based on these lemmas, we can prove our main result. \begin{theorem}\label{theo:regret} $$ \mathrm{\mathbf{Reg}}^d_T \leq \alpha_0+\alpha_1 \sum_{t=2}^T \left\|z_t-z_{t-1}\right\|_2 + \alpha_2 \sum_{t=2}^T \left\|z_t-z_{t-1}\right\|_2^2$$ where $\alpha_i>0$, $i=0,1,2$ are assessed in the proof. \end{theorem} \begin{proof} First, we compute a bound for the loss: \begin{equation*} \begin{split} &f_t(x_t)-f_t(z_t)=\\&=f_t(x_t)-f_t(z_t)\pm f_{t-1}(x_t)+ f_{t-1}(z_{t})- f_{t-1}(z_{t-1})\\ &\leq d_t(x_t)-d_t(z_t)+f_{t-1}(x_t)-f_{t-1}(z_{t-1})\\ \end{split} \end{equation*} where we have used the fact that $ f_{t-1}(z_{t})\leq f_{t-1}(z_{t-1})$. Now, applying Lemma \ref{lem:d_t} and Lemma \ref{lem:surro} we get: $f_t(x_t)-f_t(z_t)\leq \gamma_1 \left\| x_t-z_t\right\| +\gamma_2 \left\| x_t-z_t\right\|_2^2+ \frac{1}{\tau}\left\| x_{t-1}-z_{t-1}\right\|_2^2.$ Finally, summing over $t=2,\dots,T$ and exploiting Lemma \ref{lem:somme}, we obtain the thesis with $$\alpha_0= \left(\frac{1}{\tau}+\gamma_2\right)+\gamma_1 c_1+c_3+\frac{1}{\tau}\left\|x_1-z_1\right\|_2^2-\frac{1}{\tau}\left\|x_T-z_T\right\|_2^2,$$ $$\alpha_1=\left(\frac{1}{\tau}+\gamma_2\right) c_5+\gamma_1 c_2,~~~\alpha_2=\left(\frac{1}{\tau}+\gamma_2\right)c_4.$$ \end{proof} We notice that Theorem \ref{theo:regret} can be reformulated by highlighting that the dynamic regret is controlled by the evolution of the system variables $v_t, A_t$ (instead of $z_t$). \begin{corollary}\label{cor:regret} $$\mathrm{\mathbf{Reg}}^d_T \leq \alpha_0+\alpha_1 \sum_{t=2}^T \left\|\theta_t\right\| + \alpha_2 \sum_{t=2}^T \left\|\theta_t\right\|_2^2,$$ where $\theta_t= \frac{(z_M+v_M) \left\| A_t^T A_t-A_{t-1}^T A_{t-1}\right\|_2}{\mu} + \frac{1}{\mu\tau}\left\|v_t-v_{t-1}\right\|_2.$ \end{corollary} \begin{proof} We prove that $\Delta_t=\|z_t-z_{t-1}\|_2\leq \theta_t$. Given this, the thesis follows from Theorem \ref{theo:regret}. At any $t$, using the stationarity of $z_t$ and $z_{t-1}$ for $\Gamma_t$ and $\Gamma_{t-1}$, respectively: \begin{equation*} \begin{split} &\Delta_t \leq \frac{\left\|(I-\tau A_t^t A_t)z_t-(I-\tau A_{t-1}^t A_{t-1})z_{t-1}\right\|_2}{1+\mu\tau}\\&~~+\tau\frac{\left\|A_t^T A_t v_t-A_{t-1}^T A_{t-1} u_{t-1} \right\|_2}{1+\mu\tau}.\\ \end{split} \end{equation*} Adding and subtracting $(I-\tau A_{t-1}^t A_{t-1})z_{t}$ and $ \tau A_{t-1}^T A_{t-1}v_t$ within the two norms, respectively, we obtain: \begin{equation*} \begin{split} &\Delta_t \leq \frac{(z_M+v_M) \tau \left\| A_t^T A_t-A_{t-1}^T A_{t-1}\right\|_2}{1+\mu\tau}\\&~~ + \frac{\left\|z_t- z_{t-1}\right\|_2+\left\|v_t- u_{t-1}\right\|_2}{1+\mu\tau}.\\ \end{split} \end{equation*} Therefore, \begin{equation*} \begin{split} &\Delta_t \leq \frac{(z_M+v_M) \left\| A_t^T A_t-A_{t-1}^T A_{t-1}\right\|_2}{\mu}\\& + \frac{\left\|v_t- v_{t-1}\right\|_2}{\mu\tau}. \end{split} \end{equation*} \end{proof} \begin{remark} The regret analysis can be performed in an analogous way if $r>1$ IST steps are played at each time step. Starting from Lemma \ref{lem:contr}, we would have a stronger contractivity with contraction factor $(1+\mu\tau)^{-r}$, which would lead step by step to tighter constant parameters in Theorem \ref{theo:regret}. Since the purpose of this work is to describe the general behavior, a precise analysis of the improvement obtained with $r>1$ is left to future work. \end{remark} \subsection{Consequences} Theorem \ref{theo:regret} and Corollary \ref{cor:regret} state a regret bound depending uniquely on the behavior of the system (and not directly on $T$) in terms of sum of (linear and quadratic) distances between successive iterates (in terms of minima $z_t$, or $A_t$ and $v_t$). This is in line with the results of \cite{mok16}, which obtained similar (linear) bounds for generic strongly convex problems, under stronger assumptions. In particular, as in \cite{mok16}, we can observe that: \begin{itemize} \item[-] if $z_t=z$ (or $A_t=A$ and $v_t=v$) for any $t>t_0$, then $\mathrm{\mathbf{Reg}}^d_T=\mathcal{O}(1)$, which implies that $x_t$ converges to $z$; \item[-] if $\|z_t-z_{t-1}\|_2=\mathcal{O}\left(\frac{1}{t^\eta}\right)$ (or $\|A_t^T A_t-A_{t-1}^T A_{t-1}\|_2=\mathcal{O}\left(\frac{1}{t^\eta}\right)$ and $\|v_t-v_{t-1}\|_2=\mathcal{O}\left(\frac{1}{t^\eta}\right)$), with $\eta>0$, then $x_t$ converges to $z_t$; \item[-] if $\|z_t-z_{t-1}\|_2\leq C$ (or $\|A_t^T A_t-A_{t-1}^T A_{t-1}\|_2\leq C$ and $\|v_t-v_{t-1}\|_2\leq C$) where $C$ is constant, we have a steady state estimation error. \end{itemize} \section{NUMERICAL RESULTS}\label{sec:nr} \begin{figure*}[ht] \centering \includegraphics[width=0.98\columnwidth]{opwdanew.pdf}\quad \includegraphics[width=0.98\columnwidth]{pwdbnew.pdf} \caption{Estimation of $a_{1,t}$ and $b_{1,t}$, with online IST with $r=1000$ and $r=100$.} \label{fig:simulations} \end{figure*} \begin{figure}[ht] \centering \includegraphics[width=0.98\columnwidth]{supportnewnew.pdf} \caption{Signed error $\widehat{v}_{i,sm}-v_{i,sm}$ ($m=15$, $s=1,\dots,66$) for the $P+Q=20$ possible parameters (0 and 10 respectively are $a_{1,t}$ and $b_{1,t}$, while the others are the null parameters).} \label{fig:support} \end{figure} In this section, we present the implementation of the proposed method to tackle the online estimation of time-varying parameters in a time-varying autoregressive model with an exogenous input (TVARX, \cite{li11}). In a TVARX model, the input-output relationship is as follows: $y_t=\sum_{p=1}^P a_{p,t}y_{t-p}+\sum_{q=1}^Q b_{q,t}u_{t-q}+e_t$, where $u_t,y_t\in\mathbb{R}$ respectively are the measurable input and output; $e_t\in\mathbb{R}$ is the measurement error; $a_{p,t}, b_{q,t}\in\mathbb{R}$ are the time-varying parameters to be estimated. The goal is to perform a recursive identification, that is, to estimate $a_{p,t}$ and $b_{q,t}$ at time $t$, exploiting the knowledge of $u$ and $y$ and an estimation at time $t-1$. In our simulations, we assume that the dimensions $P$ and $Q$ are not known. A suitable strategy to overcome this lack is to fix a sufficiently large upper bound for them, and then look for a sparse solution, that is, to the most parsimonious model that represents the system accurately. The model and methodology presented in Sections \ref{sec:ps} and \ref{sec:pm} perfectly match this purpose. More precisely, we consider the following setting: we iteratively collect groups of $m$ output measurements $\mathbf{y}_t:=(y_t,\dots, y_{t+m})^T$, and we run $r$ steps of the online IST algorithm. It is easy to check that we can define the time-varying matrix $A_t\in\mathbb{R}^{m,P+Q}$ as follows: \begin{equation*} \left(\begin{array}{cccccc} y_{t-1}&\cdots&y_{t-P}&u_{t-1}&\cdots&u_{t-Q}\\ y_{t}&\cdots&y_{t-P+1}&u_{t}&\cdots&u_{t-Q+1}\\ \vdots&&&&&\vdots\\ y_{t+m-1}&\cdots&y_{t+m-P}&u_{t+m-1}&\cdots&u_{t+m-Q}\\ \end{array}\right). \end{equation*} Assuming $m$ sufficiently small so that $a$ and $b$ are constant between $t$ and $t+m$, we formulate the problem as follows: at each $t=sm$, $s\in\mathbb{N}$, we collect $m$ measurements $\mathbf{y}_t$ and given $A_t$ we aim to recover $$v_t=(a_{1,t},\dots, a_{P,t}, b_{1,t},\dots,b_{Q,t})^T$$ from $$\mathbf{y}_t= A_t v_t +\mathbf{e}_t$$ where $\mathbf{e}_t:=(e_t,\dots, e_{t+m})^T$. $v_t$ is assumed to be sparse with respect to overestimated dimension $P+Q$. For our simulations, we retrieve the TVARX(1,1) example considered in \cite[Section V]{li11}: $y_t=a_{1,t}y_{t-1}+b_{1,t}u_{t-1}+e_t$, with $P=Q=1$. As mentioned, we assume $P$ and $Q$ unknown and we overestimate them as $P=Q=10$. In terms of sparse signal recovery, we can say that we have to online estimate a time-varying sparse signal $v_t\in\mathbb{R}^n$ with $n=20$ and (unknown) sparsity $k=2$. Such sparse signal has constant support (but this information is assumed to be unknown and then not exploited in our procedure). We consider a time horizon of 1 second and sampling frequency of 1000 Hz. We assume that $a_{1,t}$ and $b_{1,t}$ are step-wise constant, with some abrupt changes (with respect to the example in \cite{li11}, we consider larger gaps). Specifically, we set: {\small{ \begin{equation*} a_1(t)=\left\{\begin{split} &-0.9 \text { if } t<0.5\\ &0.9 \text{ otherwise.} \end{split}\right. \end{equation*} \begin{equation*} b_1(t)=\left\{\begin{split} &0.7 \text { if } t<0.2\\ &-0.8 \text { if } 0.2 \leq t<0.4\\ &0.8 \text { if } 0.4 \leq t<0.7\\ &-0.7 \text{ otherwise.}\\ \end{split}\right. \end{equation*} }} The input $u$ is assumed to be a standard Gaussian sequence, periodic with period $m$ (so that the right part of $A_t$ is constant for any $t=ms$, $s\in\mathbb{N}$). The measurement noise is white Gaussian with SNR around 20dB. Finally, we assume $m=15$, therefore, as in CS, we have less measurements than parameters to estimate. In the literature, this is also known as Compressive System Identification \cite{tot11,san11}. In \cite{tot11}, the case of constant parameters was considered, while in \cite{san11} an extension to piecewise-constant parameters was proposed. $A_t$ is not a typical CS sensing matrix, but the randomness introduced by the chosen $u$ and the circulant structure are promising features to use it in a CS framework \cite{for10}. In \cite{tot11,san11}, the properties of $A_t$ for CS where theoretically studied, but the obtained bounds are not tight \cite[Section IV.C]{san11}. This point is then still open and will be considered for future work, as well as a thorough comparison to the approach proposed in \cite{san11}. Coming back to our experiment, we run online IST with the following design parameters: $\lambda=2\times 10^{-2}$, $\mu=10^{-6}$, $\tau=3\times 10^{-2}$; the initial condition is assumed to be zero. At each $t=ms$, we run $r$ IST steps, with $r=1000$ and $r=100$. Simulations have been performed on a 2.67 GHz CPU, where $r=1000$ and $r=100$ respectively require 10 ms and 2 ms. Since we consider blocks of $m=15$ measurements (acquired in 15 ms), in both cases the algorithm does not exceed the time of acquisition. The overall delay for acquiring the measurements and running the simulations is then around 25 ms and 17 ms, respectively. We specify that performing a complete IST at each block (and then get the minimum of $f_t$) is not feasible in real time, since the convergence on this problem requires up to 35000 iterations and 500 ms for execution on the considered CPU. We have performed 250 runs, whose average results are shown in Table \ref{tab:0}. The mean square error is defined as MSE=$\frac{1}{P+Q}\sum_{s=1}^{T/m}\sum_{i=1}^{P+Q}\left\|v_{i,sm}-\widehat{v}_{i,sm}\right\|_2^2$, where $v_t=(a_{1,t},\dots, a_{P,t}, b_{1,t},\dots,b_{Q,t})^T$ and $\widehat{v}_t$ is our estimation. As expected, we can notice an average improvement when $r$ increases. \begin{table} \begin{center} \caption{Average MSE and standard deviation over 250 runs, with $r$ IST steps at each $sm$.} \resizebox{0.5\columnwidth}{!} { \begin{tabular}{| r | c | c |} \hline IST steps & Average MSE & Std \\ \hline $r=100$ & 0.011 & 0.002\\ $r=1000$ & 0.006& 0.001\\ \hline \end{tabular} \label{tab:0}} \end{center} \vspace{-0.43cm} \end{table} In Figure \ref{fig:simulations}, we show the estimation of $a_1$ and $b_1$ in a single run. We can appreciate that when $r=1000$, $a_1$ and $b_1$ are generally well approximated and the abrupt changes are promptly detected. Some picks are visible (for instance, immediately after 0.4 ms in the $a_1$ graph), due to locally large noise (some oscillations were highlighted also in \cite{li11}). In the case $r=100$, instead the changes are less promptly detected (in particular, for $a_1$). However, a good approximation is achieved after some steps. Moreover, a smaller number of iterations makes the algorithm more conservative, thus less affected by locally large noise: for example, the pick after 0.4 ms is canceled. In Figure \ref{fig:support}, we show the signed error $\widehat{v}_{i,sm}-v_{i,sm}$, $s=1,\dots,66$. The values 0 and 10 on the $x$-axis respectively represent $a_t$ and $b_t$; the others are the null ones. Green circles and blue squares represent the estimations, at each $sm$, respectively for $r=100$ and $r=1000$. For $r=1000$, $\widehat{v}_{i,sm}-v_{i,sm}$ is smaller on the null parameters if compared to non-null parameters $a_1$ and $b_1$. This means that the support (\emph{i.e.}, the positions of the non-null components in the vector $v_t$) is well detected. \section{CONCLUSIONS} In this work, we have presented a method for online optimization for time-varying sparse problems. We have considered an Elastic-net cost functional and proposed an online IST algorithm. We have then proved that this method is successful in terms of dynamic regret. Moreover, we have shown how to apply the algorithm to a problem of recursive identification for a TVARX model, and we have presented a few numerical results. In future work, other algorithms (\emph{e.g.}, ADMM) will be considered for this problem and other sparse (possibly not strongly convex) models will be investigated. \bibliographystyle{plain}

5f428b2e109678de92f5409f17149eed8a26beb0

\section*{Acknowledgements} This work is supported in part by the Grant-in-Aid for Scientific Research No.25287049 (M.K.), No.15K05055 (M.S.) and No.25400260 (H.S.) by the Japanese Ministry of Education, Science, Sports and Culture.

4ed34f15777bbab69c2d5bdd1b9c8a873338266b

\section*{Content} \section*{Background} \subsection*{Introduction} The knowledge graph (KG) has become the preferred data model for complex knowledge domains. Accordingly Wilcke et al. published: "The knowledge graph as the default data model for learning on heterogeneous knowledge."\cite{wilcke2017kgdatamodel}. Biology and biomedical knowledge is complex and involves a plethora of entity and association types, hence is particularly suited to heterogeneous graph methodology. From such a KG, statistical knowledge can be inferred, for example, probabilistic associations between genes and phenotypic traits. In KG terms, node and edge semantics are varied and critical for precise representation of the knowledge. Methods which consider surrounding node and edge contexts support a rich and combinatorially expanding feature set. KG embedding connotes representation of entities as computable feature vectors amenable to machine learning (ML) methods\cite{goodfellow2016deep,cai2018grembed}. As both KG and ML methodology advances, the issues of embedding, representation and vectorization become crucial, as signaled by related research activity spanning computing, natural and social sciences\cite{cai2018grembed}. Deep learning is a powerful approach for representation learning on large graphs and datasets. Multi-layer\ deep neural networks entail transformations from input raw data to layered representations obviating the need for feature engineering up front. Instead a set of continuous, latent features (representations) are learned which, in the graph use case, encode localized structural topology around a given node facilitating prediction tasks based on network structure. Previous work has focused on using neural network learning models to generate node embeddings for graphs such as DeepWalk\cite{perozzi2014deepwalk}, LINE\cite{tang2015line}, and node2vec\cite{grover2016node2vec}. However these models were designed for homogeneous networks, which means that they do not explicitly encode information related to the types of nodes and edges in a heterogeneous network. Recently, metapath2vec\cite{dong2017metapath2vec} was proposed by incorporating metapaths with node semantics for node embedding learning. However, this approach has several drawbacks: 1) domain knowledge is required to define metapaths and those mentioned in \cite{dong2017metapath2vec} are symmetric paths which are unrealistic in many applications; 2) metapath2vec does not consider edge types rather only node types; and 3) metapath2vec can only consider one metapath at one time to generate random walk, it cannot consider all the metapaths at the same time during random walk. On another related track, which might be termed biomedical data science (BMDS), previous work has employed KG embedding and ML methodology with the focus on applicability and applications such as compound target bioactivity\cite{chen2012assessing, seal2015randomwalk} and disease-associated gene prioritization\cite{himmelstein2015}. Yet other efforts have simply employed off-the-shelf ML toolkits (e.g. Scikit-learn, WEKA) and methods to address biomedical informatics prediction challenges. To address the above problems, edge2vec was developed to consider edge semantics when generating node sequence using a random walk strategy. An edge-type transition matrix is defined to improve representation of node "context" and designed with an Expectation-Maximization (EM) model. In the maximization step, we use the transition matrix to generate node sequences based on random walk in a heterogeneous graph. In the expectation step, we use the generated node ‘context‘ from node embeddings as feedback to optimize the transition matrix. We also use a skip-gram sampling strategy to select partial nodes for the EM approach to make the edge2vec model run on large-scale networks to learn node embeddings in a more efficient way. In the end, the topologically similar nodes (with similar sub-structures or located near each other in the network) are with similar emebeddings; the semantically similar nodes (with same node-types or logistically related attributes) are with similar embeddings. Within biomedicine, the sciences involved in drug discovery are diverse. Drug efficacy and safety depend on calibrated modulation of complex, interrelated biomolecular pathways and targets. Prediction of compound-target bioactivity, normally non-covalent binding, remains high-challenge and high-value, both for generating novel drug leads and hypotheses, and for elucidating the mechanism of action for known compounds and drugs. With this rich knowledge domain as context, in this paper, we apply edge2vec on Chem2Bio2RDF\cite{chen2010chem2bio2rdf}, a highly heterogeneous graph integrating over 25 biomedical and drug discovery datasets. The contribution of our work is threefold. \begin{itemize} \item We define an edge-type transition matrix to represent network heterogeneity. The calculation of the matrix is mainly based on the path similarity of different edge-types. \item We develop an EM model to train a transition matrix via random walks on a heterogeneous graph as a unified framework and employ a stochastic gradient descent (SGD) method to learn node embedding in an efficient manner. The learned node vector can include not only the topological information of network structure, but also the edge type information, which indicates different relationships among nodes. \item We evaluate our model in the drug discovery domain by predicting drug-target associations using the highest available quality datasets as ground truth. Validation of the edge2vec model is addressed via three prediction tasks, all realistic biomedical discovery use cases. Validation results indicate that edge2vec adds value relative to existing methodology for drug discovery knowledge discovery. \end{itemize} In the following sections, first, we introduces edge2vec and its importance; second, we discusses related work about node embedding learning as well as heterogeneous network analysis; third, we explains edge2vec; fourth, we evaluates edge2vec based on later drug discovery; fifth, we illustrates two case studies to visualize edge2vec results, And in the end we concludes and points out future work. \subsection*{Related Work} \textbf{Network Representation:} Network representation is useful in a variety of applications such as network classification\cite{bhagat2011node,sen2008collective}, content recommendation\cite{fouss2007random,yu2014personalized,gao2018end}, community detection\cite{gao2017personalized,liu2016comparing,zhang2016others} and link prediction\cite{liben2007link}. Networks are easily and naturally represented by adjacency matrix, but such matrices are generally sparse and high dimension, thus not well suited to statistical learning\cite{perozzi2014deepwalk}. How to represent network information in low dimension is an important task. There are classical methods of network representation which is dimension reduction based on calculating eigenvector, such as LLE\cite{roweis2000nonlinear, saul2000introduction}, Laplacian Eigenmap\cite{belkin2002laplacian,tang2011leveraging}, MDS\cite{cox2000multidimensional}, IsoMap\cite{tenenbaum2000global}, and DGE\cite{chen2007directed}. However, these methods do not perform well in large-scale networks. \textbf{Representation Learning based on Deep Neural Network:} In deep learning, more and more encoder-decoder models have been proposed to solve network representation problems. By optimizing a deterministic distance measure, those models can learn a node embedding from its neighbor nodes so as to project nodes into a latent space with a pre-defined dimensionality. Recently, deep neural network\cite{collobert2008unified} based representation learning has been widely used in the natural language processing. Word2vec\cite{mikolov2010recurrent} is the deep learning model developed by Google to represent a word in a low dimension dense vector, which has proven to be successful in natural language processing\cite{pennington2014glove}. By close analogy, topological paths neighboring a node may be handled like sequences of words, and word2vec can be adapted to network representation learning to reduce computing complexity and improve performance relative to conventional approaches. Accordingly, several recent publications have proposed word2vec-based network representation learning frameworks, such as DeepWalk\cite{perozzi2014deepwalk}, GraRep\cite{cao2015grarep}, TADW\cite{cao2015grarep}, CNRL\cite{tu2016community}, LINE\cite{tang2015line}, node2vec\cite{grover2016node2vec}, and metapath2vec\cite{dong2017metapath2vec}. All of the above frameworks utilize the skip-gram model\cite{mikolov2013efficient,mikolov2013distributed} to learn a representation of a node incorporating its topological context, so nodes with similar topological information will have similar numerical representations. Node representations are learned via skip-gram model by optimizing the likelihood objective using SGD with negative sampling\cite{levy2015improving}. \textbf{Sampling Strategy:} Similar to word sequences from documents, node sequences may be sampled from the underlying network as an ordered sequence of nodes\cite{dong2017metapath2vec}. Accordingly, different network representation learning frameworks adopt different node sampling strategies. DeepWalk\cite{perozzi2014deepwalk} deploys a truncated random walk to sample node sequences, and uses the skip-gram model to learn the representation of node sequences. However, DeepWalk only considers the first-order proximity between nodes. Moreover, it applies to unweighted networks. Practically, LINE is applicable for both weighted and unweighted networks and easily scales to large-scale networks with millions of nodes. The problem is that embedding of some loosely-connected nodes, which have few connected edges, heavily depends on their connected neighbors and unconnected negative samples\cite{xu2017empirical}. Most prior methods do not give full consideration to heterogeneity of nodes and edges. Thus Deepwalk, LINE, and Node2vec are not effective for representing these heterogeneous networks. Sun et al.\cite{sun2011pathsim} introduced a metapath-based similarity measurement to find similar objects of the heterogeneous information networks. Furthermore, Dong et al. proposed metapath2vec\cite{dong2017metapath2vec} to capture heterogeneous structure and semantic correlation exhibited from large-scale networks by considering node types. However, one drawback of all previous methods is that they either only deal with homogeneous networks or do not consider edge semantics. When network contains nodes and edges with different types, the state-of-the-art embedding results are no longer effective as all of them do not consider edge semantics. To represent heterogeneity, we have developed edge2vec to learn node representations with general, systematic consideration of edge semantics. \textbf{Representation learning in biomedical domains:} In biomedical domains, there exist rich heterogeneous datasets about genes, proteins, genetic variations, chemical compounds, diseases, and drugs. Ongoing and expanding efforts to integrate and harness these datasets for data-driven discovery reflect widespread understandings of potential benefits to science and human health. For example, Chem2Bio2RDF\cite{chen2010chem2bio2rdf} integrates over 25 different datasets related to drug discovery and comprises a large scale heterogeneous network. Such repositories hold complex relationships between many entity types. Representing this semantic complexity requires suitable embedding methods informed by these rich domains. Chen et al.\cite{chen2012assessing} propose Semantic Link Association Prediction (SLAP) to predict ‘missing links‘ between drugs and targets in Chem2Bio2RDF. Subsequently, Fu et al.\cite{fu2016predicting} applied an improved PathSim method with more than 50 metapaths on an extended Chem2Bio2RDF dataset to predict drug target interactions and rank metapaths based on Gini index. Although these previous studies focus on relationships between drugs and targets, none handle edge heterogeneity in graph embedding directly and generally. We regard this as an unmet need, both for the methodological value in exploring edge semantics, and also for the more practical value of applying and validating this novel methodology via biomedical data science use cases. Thus motivated, we propose edge2vec as an improved representation learning model, well suited for discovery on biomedical knowledge graphs. \textbf{Context: applied machine learning and data science:} Machine learning is a big, diverse and rapidly advancing area of research, challenging to monitor and contextualize even to its scholars and practitioners. For biomedical data science, the challenge is compounded with applicability to complex real world datasets and tasks. This is applied machine learning, wherein the final evaluation depends on relevance and comprehensibility to the application domain. In this paper we strive to maintain this relevance and comprehensibility through concise, contextualizing notes and examples, with appropriate citations for further study. Terminology is a related challenge, for example the equivalence of "graph" and "network", so to assist we provide a glossary as supplementary material. \section*{Methods} \begin{figure} [width=0.95\columnwidth]{pipeline.png} \caption{ \csentence{An illustrative pipeline of edge2vec.}(a) a heterogeneous network with three types of nodes and two types of edges, colored by types. (b) EM framework to optimize an edge-type transition matrix $M$ and generate node random walks as well as related edge-type corpus. (c) skip-gram model is used for node embedding learning. For a node $v_{4}$, the input layer is its one-hot encoding and the output layer is the one-hot prediction for all its $K$ neighbor nodes (e.g. node $v_{1}$ and node $v_{10}$).} \label{fig:pipline} \end{figure} In this section, we introduce edge2vec. The pipeline is shown in Figure \ref{fig:pipline}. We treat heterogeneous network embedding learning as an optimization problem and design an EM framework associated with a skip gram model to solve it. See Algorithm ~\ref{al:pseudo} pseudo code for details. \subsection*{Edge-type transition matrix for network embedding} As word2vec\cite{mikolov2010recurrent} informed node2vec\cite{grover2016node2vec}, we can represent a node and its network neighborhood analogous to a word-context relationship in a text corpus. Random walk paths of nodes are akin to word sequences. We thereby convert the node embedding learning problem into a node neighborhood optimization problem: given a node, we need to maximize the probability of neighbor nodes, which is Formula 1: \begin{equation} \argmax_{\theta} \prod_{v \in V}\prod_{c \in N(v)} p(c|v;\theta) \label{eq:hnode2vec} \end{equation} where V refers to the node collection of the network G(V,E); N(v) refers to the neighbor node collection of node v; $\theta$ is the node embedding parameterization to be learned. However, this optimization only works well in homogeneous networks. As in heterogeneous networks, different types of nodes and edges occur with varying frequency. But low frequency node and edge types may be very important, depending on their semantics in the knowledge domain. For instance, in a scholarly citation network, venue nodes (i.e., conferences and journals) are fewer but more important than publication nodes. Since node2vec would treat all nodes equally, knowledge contained in the venue relationships would be lost. Likewise, throughout biomedical domains, node and edge semantics must be considered to avoid loss of critical knowledge. For one example, the edge relationship between an approved drug and its well validated protein target is highly and exceptionally informative, reflecting prodigious research efforts and expense. To address this need for edge semantics, we design an edge-type transition matrix which holds the transition weights between different edge types during the random walk process. Therefore, we consider not only the topological structure of the network but also edge semantics. Accordingly, the optimized version is shown in Formula 2: \begin{equation} \argmax_{\theta,M} \prod_{v \in V}\prod_{c \in N(v)} p(c|v;\theta;M) \label{eq:hnode2vec} \end{equation} M refers to the edge-type transition matrix. The matrix stores the random walk transition weights between different edge types. By employing the transition matrix as a prior distribution guiding the random walk process, we not only consider the distance between the next-step node and the previous-step node but also the weight between the next-step traversed edge type and the previous-step traversed edge type. Therefore, we can normalize by type so that the effect of low frequency node/edge types won't be lost by dilution among high frequency node/edge types. As shown above, the optimization function maximizes the probability of generating the node neighborhood of a given node v, thus the transition probability from the current node v to its neighbor c can be seen in Formula 3: \begin{equation} p(c|v;\theta;M) = \frac{e^{\vec{f_{v}} \cdot \vec{f_{c}}}}{\sum_{u \in V}e^{\vec{f_{u}} \cdot \vec{f_{c}}}} \label{eq:prob} \end{equation} where $\vec{f_i}$ means the current step embedding for node \textit{i} which will be updated in each batch. We calculate the inner product of two node embeddings, which are normalized by a Softmax function. We designed an EM framework to combine the update of the transition matrix M and optimization of node context into a unified framework. An edge-type transition matrix is initialized with all values set to 1, meaning initially, all edge type transitions are regarded as equally probable. Then, we iteratively generate the random walk corpus of paths, optimizing the transition matrix based on the sampled frequencies of edge type transitions. \begin{algorithm} \caption{edge2vec algorithm} \begin{algorithmic} \REQUIRE Graph$<V,E>$ $g$, Edge-type transition matrix $M$ \STATE initialize $walks$ empty, all values in $M$ as $1$, node embeddings $f$ \STATE $walks, M$ = GenerateTransitionMatrix($g,M$) \STATE $f$ = StochasticGradientDescent($walks$) \STATE \textbf{return} $f$ \\\hrulefill \STATE \textbf{GenerateTransitionMatrix}($g,M$) \STATE initialize \# of iteration N \WHILE{$N > 0$} \STATE $N \leftarrow N - 1$ \STATE \#E step \STATE $walks$ = HeteroRandomWalk($g,M$) \STATE \#M step \STATE $\vec{v_{i}}$ = vector with each dimension as the \# of edge type $i$ in each walk from $walks$ \STATE $M_{ij}$ = Sigmoid( PearsonCorrelation($\vec{v_{i}},\vec{v_{j}}$) ) \ENDWHILE \STATE return $walks$, M \STATE \hrulefill \STATE \textbf{HeteroRandomWalk}($g,M$) \FOR{node $n \in g$} \STATE initialize an empty node walk $w$, an empty edge walk $T$, given random walk length $l$ \STATE Append $n$ to $w$ \WHILE{length(w) < $l$} \IF{length$(w) == 1 $} \STATE $curr$ = $w[-1]$ \STATE Random sample node $m$ from Neighbour($curr$) based on edge weight. \STATE Append $m$ to $w$ \STATE Append EdgeType($curr,m$) to $T$ \ELSE \STATE $curr$ = $w[-1]$, $prev$ = $w[-2]$ \STATE $p_{1}$ = T[-1] \FOR{node $k \in$ Neighbour($curr$)} \STATE $p_{2}$ = EdgeType($k,curr$) \STATE $EW(k,curr) = M[p_{1}][p_{2}] \cdot W(k,curr) \cdot \alpha_{pq}(k,u)$ \#$W$(k,curr) is edge weight between node $k$ and $curr$. \STATE Random sample node $m$ from Neighbour($curr$) based on updated edge weight $EW(k,curr)$. \STATE Append $m$ to $w$ \STATE Append EdgeType($curr,m$) to $T$ \ENDFOR \ENDIF \ENDWHILE \ENDFOR \STATE \textbf{return} $T$ \end{algorithmic} \label{al:pseudo} \end{algorithm} \subsection*{Expectation-Maximization framework} \subsubsection*{Expectation step} Assume we have a set of E=$ \{ $ e\textsubscript{1}, e\textsubscript{2}, e\textsubscript{3}$ \ldots $ e\textsubscript{m}$ \} $ different edge types in a network. From the previous iteration in the EM framework, we can get a collection of random walk paths for each node as P = $ \{ $ p\textsubscript{1}, p\textsubscript{2}, $ \ldots $ p\textsubscript{n}$ \} $ . In each walk path p\textsubscript{i} (i $ \in $ $ \{ $ 1,2$ \ldots $ n$ \} $ ), it is constructed like p\textsubscript{i} = $ \{ $ n\textsubscript{1},n\textsubscript{2},n\textsubscript{3}, ...,n\textsubscript{l} $ \} $ where n\textsubscript{i} is the ith node in p\textsubscript{i} and l is a predefined walk length. Based on each path, we first extract all edges $ \{ $ T(n\textsubscript{1}, n\textsubscript{2}), T(n\textsubscript{2}, n\textsubscript{3}), $ \ldots $ , T(n\textsubscript{l-1}, n\textsubscript{l})$ \} $ in the path by locating every start node n\textsubscript{k} and end node n\textsubscript{k+1} where k $ \in $ $ \{ $ 1, 2, ..., l $-$ 1$ \} $ , e\textsubscript{k} = T(n\textsubscript{i}, n\textsubscript{j}) refers to the edge type between n\textsubscript{i} and n\textsubscript{j}. After that, we calculate the number of times each type of edge e\textsubscript{j} (e\textsubscript{j} $ \in $ E) appears in the walk path p\textsubscript{i}. The same calculation is applied to all walk paths. In the end, for each edge type e\textsubscript{j}, we get a vector representation v\textsubscript{j}, where the ith dimension in the v\textsubscript{j} refers to the number of times e\textsubscript{j} appears in walk path p\textsubscript{i} . One assumption of our model is for a pair of edge type e\textsubscript{1} and e\textsubscript{2}, the distribution of each edge type sampled from the random walk paths is a valid estimator for the transition correlation for the graph. Hence, by calculating the correlation between their associated vector v\textsubscript{i} and v\textsubscript{j} in the walks, we can regard the correlation score as their updated transition weight. Therefore, we can define the formula for updating transition matrix as Formula 4: \begin{equation} M(e_{i},e_{j}) = \text{Sigmoid}(\frac{E[\vec{(v_{i}}-\mu(\vec{v_{i}}))\vec{(v_{j}}-\mu(\vec{v_{j}}))]}{\sigma(\vec{v_{i}})\sigma(\vec{v_{j}})}) \label{eq:update} \end{equation} where $E[\cdot]$ is the expectation value and $\sigma$ is related standard derivation value. M(e\textsubscript{i} , e\textsubscript{j}) refers to the updated transition weight between edge type i and j. v\textsubscript{i} and v\textsubscript{j} are vector representation of e\textsubscript{i} and e\textsubscript{j} on all walk paths. By using Pearson correlation analysis, we can get a pairwise correlation score between two edge types to check the distribution difference. Larger weight value means larger correlation between the pair of edge types. However, as the range of the correlation score varies from -1 to +1, it makes no sense if we keep the original negative weights between a pair of edge types. Because we involve the optimized transition weights to the random walk probability, and the probability can’t be negative, thus we normalize by transformation to a Sigmoid function to solve this issue and restrict the transition probability in a range of between 0 and 1. Moreover this non-linear transformation can better help to capture the patterns of transition probability than other linear or quadratic transformation functions\cite{kedem2012non}. The definition of Sigmoid$(\cdot)$ is shown as Formula 5: \begin{equation} \text{Sigmoid}(x) = \frac{1}{1+e^{-x}} \label{eq:relu} \end{equation} In summary, the non-linear transformed correlation ensures three characteristics of the biased random walk on a heterogeneous network: First, a random walk tends to pass on edges with same edge-type. Based on the correlation calculation in Formula 4, given an edge-type, the correlation with itself is always +1, which is the highest transition weight inside the transition matrix. Second, the Sigmoid function guarantees optimization convergence of transition matrix. Transition weights are adjusted according to the Sigmoid function by training based on the correlations calculated from the random walks until a stable final value is reached. Third, only edge-types with closer relationships tend to have higher transition weights. Although some edge-types are \textit{globally} more common and likely to appear in random walks, we consider the \textit{specific} co-occurrence rates between edge-types in the same random walk. For example, if edge-type $a$ appears (20,30,40) times in three random walks, while edge-type $b$ appears (1,1,1) times in the same randoms walks. The transition weight from $b$ to $a$ is still considerably low. \subsubsection*{Maximization step} In each iteration in the EM framework, based on the updated edge-type transition matrix M in the expectation step and the network topological structure, the biased random walk process generates a new paths with information of nodes and node neighbors. The transition matrix contributes to the calculation of random walk probabilities, thereby including the influence of edge-type information in sampling, which can reduce the negative effects caused by skewed type distribution issues. Even though some types of edges appear less frequently in the network, if the transition weights between those edge-types and other edge-types are high, the edge still has a high probability to get visited during the random walk process. Another important feature is that based on the expectation step, as well as Formula 4, for an edge-type e, $M_{e,e}$ is always the largest among all possible edge-type pairs toward e, which means random walk prefers to keep the same kind of edge-type. So, during the random walk process, given the current node v and the previous node u, the probability for the next candidate node n is calculated as Formula 6 and demonstrated in Figure \ref{fig:demo}: \begin{equation} p(n|v;u;M) = \frac{w_{vn}\cdot M_{T(u,v)T(v,n)}\cdot \alpha_{pq}(n,u)}{\sum_{k \in N(v)}w_{vk}\cdot M_{T(u,v)T(v,k)}\cdot \alpha_{pq}(k,u)} \label{eq:random_walk_pr} \end{equation} \begin{figure} [width=1\columnwidth]{walk.png} \caption{Three parts of the weights to guide the biased random walk on heterogeneous networks.} \label{fig:demo} \end{figure} where T(v,u) refers to the edge-type between node v and node u. $\alpha_{pq}(k,u)$ is defined based on the distance $d_{ku}$ between next step node candidate k and previous traversed node u. The distance function is defined as Formula 7: \begin{equation} \alpha_{pq}(k,u) = \begin{cases} \frac{1}{p}, & \quad d_{ku} = 0\\ 1, & \quad d_{ku} = 1\\ \frac{1}{q}, & \quad d_{ku} = 2\ \end{cases} \label{eq:distance} \end{equation} As seen in Algorithm 1, at the beginning, we initialize walk paths as empty, all values in the transition matrix as 1, we use function $GenerateTransitionMatrix(\cdot)$ to utilize an EM framework to get walk paths and the matrix M. In maximization steps, the function takes transition matrix in the last iteration as input, invokes the $HeteroRandomWalk(\cdot)$ function to get walk paths, the probability of random walk is mainly based on Formula 6. In expectation steps, the function utilizes the updated walk paths to optimize the transition matrix by Formula 4. We can retrieve an optimized edge-type transition matrix, which holds the correlation between edge-types, via the EM framework. At the same time, we can also get the random walks as a node "corpus", which holds the correlation between nodes. We therefore represent the whole heterogeneous network as a collection of random walk paths, which can be used as the input of the next step for embedding optimization. \subsection*{Skip gram for embedding optimization} With the help of the EM framework, we can get the transition matrix M and random walks \textit{w} as the input layer to train the node embedding via a one layer neural network. To optimize the Formula 2, we use the stochastic gradient descent (SGD) method to get optimized node embeddings. Considering all nodes to maximize Formula 2 would be slow and computationally inefficient. Hence, in addition to the known neighbor node \textit{t}, we use the negative sampling method to generate \textit{k} negative nodes towards a given node \textit{v}. And the K negative nodes $u_{i}$ where $i \in \{1,2,...,k\}$ are randomly sampled from the uniformed distribution $D(t)$ with probability $P(t)$. Moreover, we take logarithm on Formula 2 to reduce calculation complexity. And the final objective function turns to be Formula 8 in the end: \begin{equation} \mathbf{O}(f) = \text{log [Sigmoid}(\vec{f_{t}}^T \vec{f_{v}})] + \sum_{i=1}^{k}E_{u_{i} \sim P(t|t \sim D(t))}\text{log [Sigmoid}(\vec{-f_{u_{i}}}^T \vec{f_{v}})] \label{eq:model_objective_function} \end{equation} The goal of the objective function is to maximize the similarity with the positive neighbour node and minimize the similarity with negative neighbor nodes. \section*{Results} In this section, we describe the biomedical dataset used to test edge2vec and demonstrate the advantage of our model in three evaluation tasks. Moreover, we have a separate section for parameter tuning to retrieve the best model in both efficacy and efficiency points of view. \subsection*{Biomedical dataset: Chem2Bio2RDF} Chem2Bio2RDF\cite{chen2010chem2bio2rdf} is a richly heterogeneous dataset integrating data from multiple public sources spanning biomedical sub-domains including bioinformatics, cheminformatics and chemical biology. The dataset includes 10 node types and 12 edge types. For details of each node/edge-type description, please refer to Table \ref{tab:data}. In total, there are 295,911 nodes and 727,997 edges, a relatively sparsely connected network. There exist multiple edge types between two given node types, for example, two edge types between node types "gene" and "compound." Node and edge type distributions are highly skewed. For instance, there are more than 20,000 compound nodes but a relative few are well studied in biological experiments, such as approved drugs, while most have few high confidence biological associations. Overall, the heterogeneity comprised by these network characteristics present significant challenges for embedding learning, and moreover, the particulars and specific semantics of this biomedical knowledge graph are essential considerations in optimizing learning power. Figure \ref{fig:graph_representation} shows the whole network structure of Chem2Bio2RDF. \begin{figure} [width=0.8\columnwidth]{data.png} \caption{Chem2Bio2RF medical data graph structure.} \label{fig:graph_representation} \end{figure} \begin{table} \centering \begin{tabular}{|l|l|l|l|l|l|l|} \hline \textbf{node type} & \textbf{number} & \textbf{edge type} & \textbf{number} & \textbf{edge type description}\\ \hline gene & 21,738 & hprd & 30,215 & protein protein interaction\\ \hline compound & 258,003 & protein & 11,258 & has pathway\\ \hline chebi& 2,777 & tissue & 10,178 & tissue gene expression\\ \hline pathway & 192 & GO id & 95,422 & has GO \\ \hline sider & 1,051 & family name & 7,181 & has gene family\\ \hline gene-family& 329 & gene & 2,929 & cause disease \\ \hline GO & 9,710 & chebi & 15,986 & has chemical ontology\\ \hline substructure & 290 & drug & 943 &has pathway \\ \hline tissue& 507 & cid & 9,004 & cause side effect\\ \hline disease & 1,284 & expression & 16,167 & compound gene expression\\ \hline & & substructure & 6,169 & has substructure \\ \hline & &chemogenomics & 522,545 & bind \\ \hline \end{tabular} \caption{Node and edge description in Chem2BioRDF } \label{tab:data} \end{table} Given the proposed edg2vec, we set up parameters with p = q = 0.25; embedding dimension d = 128; for other parameters we use the defaults from node2vec. After those parameters are assigned, we use Chem2BioRDF to train our edge2vec model. To evaluate the fitness of the generated node embeddings, we propose three evaluation tasks in the following three sections. \subsection*{Evaluation Metrics} In this paper, we evaluate our model from both classification and information retrieval viewpoints. Precision, recall, F1 score and Hamming loss are four metrics reported in classification tasks. Precision implies the ratio of correct positive results returned by the classifier; recall implies the ratio of correct positive results are returned; F1 score balances both precision and recall by taking their harmonic average. All above three metrics are in a range of 0 and 1, the higher the better. While the Hamming loss is the fraction of labels that are incorrectly predicted. The score is also in a range of 0 and 1, but the lower the better. Precision@K, recall@K, MAP, NDCG and reciprocal rank are five metrics reported in information retrieval related tasks. Precision@K and recall@K imply the precision and recall score in the Top K ranked results. MAP refers to "mean average precision", which implies the average precision score for all searching queries. NDCG refers to "normalized discounted cumulative gain", which is a metric to measure not only the accuracy of searching results but also the ranked position of correct results. Like NDCG, reciprocal rank also considers the correct results ranking positions in the returned ranking list. It is the multiplicative inverse of the rank of the first correct result among all searching queries. \subsection*{Entity multi-classification} \begin{table} \centering \begin{tabular}{|l|l|l|l|l|l|} \hline \textbf{Algorithm} & \textbf{Precision} & \textbf{Recall} &\textbf{F1 measure} & \textbf{Hamming loss} \\ \hline DeepWalk & 0.5624 & 0.5708 & 0.5650 & 0.4291 \\ \hline LINE & 0.6366 &0.6390 & 0.6279 & 0.3609 \\ \hline node2vec & 0.5652 & 0.5656 & 0.5622 & 0.4343 \\ \hline edge2vec & \textbf{0.7554}* &\textbf{0.7546}* & \textbf{0.7544}*& \textbf{0.2453}*\\ \hline \end{tabular} \caption{Classification on node labels in the medical network. Symbol "*" highlights the cases where our model significantly beats the best baseline with $p$ value smaller than 0.01.} \label{tab:multiclass} \end{table} We first propose a node multi-classification task. In this task, we take the types of nodes away so the network only has nodes, edges, and edge-types. We run edge2vec and cluster nodes based on the result of edge2vec to see whether nodes with similar types will be clustered together. In the Chem2BioRDF dataset, there are 10 different node types with different scale number. In order to build up a suitable dataset for the classification model, for each node type, we randomly sample equal number of nodes from the dataset. In this way, we have a natural baseline as precision = 0.1 for a random classifier. Each node is represented as an instance; the 128 dimension vectors are regarded as 128 different features. Its related node type is the response variable. We use a linear support vector machine as the classification model to predict the node’s labels, and use a 10-fold validation to evaluate the returned metrics. Three network embedding methods including DeepWalk, LINE and node2vec are our baseline algorithms. For node2vec, we take p = q = 0.25 which is the same setting as edge2vec. Other settings for all three algorithms are just default settings according to their related publications. For each node, after we learn its node embeddings for all baselines, we concatenate the embedding with the number of edges it has for each edge-type to integrate edge-type information into all baseline models as well. For example, if there are four edge-types in a network and a node has one edge with type 1, two edges with type 2, three edges with type 3 and zero edge with type 4, we concatenate an additional four dimensional vector (1,2,3,0) to the original learned embedding. As metapath2vec requires metapath definitions (manually curated) and thereby only uses selected metapath-pattern matched nodes for training node embeddings, metapath2vec is not comparable with other algorithms for a multi-classification task, which is also a drawback of metapath2vec. We use precision, recall, F1 score macro, and Hamming loss as four evaluation metrics. These are all commonly used evaluation metrics particularly for classification problem. Precision is the fraction of relevant instances among the retrieved instances, while recall is the fraction of relevant instances that have been retrieved over the total amount of relevant instances. F1 measure is the harmonic average of the precision and recall, which balances the two metrics. Hamming loss is the fraction of labels that are incorrectly predicted. Details of the evaluation results can be seen in Table \ref{tab:multiclass}. To verify our model's superiority, we run our model five times and calculate the performance differences between our model and the best baseline on each metric for all the runs, and apply a T-test to check whether the performance difference is significantly above 0 or not. From the evaluation results, we can find all four algorithms can predict node types far better than a random classifier. It means even we treat this heterogeneous network as a homogeneous one, there is still some meaningful information stored in these node embeddings. DeepWalk and node2vec have similar results which is no wonder because DeepWalk can be regarded as a particular node2vec model when p = q = 1. While LINE performs the best among all three baselines. It means for this medical network, local structure (one step neighbours and two step neighbours contains most information of a node). However, our proposed edge2vec model outperforms all baseline algorithms,. In all four evaluation metrics, our model has at least 20$\%$ improvement in each evaluation metric. It reflects that our model can better predict node labels via its node embedding. Moreover, in all steps of edge2vec, we only use edge-type information during the random walk to generate edge-type transition metrics, and no node type information. Therefore, we can rigorously validate model performance in node type prediction. \subsection*{Compound-gene bioactivity prediction} \begin{table} \centering \begin{tabular}{|l|l|l|l|l|l|l|} \hline \textbf{Algorithm} & \textbf{Precision} & \textbf{Recall} &\textbf{F1 measure} & \textbf{Hamming loss} & \textbf{AUROC} \\ \hline DeepWalk & 0.7787 & 0.7750 & 0.7742 & 0.2250 & 0.7660\\ \hline LINE & 0.8170 & 0.8166 & 0.8166 & 0.1833 & 0.8058\\ \hline node2vec &0.7983 & 0.7916 & 0.7904 & 0.2083 & 0.7793\\ \hline metapath2vec (Co-Ge-Co) & 0.5170 & 0.5170 & 0.5168 & 0.4830 & 0.5007 \\ \hline metapath2vec (Co-Ge-Ge-Co) & 0.4979 & 0.4980 & 0.4976 & 0.5020 & 0.4890 \\ \hline metapath2vec (Co-Dr-Ge-Dr-Co) & 0.5305 & 0.5305 & 0.5304 &0.4695 & 0.5304 \\ \hline metapath2vec++ (Co-Ge-Co) & 0.4969 & 0.4970 & 0.4965 &0.5030 & 0.4776\\ \hline metapath2vec++ (Co-Ge-Ge-Co) & 0.4854 & 0.4855 & 0.4854 & 0.5145& 0.4776\\ \hline metapath2vec++ (Co-Dr-Ge-Dr-Co) &0.5120 & 0.5120 & 0.5119 & 0.4880& 0.5102\\ \hline edge2vec & \textbf{0.9017}* & \textbf{0.9000}* & \textbf{0.8998}* &\textbf{0.1000}*& \textbf{0.8914}*\\ \hline \end{tabular} \caption{Compound-gene bioactivity prediction. Symbol "*" highlights the cases where our model significantly beats the best baseline with $p$ value smaller than 0.01.} \label{tab:link_prediction} \end{table} One high-value biomedical challenge is to predict bioactivity between a compound and protein target (often referenced by the encoding gene). Such predictions can accelerate early stage drug discovery by informing and/or replacing expensive screening campaigns via virtual screening. Therefore, we considered the real world bioactivity prediction use case for a validation task. Besides the three baselines in the first task, we add metapath2vec to our baseline as well. As metapath2vec needs to define metapaths \textit{a priori}, three metapaths, which are compound-gene-compound; compound -gene-gene-compound; compound-drug-gene-drug-compound were selected as the metapaths for our baseline metapath2vec. As metapaths need to be symmetric, we have to take the network as undirected when training metapath2vec node embeddings. Our ground truth is from another work\cite{fu2016predicting} in which the authors generated 600,000 negative compound gene pairs and 145,6222 positive pairs. These ground truth pairs do not exist in Chem2BioRDF so it can be used as ground truth to evaluate of the result of edge2vec for bioactivity prediction. As the label for a compound-gene pair in ground truth is either ‘positive‘ or ‘negative‘, the prediction task is a binary classification task. Here, we randomly select 2,000 positive pairs and 2,000 negative pairs from the ground truth. And a random classifier will have an accuracy value as 0.5 naturally. Similar to the approach in the multi-class classification task, for each compound-gene pair, we use the difference of both embeddings together to form a new 128-dimension embedding to represent the pair, and we apply a logistic regression classifier to train a prediction model. Each dimension of the pair is also regarded as a feature. The relationship between the compound and gene is a binary label for each pair. In the training and testing process, If the prediction score is above 0.5, we label the pair as "positive", otherwise as "negative". We deploy the same evaluation metrics as the multi-class classification task plus area under an ROC curve (AUROC) . The detailed result is shown in Table \ref{tab:link_prediction}. To verify our model's superiority, we run our model five times and calculate the performance differences between our model and the best baseline on each metric for all the runs, and apply a T-test to check whether the performance difference is significantly above 0 or not. In Figure \ref{fig:roc}, we also report the ROC curve for edge2vec and baseline models based on their prediction scores, where we can find our model curve significantly performs better than the baselines. \begin{figure} [width=0.8\columnwidth]{roc.png} \caption{The ROC curve of compound-gene bioactivity prediction. the metapath2vec curve refers to the best result from all six reported metapath2vec/ metapath2vec++ models in Table \ref{tab:link_prediction}.} \label{fig:roc} \end{figure} Some interesting findings are observed from the experiments. First of all, among all three well known baseline algorithms (DeepWalk, LINE and node2vec), LINE still outperforms the other two baselines. And the result of DeepWalk is similar to that of node2vec. So, edge2vec is reliable and functionally stable for the two tasks. For metapath2vec, we leverage both metapath2vec and metapath2vec++ in our baseline models. As metapath2vec relies too much on selected metapaths, none of the three metapaths performs well. Among these three metapaths, we find metapath compound-drug-gene-drug-compound works the best, implying that prediction accuracy is improved by the additional node types. Although the number of drug nodes is trivial compared with the number of compound nodes and gene nodes, drug nodes have larger effects than compounds and genes in terms of bioactivity prediction. So it is necessary to treat different types of nodes separately within an embedding model. Compared with metapath2vec, metapath2vec++ however achieves worse result in all three metapaths. edge2vec outperforms all baseline models. The F1 measure is around 0.9 which is far better than the random classifier with a score of 0.5. Also it has around 10$\%$ improvement compared to the LINE result which is the best of baseline results. \subsection*{Compound-gene search ranking} Bioactivity prediction as a binary classification task, like single point high throughput screening in the wet lab, predicts active or inactive only. This is helpful, but more useful is the capability to predict ranking of hits by a measure that increases the probability of success and overall efficiency in costly follow up efforts. Hence, this comprises our final evaluation task: compound-gene search ranking. By analogy, the number of hits returned by a search algorithm is generally less important than the ranking and particularly the top ranked hits. Thus, our final task can be described as an information retrieval or search efficiency task. To limit runtime cost, from the ground truth, we select 70 compounds, which contain more than one positive pair with a gene. For each compound, we calculate the top 100 similar nodes in Chem2BioRDF. Compared with the positive pairs of 70 compounds in ground truth, we evaluate the searching result using metrics such as precision, recall, MAP, NDCG, and mean reciprocal rank (MRR). These metrics care not only whether the bioactive genes are returned but also the ranking of the returned genes. For a compound node, if its bioactive genes shows up with a higher rank in the top 100 returned ranking list, the evaluation metrics will have larger values. After the bioactivity prediction task, we choose the best metapath among all three metapaths, which is compound-drug-gene-drug-compound. We evaluate the embedding results from LINE, node2vec, and edge2vec, as well as metapath2vec with the best metapath. Evaluation details are shown in Table \ref{tab:search}. To verify our model's superiority, we run our model five times and calculate the performance differences between our model and the best baseline on each metric for all the runs, and apply a T-test to check whether the performance difference is significantly above 0 or not. From the evaluation table, we find DeepWalk and node2vec still have similar results, and both outperform LINE. metapath2vec is almost meaningless because all returned evaluation metrics are approaching to 0, which means it can barely retrieve future positive genes to compounds. Metapaht2vec++ performs slightly better than its previous performance in other two tasks and is comparable to LINE. And overall, node2vec works the best in all baseline algorithms. Compared with all baseline algorithms, our edge2vec outperforms all the baseline methods. Although the retrieved scores are all relatively small, there is around 10$\%$ improvement in precision and a little better in the rest of evaluation metrics at least. This, edge2vec adds value in this critical task of compound-gene ranking, which can improve cost efficiency in virtual screening follow up efforts of early stage drug discovery. \begin{table} \centering \begin{tabular}{|p{2cm}|l|l|l|l|l|l|l|l|} \hline \textbf{algorithm} & \textbf{P@10} & \textbf{P@100} &\textbf{Recall@10} & \textbf{Recall@100} & \textbf{MAP} & \textbf{NDCG} & \textbf{MRR} \\ \hline DeepWalk &0.0623 & 0.0198 & 0.0725 & 0.2780 & 0.0707& 0.1444 & 0.1502 \\ \hline LINE & 0.0186 & 0.0069 & 0.0109 & 0.0360 & 0.0042 & 0.0234 & 0.0532 \\ \hline node2vec & 0.0714 & 0.0277 & 0.0859 & 0.2804 & 0.0786 & 0.1690 & 0.1676 \\ \hline metapath2vec (Co-Dr-Ge-Dr-Co) & 0.0000 & 0.0001 & 0.0000 & 0.0011 & 0.0000 & 0.0004 & 0.0001 \\ \hline metapath2vec++ (Co-Dr-Ge-Dr-Co) & 0.0157 & 0.0039 & 0.0082 & 0.0200 & 0.0040 & 0.0172 & 0.0509\\ \hline edge2vec & \textbf{0.0843}* &\textbf{ 0.0329}* & \textbf{0.0990}* & \textbf{ 0.3092}* & \textbf{0.0809} & \textbf{0.1840}* & \textbf{0.1882}* \\ \hline \end{tabular} \caption{Searching accuracy for retrieving potential compound-gene bindings. Symbol "*" highlights the cases where our model significantly beats the best baseline with $p$ value smaller than 0.01.} \label{tab:search} \end{table} \subsection*{Parameter tuning} In our EM framework for edge transition matrix training, in order to get the best fit to retrieve transition relationships between edge types, we have tried various correlation methods including cosine, Spearman, Pearson and Wilcoxon signed-rank. To standardize the returned correlation scores into a reasonable (above 0) and comparable range, we have tried various activation function such as ReLU, Sigmoid and traditional standardization methods. Our experiments show that using the combination Sigmoid activation function and Pearson correlation similarity performs the best to represent the transition relationship between edges. Hence, we decided to use this combination for building up our framework. During the transition matrix training process, there are four important parameters to be tuned. We list them with default value below. \begin{enumerate} \item{Number of walks on per node, r = 1} \item{Walk length in each random walk path, w = 50} \item{The ratio of nodes sampled for training edge transition matrix, p = 0.01} \item{The number of iterations for training edge transition matrix, N = 10} \end{enumerate} The default parameter settings are used to train our edge2vec model and compare with baseline models in previous sections. In this section, we vary each of them and fix the rest to examine the parameter sensitivity of our model. We leverage all generated results on solving node multi-class classification task and use the Macro F1 score as the judgment to evaluate related models. The result of our tuning process is shown in Figure \ref{fig:tune}. \begin{figure} [width=1.0\columnwidth]{tune.png} \caption{Parameter tuning in multi-class node classification.} \label{fig:tune} \end{figure} To test how much that numbers of walk per node can affect our model, we test five cases r = $\{1,10,20,30,40\}$ and the result is in Figure \ref{fig:tune} (a). We can see that more numbers of walks on per node leads to an increase in Macro F1 score. The reason might be that more walks on a node can better reflect the edge relationships around the node and avoid the negative influence of walk randomness. However, even though it shows a positive trend, the increase is small. So a short number of walks on per node should be able to capture enough edge relation information around the node. In Figure \ref{fig:tune} (b), with the increase of walk length, the Macro F1 score increases in the beginning and decreases later on. In general, as the random walk length increasing, it will tend to contain all types of edges. As we don't consider the sequence of the edge types in the same walk, increasing walk length can add noise obfuscating edge type transition relationships. Figure \ref{fig:tune} (c) shows the different ratio of nodes involved in the EM training process to generate edge transition matrix. It might be no need to involve all nodes when training the edge transition matrix as long as all edge types can be covered in random walks in each iteration. Although with more nodes involved, the overall trend of Macro F1 score has a positive sign, the increase of Macro F1 score is not huge and the trend even fluctuates a little bit. We thereby conclude that using a sampling of nodes to train the edge type transition matrix is sufficient and computationally efficient. Figure \ref{fig:tune} (d) refers to the influence of number of iterations in the EM process for edge type transition matrix training. We believe the reason why when N = 1 the Macro F1 score outperforms than N = 10 is by chance as when we increase the number of iterations, the overall trend of Macro F1 score also increases. From the Figure \ref{fig:tune} (d), the convergence is fast, and a few iterations can already generate a satisfactory Macro F1 score. \section*{Discussion} To discuss how we can apply our node2vec model on a biomedical data set, we conduct two case studies to show the practical values of our model. The first one is to rank the positive links between nodes, which can be used for similarity search and recommendation; the second one is to cluster and visualize similar gene nodes that belong to the same gene family. \subsection*{Ranking positive bindings for similarity search} \begin{table} \centering \begin{tabular}{|l|l|p{0.9cm}|p{0.9cm}|l|l|p{0.9cm}|p{0.9cm}|} \hline drug & gene & node2vec similarity & edge2vec similarity & drug & gene & node2vec similarity & edge2vec similarity \\ \hline DB11348 & CADPS2 & 0.1196 & 0.6199 & DB09130 & CLEC3B& 0.2177 & 0.6299 \\ \hline DB11348 & NUCB1 & 0.2221 & 0.6613 & DB09130 & HSPA13 & 0.1547 & 0.6278 \\ \hline DB11348 & TPT1 & 0.2233 & 0.6439 & DB09130 & IGLL1 & 0.1806 & 0.5856 \\ \hline DB11348 & NUCB2 &0.2668 & 0.6867 & DB09130 & SFPQ & 0.1275 & 0.5281 \\ \hline DB11348 & CIB2 & 0.2761 & 0.6946 & DB09130 & A1BG & 0.2478 & 0.6483 \\ \hline DB11348 & CAPS & 0.2827 & 0.6765& DB08818 & HMMR & 0.1624 & 0.6072 \\ \hline DB11348 & PEF1 & 0.3357 & 0.7472 & DB08818 & CD44 & 0.1408 & 0.5670 \\ \hline DB11348 & CALR3 & 0.2396 & 0.6429 & DB08818 & HAPLN1 & 0.1672 & 0.5790 \\ \hline DB11348 & CADPS & 0.1715 & 0.5657 & DB08818 & VCAN & 0.1877 & 0.6186 \\ \hline DB11348 & FBN3 & 0.3306 & 0.7369 & DB09131 & CSTB &0.1675 & 0.5999 \\ \hline \end{tabular} \caption{Compare node2vec \& edge2vec difference on classification tasks} \label{tab:intro_case} \end{table} To verify how well our model can be used for similarity search and recommendation use cases, we carried out a ranking experiments using the links identified in the existing network. We randomly selected three widely used drugs from the network, which are Hyaluronic acid (DB08818), Calcium Phosphate (DB11348), Copper (DB09130), and Cupric Chloride (DB09131). Each of them has multiple target genes to interact with. The selected pairs of drugs and target genes exist in the network, and we want to reproduce the links using the cosine similarity score based on the embedding vectors. As we can see, our proposed edge2vec embedding can represent node similarity significantly better than the node2vec embedding. The cosine similarity scores for the drug targets of calcium phosphate were all above 0.6, indicating strong similarity between the drug and the target genes. However, using node2vec embedding the cosine similarity scores between calcium phosphate and its targets were all below 0.4, and some of them demonstrated strong dissimilarity like CADPS2 as a target gene of calcium phosphate. The same findings for the other three drugs and their target genes. In addition, all of the target genes for those drugs can be identified as similar nodes with high rankings using edge2vec embeddings. Details are shown in Table \ref{tab:intro_case}. we further performed a pairwise t-test\cite{box1987guinness} study to see whether the similarity scores generated by two models are significantly different or not. If edge2vec has significantly higher similarity score than node2vec, it means our model can better predict those positive drug-target pairs in the network. In our result, the difference between two embedding approaches is 0.0103 with a p-value of 0.0001. It means our embedding approach can better rank and retrieve the existing links in the network than node2vec. \subsection*{Gene clustering analysis} \begin{figure} [width=1\columnwidth]{visualization.png} \caption{2-D PCA projection on 25 random selected genes, five each from 5 random gene families.} \label{fig:case_visualization} \end{figure} In order to further prove the usefulness of our node embedding results, we carried out a clustering analysis of gene nodes. We arbitrarily selected five gene families, which are ATP binding cassette transporters (ABC), Gap junction proteins (GJ), DEAD-box helicases (DDX), Cytochrome C Oxidase Subunits (COX), and Pseudoautosomal region 2 (PAR2). Each gene family refers to a collection of genes originated from the same root gene and performing similar biochemical functions. From each given gene family, five genes are randomly selected to perform clustering analysis. Then, we use principal component analysis (PCA) to project the default 128 dimensions of all gene embeddings into a 2-D space. Ideally, genes belonging to the same family should locate closer to each other than genes from different families. The resulting 2-D plot has shown that the twenty five genes in five gene families have been well clustered based on their node embedding vectors. Results can be visualized in Figure \ref{fig:case_visualization}, where each pentagon refers to a gene and different colors indicate different gene families. It is easy to observe that genes are clustered by family in the 2-D space. In the 2-D projection, genes in family COX are all located in the bottom right corner of the plot, and genes in family GJ tend to stay in the bottom left corner. In a word, our edge2vec model can generate node embeddings highly reflecting their family information. \section*{Conclusions} In this paper, we propose edge2vec, which incorporates edge semantics to add value over previous methods, as evaluated by knowledge discovery tasks in the domain of biomedical informatics. Edge2vec employs an EM framework associated with a one-layer neural network, to learn node embeddings which perform better than previous methods for heterogeneous networks. The novelty of our work is to generate an edge-type transition matrix so that during the process to generate the node random walk corpus, heterogeneity of the network is also considered. It can reduce the skewed type distribution issue via weighted sampling. Moreover, compared with other state-of-art heterogeneous network embedding methods such as metapath2vec, our edge2vec has no restrictions and can deal with the situation where there are multiple relationships between two types of nodes. To illustrate efficiency and accuracy of our proposed model, we evaluate it on biomedical dataset Chem2BioRDF and propose three evaluation tasks including node multi-class classification, link prediction, and search rank efficiency. Edge2vec outperforms all baseline algorithms significantly. Furthermore, we illustrate the effect of edge2vec in biomedical domains using two case studies to explain the biological meanings of the prediction. Moreover, edge2vec can work well in both undirected and unweighted networks, and computational cost is only moderately increased relative to node2vec by choice of appropriate random walk strategy in the EM framework. There are certainly promising future directions to be explored, which can be subdivided into (1) algorithmic modifications applicable to heterogeneous graphs generally,and (2) domain knowledge based enhancements applicable to characteristics of biomedical datasets and use cases. Informed by metapath2vec, we could change the objective function by using a node-type based negative sampling. Instead of random sampling from all types of nodes, we could sample negative nodes based on the ratio of each node type. Another opportunity for improvement involves adding domain knowledge into our existing model. During the random walk generation process, we have already considered both node distance (p,q) and edge-type (transition matrix M). In addition to these two attributes, we could add some pre-defined rules guiding random walks based on our domain knowledge. Another problem for such complex networks is the scale free issue, i.e. skewed degree distributions, where a relative few hub nodes account for the bulk of connections. To reduce this hub node effect in the network, we can also come up with new methods such as novel restriction rules in the random walk, or mitigate those effects by adding node degree related information to optimize a new objective function. For the node attribute prediction task, we can actually apply a semi-supervised approach: Given a sampling of nodes with known attributes, we can add this information into our embedding learning process and predict the attributes of remaining nodes. Or we can apply an unsupervised approach: given the relationship between the target attribute with other known attributes, we use this relationship information and other known attributes information in our embedding learning process, and in the end directly predict node target attributes. These enhancements could extend our edge2vec model to better address specific downstream tasks such as node attribute prediction with unsupervised or semi-supervised training strategies. \section*{List of Abbreviations} \begin{table} \centering \begin{tabular}{|p{2.5cm}|p{9cm}|} \hline \textbf{Knowledge graph} & Knowledge base represented as nodes as entities and edges as relationships. A.k.a. Knowledge network. A type of heterogeneous graph. \\ \textbf{Machine learning} & Computer system automatically learns the data pattern via designed algorithms and statistical models.\\ \hline \textbf{Heterogeneous graph} & Includes nodes of multiple classes. A bipartite graph has two classes, e.g. persons and movies. \\ \hline \textbf{Graph embedding} & Transformation into feature space represented as numerical vectors. Node embedding is a form of graph embedding, where each vector represents one node. \\ \hline \textbf{Adjacency matrix} & Graph topology square matrix, one row/column per node, typically very sparse. \\ \hline \textbf{Semantics} & Ontology defining types of entities and their relationships, in graph terms, nodes and edges, or in RDF terms, classes of entities and predicates. Formalized or not, the ontology conveys the semantics mapping data to knowledge. \\ \hline \textbf{EM}& Expectation Maximization \\ \hline \textbf{SGD} & Stochastic Gradient Descent \\ \hline \textbf{word2vec} & ML feature method developed by Google representing a word and its lexical context in a low dimensional dense vector.\\ \hline \textbf{node2vec} & Conceptually derived from word2vec, represents a node and its topological context. A path is treated like a sentence. \\ \hline \textbf{Skip gram model} & Employed by word2vec, Node2vec and related methods, to represent non-adjacent words/nodes within some context window. \\ \hline \textbf{Metapath} & Path pattern, the classes of nodes and edges, for which many instances may exist. \\ \hline \textbf{metapath2vec} & Like node2vec but with heterogeneous graphs. Metapath based random walks. \\ \hline \textbf{Transition matrix} & In the edge2vec algorithm, matrix of relative probabilities or expectation values for edge type transitions. \\ \hline \textbf{Representation learning} & Processes by which an input representation is transformed to an alternative representation, such as in layers of a neural network. \\ \hline \textbf{Biomedical Knowledge Domain} & Shorthand for a varied set of richly heterogeneous subdomains, including medical and pharmaceutical informatics, bioinformatics, and cheminformatics. \\ \hline \end{tabular} \end{table} \section*{Declarations} \subsection*{Acknowledgements} Not applicable. \subsection*{Funding} The work was supported by National Natural Science Foundation of China (No. 71573162). The funding bodies had no role in the design of this study, the collection, analysis, and interpretation of data, or the writing of this manuscript. \subsection*{Availability of data and materials} The dataset used in this work can be found at Github \footnote{https://github.com/RoyZhengGao/edge2vec}. There are three files within the zipped folder: \textit{chem2bio2rdf.txt} is the heterogeneous graph we used to train our model. In this file, each line is a RDF triplet which contains two entities and their relations. Entity type and relation type can be obtained from their RDF representations directly. \textit{negative.txt} and \textit{positive.txt} stored the negative and positive bindings between genes and compounds, which are all directly generated from the original chem2bio2rdf paper. The dataset is processed from the original dataset published at BMC Bioinformatics \footnote{https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-016-1005-x}. We use the \textit{chem2bio2rdf.txt}, \textit{internal\_testset\_label/positive.txt} and \textit{internal\_testset\_label/negative.txt} from \textit{semantic\_network\_dataset.zip} directly. The source code is stored at at Github \footnote{https://github.com/RoyZhengGao/edge2vec}. For details to run the code, please refer to the instructions in the Github link. In order to load the data into edge2vec, please convert all RDF format data points to hashed ids first, and save in csv format. Then run \textit{transition.py} to generate and store the transition matrix. Then, \textit{transition.py} is utilized to load the data and transition matrix for embedding optimization. For evaluation, please use \textit{negative.txt} and \textit{positive.txt} as ground truth files. \subsection*{Authors' Contributions} GZ prepared the dataset, programmed the algorithm, and initially drafted the manuscript. XL and YD conceived and guided the project. JJY revised the manuscript with emphasis on biomedical data science applications. GF helped to prepare the ground truth data set and comparison and helped with case study and editing. ST helped to implement the baselines. CO made contributions on the background writing. CG, BF, DW and QY are discussed the scientific ideas, reviewed and approved the final manuscript. \subsection*{Ethics approval and consent to participate} Not applicable. \subsection*{Consent for publication} Not applicable. \subsection*{Competing Interests} The authors declare that they have no competing interests. \begin{backmatter} \bibliographystyle{bmc-mathphys}

00d7ea13078c26ed3709e632abffdb2a32232f3d

\section{Introduction} \subsection{The `error term' in tropicalization} \label{subsec:error} The relationship between tropical and algebraic geometry is based on the `Maslov dequantization': \[ \log_T\left(T^a + T^b\right) \approx \max(a,b) \quad \text{ for $T \gg 1$.}\] Setting $b=0$, we can use this to arrive at the following approximation: \[ (\log T)^2\int_{-A}^A \log_T\left(1+T^a\right) da \approx (\log T)^2 \int_{-A}^A \max(0,a) da = \frac{A^2 }{2}(\log T)^2.\] However there is an error term in this approximation (see Figure \ref{fig:Maslov}): in the limit $T \to \infty$ it is given by \begin{align*} (\log T)^2 \int_{-A}^A \left(\log_T \left(1+T^a \right) - \max(0,a)\right) da &= 2 (\log T)^2\int_0^A \log_T\left(1+T^{-a}\right) da \\ &= 2 \int_{T^{-A}}^1 \frac{\log(1+x)}{x} dx \\ &= 2\sum_{k=1}^\infty \frac{(-1)^{k+1}}{k^2} +O(T^{-A}) \\ &= \zeta(2) +O(T^{-A}). \end{align*} In other words, $\zeta(2)=\pi^2/6$ arises as a subleading term in the Maslov dequantization. Going one dimension up, we can calculate the error term in the analogous approximation \[ (\log T)^3\int_U \log_T\left(1+T^{a_1} +T^{a_2}\right) da_1 da_2 \approx (\log T)^3 \int_U \max(0,a_1,a_2) da_1da_2.\] We will assume that $U \subset \R^2$ is a polygon containing the origin, and transverse to the legs of the `tropical curve' $\operatorname{Sing}(\max(0,s_1,s_2))$ (i.e. the locus where at least two of $0,s_1,s_2$ are tied for largest). The error term is equal to \[ \int_{\log T \cdot U} \left( \log\left(1+e^{s_1}+e^{s_2}\right) -\max(0,s_1,s_2) \right) ds_1ds_2. \] The integrand looks approximately like $\log\left(1+e^s\right)-\max(0,s)$ in the directions normal to the legs of the tropical curve. Thus the leading piece of the error term is equal to the total length of the tropical curve contained inside the region $\log T \cdot U$ multiplied by $\zeta(2)$, which will be linear in $\log T$. It turns out that there is also a constant term, which is equal to $\zeta(3)$ (see Proposition \ref{prop:asymptotics_of_integral}). The main idea of this paper is to use such approximations to compute asymptotic expansions for period integrals, and to relate them to the Gamma class of the mirror, which we describe in the next section. \begin{rmk} These error terms compute the volume of (parts of) amoebas; see \cite{Passare:zeta2, Passare-Rullgard:amoeba} for the study in 2 dimensions. \end{rmk} \begin{figure}[ht] \centering \includegraphics[scale=0.5]{dim2.pdf} \includegraphics[scale=0.5]{dim3.pdf} \caption{The graphs of $\log_T(1+T^x)$ and $\log_T(1+T^x+T^y)$ when $T=e$. These images were produced using Maple 2018 \cite{Maple2018}.} \label{fig:Maslov} \end{figure} \subsection{The Gamma class and mirror periods} It has been long observed that products of the characteristic numbers of a Calabi--Yau manifold by zeta values can be found in the asymptotics of periods of the mirror near the large-complex structure limit. For example, $\zeta(3)$ multiplied by the Euler number of a quintic threefold appears in the famous work of Candelas--de la Ossa--Green--Parkes \cite{CdlOGP:pair}. Later, Hosono--Klemm--Theisen--Yau \cite{HKTY:compint} observed that certain Chern numbers of Calabi--Yau complete intersection threefolds can be read off from hypergeometric solutions to the mirror Picard--Fuchs equation. This observation led Libgober \cite{Libgober:Gamma} to introduce the (inverse) Gamma class which makes sense for any almost-complex (or stably complex-oriented) manifold. The \emph{Gamma class}\footnote{When $X$ is an orbifold, the Gamma class has a component in the twisted sector. We nevertheless ignore the twisted sector component since it does not intervene in the statement of the Gamma Conjecture.} of an almost-complex manifold $X$ is defined to be the cohomology class \begin{align*} \wh{\Gamma}_X & = \prod_i \Gamma(1+\delta_i) = \exp\left(-\gamma c_1(X) + \sum_{k=2}^\infty (-1)^k \zeta(k) (k-1)! \ch_k(TX) \right) \in H^*(X,\R) \end{align*} where $\delta_i$ are the Chern roots of the tangent bundle $TX$ (such that $c(TX) = \prod_i (1+\delta_i)$) and $\gamma = \lim_{n\to \infty} (1+\frac{1}{2} +\cdots+ \frac{1}{n} -\log n)$ is the Euler constant. In terms of the Gamma class, a conjecture put forward by Hosono \cite[Conjecture 2.2]{Hosono:central} (see also \cite{Horja:hypergeometric, Enckevort-Straten, Borisov-Horja:FM, Almkvist-vanStraten-Zudilin:Apery, Golyshev:deresonating, Iritani:periods}) can be restated as follows: \begin{conje}[Gamma Conjecture in the Calabi--Yau case] \label{conje:CY_Gamma} Let $X$ be a Calabi--Yau manifold equipped with a symplectic form $\omega$ and let $\{Z_t\}_{t\in \Delta^*}$ be a family of Calabi--Yau manifolds parametrized by $t$ in a small punctured disc $\Delta^*$ that corresponds to $(X,\omega)$ under mirror symmetry. For a suitable choice of a holomorphic volume form $\Omega_t$ on $Z_t$ and of a coordinate $t$, if a Lagrangian cycle $C_t\subset Z_t$ is mirror to a coherent sheaf $E$ on $X$, then \[ \int_{C_t\subset Z_t} \Omega_t = \int_X t^{-\omega} \cdot \wh{\Gamma}_X \cdot (2\pi \mathtt{i})^{\deg/2} \ch(E) + O\left(t^\epsilon\right) \qquad \text{as $t\to 0$ in a fixed angular sector} \] for some $\epsilon>0$, where $\mathtt{i}=\sqrt{-1}$ is the imaginary unit. \end{conje} \begin{rmk} (a) The original conjecture of Hosono \cite{Hosono:central} is stated as an equality between periods and explicit hypergeometric series in the case of complete intersection Calabi--Yau manifolds. The version presented here can be obtained from the leading asymptotics of the hypergeometric series. (b) Both sides of the equality in the Gamma Conjectures are multivalued functions of $t$: on the right hand side, a choice of branch of $\log t$ is required to specify a value for $t^{-\omega}$, while on the left hand side the monodromy of the family $Z_t$ in general acts non-trivially on the homology classes of Lagrangian cycles. The family of Lagrangian cycles $C_t$ mirror to $E$ is identified over the universal cover of the punctured disc $\Delta^*$. (c) This is not a mathematically precise conjecture since it depends on mirror symmetry. In the case of Fano manifolds, there is a precise conjecture (the original Gamma Conjecture) which can be formulated purely in terms of quantum cohomology of a Fano manifold $X$ \cite{GGI:gammagrass, Galkin-Iritani:Gammamirror, Sanda-Shamoto} and is closely related to Dubrovin's conjecture \cite{Dubrovin:ICM}. (d) In the above conjecture, we implicitly assume that $t=0$ is a point of maximal degeneracy in the sense that the associated limit mixed Hodge structure is Hodge--Tate \cite{Deligne:localbehavior} and that the mirror map takes the form\footnote{If the mirror map is of the form $-\omega \log t + h + O(t)$ with $h\in H^2(X)$, then we need to replace $\wh{\Gamma}_X$ with $e^h \wh{\Gamma}_X$ in the conjecture; the class $h$ appears, for instance, when we replace $t$ with $2t$.} $-\omega \log t + O(t)$ so that $t=0$ corresponds to the large-radius limit point of $(X,\omega)$. We could further assume that the volume form $\Omega_t$ is normalized by a Hodge-theoretic condition as discussed in \cite{CdlOGP:pair, Morrison:guide, Deligne:localbehavior}. (e) Using the Gamma class, Katzarkov--Kontsevich--Pantev \cite{KKP:Hodge} and the third author \cite{Iritani:integral} introduced a rational/integral structure on the quantum cohomology, which conjecturally corresponds to the natural rational/integral structure (given by Betti cohomology) on the B-side. \end{rmk} Although Conjecture \ref{conje:CY_Gamma} is not mathematically precise, we can make it precise by specifying what we mean by a ``mirror pair'' and by fixing the correspondence between equivalence classes of cycles on the two sides: the Strominger--Yau--Zaslow (SYZ) conjecture posits that mirror pairs should carry dual (possibly singular) torus fibrations, and building on this, the Gross--Siebert program gives a geometric construction of mirror pairs in some large generality (see \cite{SYZ,Gross-Siebert:realaffine}). This determines the correspondence between Lagrangian cycles and coherent sheaves appearing in the Gamma Conjecture. The present paper aims to understand/explain the Gamma Conjecture from the viewpoint of the SYZ fibrations. \subsection{The Gamma Conjecture for Batyrev mirrors}\label{subsec:gam_bat} Let $\Delta \subset P_\R$ be a reflexive polytope, and $\nabla \subset Q_\R$ its polar dual, where $P \cong \Z^{n+1}$ and $Q := P^\vee$ and we write $P_K =P\otimes_\Z K$, $Q_K=Q\otimes_\Z K$ for a $\Z$-module $K$. Let $V$ be a subset of $\partial\nabla \cap Q$ containing all vertices of $\nabla$ and let $\lambda\colon V \to \R_+$ be a positive real-valued function.\footnote{$\R_+$ denotes the set of positive real numbers.} We assume that there exists a simplicial fan $\Sigma_\lambda$ on $Q_\R$ such that the set of one-dimensional cones of $\Sigma_\lambda$ is $\{\R_{\ge 0} \cdot q:q\in V\}$ and that $\lambda$ extends to a strictly-convex piecewise-linear function $\lambda\colon Q_\R\to \R$ with respect to the fan $\Sigma_\lambda$. We set \[ f_t(z) := \sum_{q \in V} t^{\lambda_q} \cdot z^q\] for $t \in \R_+$ and $z\in P_{\C^*}$, and \[ \mathring{Z}_t := \left\{1 = f_t(z) \right\} \subset P_{\C^*}.\] The positive real locus $C_t^+ \subset \mathring{Z}_t$ is defined to be the intersection $\mathring{Z}_t \cap P_{\R_+}$; this is homeomorphic to a real $n$-dimensional sphere for a sufficiently small $t>0$ (see Section \ref{subsec:tropical_setup}). Let $Y_\nabla$ denote the toric variety defined by the normal fan of $\nabla$ and take a partial crepant resolution $\wh{Y}_\nabla$ of $Y_\nabla$ which has at worst quotient singularities. The hypersurface $\mathring{Z}_t$ compactifies to a quasi-smooth Calabi--Yau hypersurface $Z_t\subset \wh{Y}_\nabla$. The holomorphic volume form \[ \Omega_t :=\left. \frac{d\log z_0 \wedge d\log z_1 \wedge \cdots \wedge d\log z_n}{df_t(z)}\right|_{\mathring{Z}_t} \] also extends to $Z_t$, where $(z_0,z_1,\dots,z_n)$ denotes $\C^*$-coordinates on $P_{\C^*}\cong (\C^*)^{n+1}$. On the $B$-side of mirror symmetry we will consider the period integral \[ \int_{C_t^+\subset \mathring{Z}_t} \Omega_t. \] On the $A$-side of mirror symmetry, we consider the compact convex polytope \[ \Delta_\lambda := \{ p \in P_\R: \langle q,p\rangle+\lambda_q \ge 0, \ \forall q\in V\}.\] Our assumption on $\lambda$ ensures that the slopes of the edges at each vertex form a basis of $P_\R$. We have a corresponding toric orbifold $Y_{\Delta_\lambda}$ equipped with a K\"ahler class $[\omega_\lambda] = \sum_{q\in V} \lambda_q \cdot D_q$, where $D_q$ is the toric divisor corresponding to the $q$th face $\{p \in \Delta_\lambda: \langle q,p \rangle+\lambda_q =0\}$ of $\Delta_\lambda$. The Batyrev mirror of $Z_t$ is given by a quasi-smooth Calabi--Yau hypersurface $X \subset Y_{\Delta_\lambda}$ \cite{Batyrev:dual_polyhedra}. It is expected that the large-radius limit of $X$ corresponds to the large complex structure limit $t\to 0$ for $Z_t$ and that the Lagrangian sphere $C_t^+\subset Z_t$ is mirror to the structure sheaf $\mathcal{O}_X$ of $X$. Thus it makes sense to substitute $C_t=C_t^+$ and $E=\mathcal{O}_X$ in Conjecture \ref{conje:CY_Gamma}. Our first main result is a proof of this special case of the Gamma Conjecture: \begin{main}[Gamma Conjecture for the structure sheaf on Batyrev mirror pairs] \label{thm:Bat_Gamma_struc} We have: \[ \int_{C_t^+\subset Z_t} \Omega_t = \int_X t^{-\omega} \cdot \wh{\Gamma}_X + O\left(t^\epsilon\right) \qquad \text{as $t\to +0$, for some $\epsilon>0$.} \] \end{main} Our second main result is a generalization of Theorem \ref{thm:Bat_Gamma_struc}, in which the structure sheaf $\mathcal{O}_X$ is replaced with an arbitrary ambient line bundle on $X$, i.e., one restricted from $Y_{\Delta_\lambda}$. Any line bundle on $Y_{\Delta_\lambda}$ has the form $\mathcal{O}_\nu = \mathcal{O}(-\sum_{q \in V} \nu_q D_q)$ for some $\nu \in \Z^V$; let $\mathcal{L}_\nu$ denote the restriction of $\mathcal{O}_\nu$ to $X$. We now describe the cycle $C_t^{(\nu)} \subset Z_t$ mirror to $\mathcal{L}_\nu$. Consider the polynomial function $f_{t,\theta}(z)$ on $P_{\C^*}$ \[ f_{t,\theta}(z) := \sum_{q\in V} e^{\mathtt{i} \theta_q} t^{\lambda_q} z^q \] that we obtain from $f_t(z)$ by multiplying the coefficients $t^{\lambda_q}$ by $e^{\mathtt{i} \theta_q}$ for some $\theta_q \in \R$, and the associated hypersurface \[ \mathring{Z}_{t,\theta} := \{f_{t,\theta}(z) = 1\} \subset P_{\C^*}. \] Let $C_t^{(\nu)} \subset \mathring{Z}_{t}$ be the parallel transport of the positive real cycle $C_t^+ \subset \mathring{Z}_t$ as we vary $\theta$ continuously from $\theta=0$ to $\theta = 2\pi \nu$. \begin{rmk}\label{rmk:why} Let us explain why the cycle $C_t^{(\nu)}$ is expected to be isotopic to a Lagrangian cycle mirror to $\mathcal{L}_\nu$. One expects a commutative diagram of categories as follows: \[ \xymatrix{ D^bCoh(Y_{\Delta_\lambda}) \ar[d]_-{\text{restriction}} \ar@{<->}[rr]_{\text{HMS}}^\simeq && D^b FS(P_{\C^*},f_t) \ar@{<->}[rrr]_-{\text{compactify}}^\simeq && &D^b FS(\overline{P}_{\C^*},\bar{f}_t) \ar[d]^-{L \mapsto \partial L} \\ D^bCoh(X) \ar@{<->}[rrrrr]_{\text{HMS}}^\simeq &&&&& D^b Fuk(Z_t).}\] The top arrow labelled `HMS' is homological mirror symmetry for the toric variety $Y_{\Delta_\lambda}$ and its Landau--Ginzburg mirror $(P_{\C^*},f_t)$, where we fix $t$. The other arrow on the top line identifies the Fukaya--Seidel category of this Landau--Ginzburg model with that of its fibrewise compactification (see \cite{Seidel:II.5}). The objects of $FS(P_{\C^*},f_t)$ may be taken to be certain Lagrangian submanifolds of $P_{\C^*}$ with boundary on the fibre $f_t^{-1}(1) = \mathring{Z}_t$, while the objects of $FS(\overline{P}_{\C^*},\bar{f}_t)$ are Lagrangians with boundary on the compactified fibre $\bar{f}_t^{-1}(1) = Z_t$. The left vertical arrow denotes the derived restriction functor, while the right one sends a Lagrangian to its boundary; the commutativity of this diagram appears, e.g., in \cite[Conjecture 7.7]{Auroux:Tduality}. Under the (presumably removable) assumption that $Y_{\Delta_\lambda}$ is smooth, the first author has constructed certain objects $L_\nu$ of $FS(P_{\C^*},f_t)$, and proved that they are mirror to the line bundles $\mathcal{O}_\nu$ \cite{Abouzaid:homogeneous, Abouzaid:HMS_toric} (see also the work of Fang--Liu--Treumann--Zaslow \cite{FLTZ:T-dual}, Fang \cite{Fang:central_charges}, Fang--Zhou \cite{Fang-Zhou}, and Hanlon \cite{Hanlon:monodromy}). Therefore, by commutativity of the above diagram, one expects $\mathcal{L}_\nu$ to be mirror to $\partial L_\nu \subset Z_t$. One can identify $\partial L_\nu$ with $C_t^{(\nu)}$ up to an isotopy: this becomes transparent using the tropical construction of $C_t^{(\nu)}$ given in Section \ref{sec:phase}. \end{rmk} In light of Remark \ref{rmk:why}, it makes sense to substitute $C_t = C_t^{(\nu)}$ and $E = \mathcal{L}_\nu$ in Conjecture \ref{conje:CY_Gamma}. Our second main result is a proof of this special case of the Gamma Conjecture (generalizing Theorem \ref{thm:Bat_Gamma_struc}, which is the case $\nu=0$): \begin{main}[Gamma Conjecture for ambient line bundles on Batyrev mirror pairs] \label{thm:Bat_Gamma_lb} We have: \[ \int_{C_t^{(\nu)} \subset Z_t} \Omega_t = \int_X t^{-\omega} \cdot \wh{\Gamma}_X \cdot e^{-\sum_{q \in V} 2\pi \mathtt{i} \nu_q D_q} + O\left(t^\epsilon\right) \qquad \text{as $t\to +0$, for some $\epsilon>0$.} \] \end{main} We remark that similar results have been obtained in \cite[Theorem 1.1]{Iritani:periods}; the novelty in our work is the method of proof, which relates the Gamma Conjecture to the SYZ Conjecture and the Gross--Siebert program. In fact, due to the local nature of the computations, we expect that it should not be significantly harder to implement our approach for general Gross--Siebert mirrors than for Batyrev mirrors. The Gamma Conjecture for general Gross--Siebert mirrors is open. In a different direction, we expect that it should be possible to implement our approach to prove the Gamma Conjecture for certain Lagrangian cycles $C_t$ fibring over `tropical cycles' in the base of the SYZ fibration (see \cite{Castano-Bernard2014,Ruddat2019} for the notion of `tropical cycle' in closely-related contexts, and \cite{Matessi2018a,Mikhalkin2018} for the construction of the corresponding Lagrangian cycles). Indeed this is essentially done in \cite{Ruddat2019}, in the case that the tropical cycle in the base of the SYZ fibration is 1-dimensional. In this case the interesting part of the Gamma class (i.e., the part involving zeta values) does not appear in the computation: the mirror coherent sheaf is the skyscraper sheaf of a curve, and in particular its Chern character is concentrated in degrees $\ge 2n-2$, whereas the zeta values in the Gamma class of a Calabi--Yau only appear in degrees $\ge 4$. This reflects the fact that the 1-dimensional tropical cycle can be (topologically) deformed to avoid the codimension-2 singular locus of the SYZ fibration, where the non-trivial contributions to the Gamma class are concentrated.\footnote{We should mention that the aim of \cite{Ruddat2019} is rather different from that of the current paper: the authors show that the natural coordinate on the base of the family constructed by Gross--Siebert is a canonical coordinate in the Hodge-theoretic sense.} \subsection{Proofs of Theorems \ref{thm:Bat_Gamma_struc} and \ref{thm:Bat_Gamma_lb}} We compute the asymptotics of the period integrals appearing in Theorems \ref{thm:Bat_Gamma_struc} and \ref{thm:Bat_Gamma_lb} by breaking them up into local pieces using tropical geometry. This procedure involves an extra layer of combinatorial complexity in the case of Theorem \ref{thm:Bat_Gamma_lb}, so we give the proof of Theorem \ref{thm:Bat_Gamma_struc} first in the name of transparency. The `local period integrals' that will appear are \begin{equation} \label{eqn:Ilm} I_{\ell;m_1,\ldots,m_k} := \int_{[0,\infty)^k} s_1^{m_1}\ldots s_k^{m_k} \cdot g_\ell(e^{-s_1},\ldots,e^{-s_k}) \,ds_1\cdots ds_k \qquad\text{for $\ell, m_j \in \Z_{\ge 0}$,} \end{equation} where \[g_\ell(X_1,\ldots,X_k) := \sum_{K \subset \{1,\dots,k\}} (-1)^{|K|} \cdot \left(\log \left(1+ \textstyle\sum_{ j \in K} X_j\right)\right)^\ell.\] The integral \eqref{eqn:Ilm} converges because the integrand decays exponentially at infinity, due to the bound \begin{equation} \label{eqn:boundint} g_\ell(X_1,\ldots,X_k) \le C_\ell \cdot \prod_{j=1}^k X_j \qquad \text{ on $[0,1]^k$.} \end{equation} This bound can be proved by observing that the function $g_\ell(X_1,\dots,X_k)$ is analytic in a neighbourhood of $[0,1]^k$, and vanishes along the coordinate hyperplanes $\{X_j = 0\}$, so is divisible by $\prod_{j=1}^k X_j$. We note that $g_\ell=0$ for $\ell = 0$. We define a class in $H^*(X)$ by \begin{equation} \label{eq:hG_X} \wh{G}_X = 1 + \sum_{q,J,\ell,\vec{m}} \frac{I_{\ell;\vec{m}}}{\ell! \prod_{j\in J} m_j!} \cdot (-D_q) \cdot (-\sigma)^{\ell-1} \cdot \prod_{j \in J} (-D_j)^{m_j + 1} \end{equation} where $\sigma= \sum_{j \in V} D_j$ is the first Chern class of $Y_{\Delta_\lambda}$, and the sum is over all $q \in V$, all nonempty subsets $J \subset V$ not containing $q$, $\ell \ge 1$ and $\vec{m}\in (\Z_{\ge 0})^{J}$. \begin{thm} \label{thm:periodR} Let $(X, Z_t)$ be a Batyrev mirror pair of Calabi--Yau hypersurfaces and let $C_t^+\subset Z_t$ denote the positive real locus. Then we have \[ \int_{C_t^+\subset Z_t} \Omega_t = \int_X t^{-\omega_\lambda} \cdot \wh{G}_X + O\left(t^\epsilon\right)\qquad \text{as $t\to +0$, for some $\epsilon>0$.}\] \end{thm} Theorem \ref{thm:periodR} is proved in Section \ref{sec:periodR}. The proof uses tropical geometry to decompose $C_t^+$ into pieces, so that the integrals of $\Omega_t$ over these pieces are in one-to-one correspondence with the terms on the right-hand side. \begin{thm} \label{thm:gamma} We have $\wh{G}_X = \wh{\Gamma}_X$. \end{thm} Theorem \ref{thm:gamma} is proved in Section \ref{sec:gamma}. Combining Theorems \ref{thm:periodR} and \ref{thm:gamma}, we have proved Theorem \ref{thm:Bat_Gamma_struc}. In Section \ref{sec:phase}, we generalize Theorem \ref{thm:periodR} to give a computation of period integrals over cycles $C_{t,\theta} \subset \mathring{Z}_{t,\theta}$, obtained by parallel transport of $C_t^+$; see Theorem \ref{thm:cycle_with_phase}. These include the cycles $C_t^{(\nu)}$, in the special case that all $\theta_q$ are integers. Theorem \ref{thm:Bat_Gamma_lb} is proved by combining Theorems \ref{thm:cycle_with_phase} and \ref{thm:gamma}. \subsection{Plan} Theorems \ref{thm:periodR} and \ref{thm:gamma} are proved in Sections \ref{sec:periodR} and \ref{sec:gamma} respectively; together they prove Theorem \ref{thm:Bat_Gamma_struc}. The additional ingredient needed to prove Theorem \ref{thm:Bat_Gamma_lb}, namely Theorem \ref{thm:cycle_with_phase}, is proved in Section \ref{sec:phase}. However, the geometric idea underlying our approach may not shine through the tropical combinatorics of the rigorous proofs. Therefore, in Section \ref{sec:disc} we explain the idea behind the proofs informally, emphasizing the relationship with the SYZ conjecture and the Gross--Siebert program. The reader who has no interest in informal discussions can skip Section \ref{sec:disc}. \paragraph{Acknowledgements.} We thank Denis Auroux for a helpful conversation at an early stage of this project. This work was done during the authors' stay at the Institute for Advanced Study (Fall 2016), Kyoto University (Winter 2017) and the Mathematical Sciences Research Institute (Spring 2018, supported by the National Science Foundation Grant Number DMS-1440140). M.A. was supported by the National Science Foundation through agreement number DMS-1609148 and DMS-1564172, and by the Simons Foundation through its ``Homological Mirror Symmetry'' Collaboration grant. S.G. was supported by the National Science Foundation through agreement number DMS-1128155. H.I.~was supported by JSPS KAKENHI Grant Number 16K05127, 16H06335, 16H06337, and 17H06127. N.S.~was partially supported by a Royal Society University Research Fellowship, a Sloan Research Fellowship, and by the National Science Foundation through Grant number DMS-1310604 and under agreement number DMS-1128155. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation. \section{Discussion and examples} \label{sec:disc} In this section we sketch the proof of the Gamma Conjecture for the structure sheaf on Batyrev mirrors of dimension at most $3$, emphasizing the relationship with the SYZ conjecture and the Gross--Siebert program. The discussion is not intended to be completely rigorous. Observe that the conjecture can be rewritten as \begin{equation} \label{eqn:gammai} \int_{C_t^+} \Omega_t + O\left(t^\epsilon\right) = \sum_{i=0}^n \frac{(-\log t)^{n-i}}{(n-i)!} \cdot \int_X \omega_\lambda^{n-i} \cdot \wh{\Gamma}_i \end{equation} where $\wh{\Gamma}_i$ is the degree-$2i$ component of $\wh{\Gamma}_X$. Roughly speaking we will stratify $C_t^+$ in accordance with the singularities of the SYZ fibration, and we will see that the codimension-$i$ strata give rise to terms in the asymptotic expansion of the period integral which precisely add up to the $i$th term on the right-hand side. This is a compelling picture, but unfortunately it becomes more complicated in higher dimensions (compare Remark \ref{rmk:kinks}) and we have not been able to cleanly generalize it. The remainder of the paper explains a more pedestrian version of our period computation which works in all dimensions, but which uses the embedding of Batyrev mirror pairs in toric varieties corresponding to dual reflexive polytopes. \subsection{Leading term} \label{sec:leading} We consider the map $\operatorname{Log}_t \colon P_{\C^*} \to P_\R$ \begin{align*} \operatorname{Log}_t(z_0,\ldots,z_n) = (\log_t|z_0|,\ldots,\log_t|z_n|). \end{align*} In the limit $t \to 0$, the \emph{amoeba} $\operatorname{Log}_t(\mathring{Z}_t)$ converges to the \emph{tropical amoeba}, which is a codimension-$1$ weighted balanced polyhedral complex \cite{Mikhalkin2004}. The unique compact component of the complement of the tropical amoeba is precisely the polytope $\Delta_\lambda$ that appears on the $A$-side of our mirror statement. The image of the cycle $C_t^+$ under $\operatorname{Log}_t$ converges to $\partial \Delta_\lambda$ as $t \to 0$, and the pullback of the volume form $\Omega_t$ to $C_t^+$ converges to the rescaling of the affine volume form on each face by $-\log t$. Using this we obtain that the leading term of the period integral is \[ \int_{C_t^+} \Omega_t = (-\log t)^n \cdot \operatorname{vol}\left( \partial \Delta_\lambda \right) + O\left((\log t)^{n-1}\right). \] The volume on the right-hand side coincides with the sum of symplectic volumes of boundary divisors $D_j \subset Y_{\Delta_\lambda}$ by \cite[Theorem 2.10]{Guillemin1994}. This coincides with the symplectic volume of $X$ (since $X$ is cohomologous to $\sum_j D_j$), which gives us the leading term in the Gamma Conjecture: \[ \int_{C_t^+} \Omega_t = \int_X t^{-\omega_\lambda} + O\left((\log t)^{n-1}\right). \] The sub-leading terms are related to the `bends' in $C_t^+$ where we interpolate between adjacent faces of $\partial \Delta_\lambda$, as we will see in the next sections. This is closely related to the SYZ conjecture, according to which there should exist a special Lagrangian torus fibration with singularities $Z_t \to B$, where $B\cong \partial \Delta_\lambda$ is endowed with an affine structure.\footnote{Proving the existence of such special Lagrangian torus fibrations remains a difficult question, presenting numerous challenges, see \cite{Joyce:SYZ}. In practice, our approach only requires the existence of a weak version of an SYZ fibration, similar to that appearing in the Gross--Siebert program. Nevertheless, for the purposes of this informal discussion, we will refer freely to special Lagrangian torus fibrations.} The cycle $C_t^+$ should correspond to the zero-section $B \subset Z_t$ of this fibration. The restriction of the holomorphic volume form $\Omega_t$ to the cycle $C_t^+$ is real, and should be approximately equal to the pullback of the affine volume form on $B$ (see e.g.~\cite[Sections 1-2]{Gross2013}). Thus the leading-order term of the period integral should be \[ \int_{C_t^+} \Omega_t = (-\log t)^n \cdot \operatorname{vol}(B) + O\left((\log t)^{n-1}\right).\] The mirror $X$ should admit a dual special Lagrangian torus fibration $X \to B$, and its symplectic volume should coincide with the affine volume of $B$. Thus we obtain an explanation of the leading term in the Gamma Conjecture that is similar to the previous one. As promised, the codimension-$0$ locus of the base of the SYZ fibration gave rise to the $i=0$ term on the right-hand side of \eqref{eqn:gammai}. \begin{rmk} The relationship between the leading order asymptotics of periods and tropical geometry has been studied by several people. Mikhalkin--Zharkov \cite{Mikhalkin-Zharkov:trop_Jac} introduced periods for tropical curves in terms of affine length; Iwao \cite{Iwao:int_trop} compared tropical periods for curves with the leading asymptotics of classical periods. Yamamoto \cite{Yamamoto:K3_periods} studied periods (or, radiance obstruction) of tropical K3 hypersurfaces and compared them with classical ones. \end{rmk} \subsection{K3 surfaces} Let us consider the two-dimensional case, so $Z_t$ and $X$ are K3 surfaces. There should be an SYZ fibration $p:Z_t \to B$ where $B \cong \partial \Delta_\lambda$, compare \cite{Gross2001,Gross2013,WDRuan:Lag_fibration_survey}. We have one affine coordinate chart of $B$ for each face of $\Delta_\lambda$, which has the subspace affine structure; and we also have an affine coordinate chart for each vertex, which is given by projection along the remaining `ray' emanating from the vertex (see Figure \ref{fig:tropK3}). \begin{figure}[ht] \centering \begin{tikzpicture}[x=1.2pt, y=1.2pt] \draw[very thick] (100,100)--(100,20)--(200,10)--(180,60)--(100,100); \draw[very thick] (100,100)--(200,10); \draw[->,very thick] (100,20)--(80,0); \draw[very thick] (200,10)--(220,-5); \draw[very thick] (180,60)--(200,70); \draw[->, very thick] (100,100)-- (90,130); \draw[very thin] (100,20)--(180,60); \draw[very thin] (120,40)--(100,100); \draw[very thin,->] (120,40)--(90,100); \draw[very thin,->] (120,40)--(88,72); \draw[very thin,->] (120,40)--(85,40); \draw[very thin] (120,40)--(100,20); \filldraw (100,40) circle [radius=1.5]; \filldraw (100,60) circle [radius=1.5]; \filldraw (100,80) circle [radius=1.5]; \filldraw (125,17.5) circle [radius=1.5]; \filldraw (150,15) circle [radius=1.5]; \filldraw (175,12.5) circle [radius=1.5]; \filldraw (125,77.5) circle [radius=1.5]; \filldraw (150,55) circle [radius=1.5]; \filldraw (175,32.5) circle [radius=1.5]; \filldraw (195,22.5) circle [radius=1.5]; \filldraw(190,35) circle [radius=1.5]; \filldraw(185,47.5) circle [radius=1.5]; \filldraw(120,90) circle [radius=1.5]; \filldraw(140,80) circle [radius=1.5]; \filldraw(160,70) circle [radius=1.5]; \filldraw (100,100) circle [radius=1.5]; \filldraw (100,20) circle [radius=1.5]; \filldraw (200,10) circle [radius=1.5]; \filldraw (180,60) circle [radius=1.5]; \fill[yellow, opacity=0.7] (109,29) -- (171,23) -- (109,79) -- (110,30); \shade[top color=blue, bottom color=blue!30,opacity=0.6] (100,100) -- (100,93) arc[x radius=7,y radius =7, start angle =-90, end angle =-24]--cycle; \shade[left color=blue,right color=blue!30,opacity=0.6] (100,75) arc[x radius =5, y radius =5, start angle=-90, end angle =90] -- cycle; \shade[left color=blue,right color=blue!30,opacity=0.6] (100,55) arc[x radius =5, y radius =5, start angle=-90, end angle =90] -- cycle; \shade[left color=blue,right color=blue!30,opacity=0.6] (100,35) arc[x radius =5, y radius =5, start angle=-90, end angle =90] -- cycle; \shade[bottom color=blue, top color=blue!30,opacity=0.6] (100,27) arc[x radius =7, y radius =7, start angle=90, end angle =-10] -- (100,20) -- cycle; \fill[red,opacity=0.1] (100,44) -- (104,46) -- (104,55) -- (100,56) -- cycle; \draw[red,very thick,opacity=0.4] (100,44) -- (104,46) -- (104,55) -- (100,56); \fill[red,opacity=0.2] (100,44) -- (106,44) -- (106,51) -- (100,56) --cycle; \draw[red, very thick] (100,44) -- (106,44) -- (106,51) -- (100,56); \end{tikzpicture} \caption{The tropical amoeba of a mirror quartic $\mathring{Z}_t = \{ tW_1+tW_2+tW_3+t/(W_1W_2W_3)=1\}$} \label{fig:tropK3} \end{figure} The resulting affine structure on $B$ is defined everywhere except near certain points along the edges of $\Delta_\lambda$, which correspond to the intersections of $Z_t$ with codimension-$2$ toric strata of $Y_\nabla$. Generically, there are 24 of these, so we end up with an affine structure on the 2-sphere with 24 singularities. Away from a neighbourhood of the singularities, the holomorphic volume form $\Omega_t$ is approximately equal to the flat volume form to order $O\left(t^\epsilon\right)$, so \[ \int_{p^{-1}(U)\cap C_t^+} \Omega_t = (-\log t)^2 \cdot \operatorname{vol}(U) + O\left(t^\epsilon\right).\] In a neighbourhood of a singularity, if we throw out terms of order $O\left(t^\epsilon\right)$ then the local model for $(Z_t,C_t^+,\Omega_t)$ is \begin{align*} Z_t &= \{(y_1,y_2,x) \in \C^2 \times \C^* : y_1 y_2=1+x\},\\ C_t^+ &= Z_t \cap (\R_+)^3,\\ \Omega_t &= \frac{dy_1 \wedge dy_2}{x} = d\log y_1 \wedge d\log x \end{align*} where $\{y_1y_2 = 0\}$ corresponds to the boundary divisor of $Y_\nabla$ (compare \cite{Kontsevich-Soibelman:affine}). \begin{example} Let $X \subset \C\mathbb{P}^3$ be a quartic K3 surface equipped with a symplectic form $\omega$ in the class $c_1(\C\mathbb{P}^3)$. The mirror is given by $\mathring{Z}_t=\{tW_1+tW_2+tW_3+t/(W_1W_2W_3)=1\}$. The tropical amoeba of $\mathring{Z}_t$ is shown in Figure \ref{fig:tropK3}; the $\operatorname{Log}_t$-image of the positive real cycle $C_t^+$ converges to the boundary of the simplex $\Delta_\lambda= \{w_1\ge -1, w_2\ge -1, w_3\ge -1, w_1+w_2+w_3\le 1\}$ as $t\to +0$, where $w_i= \log_t|W_i|$. We cover $\partial \Delta_\lambda$ by affine charts: on the interior of a facet (yellow region), we consider the subspace affine structure, and around a lattice point $v$ on an edge (blue region), we consider the affine structure given by the projection along the ray $\R_+ v$. The singularities of the affine structure occur somewhere between adjacent lattice points on edges. For example, consider the red region in Figure \ref{fig:tropK3}, which lies between the two affine charts $(w_2-w_1, w_2+w_3+1)$, $(w_2-w_1,w_3)$ associated with the rays $\R_+(-1,-1,1)$ and $\R_+(-1,-1,0)$. Since $tW_3=t^{1+w_3}$ and $t/(W_1W_2W_3) = t^{1-w_1-w_2-w_3}$ are exponentially small near the red region, the cycle $C_t^+$ in this region is given by the equation \[ tW_1 + t W_2 \approx 1 \quad \Longleftrightarrow \quad t^{-(w_2-w_1)} + 1 \approx t^{-w_2-1} = t^{-w_2-w_3-1}\cdot t^{w_3}. \] Setting $x=t^{-(w_2-w_1)}$, $y_1 = t^{-(w_2+w_3+1)}$, $y_2=t^{w_3}$, we find that the red region of the cycle is approximated by the positive real locus of the local model $1+x = y_1y_2$ above. Note that $(\log_t x, \log_t y_1)$ and $(\log_t x, \log_t y_2)$ give affine charts of the adjacent blue regions. \end{example} The base of the SYZ fibration at such a point is a `focus-focus singularity'. An approximation to the SYZ fibration $p:Z_t \to B$ can be written down away from the region where both $y_i$ are small. The approximation is defined using coordinates $b=-\log_t |x|$, $c_1=-\log_t |y_1|$, $c_2=-\log_t|y_2|$; we have $p \approx (b,c_1)$ away from $y_1=0$ and $p\approx (b,c_2)$ away from $y_2=0$. We observe that \[ c_1 + c_2 =- \log_t|y_1y_2| = -\log_t|1+x| \approx \max(0,b) \] when $b \gg 0$ or $b \ll 0$, so the transition maps for the approximate SYZ fibration are approximately affine-linear in these regions. The fact that these transition maps are different for large and small $b$ accounts for the non-trivial monodromy of the affine structure around the focus-focus singularity. \begin{figure}[ht] \centering \begin{tikzpicture} \node at (0,0) {\includegraphics[scale=0.4]{2dim_sing_4.pdf}}; \draw[blue,->] (3,1.5) -- (1.8,1.5) -- (1.4,1.1); \filldraw[blue] (3,1.5) node [fill=white,draw] {\scriptsize $c_1=\text{const}$}; \draw[red,->] (-2,2.5)--(-1.53,1.72); \filldraw[red] (-2.5,2.7) node [fill=white, draw] {\scriptsize $c_2=\text{const}$}; \draw[->] (-2.1,-3.25)--(-0.8,-3); \draw (-0.7,-3.2) node {\footnotesize $c_1$}; \draw[->] (-2.1,-3.25)--(-2.2,-1.25); \draw (-2,-1.4) node {\footnotesize $b$}; \draw[->] (-2.1,-3.25)--(-2.7,-1.75); \draw (-2.9,-1.9) node {\footnotesize $c_2$}; \draw[->] (5,0)--(10,0); \draw[->] (7.5,-3)--(7.5,3); \filldraw[opacity=0.3] (6,2)--(6,-2)--(7.5,-2)--(9,-0.5)--(9,2)--cycle; \draw (10.2,0) node {$b$}; \draw (7.5,3.2) node {$c_1$}; \draw (5.8,-0.2) node {$-B$}; \draw (9,-0.2) node {$B$}; \draw (7.3,2) node {$C_1$}; \draw (7.15,-2.07) node {$-C_2$}; \filldraw (7.5,0) node {$\times$} ; \draw[blue] (5,1) node {$c_1=\text{const}$}; \draw[blue] (6,1.6)--(9,1.6); \draw[blue] (6,1.2)--(9,1.2); \draw[blue] (6,0.8)--(9,0.8); \draw[blue] (6,0.4)--(9,0.4); \draw[red] (5,-1) node {$c_2=\text{const}$}; \draw[red] (6,-0.4)--(7.5,-0.4)--(9,1.1); \draw[red] (6,-0.8)--(7.5,-0.8)--(9,0.7); \draw[red] (6,-1.2)--(7.5,-1.2)--(9,0.3); \draw[red] (6,-1.6)--(7.5,-1.6)--(9,-.1); \end{tikzpicture} \caption{The left picture: the surface $c_1+c_2=-\log_t(1+t^{-b})$ with $t=1/e$. The red and blue coordinate lines show how the approximately-affine charts $(b,c_1)$, $(b,c_2)$ are glued. (This image was produced using the Wolfram Development Platform \cite{WDP:9July2018}.) The right picture: the affine manifold with singularity drawn on the $(b,c_1)$-plane (which corresponds to the red region in Figure \ref{fig:tropK3}); this arises from the left picture in the limit $t\to 0$.} \label{fig:2dim_sing} \end{figure} Figure \ref{fig:2dim_sing} shows the hypersurface $C_t^+ = \left\{c_1 + c_2 = -\log_t\left(1+t^{-b}\right)\right\} \subset \R^3$, with the horizontal coordinates corresponding to $c_1$ and $c_2$, and the vertical coordinate to $b$. On it we draw the level sets of the coordinates of the SYZ fibration, where each is defined. We cut out a region $p^{-1}(U)\cap C_t^+$ where $U$ is a neighbourhood of the singularity in the base of the SYZ fibration. It will have boundaries $-B \le b \le B$ for some large $B$; $c_1 \le C_1$ for some large $C_1$, so that $c_1$ is a coordinate of the approximate SYZ fibration along that boundary; and $c_2 \le C_2$ for some large $C_2$, for the same reason. We assume $B\ll C_1+C_2$ so that the boundaries $c_1=C_1$ and $c_2=C_2$ do not intersect. We integrate $\Omega_t = (-\log t)^2 db \wedge dc_1$ over $p^{-1}(U) \cap C_t^+$, which means we calculate the area of its projection to the $b$-$c_1$ plane. This is the area of the region $\{(b,c_1):-B \le b \le B, -\log_t(1+t^{-b}) - C_2 \le c_1 \le C_1\}$, which is clearly \[ (-\log t)^2 \int_{-B}^B \left( C_1+C_2+\log_t(1+t^{-b})\right) db. \] In contrast, the affine volume of $U$ will be \[ (-\log t)^2 \int_{-B}^B \left( C_1+C_2 -\max(0,b) \right) db. \] The difference between these two is the contribution of this region to the sub-leading terms of our period integral. It is equal to \begin{align*} (-\log t)^2 \int_{-B}^B \left( \max(0,b) + \log_t(1+t^{-b}) \right) db = - \zeta(2) + O(t^B) \end{align*} as we observed in the Introduction (see Section \ref{subsec:error}). Thus we have established that each of the $24$ singular points in the SYZ base (i.e., the codimension-$2$ strata) gives rise to a contribution of $-\zeta(2)$ to the sub-leading term in the period integral. These terms sum to \[-24 \zeta(2) = \int_X \wh{\Gamma}_2,\] using the fact that $\wh{\Gamma}_2 = -c_2(TX) =-24[{\rm pt}]$ for a $K3$ surface, which is the $i=2$ term in the right-hand side of \eqref{eqn:gammai} as promised. This completes the sketch proof of Theorem \ref{thm:gamma} in dimension $2$. The complete proof of Theorem \ref{thm:gamma} that we give in Section \ref{sec:gamma} applies even in situations where $X$ is not smooth but only quasi-smooth, which means (in this two-dimensional case) that some of the singular points in the SYZ base have collided. There is a new phenomenon here, which we briefly indicate without going into full details. We consider the tropical polynomial \[ f_a(b) := \max(-b,a,b),\] and the leading behaviour of the corresponding `error in tropicalization' integral \[ I(a,B,t) := (-\log t)^2 \int_{-B}^B \left(f_a(b)+ \log_t \left(t^{b}+t^{-a}+t^{-b} \right) \right) \, db\] as $t \to +0$, with $a,B$ held fixed and satisfying $|a|\ll B$. When $a > 0$, $f_a(b)$ has two bends and is `tropically smooth' at both (i.e., the slope changes by $1$). However when $a \le 0$, we have $f_a(b) = \max(-b,b)$ and the two bends have collided into a single bend which is not tropically smooth (the slope changes by $2$). This is reflected in the behaviour of the integral: when $a>0$, the two bends in $f_a$ each contribute $-\zeta(2)$ to the leading term of the integral, by the computation of Section \ref{subsec:error}, so $I(a,B,t) = -2\zeta(2)+O(t^\epsilon)$ for some $\epsilon>0$. When $a<0$ the terms involving $a$ contribute negligibly to the integral; after dropping these terms, a straightforward manipulation reduces the computation to that of Section \ref{subsec:error}, giving the answer $-\zeta(2)/2 + O(t^\epsilon)$. This reflects the fact that, although two separate focus-focus singularities each contribute $-\zeta(2)$ to the period integral, after they collide the contribution is only $-\zeta(2)/2$. The corresponding local model for $Z_t$ is given by $y_1y_2 = x+t^{-a}+x^{-1}$. The discontinuity of the constant term in the asymptotics of periods can be understood from the fact that the large-complex structure limit of $Z_t$ is different between $a>0$ and $a<0$. This collision of two focus-focus singularities in the SYZ base of $Z_t$ is mirror to a degeneration of $X$ so that it acquires an $A_1$ singularity. Indeed, a local picture for the development of this $A_1$ singularity is given by the family of toric varieties with moment polytopes $\{(b,c): c \ge f_a(b)\}$ as $a$ passes from positive to negative. We consider the effect of this degeneration on the $i=2$ term in the right-hand side of \eqref{eqn:gammai}, which is \[ \int_X \wh{\Gamma}_2 = -\zeta(2) \cdot \chi(X).\] For $a>0$, the local contribution to the Euler characteristic is $2$, from the two toric fixed points; for $a \le 0$ the local contribution is $1/2$, from the single toric fixed point which is an orbifold point of order $2$. Thus the effect of the collision of two focus-focus singularities on the period integral, and on the mirror integral \eqref{eqn:gammai}, is the same: $-2\zeta(2)$ gets replaced by $-\zeta(2)/2$. A similar phenomenon can be observed with the collision of $k$ focus-focus singularities, replacing $-k\zeta(2)$ with $-\zeta(2)/k$. \begin{rmk} The asymptotics becomes subtle when $a=0$. More generally, we can consider the local model $Z_t$ defined by $y_1y_2 = x + c t^{-a} + x^{-1}$. For non-zero $a$, the corresponding `error in tropicalization' integral does not depend on the coefficient $c$. For $a=0$, however, it depends analytically on $c$ as follows: \begin{align*} (-\log t)^2 \int_{-\infty}^\infty \left( \max(-b,b) + \log_t(t^b+c+t^{-b})\right) db & = -2 \int_0^1 \log(x^2+ c x + 1) \frac{dx}{x} \\ & = \arcsin^2\left(\frac{c}{2}\right) - \pi \arcsin\left(\frac{c}{2}\right) - \frac{\pi^2}{12}. \end{align*} It is interesting to note that this gives $-2 \zeta(2)$ for $c=2$. The point $(a,c)=(0,2)$ is the so-called conifold point in the complex moduli space, and should be mirror to a smoothing $T^*S^2$ of the $A_1$-singularity. The value $-2\zeta(2)$ can be interpreted as $-\chi(T^*S^2)\zeta(2)$. \end{rmk} \subsection{Threefolds} Now we consider the case where $Z_t$ is 3-dimensional. In this case there is again an SYZ fibration $p: Z_t \to B$ with $B \cong \partial \Delta_\lambda$, but the singular locus is more complicated: it generically consists of a trivalent graph lying inside the codimension-$1$ locus of $\partial \Delta_\lambda$ \cite{Gross2001,Gross2013,WDRuan:Lag_fibration_survey}, and there are two types of vertices: those lying in the interior of a codimension-$1$ face, with the three incident edges all lying in the same face (which we will call `type I'); and those lying at the intersection of three codimension-$1$ faces, with the three incident edges all lying in different faces (which we will call `type II'). \subsubsection{The edges} \label{sec:edges} Along an edge of the singular locus, the SYZ fibration is a product of the two-dimensional case previously considered with an $S^1$-fibration over an interval. Thus the integral along the edges should contribute $-\zeta(2)\cdot (-\log t) \cdot (\text{total length of edges})$, which comes out equal to \[ -\zeta(2) \cdot (-\log t) \cdot \int_X \omega_\lambda \cdot c_2(TX) = \int_X t^{-\omega_\lambda} \cdot \wh{\Gamma}_2\] as required ($c_2(TX)$ is represented by the singular locus of the fibration; see \cite[Theorem 2.17]{Gross2001}). \subsubsection{Type I vertex: $y_1y_2y_3=1+x$} \label{sec:typei} The local model near a type I vertex is \begin{align*} Z_t &= \{(y_1,y_2,y_3,x)\in \C^3\times \C^*: y_1 y_2 y_3 = 1+x\}\\ C_t^+ &= Z_t \cap (\R_+)^4\\ \Omega_t &= \frac{dy_1 \wedge dy_2 \wedge dy_3}{x} = d\log x \wedge d\log y_1 \wedge d\log y_2, \end{align*} where the boundary divisor of $Y_\nabla$ corresponds to $\{y_1y_2y_3=0\}$. The SYZ fibration is approximated using coordinates $b=-\log_t|x|$ and $c_i = -\log_t |y_i|$ as before. We set $p \approx (b,c_2,c_3)$ away from $y_1 = 0$, $p \approx (b,c_1,c_3)$ away from $y_2=0$, and $p \approx (b,c_1,c_2)$ away from $y_3=0$. Observe that $c_1+c_2+c_3 \approx \max(0,b)$ away from $b \approx 0$ as before, so once again the transition maps are affine-linear away from this area. The region $p^{-1}(U)\cap C_t^+$ will be cut out by inequalities $-B \le b \le B$, $c_i \le C_i$ as before, and we must calculate its projection to the $b$-$c_1$-$c_2$-plane. The projection to the $b$-$c_1$-$c_2$ is cut out by inequalities \[ -B \le b\le B,\quad c_1 \le C_1, \quad c_2 \le C_2, \quad c_1+c_2 \ge -\log_t(1+t^{-b}) - C_3. \] We assume $B\ll C_1+C_2+C_3$ to ensure that the fibre of this region over $b\in [-B,B]$ is nonempty. The fibre of this region over $b \in [-B,B]$ is a right-angle isosceles triangle whose sidelengths are easily calculated, which gives the total volume of the region as \[ (-\log t)^3\int_{-B}^B \frac{\left(C_1+C_2+C_3 + \log_t(1+t^{-b})\right)^2}{2} db. \] As before, we need to subtract off the affine volume of the region, which is \[ (-\log t)^3 \int_{-B}^B \frac{\left(C_1+C_2+C_3 - \max(0,b)\right)^2}{2} db. \] However even after subtracting off this volume, we will still get a divergent integral as $t$ goes to $+0$. That is because of the contributions from the edges of the discriminant locus: the three edges each contribute a term \[ -(-\log t)\cdot C_i \cdot \zeta(2) \approx (-\log t)^3 \cdot C_i \cdot \int_{-B}^B \left(\max(0,b) +\log_t(1+t^{-b}) \right) db \] to the integral. When we subtract off these contributions from the legs, we end up with the contribution which arises solely from the vertex of the discriminant locus, which is given by the integral \[ (-\log t)^3 \int_{-B}^B \frac{\left(-\log_t(1+t^{-b})\right)^2 - \max(0,b)^2}{2} db = \zeta(3) + O(t^\epsilon). \] We shall prove this later, see \eqref{eq:zeta_n}. \subsubsection{Type II vertex: $y_1y_2 = 1+x_1+x_2$} \label{sec:typeii} The local model near a type II vertex is \begin{align*} Z_t &= \{(y_1,y_2,x_1,x_2)\in \C^2\times (\C^*)^2:y_1y_2 = 1+ x_1+x_2\} \\ C_t^+ &= Z_t \cap (\R_+)^4\\ \Omega_t &= \frac{dy_1 \wedge dy_2 \wedge dx_1}{x_1x_2} = d\log x_1 \wedge d\log x_2 \wedge d\log y_1, \end{align*} where the boundary divisor of $Y_\nabla$ corresponds to $\{y_1y_2=0\}$. The SYZ fibration is approximated using coordinates $b_i =-\log_t |x_i|$ and $c_i = -\log_t |y_i|$. The region $p^{-1}(U)\cap C_t^+$ will be cut out by $(b_1,b_2) \in V$ for some region $V \subset \R^2$ enclosing the origin, together with $c_i \le C_i$, and we must calculate its projection to the $b_1$-$b_2$-$c_1$-plane. We assume that $\max(0,b_1,b_2)\ll C_1+C_2$ for $(b_1,b_2)\in V$. We find that this area is equal to \[ (-\log t)^3 \int_V \left(C_1+C_2 + \log_t(1+t^{-b_1}+t^{-b_2})\right) db_1 db_2. \] Once again we subtract off the affine volume of the region, leaving \[ (-\log t)^3\int_V \left( \max(0,b_1,b_2) + \log_t(1+t^{-b_1} + t^{-b_2}) \right) db_1 db_2. \] Next we need to subtract off the sum of the contributions from the edges of the discriminant locus, which is equal to $\zeta(2)$ multiplied by the total length $L$ of the standard tropical line $\operatorname{Sing}(\max(0,b_1,b_2))$ contained in the region $V$. We shall show in Proposition \ref{prop:asymptotics_of_integral} that \begin{equation} \label{eqn:typeII} \lim_{t\to +0} \left[ (-\log t)^3 \int_{V} \left(\max(0,b_1,b_2) + \log_t(1+t^{-b_1} + t^{-b_2}) \right) db_1 db_2 + \zeta(2)\cdot (-\log t) \cdot L \right] = -\zeta(3) \end{equation} so the contribution of a Type II vertex in the discriminant locus to the overall integral is $-\zeta(3)$. \begin{figure}[ht] \centering \begin{tikzpicture}[x=1pt, y=1pt] \filldraw[opacity=0.3] (110,10)--(110,50)--(150,90)--(190,90)--(190,50)--(150,10)--(110,10); \draw[very thick] (150,50)--(200,100); \draw[very thick] (150,50)--(100,50); \draw[very thick] (150,50)--(150,0); \begin{scope}[xshift=140] \filldraw[opacity=0.3] (110,10)--(110,50)--(110,90)--(230,90)--(230,10)--(110,10); \draw[very thick] (150,50)--(200,100); \draw[very thick] (150,50)--(100,50); \draw[very thick] (150,50)--(150,0); \end{scope} \end{tikzpicture} \caption{Regions $V$ with kinks (left) and without kinks (right) along the tropical line.} \label{fig:kink} \end{figure} \begin{rmk} \label{rmk:kinks} There is an important issue which we have glossed over in this computation: in order for \eqref{eqn:typeII} to hold, the boundary of the region $V$ should be smooth and transverse to the edges of the discriminant locus (i.e., the legs of the tropical line) where it crosses them. For example, if one takes the region $V$ shown in the left side of Figure \ref{fig:kink}, the value of the integral \eqref{eqn:typeII} will be equal to $5\zeta(3)/4$ (see the proof of Proposition \ref{prop:asymptotics_of_integral}). Some of the contribution of the vertex is `hiding in the kinks in the boundary of $V$' in this case. It turns out that in higher dimensions, the local contribution is even more strongly dependent on the shape of the region $V$. For example it is not enough that $V$ have no `kinks' where it crosses the discriminant locus: in dimensions $\ge 4$ the integral may in general depend on the angle at which $V$ intersects the singular locus. We have not found a way to organize these choices efficiently. In Section \ref{sec:periodR} we take the more pedestrian approach of decomposing the cycle into pieces in a completely canonical way, at the cost of leaving certain `kinks' in the pieces which result in a formula which is less visibly `local' in the base of the SYZ fibration. \end{rmk} \subsubsection{Proof of Theorem \ref{thm:gamma} in dimension 3} We can piece together a sketch proof of Theorem \ref{thm:gamma} in dimension $3$ from the pieces we have assembled. In Section \ref{sec:leading} we have seen that the codimension-$0$ strata of the base of the SYZ fibration contribute the $i=0$ term on the right-hand side of \eqref{eqn:gammai}; in Section \ref{sec:edges} we have seen that the edges (codimension-$2$ strata) contribute the $i=2$ term; it remains to see how the type I and type II vertices contribute the $i=3$ term. It is clear that their contribution is \[ \left( \#(\text{type I vertices}) - \#(\text{type II vertices})\right) \cdot \zeta(3),\] which we must show is equal to \[ \int_X \wh{\Gamma}_3 = -2\zeta(3) \cdot \int_X \ch_3(TX) = -\zeta(3) \cdot \int_X c_3(TX) = -\zeta(3) \cdot \chi(X). \] The answer now follows from the observation that in the stratification of $X$ according to singularities of the SYZ fibration, each stratum has an $S^1$ factor and therefore vanishing Euler characteristic except for those lying over the vertices of the discriminant locus. The mirror to a type I SYZ fibre is a type II SYZ fibre, which has Euler characteristic $-1$; whereas the mirror to a type II SYZ fibre is a type I SYZ fibre, which has Euler characteristic $+1$. Therefore we have \[ \chi(X) = \#(\text{type II vertices}) - \#(\text{type I vertices}),\] which completes the sketch of a proof. \begin{example} The SYZ fibration on the quintic threefold has ${5 \choose 3} \cdot 5 = 50$ vertices of type I and ${5 \choose 2} \cdot 5^2 = 250$ vertices of type II. The Euler characteristic of the mirror quintic is $250-50 = 200$, as one clearly sees from its Hodge diamond \cite{CdlOGP:pair}. \end{example} \section{Proof of Theorem \ref{thm:periodR}} \label{sec:periodR} In this section we prove Theorem \ref{thm:periodR}. We break up the mirror period integral into pieces corresponding to a polyhedral decomposition of $\partial \Delta_\lambda$, which is the limit shape of $C_t^+$. Then we express each piece in terms of integrals over the toric variety $Y_{\Delta_\lambda}$ by applying the Duistermaat--Heckman theorem. \subsection{Tropical setup} \label{subsec:tropical_setup} Consider the affine functions \begin{align*} \beta_q\colon P_\R \to \R, \qquad \beta_q(p) = \langle q,p \rangle + \lambda_q, \end{align*} as well as the map \begin{align*} i_t\colon P_\R \to P_{\C^*}, \qquad i_t(p_0,\dots,p_n) = \left(t^{p_0},\dots,t^{p_n}\right) \end{align*} for $t \in \R_+$, which is right-inverse to $\operatorname{Log}_t$. If we define \[B_t := \left\{ 1 = \sum_{q \in V} t^{\beta_q}\right\} \subset P_\R,\] then it is clear that $i_t(B_t) = C_t^+$. Therefore \[ \int_{C_t^+} \Omega_t = \int_{B_t} i_t^* \Omega_t.\] We observe that \[ \log_t\left|t^{\lambda_q} \cdot z^q\right| = \beta_q(\operatorname{Log}_t(z)) \] so $\beta_q$ is the `tropical monomial' corresponding to the honest monomial $t^{\lambda_q} \cdot z^q$. As a result, in the limit $t \to 0$, the \emph{amoeba} $\operatorname{Log}_t(\mathring{Z}_t)$ converges to the \emph{tropical amoeba} $\operatorname{Sing}(\min(0, \{\beta_q\}_{q\in V}))$, the non-smooth locus of the piecewise affine-linear function $\min(0,\{\beta_q\}_{q\in V})$ on $P_\R$ \cite{Mikhalkin2004}. We observe that the unique compact component of the complement of the tropical amoeba is precisely the polytope $\Delta_\lambda$ that appears on the $A$-side of our mirror statement. In the limit $t\to 0$, $B_t$ converges to the boundary $\partial\Delta_\lambda$ of the polytope. \subsection{Decomposing the domain} \label{sec:decomposing-domain} We now decompose the domain $B_t$ of our period integral into regions where the different monomials dominate. \begin{figure}[ht] \centering \begin{tikzpicture}[x=1pt, y=1pt] \filldraw[gray!30!white] (148,30)--(144,36)--(187,36)--(181,30)--(148,30); \filldraw[gray] (148,30)--(144,36) -- (136,36) -- (140,30) -- (148,30); \draw[very thick] (100,90)--(200,90)--(220,60)--(190,30)--(140,30)--(100,90); \draw[very thick] (233,25)--(196,-12)--(124,-12)--(80,55); \draw[very thick] (100,90)--(80,55); \draw[very thick] (220,60)--(233,25); \draw[very thick] (190,30)--(196,-12); \draw[very thick] (140,30)--(124,-12); \draw[dotted] (200,90)--(213,55); \draw[dotted] (80,55)--(213,55)--(233,25); \draw[very thin] (148,30)--(108,90); \draw[very thin] (136,36)--(196,36); \draw[very thin] (104,84)--(204,84); \draw[very thin] (193,90)--(215,55); \draw[very thin] (181,30)--(216,65); \draw (160,62) node {face $q$}; \draw (165,10) node {face $k$}; \draw (115,39) node {\rotatebox{-56}{face $j$}}; \end{tikzpicture} \caption{Decomposition of the cycle $B_t$. The cycle $B_t$ approaches to the boundary of the polytope $\Delta_\lambda$ as $t\to +0$. The light grey region is the limit of $B_t^{q,\{k\}}$ and the dark grey region is the limit of $B_t^{q,\{k,j\}}$; these pieces $B_t^{q,K}$ can be presented as graphs over the shaded regions.} \label{fig:piece} \end{figure} We cover $P_\R$ with the sets \[ U^q := \left\{p \in P_\R: \beta_i(p) \ge \beta_q(p) \text{ for all $i \in V$}\right\}\] for $q \in V$. Thus we can cover $B_t$ with the sets $B^q_t := U^q \cap B_t$. In the limit $t \to 0$, $B_t^q$ converges to the $q$th face of $\Delta_\lambda$. The above cover is well-adapted to consider tropical limits, but our analysis of sub-leading terms requires a further decomposition. Let us fix $\epsilon >0$, and for each $\{q\} \sqcup K \subset V$ set \[ U^{q,K} := \left\{ p \in P_\R : \begin{aligned} \beta_k(p) - \beta_q(p) & \in [0,\epsilon] & &\text{for $k\in \{q\}\cup K$} \\ \beta_m(p )- \beta_q(p) &\in [\epsilon, \infty) & & \text{for $m\in V\setminus (\{q\}\cup K)$} \end{aligned} \right\} \] In words, $U^{q,K}$ is the region where the tropical monomial $\beta_q$ is smallest (hence `dominates') and the tropical monomials $\{\beta_k\}_{k \in K}$ are not far behind. We observe that $U^q$ is covered by the sets $U^{q,K}$. Then we obtain a cover $B_t^{q,K} := U^{q,K} \cap B_t$ (see Figure \ref{fig:piece}). So our period integral is equal to \[ \sum_{q,K} \int_{B_t^{q,K}} i_t^* \Omega_t.\] If we choose $\epsilon>0$ small enough, then $B_t^{q,K}$ is nonempty for sufficiently small $t>0$ if and only if the facets $\{\beta_i=0\}\cap \Delta_\lambda$ with $i \in\{q\}\sqcup K$ have nonempty intersection, or equivalently, $\{q\}\sqcup K$ spans a cone of the fan $\Sigma_\lambda$. Starting in the next section, and going through the end of Section \ref{sec:duistermaat-heckman}, we will restrict to pairs $(q,K)$ such that the facets corresponding to $\{q\}\sqcup K$ intersect. \subsection{Approximation in each region} \label{sec:approx_each_region} Let us consider the integral over $B_t^{q,K}$. Observe that \begin{align} \label{eq:f_t_on_region} f_t(z) &= t^{\lambda_q}z^q \left (1 + \sum_{k\in K} t^{\lambda_k-\lambda_q} z^{k-q} + h_t(z)\right), \end{align} where \[ h_t(z) = \sum_{j\in V \setminus (\{q\}\sqcup K)} t^{\lambda_j-\lambda_q} \cdot z^{j-q}. \] We observe that over $i_t(B_t^{q,K})$, we have \begin{equation} \label{eqn:negl} h_t, \ z_i\frac{\partial h_t}{\partial z_i} \in O\left(t^\epsilon\right) \end{equation} because each contributing monomial is so. The idea for approximating the integral over $B_t^{q,K}$ is to `throw away' these negligible terms. In order to evaluate the integral over the region $B_t^{q,K}$, we introduce an affine coordinate system $(a,b_k,c_j)$ on $P_\R$: \begin{align*} a &= \beta_q \\ b_k &= \beta_k - \beta_q \quad\text{for $k \in K$} \\ \{c_j\} &= \text{a collection of integral linear functions completing a coordinate system}. \end{align*} The fact that it is possible to complete $\{a,b_k\}$ to a coordinate system follows from our assumption that the fan $\Sigma_\lambda$ is simplicial and that the facets corresponding to $\{q\}\sqcup K$ intersect. We also write \[ r_{q,K} \cdot da \prod_{k\in K} db_k \prod_j dc_j = \text{the standard affine volume form on $P_\R$} \] for some factor $r_{q,K}>0$ and set \begin{equation} \label{eq:residual_vol} d\operatorname{vol}_{q,K}= r_{q,K} \prod_{j} dc_j \end{equation} for the residual volume form on the $c$-plane. When $a$, $b_k$ already form a coordinte system and there are no $c$ variables, we regard $d\operatorname{vol}_{q,K}$ as a measure on the point $\{0\}=\R^0$. Note that this is different from the affine volume form induced on a subspace of the form $\{a=\text{const}, b_k= \text{const}\}$ unless the covectors $da=q$, $db_k=k-q$ are part of a $\Z$-basis of $P^\vee$. We introduce the corresponding monomials on $P_{\C^*}$: \begin{align*} w &= t^{\lambda_q} \cdot z^q \\ x_k &= t^{\lambda_k - \lambda_q} \cdot z^{k-q} \quad \text{ for $k \in K$} \\ y_j &= z^{c_j}. \end{align*} In these coordinates we have \[ f_t(w,x,y) = w \cdot \left( 1+\sum_{k \in K} x_k + h_t(x,y,w)\right). \] By the definition of $r_{q,K}$, the standard holomorphic volume form on $P_{\C^*}$ is given by $r_{q,K}\cdot d\log w\wedge \bigwedge_{k\in K} d\log x_k \wedge \bigwedge_j d\log y_j$. Thus the volume form of $\mathring{Z}_t$ is \begin{align} \label{eq:Omega_in_xy} \begin{split} \Omega_t &= r_{q,K} \cdot \left. \frac{d\log w \wedge \bigwedge_{k\in K}d\log x_k \wedge \bigwedge_j d\log y_j}{df_t(w,x,y)} \right|_{\mathring{Z}_t} \\ &= r_{q,K}\cdot \left. \frac{\bigwedge_{k\in K}d\log x_k \wedge \bigwedge_j d\log y_j}{w\cdot (\partial f_t(w,x,y)/\partial w)} \right|_{\mathring{Z}_t} \end{split} \end{align} in the region where the denominator does not vanish. On $C_t^+=i_t(B_t)=\{z\in P_{\R_+} : 1=f_t(z)\}$, we have \begin{align*} \label{eq:estimate_wf_w} w \cdot \frac{\partial f_t}{\partial w} (w,x,y) &= f_t(w,x,y) + w^2 \frac{\partial h_t}{\partial w}(w,x,y) && \\ & = 1 + O\left(t^\epsilon\right) && \text{over $i_t(B_t^{q,K})$} \end{align*} where we used \eqref{eqn:negl} and the fact that $0<w<f_t(x,y,w)=1$ on $C_t^+$. Therefore we have \begin{equation*} \Omega_t = \left( 1 + O\left(t^\epsilon\right) \right) r_{q,K} \cdot \bigwedge_{k\in K}d\log x_k \wedge \bigwedge_j d\log y_j, \end{equation*} so \begin{align} \label{eq:Omega_in_bc} \begin{split} \int_{B_t^{q,K}} i_t^* \Omega_t & = \left( 1 + O\left(t^\epsilon\right) \right) \cdot r_{q,K}\cdot \int_{B_t^{q,K}} \bigwedge_{k\in K} d\log\left(t^{b_k}\right) \wedge \bigwedge_j d\log\left(t^{c_j}\right)\\ &= \left( 1 + O\left(t^\epsilon\right) \right) \cdot (-\log t)^n \cdot \operatorname{vol}\left(\pi_{b,c}\left(B_t^{q,K}\right)\right), \end{split} \end{align} where $\pi_{b,c}$ denotes the projection to the $(b,c)$-plane and $\operatorname{vol}$ denotes the volume with respect to the product of $\prod_{k\in K} db_k$ and the residual volume form $d\operatorname{vol}_{q,K}$ in \eqref{eq:residual_vol}. See the remark below for the reason why $(-\log t)^n$ instead of $(\log t)^n$ appears in the last expression. \begin{rmk} \label{rmk:sign_orientation} We have been vague about how we choose an order of the coordinates $a$, $b_k$, $c_j$ (or $w$, $x_k$, $y_j$) and an orientation of the cycle $B_t$; strictly speaking we need them to define $\Omega_t$ and the integral. For convenience, we shall always arrange these choices so that $i_t^*\Omega_t$ defines a positive measure (density) on $B_t$. Note that the factor $(-\log t)^n$ appearing in the above formula is positive since $(-\log t)>0$ for a sufficiently small $t$. \end{rmk} \subsection{Approximation in terms of volumes of polytopes} \label{sec:approx_vol_polytope} We now approximate the affine volume of $\pi_{b,c}(B_t^{q,K})$ in terms of the volumes of certain polytopes. On $B_t^{q,K}$, the defining equation can be rewritten as \begin{align} \label{eq:defining_eqn_Bt} 1 = \sum_p t^{\beta_p} \iff 1 = t^a \cdot \bigg(1+\sum_{k \in K} t^{b_k} + \underbrace{\sum_{m \in V \setminus (\{q\} \sqcup K)} t^{\beta_m-a}}_{O(t^\epsilon)}\bigg) , \end{align} which can be used to write $a$ as a function $a_{q,K}(b,c)$ of the variables $b_k,c_j$. We observe that we have the approximation \begin{equation} \label{eq:approx_aqK} a_{q,K}(b,c) = a'_{q,K}(b) + O\left(t^\epsilon\right) \quad \text{ where} \quad a'_{q,K}(b) := -\log_t\left(1+\sum_{k \in K} t^{b_k} \right). \end{equation} Now $B_t^{q,K}$ is defined by the inequalities \begin{align*} b_k & \in [0,\epsilon] && \text{ for $k \in K$} \\ \beta_m(a,b,c) - a & \in [\epsilon,\infty)&& \text{ for $m \notin \{q\} \sqcup K$}, \\ \intertext{which means the region $\pi_{b,c}\left(B_t^{q,K}\right)$ is defined by the inequalities} b_k & \in [0,\epsilon] && \text{ for $k \in K$} \\ \beta_m(a_{q,K}(b,c),b,c) - a_{q,K}(b,c) & \in [\epsilon,\infty) && \text{ for $m \notin \{q\}\sqcup K$}. \end{align*} We will consider the fibres $F_{q,K}(b)$ of the projection \[ \pi_b: \pi_{b,c}\left(B_t^{q,K}\right) \to [0,\epsilon]^K.\] It is clear that \[ \operatorname{vol}\left(\pi_{b,c}\left(B_t^{q,K}\right)\right) = \int_{[0,\epsilon]^K} \operatorname{vol}\left(F_{q,K}(b)\right) db,\] where we use the volume form $d\operatorname{vol}_{q,K}$ \eqref{eq:residual_vol} on the $c$-plane to define $\operatorname{vol}(F_{q,K}(b))$, so our next project is to approximate the volume of the fibres $F_{q,K}(b)$. We claim that \[ \operatorname{vol}(F_{q,K}(b)) = \operatorname{vol}(F'_{q,K}(a'_{q,K}(b),b)) + O\left(t^\epsilon\right) \] where $F'_{q,K}(a,b)$ is the compact polytope in the $c$-plane defined by \[ \beta_m(a,b,c) - a \in [\epsilon,\infty) \quad \text{ for $m \notin \{q\}\sqcup K$} \] with fixed $(a,b)$. Indeed, this follows because $F_{q,K}(b)$ can be sandwiched between two perturbations of the compact polytope $F'_{q,K}(a'_{q,K}(b),b)$ where the facets have been shifted by quantities of order $O\left(t^\epsilon\right)$. We have succeeded in approximating the volume of $\pi_{b,c}(B_t^{q,K})$ in terms of the volumes of the polytopes $F'_{q,K}(a,b)$, but we would prefer to work with the volumes of the polytopes $E_{q,K}(a,b)$ defined by \[ \beta_m(a,b,c) - a \ge 0 \quad \text{ for $m \notin \{q\} \sqcup K$} \] with fixed $(a,b)$. We shall regard $F'_{q,K}(a,b)$ and $E_{q,K}(a,b)$ either as polytopes in the $c$-plane or as subsets of $P_\R$ with the values of $(a,b)$ fixed. More generally, for any subset $K' \subset V$ not containing $q$, we write $E_{q,K'}(a,\{b_k\}_{k\in K'})$ for the polytope in $P_\R$ defined by \[ \beta_m - a \ge 0 \quad \text{ for $m\notin \{q\} \sqcup K'$} \] with the values of $a=\beta_q$ and $b_k =\beta_k - \beta_q$ (with $k\in K'$) fixed. We have \[ F'_{q,K}(a,b)=E_{q,K}(a,b) \setminus \bigcup_{j\notin \{q\}\sqcup K} \bigcup_{b_j \in [0,\epsilon]} E_{q,K\sqcup\{j\}}(a,b,b_j). \] This volume can be computed by the inclusion-exclusion principle: noting that \[ \bigcap_{j\in I} \bigg(\bigcup_{b_j \in [0,\epsilon]} E_{q, K\sqcup\{j\}}(a,b,b_j)\bigg) = \bigcup_{b'\in [0,\epsilon]^I} E_{q,K\sqcup I}(a,b,b') \] for $I$ disjoint from $\{q\}\sqcup K$, we obtain \[ \operatorname{vol}(F'_{q,K}(a,b)) = \sum_{J : J \supset K, q\notin J} (-1)^{|J \setminus K|} \cdot \int_{[0,\epsilon]^{J \setminus K} } \operatorname{vol}(E_{q,J}(a,b,b')) db' \] where we write $b=(b_k)_{k\in K}$, $b'=(b_j)_{j\in J\setminus K}$, and use the volume form $d \operatorname{vol}_{q,J}$ to define $\operatorname{vol}(E_{q,J}(a,b,b'))$. This means our period integral becomes \begin{align} \label{eqn:perH} \begin{split} \int_{C_t^+} \Omega_t & = \left(1+O\left(t^\epsilon\right)\right) (-\log t)^n \left(\sum_{q,K,q\notin K} \int_{[0,\epsilon]^K} \operatorname{vol}\left(F'_{q,K}\left(a'_{q,K}(b),b\right)\right) db + O\left(t^\epsilon\right)\right) \\ &= \left(1+O\left(t^\epsilon\right)\right) (-\log t)^n \left(\sum_{q,K \subset J, q\notin J} (-1)^{|J \setminus K|} \int_{[0,\epsilon]^J} \operatorname{vol}\left(E_{q,J}\left(a'_{q,K}(b),b,b'\right)\right) db db' + O\left(t^\epsilon\right) \right). \end{split} \end{align} \subsection{Duistermaat--Heckman} \label{sec:duistermaat-heckman} We apply the Duistermaat--Heckman theorem to express the volumes of polytopes in \eqref{eqn:perH} as symplectic volumes. \begin{lem} \label{lem:DH} For positive, sufficiently small $a$ and $b_j$ with $j\in J$, we have \[ \operatorname{vol}(E_{q,J}(a,b)) = \int_{Y_{\Delta_\lambda}} \exp\left(\omega_\lambda - \sum_{j \in J} b_j \cdot D_j -a \cdot \sigma \right) \cdot D_q \cdot \prod_{j \in J} D_j , \] where $D_j \subset Y_{\Delta_\lambda}$ denotes the toric divisor corresponding to the $j$th facet $\{\beta_j=0\}\cap \Delta_\lambda$ of $\Delta_\lambda$, and $\sigma:=\sum_{j \in V} D_j$. The right-hand side vanishes when the facets corresponding to the elements of $\{q\} \sqcup J$ do not intersect. \end{lem} \begin{proof} We use the Duistermaat--Heckman theorem to identify the volume of $E_{q,J}(q,b)$ with the symplectic volume of a toric subvariety of $Y_{\Delta_\lambda}$. The polytope $E_{q,J}(a,b)$ is defined by \begin{align*} \beta_q &= a \\ \beta_j - a &= b_j && \text{for $j \in J$} \\ \beta_m - a & \ge 0 && \text{for $m \in V\setminus (\{q\}\sqcup J)$},\\ \intertext{which is equivalent to} \beta_q - a & =0 \\ \beta_j - b_j - a &= 0 && \text{for $j \in J$} \\ \beta_m - a & \ge 0 && \text{for $m \in V\setminus (\{q\} \sqcup J)$}. \intertext{ This is precisely the face of the polytope $\Delta_{\lambda'}$ corresponding to the set $\{q\} \sqcup J$, where} \lambda'_j &= \lambda_j - b_j - a && \text{for $j \in J$} \\ \lambda'_m &= \lambda_m - a && \text{for $m \in V\setminus J$}. \end{align*} When $a$ and $\{b_j\}_{j\in J}$ are sufficiently small, the combinatorial type of $\Delta_{\lambda'}$ is the same as that of $\Delta_\lambda$, and the volume of the face corresponding to $\{q\} \sqcup J$ is equal to the symplectic volume of the stratum \[ D_q \cap \bigcap_{j \in J} D_j\] with respect to a symplectic form in cohomology class \[ \left[\omega_{\lambda'}\right] = \left[\omega_\lambda \right] - \sum_{j \in J} b_j \cdot D_j - a \cdot \sum_{j \in V} D_j,\] by \cite[Theorem 2.10]{Guillemin1994}. This yields the result. The right-hand side vanishes if $D_q \cap \bigcap_{j\in J} D_j= \emptyset$, and therefore if the facets of $\Delta_\lambda$ from $\{q\}\sqcup J$ do not intersect. \end{proof} \begin{rmk} We hid some technical details when applying the Duistermaat--Heckman theorem. When the covectors $da$, $db_j$ are not part of a $\Z$-basis of $P^\vee$, the corresponding toric substack $D_q\cap \bigcap_{j\in J}D_j$ has a generic stabilizer. The order of the generic stabilizer equals the ratio between the affine volume form of the face corresponding to $\{q\}\sqcup J$ and the residual volume form $d\operatorname{vol}_{q,J}$ on the $c$-plane. Since, by definition, the integral over $D_q \cap \bigcap_{j\in J} D_j$ is the integral over the coarse moduli space divided by the order of the generic stabilizer, the volume of $E_{q,J}(a,b)$ with respect to $d\operatorname{vol}_{q,J}$ gives the correct answer. \end{rmk} We now substitute this into \eqref{eqn:perH}: we can ensure that $b_j$ in \eqref{eqn:perH} is sufficiently small by making $\epsilon>0$ small, and also ensure that $a'_{q,K}(b)$ in \eqref{eqn:perH} is sufficiently small by making $t>0$ small because of the estimate: \[ 0\le a'_{q,K}(b)= \frac{\log(1+\sum_{k\in K} t^{b_k})}{-\log t} \le \frac{\log |V|}{-\log t}. \] We now obtain: \begin{align*} \int_{C_t^+} \Omega_t &= \left(1 + O\left(t^\epsilon\right)\right) \cdot \left( \int_{Y_{\Delta_\lambda}} P_t(D)_{n+1} + O\left((-\log t)^n t^\epsilon\right) \right), \end{align*} where \begin{align*} P_t(D) &= (-\log t)^n \cdot \sum_{q,K \subset J, q\notin J} (-1)^{|J \setminus K|} \int_{[0,\epsilon]^J} e^{\omega_\lambda - \sum_{j \in J} b_j \cdot D_j + \sigma \cdot \log_t\left(1+\sum_{k \in K} t^{b_k}\right)} db \cdot D_q \prod_{j \in J} D_j \end{align*} The subscript `$n+1$' denotes the part of $P_t(D)$ in degree $2(n+1)$: that is the only part that gets hit by the integral $\int_{Y_{\Delta_\lambda}}$. The summand for $(J,q)$ automatically vanishes unless the facets corresponding to $\{q\} \sqcup J$ intersect, in particular, unless the facets corresponding to $\{q\} \sqcup K$ intersect. Therefore we can now withdraw the assumption imposed at the end of Section \ref{sec:decomposing-domain} that the facets corresponding to $\{q\}\sqcup K$ intersect and consider the sum over arbitrary $K, J, q$ with $K\subset J$ and $q\notin J$. \subsection{End of the proof} By replacing $\omega_\lambda$, $D_j$, $\sigma$ in the definition of $P_t(D)$ with $(-\log t) \omega_\lambda$, $(-\log t) D_j$, $(-\log t) \sigma$ respectively, we obtain \begin{align*} (-\log t)^n \sum_{q,K\subset J, q\notin J} (-1)^{|J\setminus K|} \int_{[0,\epsilon]^J} e^{(-\log t)(\omega_\lambda - \sum_{j\in J} b_j \cdot D_j) - \sigma \log (1+\sum_{k\in K} t^{b_k})}db \cdot (-\log t)^{|J|+1} D_q \prod_{j\in J} D_j \end{align*} The degree $2(n+1)$ part of this quantity equals $(-\log t)^{n+1} P_t(D)_{n+1}$. Making the substitution $s_j = -\log t \cdot b_j$, we find that $P_t(D)_{n+1} = Q_t(D)_{n+1}$ with \[ Q_t(D) = \sum_{q,K \subset J,q\notin J} (-1)^{|J \setminus K|} \cdot \int_{[0,-\epsilon \log t]^J} t^{-\omega_\lambda} \cdot e^{-\sum_{j \in J} s_j \cdot D_j - \sigma\cdot \log\left(1+\sum_{k \in K} e^{-s_k} \right)} ds \cdot D_q\prod_{j \in J} D_j \] where the factor $(-\log t)^{|J|}$ is absorbed by $d b$ to become $d s$. By expanding the exponential, we find that \begin{align*} Q_t(D) &= t^{-\omega_\lambda} \left(\sigma+ \sum_{q,J,\ell,\vec{m}: q\notin J, J\neq \emptyset} \frac{I^\epsilon_{\ell;\vec{m}}(t)}{\ell! \prod_{j\in J}m_j!} \cdot D_q \cdot (-\sigma)^\ell \prod_{j \in J} (-D_j)^{m_j + 1}\right), \end{align*} where the sum is over $\ell \in \Z_{\ge 0}$, $\vec{m} \in (\Z_{\ge 0})^J$, $q$, $J$ with $q\notin J$ and \[ I_{\ell;\vec{m}}^\epsilon(t) := \int_{[0,-\epsilon \log t]^J} \prod_{j\in J} s_j^{m_j} \cdot g_\ell\left(\{e^{-s_i}\}_{i \in J}\right) ds \] is an $\epsilon$-truncated version of the `local integral' $I_{\ell,\vec{m}}$ in \eqref{eqn:Ilm}. We note that the first term $\sigma=\sum_{q\in V} D_q$ arises from the case where $J=K=\emptyset$. \begin{lem} \label{lem:convergence_Ilm}We have $I^\epsilon_{\ell;\vec{m}}(t) = I_{\ell;\vec{m}} + O\left((-\log t)^{|\vec{m}|} t^\epsilon\right)$ as $t\to +0$, where $|\vec{m}|=\sum_{j\in J} m_j$. \end{lem} \begin{proof} We recall the bound \eqref{eqn:boundint}, which was used to prove exponential decay of the integrand at infinity. It gives \[\prod_{j \in J}s_j^{m_j} \cdot g_\ell\left(\{e^{-s_i}\}_{i \in J}\right) \le C_\ell \cdot \prod_{j \in J} e^{-s_j} s_j^{m_j} \qquad\text{for $s_j \ge 0$.}\] The order estimate then follows by \[ \int_{-\epsilon \log t}^\infty s^{m} e^{-s} ds = O\left((-\log t)^m t^\epsilon\right). \] \end{proof} Now observe that $I_{0,\vec{m}} = 0$, and the anticanonical hypersurface $X$ is homologous to the toric boundary divisor $\sigma=\sum_{q\in V}D_q$, so we have proved \[ \int_{C_t^+} \Omega_t = \left(1 + O\left(t^\epsilon\right)\right) \cdot \left(\int_X t^{-\omega_\lambda} \cdot \wh{G}_X + O\left((-\log t)^n t^\epsilon\right) \right) \] where $\wh{G}_X$ is given in \eqref{eq:hG_X}. Because $-\log t = O(t^{-\delta})$ for any $\delta >0$, we can absorb the error terms depending on $\log t$ by reducing $\epsilon$, and thereby obtain \[ \int_{C_t^+} \Omega_t = \int_X t^{-\omega_\lambda} \cdot \wh{G}_X + O\left(t^\epsilon\right) \] for some (new, smaller) $\epsilon > 0$. This completes the proof of Theorem \ref{thm:periodR}. \section{Proof of Theorem \ref{thm:gamma}} \label{sec:gamma} \subsection{Formula for the Gamma class} Since the Gamma class is multiplicative, the short exact sequence $0 \to TX \to TY \to NX \to 0$ gives \begin{align*} \wh{\Gamma}(TX) &= \frac{\wh{\Gamma}(TY)}{\wh{\Gamma}(NX)}. \end{align*} The Euler sequence on the toric variety $Y$ gives the following expression for its Gamma class: \[ \wh{\Gamma}(TY) = \prod_{j \in V} \Gamma(1+D_j)\] (compare \cite[Proposition 13.1.2]{CLS}). Setting $\sigma := \sum_{j\in V} D_j$ as before, we have \[ \wh{\Gamma}(NX) = \Gamma(1+\sigma)\] because $X$ is anticanonical and $K_Y = -\sigma$. Thus we have the formula \begin{equation} \label{eq:1Gamma_X} \wh{\Gamma}_X = \frac{\prod_{j \in V} \Gamma(1+D_j)}{\Gamma(1+\sigma)}. \end{equation} Substituting in the power series expansion of $\Gamma(1+z)$, we obtain the more explicit \begin{equation} \label{eq:2Gamma_X} \wh{\Gamma}_X = \exp\left(\sum_{k \ge 2} (-1)^k \cdot \frac{\zeta(k)}{k} \cdot \left(\sum_{j \in V} D_j^k - \sigma^k\right)\right). \end{equation} \subsection{The identity $\wh{G}_X = \wh{\Gamma}_X$ as formal power series} \label{subsec:G_Gamma} The expressions \eqref{eq:hG_X}, \eqref{eq:2Gamma_X} for $\wh{G}_X$ and $\wh{\Gamma}_X$ respectively define symmetric formal power series in the variables $D_j$, $j\in V$. Theorem \ref{thm:gamma} follows from the following stronger statement: \begin{prop} \label{pro:G_Gamma}We have $\wh{G}_X = \wh{\Gamma}_X$ as formal power series in $\{D_j:j\in V\}$. \end{prop} In the rest of this Section \ref{subsec:G_Gamma}, we prove Proposition \ref{pro:G_Gamma}. By \eqref{eq:hG_X}, we have \begin{align*} \wh{G}_X= 1 + \sum_{q,J,\ell,\vec{m}} (-D_q) \prod_{j\in J}(-D_j) \int_{[0,\infty)^J} \frac{(-\sigma)^{\ell-1} \prod_{j\in J}(-s_j D_j)^{m_j}}{\ell! \prod_{j\in J} m_j!} \sum_{K\subset J} (-1)^{|K|} \log^\ell\left(1+\sum_{k\in K} e^{-s_k}\right) ds^J \end{align*} where, as before, the sum is taken over all $\ell \ge 1$, $q\in V$, all nonempty subsets $J\subset V$ with $q\notin J$, and all vectors $\vec{m} \colon J \to \Z_{\ge 0}$, and we write $ds^J = \prod_{j\in J} ds_j$. We now regard $D_j$ as positive real numbers and introduce the following function $G_X(D)$ of $D=(D_j:j\in V) \in (\R_+)^V$: \[ G_X(D) := 1+\sum_{q,J} (-D_q) \prod_{j\in J}(-D_j) \int_{[0,\infty)^J} e^{-\sum_{j\in J} D_j s_j} \sum_{K\subset J} (-1)^{|K|} \frac{\left(1+\sum_{k\in K} e^{-s_k}\right)^{-\sigma}-1}{-\sigma} ds^J \] where the sum is over all $q\in V$ and all nonempty subsets $J \subset V$ not containing $q$, and $\sigma = \sum_{j\in V} D_j$. The convergence of the integral is ensured by the exponentially decaying factor $e^{-\sum_{j\in V} D_j s_j}$. It is straightforward to compute that, if the Taylor expansion of the integrand could be exchanged with the integral in the definition of $G_X(D)$, the result would be the formal power series $\wh{G}_X$. In fact we prove in Lemma \ref{lem:asymp} below that, for a fixed $D=(D_j:j\in V) \in (\R_+)^V$, we have the asymptotic expansion \begin{equation} \label{eq:asymptotic_expansion_G_X} G_X(y D) \sim \left. \wh{G}_X\right |_{D_j \rightarrow y D_j} \qquad \text{as $y \to +0$} \end{equation} where $\wh{G}_X|_{D_j \to y D_j}$ means the substitution of $yD_j$ for $D_j$ in the formal power series $\wh{G}_X$. Similarly, we have the asymptotic expansion\footnote{actually the Taylor expansion} \[ \Gamma_X(y D) \sim \left. \wh{\Gamma}_X\right |_{D_j\to y D_j} \qquad \text{as $y\to +0$} \] where $\Gamma_X(D)$ is given by \[ \Gamma_X(D) := \frac{\prod_{j\in V}\Gamma(1+D_j)}{\Gamma(1+\sigma)} \qquad \text{with $\sigma=\sum_{j\in V} D_j$}. \] Therefore it suffices to show that $G_X(D) = \Gamma_X(D)$ as functions of $D$. Note that we can interchange the integral sign with the sum over $K$ in the definition of $G_X(D)$ because of the factor $e^{-\sum_{j\in J} D_j s_j}$ (this interchange was not possible for $\wh{G}_X$). Thus: \begin{align*} G_X(D) &=1+\sum_{q,J} (-D_q) \prod_{j\in J}(-D_j) \int_{[0,\infty)^J} e^{-\sum_{j\in J} D_j s_j} \sum_{K\subset J} (-1)^{|K|} \frac{\left(1+\sum_{k\in K} e^{-s_k}\right)^{-\sigma}}{-\sigma} ds^J \\ &= 1+ \sum_{K\subset J, J\neq \emptyset, q\notin J} (-1)^{|J\setminus K|}\frac{D_q\prod_{k\in K} D_k}{\sigma} \int_{[0,\infty)^K} e^{-\sum_{k\in K} D_k s_k}\left(1+\sum_{k\in K} e^{-s_k}\right)^{-\sigma} ds^K. \end{align*} In the first line we used the fact that $\sum_{K\subset J}(-1)^K = 0$, and in the second line we interchanged the integration and summation, and then integrated $s_j$ out for $j\in J\setminus K$. Fixing an element $q\in V$ and a subset $K\subset V$ not containing $q$, we sum over subsets $J$ containing $K$ but not $q$. Using the fact that \[ \sum_{K\subset J\subset V\setminus \{q\}, J\neq \emptyset} (-1)^{|J\setminus K|} = \begin{cases} -1 & \text{if $K=\emptyset$;}\\ 0 & \text{if $|K|\le |V|-2$;} \\ 1 &\text{if $|K|=|V|-1$,} \end{cases} \] we obtain \begin{equation} \label{eq:GXD} G_X(D) =\frac{\prod_{j\in V} D_j}{\sigma} \sum_{K\sqcup \{q\} = V} \int_{[0,\infty)^K}e^{-\sum_{k\in K} D_ks_k}\left(1+\sum_{k\in K}e^{-s_k}\right)^{-\sigma} ds^K, \end{equation} where the case $K=\emptyset$ cancels the leading term $1$ and only the case $|K|=|V|-1$ remains. In order to compute the sum of integrals in Equation \eqref{eq:GXD}, we interpret the domains of integration as subsets of the projective space over the tropical numbers: concretely, we define the tropical projective space to be the quotient \[\T\mathbb{P}^{|V|-1} :=\left((\R_{\ge 0})^V\setminus \{0\}\right)/\R_+,\] where $\R_+$ acts on $(\R_{\ge 0})^V$ diagonally by scalar multiplication. We write $[u_j:j\in V]$ for the homogeneous coordinates on $\T \mathbb{P}^{|V|-1}$. This projective space is equipped with a natural volume form, which is given by the expression \begin{equation} \label{eq:volume_form_TP} d\operatorname{vol} = \prod_{j\in V\setminus \{q\}} d\log \frac{u_j}{u_q} \end{equation} for each choice of `inhomogeneous coordinates' which identify the complement of the hypersurface $\{u_q = 0\}$ with tropical affine space via the map \begin{align*} \T\mathbb{P}^{|V|-1} \setminus \{u_q= 0\}& \overset{\cong}{\longrightarrow} (\R_{\ge 0})^{|V|-1}, \\ [u_j:j\in V] & \longmapsto (t_j = u_j/u_{q}:j\in V\setminus \{q\}). \end{align*} The key point is that the equality $d \log t = - d \log 1/t $ implies that the right-hand sides of Equation \eqref{eq:volume_form_TP} for two different affine charts agree on the overlap, yielding a volume form on $\T\mathbb{P}^{|V|-1} $. \begin{lem} With respect to the volume form in Equation \eqref{eq:volume_form_TP}, we have: \[ \sum_{K\sqcup \{q\} = V} \int_{[0,\infty)^K}e^{-\sum_{k\in K} D_ks_k}\left(1+\sum_{k\in K}e^{-s_k}\right)^{-\sigma} ds^K = \int_{\T\mathbb{P}^{|V|-1}} \frac{\prod_{j\in V} u_j^{D_j}}{(\sum_{j\in V} u_j)^\sigma} d\operatorname{vol}. \] \end{lem} \begin{proof} We begin by noting that $\prod_{j\in V} u_j^{D_j}/(\sum_{j\in V} u_j)^\sigma$ is a well-defined function on $\T\mathbb{P}^{|V|-1}$ because the numerator and denominator are homogeneous functions of equal degree, and the denominator is non-vanishing. Consider the subdivision $\T\mathbb{P}^{|V|-1} = \bigcup_{q\in V} R_q$ with \[ R_q = \{[u_j:j\in V]\in \T\mathbb{P}^{|V|-1} : u_q=\max(u_j : j\in V)\}. \] Then we have: \begin{align*} \int_{\T\mathbb{P}^{|V|-1}} \frac{\prod_{j\in V} u_j^{D_j}}{(\sum_{j\in V} u_j)^\sigma} d\operatorname{vol} & = \sum_{q\in V}\int_{R_q} \frac{\prod_{j\in V} u_j^{D_j}}{(\sum_{j\in V} u_j)^\sigma} d\operatorname{vol} \\ &= \sum_{q\sqcup K = V} \int_{[0,1]^K} \frac{\prod_{k\in K} t_k^{D_k}}{(1+\sum_{k\in K} t_k)^\sigma} \prod_{k\in K}\frac{dt_k}{t_k} \end{align*} where, in the second line, we used the inhomogeneous coordinates $(t_k:k\in K)$ given by $t_k = u_k/u_q$. The conclusion follows by the change of variables $t_k = e^{-s_k}$. \end{proof} We apply the above lemma to \eqref{eq:GXD}. We rewrite the integral over $\T\mathbb{P}^{|V|-1}$ as an integral over the simplex $\nabla = \{\sum_{j\in V} u_j =1\}$, which is a slice of the diagonal action on $\R^V_{\geq 0} \setminus \{0\}$. Writing $\{D_0,\dots,D_m\}$ for $\{D_j : j\in V\}$, we find that the restriction of $d\operatorname{vol}$ to $\nabla$ is \[\frac{du_1 \ldots du_m}{u_0 u_1 \ldots u_m}.\] Thus we have \begin{align*} G_X(D) &= \frac{\prod_{j=0}^{m} D_j}{\sigma} \int_{\nabla} \prod_{j=0}^m u_j^{D_j-1} du_1\ldots du_m\\ & = \frac{\prod_{j=0}^{m} D_j}{\sigma} \cdot \frac{\Gamma(D_0) \cdots \Gamma(D_m)}{\Gamma(\sigma)} \\ & = \Gamma_X(D), \end{align*} using a well-known integral due to Dirichlet \cite{Dirichlet1839},\footnote{This integral generalizes the Euler integral of the first kind which defines the Beta function (from which it can be proved by induction), and expresses the fact that the Dirichlet distribution is normalized.} together with the identity $z\Gamma(z) = \Gamma(1+z)$ and the formula \eqref{eq:1Gamma_X}. This essentially completes the proof of Proposition \ref{pro:G_Gamma} and hence of Theorem \ref{thm:gamma}, with the only missing step being the computation of the asymptotic expansion of $G_X$, which we now provide: \begin{lem} \label{lem:asymp} For a fixed $D \in (\R_+)^V$, we have the asymptotic expansion \[ G_X(y D) \sim \left. \wh{G}_X\right |_{D_j \rightarrow y D_j} \qquad \text{as $y \to +0$}. \] \end{lem} \begin{proof} We fix $D\in (\R_+)^V$ throughout the proof. For a nonempty subset $J\subset V$, we set \begin{align*} g_J(X,y) & := h_J(X,y) \cdot \prod_{j\in J} X_j^{y D_j}, \\ h_J(X,y) & := \frac{\sum_{K\subset J} (-1)^{|K|} \left( \left(1+\sum_{j\in K} X_j\right)^{-y \sigma} -1\right)}{(-y \sigma) \cdot \prod_{j\in J} X_j} \end{align*} where $X_j$ ($j\in V$) and $y$ are variables in $(0,1]$. Via the change of variables $X_j = e^{-s_j}$, we have \[ G_X(yD) = 1+ \sum_{q,J}(-yD_q) \prod_{j\in J}(-yD_j) \int_{[0,1]^J} g_J(X,y) \prod_{j\in J} dX_j \] where the summation range is the same as before. Note that the sum over $q,J$ is finite. It suffices to show that we can exchange the Taylor expansion of $g_J(X,y)$ in $y$ with the integral over $[0,1]^J$ to get the asymptotic expansion. For this we use Taylor's theorem: \[ g_J(X,y) = \sum_{a=0}^{m-1}\frac{1}{a!} (\partial_y^a g_J)(X,0) \cdot y^a + \frac{1}{m!} (\partial_y^m g_J)(X,\xi(X)) \cdot y^m, \qquad \exists \xi(X)\in [0,y]. \] Note that each term $(\partial_y^a g_J)(X,0)$ is a linear combination of products of the integrands defining the local integrals $I_{\ell;\vec{m}}$, and hence is integrable on $[0,1]^J$ for the same reason that the local integrals are, namely the exponential decay arising from the bound \eqref{eqn:boundint}. It remains to show that $|(\partial_y^m g_J)(X,y)|$ is bounded by an integrable function of $X$ (on $[0,1]^J$) which is independent of $y\in [0,1]$. As in the proof of the bound \eqref{eqn:boundint}, we can see that $h_J(X,y)$ extends to a smooth (even analytic) function in a neighbourhood of $[0,1]^J \times [0,1]$; the numerator in the definition of $h_J$ vanishes along $X_j=0$ for each $j\in J$ and thus $h_J$ does not have poles along $X_j=0$. Thus there exist smooth functions $f_a(X,y)$ in a neighbourhood of $[0,1]^J\times [0,1]$ such that \[ \partial_y^m g_J(X,y) = \sum_{a=0}^m f_a(X,y) \cdot \left(\textstyle\sum_{j\in J} D_j \log X_j\right)^a \prod_{j\in J}X_j^{y D_j} \] and therefore we find a constant $C>0$ with \[ \left | \partial_y^m g_J(X,y) \right |\le C \sum_{a=0}^m \left|\textstyle\sum_{j\in J} D_j \log X_j\right|^a \] for $(X,y) \in (0,1]^J \times (0,1]$. The right-hand side is integrable on $[0,1]^J$ with respect to $X$, and the lemma follows. \end{proof} \subsection{Examples of the local integrals} Recall that $\wh{\Gamma}_X$ is expanded in the $\zeta$-values $\zeta(k)$ with $k\ge 2$ (see \eqref{eq:2Gamma_X}). The identity $\wh{G}_X = \wh{\Gamma}_X$ determines some of the local integrals $I_{\ell;\vec{m}}$ \eqref{eqn:Ilm} in terms of $\zeta(k)$. In general, however, the identity only shows that certain polynomial expressions in the local integrals equal $\zeta(k)$; it seems that individual local integrals cannot necessarily be written as polynomials in $\zeta(k)$. A \emph{local integral of weight $k$} is a real number belonging to the set \[ \{I_{\ell;\vec{m}} : \ell + |\vec{m}| + \dim(\vec{m}) = k, \ \ell\ge 1, \ \dim(\vec{m})\ge 1\} \] where we set $\dim(\vec{m})=p$ and $|\vec{m}|=\sum_{i=1}^p m_i$ for $\vec{m} =(m_1,\dots,m_p) \in (\Z_{\ge 0})^p$. We can easily see that there are $\pi(k-1) + \pi(k-2) + \cdots +\pi(1)$ many local integrals of weight $k$, where $\pi(j)$ denotes the number of partitions of $j\in\N$. On the other hand, we obtain $\pi(k)-1$ relations\footnote{Note that $\pi(k)$ is the dimension of the space of symmetric functions of degree $k$. We have one fewer relations since the coefficients in front of $D_j^k$ of the degree-$k$ parts of both $\log \wh{G}_X$ and $\log \wh{\Gamma}_X$ vanish.} in weight $k$ from the identity $\wh{G}_X = \wh{\Gamma}_X$. Therefore, as $k$ grows, the number of local integrals becomes far greater than the number of relations among them. \begin{table}[h] \begin{center} \begin{tabular}{r|cccccccccccc} weight $k$ & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 \\ \hline $\#$ of local integrals & 1 & 3 & 6 & 11 & 18 & 29 & 44 & 66 & 96 & 138 & 194 & 271 \\ $\#$ of relations & 1 & 2 & 4 & 6 & 10 & 14 & 21 & 29 & 41 & 55 & 76 & 100 \end{tabular} \end{center} \end{table} The local integral $I_{1;m}$ of weight $m+2$ can be computed explicitly. \begin{align} \label{eq:I1m} \begin{split} I_{1;m} & = -\int_0^\infty s^m \log(1+e^{-s}) ds \\ & = \sum_{n=1}^\infty\int_0^\infty s^m \frac{(-1)^n}{n} e^{-ns} ds = m! \sum_{n=1}^\infty \frac{(-1)^n}{n^{m+2}} \\ &= - m! \left(1-\frac{1}{2^{m+1}}\right) \zeta(m+2). \end{split} \end{align} \begin{rmk} Viewing $\wh{G}_X$, $\wh{\Gamma}$ as functions of $D=(D_1,\dots,D_m)$ and writing $\wh{G}_X= G_X(D_1,\dots,D_m)$, $\wh{\Gamma}_X=\Gamma_X(D_1,\dots,D_m)$, we have $G_X(D_1,\dots,D_{m-1}) = G_X(D_1,\dots,D_{m-1},0)$ and $\Gamma_X(D_1,\dots,D_{m-1}) =\Gamma_X(D_1,\dots,D_{m-1},0)$. Therefore we can regard $\wh{G}_X$, $\wh{\Gamma}_X$ as symmetric functions in infinitely many variables $(D_1,D_2,D_3,\dots)$, and we obtain the maximal number of relations by doing so. \end{rmk} \subsubsection*{In weight 2} We have only one local integral $I_{1;0} = -\frac{1}{2} \zeta(2)$. \subsubsection*{In weight 3} We have 3 local integrals $I_{2;0}$, $I_{1;0,0}$, $I_{1;1}$. The identity $\wh{G}_X = \wh{\Gamma}_X$ together with \eqref{eq:I1m} shows: \begin{align} \label{eq:local_integral_wt_3} \begin{split} I_{1;1} & = -\frac{3}{4} \zeta(3), \\ I_{2;0} & = -\frac{1}{4} \zeta(3), \\ I_{1;0,0} & = -\frac{5}{12} \zeta(3). \end{split} \end{align} \subsubsection*{In weight 4} We have 6 local integrals $I_{3;0}$, $I_{2;1}$, $I_{2;0,0}$, $I_{1;2}$, $I_{1;1,0}$, $I_{1;0,0,0}$. We obtain $I_{1;2} = -\frac{7}{4} \zeta(4)$ from \eqref{eq:I1m} and the following four relations from $\wh{G}_X = \wh{\Gamma}_X$: \begin{align} \label{eq:relation_wt_4} \begin{split} \frac{1}{2} I_{1;2} + \frac{1}{2} I_{2;1} + \frac{1}{3} I_{3;0} + \zeta(4) &= 0, \\ 4 I_{1;0}^2 + 2 I_{1;2} + \zeta(4) & = 0, \\ \frac{1}{2} I_{1;2} + \frac{3}{2} I_{2;1} - 2 I_{1;1,0} - \frac{3}{2} I_{2;0,0} + \zeta(4) & = 0, \\ 2 I_{1;2} - 8 I_{1;1,0} + 4 I_{1;0,0,0} + \zeta(4) & =0, \end{split} \end{align} where the second equation reduces to the well-known identity $\zeta(4) = \frac{2}{5} \zeta(2)^2$. \subsubsection*{Other examples} Considering the case where $\{D_j : j \in V\} = \{D_1,D_2\}$ and comparing the coefficient of $D_1^{n-1} D_2$ of the identity $\wh{G}_X= \wh{\Gamma}_X$, we get a linear relation among the local integrals: \[ (n-1)! \zeta(n) + \sum_{i=1}^{n-2} \binom{n-1}{i} I_{n-1-i;i} + 2 I_{n-1,0} = 0, \] or equivalently, \begin{equation} \label{eq:zeta_n} \zeta(n) = \frac{1}{(n-1)!} \int_{-\infty}^\infty \left(\log(1+e^s)\right)^{n-1} - (\max(0,s))^{n-1} ds. \end{equation} This generalizes the first equation of \eqref{eq:relation_wt_4}. There are other examples of the local integrals which can be expressed in terms of the $\zeta$-values: \begin{align*} I_{2;2} & = \frac{29}{8} \zeta(5) - 2 \zeta(2) \zeta(3), \\ I_{2;4} & = \frac{753}{8} \zeta(7) - 42 \zeta(3) \zeta(4) - 24 \zeta(2) \zeta(5). \end{align*} \subsection{Proof of \eqref{eqn:typeII}} We use the above results for the local integrals to prove \eqref{eqn:typeII}. \begin{prop} \label{prop:asymptotics_of_integral} Let $V$ be a bounded domain in $\R^2$ containing the origin such that $\partial V$ is affine-linear in a neighbourhood of $\operatorname{Sing}(\max(0,b_1,b_2))$ and intersects it transversally. Let $L$ be the total affine length of the intersection $V \cap \operatorname{Sing}(\max(0,b_1,b_2))$. Then we have \begin{align*} (-\log t)^3 \int_V \left(-\log_t(1+t^{-b_1}+t^{-b_2}) - \max(0,b_1,b_2)\right) db_1db_2 = \zeta(2) \cdot (-\log t) \cdot L + \zeta(3) + O\left(t^\epsilon\right) \end{align*} as $t\to +0$, for some $\epsilon>0$ depending only on $V$. \end{prop} \begin{proof} We decompose $V$ into small pieces $V_1,V_2,V_3,\dots$ and evaluate the integral locally. The integrand $-\log_t(1+t^{-b_1}+t^{-b_2}) -\max(0,b_1,b_2)$ is exponentially close to zero away from the tropical line $\operatorname{Sing}(\max(0,b_1,b_2))$ (see Figure \ref{fig:Maslov}). Hence, for any bounded domain $V_1$ with $\overline{V_1}\cap \operatorname{Sing}(0,b_1,b_2)=\emptyset$, we have \[ (-\log t)^3 \int_{V_1} (-\log_t(1+t^{-b_1}+t^{-b_2}) - \max(0,b_1,b_2)) db_1db_2 = O(t^\epsilon) \] for some $\epsilon>0$ depending on $V_1$. Consider a domain $V_2\subset \R^2$ such that $\overline{V_2}$ intersects the tropical line $\operatorname{Sing}(\max(0,b_1,b_2))$ only along the edge $\{b_1=0, b_2<0\}$. We again assume that $\partial V_2$ is affine-linear in a neighbourhood of the edge and is transverse to it. Since the contribution (to the integral) away from the edge $\{b_1=0,b_2<0\}$ is exponentially small, we may assume that $V_2$ is of the form \[ -B<b_1<B, \quad f(b_1)<b_2<g(b_1) \] for some negative affine-linear functions $f,g$. Because $t^{-b_2}$ is exponentially small in this region, we have \begin{align*} (-\log t)^3\int_{V_1} & \left(-\log_t(1+t^{-b_1} + t^{-b_2})-\max(0,b_1,b_2) \right) db_1db_2 \\ & = (-\log t)^3 \int_{-B}^B (-\log_t(1+t^{-b_1})-\max(0,b_1)) (g(b_1)-f(b_1))db_1+O\left(t^\epsilon\right) \\ & = (-\log t) \cdot L_2 \cdot \int_{ B\log t}^{-B \log t} \left( \log(1+e^x) - \max(0,x)\right) dx + O\left(t^\epsilon\right) \\ & = (-\log t) \cdot L_2 \cdot \zeta(2) + O\left(t^\epsilon\right) \end{align*} where $L_2 = g(0)-f(0)$ is the affine length of $V_2 \cap \operatorname{Sing}(\max(0,b_1,b_2))$ and we used \eqref{eq:zeta_n} in the last step. By the symmetry of the integrand under the affine transformation $(b_1,b_2)\to (-b_2,b_1-b_2)$, the same is true for regions intersecting the other edges. It now suffices to prove the statement for one particular $V$. Let $V$ be the rectangular region given by $-B<b_1<2B$, $-B<b_2<B$ with $B>0$ (see Figure \ref{fig:domain_decomp}). We decompose it into seven regions $W_1,\dots,W_7$ as shown. We have \begin{align*} (-\log t)^3 \int_{W_1} &\left(- \log_t(1+t^{-b_1}+t^{-b_2}) -\max(0,b_1,b_2)\right)db_1db_2 \\ & = \int_{[0,(-\log t)\cdot B]^2} \log(1+e^{-s_1} + e^{-s_2}) ds_1 ds_2 \\ & = \int_{[0,(-\log t)\cdot B]^2} \left(\log(1+e^{-s_1} + e^{-s_2}) - \log(1+e^{-s_1}) - \log(1+e^{-s_2})\right) ds_1 ds_2 \\ & \qquad + 2 (-\log t) \cdot B \cdot \int_{[0,(-\log t)\cdot B]} \log(1+e^{-s}) ds \\ & =I_{1;0,0} - 2 (-\log t) \cdot B \cdot I_{1;0} + O\left(t^\epsilon\right) = -\frac{5}{12} \zeta(3) + (-\log t) \cdot B \cdot \zeta(2) + O\left(t^\epsilon\right), \end{align*} where we used \eqref{eq:I1m} and \eqref{eq:local_integral_wt_3} in the last step. The integrals over the regions $W_2,W_3$ are the same by the affine symmetry. The integral over $W_4$ is given by \begin{align*} (-\log t)^3 \int_{W_4} &\left(-\log_t(1+t^{-b_1}+t^{-b_2})-\max(0,b_1,b_2) \right) db_1db_2 \\ & = (-\log t)^3 \int_{0\le b_1 \le B, B\le b_1-b_2 \le B+b_1} \log_t(1+ t^{b_1} +t^{b_2-b_1}) db_1db_2 \\ & = \int_{[0,(-\log t)\cdot B]} x \log(1+e^{-x}) dx + O\left(t^\epsilon\right) \qquad \text{($b_2$ integrated out)} \\ & = -I_{1;1} + O\left(t^\epsilon\right) = \frac{3}{4} \zeta(3) + O\left(t^\epsilon\right), \end{align*} where we used \eqref{eq:I1m} in the last step. The integrals over $W_5,W_6$ are the same by the affine symmetry. The integral over $W_7$ is of order $O\left(t^\epsilon\right)$. The conclusion follows by summing up these contributions. \end{proof} \begin{figure} \centering \begin{tikzpicture}[x=1pt,y=1pt] \draw (110,10)--(110,50)--(110,90)--(230,90)--(230,10)--(110,10); \draw (110,10)--(110,50)--(150,90)--(190,90)--(190,50)--(150,10)--(110,10); \draw (230,90)--(190,50)--(190,10); \draw (130,30) node {$W_1$}; \draw (150,70) node {$W_2$}; \draw (170,50) node {$W_3$}; \draw (120,78) node {$W_5$}; \draw (178,20) node {$W_4$}; \draw (205,78) node {$W_6$}; \draw (210,40) node {$W_7$}; \draw[very thick] (150,50)--(200,100); \draw[very thick] (150,50)--(100,50); \draw[very thick] (150,50)--(150,0); \end{tikzpicture} \caption{Decomposition of the domain $V$} \label{fig:domain_decomp} \end{figure} \begin{rmk} If we only assume that $\partial V$ is \emph{smooth} (instead of linear) in a neighbourhood of the tropical line $\operatorname{Sing}(\max(0,b_1,b_2))$ and intersects it transversally, we get the same result except that the error term $O\left(t^\epsilon\right)$ in the right-hand side must be replaced with $O((-\log t)^{-1})$. This is because the 2-jet of $g-f$ (in the above proof) contributes to the term of order $(-\log t)^{-1}$. More precisely, we are able to show the identity of distributions: \[ -\log_t(1+t^{-b_1}+t^{-b_2}) = \max(0,b_1,b_2) - \frac{\zeta(2)}{ (-\log t)^{2}} \delta_{\operatorname{Sing}(\max(0,b_1,b_2))} -\frac{\zeta(3)}{(-\log t)^{3}} \delta_{(0,0)} + O\left(\frac{1}{(-\log t)^{4}}\right) \] where $\delta_A$ means the delta measure supported on $A$ (with $A$ equipped with the affine measure). We plan to explore such distributions on tropical spaces in a future paper. \end{rmk} \section{Periods of cycles that are mirror to line bundles} \label{sec:phase} \subsection{Varying the phases} Let $t>0$ be a small positive number and let $\theta = (\theta_q)_{q\in V}$ be a collection of real numbers. Recall that we define \[ f_{t,\theta}(z) := \sum_{q\in V} e^{\mathtt{i} \theta_q} t^{\lambda_q} z^q \] and \[ \mathring{Z}_{t,\theta} := \{f_{t,\theta}(z) = 1\} \subset P_{\C^*}. \] Let $C_{t,\theta}\subset \mathring{Z}_{t,\theta}$ be the parallel transport of the positive real cycle $C_t^+ \subset \mathring{Z}_t$ as we vary $\theta$ continuously from $\theta=0$. Let $\Omega_{t,\theta}$ be the holomorphic volume form on $\mathring{Z}_{t,\theta}$ defined by \[ \Omega_{t,\theta} = \left. \frac{d\log z_0 \wedge d\log z_1 \wedge \cdots \wedge d\log z_n}{df_{t,\theta}}\right|_{\mathring{Z}_{t,\theta}} \] where $(z_0,\dots,z_n)$ are coordinates on $P_{\C^*} \cong (\C^*)^{n+1}$. In this section, we prove the following generalization of Theorem \ref{thm:periodR}: \begin{thm} \label{thm:cycle_with_phase} We have \[ \int_{C_{t,\theta}\subset Z_{t,\theta}} \Omega_{t,\theta} = \int_X t^{-\omega_\lambda} \cdot e^{-\sum_{q\in V} \mathtt{i} \theta_q D_q} \cdot \wh{G}_X + O\left(t^\epsilon\right) \qquad \text{as $t\to +0$, for some $\epsilon>0$.} \] \end{thm} We can formally obtain this from Theorem \ref{thm:periodR} by substituting $\lambda_q+\mathtt{i}(\theta_q/\log t)$ for $\lambda_q$. Note however that this does not logically follow from Theorem \ref{thm:periodR} since in general `analytic continuation' does not commute with `asymptotic expansion'. We will justify this substitution via a tropical calculation similar to Section \ref{sec:periodR}. When $\theta_q = 2\pi \nu_q$, we have $C_t^{(\nu)} = C_{t,\theta}$. Therefore, Theorem \ref{thm:Bat_Gamma_lb} follows by combining Theorems \ref{thm:cycle_with_phase} and \ref{thm:gamma}. \subsection{A tropical construction of the cycle $C_{t,\theta}$} \label{sec:construction_Cttheta} We first construct a cycle approximately contained in $\mathring{Z}_{t,\theta}$ by ``sliding'' the positive real cycle $B_t=\operatorname{Log}_t(C_t^+)$ in the purely imaginary direction. Then we modify it to an actual cycle $C_{t,\theta}$ in $\mathring{Z}_{t,\theta}$. A similar construction for ``semi-tropical'' cycles appears in the work of Abouzaid \cite{Abouzaid:homogeneous, Abouzaid:HMS_toric}, Fang--Liu--Treumann--Zaslow \cite{FLTZ:T-dual}, and Hanlon \cite{Hanlon:monodromy}. Let $\kappa>0$ be a positive real number and let $N_\kappa(\Delta_\lambda)\subset P_\R$ denote the $\kappa$-neighbourhood of $\Delta_\lambda$ \[ N_\kappa(\Delta_\lambda) = \{ p\in P_\R: \beta_q(p) \ge -\kappa, \ \forall q\in V\}. \] We choose $\kappa>0$ sufficiently small so that the following holds: \begin{itemize} \item[(i)] for every subset $K\subset V$, if the regions $N_\kappa(\Delta_\lambda) \cap\{\beta_q \le \kappa\}$ with $q\in K$ intersect, then the facets $\Delta_\lambda\cap \{\beta_q=0\}$ with $q\in K$ intersect; in this case $K$ is linearly independent as a subset of $Q_\R$. \item[(ii)] for every $p\in P_\R$, $K=\{k \in V: \beta_k(p) \le \min_{q\in V}(\beta_q(p))+\kappa\}$ is linearly independent as a subset of $Q_\R$. \end{itemize} We choose a smooth function $\phi\colon N_\kappa(\Delta_\lambda) \to P_\R$ such that the following holds for every $q\in V$: \begin{equation} \label{eq:bdry_cond_phi} \langle q,\phi(p)\rangle = - \theta_q \qquad \text{whenever} \quad \beta_q(p)\le \kappa. \end{equation} Such a function exists because of the above condition (i). Define a map $\Phi_t\colon N_\kappa(\Delta_\lambda) \to P_\C$ by \[ \Phi_t(p) = p + \mathtt{i} \frac{\phi(p)}{\log t}. \] We naturally extend the functions $\beta_q \colon P_\R \to \R$, $i_t \colon P_\R \to P_{\C^*}$ to the functions $\beta_q\colon P_\C\to \C$, $i_t \colon P_\C \to P_{\C^*}$ by the formulae: \[ \beta_q(p) = \langle q, p\rangle + \lambda_q, \qquad i_t(p_0,p_1,\dots,p_n) = (t^{p_0},t^{p_1},\dots,t^{p_n}). \] The following is immediate from the definition. \begin{lem} \label{lem:approx_cycle} For $p \in N_\kappa(\Delta_\lambda)$, we have that $|f_{t,\theta}(i_t(\Phi_t(p))) - f_t(i_t(p))| \le 2|V| t^\kappa$. \end{lem} \begin{proof} We have \begin{align*} f_t(i_t(p)) & = \sum_{q\in V} t^{\beta_q(p)} \\ f_{t,\theta}(i_t(\Phi_t(p))) &= \sum_{q\in V} e^{\mathtt{i}(\theta_q+\langle q,\phi(p)\rangle)} t^{\beta_q(p)}. \end{align*} By the condition \eqref{eq:bdry_cond_phi}, the summands with $\beta_q(p) \le \kappa$ coincide. Since the summands with $\beta_q(p)>\kappa$ have norm bounded by $t^\kappa$, the conclusion follows. \end{proof} The lemma implies that the cycle $i_t(\Phi_t(B_t))$ is approximately contained in the hypersurface $\mathring{Z}_{t,\theta}$. We shall make it an actual cycle in $\mathring{Z}_{t,\theta}$ by a $C^1$-small perturbation. In the following, we will fix a Hermitian norm on $P_\C\cong \C^{n+1}$. \begin{prop} \label{prop:delta} Set $R_t=\{p\in P_\R: \frac{1}{2} \le f_t(i_t(p))\le \frac{3}{2}\}$. For sufficiently small $t>0$, there exists a smooth function $\delta_t \colon R_t \to P_\C$ such that \begin{itemize} \item[(a)] we have $f_{t,\theta}(i_t(\Phi_t(p)+\delta_t(p)))= f_t(i_t(p))$ for $p\in R_t$, and \item[(b)] $\|\delta_t\|_{C^1}=O(t^\kappa)$, where $\|\cdot\|_{C^1}$ denotes the $C^1$-norm over the region $R_t$. \end{itemize} \end{prop} Using the function $\delta_t$ in this proposition, we define \[ C_{t,\theta} := i_t(\widetilde{\Phi}_t(B_t)) \] where $\widetilde{\Phi}_t\colon R_t \to P_\C$ is defined by $\widetilde{\Phi}_t(p) = \Phi_t(p) + \delta_t(p)$. This is a cycle contained in $\mathring{Z}_{t,\theta}$ and is homeomorphic to a sphere. \begin{example} We give an example of the phase-shifting function $\phi\colon N_\kappa(\Delta_\lambda) \to P_\R$ in the case of $Y_{\Delta_\lambda} = \C\mathbb{P}^2$. In Figure \ref{fig:phase}, we present $\phi$ as a vector field near the boundary of the moment polytope $\Delta_\lambda$; the boundary condition \eqref{eq:bdry_cond_phi} says that $p+\phi(p)$ with $p\in \partial \Delta_\lambda$ lies on the boundary of a shifted polytope (dotted line). Following the method in \cite{FLTZ:T-dual}, we can give such $\phi$ via a piecewise linear function $f$ on the fan (drawn in thin lines) dual to the polytope $\Delta_\lambda$: $\phi$ can be given as the gradient of a smoothing of the piecewise-linear function $f$ that takes the values $\theta_1$, $\theta_2$, $\theta_3$, respectively, at the primitive generators $(-1,0)$, $(0,-1)$, $(1,1)$ of the 1-dimensional cones. \begin{figure}[h] \centering \begin{tikzpicture} \draw[very thick] (-1,-1) -- (2,-1) -- (-1,2) -- (-1,-1); \draw[very thin] (0,0)--(2,2); \draw[very thin] (-3,0) -- (0,0) -- (0,-2); \draw[dotted] (-1.5,-0.5) -- (1,-0.5) -- (-1.5,2)--(-1.5,-0.5); \draw[thin, ->] (-1,-0.4) -- (-1.5,0.1); \draw[thin, ->] (-1,-0.7) -- (-1.5, -0.2); \draw[thin, ->] (-1,-1) -- (-1.5,-0.5); \draw[thin, ->] (-0.6,-1) -- (-1.1,-0.5); \draw[thin, ->] (-0.2,-1) -- (-0.7,-0.5); \draw[thin, ->] (1,-1) -- (0,-0.5); \draw[thin,->] (1.5,-1) -- (0.5,-0.5); \draw[thin,->] (2,-1) -- (1,-0.5); \draw[thin, ->] (1.6,-0.6) -- (0.6,-0.1); \draw[thin, ->] (1.2,-0.2) -- (0.2,0.3); \draw[thin, ->] (0.2,0.8) -- (-0.3,0.8); \draw[thin, ->] (-0.1,1.1) -- (-0.6,1.1); \draw[thin, ->] (-0.4,1.4) -- (-0.9,1.4); \draw[thin,->] (-0.7,1.7) -- (-1.2,1.7); \draw[thin, ->] (-1,2) -- (-1.5,2); \draw[thin,->] (-1,1.65) -- (-1.5,1.65); \draw[thin, ->] (-1,1.3) -- (-1.5,1.3); \draw[thin, ->] (-1,0.9) -- (-1.5,0.9); \draw[thin, ->] (-1,0.5) -- (-1.5,0.5); \draw (-1.5,-1.5) node {\scriptsize $\phi = (-\theta_1,-\theta_2)$}; \draw (3,-0.2) node {\scriptsize $\phi = (\theta_2+\theta_3, -\theta_2)$}; \draw (-3,1) node {\scriptsize $\phi = (-\theta_1, \theta_1+\theta_3)$}; \end{tikzpicture} \caption{A phase-shifting function $\phi$ as a vector field ($Y_{\Delta_\lambda}= \C\mathbb{P}^2$).} \label{fig:phase} \end{figure} \end{example} We end this subsection with the proof of Proposition \ref{prop:delta}. It is elementary and based on a standard method in tropical geometry. The uninterested reader can safely skip the proof since the details will not be used later. \begin{proof}[Proof of Proposition \ref{prop:delta}] The region $R_t$ is contained in $N_\kappa(\Delta_\lambda)$ for sufficiently small $t>0$ because $f_t(i_t(p))\ge t^{-\kappa}$ for $p\in P_\R \setminus N_\kappa(\Delta_\lambda)$. In particular, $\Phi_t(p)$ is defined for $p\in R_t$. Define a holomorphic function $g_t\colon P_\C \to \C$ by $g_t(p) = f_{t,\theta}(i_t(p)) = \sum_{q\in V} e^{\mathtt{i}\theta_q} t^{\beta_q(p)}$. By Lemma \ref{lem:approx_cycle}, $\Phi_t(p)$ with $p\in R_t$ is ``close'' to the fibre of the map $g_t$ at $f_t(i_t(p))$. We will flow $\Phi_t(p)$ to a nearby point $\Phi_t(p)+\delta_t(p)$ lying in the fibre $g_t^{-1}(f_t(i_t(p)))$ by the gradient vector field of $g_t$. Let \[ \grad g_t = \left(\overline{\parfrac{g_t}{a_0}},\dots, \overline{\parfrac{g_t}{a_n}}\right) \] be the gradient vector field of $g_t$ on $P_\C$ and define $\xi \colon R_t\to \C$ by \[ \xi_t(p) = f_t(i_t(p)) - g_t(\Phi_t(p)). \] Lemma \ref{lem:approx_cycle} gives $|\xi_t(p)| \le 2 |V|t^\kappa$. Consider the differential equation for an unknown function $c(p;s) \colon R_t\times [0,1] \to P_\C$ \[ \frac{d}{ds} c(p;s) = \xi_t(p) \cdot \frac{\grad g_t}{|\grad g_t|^2}(c(p;s)) \] with the initial condition $c(p;0) = \Phi_t(p)$. Suppose we have a global solution $c(p;s)$. Then the differential equation shows that $g_t(c(p;s))=g_t(\Phi_t(p))+s \xi_t(p)$ and therefore $c(p;1)$ lies in the fiber $g_t^{-1}(f_t(i_t(p)))$. Setting $c(p;1) = \Phi_t(p) + \delta_t(p)$, we have part (a) of the proposition. We will show that a solution exists and that the $C^1$-norm of $\delta_t$ is of order $t^\kappa$. For $p\in R_t$, let $s_p$ be the supremum of $\overline{s}\in [0,1]$ such that the flow $s\mapsto c(p;s)$ exists on the interval $[0,\overline{s})$. As discussed, $g_t(c(p;s))$ with $s\in [0,s_p)$ lies in the interval connecting $g_t(\Phi_t(p))$ and $f_t(i_t(p))$. Since $|f_t(i_t(p))-g_t(\Phi_t(p))| \le 2|V|t^\kappa$ by Lemma \ref{lem:approx_cycle} and $\frac{1}{2}\le f_t(i_t(p))\le \frac{3}{2}$, we may assume that $|g_t(c(p;s))|\ge \frac{1}{3}$ for sufficiently small $t>0$. Thus by Lemma \ref{lem:grad} below, we have \begin{equation} \label{eq:grad_g_estimate} |(\grad g_t)(c(p;s))| \ge (-\log t) \rho_0>0 \end{equation} for some constant $\rho_0>0$ for a sufficiently small $t>0$. Therefore, the differential equation implies \[ \left| \frac{d}{ds} c(p;s) \right| = \frac{|\xi_t(p)|}{|(\grad g_t)(c(p;s))|} \le \frac{|\xi_t(p)|}{(-\log t) \rho_0} \le 2 \rho_0^{-1} |V| (-\log t)^{-1} t^\kappa \] for $s\in [0,s_p)$. This implies that the limit $\lim_{s\to s_p-0}c(p;s)$ exists. If $s_p<1$, the solution can be extended to a larger interval, contradicting the assumption. Therefore we must have $s_p=1$ and the solution exists on the interval $[0,1]$. By integrating the above estimate, we get the bound (for $s\in [0,1]$): \begin{equation} \label{eq:estimate_cas} |c(p;s)-c(p;0)| \le \rho_1 (-\log t)^{-1}t^\kappa \end{equation} with $\rho_1 = 2 \rho_0^{-1} |V|$. This gives a $C^0$-bound for $\delta_t(p) =c(p;1)-c(p;0)$. To obtain a $C^1$-bound for $\delta_t$, we use the differential equation \begin{equation} \label{eq:linear_DE} \frac{d}{ds} \parfrac{c(p;s)}{p_i} = \parfrac{\xi_t}{p_i}(p) \frac{\grad g_t}{|\grad g_t|^2}(c(p;s)) + \xi_t(p) \left( F(c(p;s)) \parfrac{c(p;s)}{p_i} + G(c(p;s)) \parfrac{\overline{c(p;s)}}{p_i}\right) \end{equation} where $F(c)$, $G(c)$ are square matrices of size $n+1$ (viewed as endomorphisms of $P_\C$) whose $(j,k)$-entries are given by \begin{align*} F(c)_{jk} & = -\frac{1}{|\grad g_t|^{4}} \overline{\parfrac{g_t}{c_j}} \sum_{l=0}^n \overline{\parfrac{g_t}{c_l}} \parfrac{^2g_t}{c_l\partial c_k} \\ G(c)_{jk} & = \frac{1}{|\grad g_t|^2} \overline{\parfrac{^2g_t}{c_j \partial c_k}} - \frac{1}{|\grad g_t|^4} \overline{\parfrac{g_t}{c_j}} \sum_{l=0}^n \parfrac{g_t}{c_l}\overline{\parfrac{^2 g_t}{c_l \partial c_k}}. \end{align*} Given the function $c(p;s)$, \eqref{eq:linear_DE} can be viewed as a linear differential equation for $x(s)=\parfrac{c(p;s)}{a_i}$ as appears in Lemma \ref{lem:Gronwall} below. In view of Lemma \ref{lem:Gronwall}, it suffices to establish the following inequalities for sufficiently small $t>0$: \begin{align} \label{eq:estimate_coeff} \begin{split} \max\left( \left|\parfrac{\xi_t}{p_i}(p) \frac{\grad g_t}{|\grad g_t|^2}(c(p;s)) \right|, |\xi_t(p) F(c(p;s))_{jk}|, |\xi_t(p) G(c(p;s))_{jk}| \right) & \le \rho_2 t^\kappa \\ \left|\parfrac{c(p;0)}{p_i}\right| & \le \rho_2 \end{split} \end{align} where $\rho_2>0$ is a constant independent of $p,s,t$. The second inequality follows from the initial condition $c(p;0) = \Phi_t(p) = p + \mathtt{i} \phi(p)/\log t$. Similarly to the proof of Lemma \ref{lem:approx_cycle}, we have \[ \left|\parfrac{\xi_t}{p_i}(p) \right| = (-\log t) \left| \sum_{q\in V, \beta_q(p)>\kappa} q_i (1 - e^{\mathtt{i} (\theta_q + \langle q, \phi(p)\rangle)}) t^{\beta_q(p)} \right| \le \rho_3 (-\log t) t^\kappa \] for some constant $\rho_3>0$. Therefore, using the estimate \eqref{eq:grad_g_estimate}, we get \begin{equation} \label{eq:estimate_derxi_grad} \left|\parfrac{\xi_t}{p_i}(p) \frac{\grad g_t}{|\grad g_t|^2}(c(p;s)) \right| \le \rho_0^{-1} \rho_3 t^\kappa. \end{equation} Since $p\in R_t$, we have $f_t(i_t(p))\le \frac{3}{2}$ and hence $\beta_q(p) \ge \log_t (3/2)$. This together with the estimate \eqref{eq:estimate_cas} gives \begin{align*} \Re (\beta_q(c(p;s))) & \ge \Re(\beta_q(c(p;0))) - |q| \rho_1 (-\log t)^{-1} t^\kappa \\ & \ge \log_t(3/2) - \rho_4 (-\log t)^{-1} t^\kappa. \end{align*} with $\rho_4 := \rho_1 \max_{q\in V}(|q|) $, where we used $\Re(\beta_q(c(p;0)))= \Re(\beta_q(\Phi(p)))=\beta_q(p)$. Therefore we have \begin{align*} \left| \parfrac{g_t}{c_j}(c(p;s)) \right | & = (-\log t) \left| \sum_{q\in V} e^{\mathtt{i}\theta_q} q_j t^{\beta_q(c(p;s))} \right| \\ & \le (-\log t) \sum_{q\in V} |q_j| t^{\Re(\beta_q(c(p;s)))} \\ & \le (-\log t) \rho_5 \cdot t^{\log_t(3/2) - \rho_4 (-\log t)^{-1} t^\kappa} \le (-\log t) \rho_6 \end{align*} for some constant $\rho_5>0$ and $\rho_6 = \frac{3}{2} \rho_5 \exp(\rho_4)$. Similarly we have \[ \left| \parfrac{^2 g_t}{c_j \partial c_k}(c(p;s))\right| \le (-\log t)^2 \rho_7 \] for some constant $\rho_7>0$. These estimates together with \eqref{eq:grad_g_estimate} and $|\xi_t(p)|\le 2|V| t^\kappa$ imply \begin{equation} \label{eq:estimate_xiF_xiG} \max\left(|\xi_t(p) F(c(p;s))_{jk}|, |\xi_t(p) G(c(p;s))_{jk}|\right) \le \rho_8 t^\kappa \end{equation} for some constant $\rho_8>0$. The estimates \eqref{eq:estimate_derxi_grad}, \eqref{eq:estimate_xiF_xiG} imply the first inequality of \eqref{eq:estimate_coeff}. The proposition is proved. \end{proof} \begin{rmk} By a similar argument, we can prove that the function $\delta_t$ constructed in the above proof is small in the $C^\infty$-topology, i.e.~$\|\delta_t\|_{C^m} =O(t^\kappa)$ for all $m\ge 0$, although we do not need this result. \end{rmk} We end this subsection with the two lemmas used in the above proof. \begin{lem} \label{lem:grad} Suppose $0<t<1$ and let $g_t\colon P_\C \to \C$ be the function defined by $g_t(p) = \sum_{q\in V} e^{\mathtt{i} \theta_q} t^{\beta_q(p)}$. There exist constants $\rho_1,\rho_2>0$ independent of $p\in P_\C$ such that \[ \frac{|\grad g_t(p)|}{|g_t(p)|} \ge (-\log t)(\rho_1 - \rho_2 t^\kappa) \] whenever $g_t(p) \neq 0$. In particular, the hypersurface $\mathring{Z}_{t,\theta}$ is smooth for sufficiently small $t>0$. \end{lem} \begin{proof} Fix $p\in P_\C$. Set $\beta_{q_0}(p) = \min_{q\in V}\beta_q(p)$ and $K=\{q\in V: \beta_k(p) \le \beta_{q_0}(p)+\kappa\}$. By the condition (ii) above, $K$ is linearly independent as a subset of $Q_\R$. We have \begin{align*} |\grad g_t(p)| &= (-\log t)\left| \sum_{q\in V} q e^{\mathtt{i} \theta_q} t^{\beta_q(p)}\right | \\ & \ge (-\log t) \cdot |t^{\beta_{q_0}(p)}| \left( \left| \sum_{q\in K} q e^{\mathtt{i} \theta_q} t^{\beta_q(p) - \beta_{q_0}(p)}\right| - \left|\sum_{q\in V\setminus K} q e^{\mathtt{i} \theta_q} t^{\beta_q(p) - \beta_{q_0}(p)}\right|\right) \\ & \ge (-\log t) \cdot |t^{\beta_{q_0}(p)}| \left( \rho_1 \sum_{q\in K} |t^{\beta_q(p)-\beta_{q_0}(p)}|-\rho_3 t^\kappa\right) \end{align*} for some constants $\rho_1,\rho_3>0$, where we used the fact that $K$ is linearly independent and that all norms on a finite dimensional vector space are equivalent. Similarly we have \[ |g_t(p)| \le |t^{\beta_{q_0}(p)}| \left(\sum_{q\in K} |t^{\beta_q(p)-\beta_{q_0}(p)}|+\rho_4 t^\kappa \right) \] for some constant $\rho_4>0$. Combining these inequalities, we get \begin{align*} \frac{|\grad g_t(p)|}{|g_t(p)|} &\ge (-\log t)\frac{\rho_1-\rho_3 t^\kappa/\sum_{q\in K}|t^{\beta_q(p)-\beta_{q_0}(p)}|}{1+\rho_4 t^\kappa/\sum_{q\in K}|t^{\beta_q(p)-\beta_{q_0}(p)}|}\\ & \ge (-\log t) \left(\rho_1 - (\rho_1\rho_4+\rho_3) \frac{t^\kappa}{\sum_{q\in K}|t^{\beta_q(p)-\beta_{q_0}(p)}|}\right). \end{align*} Setting $\rho_2 = \rho_1 \rho_4 + \rho_3$, we obtain the conclusion since $\sum_{q\in K} |t^{\beta_q(p)-\beta_{q_0}(p)}| \ge 1$. \end{proof} \begin{lem} \label{lem:Gronwall} Let $x(s)$ be a vector-valued function satisfying the differential equation \[ \frac{dx}{ds}(s) = y(s) + A(s) x(s) \] where $y(s)$ is a vector-valued smooth function and $A(s)$ is a matrix-valued smooth function. Then we have \[ |x(1) - x(0)| \le \left (\int_0^1 |y(s) + A(s)x(0)| ds \right) e^{\int_0^1 |A(s)|ds} \] where $|A(s)|$ denotes the operator norm. \end{lem} \begin{proof} We have the integral inequality: \begin{align*} |x(s)-x(0)| & = \left| \int_0^s y(u) + A(u) x(0) + A(u) (x(u)-x(0)) du\right| \\ & \le \int_0^1 |y(u)+A(u) x(0)| du + \int_0^s |A(u)| \cdot |x(u) - x(0)| du. \end{align*} The lemma follows by the Gronwall inequality. \end{proof} \subsection{Complex volume of a polytope} We introduce a complex volume of an $m$-dimensional polytope `enclosed' by a complex hyperplane arrangement in $\C^m$. The complex volume will appear in the calculation of periods of $C_{t,\theta}$. Let $P\subset \R^m$ be a compact convex polytope (with non-empty interior) equipped with an orientation (as a manifold with corners). Let $\{F_i\}_{i\in I}$ be the set of facets (i.e.~faces of codimension one) of $P$. We assume that if $\bigcap_{i \in K} F_i \neq \emptyset$ for $K\subset I$, then $\bigcap_{i \in K} F_i$ is a face of codimension $|K|$. Suppose that we have a collection of complex affine hyperplanes $\{H_i\}_{i \in I}$ in $\C^m$ labelled by the same index set $I$. We write $H_i=(\alpha_i+ \mu_i=0)$, where $\alpha_i\colon \C^m \to \C$ is a linear function and $\mu_i\in \C$ is a constant. We require that the hyperplane arrangement $\{H_i\}_{i \in I}$ satisfies the following (open) condition: for every subset $K\subset I$ such that $\bigcap_{i \in K} F_i \neq \emptyset$, the hyperplanes in the collection $\{H_i\}_{i \in K}$ intersect transversally along a codimension $|K|$ affine subspace. Consider a smooth map \[ \Phi \colon P \to \C^m \] satisfying \[ \Phi(F_i) \subset H_i \] for all $i \in I$. Note that this condition (along with the requirement on $\{H_i\}_{i\in I}$ imposed above) determines the image of a vertex in $P$ under $\Phi$. Fix a holomorphic volume form $d\operatorname{vol} = r dc_1 \wedge \cdots \wedge dc_m$ on $\C^m$, for some $r \in \C^*$. The \emph{complex volume} of $(P,\Phi)$ is defined to be \begin{equation} \label{eq:complex_volume} \operatorname{vol}_\C(P,\Phi) = \int_P \Phi^* (d\operatorname{vol}). \end{equation} \begin{lem} \label{lem:complex_volume} For a fixed volume form $d \operatorname{vol}$, the complex volume $\operatorname{vol}_\C(P,\Phi)$ depends only on the hyperplane arrangement $\{H_i\}_{i \in I}$ on $\C^m$ indexed by facets of $P$. Moreover, it is a polynomial function of $\{\mu_i\}_{i\in I}$. \end{lem} \begin{proof} This follows from a repeated application of the Stokes theorem. More generally, for a polynomial $m$-form $\omega$ on $\C^m$, we can see that the integral \[ \int_P \Phi^*\omega \] does not depend on the choice of $\Phi$ satisfying the above condition. Since any polynomial $m$-form on $\C^m$ is exact, by Stokes theorem, we can write it as the sum of integrals $\int_{F_i} (\Phi|_{F_i})^*\eta_i$ over $i \in I$ with $\eta_i$ being a polynomial $(m-1)$-form on $H_i$. The conclusion follows by induction on the dimension. \end{proof} \subsection{Computing periods of $C_{t,\theta}$ tropically} By construction, the cycle $C_{t,\theta}$ is identified with the positive real cycle $C_t^+$ via the map $\widetilde{\Phi}_t=\Phi_t+\delta_t$. The tropical decomposition of $C_t^+$ in Section \ref{sec:decomposing-domain} then induces a decomposition of $C_{t,\theta}$. We can decompose and calculate the period of $C_{t,\theta}$ in almost the same way as before, with the only significant difference being that we use the complex volume (instead of the real volume) of a polytope and apply a complexified Duistermaat--Heckman theorem. We use the notation in Section \ref{sec:periodR}. Let $\kappa>0$ be as in Section \ref{sec:construction_Cttheta} and let $\epsilon>0$ be a sufficiently small number satisfying $0<\epsilon<\kappa/2$. Let $B_t^{q,K}\subset B_t$ be the region introduced in Section \ref{sec:decomposing-domain} defined by this $\epsilon$. We have \[ \int_{C_{t,\theta}} \Omega_{t,\theta} = \sum_{q,K} \int_{B^{q,K}_t} \widetilde{\Phi}_t^*i_t^* \Omega_{t,\theta} \] where, as before, it suffices to consider the sum over all $q\in V$ and a subset $K\subset V$ not containing $q$ such that the facets of $\Delta_\lambda$ corresponding to $K\sqcup \{q\}$ have nonempty intersection. Let $(a,b_k,c_j)$ be affine coordinates on $P_\R$ from Section \ref{sec:approx_each_region} associated with the choice of $q$ and $K$. We naturally extend these coordinates to complex coordinates on $P_\C$. Similarly to equations \eqref{eq:f_t_on_region}--\eqref{eqn:negl}, we have \[ f_{t,\theta}\circ i_t(p) = e^{\mathtt{i}\theta_q} t^a \left( 1+ \sum_{k\in K} e^{\mathtt{i}(\theta_k-\theta_q)}t^{b_k} + h_t(p) \right) \] where $h_t(p) = \sum_{m\in V\setminus (\{q\}\sqcup K)}e^{\mathtt{i}(\theta_i - \theta_q)} t^{\beta_m(p) - a}$ satisfies the uniform estimate \[ h_t(p) \in O(t^\epsilon), \quad \parfrac{h_t}{p_i}(p) \in O((-\log t) t^\epsilon) \] over $\widetilde{\Phi}_t(B^{q,K}_t)$; this follows from the $C^0$-estimate for $\delta_t$ in Proposition \ref{prop:delta}. Hence a calculation similar to equations \eqref{eq:Omega_in_xy}--\eqref{eq:Omega_in_bc} shows \begin{align*} i_t^*\Omega_{t,\theta} & = \left. r_{q,K} (-\log t)^{n+1} \frac{da\wedge \bigwedge_{k\in K} db_k \wedge \bigwedge dc_j}{d(f_{t,\theta}\circ i_t)}\right|_{(f_{t,\theta} \circ i_t)^{-1}(1)} \\ & = (1+O(t^\epsilon))\cdot (-\log t)^n \cdot r_{q,K} \bigwedge_{k\in K} db_k \wedge \bigwedge _j dc_j \end{align*} over $\widetilde{\Phi}_t(B^{q,K}_t)$, where $r_{q,K}>0$ is the number in Section \ref{sec:approx_each_region}. Consider the Riemannian metric $g$ on $B_t$ induced from the Euclidean metric on the ambient space $P_\R \cong \R^{n+1}$. Then the volume of $B_t$ with respect to $g$ is bounded as $t\to +0$; this follows from the leading asymptotics of periods $\int_{B_t} i_t^*\Omega_t$ in Theorem \ref{thm:periodR} and the estimate \begin{align*} \int_{B_t} d\operatorname{vol}_g & = \int_{B_t} |d(f_t \circ i_t)|\frac{dp_0 \cdots dp_n}{d(f_t \circ i_t)} \le \frac{\rho}{(-\log t)^n} \int_{B_t} i_t^*\Omega_t \end{align*} for some constants $\rho>0$, where we used the estimate $|d(f_t\circ i_t)| \le \rho (-\log t)$ over $B_t$. When we define the $C^1$-norm of $\delta_t|_{B_t}$ using the metric $g$, we have $\|\delta_t|_{B_t}\|_{C^1} \le \|\delta_t\|_{C^1}$; hence by Proposition \ref{prop:delta}, $\|\delta_t|_{B_t}\|_{C^1} = O(t^\kappa)$. Moreover it is easy to show that the $C^1$-norm of $\Phi_t$ (and hence of $\Phi_t|_{B_t}$) is bounded. From these facts, we obtain the approximation: \begin{align*} \int_{B^{q,K}_t} \widetilde{\Phi}_t^*i_t^*\Omega_{t,\theta} & = (1+O(t^\epsilon)) \cdot (-\log t)^n \cdot r_{q,K} \int_{B_t^{q,K}} \widetilde{\Phi}_t^*\left(\textstyle{\bigwedge_{k\in K} db_k \wedge \bigwedge _j dc_j}\right) \\ & = (1+O(t^\epsilon)) \cdot (-\log t)^n \cdot \left( r_{q,K} \int_{B_t^{q,K}} \Phi_t^*\left(\textstyle{ \bigwedge_{k\in K} db_k \wedge \bigwedge _j dc_j}\right) + O(t^\kappa) \right). \end{align*} This corresponds to \eqref{eq:Omega_in_bc} in Section \ref{sec:approx_each_region}. Over $B^{q,K}_t$, we have $1= \sum_{k\in V} t^{\beta_k(p)} \le |V| t^a$, and thus $0\le a\le \log |V|/(-\log t)$. Therefore, by taking $t>0$ sufficiently small, we may assume that $0\le \beta_q=a \le \epsilon<\kappa$ and that $\beta_k = a+b_k \le 2 \epsilon < \kappa$ for $k\in K$ over $B^{q,K}_t$. This implies \[ \Phi_t^*b_k= b_k+ \mathtt{i} \frac{\theta_k-\theta_q}{-\log t} \qquad \text{over $B^{q,K}_t$} \] for $k\in K$ by the condition \eqref{eq:bdry_cond_phi}. Therefore we have \begin{align} \label{eq:integral_vol_FqK(b)} \begin{split} r_{q,K} \int_{B_t^{q,K}} \Phi_t^*\left(\textstyle{ \bigwedge_{k\in K} db_k \wedge \bigwedge _j dc_j}\right) & = \int_{[0,\epsilon]^K} d\tilde{b} \int_{B_t^{q,K}\cap \{b=\tilde{b}\}} \Phi_t^*(d\operatorname{vol}_{q,K}) \\ & = \int_{[0,\epsilon]^K} d\tilde{b} \int_{F_{q,K}(\tilde{b})} (\Phi_t \circ s_{\tilde{b}})^*(d\operatorname{vol}_{q,K}) \end{split} \end{align} where $B_t^{q,K} \cap \{b=\tilde{b}\}$ denotes the subset of $B_t^{q,K}$ where the values of the coordinates $b=(b_k)_{k\in K}$ equal $\tilde{b}\in [0,\epsilon]^K$, $F_{q,K}(\tilde{b}) = \pi_c(B_t^{q,K} \cap \{b=\tilde{b}\})$ is its projection to the $c$-plane as appears in Section \ref{sec:approx_vol_polytope}, and $s_{\tilde{b}}$ denotes the (unique) section of the projection $\pi_c\colon B_t^{q,K} \cap \{b= \tilde{b}\} \to F_{q,K}(\tilde{b})$; we have $s_{\tilde{b}}(c)= (a_{q,K}(\tilde{b},c),\tilde{b},c)$. Also $d\operatorname{vol}_{q,K} = r_{q,K} \bigwedge_j dc_j$ (see \eqref{eq:residual_vol}) is now regarded as a holomorphic form on $P_\C$. Recall the function $a_{q,K}'(b)$ from \eqref{eq:approx_aqK}. It differs from $a_{q,K}(b,c)$ by a function of order $O(t^\epsilon)$. Moreover, we have the uniform estimate over the region $\pi_{b,c}(B^{q,K}_t)$ \[ \parfrac{a_{q,K}(b,c)}{c_j} = O(t^\epsilon). \] In fact, by differentiating \eqref{eq:defining_eqn_Bt} with respect to $c_j$, we get \[ \parfrac{a_{q,K}}{c_j}(b,c) \left( 1+\sum_{m\notin \{q\}\sqcup K}\left(\partial_a \beta_m-1\right)t^{\beta_m(a_{q,K}(b,c),b,c)} \right) = - \sum_{m\notin\{q\}\sqcup K} (\partial_{c_j}\beta_m) t^{\beta_m(a_{q,K}(b,c),b,c)} \] where we regard $\beta_m=\beta_m(a,b,c)$ as an affine linear form in $a,b,c$; the above estimate follows by $\beta_m(a_{q,K}(b,c),b,c)\ge a_{q,K}(b,c)+\epsilon>\epsilon$ over $\pi_{b,c}(B^{q,K}_t)$. By this estimate, we can replace the section $s_b$ in \eqref{eq:integral_vol_FqK(b)} with $s_b'(c) = (a_{q,K}'(b),b,c)$ with error terms of order $O(t^\epsilon)$. Furthermore, by the same argument as in Section \ref{sec:approx_vol_polytope}, we can replace $F_{q,K}(b)$ with the polytope $F_{q,K}'(a'_{q,K}(b),b)$ with error terms of order $O(t^\epsilon)$ again. Namely, we have \begin{align*} \int_{F_{q,K}(b)} (\Phi_t\circ s_b)^*d\operatorname{vol}_{q,K} & = \int_{F_{q,K}(b)} (\Phi_t\circ s_b')^*(d\operatorname{vol}_{q,K}) + O(t^\epsilon) \\ & = \int_{F_{q,K}'(a_{q,K}'(b),b)} (\Phi_t\circ s_b')^*(d\operatorname{vol}_{q,K}) + O(t^\epsilon) \\ & = \operatorname{vol}_\C\left(F_{q,K}'(a_{q,K}'(b),b), \Phi_t\right). \end{align*} In the last line, we identified the polytope $F_{q,K}'(a_{q,K}'(b),b)$ with its image by $s_b'$ (as we did in Section \ref{sec:approx_vol_polytope}) and considered its complex volume \eqref{eq:complex_volume} with respect to $d\operatorname{vol}_{q,K}$. Note that $\Phi_t(F_{q,K}'(a_{q,K}'(b),b))$ is contained in the affine subspace of $P_\C$ where the values of the complex affine-linear functions $a$, $b_k$ are fixed (again by the condition \eqref{eq:bdry_cond_phi}). Applying the inclusion-exclusion principle for the complex volumes of polytopes as we did in Section \ref{sec:approx_vol_polytope}, we arrive at the formula\footnote{We express the complex volume of $F_{q,K}'(a,b)$ as a signed sum of the complex volumes of $\bigcup_{b'\in [0,\epsilon]^I} E_{q,K\sqcup I}(a,b,b')$. Then we use the fact that $\Phi_t^*(b'_i)$ (with $i\in I$) on $\bigcup_{b'\in [0,\epsilon]^I} E_{q,K\sqcup I}(a,b,b')$ equals $b'_i+\mathtt{i} (\phi_i-\phi_q)/(-\log t)$ for $a=a'_{q,K}(b)$ to factor out $db'$.}: \[ \int_{C_{t,\theta}} \Omega_{t,\theta} = (1+O(t^\epsilon)) \ (-\log t)^n \ \left( \sum_{q,K\subset J,q\notin J} (-1)^{|J\setminus K|}\int_{[0,\epsilon]^J} \operatorname{vol}_\C\left(E_{q,J}(a_{q,K}'(b),b,b'),\Phi_t\right) db db' + O(t^\epsilon)\right) \] where we write $b=(b_j)_{j\in K}$ and $b'=(b'_j)_{j\in J\setminus K}$ and used the holomorphic form $d \operatorname{vol}_{q,J}$ to define the complex volume for $(E_{q,J}(a_{q,K}'(b),b,b'),\Phi_t)$. This generalizes \eqref{eqn:perH}. Finally we apply a complex version of the Duistermaat--Heckman theorem. Consider the polytope $E_{q,J}(a,b)$ with $0<a, b_j<\epsilon$. The image of $E_{q,J}(a,b)$ under $\Phi_t$ is contained in the complex affine subspace of $P_\C$ defined by \begin{align*} \beta_q & = a + \mathtt{i} \frac{\theta_q}{-\log t}, \\ \beta_j & = a+b_j+ \mathtt{i}\frac{\theta_j}{-\log t} \qquad j\in J. \end{align*} by \eqref{eq:bdry_cond_phi}. The facets of $E_{q,J}(a,b)$ are given by $E_{q,J}(a,b) \cap \{\beta_m = a\}$ for $m\notin \{q\} \sqcup J$. These facets map (under $\Phi_t$) to the complex affine hyperplane given by \[ \beta_m-\beta_q= \mathtt{i} \frac{\phi_m-\phi_q}{-\log t} \] by \eqref{eq:bdry_cond_phi}. Therefore its complex volume $\operatorname{vol}_\C(E_{q,K}(a,b),\Phi_t)$ is a polynomial function of the constant terms of these complex affine-linear forms by Lemma \ref{lem:complex_volume}. By analytic continuation of Lemma \ref{lem:DH}, we get \[ \operatorname{vol}_\C(E_{q,J}(a,b),\Phi_t) = \int_{Y_{\Delta_\lambda}} \exp\left(\omega_{\lambda}+\mathtt{i} \frac{\sum_{k\in V} \theta_k D_k}{\log t} - \sum_{j\in J} b_j D_j - a \sigma\right). \] We obtain this from Lemma \ref{lem:DH} by substituting $\lambda_k + \mathtt{i} (\theta_k/\log t)$ for $\lambda_k$ for all $k\in V$. The rest of the argument works in exactly the same way as before, just by replacing $\lambda_k$ with $\lambda_k+\mathtt{i}(\theta_k/\log t)$, and we arrive at Theorem \ref{thm:cycle_with_phase}. \bibliographystyle{amsalpha}

dd1df8ef4448038adda9c8b1a72ce0d1e0d480ec

\section{Introduction} Over the past decades, deep neural networks (DNN) experienced unprecedented rapid development in company with the data explosion, and achieves impressive success in matching human intelligence in many applications, such as IBM's Watson \cite{kelly2013smart}, DeepMind's Atari \cite{mnih2015human} and AlphaGo \cite{alphago}. However, extensive studies reveal that DNN is vulnerable to adversarial attacks \cite{szegedy2013intriguing,goodfellow2014explaining,Papernot2017,pei2017deepxplore,tian2017deeptest,madry2017towards}, where imperceptible perturbations to an input can cause DNN to misclassify images with high confidence. This could impede DNN's application especially in safety critical scenarios, and cause losses and even severe tragedies if flawed DNNs are deployed to safety-critical systems (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, the recent Tesla autonomous driving accident \cite{tesla}). Such DNN vulnerabilities have since been also studied in other domains such as reinforcement learning \cite{behzadan2017vulnerability,huang2017adversarial}, speech recognition \cite{adv_speech} and some natural language processing (NLP) tasks \cite{Alzantot,advcompre,papernot2016crafting,kuleshov2018adversarial}. \begin{figure}[t] \centering \includegraphics[width=\columnwidth]{image/dnc_teaser.pdf} \caption{(Top) Control signals from DNC with benign input can be compared with (Bottom) control signals from DNC with adversarial input.} \label{fig:teaser} \end{figure} Differential neural computer (DNC) \cite{graves2016hybrid} was recently proposed as an extension of neural Turing machine (NTM) \cite{NTM}, with the addition of memory attention mechanisms that control where data is stored in the memory and temporal attention that records the order of events. Unlike the common feed-forward DNN and recurrent neural networks, DNC is Turing complete and adopts the Von Neumann architecture with a scalable memory module. DNC has achieved superior performance on several complex tasks that demand the representation of data structure and storage of data over long timescales, as demonstrated on bAbI question answering (QA) tasks, graph experiments (London Underground and family tree) and a block puzzle solving task \cite{graves2016hybrid}. Many adversarial attack studies have been conducted on feed-forward DNN and some on recurrent neural networks. Even with DNC's promise as a more universal approach to solving machine learning tasks, there lacks a in-depth study on whether such adversarial attacks also applied to it. Unlike common DNN, DNC separates its computational and memory capabilities into a DNN central controller and memory module respectively. This brings an opportunity to study how adversarial attacks affect the computational and memory components of a neural network separately. With the augmented memory enhancement, DNC tends to be more robust than a NTM. However, it is still unknown how the memory contributes to DNC's robustness against adversarial attacks. Gradient based adversarial techniques, from computer vision domain, do not apply to our experiments due to discrete nature of input in the NLP domain. The replacement of a single word can drastically alter the semantics of a text or introduce grammatical errors. Previous NLP adversarial methods either erase or change words directly with domain-specific rules or require human intervention \cite{Alzantot,kuleshov2018adversarial,advcompre}. To address these challenges, we propose scalable adversarial strategies that rely on metamorphic relations to generate attacks which are grammatically sound and preserve correct answers in QA tasks. In this paper, we demonstrate that DNC can be vulnerable to our adversarial attacks on bAbI QA tasks, where it has originally achieved near-perfect performance \cite{rsDNC}. We also find that a larger size of the memory module provides limited benefit in DNC's robustness against such attacks. The effectiveness of such attacks are determined by their position, type of content and length, as shown by our experiments. Finally, we analyze the activities of DNC controller control signals and find that adversarial attacks disrupt the read, write and erase functions of the DNC. The major contributions of this paper are the following: (1) \textbf{First adversarial attack on DNC}, a neural network architecture that display state-of-the-art performance in various tasks, that demonstrates its vulnerabilities. (2) \textbf{Pick-n-Plug and Pick-Permute-Plug}: Using two new automated and scalable strategies to generate grammatically correct adversarial attacks in the NLP QA domain. (3) Evidence that larger memory size provides limited benefit in resisting such adversarial attacks. (4) Analysis of DNC's control signals, as illustrated in Figure~\ref{fig:teaser}, which shows that adversarial attacks disrupt the read, write and erase functions of the DNC. \section{Background and Related Work} \subsubsection{Adversarial Attack} Adversarial attacks were first discovered in computer vision domains \cite{szegedy2013intriguing}. Carefully perturbed images, with changes imperceptible to humans, can easily fool DNNs. Since then, we have witnessed an arms race between attackers \cite{iclr18-ae-boundary-analysis,iclr18-a-boundary-attack,iclr18-ae-natural} and defenders \cite{iclr18-b-counter,iclr18-b-defense-gan,iclr18-b-thermo}. The presence of adversarial examples have permeated into various computer vision tasks apart from visual classification, such as face recognition \cite{sharif2016accessorize}, object detection \cite{Xie_2017_ICCV}, semantic segmentation \cite{fischer2017adversarial}, generative modeling \cite{kos2018adversarial}, robustness testing \cite{felix_ase18_gauge,felix_arxiv18_fuzz}, \emph{etc}\onedot} \def\vs{\emph{vs}\onedot. Although adversarial attack techniques are extensively studied in computer vision domain, there is limited work conducted in NLP domain \cite{advcompre,Alzantot,kuleshov2018adversarial,advrnnpapernot}. One challenge lies in the discrete nature of word inputs in NLP which makes implementation of gradient-based perturbation methods challenging. Different from adversarial attack for images where small pixel changes are very unlikely to alter the correct class of an image, a change of word in a body of text may completely change the meaning of it under a particular NLP task or introduce grammatical errors. For adversarial attacks in NLP, Jia \cite{advcompre} proposed to add distracting sentences to the original text, which are generated from the task's question with rules. However, the grammatical correctness and the preservation of correct answer rely on manual checks, which makes such method difficult to scale. Word substitution based attacks \cite{Alzantot,papernot2016crafting,kuleshov2018adversarial} were also proposed, by changing words in original text with synonyms. In \cite{Alzantot}, a portion of the substituted text are interpreted by humans to be a different class, highlighting the challenge of creating adversaries that avoid changing the meaning of original text. In this paper, we attempt to overcome these challenges with adversaries generated based on metamorphic relations. Our adversarial methods are scalable and do not disrupt the information from the original text for QA tasks. \subsubsection{Differential Neural Computer (DNC)} A DNC is a DNN augmented with an external memory module in the form of a matrix $M \in \mathbb{R^{\mathrm{N \times W}}}$ \cite{graves2016hybrid}. The DNN of DNC acts as the controller module, whose operations can be learn with gradient descent, while the external memory matrix serves as a module for data storage. The DNC's controller and memory module are like its CPU and RAM respectively. The DNC's memory can be written to and accessed by the controller at each input time-step. At a particular time-step, the DNC controller takes in an input vector $\mathbf{x}_t \in \mathbb{R^{\mathrm{X}}}$, and a set of read vectors from the previous time-step $\boldsymbol{\mu}_{t-1} \in \mathbb{R^{\mathrm{P}}}$. \begin{align} (\mathbf{v}_t, \boldsymbol{\xi}_t)=\mathrm{Controller}([\mathbf{x}_t, \boldsymbol{\mu}_{t-1}], \theta_c) \end{align} $\theta_c$ is the controller's trainable weight parameters, $\mathbf{v}_t \in \mathbb{R^{\mathrm{C}}}$ is the controller output while $\xi_t$ is a set of control signals. The controller uses its write and read heads in order to manage the memory matrix. At each time-step, the controller's set of control signals $\boldsymbol{\xi}_t$ represents the operations of these heads. These control signals can be categorized into gates, keys or vectors. Their values determine how, where and what is being read, written and erased from the memory matrix. A series of operations in the memory module with the control signals and its current memory matrix $M_t$ erase and write new data and produces a concatenation of read vectors $\boldsymbol{\mu}_t \in \mathbb{R^{\mathrm{P}}}$. \begin{align} \boldsymbol{\mu}_t = \mathrm{MemoryModule}(\boldsymbol{\xi}_t, M_t) \end{align} The final output of the DNC is a sum of weighted controller output and weighted concatenation of read vectors from the memory module. \begin{align} \mathbf{y}_t=W_v \mathbf{v}_t + W_{\mu} \boldsymbol{\mu}_t + \mathbf{b}_t \end{align} where $W_v \in \mathbb{R^{\mathrm{Y \times C}}}$, $W_{\mu} \in \mathbb{R^{\mathrm{Y \times P}}}$ and $\mathbf{b}_t \in \mathbb{R^{\mathrm{Y}}}$ are trainable weights of the DNC. There are two mechanisms for writing into memory $M_t$: least used location and content-based addressing. Least used location mechanism find least utilized memory locations for new information to be written in. Content-based addressing find memory locations which have the highest similarity with the controller's write keys. Allocation gate determines how much of these two mechanisms to use in order to write new information. For reading, there are also two mechanisms in the original DNC paper. The first is content-based addressing which is very similar to that of the writing operation. Read keys, rather than write keys, are used to find locations to read from. The second mechanism, called temporal linkage mechanism, helps retrieve information in chronological order of when they are written. In an improved version of DNC, this mechanism is dropped for bAbI tasks to increase its memory efficiency \cite{rsDNC}. Free gate values are used to forget data which were recently accessed by the controller read heads. \subsubsection{bAbI Dataset} DNC that we are evaluating has shown near-perfect performance on the bAbI dataset which makes it a good candidate to evaluate the DNC's vulnerability to attacks. The bAbI dataset has 20 question \& answer tasks of different themes to evaluate a range of logical reasoning capabilities. Each task contains stories where each story has one or more following questions with answers that can be derived from the story. After removing digits from the stories, all words are converted into lower case and splitted into word tokens. The whole vocabulary contains 156 unique words and three symbols: `?', `.', `-'. The `-' is used to indicate positions in the stories where answers are required. The performance of bAbI tasks is evaluated by word error rate (WER) which is the rate of incorrect answers over total number of answers. \section{Metamorphic Relation-Based Adversaries}\label{sec:method} In a general classification task, a successful adversarial attack is a modified input $x'$ of an original input $x$ such that it causes the originally correct prediction $y = f(x)$ from a classifer $f$ to be incorrect such that $y' = f(x')$ and $y' \neq y$. This definition can be extend to NLP question answering (QA) tasks such that $x$ and $x'$ can be generalized to $X$ and $X'$, where $X = [\mathbf{x}_1, \mathbf{x}_2, \ldots, \mathbf{x}_m]$, $X' = [\mathbf{x}'_1, \mathbf{x}'_2, \ldots, \mathbf{x}'_{m'}]$ and $\mathbf{x}_i, \mathbf{x}_i' \in \mathbb{R^{\mathrm{X}}}$ are one-hot vectors of words in the original and adversarial input sequence respectively. $y$ can also be extended to become $Y$, where $Y=[\mathbf{y}_1,\ldots]$ which is a sequence of correct answer one-hot word vectors $\mathbf{y}_i$. \subsubsection{Metamorphic Transformation} To generate adversarial attacks which do not change the original answer to bAbI QA tasks, we draw inspiration from metamorphic relations. An example of metamorphic relations for sine function is $\sin(x + \pi) = \sin(x)$. Metamorphic relations have been used in testing software \cite{chen1998metamorphic} and supervised classifiers \cite{xie2011testing}. Here, we define a metamorphic transformation $T$ as a function that maps an input $x$ to $x'$ which satisfy a metamorphic relation with $f$. More formally, \begin{align} x' = T(x)~~~~~~~~~~~~f(x) = f(x') \nonumber \end{align} $x$ is the original input and $x'$ is the output of a metamorphic transformation of $x$. In the example of {\it sine} function where $f(x) = \sin{(x)}$, a valid metamorphic transformation is $T(x) = x + \pi$. Similarly, in the QA tasks, the input $x$ and $x'$ can be generalized to $X$ and $X'$, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, $X' = T(X)$, $f(X) = f(X')$, where $f$ is an oracle that is \emph{always correct}, $X = [\mathbf{x}_1, \mathbf{x}_2, ..., \mathbf{x}_m]$, $X' = [\mathbf{x}'_1, \mathbf{x}'_2, ..., \mathbf{x}'_{m'}]$ and $\mathbf{x}_i, \mathbf{x}_i' \in \mathbb{R^{\mathrm{X}}}$ are one-hot vectors of words in the original and transformed adversarial input sequence respectively. For any input $X'$ generated by the metamorphic transformation on $X$, the answer $f(X')$ would remain unchanged from $f(X)$ under the oracle $f$. Consider a DNN model for a QA task as $f'$ where its prediction of an input sequence $X$ is $f'(X)$. If an adversarial input $X'=T(X)$ is generated with a metamorphic transformation $T$ (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, $f(X) = f(X')$) such that $f'(X) \neq f'(X')$, this would be a successful adversarial attack. \subsubsection{Pick-n-Plug} A metamorphic transformation can be composed of a series of $n$ operator functions $g_1, g_2,\ldots,g_n$, such that: $T(X) = g_n(\ldots g_2(g_1(X))\ldots)$. We propose Pick-n-Plug which relies on metamorphic transformation $T_{\mathrm{pick-n-plug}}$ to generate adversarial attacks. It consists of a pick operator $g_{\mathrm{pick}}$ to draw adversarial sentences from a particular task (source task) and plug operator $g_{\mathrm{plug}}$ to inject these sentences into a story from another task (target task), without changing its correct answers. The $g_{\mathrm{pick}}$ step to draw sentences can be a random search, as we have demonstrated in our experiments, or other search methods. Pick-n-Plug requires no human intervention in generating adversarial attacks while ensuring grammatical correctness. Figure~\ref{fig:adv_example} shows an example of successful Pick-n-Plug attack where task \#19 and \#3 are the target and source task respectively. More formally, \begin{gather} T_{\mathrm{pick-n-plug}}(X) = g_{\mathrm{plug}}(g_{\mathrm{pick}}(X)) = X' \mathrm{~~where} \nonumber\\ g_{\mathrm{pick}}(X) = (X, [S_1, \ldots, S_l]) \nonumber\\ g_{\mathrm{plug}}(X, [S_1, \ldots, S_l]) = X' \nonumber \end{gather} where $S_i = [\mathbf{x}^{S_i}_1, \mathbf{x}^{S_i}_2, \ldots, \mathbf{x}^{S_i}_j]$ is a sequence of word vectors in an adversarial sentence, $l$ is the number of adversarial sentences picked from the source task. $X'$ is an adversarial input as the final output of the Pick-n-Plug metamorphic transformation $T_{\mathrm{pick-n-plug}}$. In our adversarial attacks, we identify a pair of a target task and a source task. One or more sentences $S_1, \ldots, S_l$ from the source task are picked, with an operator $g_{\mathrm{pick}}$ and then plugged $g_{\mathrm{plug}}$ into a story from the target task, with the aim of changing the DNC's answer from correct to incorrect. To maintain coherence within the injected text, blocks of consecutive sentences from the source task are used to attack target stories, with the first sentence in the block being randomly sampled. $T_{\mathrm{pick-n-plug}}$ relies on a metamorphic relation between sentences and questions from several pairs of tasks such as task \#19 and \#3. Sentences from one task do not interfere with the information expressed by sentences from another task, and hence do not change the correct answer to their questions. For example, the directional information contained in Sentence 1: `\textit{The office is south of the hallway.}' from task \#19 is still preserved even when Sentence 2: `\textit{Mary journeyed to the hallway.}' from task \#3 is in the same story. To illustrate, for a story containing Sentence 1, if the correct answer to the question: `\textit{How do you go from the office to the hallway?}' is `\textit{north}', adding Sentence 2 `\textit{Mary journeyed to the hallway.}' to the story should not change the correct answer. Pairs of unrelated text corpora would potentially have this property. While using the Pick-n-Plug to attack a target task, the choice of source task, number of injected sentences, and position of the adversarial injection can be varied. In the following sections of this paper, we show the effect of these factors on the DNC's performance. \begin{figure} \begin{minipage}[c]{0.49\columnwidth} \centering \includegraphics[width=0.8\columnwidth]{image/adversarial_example_fx.pdf}\\~\\ \end{minipage}\hfill \begin{minipage}[c]{0.51\columnwidth} \caption{A successful Pick-n-Plug attack on target task \#19 with source task \#3, where the adversarial sentences (in bold) are picked from task \#3 and injected right before the question, after the story. In the presence of these adversarial sentences the DNC predicts `s', `e' which stands for `South' and then `East'. The DNC correctly predicts `n', `n' (`North' then `North') without adversarial sentences.}\label{fig:adv_example} \end{minipage} \vspace{-10mm} \end{figure} \subsubsection{Pick-Permute-Plug} We also propose Pick-Permute-Plug with metamorphic transformation $T_{\mathrm{pick-permute-plug}}$ to extend the adversarial capability of Pick-n-Plug. In Pick-n-Plug, the diversity of adversarial injected sentences is restricted by the text of the source task. With an additional $g_{\mathrm{permute}}$ operator after picking sentences ($g_{\mathrm{pick}}$) from a source task, words in a particular adversarial sentence can be permuted with `synonyms' to generate much wider range of possible attacks. Since only words in the in adversarial sentences are permuted, the correct answer is still preserved. $g_{\mathrm{permute}}$ can optionally rely on the original input $X$ to influence $S'_1, .., S'_l$ according to conditions such as to enforce similar grammatical tense or style of writing. The Pick-Permute-Plug metamorphic transformation can be summarized as: \begin{gather} T_{\mathrm{pick-permute-plug}}(X) = g_{\mathrm{plug}}(g_{\mathrm{permute}}(g_{\mathrm{pick}}(X))) = X'\nonumber\\ \mathrm{where~~~~}g_{\mathrm{pick}}(X) = (X, [S_1, \ldots, S_l])\nonumber\\ g_{\mathrm{permute}}(X, [S_1, \ldots, S_l]) = (X, [S'_1, \ldots, S'_l])\nonumber\\ g_{\mathrm{plug}}(X, [S'_1, \ldots, S'_l]) = X'\nonumber \end{gather} and $S'_i = [\mathbf{x}^{S'_i}_1, \mathbf{x}^{S'_i}_2, \ldots, \mathbf{x}^{S'_i}_j]$ is an adversarial sentence from the ${permute}$ step such that $S'_i \neq S_i$ if one or more of its words have been substituted with synonyms. In the same example above, the word `\textit{hallway}' in Sentence 2: `\textit{Mary journeyed to the hallway.}' can be substituted with `\textit{office}' under the $g_{\mathrm{permute}}$ operator to generate Sentence 3: `\textit{Mary journeyed to the office.}', before injecting into the story. For a story containing Sentence 1, if the correct answer to the question: `\textit{How do you go from the office to the hallway?}' is `\textit{north}', adding Sentence 3 to the story should also not change the correct answer. Other words can also be selected to be permuted such as substituting the name word '\textit{Mary}' in Sentence 3 with '\textit{John}'. The added flexibility allows for more control of the target DNC's predictions and behaviors with wider range of possible changes in the input sequence. In practice, the $permute$ step could be executed by greedily permuting synonyms over the Pick-Permute-Plug process iteratively with respect to the DNC's output confidence, in order to induce prediction of a target output with high confidence. \section{Experiments}\label{sec:exp} \subsection{DNC Robustness against Adversarial Framework} \subsubsection{Experiment Setup} We conduct in-depth evaluation on the robustness of DNC using our adversarial framework. The DNC was jointly trained on en-10k data subset of all 20 bAbI tasks and obtained near-perfect performance in all the tasks. We conduct the following evaluations on the DNC model that was pretrained on all tasks and publicly released in \cite{rsDNC}. In our adversarial attacks, we select 4 representative tasks (\#3, \#15, \#16 and \#19) of bAbI tasks to form 12 target-source task pairs and evaluate the DNC’s robustness on them. \begin{table}[!htbp] \centering \caption{Word error rate (\%) of DNC on task \# 3, 15, 16 and 19 under Plug-n-Pick attacks. Test error rate without adversarial sentences for task \# 3, 15, 16 and 19 are 1.6\%, 0\% , 0\% and 0\%, respectively. `Full Block' adversarial injection means that a complete sequence of sentences from a sampled source task story is injected as adversarial sentences.} \begin{tabular}{cc|cccc|c} \toprule \ \multirow{2}{*}{Source} & \multirow{2}{*}{Position} & \multicolumn{4}{c|}{\# of Adversarial Sentences} & \multirow{2}{*}{Full Block} \\ \cline{3-6} & & 1 & 2 & 3 & 4 &\\\hline \multicolumn{7}{c}{\emph{Under Adversarial Attacks on Task 3}} \\ \multirow{2}{*}{\#15} & before\_story & 3.6 & 2.8 & 3.3 & 4.8 & 21.0\\ & before\_question & 2.7 & 2.9 & 3.6 & 5.9 & 22.4\\ \multirow{2}{*}{\#16} & before\_story & 4.6 & 3.9 & 4.9 & 8.1 & 23.9\\ & before\_question & 2.5 & 2.6 & 4.4 & 6.3 & 26\\ \multirow{2}{*}{\#19} & before\_story & 4.5 & 6.7 & 8.3 & 9.5 & 12.0\\ & before\_question & \textbf{6.4} & \textbf{10.5} & \textbf{15.2} & \textbf{20.4} & \textbf{26.5}\\ \bottomrule \multicolumn{7}{c}{\emph{Under Adversarial Attacks on Task 15}} \\ \multirow{2}{*}{\#3} & before\_story & 0.6 & 5 & 9.6 & 13.6 & 88.4\\ & before\_question & 0.7 & 3.9 & 8.4 & 13.9 & \textbf{98.5}\\ \multirow{2}{*}{\#16} & before\_story & 1.9 & 6.6 & 13.3 & 24.2 & 56.4\\ & before\_question & 1.4 & 5.4 & 11.9 & 23.1 & 60.3\\ \multirow{2}{*}{\#19} & before\_story & \textbf{4.4} & \textbf{22.4} & \textbf{40.9} & 53.7 & 64.2\\ & before\_question & 3.3 & 17.8 & 37.5 & \textbf{55} & 65.8\\ \bottomrule \multicolumn{7}{c}{\emph{Under Adversarial Attacks on Task 16}} \\ \multirow{2}{*}{\#3} & before\_story & 0.4 & 0.9 & 0.7 & 0.7 & 24.5\\ & before\_question & 0.1 & 1.1 & 1.3 & 1.3 & \textbf{94.2}\\ \multirow{2}{*}{\#15} & before\_story & \textbf{2.8} & 2 & 3.9 & \textbf{7.7} & 27.4\\ & before\_question & 0.2 & 1 & 1.8 & 2.5 & 12.9\\ \multirow{2}{*}{\#19} & before\_story & 1.3 & 1.9 & 2.6 & 4.3 & 5.4\\ & before\_question & 0.6 & \textbf{2.5} & \textbf{4.1} & 6.9 & 10.8\\ \bottomrule \multicolumn{7}{c}{\emph{Under Adversarial Attacks on Task 19}} \\ \multirow{2}{*}{\#3} & before\_story & 0.2 & 0.3 & 0.5 & 0.85 & 10.2\\ & before\_question & 0.3 & 0.5 & 1.05 & 1.75 & \textbf{51.4}\\ \multirow{2}{*}{\#15} & before\_story & 0.55 & 1.2 & 4 & 7.15 & 28.6\\ & before\_question & \textbf{0.9} & \textbf{2.95} & \textbf{8.55} & \textbf{14.8} & 37.0\\ \multirow{2}{*}{\#16} & before\_story & 0.4 & 2.4 & 4.45 & 6.75 & 19.1\\ & before\_question & 0.6 & 2.85 & 5.5 & 9.1 & 36.1\\ \bottomrule \end{tabular} \label{tab:t3_t15_t16_t19_adv} \end{table} The pick step is implemented as a random search for non-question sentences. To investigate the upper limit of the Pick-n-Plug adversarial effect, we also consider a scenario where bodies of text from the source task are picked from a story's start to the right before its first question, with no limit on the number of sentences. \subsubsection{Results \& Discussion} \subsubsection{Injection Position} For both target tasks and all source tasks, when the number of adversarial sentences are larger, the adversarial sentences inserted right before questions have created stronger attack than sentences inserted at the start of the story, shown in Table~\ref{tab:t3_t15_t16_t19_adv}. For some cases where the number of adversarial sentences are smaller, the adversarial effect at the position before the story is larger. Intuitively, the effect of adversarial sentences inserted at the beginning of the story can be thought as to kick-start the DNC in the wrong direction to focus on details of the story that are not relevant in correctly answering the question. In contrast, the adversarial sentences right before the question might cause the DNC to erase data in its memory that is important to correctly answer the question at the end, as a price of storing data from the relatively more recent adversarial sentences. This implies that, as the length of adversarial sentences increases, the effect of the adversarial sentences in overwriting relevant data outweighs the effect of starting off the DNC's attention to the less relevant direction. \subsubsection{Source Task} Adversarial sentences from source task \#19 generally degrade the DNC's performance more than adversarial sentences from other source task (\#3, 15 and 16). A possible explanation may be that the distracting strength of adversarial sentences lies in the amount of information they carry. In task \#19, directional relationships between two location are expressed in each sentence. This translates to change to two entities’ attribute per sentence. For example, a sentence from task \#19 like ``\textit{The office is east of the garden.}'' encodes that for the `\textit{office}' entity's attribute that is east of `\textit{garden}’ and the `garden' entity's attribute that it is west of `\textit{office}'. In task \#3, \#15 and \#16, only one attribute of a single object such as its location (\#3) or its identity (\#15) is altered per sentence. This means that sentences from task \#19 carry almost twice the amount of information about entities than sentences from \#3, \#15 and \#16, explaining the stronger adversarial effect. It implies that adversarial sentences, within a word limit, generally have a more potent effect if the amount of information encoded is maximized in these sentences. \subsubsection{Number of Adversarial Sentences} For target tasks and all source tasks, adversarial blocks with more sentences generally contribute to more powerful adversarial attack in causing the DNC to predict incorrect answers. It can be interpreted intuitively that the adversarial effect of DNC storing irrelevant data and overwriting important information gets more pronounced with increasing amount of adversarial information. In the case where the number of sentences is limited to 4, the DNC's performance in task \#15 degrade from $0$\% to $55$\% when attacked with task \#19 adversarial sentences injected right before the question. In the case where blocks of contiguous sentences from source task stories are injected without sentence limit, the DNC degrades from a perfect test performance of $0$\% error to $98.5$\% in target task \#15. \subsection{Role of DNC Memory Module in Robustness} The memory module in DNC contributes its state-of-the-art performance in bAbI tasks, but is still unknown what its role is in resisting adversarial attacks. Since one possible effect of adversarial sentences is to overwrite relevant stored information, a bigger size of the DNC's memory module might mitigate this effect by having more free space to write new information rather than overwriting these important data. In the DNC architecture, its controller is meant to work with memory module of different sizes as long as the dimensionality of each memory row is the same. Since the learned parameters of DNC's controller are compatible with any number of rows the memory module can have, we use the same DNC's controller for all memory sizes. We carry out adversarial attacks with Pick-n-Plug on DNC of memory size 0.5x, 0.75x, 1x, 2x, 4x, 8x, 16x and 32x of the original 192 memory rows. In one of the attacks, blocks of 4 adversarial sentences are injected before the question of the target task. In a stronger attack, one separate block of 4 adversarial sentences is injected before the question and another before the start of the target story. We carry out these attacks on target task \#3 using adversarial source task \#19 and vice versa. The performance of DNC with these memory sizes are also evaluated on clean test dataset. \subsubsection{Results \& Discussion} Our evaluation results show that memory size plays a limited role in DNC's resistance against adversarial attacks from Pick-n-Plug, as shown in Figure~\ref{fig:memory}. The error rates of DNC drip to a minimum ($9$\% to $24$\% reduction in error from the original memory size) in our experiments when memory size is either 2x or 1.5x of its original. However, as the memory size increases further, the adversarial error rates increase even past the original error rate. This implies that there is still a significant robustness gap that larger DNC's memory size cannot close. The error rate of the DNC stands at $15.6$\% at its maximum robustness when task \#3 is attacked by block of 4 adversarial sentences injected before the question. Its performance on clean test samples is also degraded, also shown in Figure~\ref{fig:memory}. This seems to occur earlier than the degradation of performance under adversarial attacks. An explanation for this might be that the DNC controller was not trained to optimally handle memory modules of sizes too far from its original size, even though they are compatible. This degradation due to unfamiliarity of DNC's controller with increasing memory size may outweigh the robustness benefit that larger memory size can offer, resulting in a maximum robust performance point close to the original memory size, 1.5 to 2x in our experiments. \begin{figure} \centering \includegraphics[width=\columnwidth]{image/memory} \caption{(L) Word error rate (\%) of DNC with different sizes of augmented memory module on target task \#3 with adversarial source task \#19, and (R) on target task \#19 with adversarial source task \#3.} \label{fig:memory} \vspace{-4mm} \end{figure} \subsection{Adversarial Effect on DNC Controller Activities} We investigate the behaviors of DNC controller under different inputs, 1) clean input (CE), 2) unsuccessful adversarial attack (UAA), and 3) successful adversarial attack (SAA), by probing its control signals. The DNC controller outputs control signals to determine how, where and what content that is read, written and erased from the DNC memory module. These control signals comprise 3 gates, 2 keys and 2 vectors. Gate values are scalar that range from 0 to 1, while keys and vectors are W-dimensional vectors with real values. The free gate value determines how much of recently accessed information to forget. Allocation gate shows how much the location of written content is influenced by availability of space in that location rather than the relevance with the information already stored in that location. The write gate determines how much to write into memory at a particular time step. The write key and read key guide to memory locations with highest similarity to write new data and read stored information, respectively. Write vector and erase vector are the two vector values which determine what content to write into and erase from the memory, respectively. We can probe how DNC controller's behaviors change by comparing how the control signals deviate under one input sequence to another. In our experiments, we compare the signals under different inputs with pairwise similarity metrics in such fashion: clean example with unsuccessful adversarial attack (CE-UAA), clean example with successful adversarial attack (CE-SAA), and unsuccessful adversarial attack with successful adversarial attack (UAA-SAA). The UAA is an example where the DNC's input sequence contains injected adversarial sentences and it still predicts the correct answer to the question. In SAA, the DNC predicts the incorrect answer. For similarity comparison of 2 sequences of scalar values, like the gate values, we can use normalized KL-divergence. We can use cosine similarity to compare vector-based control signals like the write/read keys and write/erase vectors. \begin{figure*} \centering \includegraphics[width=0.65\columnwidth]{image/DNC_free_gates.pdf} \includegraphics[width=0.65\columnwidth]{image/DNC_alloc_gates.pdf} \includegraphics[width=0.65\columnwidth]{image/DNC_write_gates.pdf} \caption{(Top-Down) The free gate, allocation gate, and write gate values of DNC's read heads when the input sequence is (Row 1) an unsuccessful adversarial input, (Row 2) a clean example, and (Row 3) a successful adversarial input. } \label{fig:combo} \end{figure*} Every DNC input sequence in our experiments contain a story part and a question part. For an adversarial input sequences, we can further differentiate the story part into its clean and adversarial segments, to have a more fine grained picture of the control signals. We compared the DNC control signals under these separate parts of input sequence. Here, we describe how we sampled a CE, UAA and SAA that are closely related for a meaningful comparison. First, an UAA is sampled from a Pick-n-Plug run on target task \#19 that picks 4 adversarial sentences from source task \#3, and inject them right before every question. To generate closely related versions of this UAA sample, we conduct a run of Pick-Permute-Plug where 4 location words in this UAA's adversarial sentences are permuted with a set of 8 synonyms in a brute force manner, while the rest of the adversarial sentences remains the same. Among these new adversarial examples which successfully caused the DNC to predict incorrect answer, a SAA was sampled. The CE equivalent of the UAA and SAA is one with all 4 adversarial sentences removed. All words in the story and question segments of CE, UAA and SAA are the same. In the adversarial segments of UAA and SAA, up to 4 of the location words can differ while the other words and length of the word sequences are the same. \subsubsection{Results \& Discussion} For all pair-wise comparisons and for all of DNC's controller keys and vectors~(see Table~\ref{tab:cos_sim_combo}), the mean cosine similarities at the story segment are significantly higher than the mean cosine similarities at the question segment. This indicates that the main adversarial effect in disrupting the DNC's controller keys and vectors emerge mainly at sections after the adversarial sentences are injected rather than before that, despite DNC controller's bidirectional architecture. When compared with CE, the cosine similarities of controller keys and vectors from SAA is lower than that from UAA. This suggests that the keys and vectors from a clean example are disrupted more under a successful attack than an unsuccessful one. When comparing the DNC's controller keys and vectors under UAA and SAA, the cosine similarities between all of them - the write keys, read keys, write vectors and erase vectors - are lowest in the adversarial segment. Upon closer look into the cosine similarities of these keys and vectors in the adversarial segment, we can see from Figure \ref{fig:cos_sim_UAA_SAA} that there are some sharp dips in the similarities of the keys and vectors. These dips are at the time-step where the input words are different. The four dips correspond to [`\textit{hallway}', `\textit{bathroom}', `\textit{hallway}', `\textit{park}'] in the UAA and [`\textit{kitchen}', `\textit{hallway}', `\textit{bedroom}', `\textit{garden}'] in the SAA. Looking at the cosine similarities at question segment, there is a sharp dip at the end where the answer is expected from the DNC, while the similarities are relatively stable at previous time steps. This suggests that the adversarial sentences has a latent effect on the keys and vectors which emerges in critical segments such as when information is retrieved to answer a question. This latent effect is very likely to be strengthened by the DNC memory module since it excels in storage of information. These observations indicate that disruptions to these keys and vectors, which are involved in DNC's write, read and erase operations, play a part of the overall adversarial effect from a successful attack. For gate values, we find no obvious difference between the patterns of the gate values when DNC is presented with a CE, UAA and SAA from a general view (see Figure~\ref{fig:combo}). \begin{figure} \centering \includegraphics[width=\columnwidth]{image/UAA_SAA_cos_sim_full_colored_fx2.pdf} \caption{Cosine similarity of DNC's keys and vectors, between when the DNC is presented with a unsuccessful adversarial attack and when it is presented with a successful adversarial attack. (Top left) The cosine similarity values at the story, (Top right) adversarial portion, and (Bottom) question of a sample QA from task \#19.} \label{fig:cos_sim_UAA_SAA} \end{figure} When the KL-divergence is used to compare the gate values, significant patterns appear at the different segments of the input sequences. For all 3 type of gate values at the story segment~(see Table~\ref{tab:kl_combo}), the KL-divergence of all pair-wise comparisons are significantly lower than the KL-divergence at the question segment, with 2 to 3 order of magnitude difference. This implies that the main adversarial effect on DNC's 3 gate values emerge after the injection of adversarial sentences, rather than before that. \begin{table}[!htbp] \centering \caption{Mean cosine similarity between DNC's keys and vectors as a comparison how different these control signals are with different inputs. The input can be a clean example (CE), unsuccessful adversarial attack (UAA) or successful advers. attack (SAA).} \begin{tabular}{clccc} \toprule ~ & ~ & Story & Adversary & Question \\ \cline{1-5} \multirow{3}{*}{Write Keys} & CE-UAA & 0.9969 & - & 0.9790\\ ~ & CE-SAA & 0.9962 & - & \textbf{0.9627}\\ \cline{2-5} ~ & UAA-SAA & 0.9997 & \textbf{0.8593} & 0.9689\\ \bottomrule \multirow{3}{*}{Read Keys} & CE-UAA & 0.9981 & - & 0.9749\\ ~ & CE-SAA & 0.9980 & - & \textbf{0.9652}\\ \cline{2-5} ~ & UAA-SAA & 0.9999 & \textbf{0.8417} & 0.9737\\ \bottomrule \multirow{3}{*}{Write Vectors} & CE-UAA & 0.9988 & - & 0.9799\\ ~ & CE-SAA & 0.9988 & - & \textbf{0.9653}\\ \cline{2-5} ~ & UAA-SAA & 0.9999 & \textbf{0.7400} & 0.9761\\ \bottomrule \multirow{3}{*}{Erase Vectors} & CE-UAA & 0.9996 & - & 0.9975\\ ~ & CE-SAA & 0.9995 & - & \textbf{0.9942}\\ \cline{2-5} ~ & UAA-SAA & 1.000 & \textbf{0.9660} & 0.9957\\ \bottomrule \end{tabular} \label{tab:cos_sim_combo} \end{table} \begin{table}[!htbp] \centering \caption{KL-divergence of DNC controller's gate values to compare how they changes with different inputs.The input can be a clean example (CE), unsuccessful adversarial attack (UAA) or successful adversarial attack (SAA). Free gate values from only 1 of the 4 DNC's read heads are compared here. The KL-divergences of free gate values from the other 3 read heads show similar spread of maximum values.} \begin{tabular}{clccc} \toprule ~ & ~ & Story & Adversary & Question \\\hline \multirow{6}{*}{Free Gate} & CE-UAA & 3.735E-05 & - & 0.002472\\ ~ & UAA-CE & 3.710E-05 & - & 0.002515\\ ~ & CE-SAA & 2.526E-05 & - & \textbf{0.02041}\\ ~ & SAA-CE & 2.529E-05 & - & 0.01566\\ \cline{2-5} ~ & UAA-SAA & 2.212E-06 & 0.005523 & \textbf{0.02470}\\ ~ & SAA-UAA & 2.234E-06 & 0.006529 & 0.01791\\ \bottomrule \multirow{6}{*}{Alloc Gate} & CE-UAA & 0.0003625 & - & 0.001112\\ ~ & UAA-CE & 0.0003883 & - & 0.001184\\ ~ & CE-SAA & 0.0003089 & - & \textbf{0.01512}\\ ~ & SAA-CE & 0.0003319 & - & 0.01148\\ \cline{2-5} ~ & UAA-SAA & 3.730E-06 & \textbf{0.2279} & 0.02437\\ ~ & SAA-UAA & 3.760E-06 & 0.1596 & 0.01682\\ \bottomrule \multirow{6}{*}{Write Gate} & CE-UAA & 0.0003325 & - & 0.005531\\ ~ & UAA-CE & 0.0003171 & - & 0.005586\\ ~ & CE-SAA & 0.0002664 & - & \textbf{0.007648}\\ ~ & SAA-CE & 0.0002548 & - & 0.007347\\ \cline{2-5} ~ & UAA-SAA & 1.369E-05 & \textbf{0.2408} & 0.001681\\ ~ & SAA-UAA & 1.389E-05 & 0.1101 & 0.001621\\ \bottomrule \end{tabular} \label{tab:kl_combo} \end{table} At the question segment, the KL-divergence of DNC's gate values from CE under SAA is higher than that under UAA for all the 3 gate types. This suggests that the gate values deviate more from a clean input under a successful attack than an unsuccessful one. These two observations on the gate values, in the story and question segments, resonates with the above observations on the controller keys and vectors in these word segments. When comparing the DNC's gate values under UAA and SAA, the KL-divergence of the allocation and write gate values in the adversarial segment is the largest. This might be due to the presence of 4 different location words in the adversarial sentences of UAA and SAA. In contrast, the KL-divergence of free gate values is larger in the question segment than in the story or adversarial segment, despite the fact that the words in UAA's and SAA's question segment are the same. This implies that the adversarial sentences have different effects on the gate control signals. These observations suggest that disruptions to these gates, which are responsible for the DNC's erase and write operations, play a part of the overall adversarial effect from a successful attack. \section{Conclusions}\label{sec:concl} In this paper, we propose a metamorphic relation based adversarial attack framework for DNCs, and show its attacks can cause multi-faceted disruptive effects on the read, write and erase functions of the DNC. Our in-depth evaluation on bAbI logical question answering task demonstrates that the current DNCs still face robustness issues, despite with larger memory size. We hope our framework motivates more work about adversarial attacks in NLP and more extensive studies on DNCs, towards constructing robust, automatically programmed, general purpose learning machines. \bibliographystyle{IEEEtran}

acba908e1126394863ea9dae77cf5a904e6f95be

\section{Introduction} The identification of governing equations from data plays an increasingly important role in the field of nonlinear dynamical systems. After all, in many practically relevant situations, the actual underlying model that captures a dynamical system is unknown, while the only information available is data arising from the natural time evolution. In this instance of system identification, given time-dependent measurement data, the task at hand is to reconstruct the underlying system, merely assuming that the governing equations can be accurately approximated by linear combinations of preselected basis functions. We consider autonomous dynamical systems given by ordinary differential equations (ODEs). Many physical --~even chaotic~-- systems have a simple ODE structure comprising only a small number of terms, e.g., monomials or trigonometric functions and combinations thereof. That is, if we choose a large set of basis functions, then a system may be described in terms of only a few of these basis functions, or -- in other words -- the governing equations are sparse in the high-dimensional space of selected basis functions, cf.~\cite{SPROTT1994, SOCOLAR2006, TESCHL2012}. Essentially, we are interested in the relationship between two data sets, namely measurements of $X(t)$ and measurements of $\dot{X}(t)$ at certain time points $t$. The methods we will discuss below (as well as the approach we propose), however, can be applied to a wider range of problems. For instance, we could also consider discrete dynamical systems or stochastic differential equations (SDEs). The so-called SINDy approach \cite{BRUNTON2016} has also already been extended to identify the governing equations of SDEs, see~\cite{BONINSEGNA2018}. The challenge here is to recover the dynamics (i.e., to recover $F$ such that $\dot{X}(t) = F(X(t))$) from a rather small number of measurements of the time-dependent states $X(t)$ and the corresponding derivatives. This can be accomplished by using different approaches, see, for example, \cite{SJOBERG1995}. One of them is \emph{symbolic regression} \cite{KOZA1994}, where evolutionary algorithms reveal appropriate models that fit the given measurement data by combining mathematical expressions. Unfortunately, symbolic regression is in general too expensive and the convergence might be slow for high-dimensional data sets because discovering the governing equations from measurement data in a high-dimensional space $\mathbb{R}^d$ can be prohibitively expensive in terms of memory consumption and computational costs. This phenomenon is often referred to as the \emph{``curse of dimensionality''}. Also neural networks can be used to approximate the governing equations of dynamical systems (and, in fact, to approximate any multidimensional function), see~\cite{MUZHOU2011}. As for the symbolic regression, knowledge about the governing equations is not needed a priori for these purely data-driven methods. However, the approximation based on neural networks does not necessarily lead to interpretable representations of the governing equations since the given data is only interpolated by using artificial activation functions, e.g., radial basis functions. In~\cite{BRUNTON2016}, Brunton et al.\ proposed an algorithm for the \emph{sparse identification of nonlinear dynamical systems} (SINDy). Given data measurements of states and corresponding derivatives, a certain set of basis functions has to be chosen a priori. Based on sequential least-squares approximations and an iterative selection of relevant basis functions -- which is comparable to \emph{iterative hard thresholding} \cite{BLUMENSATH2012} -- the application of SINDy often results in a sparse representation of the underlying system, provided that we choose suitable basis functions. We could also add an $L^1$-regularization term to the regression problem if sparsity of the solution should be enforced. In the community of compressed sensing, this method is known as the \emph{least absolute shrinkage and selection operator} (LASSO), see~\cite{TIBSHIRANI1996}. However, for large data sets (particularly for large dimensions $d$) the above methods still suffer from the curse of dimensionality. Thus, there is a need for efficient methods that are able to recover the governing equations of high-dimensional nonlinear dynamical systems. In this work, we combine the data-driven discovery of the underlying dynamics with the computation of pseudoinverses of tensor-network decompositions. Tensors are generalizations of vectors and matrices represented by multidimensional arrays with multiple indices. The interest in low-rank tensor decompositions has been growing rapidly over the last years since several tensor formats such as the canonical format \cite{CARROLL1970, HARSHMAN1970}, the Tucker format \cite{TUCKER1963,TUCKER1964}, and the hierarchical Tucker format \cite{HACKBUSCH2009, ARNOLD2013} have shown that it is possible to mitigate the curse of dimensionality for many high-dimensional problems. Tensor-based methods have been successfully applied to many different application areas, e.g., quantum physics \cite{WHITE1992, MEYER2009,Orus-AnnPhys-2014, VerstraeteBig, EisertTensors}, chemical reaction dynamics \cite{JAHNKE2008, GELSS2016}, machine learning \cite{BEYLKIN2009,NOVIKOV2015}, and high-dimensional data analysis \cite{KLUS2016, KLUS2018}. It turned out that one of the most promising tensor formats is the so-called tensor-train format (TT format) \cite{OSELEDETS2009a, OSELEDETS2009b, OSELEDETS2011} or, equivalently, the matrix product state (MPS) format \cite{MPSSurvey,MPSRev}. It is a special case of the hierarchical Tucker format and combines the advantages of the canonical format and the Tucker format, i.e., the storage consumption of a tensor train does not depend exponentially on the number of dimensions and there exist robust algorithms for the computation of best approximations. The fact that the TT format has emerged independently in several fields of science signifies its importance. In quantum physics, the MPS representation, on the one hand, dates back to work on the density-matrix renormalization group \cite{WHITE1992}. On the other hand, pure finitely correlated states \cite{raey} are again basically quantum states in the TT format, originating in mathematical physics in efforts to generalize notions of hidden Markov models. Within numerical mathematics, the use of tensor network decompositions and approximations is rather new, which has led to an exciting development. However, many questions are still open, specifically when it comes to the convergence when approximating solutions of systems of linear equations given in the TT format, cf.~\cite{HOLTZ2012}. For an overview of different low-rank tensor approximation approaches, we refer to~\cite{GRASEDYCK2013,GELSS2017b,Orus-AnnPhys-2014}. The new role of tensor decompositions for data analysis techniques and especially compressed sensing was already discussed in, e.g.,~\cite{KLUS2018, RAUHUT2015, CICHOCKI2015}. Using the TT format we shift the focus from \emph{sparse} structures to \emph{low-rank} or \emph{data-sparse structures}. That is, instead of determining sparse vectors or matrices, we aim at recovering dynamical systems by utilizing tensor decompositions with a moderate number of core elements. Applying the approach proposed in~\cite{KLUS2018} to construct pseudoinverses in the TT format, we can extend the method to identify governing equations so that low-rank representations of the data can be utilized to reduce the computational complexity as well as the memory requirements. Referring to SINDy, we call our method MANDy, which is short for \emph{multidimensional approximation of nonlinear dynamical systems}. Data-driven discovery of dynamical systems will continue to play an increasingly important role. Here, we will show that rewriting mathematical methods by exploiting low-rank tensor decompositions enables the reconstruction of high-dimensional systems which cannot be analyzed by conventional methods. The main contributions of this paper are: \begin{itemize}[leftmargin=*,itemsep=0ex] \item combination of data-driven methods with tensor decomposition techniques for the recovery of governing equations, reducing computational costs as well as storage consumption, \item extension of the pseudoinverse computation in the TT format (developed for \emph{tensor-based dynamic mode decomposition} in~\cite{KLUS2018}), \item derivation of different basis decompositions in the TT format, \item demonstration on realistic problems from physics and engineering. \end{itemize} This work is organized as follows: In Section \ref{sec: notation}, we introduce SINDy and give a brief overview of the TT format and pseudoinverses in the TT format. In Section \ref{sec: tensor-based SINDy}, we derive the tensor-based reformulation of SINDy. Numerical results are presented in Section \ref{sec: results}. Section \ref{sec: conclusion} concludes with a brief summary and a future outlook. \section{Notation and preliminaries}\label{sec: notation} In this section, we will introduce an approach called SINDy to identify ordinary differential equations from data, the tensor-train format, and an algorithm to compute pseudoinverses in the tensor-train format. By combining these techniques, we are able to discover the governing equations of high-dimensional dynamical systems. \subsection{Sparse identification of nonlinear dynamics} \label{sec: recovery} We will consider autonomous ordinary differential equations of the form $\dot{X}(t) = F(X(t))$, where $X(t) \in \mathbb{R}^d$ is the state of the system at time $t$ and $F \colon \mathbb{R}^d \rightarrow \mathbb{R}^d$ is a function. Assuming we have $ m $ measurements of the state of the system, given by $ X_k $, $ k = 1,\dots, m $, and the corresponding time derivatives, given by $ Y_k = \dot{X}_k $, the goal is to reconstruct the function $ F $. To this end, we choose a set of basis functions (also called dictionary) $ \mathbb{D} = \{\psi_1, \psi_2, \dots, \psi_p\} $, with $ \psi_j \colon \mathbb{R}^d \rightarrow \mathbb{R} $, and define the vector-valued function $ \Psi \colon \R^d \to \R^p $ by \begin{equation*} \Psi(X) = \begin{bmatrix} \psi_1(X) & \psi_2(X) & \dots & \psi_p(X) \end{bmatrix}^T. \end{equation*} The transformed data matrix $ \Psi(\mathcal{X}) $ is then defined by \begin{equation}\label{eq: basis matrix} \Psi(\mathcal{X}) = \begin{bmatrix} \Psi(X_1) & \dots & \Psi(X_m) \end{bmatrix} \in \R^{p \times m}. \end{equation} We now wish to determine the coefficient matrix \begin{equation*} \Xi = \begin{bmatrix} \xi_1 & \xi_2 & \dots & \xi_d \end{bmatrix} \in \R^{p \times d} \end{equation*} such that the cost function $ \big\lVert \mathcal{Y} - \Xi^T \Psi(\mathcal{X}) \big\rVert_F $ is minimized. Each column vector $ \xi_i $ of $ \Xi $ then corresponds to one function $ F_i $ via \begin{equation*} \left(Y_k\right)_i = F_i(X_k) = \xi_i^T \Psi(X_k), \end{equation*} where the entries of $ \xi_i $ determine which basis functions are used for the reconstruction. Thus, we obtain a model of the form $ \dot{X}(t) = \Xi^T \Psi(X(t)) $. One approach to compute an optimal coefficient matrix in the least-squares sense is the use of the pseudoinverse of $\Psi(\mathcal{X})$. That is, we determine $\Xi$ by $ \Xi^T = \mathcal{Y} \cdot (\Psi(\mathcal{X}))^+ $, where $ ^+ $ denotes the pseudoinverse. Additionally, SINDy aims at finding a sparse coefficient matrix $ \Xi $ \cite{BRUNTON2016}. This is accomplished by iteratively removing basis functions corresponding to small entries that might not be required for the reconstruction. Provided that the governing equations can be expressed in terms of the basis functions, SINDy may completely recover the dynamical system. The iterative elimination of basis functions could be omitted if the number of snapshots is large enough so that the solution of the initial least-squares problem is already close to the true solution. Alternatively, a LASSO-based (see, e.g.,~\cite{TIBSHIRANI1996}) approach can be used to identify the coefficients. If the time derivatives are not available and need to be approximated by finite differences, the resulting $ \dot{X} $ data might be noisy and necessitate denoising techniques. Also measurement data will in general contain noise and require regularization. For more details, see~\cite{BRUNTON2016} and references therein. \begin{example}[Illustration of SINDy] \label{ex:SINDy examples} To illustrate the sparse identification process, let us begin with a simple example. Consider Chua's circuit, see, e.g.,~\cite{KILIC2010}, given by \begin{equation}\label{eq: chua ODE} \begin{split} \dot{x}_1 &= \alpha (x_2 - x_1 - g(x_1)), \\ \dot{x}_2 &= x_1 - x_2 + x_3, \\ \dot{x}_3 &= -\beta x_2, \end{split} \end{equation} where $ \alpha $ and $ \beta $ are real parameters and $ g \colon \mathbb{R} \to \mathbb{R} $ is a nonlinear function. We will set $ \alpha = 10 $, $ \beta = 14.87 $, and $ g(z) = \delta_1 \, z + \delta_2 \, z \abs{z} $, with $ \delta_1 = -\frac{8}{7} $ and $ \delta_2 = \frac{4}{63} $. Let us try to recover the governing equations of Chua's circuit from data. We simulate the system for $ t = 0, \dots, 20 $ with a step size of $ h = 0.01 $ and the initial condition $X_1 = \left[ -1.13 , 0.004, 0.45\right]^T$, thus $ \mathcal{X}, \mathcal{Y} \in \mathbb{R}^{3 \times 2000} $. Here, we use the exact derivatives. \begin{enumerate}[label=(\roman*), leftmargin=0.0em, itemindent=1.5em] \item Using a basis comprising monomials of order up to and including two in each dimension, i.e., \begin{equation}\label{chua: first try} \Psi_1(X) = \begin{bmatrix} \; 1 & x_1 & x_1^2 & x_2 & x_1 x_2 & x_1^2 x_2 & \dots & x_2^2 x_3^2 & x_1 x_2^2 x_3^2 & x_1^2 x_2^2 x_3^2 \; \end{bmatrix}^T, \end{equation} we obtain the following coefficient matrix \begin{equation*} \Xi_1^T = \hspace*{-0.7em} \kbordermatrix{ & 1 & x_1 & x_1^2 & x_2 & x_1 x_2 & x_1^2 x_2 & x_2^2 & x_1 x_2^2 & x_1^2 x_2^2 & x_3 & \dots \\ & -0.06 & 1.10 & 0.35 & 9.61 & -1.77 & -1.97 & 1.05 & 3.85 & 1.07 & -0.07 & \dots \\ & 0 & 1 & 0 & -1 & 0 & 0 & 0 & 0 & 0 & 1 & \dots \\ & 0 & 0 & 0 & -14.87 & 0 & 0 & 0 & 0 & 0 & 0 & \dots \\ }. \end{equation*} The coefficients for the second and third function are identified correctly, whereas the first equation cannot be represented by the basis functions and all coefficients are unequal to zero, approximating the missing basis function. The resulting trajectories are shown in Figure~\ref{fig:Chua SINDy} (a). \item If we use a different set of basis functions, given by \begin{equation}\label{chua: second try} \Psi_2(X) = \begin{bmatrix} \; 1 & x_1 & x_2 & x_3 & \abs{x_1} & x_1 \abs{x_1} & x_2 \abs{x_1} & x_3 \abs{x_1} & \dots & x_3 \abs{x_3} & x_3 \abs{x_3} \; \end{bmatrix}^T \end{equation} the system is recovered correctly as \begin{equation*} \Xi_2^T = \hspace*{-0.7em} \kbordermatrix{ & 1 & x_1 & x_2 & x_3 & \abs{x_1} & x_1 \abs{x_1} \\ & 0 & 1.42 & 10 & 0 & 0 & -0.6349 & \dots \\ & 0 & 1 & -1 & 1 & 0 & 0 & \dots \\ & 0 & 0 & -14.87 & 0 & 0 & 0 & \dots }. \end{equation*} All remaining coefficients are numerically zero. The simulation results are shown in Figure~\ref{fig:Chua SINDy}~(b). \xqed{$\triangle$} \end{enumerate} \begin{figure}[htb] \centering \begin{minipage}{0.49\textwidth} \centering \subfiguretitle{(a)} \includegraphics[width=\textwidth]{pictures/example_chua_1} \end{minipage} \begin{minipage}{0.49\textwidth} \centering \subfiguretitle{(b)} \includegraphics[width=\textwidth]{pictures/example_chua_2} \end{minipage} \caption{Simulation of Chua's circuit and the identified systems for $ t = 0, \dots, 100 $ with a step size of $ h = 0.01 $. (a) Using a basis comprising only monomials leads to incorrect simulation results, the identified system does not capture the full dynamics. (b) Constructing a basis that contains also $ x_1 \abs{x_1} $ leads to correct results, the original and identified system are identical up to numerical errors.} \label{fig:Chua SINDy} \end{figure} \end{example} \begin{remark}[Dependence of SINDy on basis functions] The example illustrates that in order to obtain an accurate representation of the system, a priori knowledge about the governing equations might be required, which is in general not available when the method is applied to measurement data or data stemming from black-box models. If the basis functions are not chosen appropriately, then SINDy will not faithfully reproduce the dynamics of the system. \end{remark} \subsection{The tensor-train format} Tensors of order $ d $ are multidimensional arrays $\mathbf{T} \in \mathbb{R}^{n_1 \times \dots \times n_d}$. Figure~\ref{fig: tensors} shows a number of simple examples. The different dimensions $n_i \in \mathbb{N}$ for $i=1, \dots , d$ are called \emph{modes}. Elements of tensors are accessed by $\mathbf{T}_{j_1, \dots, j_d} \in \mathbb{R}$, with $1 \leq j_i \leq n_i$. If we fix certain indices, colons are used to indicate the free modes (similar to Matlab's or Python's colon notation). The storage consumption of a tensor can be estimated as $O(n^d)$, where $n$ is the maximum of all mode sizes $ n_1, \dots, n_d $. Storing a $d$-dimensional tensor may be infeasible for large $d$ since the number of elements of a tensor grows exponentially with the order. This is typically called the \emph{curse of dimensionality}. However, it is possible to mitigate this by exploiting low-rank tensor approximations. If the underlying correlation structure admits such a decomposition, an enormous reduction in complexity can be achieved. The basic idea is to decompose tensors $\mathbf{T} \in \mathbb{R}^{n_1 \times \dots \times n_d}$ into different components using tensor products, see~\cite{WILSON1901}. \begin{figure}[htbp] \centering \begin{subfigure}[b]{0.2\textwidth} \centering \caption{} \begin{tikzpicture} \node (1) [rectangle] at (0,0) {$\begin{bmatrix} 1 \\ 0 \\ 1 \\ 1 \\0 \end{bmatrix}$}; \end{tikzpicture} \end{subfigure} \hfill \begin{subfigure}[b]{0.3\textwidth} \centering \caption{} \begin{tikzpicture} \node (1) [rectangle] {$\begin{bmatrix} 1 & 0 & 1 & 0\\ 0 & 1 & 1 & 0 \\ 1 & 1 & 0 & 1 \\ 1 & 0 & 1 & 0 \\ 0 & 1 & 0 & 1\end{bmatrix}$}; \end{tikzpicture} \end{subfigure} \hfill \begin{subfigure}[b]{0.4\textwidth} \centering \caption{} \begin{tikzpicture}[ myrect/.style={ rectangle, draw, inner sep=0pt, fit=#1} ] \def0.38{0.38} \def0.475{0.475} \node (1) [draw=black, rectangle] at (0.38+0.38,0.475+0.475) {$\begin{bmatrix} 1 & 0 & 1 & 1 \\ 1 & 0 & 0 & 0 \\ 0 & 1 & 1 & 0\end{bmatrix}$}; \node (3) [rectangle] at (0,0) {\phantom{\phantom{$\begin{bmatrix} 1 & 1 & 0 & 1 \\ 0 & 1 & 1 & 0 \\ 1 & 0 & 0 & 1\end{bmatrix}$}}}; \draw[-,dotted] (1.south west)--(3.south west); \node (2) [fill=white, opacity=0.9, rectangle] at (0.38,0.475) {\phantom{$\begin{bmatrix} 0 & 1 & 1 & 0 \\ 1 & 0 & 0 & 1 \\ 1 & 1 & 1 & 1\end{bmatrix}$}}; \node (2-2) [draw=black, rectangle] at (0.38,0.475) {$\begin{bmatrix} 0 & 1 & 1 & 0 \\ 1 & 0 & 0 & 1 \\ 1 & 1 & 1 & 1\end{bmatrix}$}; \node (3) [fill=white, opacity=0.9, rectangle] at (0,0) {\phantom{$\begin{bmatrix} 1 & 1 & 0 & 1 \\ 0 & 1 & 1 & 0 \\ 1 & 0 & 0 & 1\end{bmatrix}$}}; \node (3-2) [draw=black, rectangle] at (0,0) {$\begin{bmatrix} 1 & 1 & 0 & 1 \\ 0 & 1 & 1 & 0 \\ 1 & 0 & 0 & 1\end{bmatrix}$}; \draw[-,dotted] (1.north east)--(3.north east); \draw[-,dotted] (1.north west)--(3.north west); \draw[-,dotted] (1.south east)--(3.south east); \end{tikzpicture} \end{subfigure} \caption[Low-dimensional tensors represented by arrays]{Low-dimensional tensors represented by arrays: (a) A tensor of order 1 is a vector. (b) A tensor of order 2 is a matrix. (c) A tensor of order 3 can be visualized as layers of matrices.} \label{fig: tensors} \end{figure} Examples of tensor formats are \emph{rank-one tensors} \cite{FRIEDLAND2013, HACKBUSCH2012} and the \emph{canonical format} \cite{HITCHCOCK1927, KOLDA2009}. A tensor $\mathbf{T} \in \R^{n_1 \times \dots \times n_d}$ of order $d$ is called a rank-one tensor if it can be written as the tensor product of $d$ vectors. A tensor in the canonical format is simply the sum of rank-one tensors. For our purposes, we will rely on the \emph{tensor-train format} (TT format) \cite{OSELEDETS2009b,OSELEDETS2011}, a special case of the \emph{hierarchical Tucker format} \cite{HACKBUSCH2009,ARNOLD2013}. A tensor in the TT format is given by \begin{equation} \mathbf{T} = \sum_{k_0=1}^{r_0} \cdots \sum_{k_d=1}^{r_d} \mathbf{T}^{(1)}_{k_0,:,k_1} \otimes \dots \otimes \mathbf{T}^{(d)}_{k_{d-1},:,k_d}. \end{equation} The TT ranks $r_0 , \dots , r_d$, where $r_0 = r_d = 1$, determine the storage consumption of a tensor train and its complexity. Lower ranks imply a lower memory consumption and lower computational costs. Therefore, our aim is to compute low-rank approximations of high-dimensional tensors in the TT format. Figure~\ref{fig: tensor trains} shows the graphical representation -- motivated by~\cite{HOLTZ2012} -- of a tensor train $\mathbf{T} \in \mathbb{R}^{n_1 \times \dots \times n_d}$. A core is depicted by a circle with different arms indicating the modes of the tensor and the rank indices. For more information, we refer to~\cite{GELSS2017b, HACKBUSCH2012}. \begin{figure}[htbp] \centering \begin{tikzpicture} \draw[black] (0,0) -- node [label={[shift={(0,-0.15)}]$r_1$}] {} ++ (1.5,0) ; \draw[black] (1.5,0) -- node [label={[shift={(0,-0.15)}]$r_2$}] {} ++ (1.5,0) ; \draw[black] (3,0) -- ++ (1,0) ; \draw[black, dotted] (4,0) -- ++ (1,0) ; \draw[black] (5,0) -- ++ (1,0) ; \draw[black] (6,0) -- node [label={[shift={(0,-0.19)}]$r_{d-1}$}] {} ++ (1.5,0) ; \draw[black] (0,0) -- node [label={[shift={(0,-1)}]$n_1$}] {} ++ (0,-0.7) ; \draw[black] (1.5,0) -- node [label={[shift={(0,-1)}]$n_2$}] {} ++ (0,-0.7) ; \draw[black] (3,0) -- node [label={[shift={(0,-1)}]$n_3$}] {} ++ (0,-0.7) ; \draw[black] (6,0) -- node [label={[shift={(0,-1)}]$n_{d-1}$}] {} ++ (0,-0.77) ; \draw[black] (7.5,0) -- node [label={[shift={(0,-1)}]$n_d$}] {} ++ (0,-0.71) ; \node[draw,shape=circle,fill=Gray, scale=0.65] at (0,0){}; \node[draw,shape=circle,fill=Gray, scale=0.65] at (1.5,0){}; \node[draw,shape=circle,fill=Gray, scale=0.65] at (3,0){}; \node[draw,shape=circle,fill=Gray, scale=0.65] at (6,0){}; \node[draw,shape=circle,fill=Gray, scale=0.65] at (7.5,0){}; \end{tikzpicture} \caption{Graphical representation of tensor trains. Here, the first and the last core are considered as matrices, the other cores are tensors of order 3.} \label{fig: tensor trains} \end{figure} We represent the TT cores as two-dimensional arrays containing vectors as elements. For $\mathbf{T} \in \mathbb{R}^{n_1 \times \dots \times n_d}$ with cores $\mathbf{T}^{(i)} \in \mathbb{R}^{r_{i-1} \times n_i \times r_i}$, a single core is written as \begin{equation} \left\llbracket \mathbf{T}^{(i)} \right\rrbracket = \left\llbracket \begin{matrix} & \mathbf{T}^{(i)}_{1,:,1} & \cdots & \mathbf{T}^{(i)}_{1,:,r_i} & \\ & & & & \\ & \vdots & \ddots & \vdots & \\ & & & & \\ & \mathbf{T}^{(i)}_{r_{i-1},:,1} & \cdots & \mathbf{T}^{(i)}_{r_{i-1},:,r_i} & \end{matrix}\right\rrbracket. \label{eq: core notation - single core} \end{equation} We then use the notation $\mathbf{T} = \left\llbracket \mathbf{T}^{(1)}\right\rrbracket \otimes \dots \otimes \left\llbracket \mathbf{T}^{(d)}\right\rrbracket$ for representing tensor trains $\mathbf{T}$, see~\cite{GELSS2017, KAZEEV2012}. This can be regarded as a generalization of the standard matrix multiplication where the cores contain vectors as elements instead of scalar values. Just like multiplying two matrices, we compute the tensor products of the corresponding elements and then sum over the columns and rows, respectively. In order to construct \emph{matricizations} and \emph{vectorizations} --~also called \emph{tensor unfoldings}~\cite{HOLTZ2012}~-- we define a bijection $\phi_N$ for the mode set $N = (n_1, \dots, n_d)^T \in \mathbb{N}^d$ with \begin{equation*} \begin{gathered} \phi_N \colon \{1, \ldots, n_1 \} \times \dots \times \{1, \ldots, n_d \} \rightarrow \{1 , \ldots , \prod_{k=1}^{d} n_k\},\\ (j_1, \dots, j_d) \mapsto \overline{j_1, \dots, j_d} := \phi_N(j_1, \dots, j_d), \end{gathered} \end{equation*} where $\overline{j_1, \dots, j_d}$ is the single-index representation of the corresponding multi-index $(j_1, \dots, j_d)$. Typically, bijections based on reverse lexicographic ordering (column-major order in Matlab and Fortran) or colexicographic ordering (row-major order in Numpy and C\texttt{++}) are used, see, e.g.,~\cite{CICHOCKI2016}. \begin{definition}[Matricization and vectorization] \label{def: matricization} Let $N = (n_1 , \ldots , n_d)^T$ be a mode set and $\mathbf{T} \in \mathbb{R}^{n_1 \times \dots \times n_d}$ a tensor. For two ordered subsets $N' = (n_1, \dots, n_l)^T$ and $N'' = (n_{l+1} , \ldots , n_d)^T$ of $N$ with $1 \leq l < d$, the \emph{matricization} of $\mathbf{T}$ with respect to $N'$ and $N''$ is given by \begin{equation} \left( \mat{\mathbf{T}}{N'}{N''} \right)_{\overline{j_1 , \ldots , j_l} , \overline{j_{l+1}, \ldots, j_{d}}} = \mathbf{T}_{j_1, \ldots, j_d}. \label{eq_matricization} \end{equation} The \emph{vectorization} of $\mathbf{T}$ is given by \begin{equation*} \mathrm{vec}(\mathbf{T}) := \left( \mat{\mathbf{T}}{N}{~} \right)_{ \overline{j_1 , \ldots , j_d} } = \mathbf{T}_{j_1 , \ldots , j_d}. \end{equation*} \end{definition} Definition \ref{def: matricization} can also be generalized to arbitrary subsets of the mode set $N$, see, e.g.,~\cite{GELSS2017b}. We will here, however, only need the special case described above. Given a TT core $\mathbf{T}^{(i)} \in \R^{r_{i-1} \times n_i \times r_i}$, we now define \begin{equation} \mathcal{L}\left( \mathbf{T}^{(i)}\right) = \mat{\mathbf{T}^{(i)}}{r_{i-1} , n_i}{r_i} \qquad \textrm{and} \qquad \mathcal{R}\left( \mathbf{T}^{(i)}\right) = \mat{\mathbf{T}^{(i)}}{r_{i-1} }{n_i, r_i}. \label{eq: left-/right-unfolding} \end{equation} The matricization $\mathcal{L} \left( \mathbf{T}^{(i)}\right)$ is called the \emph{left-unfolding} of $\mathbf{T}^{(i)}$ and $\mathcal{R} \left( \mathbf{T}^{(i)}\right)$ the \emph{right-unfolding} of $\mathbf{T}^{(i)}$, see~\cite{HOLTZ2012}. A TT core is called \emph{left-orthonormal} if its left-unfolding is orthonormal with respect to the columns, i.e., \begin{equation*} \left( \mathcal{L} \left( \mathbf{T}^{(i)}\right) \right)^T \cdot \mathcal{L} \left( \mathbf{T}^{(i)}\right) = \mat{\mathbf{T}^{(i)}}{r_i}{r_{i-1} , n_i} \cdot \mat{\mathbf{T}^{(i)}}{r_{i-1} , n_i}{r_i} = I \in \R^{r_i \times r_i}, \end{equation*} and \emph{right-orthonormal} if its right-unfolding is orthonormal with respect to the rows, i.e., \begin{equation*} \mathcal{R}\left( \mathbf{T}^{(i)}\right) \cdot \left( \mathcal{R}\left( \mathbf{T}^{(i)}\right) \right)^T = \mat{\mathbf{T}^{(i)}}{r_{i-1}}{n_i, r_i} \cdot \mat{\mathbf{T}^{(i)}}{n_i, r_i}{r_{i-1}} = I \in \R^{r_{i-1} \times r_{i-1}}. \end{equation*} Algorithms for the left- and right-orthonormalization, respectively, can be found in~\cite{OSELEDETS2011, GELSS2017b}. Note that a tensor $\mathbf{T}$ remains the same if we apply an (exact) orthonormalization procedure to it. The algorithms simply compute a different but equivalent representation. \subsection{Pseudoinverses in the TT format}\label{sec: pseudoinverse} We now explain how to compute pseudoinverses of matricizations of tensors in the TT format. Later, we aim at applying this technique to the tensorized counterpart of the matrix $\Psi(\mathcal{X})$ defined in \eqref{eq: basis matrix}. It was shown that the pseudoinverse of certain tensor unfoldings can be computed directly in the TT format. We will briefly recapitulate the main idea, details can be found in~\cite{KLUS2018}. Given a tensor $\mathbf{T} \in \mathbb{R}^{n_1 \times \dots \times n_d \times m}$ (e.g., a tensor containing $m$ snapshots of a $d$-dimensional system with dimensions $n_1, \dots , n_d$) in the TT format, we consider the matricization \begin{equation} \label{eq: matricization of T} T = \mat{\mathbf{T}}{n_1 , \dots , n_d}{m}. \end{equation} The standard way to calculate the pseudoinverse $T^+$ is based on using its singular value decomposition (SVD), i.e., $ T = U \, \Sigma \, V^T $ with $U^T U = V^T V = I $. The pseudoinverse of $T$ is then given by \begin{equation}\label{eq: pseudoinverse in full format} T^+ = V \, \Sigma^{-1} U^T. \end{equation} \begin{figure}[htbp] \centering \begin{tikzpicture} \def-3{0} \draw[black] (-1.6,0.55) -- ++ (0,-7.6) ; \node[anchor=east] at (-2.3,-3-0.3){Initial tensor train}; \draw[black] (0,-3) -- node [label={[shift={(0,-1)}]$n_1$}] {} ++ (0,-0.7) ; \draw[black] (1,-3) -- node [label={[shift={(0,-1)}]$n_2$}] {} ++ (0,-0.7) ; \draw[black] (3,-3) -- node [label={[shift={(0,-1)}]$n_d$}] {} ++ (0,-0.7) ; \draw[black] (5,-3) -- node [label={[shift={(0,-0.95)}]$m$}] {} ++ (0,-0.7) ; \draw[black] (0,-3) -- node [label={[shift={(0,0)}]$1$}] {} ++ (-0.7,0) ; \draw[black] (0,-3) -- node [label={[shift={(0,0)}]$r_1$}] {} ++ (1,0) ; \draw[black] (1,-3) -- node [] {} ++ (0.66,0) ; \draw[black, dotted] (1.66,-3) -- node [] {} ++ (0.66,0) ; \draw[black] (2.33,-3) -- node [] {} ++ (0.66,0) ; \draw[black] (3,-3) -- node [label={[shift={(0,0)}]$r_d$}] {} ++ (2,0); \draw[black] (5,-3) -- node [label={[shift={(0,0)}]$1$}] {} ++ (0.7,0) ; \node[draw,shape=circle,fill=YellowOrange, scale=0.65] at (0,-3){}; \node[draw,shape=circle,fill=YellowOrange, scale=0.65] at (1,-3){}; \node[draw,shape=circle,fill=YellowOrange, scale=0.65] at (3,-3){}; \node[draw,shape=circle,fill=cyan, scale=0.65] at (5,-3){}; \def-3{-2.5} \node[anchor=east] at (-2.3,-3-0.2){Left- and right-}; \node[anchor=east] at (-2.3,-3-0.6){orthonormalization}; \draw[black] (0,-3) -- node [label={[shift={(0,-1)}]$n_1$}] {} ++ (0,-0.7) ; \draw[black] (1,-3) -- node [label={[shift={(0,-1)}]$n_2$}] {} ++ (0,-0.7) ; \draw[black] (3,-3) -- node [label={[shift={(0,-1)}]$n_d$}] {} ++ (0,-0.7) ; \draw[black] (5,-3) -- node [label={[shift={(0,-0.95)}]$m$}] {} ++ (0,-0.7) ; \draw[black] (0,-3) -- node [label={[shift={(0,0)}]$1$}] {} ++ (-0.7,0) ; \draw[black] (0,-3) -- node [label={[shift={(0,0)}]$s_1$}] {} ++ (1,0) ; \draw[black] (1,-3) -- node [] {} ++ (0.66,0) ; \draw[black, dotted] (1.66,-3) -- node [] {} ++ (0.66,0) ; \draw[black] (2.33,-3) -- node [] {} ++ (0.66,0) ; \draw[black] (3,-3) -- node [label={[shift={(0,0)}]$s_d$}] {} ++ (1,0); \draw[black] (4,-3) -- node [label={[shift={(0,0)}]$s_d$}] {} ++ (1,0); \draw[black] (5,-3) -- node [label={[shift={(0,0)}]$1$}] {} ++ (0.7,0) ; \node[draw,shape=semicircle,rotate=135 ,fill=white, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (0,-3){}; \node[draw,shape=semicircle,rotate=315 ,fill=YellowOrange, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (0,-3){}; \node[draw,shape=semicircle,rotate=135 ,fill=white, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (1,-3){}; \node[draw,shape=semicircle,rotate=315 ,fill=YellowOrange, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (1,-3){}; \node[draw,shape=semicircle,rotate=135 ,fill=white, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (3,-3){}; \node[draw,shape=semicircle,rotate=315 ,fill=YellowOrange, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (3,-3){}; \node[draw,shape=semicircle,rotate=45 ,fill=cyan, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (5,-3){}; \node[draw,shape=semicircle,rotate=225 ,fill=white, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (5,-3){}; \node[draw,shape=circle,fill=Gray, scale=0.5] at (4,-3){}; \node[] at (4,-3-0.4){$\Sigma$}; \def-3{-6} \node[anchor=east] at (-2.3,-3+0.4){Pseudoinverse}; \draw[black] (2,-3) -- node [label={[shift={(0,-1)}]$n_1$}] {} ++ (0,-0.7) ; \draw[black] (3,-3) -- node [label={[shift={(0,-1)}]$n_2$}] {} ++ (0,-0.7) ; \draw[black] (5,-3) -- node [label={[shift={(0,-1)}]$n_d$}] {} ++ (0,-0.7) ; \draw[black] (0,-3) -- node [label={[shift={(0,-0.95)}]$m$}] {} ++ (0,-0.7) ; \draw[black] (0,-3) -- node [label={[shift={(0,0)}]$1$}] {} ++ (2,0); \draw[black] (2,-3) -- node [label={[shift={(0,0)}]$s_1$}] {} ++ (1,0); \draw[black] (3,-3) -- node [] {} ++ (0.66,0) ; \draw[black, dotted] (3.66,-3) -- node [] {} ++ (0.66,0) ; \draw[black] (4.33,-3) -- node [] {} ++ (0.66,0) ; \draw[black, rounded corners=0.1cm] (0,-3) -- node [label={[shift={(0,0)}]$s_d$}] {} ++ (-0.7,0) -- ++ (0,1.5) -- ++ (6.4,0) -- ++ (0,-1.5) -- node [label={[shift={(0,0)}]$s_d$}] {} ++ (-0.7,0); \node[draw,shape=semicircle,rotate=135 ,fill=white, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (2,-3){}; \node[draw,shape=semicircle,rotate=315 ,fill=YellowOrange, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (2,-3){}; \node[draw,shape=semicircle,rotate=135 ,fill=white, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (3,-3){}; \node[draw,shape=semicircle,rotate=315 ,fill=YellowOrange, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (3,-3){}; \node[draw,shape=semicircle,rotate=135 ,fill=white, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (5,-3){}; \node[draw,shape=semicircle,rotate=315 ,fill=YellowOrange, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (5,-3){}; \node[draw,shape=semicircle,rotate=45 ,fill=cyan, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (0,-3){}; \node[draw,shape=semicircle,rotate=225 ,fill=white, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (0,-3){}; \node[draw,shape=circle,fill=Gray, scale=0.5] at (2.5,-3+1.5){}; \node[] at (2.5,-3+1.5-0.4){$\Sigma^{-1}$}; \end{tikzpicture} \caption[]{Computation of the pseudoinverse of a tensor train: After left- and right-orthonormalization of the initial tensor train (half-filled circles), the pseudoinverse can be represented as a cyclic tensor train with reordered cores. Note that we here depict the special case of the pseudoinversion of a tensor $\mathbf{T} \in \mathbb{R}^{n_1 \times \dots \times n_d \times m}$, the general case is considered in~\cite{KLUS2018}.} \label{fig: Pseudoinverse} \end{figure} In order to directly construct the pseudoinverse $T^+$ from a TT representation of $\mathbf{T}$, we apply the two orthonormalization procedures. That is, we left-orthonormalize the TT cores from $\mathbf{T}^{(1)}$ to $\mathbf{T}^{(d)}$ and right-orthonormalize the core $\mathbf{T}^{(d+1)}$. Doing this, we obtain a global SVD of the matricization $T$. Similar to \eqref{eq: pseudoinverse in full format}, the pseudoinverse $T^+$ can then be obtained by reordering the cores, see Figure~\ref{fig: Pseudoinverse}. A detailed description of the corresponding algorithms can be found in~\cite{KLUS2018}, where we have shown how to generalize the orthonormalization procedure to arbitrary matricizations of tensor trains. Algorithm~\ref{alg:Pseudoinverse} describes the pseudoinversion procedure for a tensor train $\mathbf{T} \in \mathbb{R}^{n_1 \times \dots \times n_d \times m}$ in detail. \begin{algorithm}[htbp] \caption{Pseudoinversion of tensor trains.} \label{alg:Pseudoinverse} \begin{algorithmic}[1] \Require \parbox[t]{\dimexpr\linewidth-31.5pt}{Tensor train $\mathbf{T} \in \mathbb{R}^{n_1 \times \dots \times n_d \times m}$.\vphantom{$T = \mat{\mathbf{T}}{n_1, \dotsc, n_d}{ m}$}} \Ensure Pseudoinverse of $T = \mat{\mathbf{T}}{n_1, \dotsc, n_d}{ m}$. \algrule \State Left-orthonormalize $\mathbf{T}^{(1)}, \dotsc, \mathbf{T}^{(d-1)}$ and right-orthonormalize $\mathbf{T}^{(d+1)}$. \State Compute SVD of $\mathcal{L} \left( \mathbf{T}^{(d)} \right)$, i.e., $\mathcal{L} \left( \mathbf{T}^{(d)} \right) = U \Sigma V^T$ with $\Sigma \in \mathbb{R}^{s \times s}$. \State Define $\mathbf{U} \in \mathbb{R}^{r_{d-1} \times n_d \times s}$ as a reshaped version of $U$ with $\mathbf{U}_{k,x,l } = U_{\overline{k, x}, l}$. \State Define $\mathbf{V} \in \mathbb{R}^{s \times m }$ by $\mathcal{R} \left( \mathbf{V} \right) = V^T \cdot \mathcal{R} \left( \mathbf{T}^{(d+1)} \right)$. \State Set $\mathbf{T}^{(d)}$ to $\mathbf{U}$, $\mathbf{T}^{(d+1)}$ to $\mathbf{V}$, and $r_d$ to $s$. \State Define $\tilde{U} = \mat{\left( \sum_{k_0 =1}^{r_0} \cdots \sum_{k_{d-1}=1}^{r_{d-1}} \mathbf{T}^{(1)}_{k_0, :, k_1 } \otimes \dots \otimes \mathbf{T}^{(d)}_{k_{d-1}, :, : } \right)}{n_1, \dotsc, n_d}{r_d}$. \State Define $\tilde{V} = \mat{\mathbf{T}^{(d+1)}}{m}{r_d}$. \State Define $ T^+ = \tilde{V} \, \Sigma^{-1} \, \tilde{U}^T$. \end{algorithmic} \end{algorithm} An important aspect is that we do not need to compute the pseudoinverse of $T$ explicitly. Instead, we only orthonormalize the TT cores and compute the matrix $\Sigma$ by executing the lines 1 to 5 of Algorithm~\ref{alg:Pseudoinverse}. We then store the representation \begin{equation*} \begin{split} \mathbf{T}^+ = \sum_{k_1 = 1}^{r_1} \dots \sum_{k_{d} = 1}^{r_{d}} \sigma^{-1}_{ k_d} \cdot \mathbf{T}^{(d+1)}_{k_{d}, :, 1 } \otimes \mathbf{T}^{(1)}_{1, :, k_{1} } \otimes \dots \otimes \mathbf{T}^{(d)}_{k_{d-1}, :, k_d } \in \mathbb{R}^{m \times n_1 \times \dots \times n_d}, \end{split} \end{equation*} which can be either regarded as the sum of $r_d$ tensor trains scaled by $\sigma^{-1}_{1}, \dotsc, \sigma^{-1}_{r_d} $ or as a cyclic tensor train \cite{HACKBUSCH2012} as depicted in Figure~\ref{fig: Pseudoinverse}. It holds that \begin{equation*} \mat{\mathbf{T}^+}{m}{n_1, \dots, n_d} = \left( \mat{\mathbf{T}}{n_1, \dots, n_d}{m}\right)^+. \end{equation*} For the orthonormalization procedures, we only consider compact/reduced SVDs, i.e., only the nonzero singular values and corresponding singular vectors are stored. It is also possible to use truncated SVDs within Algorithm \ref{alg:Pseudoinverse}, i.e., we discard all singular values $\sigma_k$ with $\sigma_k / \sigma_\textrm{max} < \varepsilon$, where $\sigma_\textrm{max}$ is the largest singular value and $\varepsilon$ a given threshold. In this way, we can reduce the computational costs and the storage consumption even further. \begin{lemma}[Complexity of the pseudoinverse computation]\label{lemma: pseudoinverse - effort} The computational effort of Algorithm \ref{alg:Pseudoinverse} can be estimated as $O(d \cdot n \cdot r^3 + m \cdot r^2)$, where $n$ is the maximum of the first $d$ modes and $r$ the maximum of all TT ranks. \end{lemma} \begin{proof} The overall computational costs of Algorithm \ref{alg:Pseudoinverse} can be estimated by the number of applied SVDs. The complexity of calculating an SVD of a left-/right-unfolding can be estimated as $O(n \cdot r^3)$. Since we compute the left-orthonormalizations of the first $d$ cores and the right-orthonormalization of $\mathcal{R}(\mathbf{T}^{(d+1)}) \in \mathbb{R}^{r \times m}$, we obtain the estimation $O(d \cdot n \cdot r^3 + m \cdot r^2)$, assuming that $r \leq m$. \end{proof} \section{Tensor-based reformulation of SINDy}\label{sec: tensor-based SINDy} After introducing SINDy and the TT format, we will now show how to combine the data-driven recovery of dynamical systems with tensor decompositions. We restrict ourselves to a set of basis functions $ \mathbb{D} = \{\psi_1 , \dots , \psi_p\}$, $\psi_j \colon \mathbb{R} \rightarrow \mathbb{R}$, and corresponding basis matrices whose entries are given by products of the given basis functions applied to the different dimensions of all snapshots. Note that the basis functions $\psi_j$, $j = 1, \dots , p$, are now defined on $\mathbb{R}$ and not on $\mathbb{R}^d$ as in Section~\ref{sec: recovery}. Exploiting the TT format for the construction of the tensorized counterpart of the basis matrix $\Psi(\mathcal{X})$, we are able to express large numbers of combinations of the basis functions $\psi_1, \dots , \psi_p$ in a compact way as tensor-products of the one-dimensional basis functions. By doing this, we can not only reduce the storage consumption of the basis matrix but may also lower the computational costs compared to the conventional SINDy-like implementation. In what follows, we will describe different approaches for the tensor-based construction and show how to solve the minimization problem \begin{equation}\label{eq: min problem in TT} \min_{\mathbf{\Xi} } \lVert \mathcal{Y} - \mathbf{\Xi}^T \mathbf{\Psi}(\mathcal{X}) \rVert_F \end{equation} directly in the TT format. Here, $\mathbf{\Xi}$ and $\mathbf{\Psi}(\mathcal{X})$ denote the tensorized counterparts of the matrices $\Xi$ and $\Psi(\mathcal{X})$ introduced in Section~\ref{sec: notation}. As mentioned in the introduction, we call our method MANDy. \subsection{Basis decomposition} We will consider two different types of basis decompositions. Let $X = [x_1, \dots, x_d]^T \in \mathbb{R}^d$ be a vector and $\psi_j : \mathbb{R} \rightarrow \mathbb{R}$, $j = 1, \dots, p$, basis functions. We consider the two rank-one decompositions \begin{equation} \label{eq: psi f,x} \mathbf{\Psi}_{\textrm{cm}}(X) = \mathbf{\Psi}_{\textrm{cm}}^{(1)}(X) \otimes \dots \otimes \mathbf{\Psi}_{\textrm{cm}}^{(d)}(X) = \begin{bmatrix} \psi_1 (x_1) \\ \vdots \\ \psi_p(x_1) \end{bmatrix} \otimes \dots \otimes \begin{bmatrix} \psi_1 (x_d) \\ \vdots \\ \psi_p(x_d) \end{bmatrix} \in \mathbb{R}^{p \times \dots \times p} \end{equation} and \begin{equation} \label{eq: psi x,f} \mathbf{\Psi}_{\textrm{fm}}(X) = \mathbf{\Psi}_{\textrm{fm}}^{(1)}(X) \otimes \dots \otimes \mathbf{\Psi}_{\textrm{fm}}^{(p)}(X) = \begin{bmatrix} \psi_1 (x_1) \\ \vdots \\ \psi_1(x_d) \end{bmatrix} \otimes \dots \otimes \begin{bmatrix} \psi_p (x_1) \\ \vdots \\ \psi_p(x_d) \end{bmatrix} \in \mathbb{R}^{d \times \dots \times d}. \end{equation} That is, for \eqref{eq: psi f,x} we apply all basis functions to one specific element of the vector $X$ in each core while we apply only one basis function to all elements of $X$ in each core of \eqref{eq: psi x,f}. Analogously to the row and column major order of multidimensional arrays, we call $\mathbf{\Psi}_\textrm{cm}$ a \emph{coordinate-major decomposition} and $\mathbf{\Psi}_\textrm{fm}$ a \emph{function-major decomposition}. It holds that \begin{equation*} \left( \mathbf{\Psi}_\textrm{cm}(X) \right)_{j_1 , \dots , j_d} = \psi_{j_1}(x_1) \cdot \ldots \cdot \psi_{j_d}(x_d) \quad \text{and} \quad \left( \mathbf{\Psi}_\textrm{fm}(X) \right)_{j_1, \dots, j_p} = \psi_1(x_{j_1}) \cdot \ldots \cdot \psi_p(x_{j_p}). \end{equation*} Using the decompositions \eqref{eq: psi f,x} and \eqref{eq: psi x,f}, we can express the tensors $\mathbf{\Psi}_\textrm{cm}(X)$ and $\mathbf{\Psi}_\textrm{fm}(X)$ in a storage-efficient way. The memory consumption of the full representations of the tensors is $O(p^d)$ and $O(d^p)$, respectively, whereas the memory consumption of both rank-one representations is $O(p \cdot d)$. \begin{remark}[Choice of rank-one decompositions] What we present here are only two types of rank-one decompositions for a set of basis functions. In fact, other decompositions are also possible and specific problems might necessitate more complex representations in the future. However, we will focus on decompositions of the form \eqref{eq: psi f,x} or \eqref{eq: psi x,f}. \end{remark} \subsection{Multidimensional approximation of nonlinear dynamical systems} For $m$ different vectors $X_k = \left[x_{k,1}, \dots , x_{k,d}\right]^T \in \mathbb{R}^d$, $k = 1, \dots , m$, stored in a matrix $\mathcal{X}$ (see Section \ref{sec: recovery}), we aim at using the decompositions \eqref{eq: psi f,x} and \eqref{eq: psi x,f} to construct counterparts of the matrices \begin{equation}\label{eq: matrix coordinate major} \Psi_\textrm{cm}(\mathcal{X}) = \begin{bmatrix} \Psi_\textrm{cm}(X_1) & \Psi_\textrm{cm}(X_2) & \dots & \Psi_\textrm{cm}(X_m) \end{bmatrix} \in \mathbb{R}^{p^d \times m} \end{equation} and \begin{equation}\label{eq: matrix function major} \Psi_\textrm{fm}(\mathcal{X}) = \begin{bmatrix} \Psi_\textrm{fm}(X_1) & \Psi_\textrm{fm}(X_2) & \dots & \Psi_\textrm{fm}(X_m) \end{bmatrix} \in \mathbb{R}^{d^p \times m} \end{equation} directly in the TT format. Here, $\Psi_\textrm{cm}(X_k) \in \mathbb{R}^{p^d}$ and $\Psi_\textrm{fm}(X_k) \in \mathbb{R}^{d^p}$ denote the vectorizations $\textrm{vec}\left(\mathbf{\Psi}_\textrm{cm}(X_k)\right)$ and $\textrm{vec}\left(\mathbf{\Psi}_\textrm{fm}(X_k)\right)$, respectively. Collecting the rank-one decompositions given in \eqref{eq: psi f,x} for all vectors $X_1, \dots, X_m$ in a single TT decomposition, we obtain \begin{equation}\label{eq: TT coordinate major} \begin{split} \mathbf{\Psi}_\textrm{cm}(\mathcal{X}) &= \left \llbracket \mathbf{\Psi}_\textrm{cm} ^{(1)} (\mathcal{X})\right \rrbracket \otimes \left \llbracket \mathbf{\Psi}_\textrm{cm} ^{(2)} (\mathcal{X})\right \rrbracket \otimes \dots \otimes \left \llbracket \mathbf{\Psi}_\textrm{cm} ^{(d)} (\mathcal{X})\right \rrbracket \otimes \left \llbracket \mathbf{\Psi}_\textrm{cm} ^{(d+1)} (\mathcal{X})\right \rrbracket \\ &= \left \llbracket \begin{matrix} \mathbf{\Psi}_\textrm{cm}^{(1)} (X_1) & \cdots & \mathbf{\Psi}_\textrm{cm}^{(1)} (X_m) \end{matrix} \right\rrbracket \otimes \left\llbracket\begin{matrix} \mathbf{\Psi}_\textrm{cm}^{(2)} (X_1) & & 0 \\ & \ddots & \\ 0 & & \mathbf{\Psi}_\textrm{cm}^{(2)} (X_m) \end{matrix} \right\rrbracket \otimes \cdots \\ & \qquad \cdots \otimes \left \llbracket \begin{matrix} \mathbf{\Psi}_\textrm{cm}^{(d)} (X_1) & & 0 \\ & \ddots & \\ 0 & & \mathbf{\Psi}_\textrm{cm}^{(d)} (X_m) \end{matrix} \right\rrbracket \otimes \left \llbracket \begin{matrix} e_1 \\ \vdots \\ e_m \end{matrix} \right\rrbracket \in \mathbb{R}^{p \times \dots \times p \times m}, \end{split} \end{equation} where $e_k$, $k = 1, \dots , m$, denote the unit vectors of the standard basis in the $m$-dimensional Euclidean space. Analogously, we can write \begin{equation}\label{eq: TT function major} \begin{split} \mathbf{\Psi}_\textrm{fm}(\mathcal{X}) &= \left \llbracket \mathbf{\Psi}_\textrm{fm} ^{(1)} (\mathcal{X})\right \rrbracket \otimes \left \llbracket \mathbf{\Psi}_\textrm{fm} ^{(2)} (\mathcal{X})\right \rrbracket \otimes \dots \otimes \left \llbracket \mathbf{\Psi}_\textrm{fm} ^{(p)} (\mathcal{X})\right \rrbracket \otimes \left \llbracket \mathbf{\Psi}_\textrm{fm} ^{(p+1)} (\mathcal{X})\right \rrbracket \\ &= \left \llbracket \begin{matrix} \mathbf{\Psi}_\textrm{fm}^{(1)} (X_1) & \cdots & \mathbf{\Psi}_\textrm{fm}^{(1)} (X_m) \end{matrix} \right\rrbracket \otimes \left \llbracket \begin{matrix} \mathbf{\Psi}_\textrm{fm}^{(2)} (X_1) & & 0 \\ & \ddots & \\ 0 & & \mathbf{\Psi}_\textrm{fm}^{(2)} (X_m) \end{matrix} \right\rrbracket \otimes \cdots \\ & \qquad \cdots \otimes \left \llbracket \begin{matrix} \mathbf{\Psi}_\textrm{fm}^{(p)} (X_1) & & 0 \\ & \ddots & \\ 0 & & \mathbf{\Psi}_\textrm{fm}^{(p)} (X_m) \end{matrix} \right\rrbracket \otimes \left \llbracket \begin{matrix} e_1 \\ \vdots \\ e_m \end{matrix} \right\rrbracket \in \mathbb{R}^{d \times \dots \times d \times m}. \end{split} \end{equation} Both TT decompositions \eqref{eq: TT coordinate major} and \eqref{eq: TT function major} have a TT rank of $m$ and it holds that \begin{equation*} \mat{\mathbf{\Psi}_\textrm{cm}(\mathcal{X})}{p , \dots, p}{m} = \Psi_\textrm{cm}(\mathcal{X}) \quad \textrm{and} \quad \mat{\mathbf{\Psi}_\textrm{fm}(\mathcal{X})}{d , \dots, d}{m} = \Psi_\textrm{fm}(\mathcal{X}). \end{equation*} We could also express the basis tensors in the canonical format, but using the TT format enables us to construct the pseudoinverse of $\mathbf{\Psi}_\textrm{cm}(\mathcal{X})$ and $\mathbf{\Psi}_\textrm{fm}(\mathcal{X})$ directly as a TT decomposition, see Section \ref{sec: pseudoinverse}. That is, we solve the minimization problem \eqref{eq: min problem in TT} by computing the pseudoinverse of $\mathbf{\Psi}_\textrm{cm}(\mathcal{X})$ or $\mathbf{\Psi}_\textrm{fm}(\mathcal{X})$, respectively, in the TT format and obtain \begin{equation}\label{eq: solution in TT} \mathbf{\Xi}^T = \mathcal{Y} \cdot \mathbf{\Psi}_{\textrm{cm}/\textrm{fm}}(\mathcal{X})^+ \end{equation} with \begin{equation*} \mat{\mathbf{\Xi}^T}{d}{p/d, \dots , p/d} = \mathcal{Y} \cdot \mat{\mathbf{\Psi}_{\textrm{cm}/\textrm{fm}}(\mathcal{X})^+}{m}{p/d, \dots, p/d} = \mathcal{Y} \cdot \Psi_{\textrm{cm}/\textrm{fm}}(\mathcal{X})^+. \end{equation*} For a detailed description of the tensor contraction in \eqref{eq: solution in TT}, see Figure~\ref{fig: contraction}. \begin{figure}[htbp] \centering \begin{tikzpicture} \def-3{0} \node[anchor=east] at (-2.3,-3+0.4){$\mathbf{\Xi}^T$}; \node[anchor=east] at (-1.4,-3+0.3){$=$}; \draw[black] (1,-3) -- node [label={[shift={(0,-1.1)}]$p/d$}] {} ++ (0,-0.7) ; \draw[black] (2,-3) -- node [label={[shift={(0,-1.1)}]$p/d$}] {} ++ (0,-0.7) ; \draw[black] (4,-3) -- node [label={[shift={(0,-1.1)}]$p/d$}] {} ++ (0,-0.7) ; \draw[black] (0,-3) -- node [label={[shift={(-0.3,-0.25)}]$m$}] {} ++ (0,-1) ; \draw[black] (0,-3-1) -- node [label={[shift={(0,-1)}]$d$}] {} ++ (0,-0.7) ; \node[draw,shape=circle, fill=white, inner sep = 2pt] at (0,-3-1){$\mathcal{Y}$}; \draw[black] (0,-3) -- node [label={[shift={(0,0)}]$1$}] {} ++ (1,0); \draw[black] (1,-3) -- node [label={[shift={(0,0)}]$r_1$}] {} ++ (1,0); \draw[black] (2,-3) -- node [] {} ++ (0.66,0) ; \draw[black, dotted] (2.66,-3) -- node [] {} ++ (0.66,0) ; \draw[black] (3.33,-3) -- node [] {} ++ (0.66,0) ; \draw[black, rounded corners=0.1cm] (0,-3) -- node [label={[shift={(0,-0.1)}]$r_{d/p}$}] {} ++ (-1,0) -- ++ (0,1.5) -- ++ (6,0) -- ++ (0,-1.5) -- node [label={[shift={(0,-0.1)}]$r_{d/p}$}] {} ++ (-1,0); \node[draw,shape=semicircle,rotate=135 ,fill=white, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (1,-3){}; \node[draw,shape=semicircle,rotate=315 ,fill=YellowOrange, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (1,-3){}; \node[draw,shape=semicircle,rotate=135 ,fill=white, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (2,-3){}; \node[draw,shape=semicircle,rotate=315 ,fill=YellowOrange, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (2,-3){}; \node[draw,shape=semicircle,rotate=135 ,fill=white, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (4,-3){}; \node[draw,shape=semicircle,rotate=315 ,fill=YellowOrange, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (4,-3){}; \node[draw,shape=semicircle,rotate=45 ,fill=cyan, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (0,-3){}; \node[draw,shape=semicircle,rotate=225 ,fill=white, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (0,-3){}; \node[draw,shape=circle,fill=Gray, scale=0.5] at (2,-3+1.5){}; \node[] at (2,-3+1.5-0.4){$\Sigma^{-1}$}; \def-3{-3} \def-2{-2} \node[anchor=east] at (-1.7+-2,-3-0.2){$\mathbf{\Xi}^{\phantom{T}}$}; \node[anchor=east] at (-0.85+-2,-3-0.3){$=$}; \draw[black] (0+-2,-3) -- node [label={[shift={(0,-1.1)}]$p/d$}] {} ++ (0,-0.7) ; \draw[black] (1+-2,-3) -- node [label={[shift={(0,-1.1)}]$p/d$}] {} ++ (0,-0.7) ; \draw[black] (3+-2,-3) -- node [label={[shift={(0,-1.1)}]$p/d$}] {} ++ (0,-0.7) ; \draw[black] (6+-2,-3) -- node [label={[shift={(0,-1)}]$d$}] {} ++ (0,-0.7) ; \draw[black] (0+-2,-3) -- node [label={[shift={(0,0)}]$r_1$}] {} ++ (1,0); \draw[black] (3+-2,-3) -- node [label={[shift={(0,-0.1)}]$r_{d/p}$}] {} ++ (1,0); \draw[black] (1+-2,-3) -- node [] {} ++ (0.66,0) ; \draw[black, dotted] (1.66+-2,-3) -- node [] {} ++ (0.66,0) ; \draw[black] (2.33+-2,-3) -- node [] {} ++ (0.66,0) ; \node[draw,rounded corners, fill=white, inner sep = 4pt] at (6+-2,-3){\raisebox{-1em}{$\Sigma^{-1} \cdot \mathbf{\Psi}_{\textrm{cm}/\textrm{fm}}^{(d/p +1)}(\mathcal{X}) \cdot \mathcal{Y}^T$}}; \node[draw,shape=semicircle,rotate=135 ,fill=white, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (0+-2,-3){}; \node[draw,shape=semicircle,rotate=315 ,fill=YellowOrange, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (0+-2,-3){}; \node[draw,shape=semicircle,rotate=135 ,fill=white, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (1+-2,-3){}; \node[draw,shape=semicircle,rotate=315 ,fill=YellowOrange, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (1+-2,-3){}; \node[draw,shape=semicircle,rotate=135 ,fill=white, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (3+-2,-3){}; \node[draw,shape=semicircle,rotate=315 ,fill=YellowOrange, anchor=south,inner sep=2pt, outer sep=0pt, scale=0.75] at (3+-2,-3){}; \end{tikzpicture} \caption[]{Construction of the coefficient tensor $\mathbf{\Xi}$: After the left-orthonormalization of the first $d$ or $p$ cores, respectively, of $\mathbf{\Psi}_{\textrm{cm}/\textrm{fm}} (\mathcal{X})$, the tensor $\mathbf{\Xi}$ can be obtained by contracting the pseudoinverse with the data matrix $\mathcal{Y}$. Since $\mathbf{\Xi}^T$ is a cyclic tensor train, the tensor $\mathbf{\Xi}$ can be expressed as a standard tensor train where we only replace the last core.} \label{fig: contraction} \end{figure} Storing the TT cores of \eqref{eq: TT coordinate major} and \eqref{eq: TT function major} in a sparse format, the memory consumption of both representations can be estimated as $O(p \cdot d \cdot m)$. The memory consumption of the corresponding matricizations would be $O(p^d \cdot m)$ and $O(d^p \cdot m)$, respectively. This is the first main advantage of the proposed decompositions: If the number of snapshots is not too large ($m \ll p^d$ or $m \ll d^p$), we are able to efficiently store all entries of the matrices \eqref{eq: matrix coordinate major} and \eqref{eq: matrix function major} in the TT decompositions \eqref{eq: TT coordinate major} and \eqref{eq: TT function major}, respectively. The second advantage is then the fast computation of the pseudoinverse. The computational effort needed to compute an SVD of the matrices \eqref{eq: matrix coordinate major} and \eqref{eq: matrix function major} and to construct the pseudoinverse given in \eqref{eq: pseudoinverse in full format} can be estimated as $O(p^d \cdot m^2)$ and $O(d^p \cdot m^2)$, respectively. In contrast to that, we only need $O(p \cdot d \cdot m^3)$ steps to compute the pseudoinverse of the TT representations \eqref{eq: TT coordinate major} and \eqref{eq: TT function major}, see Lemma \ref{lemma: pseudoinverse - effort}. On the other hand, if $m \gtrsim p^d$ (or $m \gtrsim d^p$), then the tensor-train based approach is computationally more expensive than the classical methods. However, we still benefit from the reduced storage consumption of the basis tensor using MANDy. \begin{remark}[Right-orthonormality of the last core] Note that it is not necessary to right-orthonormalize the last core of $\mathbf{\Psi}_{\textrm{cm}/\textrm{fm}} (\mathcal{X})$ for the construction of its pseudoinverse as shown in Algorithm \ref{alg:Pseudoinverse} since the last core is already right-orthonormal by construction, cf.~\eqref{eq: TT coordinate major} and~\eqref{eq: TT function major}. \end{remark} \begin{remark}[Special case of monomial basis functions] For the special case of monomial basis functions, i.e., for a dictionary given by $\mathbb{D} = \{\psi_1, \dots, \psi_p\} = \{1, x, x^2, \dots , x^{p-1}\}$ with $p = 2^q$, we can decompose the components of the coordinate-major rank-one tensor given in \eqref{eq: psi f,x} even further by writing \begin{equation*} \begin{bmatrix} \psi_1(x_i) \\ \psi_2(x_i) \\ \vdots \\ \psi_p(x_i) \end{bmatrix} = \begin{bmatrix} 1 \\ x_i \\ \vdots \\ x_i^{2 ^q -1} \end{bmatrix} \hat{=} \begin{bmatrix} 1 \\ x_i \end{bmatrix} \otimes \begin{bmatrix} 1 \\ x_i^2 \end{bmatrix} \otimes \begin{bmatrix} 1 \\ x_i^4 \end{bmatrix} \otimes \dots \otimes \begin{bmatrix} 1 \\ x_i^{2^{q-1}} \end{bmatrix}, \end{equation*} which can be easily shown using $\sum_{l=0}^{q-1} 2^l = 2^{q}-1$. We will not make use of this kind of sub-decomposition in this work. Nevertheless, it may be advantageous for the identification of governing equations involving higher-order monomials. \end{remark} \begin{remark}[Constraints on the coefficients] Even when using the TT format for the construction of the tensors $\mathbf{\Psi}_\textrm{cm}(\mathcal{X})$ and $\mathbf{\Psi}_\textrm{fm}(\mathcal{X})$, respectively, it is possible to directly include linear (dimensionwise) constraints on the coefficients stored in $\mathbf{\Xi}$. A tensor train that encodes the extra conditions on the coefficients can be seen as an additional snapshot. Together with a corresponding vector appended to $\mathcal{Y}$, we can incorporate the (weighted) contraints. \end{remark} \section{Numerical results}\label{sec: results} In this section, we will present three numerical examples for the application of MANDy, namely Chua's circuit (which was already introduced in Example \ref{ex:SINDy examples}), the Fermi--Pasta--Ulam--Tsingou problem~\cite{FERMI1955}, and the Kuramoto model~\cite{ACEBRON2005}. The numerical experiments have been performed on a Linux machine with 128 GB RAM and an Intel Xeon processor with a clock speed of 3 GHz and 8 cores. The algorithms have been implemented in Python~3.6 and collected in the toolbox Scikit-TT available on GitHub: \url{https://github.com/PGelss/scikit_tt}. \subsection{Chua's circuit}\label{sec: chua} As a first experiment, let us consider our example from Section \ref{sec: recovery} again with the intent to illustrate how using tensor products of simple functions allows for the construction of rich sets of basis functions. The first set of basis functions \eqref{chua: first try} used in Example \ref{ex:SINDy examples} corresponds to the coordinate-major rank-one decomposition \begin{equation*} \mathbf{\Psi}_1(X) = \begin{bmatrix} 1 \\ x_1 \\ \, x_1^2 \, \end{bmatrix} \otimes \begin{bmatrix} 1 \\ x_2 \\ \, x_2^2 \, \end{bmatrix} \otimes \begin{bmatrix} 1 \\ x_3 \\ \, x_3^2 \, \end{bmatrix} \in \R^{3 \times 3 \times 3}. \end{equation*} As we have seen, these ansatz functions do not lead to a correct recovery of the system dynamics. Instead we have to use a basis set including the absolute values of the coordinates. The basis vector given in \eqref{chua: second try} then corresponds to the function-major rank-one decomposition \begin{equation}\label{eq: chua function-major} \mathbf{\Psi}_2(X) = \begin{bmatrix} 1 \\ x_1 \\ x_2 \\ \, x_3 \, \end{bmatrix} \otimes \begin{bmatrix} 1 \\ \abs{x_1} \\ \abs{x_2} \\ \, \abs{x_3} \, \end{bmatrix} \in \R^{4 \times 4}. \end{equation} As for the matrix-based approach in Example \ref{ex:SINDy examples}, the approximate coefficient tensor is numerically equal to the exact coefficient tensor (see Appendix \ref{app: chua}) when we apply MANDy with the basis decomposition $\mathbf{\Psi}_2$ and choose the same parameters and number of snapshots. This small example already shows the advantage of using tensor decompositions in terms of storage consumption. For the sparse TT representation of the basis tensor (see \eqref{eq: psi x,f} and \eqref{eq: TT function major}), we only have to store $18000$ entries whereas the matricized counterpart would have $32000$ entries. \subsection{Fermi--Pasta--Ulam--Tsingou problem} Let us now consider the Fermi--Pasta--Ulam--Tsingou model, which was introduced by the eponymous physicists and mathematicians in 1953. The underlying model represents a vibrating string by a system of coupled oscillators fixed at the respective end points of the string. \begin{figure}[htbp] \centering \begin{tikzpicture} \def1.5{1.5} \def1.5{1.5} \def0{0} \draw[] (5-3*1.5,0+0.5*1.5) -- ++ (0,-1.5); \draw[dashed] (5-2*1.5,0+0.5*1.5) -- ++ (0,-1.5); \draw[dashed] (5-1*1.5,0+0.5*1.5) -- ++ (0,-1.5); \draw[dashed] (5+1*1.5,0+0.5*1.5) -- ++ (0,-1.5); \draw[dashed] (5+2*1.5,0+0.5*1.5) -- ++ (0,-1.5); \draw[] (5+3*1.5,0+0.5*1.5) -- ++ (0,-1.5); \node[shape=circle,inner sep=0.1cm, outer sep=0, draw, fill=black] at (5-2*1.5+0.2,0) {}; \node[shape=circle,inner sep=0.1cm, outer sep=0, draw, fill=black] at (5-1*1.5-0.4,0) {}; \node[shape=circle,inner sep=0.1cm, outer sep=0, draw, fill=black] at (5+1*1.5-0.2,0) {}; \node[shape=circle,inner sep=0.1cm, outer sep=0, draw, fill=black] at (5+2*1.5+0.3,0) {}; \FPeval{\xa}{5-3*1.5} \FPeval{\xb}{5-2*1.5+0.2-0.1} \draw[] (\xa,0) -- ++ (0.166*\xb-0.166*\xa,0); \draw[] (\xa+0.166*\xb-0.166*\xa,0) -- ++ (0.0833*\xb-0.0833*\xa,0.1*1.5); \draw[] (\xa+0.166*\xb-0.166*\xa+0.0833*\xb-0.0833*\xa,0+0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,-0.2*1.5); \draw[] (\xa+0.333*\xb-0.333*\xa+0.0833*\xb-0.0833*\xa,0-0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,0.2*1.5); \draw[] (\xa+0.5*\xb-0.5*\xa+0.0833*\xb-0.0833*\xa,0+0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,-0.2*1.5); \draw[] (\xa+0.666*\xb-0.666*\xa+0.0833*\xb-0.0833*\xa,0-0.1*1.5) -- ++ (0.0833*\xb-0.0833*\xa,0.1*1.5); \draw[] (\xa+0.666*\xb-0.666*\xa+0.166*\xb-0.166*\xa,0) -- ++ (0.166*\xb-0.166*\xa,0); \FPeval{\xa}{5-2*1.5+0.2+0.1} \FPeval{\xb}{5-1*1.5-0.4-0.1} \draw[] (\xa,0) -- ++ (0.166*\xb-0.166*\xa,0); \draw[] (\xa+0.166*\xb-0.166*\xa,0) -- ++ (0.0833*\xb-0.0833*\xa,0.1*1.5); \draw[] (\xa+0.166*\xb-0.166*\xa+0.0833*\xb-0.0833*\xa,0+0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,-0.2*1.5); \draw[] (\xa+0.333*\xb-0.333*\xa+0.0833*\xb-0.0833*\xa,0-0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,0.2*1.5); \draw[] (\xa+0.5*\xb-0.5*\xa+0.0833*\xb-0.0833*\xa,0+0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,-0.2*1.5); \draw[] (\xa+0.666*\xb-0.666*\xa+0.0833*\xb-0.0833*\xa,0-0.1*1.5) -- ++ (0.0833*\xb-0.0833*\xa,0.1*1.5); \draw[] (\xa+0.666*\xb-0.666*\xa+0.166*\xb-0.166*\xa,0) -- ++ (0.166*\xb-0.166*\xa,0); \FPeval{\xa}{5-1*1.5-0.4+0.1} \FPeval{\xb}{5-1*1.5-0.2+0.1+0.85} \draw[] (\xa,0) -- ++ (0.166*\xb-0.166*\xa,0); \draw[] (\xa+0.166*\xb-0.166*\xa,0) -- ++ (0.0833*\xb-0.0833*\xa,0.1*1.5); \draw[] (\xa+0.166*\xb-0.166*\xa+0.0833*\xb-0.0833*\xa,0+0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,-0.2*1.5); \draw[] (\xa+0.333*\xb-0.333*\xa+0.0833*\xb-0.0833*\xa,0-0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,0.2*1.5); \draw[] (\xa+0.5*\xb-0.5*\xa+0.0833*\xb-0.0833*\xa,0+0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,-0.2*1.5); \draw[] (\xa+0.666*\xb-0.666*\xa+0.0833*\xb-0.0833*\xa,0-0.1*1.5) -- ++ (0.0833*\xb-0.0833*\xa,0.1*1.5); \draw[] (\xa+0.666*\xb-0.666*\xa+0.166*\xb-0.166*\xa,0) -- ++ (0.166*\xb-0.166*\xa,0); \draw[dotted] (5-1*1.5-0.2+0.1+0.85+0.1,0) -- (5+1*1.5-0.1-0.1-0.85-0.1,0); \FPeval{\xa}{5+1*1.5-0.1-0.1-0.85} \FPeval{\xb}{5+1*1.5-0.2-0.1} \draw[] (\xa,0) -- ++ (0.166*\xb-0.166*\xa,0); \draw[] (\xa+0.166*\xb-0.166*\xa,0) -- ++ (0.0833*\xb-0.0833*\xa,0.1*1.5); \draw[] (\xa+0.166*\xb-0.166*\xa+0.0833*\xb-0.0833*\xa,0+0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,-0.2*1.5); \draw[] (\xa+0.333*\xb-0.333*\xa+0.0833*\xb-0.0833*\xa,0-0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,0.2*1.5); \draw[] (\xa+0.5*\xb-0.5*\xa+0.0833*\xb-0.0833*\xa,0+0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,-0.2*1.5); \draw[] (\xa+0.666*\xb-0.666*\xa+0.0833*\xb-0.0833*\xa,0-0.1*1.5) -- ++ (0.0833*\xb-0.0833*\xa,0.1*1.5); \draw[] (\xa+0.666*\xb-0.666*\xa+0.166*\xb-0.166*\xa,0) -- ++ (0.166*\xb-0.166*\xa,0); \FPeval{\xa}{5+1*1.5-0.2+0.1} \FPeval{\xb}{5+2*1.5+0.3-0.1} \draw[] (\xa,0) -- ++ (0.166*\xb-0.166*\xa,0); \draw[] (\xa+0.166*\xb-0.166*\xa,0) -- ++ (0.0833*\xb-0.0833*\xa,0.1*1.5); \draw[] (\xa+0.166*\xb-0.166*\xa+0.0833*\xb-0.0833*\xa,0+0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,-0.2*1.5); \draw[] (\xa+0.333*\xb-0.333*\xa+0.0833*\xb-0.0833*\xa,0-0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,0.2*1.5); \draw[] (\xa+0.5*\xb-0.5*\xa+0.0833*\xb-0.0833*\xa,0+0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,-0.2*1.5); \draw[] (\xa+0.666*\xb-0.666*\xa+0.0833*\xb-0.0833*\xa,0-0.1*1.5) -- ++ (0.0833*\xb-0.0833*\xa,0.1*1.5); \draw[] (\xa+0.666*\xb-0.666*\xa+0.166*\xb-0.166*\xa,0) -- ++ (0.166*\xb-0.166*\xa,0); \FPeval{\xa}{5+2*1.5+0.15+0.1} \FPeval{\xb}{5+3*1.5} \draw[] (\xa,0) -- ++ (0.166*\xb-0.166*\xa,0); \draw[] (\xa+0.166*\xb-0.166*\xa,0) -- ++ (0.0833*\xb-0.0833*\xa,0.1*1.5); \draw[] (\xa+0.166*\xb-0.166*\xa+0.0833*\xb-0.0833*\xa,0+0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,-0.2*1.5); \draw[] (\xa+0.333*\xb-0.333*\xa+0.0833*\xb-0.0833*\xa,0-0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,0.2*1.5); \draw[] (\xa+0.5*\xb-0.5*\xa+0.0833*\xb-0.0833*\xa,0+0.1*1.5) -- ++ (0.166*\xb-0.166*\xa,-0.2*1.5); \draw[] (\xa+0.666*\xb-0.666*\xa+0.0833*\xb-0.0833*\xa,0-0.1*1.5) -- ++ (0.0833*\xb-0.0833*\xa,0.1*1.5); \draw[] (\xa+0.666*\xb-0.666*\xa+0.166*\xb-0.166*\xa,0) -- ++ (0.166*\xb-0.166*\xa,0); \draw[] (5-2*1.5,0-0.5*1.5-0.15) -- (5-2*1.5+0.2,0-0.5*1.5-0.15) node[below, midway] {$x_1$}; \draw[] (5-2*1.5,0-0.5*1.5-0.15+0.05) -- ++ (0,-0.1); \draw[] (5-2*1.5+0.2,0-0.5*1.5-0.15+0.05) -- ++ (0,-0.1); \draw[] (5-1*1.5,0-0.5*1.5-0.15) -- (5-1*1.5-0.4,0-0.5*1.5-0.15) node[below, midway] {$x_2$}; \draw[] (5-1*1.5,0-0.5*1.5-0.15+0.05) -- ++ (0,-0.1); \draw[] (5-1*1.5-0.4,0-0.5*1.5-0.15+0.05) -- ++ (0,-0.1); \draw[] (5+1*1.5,0-0.5*1.5-0.15) -- (5+1*1.5-0.2,0-0.5*1.5-0.15) node[below, midway] {$x_{d-1}$}; \draw[] (5+1*1.5,0-0.5*1.5-0.15+0.05) -- ++ (0,-0.1); \draw[] (5+1*1.5-0.2,0-0.5*1.5-0.15+0.05) -- ++ (0,-0.1); \draw[] (5+2*1.5,0-0.5*1.5-0.15) -- (5+2*1.5+0.3,0-0.5*1.5-0.15) node[below, midway] {$x_d$}; \draw[] (5+2*1.5,0-0.5*1.5-0.15+0.05) -- ++ (0,-0.1); \draw[] (5+2*1.5+0.3,0-0.5*1.5-0.15+0.05) -- ++ (0,-0.1); \end{tikzpicture} \caption{Fermi--Pasta--Ulam--Tsingou model: Representation of a vibrating string by a set of masses coupled by springs.} \label{fig: fpu model} \end{figure} Here, we consider a dynamical system described by a second-order differential equation $\ddot{X}(t) = F(X(t))$ where the right-hand side does not depend on $\dot{X}$. A very large number of problems in astronomy, molecular dynamics, and other areas of physics are of this form, as this type of differential equation is an immediate consequence of Newtonian mechanical laws. Moreover, it makes no difference for our methods whether we have given data measurements $[ X_1, \dots, X_m]$ and $[ \dot{X}_1, \dots, \dot{X}_m]$ or $[ X_1, \dots, X_m]$ and $[ \ddot{X}_1, \dots, \ddot{X}_m]$. Let us consider the model variant with cubic forcing terms, which is of the form \begin{equation}\label{eq: FPU} \ddot{x}_i = (x_{i+1} - 2 x_i + x_{i-1}) + \beta \left[(x_{i+1} - x_i)^3 - (x_i - x_{i-1})^3 \right], \end{equation} $ i = 1, \dots, d $, where $ x_i $ represents the displacement of the $ i $th oscillator from its original position, see Figure~\ref{fig: fpu model}. We assume the ends of the chain to be fixed, i.e., $ x_0 = x_{d+1} = 0 $. The parameter $ \beta \in \mathbb{R} $ represents the nonlinear force between the oscillators. For our first numerical experiment, we consider $d=10$ oscillators. We compare our proposed method with the matrix-based solution of the least-squares problem given in \eqref{eq: min problem in TT}. For this purpose, we choose random displacements in $[-0.1, 0.1]$ for each oscillator and compute the (exact) second derivatives using \eqref{eq: FPU} with $\beta = 0.7$. As basis functions we choose $\mathbb{D} = \{1, x, x^2, x^3\}$. This then results in the representation of $4^{10}$ combinations of function evaluations for every snapshot if we use the coordinate-major ansatz, see \eqref{eq: psi f,x}, such that \begin{equation*} \mathbf{\Psi}_\textrm{cm}(\mathcal{X}) \in \mathbb{R}^{4 \times \dots \times 4 \times m} \quad \textrm{for} \quad \mathcal{X} = \begin{bmatrix} X_1 & X_2 & \dots & X_m \end{bmatrix} \in \mathbb{R}^{10 \times m}, \end{equation*} where $m$ is the number of considered snapshots. \begin{figure}[htbp] \centering \begin{minipage}{0.49\textwidth} \centering \subfiguretitle{(a)} \includegraphics[height=140px]{pictures/example_fpu_1_a} \end{minipage} \hfill \begin{minipage}{0.49\textwidth} \centering \subfiguretitle{(b)} \includegraphics[height=140px]{pictures/example_fpu_1_b} \end{minipage} \caption{Application of MANDy to the Fermi--Pasta--Ulam--Tsingou problem: (a) CPU times needed for the computation of the coefficient tensor $\mathbf{\Xi}$ and the matricized counterpart $\Xi$, respectively. For $m = 6000$, the CPU time of the matrix-based approach is extrapolated since storing the matrices for calculations with $m > 5000$ would require too much memory. (b) Relative errors between the approximate solutions and the exact solution. Results are shown for the tensor-based approach (using different thresholds for orthonormalizations) as well as for the matrix-based approach. For $\varepsilon \leq 10^{-9}$, the relative errors of MANDy are numerically indistinguishable from those of the SINDy-like approach.} \label{fig: fpu 1} \end{figure} Figure~\ref{fig: fpu 1} shows the CPU times and relative errors of MANDy for varying $m$. We see that we are able to reduce the time needed for computing the coefficient tensor/matrix significantly -- at least within the considered range of snapshot numbers with $m \ll 4^{10}$. For MANDy, we also included the construction of the tensor train $\mathbf{\Psi}_\textrm{cm}(\mathcal{X})$ in the CPU times, whereas the runtimes for the matrix-based approach only consist of the times needed to solve the minimization problem. Without using truncated SVDs for the orthonormalization procedures, we obtain a speed-up of approximately up to five. The runtimes decrease even further when we set a threshold $\varepsilon > 0$ for the SVDs. \begin{figure}[htb] \centering \begin{minipage}{0.95\textwidth} \centering \includegraphics[height=140px]{pictures/example_fpu_2} \end{minipage} \caption{Relative errors for different parameters: The relative errors between the approximate and the exact solutions for the coefficient tensor depending on $d$ and $m$ are shown. All parameter combinations for $m$ and $d$ to the right of the dashed line could only be handled using the tensor-based method MANDy.} \label{fig: fpu 2} \end{figure} At the same time, we obtain nearly the same relative errors between the approximate and the exact solution, cf.~Appendix~\ref{app: FPU}. As shown in Figure \ref{fig: fpu 1} (b), there is in fact no noticeable difference between the error produced by the matrix-based approach and the tensor-based approach with thresholds of $\varepsilon \leq 10^{-9}$. Moreover, we can consider higher numbers of snapshots and dimensions using MANDy. This is shown more clearly in Figure~\ref{fig: fpu 2}, where we plot the relative errors in dependence of $m$ and $d$. In particular, we were not able to compute a matrix-based solution using standard SINDy for $d \geq 12$ since the basis matrix $\Psi_\textrm{cm}(\mathcal{X})$ simply becomes too large. Using MANDy, we can even approximate the solution with a relative error smaller than $10^{-1}$ for $d = 20$ and $m = 6000$ (which implies a basis tensor $\mathbf{\Psi}_\textrm{cm}(\mathcal{X})$ with $4^{20} \cdot 6000 \approx 6.6 \cdot 10^{15}$ elements). \subsection{Kuramoto model} As a third example for the application of MANDy, we consider the Kuramoto model. The model describes the behavior of a large number of coupled oscillators on a circle. First introduced by Yoshiki Kuramoto in 1975, see~\cite{KURAMOTO1975}, it has been studied extensively over the last decades. \def\oscillator[#1,#2](#3){ \FPeval{\xa}{2.25*cos(#3)+#1} \FPeval{\ya}{2.25*sin(#3)+#2} \draw[dashed] (#1,#2) -- (\xa,\ya); \node[shape=circle, draw=black, fill=black, inner sep=0.075cm] at (\xa,\ya) {}; } \begin{figure}[htbp] \centering \begin{tikzpicture} \def0{0} \def0{0} \draw[dotted] (0,0) -- ++ (2.25,0); \draw[fill=Gray!20!white, draw=none] (0,0) ++ (14.3239:2.25) arc (14.3239:0:2.25) -- (0,0); \node[] at (0+1.85,0+0.195) {$x_i$}; \draw[<-, > = latex] (0,0) ++ (14.3239+5:2.625) arc (14.3239+5:14.3239-5:2.625); \node[] at (0+2.85,0+0.6375) {$\omega_i$}; \node[draw=black, shape=circle, inner sep=1.59cm] at (0,0) {~}; \oscillator[0,0](0.25) \oscillator[0,0](0.45) \oscillator[0,0](0.7) \oscillator[0,0](0.88) \oscillator[0,0](1) \oscillator[0,0](1.2) \oscillator[0,0](1.5708) \oscillator[0,0](2) \oscillator[0,0](2.5) \oscillator[0,0](2.8) \oscillator[0,0](3) \oscillator[0,0](3.2) \oscillator[0,0](3.66) \oscillator[0,0](4) \oscillator[0,0](4.4) \oscillator[0,0](5) \oscillator[0,0](5.33) \oscillator[0,0](5.75) \oscillator[0,0](6) \end{tikzpicture} \caption[Kuramoto model]{Kuramoto model: Simulation of coupled oscillators on a ring.} \label{fig: kuramoto model} \end{figure} The governing equations -- including an external forcing, see~\cite{ACEBRON2005} -- can be written as \begin{equation}\label{eq: kuramoto ODE} \frac{\mathrm{d} x_i}{\mathrm{d} t} = \omega_i + \frac{K}{d} \sum_{j=1}^d \sin(x_j - x_i) + h \sin(x_i),\quad i = 1, \dots, d, \end{equation} where $ d $ is the number of oscillators, $ K $ the coupling strength, $h$ the external forcing parameter, $ \omega_i $ the $ i $th natural frequency, and $x_i$ the angular position of the $i$th oscillator, see Figure~\ref{fig: kuramoto model}. Here, we assume an identical coupling between all oscillators as well as a constant forcing parameter. In detail, we consider $d=100$ oscillators and set $K = 2$ and $h = 0.2$. Furthermore, we choose intrinsic frequencies equidistantly distributed in $[-5,5]$. Now, we intend to identify the governing equations from simulation data. We choose the set of basis functions given by $\mathbb{D} =\{\sin(x), \cos(x)\}$ and use the function-major decomposition \eqref{eq: psi x,f} which leads to tensors of the form \begin{equation}\label{eq: kuramoto decomposition} \mathbf{\Psi}_{\textrm{fm}}(X) = \begin{bmatrix} 1 \\ \sin(x_1) \\ \vdots \\ \sin(x_d) \end{bmatrix} \otimes \begin{bmatrix} 1 \\ \cos(x_1) \\ \vdots \\ \cos(x_d) \end{bmatrix} \in \mathbb{R}^{d +1 \times d+1}. \end{equation} Similar to the dictionary used for Chua's circuit, see Section \ref{sec: chua}, we added the basis function $1$ to both cores in order to ensure that also functions of the form $\sin(x_i)$ and $\cos(x_i)$ appear in the large basis set. Note that $ \sin(x \pm y) = \sin(x) \cos(y) \pm \cos (x) \sin(y)$ so that the system can indeed be represented by the chosen basis. We simulate the Kuramoto model from $ t_0 = 0 $ to $ t_1 = 1020 $ using an implementation of the BDF method as described in~\cite{SHAMPINE1997} and take $10$ snapshots within each second of simulation time, i.e., the time points corresponding to the snapshots are $\{0, 0.1 , 0.2, \dots , 1019.9, 1020\}$. That is, we consider $10201$ data points which would for the matrix case mean that the matrix $\Psi_\textrm{fm}(\mathcal{X})$ is square. As the initial distribution, we randomly place the oscillators on the ring. Then, based on the obtained data points $X_1 , \dots , X_{10201}$ and corresponding derivatives $\dot{X}_1 , \dots , \dot{X}_{10201}$, we recover the dynamics using MANDy with a threshold of $\varepsilon = 10^{-16}$. As in the previous examples, we are able to compare our result with the exact solution given in Appendix \ref{app: kuramoto}. We repeated the experiment several times and never observed relative errors larger than $10^{-4}$. Applying our recovered dynamics to another randomly chosen initial distribution of the oscillators, we see that we are able to approximate the dynamics of the Kuramoto model accurately within a time interval of nearly up to $100$ seconds, see Figure \ref{fig: kuramoto}. In terms of matrix representations, the reason for the inaccuracy is that the matrix $\Psi_\textrm{fm}(\mathcal{X})$ does not have full rank. Even if the computational overhead is bigger using MANDy compared to the classical method, we are able to reduce the storage consumption by a factor of more than $50$. Here, a sparse representation of the tensor cores of $\mathbf{\Psi}_\textrm{fm}(\mathcal{X})$ enables us to significantly reduce the memory consumption. \begin{figure}[htb] \centering \includegraphics[width=0.9\textwidth]{pictures/kuramoto} \caption[Application of MANDy to the Kuramoto model]{Application of MANDy to the Kuramoto model: The recovered dynamics are applied to another random initial distribution of the oscillators and then compared with the exact dynamics. Here, we only visualize the oscillators with indices $i = 1, 11 , \dots , 91$. The solid dots represent the positions according to the approximated dynamics whereas the circles with the same color represent the positions according to the real dynamics.} \label{fig: kuramoto} \end{figure} \section{Conclusion and outlook}\label{sec: conclusion} \subsection{Conclusion} \label{sec: actualconclusion} In this work, we have proposed an approach for the data-driven recovery of governing equations of nonlinear dynamical systems. The presented approach -- called MANDy, short for \emph{multidimensional approximation of nonlinear dynamical systems} -- combines data-driven methods with tensor-train decompositions. The aim of this approach is to reduce the memory consumption as well as the computational costs significantly and to mitigate the curse of dimensionality. After explaining the SINDy method for identifying governing equations based on a given basis set as well as the TT format, we have first shown how to use tensor products of simple functions for the construction of rich sets of basis functions. We have then explained how to rewrite the data-driven recovery in terms of tensor decompositions and the tensor-based computation of pseudoinverses. The results, which have been illustrated with several examples of dynamical systems, clearly demonstrate that the proposed algorithm needs less computational costs than the classical method if the number of data points is small enough. However, even if a large number of snapshots is needed for the recovery of the governing equations, we still benefit from the reduction of the memory consumption. The work presented in this paper constitutes a first step towards tensor-based techniques for the reconstruction of unknown systems purely from data measurements. \subsection{Outlook} \label{sec: structure} The investigations carried out in this work offer a number of further promising considerations. The above numerical experiments show convincingly that one can expect a significant improvement of MANDy over SINDy, not only concerning the memory requirements used in the recovery, but also in the computational effort. This advantage originates from exploiting the underlying data structure by using low-rank tensor decompositions. Key to this is an understanding of the correlations in the matrices that are approximated in the TT format. Appendix \ref{app: correlation} provides some details on the correlation structure required to obtain accurate TT approximations, requiring further research on the local correlation structure. Other future research will include the investigation of perturbations of the data and their influence on the reconstructed systems. We have observed that MANDy (as well as methods like SINDy) is highly sensitive to noisy data, even if we only add a slight Gaussian noise to the measurements. In this context, the inclusion of approximate derivatives for the recovery will also be of interest. A common way of adding noise is to take data generated by $\dot{X}(t) = F(X(t))+\xi$, where $\xi\in \mathbb{R}^d$ is a vector capturing errors, either drawn i.i.d.\ reflecting Gaussian noise, or generally with $\| \xi \|_F < \eta $ for some constant $\eta>0$. One might even hope for proving recovery guarantees as they are common in the context of compressed sensing \cite{CompressedSensingIntroRauhut,CompressedSensingIntroBoche}, given certain suitably structured data and noise levels. Moreover, the construction of elaborate sets of basis functions has to be examined more closely in the future. Here, we used a priori knowledge for the construction of the dictionary, which is generally not known in practice. Additionally, parts of this work -- in particular the proposed basis decompositions -- may be used for the development of other tensor-based counterparts of methods such as EDMD or kernel EDMD. Accepting that the improved performance originates from exploiting the correlation structure, one can hope that other tensor networks are suitable for other kinds of data. Hierarchical tensor formats \cite{HACKBUSCH2009, ARNOLD2013,PhysRevB.90.125154} and entangled project pair states \cite{Orus-AnnPhys-2014,VerstraeteBig} may be suitable as well. The hope is that the present work stimulates such endeavors. \section*{Acknowledgements} This research has been partially funded by Deutsche Forschungsgemeinschaft (DFG) through grant CRC 1114 and EI 519/9-1, as well as the Templeton Foundation and the European Research Council (ERC). {\small \bibliographystyle{unsrt}

f4363af30a48216434a6122e3b00d83924375af6

\section{Introduction}\label{sec:Intro} On a rectangle $\hat E$, serendipity finite elements ${\cal S}_r(\hat E)$ \cite{Strang_Fix_1973,Ciarlet_1978,Brenner_Scott_1994} and Brezzi-Douglas-Marini mixed finite elements BDM$_r(\hat E)$ \cite{BDM_1985} appear in the periodic table of the finite elements as given by Arnold and Logg \cite{Arnold_Logg_2014_periodicTable} (where they are denoted ${\cal S}_r\Lambda^0$ and ${\cal S}_r\Lambda^1$, respectively). They should be studied together, since they are related by a de Rham complex \cite{Arnold_Falk_Winter_2006, Arnold_2013, AFW_2010_feec} \begin{equation}\label{eq:deRhamBDM} \mathbb R \DOTSB\lhook\joinrel\longrightarrow {\cal S}_{r+1}(\hat E) \overset{\textrm{curl}\,}{\longlongrightarrow} \textrm{BDM}_r(\hat E) \overset{\textrm{div}\,}{\longlongrightarrow} \mathbb P_{r-1}(\hat E) \longrightarrow 0, \end{equation} which implies that $\textrm{BDM}_r(\hat E)=\textrm{curl}\,{\cal S}_{r+1}(\hat E)\oplus{\mathbf x}\mathbb P_{r-1}(\hat E)$, where $\mathbb P_{s}(\Hat E)$ are polynomials of degree $s$. Over a rectangular mesh, the serendipity elements merge into $H^1$ conforming spaces of scalar functions, and the BDM elements merge into $H(\textrm{div})=\big\{{\mathbf v}\in(L^2)^2:\nabla\cdot{\mathbf v}\in L^2\big\}$ conforming spaces of vector functions. In this paper, we define new (we call them \emph{direct}) serendipity and mixed finite elements on a general nondegenerate, convex quadrilateral $E$. These new elements generalize the complex \eqref{eq:deRhamBDM}, and they maintain optimal order approximation properties while possessing minimal local dimension. The serendipity finite elements on rectangles ${\cal S}_r(\hat E)$, especially the 8-node biquadratic ($r=1$) and the 12-node bicubic ($r=2$) ones, have been well studied for many years. They appear in almost any introductory reference on finite elements, e.g.,~\cite{Strang_Fix_1973,Ciarlet_1978,Brenner_Scott_1994}, and they are provided by software packages both in academia~\cite{DHJKLLS_2003_fenics} and industry~\cite{Hibbitt_Karlsson_Sorensen_2001_abaqus}. Compared with the full tensor product Lagrange finite elements $\mathbb P_{r,r}(\hat E)$, serendipity finite elements use fewer degrees of freedom, and they are usually more efficient. It was not until recently, however, that a general definition of the serendipity finite element spaces of arbitrary order on rectangles in any space dimension was given by Arnold and Awanou~\cite{Arnold_Awanou_2011, Arnold_Awanou_2014} (see also \cite{Gillette_Kloefkorn_2018_trimmed}). The serendipity finite element spaces work very well on computational meshes of rectangular elements, but it is well known that their performance is degraded on quadrilaterals when the space is mapped from a rectangle, when $r\ge2$. This is not the case for tensor product Lagrange finite elements~\cite{Lee_Bathe_1993,Kaliakin_2001,ABF_2002}. To be more precise, mapped serendipity elements of index $r$ do not approximate to optimal order $r+1$ on $E$, but the image of the full space of tensor product polynomials $\mathbb P_{r,r}(\hat E)$ maintains accuracy on $E$. We note that Rand, Gillette, and Bajaj \cite{Rand_Gillette_Bajaj_2014} recently introduced a new family of Serendipity finite elements based on generalized barycentric coordinates of index $r=2$ that is accurate to order three on any convex, planar polygon. A generalization to any order of approximation was given by Floater and Lai \cite{Floater_Lai_2016}, but on quadrilaterals, they require $\dim\mathbb P_r+r$ shape functions, which is more than the minimal required when $r>2$. There are many families of mixed finite elements on rectangles, beginning with those of Raviart and Thomas \cite{Raviart_Thomas_1977} and generalized by N\'ed\'elec \cite{Nedelec_1980}. These and the BDM$_r$ finite elements are extended to quadrilaterals using the Piola transform \cite{Thomas_1977, Raviart_Thomas_1977}. For most spaces, this creates a consistency error and consequent loss of approximation of the divergence \cite{Thomas_1977, Brezzi_Fortin_1991, ABF_2005, Boffi_Brezzi_Fortin_2013, Arbogast_Correa_2016}. The construction of mixed finite elements on quadrilaterals that maintain optimal order accuracy is considered in many papers. Most address only low order cases (see, e.g., \cite{Shen_1994, Shen_1992_phd, Boffi_Kikuchi_Schoberl_2006, Bochev_Ridzal_2008, Duan_Liang_2004, Kwak_Pyo_2011}). The exceptions we are aware of are the families of finite elements of Arnold, Boffi, and Falk (ABF$_r$(E))~\cite{ABF_2005}, Siqueira, Devloo, and Gomes \cite{Siqueira_Devloo_Gomes_2013}, and Arbogast and Correa (AC$_r(E)$ and AC$_r^{\textrm{red}}(E)$) \cite{Arbogast_Correa_2016}. The ABF elements are defined for rectangles and extended to quadrilaterals in the usual way (i.e., by mapping via the Piola transformation). They rectify the problem of poor divergence approximation by including more degrees of freedom in the space, so that approximation properties are maintained after Piola mapping. The spaces of \cite{Siqueira_Devloo_Gomes_2013} also involve the Piola map, but in a unique way. They also add shape functions to their space to obtain accuracy. The AC elements use a different strategy. These elements are defined by using vector polynomials directly on the element (i.e., without being mapped) and supplemented by two vector shape functions defined on a reference square and mapped via Piola. The AC spaces have minimal local dimension. In this paper, we introduce new families of \emph{direct} serendipity and mixed finite elements that have optimal approximation properties and maintain minimal local dimension. They are \emph{direct} in the sense that the shape functions contain a full set of polynomials defined directly on the element, as in the AC spaces. Because there are not enough degrees of freedom to achieve $H^1$ or $H(\textrm{div})$ conformity over meshes of quadrilaterals, two supplemental functions need to be added to each element, as is done for the AC spaces. The families of direct serendipity elements have the same number of degrees of freedom as the corresponding classical serendipity element, and they take the form \begin{equation}\label{eq:serendipityForm} {\cal{DS}}_r(E) = \mathbb P_r(E)\oplus\mathbb S_r^{\cal{DS}}(E),\quad r\geq2. \end{equation} Each family is defined by the choice of the two supplemental functions spanning $\mathbb S_r^{\cal{DS}}(E)$. We give a very general and explicit construction for these supplements. They can be defined directly on $E$, or they can be defined on $\hat E$ and mapped to $E$. There are two classes of families of direct mixed elements, which correspond to full and reduced $H(\textrm{div})$-approximation. For index $r$, a vector function is approximated to order $r+1$ accuracy, but the divergence of the vector is approximated to order $r$ or $r-1$ for full and reduced $H(\textrm{div})$-approximation, respectively. Each class of direct mixed elements has the same optimal number of degrees of freedom as the AC elements of that class. They take a form similar to \eqref{eq:serendipityForm}, which is \begin{equation} {\mathbf V}_r^{\textrm{red}}(E) = \mathbb P_r^2(E)\oplus\mathbb S_r^{\mathbf V}(E),\quad {\mathbf V}_r^{\textrm{full}}(E) = {\mathbf V}_r^{\textrm{red}}(E)\oplus{\mathbf x}\tilde\mathbb P_r(E),\quad r\geq1, \end{equation} where $\tilde\mathbb P_r$ are homogeneous polynomials of degree~$r$. Again, each family is defined by the choice of the two supplemental functions spanning $\mathbb S_r^{\mathbf V}(E)$. The serendipity and mixed families are related by de Rham theory: \begin{equation} \textrm{curl}\,\mathbb S_{r+1}^{\cal{DS}}(E)=\mathbb S_r^{\mathbf V}(E). \end{equation} We define one family of direct serendipity elements that is the precursor of the full and reduced AC spaces. We also define many \emph{fully} direct serendipity elements that use no mappings to define $\mathbb S_r^{{\cal{DS}}}(E)$, which in turn generate new full and reduced \emph{fully} direct mixed spaces that use no mappings whatsoever. Moreover, a second de Rham complex involving the gradient and curl operators provides new $H(\textrm{curl})=\big\{{\mathbf v}\in(L^2)^2:\textrm{curl}\,{\mathbf v}\in L^2\big\}$ elements as well. We set some basic notation in the next section. For any index $r\ge2$, we construct new families of direct serendipity elements in Sections~\ref{sec:DS}--\ref{sec:mappingSupp} for which the supplements either do not or do involve mappings, respectively. Through the de Rham theory, these lead to the AC and new direct mixed elements in Section~\ref{sec:deRham}. We discuss the stability and convergence properties of the new elements in Section~\ref{sec:properties}, and give some numerical results illustrating their performance in Section~\ref{sec:numerics}. A summary of our results and conclusions, as well as $H(\textrm{curl})$ elements, are given in the final section. \section{Some notation}\label{sec:notation} Let $\mathbb P_r(\omega)$ denote the space of polynomials of degree up to $r$ on $\omega\subset\mathbb R^d$, where $d=0$ (a point), $1$, or~$2$. Recall that \begin{equation}\label{eq:dimPo} \dim\mathbb P_r(\mathbb R^d) = \twoVec(r+d,d) = \frac{(r+d)!}{r!\,d!}. \end{equation} Let $\tilde\mathbb P_r(\omega)$ denote the space of homogeneous polynomials of degree $r$ on $\omega$. Then \begin{equation}\label{eq:dimTildePo} \dim\tilde\mathbb P_r(\mathbb R^d) = \twoVec(r+d-1,d-1) = \frac{(r+d-1)!}{r!\,(d-1)!}. \end{equation} Let the element $E\subset\mathbb R^2$ be a closed, nondegenerate, convex quadrilateral. By nondegenerate, we mean that $E$ does not degenerate to a triangle, line segment, or point. Let the reference element $\hat E$ be $[-1,1]^2$. Define the bilinear and bijective map ${\mathbf F}_{\!E}:\hat E\to E$ that maps the vertices of $\hat E$ to those of $E$. We identify ``vertical'' and ``horizontal'' pairs of opposite edges and number them consecutively as shown in Figure~\ref{fig:numbering}. Let $\nu_i$ denote the unit outer normal to edge $i$ (denoted $e_i$), $i=1,2,3,4$, and identify the vertices as ${\mathbf x}_{v,13}=e_1\cap e_3$, ${\mathbf x}_{v,14}=e_1\cap e_4$, ${\mathbf x}_{v,23}=e_2\cap e_3$, and ${\mathbf x}_{v,24}=e_2\cap e_4$. \begin{figure}[ht]\centering \setlength\unitlength{3.2pt} \begin{picture}(46,32)(-6,-6)\small \thicklines \put(0,0){\line(0,1){20}} \put(0,0){\line(1,0){20}} \put(20,0){\line(0,1){20}} \put(0,20){\line(1,0){20}} \put(0,10){\circle*{1}}\put(0,10){\vector(-1,0){4}} \put(10,0){\circle*{1}}\put(10,0){\vector(0,-1){4}} \put(20,10){\circle*{1}}\put(20,10){\vector(1,0){4}} \put(10,20){\circle*{1}}\put(10,20){\vector(0,1){4}} \put(-4.7,8.5){\makebox(0,0){$\hat\nu_1$}} \put(11.2,-5){\makebox(0,0){$\hat\nu_3$}} \put(25,8.3){\makebox(0,0){$\hat\nu_2$}} \put(8,25){\makebox(0,0){$\hat\nu_4$}} \put(0,0){\circle*{1}}\put(-2,-2){\makebox(0,0){$(-1,-1)$}} \put(20,0){\circle*{1}}\put(22,-2){\makebox(0,0){$(1,-1)$}} \put(20,20){\circle*{1}}\put(22,22){\makebox(0,0){$(1,1)$}} \put(0,20){\circle*{1}}\put(-2,22){\makebox(0,0){$(-1,1)$}} \put(10,10){\makebox(0,0){$\hat E$}} \put(1.7,14){\makebox(0,0){$\hat e_1$}} \put(6.5,-1.6){\makebox(0,0){$\hat e_3$}} \put(18.4,7){\makebox(0,0){$\hat e_2$}} \put(14,21.6){\makebox(0,0){$\hat e_4$}} \put(36,14){\makebox(0,0){$\overset{\text{\normalsize${\mathbf F}_{\!E}$}} {-\!\!\!-\!\!\!\longrightarrow}$}} \end{picture}\qquad \raisebox{6pt}{\begin{picture}(34.5,32)(-5,-6)\small \thicklines \put(0,0){\line(1,3){4}} \put(0,0){\line(1,0){24}} \put(24,0){\line(-1,4){4}} \put(4,12){\line(4,1){16}} \put(2,6){\circle*{1}}\put(2,6){\vector(-3,1){4.74}} \put(12,0){\circle*{1}}\put(12,0){\vector(0,-1){5}} \put(22,8){\circle*{1}}\put(22,8){\vector(4,1){4.85}} \put(12,14){\circle*{1}}\put(12,14){\vector(-1,4){1.21}} \put(-4,7.5){\makebox(0,0){$\nu_1$}} \put(10.8,-5.5){\makebox(0,0){$\nu_3$}} \put(28.4,9){\makebox(0,0){$\nu_2$}} \put(10.8,20){\makebox(0,0){$\nu_4$}} \put(0,0){\circle*{1}}\put(-1.8,-1.6){\makebox(0,0){${\mathbf x}_{v,13}$}} \put(24,0){\circle*{1}}\put(25.6,-1.6){\makebox(0,0){${\mathbf x}_{v,23}$}} \put(20,16){\circle*{1}}\put(21.6,17.6){\makebox(0,0){${\mathbf x}_{v,24}$}} \put(4,12){\circle*{1}}\put(2.4,13.6){\makebox(0,0){${\mathbf x}_{v,14}$}} \put(12,7){\makebox(0,0){$E$}} \put(4.6,8.8){\makebox(0,0){$e_1$}} \put(8,-1.2){\makebox(0,0){$e_3$}} \put(21.5,4){\makebox(0,0){$e_2$}} \put(16,16.5){\makebox(0,0){$e_4$}} \end{picture}} \caption{A reference element $\hat E=[-1,1]^{2}$ and quadrilateral $E$, with edges $\hat e_i$ and $e_i$, outer unit normals $\hat\nu_i$ and $\nu_i$, and vertices $(-1,-1)$ and ${\mathbf x}_{v,13}$, etc., respectively. \label{fig:numbering}} \end{figure} We define the linear polynomial $\lambda_i({\mathbf x})$ giving the distance of ${\mathbf x}\in\mathbb R^2$ to edge~$e_i$ in the normal direction as \begin{align} \label{eq:lambda_i} \lambda_i({\mathbf x}) &= - ({\mathbf x}-{\mathbf x}_i)\cdot\nu_i, \quad i=1,2,3,4, \end{align} where ${\mathbf x}_i\in e_i$ is any point on the edge. If ${\mathbf x}$ is in the interior of $E$, these functions are strictly positive, and each vanishes on the edge which defines it. We denote by $F_{\!E}^0$ the map taking a function $\hat\phi$ defined on $\hat E$ to a function $\phi$ defined on $E$ by the rule \begin{equation}\label{eq:defF0} \phi({\mathbf x}) = F_{\!E}^0(\hat\phi)({\mathbf x}) = \hat\phi(\hat{\mathbf x}), \end{equation} where ${\mathbf x}={\mathbf F}_{\!E}(\hat{\mathbf x})$. We denote by ${\mathbf F}_{\!E}^1$ the Piola map taking a vector function $\hat\pmb\psi$ defined on $\hat E$ to a vector function $\pmb\psi$ defined on $E$ by the rule \begin{equation}\label{eq:defF1} \pmb\psi({\mathbf x}) = \frac1{J_E}DF_{\!E}(\hat{\mathbf x})\,\hat\pmb\psi(\hat{\mathbf x}), \end{equation} where $DF_{\!E}(\hat{\mathbf x})$ is the Jacobian matrix of ${\mathbf F}_{\!E}$ and $J_E$ is its absolute determinant. Recall Ciarlet's definition~\cite{Ciarlet_1978} of a finite element. \begin{definition}[Ciarlet 1978]\label{defn:ciarlet} Let \begin{enumerate} \parskip=0pt\itemsep=0pt\itemOld[$1.$] $E\subset \mathbb R^d$ be a bounded closed set with nonempty interior and a Lipschitz continuous boundary, \parskip=0pt\itemsep=0pt\itemOld[$2.$] ${\cal P}$ be a finite-dimensional space of functions on $E$, and \parskip=0pt\itemsep=0pt\itemOld[$3.$] ${\cal N} = \{ N_1, N_2,\ldots, N_k \}$ be a basis for ${\cal P}'$. \end{enumerate} Then $(E, {\cal P}, {\cal N})$ is called a \emph{finite element.} \end{definition} Our task is to define the \emph{shape functions} ${\cal P}$ and the \emph{degrees of freedom} (DoFs) ${\cal N}$. The DoFs give a basis for ${\cal P}'$ provided that we have unisolvence of the shape functions (i.e., for $\phi\in{\cal P}$, $N_j(\phi)=0$ for all $j$ implies that $\phi=0$). To achieve optimal approximation properties, we will require that ${\cal P}\supset\mathbb P_r(E)$ for each index~$r$. That is, the polynomials will be directly included within the function space, and hence we call our new finite elements \emph{direct serendipity} and \emph{direct mixed} elements. Let $\Omega\subset\mathbb R^2$ be a polygonal domain, and let ${\cal T}_h$ be a conforming finite element partition or mesh of $\Omega$ into nondegenerate, convex quadrilaterals of maximal diameter $h>0$. To obtain approximation properties globally, we need to assume that the mesh is uniformly shape regular \cite[pp.~104--105]{Girault_Raviart_1986}, which means the following. For any $E\in{\cal T}_h$, denote by $T_i$, $i=1,2,3,4$, the subtriangle of $E$ with vertices being three of the four vertices of $E$. Define the parameters \begin{align} \label{eq:hE} h_E &= \text{diameter of }E, \\ \label{eq:rhoE} \rho_E &= 2\,\min_{1\leq i\leq 4}\{ \text{diameter of largest circle inscribed in }T_i \}. \end{align} Uniform shape regularity of the meshes means that there exists $\sigma_*>0$ such that the ratio $\displaystyle{\rho_E}/{h_E}\geq \sigma_*>0$ for all $E\in{\cal T}_h$, where $\sigma_*$ is independent of ${\cal T}_h$. The DoFs must be defined so that the shape functions on adjoining elements merge together. For serendipity spaces, we want the global space to reside in $H^1(\Omega)$, so the elements must merge continuously across each edge $e$. For mixed spaces, the vector variable must lie in $H(\textrm{div};\Omega)$, which means that the normal components (fluxes) of the vectors on an edge $e$ in adjacent elements must be continuous. \section{Fully direct serendipity elements in two space dimensions}\label{sec:DS} It is shown in~\cite{ABF_2002} that when $d=2$, the convergence of the linear serendipity finite element space ($r=1$) does not degenerate on quadrilaterals. The parametric serendipity element ${\cal S}_1(E)$ is the tensor product space of bilinear functions $\mathbb P_{1,1}(\hat E)$ on $\hat E$ mapped to $E$ by $F_{\!E}^0$, and, in fact, \begin{align}\label{eq:S1} {\cal S}_1(E) &= \textrm{span}\{F_{\!E}^0(1),F_{\!E}^0(\hat x),F_{\!E}^0(\hat y),F_{\!E}^0(\hat x\hat y)\}\\ \nonumber &= \textrm{span}\{1,x,y,F_{\!E}^0(\hat x\hat y)\} = \mathbb P_1(E)\oplus\textrm{span}\{F_{\!E}^0(\hat x\hat y)\} \end{align} has the form of a direct serendipity space. Therefore, we only develop our new direct serendipity finite elements ${\cal{DS}}_r(E)$ for indices $r \geq 2$. Our dual objectives are that $\mathbb P_r(E)\subset{\cal{DS}}_r(E)$ and that shape functions on adjoining elements merge continuously, i.e., so the space over $\Omega$ satisfies ${\cal{DS}}_r(\Omega)\subset H^1(\Omega)$. These objectives require us to consider the lower dimensional geometric objects within $E$ (as in \cite{Arnold_Awanou_2011}). The minimal number of DoFs associated to each lower dimensional object must correspond to the dimension of the polynomials that restrict to that object. These numbers are given in Table~\ref{tab:geometricDecomp}. A quadrilateral has 4 vertices, 4 edges, and one cell of dimension 0, 1, and 2, respectively. Each vertex requires $\dim\mathbb P(\mathbb R^0)=1$ DoF, each edge requires $\dim\mathbb P_{r-2}(\mathbb R)=r-1$ DoFs (not counting the vertices), and each cell requires $\dim\mathbb P_{r-4}(\mathbb R^2)=\twoVec(r-2,2)$ DoFs (not counting the edges and vertices). The total number of DoFs is then $D_r$, where \begin{align*} D_r = 4 + 4(r-1) + \frac12(r-2)(r-3) = \frac12(r+2)(r+1) + 2 = \dim\mathbb P_r(E) + 2, \end{align*} and so to define ${\cal{DS}}_r(E)$, we will supplement $\mathbb P_r(E)\subset{\cal{DS}}_r(E)$ with the span of two functions. We have many choices for the supplemental functions, the span of which is denoted $\mathbb S_r^{\cal{DS}}(E)$. Each choice gives rise to a distinct family of direct serendipity elements of index $r\geq2$; that is, the shape functions (${\cal P}$ in Definition~\ref{defn:ciarlet}) are \begin{equation}\label{eq:cDS=P+S} {\cal{DS}}_r(E)=\mathbb P_r(E)\oplus\mathbb S_r^{\cal{DS}}(E). \end{equation} In this section, we develop supplemental spaces that are unmapped (i.e., these new serendipity spaces are fully direct---even the supplements are defined directly on $E$). \begin{table}[ht] \caption{ Geometric decomposition and degrees of freedom (DoFs) associated to each geometric object of a quadrilateral for a serendipity element of index $r \geq 2$.\label{tab:geometricDecomp}} \centerline{\begin{tabular}{ccccc} Dimension & Object & Object & DoFs per & Total \\ & Name & Count & Object & DoFs \\ \hline 0 & vertex & 4 & 1 & 4\\ 1 & edge & 4 & $r-1$ & $4(r-1)$ \\ 2 & cell & 1 & $\frac12(r-2)(r-3)$ & $\frac12(r-2)(r-3)$\\ \end{tabular}} \end{table} We define the DoFs (${\cal N}$ in Definition~\ref{defn:ciarlet}) as a set of nodal functionals $N_j$ defined at a nodal point ${\mathbf x}_{n,j}$, i.e., \begin{equation}\label{eq:nodalFunctionals} {\cal N} = \{N_j : N_j(\phi) = \phi({\mathbf x}_{n,j})\text{ for all }\phi({\mathbf x}),\ j=1,2,\ldots,D_r\}. \end{equation} As depicted in Figure~\ref{fig:nodalDoFs}, for vertex DoFs, the nodal points are exactly the vertices ${\mathbf x}_{v,13}$, ${\mathbf x}_{v,14}$, ${\mathbf x}_{v,23}$, and ${\mathbf x}_{v,24}$ of~$E$. For edge DoFs, the nodal points plus vertices are equally distributed on each edge. There are $r-1$ nodal points on the interior of each egde, which can be denoted ${\mathbf x}_{e_i,j}$, $j=1,\ldots,r-1$ for nodal points that lie on edge $e_i$, $i=1,2,3,4$. The interior cell DoFs can be set, for example, on points of a triangle~$T$ strictly inside $E$, where the set of nodal points is the same as the nodes of the Lagrange element of order $r-4$ on the triangle~$T$. \begin{figure}[htbp]\centering{\setlength\unitlength{3.2pt} \begin{picture}(25,18)(-0.5,-0.5 \thicklines \put(5,16){\makebox(0,0){$r=2$}} \put(0,0){\line(1,3){4}} \put(0,0){\line(1,0){24}} \put(24,0){\line(-1,4){4}} \put(4,12){\line(4,1){16}} \put(2,6){\circle*{1}} \put(12,0){\circle*{1}} \put(22,8){\circle*{1}} \put(12,14){\circle*{1}} \put(0,0){\circle*{1}} \put(24,0){\circle*{1}} \put(20,16){\circle*{1}} \put(4,12){\circle*{1}} \end{picture}\quad \begin{picture}(25,18)(-0.5,-0.5 \thicklines \put(5,16){\makebox(0,0){$r=3$}} \put(0,0){\line(1,3){4}} \put(0,0){\line(1,0){24}} \put(24,0){\line(-1,4){4}} \put(4,12){\line(4,1){16}} \put(1.333,4){\circle*{1}} \put(2.667,8){\circle*{1}} \put(8,0){\circle*{1}} \put(16,0){\circle*{1}} \put(22.667,5.333){\circle*{1}} \put(21.333,10.667){\circle*{1}} \put(9.333,13.333){\circle*{1}} \put(14.667,14.667){\circle*{1}} \put(0,0){\circle*{1}} \put(24,0){\circle*{1}} \put(20,16){\circle*{1}} \put(4,12){\circle*{1}} \end{picture} \qua \begin{picture}(25,18)(-0.5,-0.5 \thicklines \put(5,16){\makebox(0,0){$r=4$}} \put(0,0){\line(1,3){4}} \put(0,0){\line(1,0){24}} \put(24,0){\line(-1,4){4}} \put(4,12){\line(4,1){16}} \put(1,3){\circle*{1}} \put(2,6){\circle*{1}} \put(3,9){\circle*{1}} \put(6,0){\circle*{1}} \put(12,0){\circle*{1}} \put(18,0){\circle*{1}} \put(23,4){\circle*{1}} \put(22,8){\circle*{1}} \put(21,12){\circle*{1}} \put(8,13){\circle*{1}} \put(12,14){\circle*{1}} \put(16,15){\circle*{1}} \put(12,7){\circle*{1}} \put(0,0){\circle*{1}} \put(24,0){\circle*{1}} \put(20,16){\circle*{1}} \put(4,12){\circle*{1}} \end{picture}\quad \begin{picture}(25,18)(-0.5,-0.5 \thicklines \put(5,16){\makebox(0,0){$r=5$}} \put(0,0){\line(1,3){4}} \put(0,0){\line(1,0){24}} \put(24,0){\line(-1,4){4}} \put(4,12){\line(4,1){16}} \put(0.8,2.4){\circle*{1}} \put(1.6,4.8){\circle*{1}} \put(2.4,7.2){\circle*{1}} \put(3.2,9.6){\circle*{1}} \put(4.8,0){\circle*{1}} \put(9.6,0){\circle*{1}} \put(14.4,0){\circle*{1}} \put(19.2,0){\circle*{1}} \put(20.8,12.8){\circle*{1}} \put(21.6,9.6){\circle*{1}} \put(22.4,6.4){\circle*{1}} \put(23.2,3.2){\circle*{1}} \put(7.2,12.8){\circle*{1}} \put(10.4,13.6){\circle*{1}} \put(13.6,14.4){\circle*{1}} \put(16.8,15.2){\circle*{1}} \put(12,9){\circle*{1}} \put(9,6){\circle*{1}} \put(15,6){\circle*{1}} \put(0,0){\circle*{1}} \put(24,0){\circle*{1}} \put(20,16){\circle*{1}} \put(4,12){\circle*{1}} \end{picture} } \caption{ The nodal points for the DoFs of the direct serendipity finite element for small $r$. \label{fig:nodalDoFs}} \end{figure} \subsection{Vertices}\label{sec:vertexDoFs} For the vertices, $r\ge2$, so we can define the shape functions \begin{equation}\label{eq:shape-vertices} \begin{alignedat}2 \phi_{v,13}({\mathbf x}) &= \lambda_2({\mathbf x})\lambda_4({\mathbf x}),&\quad\phi_{v,14}({\mathbf x}) &= \lambda_2({\mathbf x})\lambda_3({\mathbf x}),\\ \phi_{v,23}({\mathbf x}) &= \lambda_1({\mathbf x})\lambda_4({\mathbf x}),&\quad\phi_{v,24}({\mathbf x}) &= \lambda_1({\mathbf x})\lambda_3({\mathbf x}). \end{alignedat} \end{equation} These four functions are clearly linearly independent and unisolvent with respect to the vertex DoFs. All other shape functions will be defined so as to vanish at the vertices, so these four will be independent of the rest. \subsection{Interior cell}\label{sec:cell} For the entire cell $E$, we need interior shape functions only when $r\ge4$ (recall Table~\ref{tab:geometricDecomp}). We let the shape functions be defined by \begin{equation}\label{eq:shape-cell} \big\{\phi_{E,j}({\mathbf x}) : j=1,\ldots,\tfrac12(r-2)(r-3)\big\} = \lambda_1\lambda_2\lambda_3\lambda_4\mathbb P_{r-4}. \end{equation} These shape functions are linearly independent and vanish if the cell DoFs vanish. Moreover, these functions vanish on all four edges. Therefore, if unisolvent shape functions can be defined for the edge DoFs, then the set of all our shape functions will be unisolvent for the entire set of DoFs. \subsection{Edges}\label{sec:edges} We define distinct families of fully direct serendipity elements depending on the choice of the two supplemental functions used. These will be defined by a choice of four functions, which are oriented ``horizontally'' or ``vertically,'' in the sense that their zero sets are horizontal or vertical (as oriented by the bilinear map ${\mathbf F}_{\!E}$, see Figure~\ref{fig:numbering}). Two of the functions are linear polynomials, denoted $\lambda_H$ and $\lambda_V$. The other two functions should be bounded, and they are denoted $R_V$ and $R_H$. The supplemental space is then defined as \begin{equation} \label{eq:supplementSpace} \mathbb S_r^{{\cal{DS}}}(E) = \textrm{span}\{\lambda_3\lambda_4\lambda_H^{r-2}R_V,\lambda_1\lambda_2\lambda_V^{r-2}R_H\}. \end{equation} The choice of the linear function $\lambda_H$ is based on the edges $e_1$ and $e_2$. As shown in Figure~\ref{fig:lambda-h}, let ${\cal L}_1$ and ${\cal L}_2$ be the infinite lines containing the edges $e_1$ and $e_2$, respectively. When $e_1$ and $e_2$ are parallel, the only requirement for the choice of $\lambda_H$ is that its zero line intersects both ${\cal L}_1$ and ${\cal L}_2$. When $e_1$ and $e_2$ are not parallel, ${\cal L}_1$ and ${\cal L}_2$ intersect in a point ${\mathbf x}_{12}$. Then the only requirements for the choice of $\lambda_H$ is that $\lambda_H({\mathbf x}_{12})\ne0$ and that its zero line intersects both ${\cal L}_1$ and ${\cal L}_2$ on the half-lines emanating from ${\mathbf x}_{12}$ and either containing $e_1$ and $e_2$, respectively, or not containing $e_1$ and $e_2$, respectively (i.e., the zero line of $\lambda_H$ intersects the lines containing $e_1$ and $e_2$ either above or below ${\mathbf x}_{12}$). To be more precise in the case when $e_1$ and $e_2$ are not parallel, we can expand $\lambda_H\in\mathbb P_1(\mathbb R^2)$ in the basis defined by $\{1,\lambda_1,\lambda_2\}$, so there are constants $\alpha_H$, $\beta_H$, and $\gamma_H$ such that \begin{equation} \label{eq:lambda-h} \lambda_H({\mathbf x}) = \alpha_H\lambda_1({\mathbf x}) + \beta_H\lambda_2({\mathbf x}) + \gamma_H = -({\mathbf x} - {\mathbf x}_H)\cdot(\alpha_H\nu_1+\beta_H\nu_2),\quad e_1\nparallel e_2, \end{equation} where ${\mathbf x}_H$ is any point on the zero line. The requirements are that $\alpha_H$, $\beta_H$, and $\gamma_H$ are nonzero and that $\alpha_H$ and $\beta_H$ have the same sign. Without loss of generality, we may assume that $\alpha_H$ and $\beta_H$ are positive. In a similar way, $\lambda_V$ is chosen to intersect the lines extending $e_3$ and $e_4$, and when they are not parallel, either strictly to the left or right of the intersection point ${\mathbf x}_{34}$. When $e_3$ and $e_4$ are not parallel, \begin{equation} \label{eq:lambda-v} \lambda_V({\mathbf x}) = \alpha_V\lambda_1({\mathbf x}) + \beta_V\lambda_2({\mathbf x}) + \gamma_VSee = -({\mathbf x} - {\mathbf x}_V)\cdot(\alpha_V\nu_1+\beta_V\nu_2),\quad e_3\nparallel e_4, \end{equation} where ${\mathbf x}_V$ is any point on the zero line, $\alpha_V>0$, $\beta_V>0$, and $\gamma_V\ne0$. We remark that a simple choice is to take \begin{equation}\label{eq:simpleLambdaHV} \lambda_H^{\text{simple}} = \lambda_3 - \lambda_4 \quad\text{and}\quad \lambda_V^{\text{simple}} = \lambda_1 - \lambda_2. \end{equation} \begin{figure} \centering{\setlength\unitlength{1.8pt}\begin{picture}(70,45)(9,2)\small \thicklines \multiput(10,10)(1.0,1.0){37}{\circle*{0.2}} \multiput(75,5)(-1.0,1.0){42}{\circle*{0.2}} \put(20,20){\line(1,1){10}} \put(70,10){\line(-1,1){20}} \put(20,20){\line(5,-1){50}} \put(30,30){\line(1,0){20}} \put(40,40){\circle*{1}}\put(46,40){\makebox(0,0){${\mathbf x}_{12}$}} \put(23,27){\makebox(0,0){$e_1$}} \put(16,12){\makebox(0,0){${\cal L}_1$}} \put(62,22){\makebox(0,0){$e_2$}} \put(76,9){\makebox(0,0){${\cal L}_2$}} \put(50,20){\makebox(0,0){$E$}} \thinlines \put(12,16){\line(4,1){49}} \put(70,30){\rotatebox{13}{\makebox(0,0){$\lambda_H\!=\!0$}}} \put(40,5){\makebox(0,0){$e_1\nparallel e_2$}} \end{picture}\qquad\qquad\begin{picture}(50,50)(4,2)\small \thicklines \multiput(10,10)(0,1.0){37}{\circle*{0.2}} \multiput(30,5)(0,1.0){46}{\circle*{0.2}} \put(10,20){\line(0,1){20}} \put(30,10){\line(0,1){35}} \put(10,20){\line(2,-1){20}} \put(10,40){\line(4,1){20}} \put(7,30){\makebox(0,0){$e_1$}} \put(14,12){\makebox(0,0){${\cal L}_1$}} \put(33,27.5){\makebox(0,0){$e_2$}} \put(34,9){\makebox(0,0){${\cal L}_2$}} \put(20,28){\makebox(0,0){$E$}} \thinlines \put(4,16){\line(4,1){34}} \put(46,26){\rotatebox{13}{\makebox(0,0){$\lambda_H\!=\!0$}}} \put(20,5){\makebox(0,0){$e_1\parallel e_2$}} \end{picture}} \caption{Illustration of the zero lines of $\lambda_H$ and the point ${\mathbf x}_{12}$, if it exists.\label{fig:lambda-h}} \end{figure} The functions $R_V$ and $R_H$ are defined to satisfy the properties \begin{alignat}3 \label{eq:RVe} R_V({\mathbf x})|_{e_1} &= -\eta_V&&\quad\text{and}\quad&R_V({\mathbf x})|_{e_2} &= \xi_V,\\ \label{eq:RHe} R_H({\mathbf x})|_{e_3} &= -\eta_H&&\quad\text{and}\quad&R_H({\mathbf x})|_{e_4} &= \xi_H, \end{alignat} for some positive constants $\eta_V$, $\xi_V$, $\eta_H$, and $\xi_H$. For example, one choice is to let \begin{align} \label{eq:rational-v} R_V^{\text{simple}}({\mathbf x}) &= \frac{\lambda_1({\mathbf x}) - \lambda_2({\mathbf x})}{\xi_V^{-1}\lambda_1({\mathbf x}) + \eta_V^{-1}\lambda_2({\mathbf x})},\\ \label{eq:rational-h} R_H^{\text{simple}}({\mathbf x}) &= \frac{\lambda_3({\mathbf x}) - \lambda_4({\mathbf x})}{\xi_H^{-1}\lambda_3({\mathbf x}) + \eta_H^{-1}\lambda_4({\mathbf x})} \end{align} (note that the denominators do not vanish on $E$). We now define the shape functions associated with the edge DoFs. Let \begin{align} \label{eq:lambda12} \lambda_{12} &= \begin{cases} \alpha_H\xi_V\lambda_1-\beta_H\eta_V\lambda_2, & e_1\nparallel e_2,\\ \xi_V\lambda_1 - \eta_V\lambda_2, & e_1\parallel e_2, \end{cases}\\ \label{eq:lambda34} \lambda_{34} &= \begin{cases} \alpha_V\xi_H\lambda_3-\beta_V\eta_H\lambda_4, & e_3\nparallel e_4,\\ \xi_H\lambda_3 - \eta_H\lambda_4, & e_3\parallel e_4, \end{cases} \end{align} which also have horizontal and vertical zero sets, respectively. There are $2(r-1)$ shape functions associated to the edges $e_1$ and $e_2$, and they are \begin{alignat}2 \label{eq:shape-h} \phi_{H,j}({\mathbf x}) &= \lambda_3({\mathbf x})\lambda_4({\mathbf x})\lambda_H^{j}({\mathbf x}), &&\quad j=0,1,\ldots,r-2,\\ \label{eq:shape-hpm} \phi_{H,r-1+j}({\mathbf x}) &=\lambda_3({\mathbf x})\lambda_4({\mathbf x})\lambda_{13}({\mathbf x})\lambda_H^{j}({\mathbf x}), &&\quad j=0,1,\ldots,r-3,\\ \label{eq:shape-hr} \phi_{H,2r-3}({\mathbf x}) &=\lambda_3({\mathbf x})\lambda_4({\mathbf x})R_V({\mathbf x})\lambda_H^{r-2}({\mathbf x}). \end{alignat} In a similar way, we define shape functions associated with edges $e_3$ and $e_4$ to be \begin{alignat}2 \label{eq:shape-v} \phi_{V,j}({\mathbf x}) &= \lambda_1({\mathbf x})\lambda_2({\mathbf x})\lambda_V^{j}({\mathbf x}), &&\quad j=0,1,\ldots,r-2,\\ \label{eq:shape-vpm} \phi_{V,r-1+j}({\mathbf x}) &=\lambda_1({\mathbf x})\lambda_2({\mathbf x})\lambda_{24}({\mathbf x})\lambda_V^{j}({\mathbf x}), &&\quad j=0,1,\ldots,r-3,\\ \label{eq:shape-vr} \phi_{V,2r-3}({\mathbf x}) &=\lambda_1({\mathbf x})\lambda_2({\mathbf x})R_H({\mathbf x})\lambda_V^{r-2}({\mathbf x}). \end{alignat} The edge shape functions are regular polynomials of degree~$r$ except the last two functions in each direction, which may be rational functions, for example. However, all shape functions restrict to polynomials of degree~$r$ on the edges. \subsection{Unisolvence} The space of shape functions is \begin{align}\label{eq:shapeSpace} {\cal{DS}}_{r}(E) &= {\rm{span}}\big\{ \phi_{v,13}({\mathbf x}), \phi_{v,14}({\mathbf x}), \phi_{v,23}({\mathbf x}), \phi_{v,24}({\mathbf x}),\\ \nonumber &\qquad\qquad\phi_{H,j}({\mathbf x}), \phi_{V,j}({\mathbf x})\ (j=0,1,\ldots,2r-3),\\ \nonumber &\qquad\qquad\phi_{E,k}({\mathbf x})\ (k=1,\ldots,\tfrac12(r-2)(r-3))\big\}\\ \nonumber &\subset \mathbb P_r(E)\oplus\mathbb S_r^{\cal{DS}}(E). \end{align} In this subsection, we show the unisolvence of the degrees of freedom, which will then show that ${\cal{DS}}_{r}(E)=\mathbb P_r(E)\oplus\mathbb S_r^{\cal{DS}}(E)$ and complete the requirements of Ciarlet's Definition~\ref{defn:ciarlet} for ${\cal{DS}}_r(E)$ to be a well defined finite element. But first, we require a lemma. \begin{lemma}\label{lem:equ-space} Let $k\geq1$. For any $\eta_V>0$ and $\xi_V>0$, there exists a function $\tilde R_V$, defined by \eqref{eq:specialRV} and \eqref{eq:specialRVparallel}, with the properties \eqref{eq:RVe} such that the two function spaces \begin{align*} {\cal A}_k^1 &= {\rm{span}}\{ 1,\lambda_H,\ldots,\lambda_H^k, \lambda_{12},\lambda_H\lambda_{12}, \ldots, \lambda_H^{k-1}\lambda_{12},\lambda_H^k\tilde R_V\},\\ {\cal A}_k^2 &= {\rm{span}}\big\{\{1,\lambda_H,\ldots,\lambda_H^k\}\otimes \{1,\tilde R_V\}\big\} \end{align*} are identical. Moreover, for any $\eta_H>0$ and $\xi_H>0$, there exists an $\tilde R_H$, defined by \eqref{eq:specialRH}, with the properties \eqref{eq:RHe} such that the two function spaces \begin{align*} {\cal A}_k^3 &= {\rm{span}}\{ 1,\lambda_V,\ldots,\lambda_V^k, \lambda_{34},\lambda_V\lambda_{34}, \ldots, \lambda_V^{k-1}\lambda_{34},\lambda_V^k\tilde R_H \},\\ {\cal A}_k^4 &= {\rm{span}}\big\{\{1,\lambda_V,\ldots,\lambda_V^k\}\otimes\{1,\tilde R_H\}\big\} \end{align*} \end{lemma} are identical. \begin{proof} We show that ${\cal A}_k^1={\cal A}_k^2$. Assume that $e_1\nparallel e_2$, so that $\lambda_H$ has the form \eqref{eq:lambda-h} for some $\alpha_H>0$, $\beta_H>0$, and $\gamma_H\ne0$. We define $\tilde R_V$ satisfying \eqref{eq:RVe} as \begin{equation}\label{eq:specialRV} \tilde R_V = \frac{\alpha_H\xi_V\lambda_1-\beta_H\eta_V\lambda_2}{\alpha_H\lambda_1+\beta_H\lambda_2} = \frac{\lambda_{12}}{\lambda_H-\gamma_H},\quad e_1\nparallel e_2. \end{equation} Because $\alpha_H$ and $\beta_H$ are both positive, the denominator is not zero on $E$. We compute \begin{align}\label{eq:RVinSpan} \tilde R_V &= \frac{\lambda_{12}}{\lambda_H-\gamma_H} = - \frac{1}{\gamma_H}\lambda_{12} + \frac{1}{\gamma_H}\frac{\lambda_{12}\lambda_H}{\lambda_H-\gamma_H} = - \frac{1}{\gamma_H}\lambda_{12} + \frac{1}{\gamma_H}\lambda_H\tilde R_V. \end{align} We show that for any $\ell\geq0$, \begin{equation}\label{eq:RV-inSpanInduction} \tilde R_V = -\sum_{j=1}^{\ell}\frac{1}{\gamma_H^j}\lambda_{12}\lambda_H^{j-1} + \frac{1}{\gamma_H^{\ell}}\lambda_H^{\ell}\tilde R_V. \end{equation} The relation holds trivially for $\ell=0$, and \eqref{eq:RVinSpan} shows the result for $\ell=1$. Assuming by induction that \eqref{eq:RV-inSpanInduction} holds for $\ell-1$, we compute (using \eqref{eq:RVinSpan}) \begin{align*} \tilde R_V &= -\sum_{j=1}^{\ell-1}\frac{1}{\gamma_H^j}\lambda_{12}\lambda_H^{j-1} + \frac{1}{\gamma_H^{\ell-1}}\lambda_H^{\ell-1}\tilde R_V\\ &= -\sum_{j=1}^{\ell-1}\frac{1}{\gamma_H^j}\lambda_{12}\lambda_H^{j-1} + \frac{1}{\gamma_H^{\ell-1}}\lambda_H^{\ell-1} \bigg(-\frac{1}{\gamma_H}\lambda_{12} + \frac{1}{\gamma_H}\lambda_H\tilde R_V\bigg)\\ &= -\sum_{j=1}^{\ell}\frac{1}{\gamma_H^j}\lambda_{12}\lambda_H^{j-1} + \frac{1}{\gamma_H^{\ell}}\lambda_H^{\ell}\tilde R_V, \end{align*} and \eqref{eq:RV-inSpanInduction} holds for all $\ell$. Therefore, for any $0\leq m\leq k$, taking $\ell=k-m$, \begin{equation*} \lambda_H^m\tilde R_V = -\sum_{j=1}^{k-m}\frac{1}{\gamma_H^j}\lambda_{12}\lambda_H^{j-1+m} + \frac{1}{\gamma_H^{k-m}}\lambda_H^{k}\tilde R_V \in {\cal A}_k^1, \end{equation*} and we conclude that ${\cal A}_k^2\subset{\cal A}_k^1$. But clearly $\dim{\cal A}_k^2=2(k-1)$ and $\dim{\cal A}_k^1\leq2(k-1)$, and so the spaces are in fact equal. If $e_1\parallel e_2$, then $\lambda_{12}=\xi_V\lambda_1-\eta_V\lambda_2$ and $\lambda_1+\lambda_2=\delta_H>0$ is a constant. We define \begin{equation}\label{eq:specialRVparallel} \tilde R_V = \frac{\xi_V\lambda_1-\eta_V\lambda_2}{\lambda_1+\lambda_2} = \frac{1}{\delta_H}\lambda_{12},\quad e_1\parallel e_2, \end{equation} which satisfies \eqref{eq:RVe}. In this case, it is trivial that ${\cal A}_k^2={\cal A}_k^1$. By symmetry, ${\cal A}_k^3={\cal A}_k^4$, where now \begin{equation}\label{eq:specialRH} \tilde R_H = \begin{cases} \dfrac{\lambda_{34}}{\lambda_V-\gamma_V}, & e_3\nparallel e_4,\\ \dfrac{1\vphantom{H^H}}{\delta_V}\lambda_{34}, & e_3\parallel e_4, \end{cases} \end{equation} where $\lambda_3+\lambda_4=\delta_V>0$ is a constant when $e_3\parallel e_4$. \end{proof} \begin{theorem}\label{thm:unisolvent} Let $(E,{\cal{DS}}_r(E),{\cal N})$ be the $r$-th order direct serendipity finite element defined by \eqref{eq:shapeSpace}, i.e., by \eqref{eq:shape-vertices}, \eqref{eq:shape-cell}, and \eqref{eq:shape-h}--\eqref{eq:shape-vr}$)$ and \eqref{eq:nodalFunctionals}. If $\phi\in{\cal{DS}}_r(E)$ and $N_k(\phi)=0$, for all $k= 1,2,\ldots,D_r$, then $\phi=0$. Moreover, \begin{equation}\label{eq:DSR} {\cal{DS}}_r(E)=\mathbb P_r(E)\oplus\mathbb S_r^{\cal{DS}}(E), \end{equation} where $\mathbb S_r^{\cal{DS}}(E)$ is defined in \eqref{eq:supplementSpace}. \end{theorem} \begin{proof} As noted in Subsection~\ref{sec:cell}, by construction, the full set of DoFs are unisolvent for ${\cal{DS}}_r(E)$ if the edge DoFs are unisolvent for the edge shape functions $\phi_{H,j}$ and $\phi_{V,j}$. We temporarily replace $R_V$ by $\tilde R_V$ defined in \eqref{eq:specialRV} or \eqref{eq:specialRVparallel}. In that case, \begin{equation*} \{\phi_{H,j} : j=0,1,\ldots,r-2\} = \lambda_3\lambda_4{\cal A}_{r-2}^1 = \lambda_3\lambda_4{\cal A}_{r-2}^2, \end{equation*} by the lemma, and the representation of the space using ${\cal A}_{r-2}^2$ clearly shows that the DoFs on the edges $e_1$ and $e_2$ are unisolvent for $\phi_{H,j}$. That is, for an edge shape function $\phi({\mathbf x})$ with vanishing DoFs, we can use ${\cal A}_{r-2}^2$ to expand it as \begin{equation*} \phi({\mathbf x}) = \lambda_3({\mathbf x})\lambda_4({\mathbf x})\sum_{\ell=0}^{r-2}\big(a_\ell + R_V({\mathbf x})\,b_\ell\big)\lambda_H^\ell({\mathbf x}), \end{equation*} for some constants $a_\ell$ and $b_\ell$. On either edge $e_1$ or $e_2$ , $\phi({\mathbf x})$ is a polynomial of degree $r$, which must vanish due to the vanishing of the DoFs. Therefore, $a_\ell + R_V({\mathbf x})\,b_\ell$ must vanish on each edge, and we conclude that both $\eta_V a_\ell-\xi_V b_\ell=0$ and $\eta_V a_\ell+\xi_Vb_\ell=0$, i.e., $a_\ell=b_\ell=0$, and so $\phi({\mathbf x})=0$. It should be clear that we can return to the original $R_V$ and draw the same conclusion, since the DoFs of any function in the argument are unchanged by this replacement. (Of course, we no longer have that ${\cal A}_{r-2}^1={\cal A}_{r-2}^2$, but only that their DoFs agree.) Similarly we conclude unisolvence for $\phi_{V,j}$. We conclude that $\dim{\cal{DS}}_r(E)=D_r=\dim\mathbb P_r(E)+2$. Since we clearly added only two shape functions not in $\mathbb P_r(E)$, the full space of polynomials is contained in ${\cal{DS}}_r(E)$, and ${\cal{DS}}_r(E)=\mathbb P_r(E)\oplus\mathbb S_r^{\cal{DS}}(E)$. \end{proof} \subsection{Implementation as an $H^1$-Conforming Space}\label{sec:implementation} The global direct serendipity finite element space of index $r\ge2$ over ${\cal T}_h$ is \begin{equation} {\cal{DS}}_r=\{v_h\in{\cal C}^0(\Omega)\;:\; v_h|_E\in{\cal{DS}}_r(E)\ \forall E\in {\cal T}_h\}\subset H^1(\Omega). \end{equation} To implement $H^1$-conforming direct serendipity elements on the mesh ${\cal T}_h$ over $\Omega$, we need to find a proper basis for the finite element space, i.e., one that is continuous. We observe that the interior cell shape functions, after extension by zero, are in $H^1(\Omega)$ and so cause no difficulty. The vertex and edge shape functions with DoFs on the boundaries of the elements must be merged continuously. The simplest way to do this is to create a local nodal basis on every $E\in{\cal T}_h$ for these $4r$ shape functions. Let $\phi$ be any one of the shape functions of ${\cal{DS}}_r(E)$ defined above in \eqref{eq:shape-vertices}, \eqref{eq:shape-cell}, and \eqref{eq:shape-h}--\eqref{eq:shape-vr}. To reduce rounding errors in numerical calculations, we scale it so \begin{align*} \phi \quad\text{is replaced by}\quad \frac{\phi}{d_E^{n_\phi}}, \end{align*} where $d_E = \sqrt{|E|}$ and $n_\phi$ is the degree of $\phi$ when it is a polynomial and $n_{\phi_{H,2r-3}}=n_{\phi_{V,2r-3}}=r$. In general, we can find the local basis $\{\varphi_1,\ldots,\varphi_{4r}\}$ by solving a small local linear system. We order the shape functions with the vertex DoFs first ($\phi_1$ to $\phi_4$), the horizontal edge DoFs on $e_1$ and $e_2$ next ($\phi_5$ to $\phi_{2r+2}$), and finally the vertical edge DoFs on $e_3$ and $e_4$ ($\phi_{2r+3}$ to $\phi_{4r}$). We also order the DoFs similarly. Construct the $4r\times 4r$ matrix ${\bf A}=(a_{ij})$ of the DoFs, i.e., ${\bf A}_{ij} = N_j(\phi_i)$ for all $i,j\leq 4r$. This matrix has a simple block structure based on the DoFs on the vertices, edges $e_1$ and $e_2$, and edges $e_3$ and $e_4$, which is \begin{align} \bf{A} = \left( \begin{matrix} {\bf{A}}_{11} & {\bf{A}}_{12} & {\bf{A}}_{13}\\ {\bf0} & {\bf{A}}_{22} & {\bf0}\\ {\bf0} & {\bf{0}} & {\bf{A}}_{33} \end{matrix} \right), \end{align} where ${\bf{A}}_{11}$ is of size $4\times 4$ and ${\bf{A}}_{22}$ and ${\bf{A}}_{33}$ are of size $2(r-1)\times 2(r-1)$. From Theorem~\ref{thm:unisolvent} (unisolvence), we know that $\bf A$ is invertible. Let ${\bf{A}}^{-1}={\bf{B}}=(b_{ij})$ and define \begin{equation*} \varphi_i = \displaystyle\sum_{j=1}^{4r} b_{ij}\phi_j \quad\implies\quad N_k(\varphi_i) = \sum_{j=1}^{4r}b_{ij} N_k(\phi_j) =\sum_{j=1}^{4r}b_{ij} a_{jk} = \delta_{ik}, \end{equation*} and we have our nodal basis. Graphical depiction of the basis in special cases can be found in \cite{Tao_2017_phd, Arbogast_Tao_2017_serendipity}. Visually, there is nothing unusual about these basis functions. If we take the choice outlined in Lemma~\ref{lem:equ-space}, we can write down the nodal basis explicitly, using the facts that ${\cal A}_{r-2}^1={\cal A}_{r-2}^2$ and ${\cal A}_{r-2}^3={\cal A}_{r-2}^4$. The edge basis functions become, for $j=1,\ldots,r-1$, \begin{align} \label{eq:ds-baseE1} \varphi_{e_1,j}({\mathbf x}) &= \frac{\lambda_3({\mathbf x})\lambda_4({\mathbf x})}{\lambda_3({\mathbf x}_{e_1,j})\lambda_4({\mathbf x}_{e_1,j})} \frac{\xi_V - \tilde R_V({\mathbf x})}{\xi_V + \eta_V} \prod_{\shortstack{\scriptsize$k=1$\\\scriptsize$k\ne j$}}^{r-1} \frac{\lambda_H({\mathbf x}) - \lambda_H({\mathbf x}_{e_1,k})}{\lambda_H({\mathbf x}_{e_1,j})-\lambda_H({\mathbf x}_{e_1,k})},\\ \varphi_{e_2,j}({\mathbf x}) \label{eq:ds-baseE2} &= \frac{\lambda_3({\mathbf x})\lambda_4({\mathbf x})}{\lambda_3({\mathbf x}_{e_2,j})\lambda_4({\mathbf x}_{e_2,j})} \frac{\tilde R_V({\mathbf x}) + \eta_V}{\xi_V + \eta_V} \prod_{\shortstack{\scriptsize$k=1$\\\scriptsize$k\ne j$}}^{r-1} \frac{\lambda_H({\mathbf x}) - \lambda_H({\mathbf x}_{e_2,k})}{\lambda_H({\mathbf x}_{e_2,j})-\lambda_H({\mathbf x}_{e_2,k})},\\ \varphi_{e_3,j}({\mathbf x}) \label{eq:ds-baseE3} &= \frac{\lambda_1({\mathbf x})\lambda_2({\mathbf x})}{\lambda_1({\mathbf x}_{e_3,j})\lambda_2({\mathbf x}_{e_3,j})} \frac{\xi_H - \tilde R_H({\mathbf x})}{\xi_H + \eta_H} \prod_{\shortstack{\scriptsize$k=1$\\\scriptsize$k\ne j$}}^{r-1} \frac{\lambda_V({\mathbf x}) - \lambda_V({\mathbf x}_{e_3,k})}{\lambda_V({\mathbf x}_{e_3,j})-\lambda_V({\mathbf x}_{e_3,k})},\\ \varphi_{e_4,j}({\mathbf x}) \label{eq:ds-baseE4} &= \frac{\lambda_1({\mathbf x})\lambda_2({\mathbf x})}{\lambda_1({\mathbf x}_{e_4,j})\lambda_2({\mathbf x}_{e_4,j})} \frac{\tilde R_H({\mathbf x}) + \eta_H}{\xi_H + \eta_H} \prod_{\shortstack{\scriptsize$k=1$\\\scriptsize$k\ne j$}}^{r-1} \frac{\lambda_V({\mathbf x}) - \lambda_V({\mathbf x}_{e_4,k})}{\lambda_V({\mathbf x}_{e_4,j})-\lambda_V({\mathbf x}_{e_4,k})}. \end{align} The vertex basis function $\varphi_{v,13}({\mathbf x})$ can be computed by first defining \begin{equation*} \tilde\phi_{v,13}({\mathbf x}) = \frac{\lambda_2({\mathbf x})\lambda_4({\mathbf x})}{\lambda_2({\mathbf x}_{v,13})\lambda_4({\mathbf x}_{v,13})} \end{equation*} and then defining \begin{align} \label{eq:ds-baseV13} \varphi_{v,13}({\mathbf x}) &= \tilde\phi_{v,13}({\mathbf x}) - \sum_{k\in\{1,3\}}\sum_{j=1}^{r-1} \tilde\phi_{v,13}({\mathbf x}_{e_k,j})\,\varphi_{e_k,j}({\mathbf x}). \end{align} The basis functions $\varphi_{v,14}({\mathbf x})$, $\varphi_{v,23}({\mathbf x})$, and $\varphi_{v,24}({\mathbf x})$ can be defined similarly. \section{Serendipity supplements based on mapping from a reference element}\label{sec:mappingSupp} The supplemental functions $\mathbb S_r^{\cal{DS}}(E)$ used in the definition of ${\cal{DS}}_r(E)$ in \eqref{eq:shapeSpace} can be defined in terms of the bilinear map ${\mathbf F}_{\!E}:\hat E\to E$ and $F_{\!E}^0$ discussed in Section~\ref{sec:notation}. For example, since $\hat E=[-1,1]^2$, one can define \begin{align} \label{eq:rational-h-v-piola} R_V({\mathbf x}) = F_{\!E}^0(\hat x_1) \quad\text{and}\quad R_H({\mathbf x}) = F_{\!E}^0(\hat x_2), \end{align} for which $\eta_V=\xi_V=\eta_H=\xi_H=1$. We can also use the map to define the entire supplemental functions \eqref{eq:rational-h} and \eqref{eq:rational-v} themselves. For example, we can substitute the definitions \begin{align} \label{eq:shape-hr-map} \phi_{H,2r-3}({\mathbf x}) &= F_{\!E}^0\big((1-\hat x_2^2)\hat x_1\hat x_2^{r-2}\big),\\ \label{eq:shape-vr-map} \phi_{V,2r-3}({\mathbf x}) &= F_{\!E}^0\big((1-\hat x_1^2)\hat x_2\hat x_1^{r-2}\big), \end{align} giving direct serendipity elements with mapped supplements. We must show unisolvence with this substitution. We proceed to show this property for edges $e_1$ and $e_2$. The other two edges will then have this property by symmetry. Easily, \begin{equation*} \phi_{H,2r-3}({\mathbf x}) = F_{\!E}^0(1-\hat x_2^2)\,F_{\!E}^0(\hat x_1)\,\big(F_{\!E}^0(\hat x_2)\big)^{r-2} = F_{\!E}^0(1-\hat x_2^2)\,R_V\,(\lambda_H^*)^{r-2}, \end{equation*} where $R_V$ is defined in \eqref{eq:rational-h-v-piola} and \begin{equation*} \lambda_H^*=F_{\!E}^0(\hat x_2) \end{equation*} is a \emph{nonlinear} function. Because ${\mathbf F}_{\!E}$ is a bilinear map, on the edges $e_1$ and $e_2$, $\lambda_H^*$ is linear and $F_{\!E}^0(1-\hat x_2^2)$ is quadratic. However, these may be \emph{different} linear and quadratic functions on each edge. The function $\lambda_H^*=F_{\!E}^0(\hat x_2)$ has the zero set being the line joining the center of $e_1$ to the center of $e_2$. Let ${\mathbf x}_H$ be any point on this line and $\nu_H$ denote a unit normal to the line. Define \begin{equation}\label{eq:lambdaHnuH} \lambda_H({\mathbf x}) = -({\mathbf x} - {\mathbf x}_H)\cdot\nu_H. \end{equation} If $e_1$ and $e_2$ are not parallel, then there exist (up to sign, so without loss of generality) $\alpha_H>0$, $\beta_H>0$, and $\gamma_H\ne0$ such that \begin{equation*} \lambda_H({\mathbf x}) = \alpha_H\lambda_1({\mathbf x}) + \beta_H\lambda_2({\mathbf x}) + \gamma_H. \end{equation*} If $e_1$ and $e_2$ are parallel, set $\alpha_H=\beta_H=1$ and $\gamma_H=0$ to obtain the same representation of $\lambda_H$. In either case, we define \begin{equation*} \lambda_{12} = \alpha_H\lambda_1 - \beta_H\lambda_2 \quad\text{and}\quad \tilde R_V=\frac{\alpha_H\lambda_1 - \beta_H\lambda_2}{\alpha_H\lambda_1 + \beta_H\lambda_2}. \end{equation*} These functions satisfy the requirements of Lemma~\ref{lem:equ-space}. Because $\lambda_H^*$ is linear on $e_1$ and $e_2$, there are nonzero constants $a$ and $b$ of the same sign such that \begin{equation} \lambda_H^*\big|_{e_1} = a\lambda_H\big|_{e_1} \quad\text{and}\quad \lambda_H^*\big|_{e_2} = b\lambda_H\big|_{e_2}. \end{equation} Therefore, on the sides $e_1\cup e_2$, \begin{align}\label{eq:piolaDS-edgeFcn-equ} &\lambda_H({\mathbf x})^{r-2}\tilde R_V({\mathbf x})\\ \nonumber &\quad= \frac{a^{r-2} - b^{r-2}}{a^{r-2} + b^{r-2}}\lambda_H({\mathbf x})^{r-2} + \frac{2}{a^{r-2} + b^{r-2}}(\lambda_H({\mathbf x})^*)^{r-2}\tilde R_V({\mathbf x}), \quad{\mathbf x}\in e_1\cup e_2. \end{align} We define the sets \begin{align*} {\cal A}^1 &= {\rm{span}}\{ 1,\lambda_H,\ldots,\lambda_H^{r-2}, \lambda_{12},\lambda_H\lambda_{12}, \ldots, \lambda_H^{r-3}\lambda_{12},\lambda_H^{r-2}\tilde R_V\},\\ {\cal A}^2 &= {\rm{span}}\big\{\{1,\lambda_H,\ldots,\lambda_H^{r-2}\} \otimes\{1,\tilde R_V\}\big\},\\ {\cal A}^{1,\sim} &= {\rm{span}}\{ 1,\lambda_H,\ldots,\lambda_H^{r-2}, \lambda_{12},\lambda_H\lambda_{12}, \ldots, \lambda_H^{r-3}\lambda_{12},(\lambda_H^{*})^{r-2}\tilde R_V\},\\ {\cal A}^{1,*} &= {\rm{span}}\{ 1,\lambda_H,\ldots,\lambda_H^{r-2}, \lambda_{12},\lambda_H\lambda_{12}, \ldots, \lambda_H^{r-3}\lambda_{12},(\lambda_H^{*})^{r-2}R_V\}, \end{align*} The later set ${\cal A}^{1,*}$, times $F_{\!E}^0(1-\hat x_2^2)$, defines the shape functions for our direct serendipity finite element based on the mapped supplement \eqref{eq:shape-hr-map}. However, we can replace $R_V$ by $\tilde R_V$, since we consider only DoFs. That is, ${\cal A}^{1,\sim}$ and ${\cal A}^{1,*}$ are equivalent for our purposes. Lemma~\ref{lem:equ-space} shows that ${\cal A}^1={\cal A}^2$, which is unisolvent. Moreover, \eqref{eq:piolaDS-edgeFcn-equ} shows that ${\cal A}^1\subset{\cal A}^{1,\sim}$, which have the same dimension and so are equal, and hence we have unisolvence. Unisolvence is maintained after multiplication by $F_{\!E}^0(1-\hat x_2^2)$, since this modification concerns the fact that there are $c_1>0$ and $c_2>0$ so that \begin{equation}\label{eq:mappedBubblePart} F_{\!E}^0(1-\hat x_2^2)\big|_{e_j}=c_j\lambda_3\lambda_4,\quad j=1,2. \end{equation} We conclude that the direct serendipity element with the mapped supplements \eqref{eq:shape-hr-map}--\eqref{eq:shape-vr-map}, i.e., \begin{equation}\label{eq:DSmap} {\cal{DS}}_r^{\text{map}}(E) = \mathbb P_r(E)\oplus\textrm{span}\{F_{\!E}^0\big((1-\hat x_2^2)\hat x_1\hat x_2^{r-2}\big), F_{\!E}^0\big((1-\hat x_1^2)\hat x_2\hat x_1^{r-2}\big)\}, \end{equation} is well defined. \section{The de Rham complex and mixed finite elements}\label{sec:deRham} The de Rham complex of interest here is \begin{equation}\label{eq:deRham} \mathbb R \DOTSB\lhook\joinrel\longrightarrow H^1 \overset{\textrm{curl}\,}{\longlongrightarrow} H(\textrm{div}) \overset{\textrm{div}\,}{\longlongrightarrow} L^2 \longrightarrow 0, \end{equation} where the curl (or rot) of a scalar function $\phi({\mathbf x})=\phi(x_1,x_2)$ is $\textrm{curl}\,\phi = \bigg(\dfrac{\partial\phi}{\partial x_2},-\dfrac{\partial\phi}{\partial x_1}\bigg)$. From right to left, the image of one linear map is the kernel of the next. On rectangular elements, it is known \cite{Arnold_Awanou_2011,Arnold_Awanou_2014} that the serendipity space ${\cal S}_{r+1}$ is the precursor of the Brezzi-Douglas-Marini space BDM$_r$ \cite{BDM_1985} for $r\geq1$; that is, on the reference square $\hat E$, \eqref{eq:deRhamBDM} holds. \subsection{Full and reduced AC spaces}\label{sec:ACspaces} We have the following extension of \eqref{eq:deRhamBDM} to quadrilateral elements $E$. The direct serendipity spaces ${\cal{DS}}_r^{\text{map}}$ using the mapped supplements \eqref{eq:shape-hr-map}--\eqref{eq:shape-vr-map} is the precursor of the reduced $H(\textrm{div})$-approximating Arbogast-Correa space AC$_r^{\textrm{red}}$ \cite{Arbogast_Correa_2016}, $r\geq1$, defined on meshes of convex quadrilaterals: \begin{equation}\label{eq:deRhamACred} \mathbb R \DOTSB\lhook\joinrel\longrightarrow {\cal{DS}}_{r+1}^{\text{map}}(E) \overset{\textrm{curl}\,}{\longlongrightarrow} \textrm{AC}_r^{\textrm{red}}(E) \overset{\textrm{div}\,}{\longlongrightarrow} \mathbb P_{r-1}(E) \longrightarrow 0. \end{equation} Moreover, the full $H(\textrm{div})$-approximating space AC$_r$, for $r\geq1$, satisfies \begin{equation}\label{eq:deRhamAC} \mathbb R \DOTSB\lhook\joinrel\longrightarrow {\cal{DS}}_{r+1}^{\text{map}}(E) \overset{\textrm{curl}\,}{\longlongrightarrow} \textrm{AC}_r(E) \overset{\textrm{div}\,}{\longlongrightarrow} \mathbb P_{r}(E) \longrightarrow 0. \end{equation} This observation is clear once one realizes three sets of facts. First, the direct serendipity elements based on \eqref{eq:shape-hr-map}--\eqref{eq:shape-vr-map} have the structure \begin{align} {\cal{DS}}_{r+1}^{\text{map}}(E) &= \mathbb P_{r+1}(E)\oplus\mathbb S_{r+1}^{{\cal{DS}},\text{map}}(E),\\ \mathbb S_{r+1}^{{\cal{DS}},\text{map}}(E) &= \textrm{span}\big\{F_{\!E}^0\big((1-\hat x_2^2)\hat x_1\hat x_2^{r-1}\big), F_{\!E}^0\big((1-\hat x_1^2)\hat x_2\hat x_1^{r-1}\big)\big\}. \end{align} Second, the AC elements have the structure \begin{align} \label{eq:AC} \textrm{AC}_r(E) &= \textrm{AC}_r^{\textrm{red}}(E)\oplus{\mathbf x}\tilde\mathbb P_{r}(E),\\ \label{eq:ACred} \textrm{AC}_r^{\textrm{red}}(E) &= \mathbb P_r^2(E)\oplus\mathbb S_{r+1}^{\textrm{AC}}(E),\\ \label{eq:ACsupp} \mathbb S_{r+1}^{\textrm{AC}}(E) &= \textrm{span}\big\{{\mathbf F}_{\!E}^1\,\textrm{curl}\big((1-\hat x_2^2)\hat x_1\hat x_2^{r-1}\big), {\mathbf F}_{\!E}^1\,\textrm{curl}\big((1-\hat x_1^2)\hat x_2\hat x_1^{r-1}\big)\big\}, \end{align} where ${\mathbf F}_{\!E}^1$ is the Piola mapping from $E$ to $\hat E$. Finally, we have the fairly well-known helmholtz-like decomposition (see, e.g., \cite{Arbogast_Correa_2016}) \begin{equation} \mathbb P_{r}^2(E) = \textrm{curl}\,\mathbb P_{r+1}^2(E)\oplus{\mathbf x}\tilde\mathbb P_{r-1}(E), \end{equation} the relation between the $\textrm{curl}$ operator and the bilinear and Piola maps \begin{equation} \textrm{curl}\,F_{\!E}^0 = {\mathbf F}_{\!E}^1\,\textrm{curl}, \end{equation} and the fact that the $\textrm{div}$ operator takes ${\mathbf x}\mathbb P_{k}$ one-to-one and onto $\mathbb P_{k}$ for any $k\geq0$. Now we see that \begin{align} \textrm{curl}\,\mathbb S_{r+1}^{{\cal{DS}},\text{map}}(E) &= \textrm{span}\big\{\textrm{curl}\,F_{\!E}^0\big((1-\hat x_2^2)\hat x_1\hat x_2^{r-1}\big), \textrm{curl}\,F_{\!E}^0\big((1-\hat x_1^2)\hat x_2\hat x_1^{r-1}\big)\big\}\\ \nonumber &= \textrm{span}\big\{{\mathbf F}_{\!E}^1\,\textrm{curl}\big((1-\hat x_2^2)\hat x_1\hat x_2^{r-1}\big), {\mathbf F}_{\!E}^1\,\textrm{curl}\big((1-\hat x_1^2)\hat x_2\hat x_1^{r-1}\big)\big\}\\ \nonumber &= \mathbb S_r^{\textrm{AC}}(E), \end{align} \vspace*{-20pt}\par\noindent and so \begin{equation} \textrm{curl}\,{\cal{DS}}_{r+1}^{\text{map}}(E) = \textrm{curl}\,\mathbb P_{r+1}(E)\oplus\mathbb S_r^{\textrm{AC}}(E) \end{equation} is in the kernel of the operator $\textrm{div}$. Finally, \begin{align} \textrm{AC}_r^{\textrm{red}}(E) &= \textrm{curl}\,\mathbb P_{r+1}(E)\oplus\mathbb S_r^{\textrm{AC}}(E)\oplus{\mathbf x}\tilde\mathbb P_{r-1}(E),\\ \textrm{AC}_r(E) &= \textrm{curl}\,\mathbb P_{r+1}(E)\oplus\mathbb S_r^{\textrm{AC}}(E)\oplus{\mathbf x}\tilde\mathbb P_{r}(E), \end{align} satisfy the properties of the de Rham complex \eqref{eq:deRhamACred}--\eqref{eq:deRhamAC}. We remark that it is easy to check that ${\cal S}_1(E)$ (see \eqref{eq:S1}) precedes the element AC$_0$(E) in the de Rham sequence \eqref{eq:deRhamAC}. \subsection{Direct mixed finite elements on quadrilaterals}\label{sec:newMixed} The de Rham theory provides a way of constructing a mixed finite element space ${\mathbf V}_r$ based on a well defined direct serendipity space. Tangential derivatives of functions in ${\cal{DS}}_{r+1}(E)$ along the edges map by the $\textrm{curl}$ operator to normal derivatives; that is, if we define the unit tangential vector \begin{equation} \label{eq:tau} \tau_i=(-\nu_{i,2},\nu_{i,1})\quad\text{on }e_i, \end{equation} then for $\phi\in{\cal{DS}}_{r+1}(E)$, \begin{equation}\label{eq:tangentialToNormalDer} \nabla\phi\cdot\tau_i\big|_{e_i} = \textrm{curl}\,\phi\cdot\nu_i\big|_{e_i}. \end{equation} Since $\textrm{curl}\,{\cal{DS}}_{r+1}(E)$ spans $\mathbb P_{r}(e_i)$ independently of the other sides, the same is true of the normal derivatives of ${\mathbf V}_r(E)$. In fact, for $r\geq1$, we have de Rham complexes for both full and reduced direct $H(\textrm{div})$-approximating mixed elements: \begin{align} \label{eq:deRhamV} \mathbb R \DOTSB\lhook\joinrel\longrightarrow {\cal{DS}}_{r+1}(E) \overset{\textrm{curl}\,}{\longlongrightarrow} &{\mathbf V}_r^{\textrm{full}}(E) \overset{\textrm{div}\,}{\longlongrightarrow} \mathbb P_{r}(E) \longrightarrow 0,\\ \label{eq:deRhamVred} \mathbb R \DOTSB\lhook\joinrel\longrightarrow {\cal{DS}}_{r+1}(E) \overset{\textrm{curl}\,}{\longlongrightarrow} &{\mathbf V}_r^{\textrm{red}}(E) \overset{\textrm{div}\,}{\longlongrightarrow} \mathbb P_{r-1}(E) \longrightarrow 0, \end{align} for any variant of our new direct serendipity spaces. To see this fact, we need to decompose ${\mathbf V}_r(E)$ (i.e, ${\mathbf V}_r^{\text{full}}(E)$ or ${\mathbf V}_r^{\text{red}}(E)$). According to \cite{Arbogast_Correa_2016}, a reduced or full $H(\textrm{div})$-approximating mixed finite element space defined directly on a quadrilateral $E$ of minimal local dimension takes the form (${\cal P}$ in Definition~\ref{defn:ciarlet}) \begin{alignat}2 \label{eq:generalAC} {\mathbf V}_r^{\text{full}}(E) &= \mathbb P_r^2(E)\oplus{\mathbf x}\tilde\mathbb P_{r}\oplus\mathbb S_r^{\mathbf V}(E) &&= \textrm{curl}\,\mathbb P_{r+1}(E)\oplus{\mathbf x}\mathbb P_{r}\oplus\mathbb S_r^{\mathbf V}(E),\\ \label{eq:generalACred} {\mathbf V}_r^{\text{red}}(E) &= \mathbb P_r^2(E)\oplus\mathbb S_r^{\mathbf V}(E) &&= \textrm{curl}\,\mathbb P_{r+1}(E)\oplus{\mathbf x}\mathbb P_{r-1}\oplus\mathbb S_r^{\mathbf V}(E), \end{alignat} where the choice of $\mathbb S_r^{\mathbf V}(E)$ is given by taking \eqref{eq:ACsupp}. However, it is noted that other supplemental functions could be used \cite[near (3.15)]{Arbogast_Correa_2016}. Their normal components must lie in $\mathbb P_r(e_i)$ on each edge $e_i$ and, if they are mapped by the Piola transform, they must contain a nontrivial component of the DoFs of $\textrm{curl}\,\hat x^{r+1}\hat y$ and $\textrm{curl}\,\hat x\hat y^{r+1}$. As given in \cite{Arbogast_Correa_2016}, the DoFs (${\cal N}$ in Definition~\ref{defn:ciarlet}) for $\pmb\psi\in{\mathbf V}_r^{\text{full}}(E)$ ($s=r$) or $\pmb\psi\in{\mathbf V}_r^{\text{red}}(E)$ ($s=r-1$) are \begin{alignat}2 \label{eq:vdofEdge} &\int_{e_i}\pmb\psi\cdot\nu_i\,p\,dx,&&\quad\forall p\in\mathbb P_r(e_i),\ i=1,2,3,4,\\ \label{eq:vdofDiv} &\int_{E}\pmb\psi\cdot\nabla q\,dx,&&\quad\forall q\in\mathbb P_{s}(E),\\ \label{eq:vdofCell} &\int_{E}\pmb\psi\cdot{\mathbf v}\,dx,&&\quad\forall p\in\mathbb B_r^{\mathbf V}(E), \end{alignat} where the $H(\textrm{div})$ bubble functions are \begin{equation} \label{eq:vBubbles} \mathbb B_r^{\mathbf V}(E) = \textrm{curl}\big(\lambda_1\lambda_2\lambda_3\lambda_4\mathbb P_{r-3}(E)\big). \end{equation} The DoFs \eqref{eq:vdofDiv} are determined entirely by the part of ${\mathbf V}_r^{\text{full}}(E)$ or ${\mathbf V}_r^{\text{red}}(E)$ in the decomposition \eqref{eq:generalAC}--\eqref{eq:generalACred} that is ${\mathbf x}\mathbb P_{r}$ or ${\mathbf x}\mathbb P_{r-1}$, respectively. The DoFs \eqref{eq:vdofCell} correspond to the interior cell direct serendipity DoFs. In fact, $\mathbb B_r^{\mathbf V}(E)$ is exactly the $\textrm{curl}$ of the span of the cell shape functions \eqref{eq:shape-cell} for ${\cal{DS}}_{r+1}(E)$. The DoFs \eqref{eq:vdofEdge} correspond to the edge and vertex DoFs of ${\cal{DS}}_{r+1}(E)$. We can use any of our direct serendipity spaces to define the supplemental space $\mathbb S_r^{\mathbf V}(E)$ needed by ${\mathbf V}_r^{\text{full}}(E)$ or ${\mathbf V}_r^{\text{red}}(E)$. The rest of the space is composed of polynomials, and so need not be defined by ${\cal{DS}}_{r+1}(E)$ through the $\textrm{curl}$ operator, although this strategy could be used to help construct a basis for ${\mathbf V}_r(E)$ respecting the DoFs. On $E$ when $r\geq1$, our new mixed spaces use the supplemental space \begin{equation} \mathbb S_{r}^{\mathbf V}(E)=\textrm{curl}\,\mathbb S_{r+1}^{\cal{DS}}(E). \end{equation} In particular, as we saw, $\textrm{curl}\,\mathbb S_{r+1}^{{\cal{DS}},\text{map}}(E)$ gives the supplements for the elements AC$_r(E)$. If we use the fully direct serendipity supplements from \eqref{eq:supplementSpace}, we obtain new families of fully direct mixed elements. The computations are not difficult. Note that \begin{equation} \textrm{curl}\,\lambda_j = -\textrm{curl}\big(({\mathbf x}-{\mathbf x}_j)\cdot\nu_j\big) = \tau_j. \end{equation} Suppose that $\lambda_H$ is represented by \eqref{eq:lambdaHnuH} (i.e., the zero line is orthogonal to $\nu_H$) and define $\tau_H=(-\nu_{H,2},\nu_{H,1})$, so $\textrm{curl}\,\lambda_H=\tau_H$. If we use \eqref{eq:rational-v}--\eqref{eq:rational-h} to define $R_V$ and $R_H$, the supplemental space is (recall \eqref{eq:supplementSpace}) \begin{align} &\mathbb S_{r}^{\mathbf V}(E) = \textrm{span}\{\pmb\sigma_{r,1},\pmb\sigma_{r,2}\},\\ &\pmb\sigma_{r,1} = \textrm{curl}(R_V^{\text{simple}}\lambda_H^{r-1}\lambda_3\lambda_4) = \lambda_H^{r-1}\lambda_3\lambda_4 \frac{(\xi_V^{-1}+\eta_V^{-1})(\lambda_2\tau_1-\lambda_1\tau_2)}{(\xi_V^{-1}\lambda_1+\eta_V^{-1}\lambda_2)^2}\\ \nonumber &\qquad\quad + (r-1)R_V^{\text{simple}}\lambda_H^{r-2}\lambda_3\lambda_4\tau_H + R_V^{\text{simple}}\lambda_H^{r-1}(\lambda_4\tau_3+\lambda_3\tau_4),\\ &\pmb\sigma_{r,2} = \textrm{curl}(R_H^{\text{simple}}\lambda_V^{r-1}\lambda_1\lambda_2) = \lambda_V^{r-1}\lambda_1\lambda_2 \frac{(\xi_H^{-1}+\eta_H^{-1})(\lambda_4\tau_3-\lambda_3\tau_4)}{(\xi_H^{-1}\lambda_3+\eta_H^{-1}\lambda_4)^2}\\ \nonumber &\qquad\quad + (r-1)R_H^{\text{simple}}\lambda_V^{r-2}\lambda_1\lambda_2\tau_V + R_H^{\text{simple}}\lambda_V^{r-1}(\lambda_2\tau_1+\lambda_1\tau_2). \end{align} The normal flux on each edge is easy to compute. For $\pmb\sigma_{r,1}$, we have \begin{align*} \pmb\sigma_{r,1}\cdot\nu_1\big|_{e_1} &= -\eta_V\lambda_H^{r-2}\big((r-1)\lambda_3\lambda_4\tau_H\cdot\nu_1 + \lambda_H\,(\lambda_4\tau_3\cdot\nu_1+\lambda_3\tau_4\cdot\nu_1)\big),\\ \pmb\sigma_{r,1}\cdot\nu_2\big|_{e_2} &= \xi_V\lambda_H^{r-2}\big((r-1)\lambda_3\lambda_4\tau_H\cdot\nu_2 + \lambda_H\,(\lambda_4\tau_3\cdot\nu_2+\lambda_3\tau_4\cdot\nu_2)\big),\\ \pmb\sigma_{r,1}\cdot\nu_3\big|_{e_3} &= \pmb\sigma_{r,1}\cdot\nu_4\big|_{e_4} = 0. \end{align*} It is readily apparent that, indeed, the normal fluxes are in $\mathbb P_r(e_i)$ for each $i$. \subsection{Implementation as an $H(\textrm{div})$-conforming mixed space}\label{sec:implementationMixed} The mixed space of vector functions ${\mathbf V}_r$ over $\Omega$ is defined by merging continuously the normal fluxes across each edge $e$ of the mesh ${\cal T}_h$. That is, \begin{align} {\mathbf V}_r^{\text{full}} &= \big\{{\mathbf v}\in H(\textrm{div};\Omega)\;:\;{\mathbf v}\big|_E\in{\mathbf V}_r^{\text{full}}(E)\text{ for all }E\in{\cal T}_h\big\},\\ {\mathbf V}_r^{\text{red}} &= \big\{{\mathbf v}\in H(\textrm{div};\Omega)\;:\;{\mathbf v}\big|_E\in{\mathbf V}_r^{\text{red}}(E)\text{ for all }E\in{\cal T}_h\big\}. \end{align} This can be done locally by constructing a local basis respecting the (edge) DoFs, in a way similar to that described for the serendipity elements in Section~\ref{sec:implementation}. However, in practical implementation, the hybrid form of the mixed method is often used \cite{Arnold_Brezzi_1985}. In that case, the elements are simply concatenated and no DoF-basis is required. The Lagrange multiplier space, used to enforce the normal flux continuity, is simply \begin{equation} \Lambda_r = \big\{\lambda\in L^2\big(\cup_{E\in{\cal T}_h}\partial E\big)\;:\; \lambda\big|_e\in\mathbb P_r(e)\text{ for each edge }e\text{ of }{\cal T}_f\big\}. \end{equation} The mixed space of vector functions ${\mathbf V}_r^{\text {full}}$ or ${\mathbf V}_r^{\text{red}}$ is normally paired with a space approximating scalar functions \begin{equation} W_s = \big\{w\in L^2(\Omega)\;:\;w\big|_E\in\mathbb P_s(E)\text{ for all }E\in{\cal T}_h\big\}, \end{equation} denoted $W_r^{\text {full}}=W_r$ or $W_r^{\text{red}}=W_{r-1}$, respectively. These spaces are the divergences of the corresponding vector function spaces. \section{Stability and convergence properties}\label{sec:properties} In this section, we summarize the stability and convergence theory for our new direct finite elements. For the most part, we work over the entire domain $\Omega$. \subsection{Direct serendipity element properties}\label{sec:serendipityProperties} In Section~\ref{sec:implementation}, we discussed creating a local nodal basis for some of the shape functions of ${\cal{DS}}_r(E)$. By Theorem~\ref{thm:unisolvent}, there exists a fully nodal basis; that is, one for which every basis function vanishes at all but one nodal point. We denote it as $\{\varphi_1,\ldots,\varphi_{\dim{\cal{DS}}_r(E)}\}$. \begin{definition} Given the $r$-th order direct serendipity element $(E, {\cal{DS}}_r(E), {\cal N})$ and the nodal basis of ${\cal{DS}}_r(E)$, $\{\varphi_1,\ldots,\varphi_{\dim{\cal{DS}}_r(E)}\}$, let the operator ${\cal I}_E:L^2(E)\cap{\cal C}^0(E)\longrightarrow {\cal{DS}}_r(E)$ be interpolation. That is, for $\phi\in L^2(E)\cap {\cal C}^0(E)$, \[ {\cal I}_E\,\phi = \sum_{j=1}^{\dim {\cal{DS}}_r(E)} N_j(\phi)\,\varphi_j = \sum_{j=1}^{\dim {\cal{DS}}_r(E)} \phi({\mathbf x}_j)\,\varphi_j\in{\cal{DS}}_r(E). \] Given the finite element space ${\cal{DS}}_r$ over $\Omega$, let the operator ${\cal I}_h$ be global interpolation. That is, for a given function $\phi\in L^2(\Omega)\cap {\cal C}^0(\Omega)$, ${\cal I}_h\,v\in{\cal{DS}}_r$ and ${\cal I}_h\,\phi\big|_E = {\cal I}_E\,\phi$. \end{definition} By Theorem~\ref{thm:unisolvent}, the local interpolation operator preserves polynomials, so we have an important property~\cite[pp.~121--123]{Ciarlet_1978} expressed in the following lemma. \begin{lemma} \label{thm:poly-preserve} The interpolation operator ${\cal I}_E$ is polynomial preserving, i.e., $\forall \psi \in \mathbb P_r(E)$, ${\cal I}_E\,\psi = \psi$. Moreover, $\|{\cal I}_E\|$ is bounded in the $L^2$-norm. \end{lemma} With this lemma and Theorem~\ref{thm:unisolvent}, we have the analogue of the Bramble-Hilbert~\cite{Bramble_Hilbert_1970} or Dupont-Scott~\cite{Dupont_Scott_1980} lemma for local and global error estimation, provided the mesh is shape regular. \begin{lemma}\label{lem:bramble} There exists a constant $C>0$ such that for all functions $\phi\in H^{s+1}(E)$ $(H^1(E)\cap{\cal C}^0(E)$ if $s=0)$, \begin{align} |\phi-{\cal I}_E\,\phi|_{m,E}\leq C\,h_E^{s+1-m}\,|\phi|_{s+1,E}, \quad m=0,1\text{ and }s=0,1,\ldots,r, \end{align} where $|\cdot|_{m,E}$ is the $H^m(E)$ seminorm. Moreover, suppose that ${\cal T}_h$ is uniformly shape regular as $h\to0$. Then there exists a constant $C>0$, independent of $h$, such that for all functions $\phi\in H^{s+1}(\Omega)$ $(\phi\in H^1(\Omega)\cap {\cal C}^0(\overline\Omega)$ if $s=0)$, \begin{align} |\phi - {\cal I}_h\,\phi|_{m,\Omega}\leq C\,h^{s+1-m}\,|\phi|_{s+1,\Omega}, \quad m=0,1\text{ and }s=0,1\ldots,r. \end{align} \end{lemma} \subsection{Direct mixed finite element properties}\label{sec:mixedProperties} As was done by Raviart and Thomas \cite{Raviart_Thomas_1977} for their mixed spaces, we can define a projection operator, usually denoted $\pi$, mapping $H(\textrm{div};\Omega)\cap(L^{2+\epsilon}(\Omega))^2$, $\epsilon>0$, onto ${\mathbf V}_r$ (full or reduced) that has several important properties. The operator $\pi$ is pieced together from locally defined operators $\pi_E$. Following \cite{Arbogast_Correa_2016}, for ${\mathbf v}$, we define $\pi_E{\mathbf v}$ in terms of the DoFs \eqref{eq:vdofEdge}--\eqref{eq:vdofCell}. The operator $\pi$ satisfies the commuting diagram property~\cite{Douglas_Roberts_1985}, which is to say that \begin{equation} {\cal P}_{W_s}\nabla\cdot{\mathbf v}=\nabla\cdot\pi{\mathbf v}, \end{equation} where ${\cal P}_{W_s}$ is the $L^2$-orthogonal projection operator onto $W_s=\nabla\cdot{\mathbf V}_r$ and $s=r$ for full spaces and $s=r-1$ for reduced. Since $\pi_E$ is bounded in, say, $H^1$, we obtain the following approximation results \cite{Bramble_Hilbert_1970, Dupont_Scott_1980, Brezzi_Fortin_1991, Ciarlet_1978, Boffi_Brezzi_Fortin_2013, Arbogast_Correa_2016}. \begin{lemma}\label{lem:brambleMixed} Suppose that ${\cal T}_h$ is uniformly shape regular as $h\to0$. Then there is a constant $C>0$, independent of $h$, such that \begin{alignat}2 \label{eq:approx_u} \|{\mathbf v}-\pi{\mathbf v}\|_{0,\Omega} &\le C\,\|{\mathbf v}\|_{k,\Omega}\,h^{k},&&\quad k=1,\ldots,r+1,\\ \label{eq:approx_divu} \|\nabla\cdot({\mathbf v}-\pi{\mathbf v})\|_{0,\Omega} &\le C\,\|\nabla\cdot{\mathbf u}\|_{k,\Omega}\,h^{k},&&\quad k=0,1,\ldots,s+1,\\ \label{eq:approx_p} \|p-{\cal P}_{W_s}p\|_{0,\Omega} &\le C\,\|p\|_{k,\Omega}\,h^{k},&&\quad k=0,1,\ldots,s+1, \end{alignat} where $s=r\ge1$ and $s=r-1\ge0$ for full and reduced $H(\textrm{div})$-approximation, respectively, and $\|\cdot\|_{m,\Omega}$ is the $H^m(\Omega)$ norm. Moreover, the discrete inf-sup condition \begin{equation} \label{eq:inf-sup} \sup_{{\mathbf v}_h\in{\mathbf V}_r}\frac{(w_h,\nabla\cdot{\mathbf v}_h)}{\|{\mathbf v}_h\|_{H(\text{\rm div})}} \ge \gamma\,\|w_h\|_{0,\Omega},\quad\forall w_h\in W_s, \end{equation} holds for some $\gamma>0$ independent of $h$. \end{lemma} \subsection{Application to second order elliptic equations}\label{sec:ellipticProperties} Consider a uniformly elliptic problem with a homogeneous Dirichlet boundary condition \begin{alignat}{2} \label{eq:elliptic-pde} -\nabla\cdot({\mathbf a} \nabla p) &= f &&\quad\text{in }\Omega,\\ \label{eq:elliptic-dirBC} p &= 0 &&\quad\text{on }\partial\Omega, \end{alignat} where the second order tensor ${\mathbf a}({\mathbf x})$ is uniformly positive definite and bounded, and $f\in L^2(\Omega)$. The boundary value problem can be written in the weak form: Find $p\in H_0^1(\Omega)$ such that \begin{align} \label{eq:weak-bvp} ({\mathbf a}\nabla p,\nabla q) = (f,q), \quad\forall q\in H_0^1(\Omega), \end{align} where $(\cdot,\cdot)$ is the $L^2(\Omega)$ inner product. Setting \begin{equation} \label{eq:elliptic-flux} {\mathbf u} = -{\mathbf a}\nabla p, \end{equation} we also have the mixed weak form: Find ${\mathbf u}\in H(\textrm{div};\Omega)$ and $p\in L^2(\Omega)$ such that \begin{alignat}3 \label{eq:mixed-darcy} &({\mathbf a}^{-1}{\mathbf u},{\mathbf v}) - (p,\nabla\cdot{\mathbf v}) &&=0, &&\quad\forall{\mathbf v}\in H(\textrm{div};\Omega),\\ \label{eq:mixed-conservation} &(\nabla\cdot{\mathbf u},w) &&= (f,w), &&\quad\forall w\in L^2(\Omega). \end{alignat} Define the global finite element space over ${\cal T}_h$ \begin{equation*}\label{eq:x0h-in-h01} X_{0,h}=\{v_h\in{\cal{DS}}_r\;:\; v_h=0\text{ on }\partial\Omega\}\subset H_0^1(\Omega). \end{equation*} We then obtain the Galerkin approximation: Find $p_h\in X_{0,h}$ such that \begin{align} \label{eq:approx-bvp} ({\mathbf a}\nabla p_h,q_h) = (f,q_h), \quad\forall q_h\in X_{0,h}. \end{align} Combining C\'ea's lemma \cite{Ciarlet_1978,Brenner_Scott_1994} and the global projection estimate Lemma~\ref{lem:bramble}, we obtain an $H^1$-error estimate for the problem. Since $\Omega$ is a polygonal domain, $\partial\Omega$ is a Lipschitz boundary. If we assume that $\Omega$ is also convex, we have elliptic regularity of the solution~\cite[Theorem~4.3.1.4]{Grisvard_1985}, and the Aubin-Nitsche duality principle \cite{Ciarlet_1978,Brenner_Scott_1994} gives an $L^2$-error estimate. \begin{theorem}\label{thm:convergence} Let $\Omega$ be a convex polygonal domain and let ${\cal T}_h$ be uniformly shape regular. There exists a constant $C>0$, independent of $h$, such that \begin{align} \|p-p_h\|_{m,\Omega}&\leq C\,h^{s+1-m}\,|p|_{s+1,\Omega},\quad s=0,1,\ldots,r,\quad m=0,1, \end{align} where $p_h$ satisfies \eqref{eq:approx-bvp}. \end{theorem} We also have the mixed full ($s=r$) and reduced ($s=r-1$) $H(\textrm{div})$-approximation: Find $({\mathbf u}_h,p_h)\in{\mathbf V}_r\times W_s$ such that \begin{alignat}3 \label{eq:mixed-darcy-approx} &({\mathbf a}^{-1}{\mathbf u}_h,{\mathbf v}_h) - (p_h,\nabla\cdot{\mathbf v}_h) &&=0, &&\quad\forall{\mathbf v}_h\in{\mathbf V}_r,\\ \label{eq:mixed-conservation-approx} &(\nabla\cdot{\mathbf u}_h,w_h) &&= (f,w_h), &&\quad\forall w_h\in W_s. \end{alignat} \begin{theorem}\label{thm:convergenceMixed} Let $\Omega$ be a convex polygonal domain and let ${\cal T}_h$ be uniformly shape regular. There exists a constant $C>0$, independent of $h$, such that \begin{alignat}2 \|{\mathbf u}-{\mathbf u}_h\| &\le C\|{\mathbf u}\|_{k}h^{k},&&\quad k=1,\ldots,r+1,\\ \|\nabla\cdot({\mathbf u}-{\mathbf u}_h)\| + \|p-p_h\| &\le C\|{\mathbf u}\|_{k}h^{k},&&\quad k=0,1,\ldots,s+1, \end{alignat} where $({\mathbf u}_h,p_h)$ satisfies \eqref{eq:mixed-darcy-approx}--\eqref{eq:mixed-conservation-approx}. \end{theorem} \section{Numerical results}\label{sec:numerics} In this section, we consider the test problem~\eqref{eq:elliptic-pde}--\eqref{eq:elliptic-dirBC} defined on the unit square $\Omega = [0,1]^2$ with the coefficient ${\mathbf a}$ being the $2\times2$ identity matrix, i.e., we solve the Poisson equation. The exact solution is $u(x,y) = \sin(\pi x)\sin(\pi y)$ and the source term is $f(x,y) = 2\pi^2\sin(\pi x)\sin(\pi y)$. Solutions are computed on three different sequences of meshes. The first sequence, ${\cal T}_h^1$, is a uniform mesh of $n^2$ square elements (two sets of parallel edges per element). The second sequence, ${\cal T}_h^2$, is a mesh of $n^2$ trapezoids of base $h$ and one pair of parallel edges of size $0.75h$ and $1.25h$, as proposed in~\cite{ABF_2002}. The third sequence, ${\cal T}_h^3$, is chosen so as to have no pair of edges parallel. The first $4\times 4$ meshes for each sequence are shown in Figure~\ref{fig:meshes}. Finer meshes are constructed by repeating the same pattern over the domain. Our computer program uses the deal.II library \cite{BDHHKKMTW_2016_dealII84}. \begin{figure}[ht] \centerline{ \parbox{.16\linewidth}{\includegraphics[width=\linewidth]{grid-T1.eps}\newline\centerline{${\cal T}_h^1$}} \qquad\qquad\parbox{.16\linewidth}{\includegraphics[width=\linewidth]{grid-T2.eps}\newline\centerline{${\cal T}_h^2$}} \qquad\qquad\parbox{.16\linewidth}{\includegraphics[width=\linewidth]{grid-T3.eps}\newline\centerline{${\cal T}_h^3$}} } \caption{The three $4\times4$ base meshes. Finer meshes are constructed by repeating the base mesh pattern over the domain. The meshes have 2, 1, and 0 parallel edges per element, respectively.}\label{fig:meshes} \end{figure} \subsection{Fully direct serendipity spaces} In this section we present convergence studies for the fully direct serendipity spaces ${\cal{DS}}_r$ using the elements defined in~\eqref{eq:shapeSpace}. We compare the results with those of the regular serendipity spaces ${\cal S}_r$ and the spaces of elements given by mapping the local tensor product space $\mathbb P_{r,r}(\hat E)$ to the mesh elements (hereafter simply called the $\mathbb P_{r,r}$ space). As described above, one may need to consider whether opposite faces are parallel to construct a fully direct serendipity element. We now give a simple choice of element that avoids this difficulty. First, we can take \eqref{eq:simpleLambdaHV} for $\lambda_H$ and $\lambda_V$. Let $R_V$ and $R_H$ be defined by \eqref{eq:rational-v}--\eqref{eq:rational-h}, where we define $\nu_H=(\nu_3-\nu_4)/|\nu_3-\nu_4|$ and $\nu_V=(\nu_1-\nu_2)/|\nu_1-\nu_2|$, and set \begin{equation}\label{def:xi-eta-vh} \begin{alignedat}2 \xi_V^{-1} &= \sqrt{1-(\nu_H\cdot\nu_1)^2},\qquad &\eta_V^{-1} &= \sqrt{1-(\nu_H\cdot\nu_3)^2},\\ \xi_H^{-1} &= \sqrt{1-(\nu_V\cdot\nu_2)^2},\qquad &\eta_H^{-1} &= \sqrt{1-(\nu_V\cdot\nu_4)^2}. \end{alignedat} \end{equation} For an $n\times n$ mesh, the total number of degrees of freedom for $\mathbb P_{r,r}$ is $(nr+1)^2 = {\cal O}(r^2n^2)$, and for ${\cal S}_r$ and ${\cal{DS}}_r$ it is \begin{align*} \dim({\cal S}_r) = \dim({\cal{DS}}_r) &= (\text{number of vertices}) + (\text{number of edges})(r-1)\\ &\qquad\quad + (\text{number of cells})\tfrac12{(r-2)(r-3)} \\ &=(n+1)^2 + 2n(n+1)(r-1) + n^2 \tfrac12{(r-2)(r-3)}\\ &=\tfrac12(r^2 - r + 4)n^2 + 2rn + 1 = {\cal O}\left(\tfrac12{(r^2-r+4)n^2}\right). \end{align*} Therefore, the total number of degrees of freedom for a serendipity space is asymptotically about half the size of that for a tensor product space of the same order. We report the $L^2$-errors and the orders of the convergence of the spaces $\mathbb P_{r,r}$, ${\cal S}_r$ and ${\cal{DS}}_r$ for $r=2,3,4,5$ on mesh sequence ${\cal T}_h^1$ in Table~\ref{tab:t1-l2}. The errors and convergence rates in the $H^1$-seminorm are presented in Table~\ref{tab:t1-h1}. Since ${\cal T}_h^1$ is a sequence of square meshes, the direct serendipity space ${\cal{DS}}_r$ and the regular serendipity space ${\cal S}_r$ coincide on ${\cal T}_h^1$. All three families of spaces show an $(r+1)$-st order convergence in the $L^2$-norm and an $r$-th order convergence in the $H^1$-seminorm, as we should expect from theory. The errors for $\mathbb P_{r,r}$ are smaller than that for ${\cal{DS}}_r={\cal S}_r$, but $\mathbb P_{r,r}$ uses many more degrees of freedom \begin{table}[!ht] \caption{$L^2$-errors and convergence rates for $\mathbb P_{r,r}$, ${\cal{DS}}_r$, and ${\cal S}_r$ spaces on square meshes.}\label{tab:t1-l2} \begin{center} \begin{tabular}{c|cc|cc|cc|cc} \hline & \multicolumn{2}{c|}{$r=2$} & \multicolumn{2}{c|}{$r=3$} & \multicolumn{2}{c|}{$r=4$} & \multicolumn{2}{c}{$r=5$} \\ $n$ & error & rate & error & rate & error & rate & error & rate \\ \hline \multicolumn{9}{c}{$\vphantom{\int^H} \mathbb P_{r,r}$ on ${\cal T}_h^1$ meshes}\\ \hline \08 & 2.451e-04 & 2.99 & 5.564e-06 & 3.99 & 1.054e-07 & 4.99 & 1.688e-09 & 6.00 \\ 12 & 7.282e-05 & 2.99 & 1.101e-06 & 4.00 & 1.389e-08 & 5.00 & 1.483e-10 & 6.00 \\ 16 & 3.075e-05 & 3.00 & 3.486e-07 & 4.00 & 3.298e-09 & 5.00 & 2.640e-11 & 6.00 \\ 24 & 9.116e-06 & 3.00 & 6.890e-08 & 4.00 & 4.344e-10 & 5.00 & 2.420e-12 & 5.89 \\ \hline \multicolumn{9}{c}{$\vphantom{\int^H} {\cal S}_{r}={\cal{DS}}_{r}$ on ${\cal T}_h^1$ meshes}\\ \hline \08 & 2.457e-04 & 2.99 & 1.805e-05 & 4.09 & 1.422e-06 & 5.01 & 6.440e-08 & 5.93 \\ 12 & 7.289e-05 & 3.00 & 3.497e-06 & 4.05 & 1.870e-07 & 5.00 & 5.739e-09 & 5.96 \\ 16 & 3.076e-05 & 3.00 & 1.099e-06 & 4.02 & 4.437e-08 & 5.00 & 1.027e-09 & 5.98 \\ 24 & 9.118e-06 & 3.00 & 2.161e-07 & 4.01 & 5.841e-09 & 5.00 & 9.049e-11 & 5.99 \\ \hline \end{tabular} \end{center} \end{table} \begin{table}[!ht] \caption{$H^1$-seminorm errors and convergence rates for $\mathbb P_{r,r}$, ${\cal{DS}}_r$, and ${\cal S}_r$ spaces on square meshes.}\label{tab:t1-h1} \begin{center} \begin{tabular}{c|cc|cc|cc|cc} \hline & \multicolumn{2}{c|}{$r=2$} & \multicolumn{2}{c|}{$r=3$} & \multicolumn{2}{c|}{$r=4$} & \multicolumn{2}{c}{$r=5$} \\ $n$ & error & rate & error & rate & error & rate & error & rate \\ \hline \multicolumn{9}{c}{$\vphantom{\int^H} \mathbb P_{r,r}$ on ${\cal T}_h^1$ meshes}\\ \hline \08 & 1.276e-02 & 2.00 & 4.233e-04 & 3.00 & 1.047e-05 & 4.00 & 2.066e-07 & 5.00 \\ 12 & 5.673e-03 & 2.00 & 1.255e-04 & 3.00 & 2.070e-06 & 4.00 & 2.723e-08 & 5.00 \\ 16 & 3.191e-03 & 2.00 & 5.295e-05 & 3.00 & 6.549e-07 & 4.00 & 6.462e-09 & 5.00 \\ 24 & 1.418e-03 & 2.00 & 1.569e-05 & 3.00 & 1.294e-07 & 4.00 & 8.511e-10 & 5.00 \\ \hline \multicolumn{9}{c}{$\vphantom{\int^H} {\cal S}_{r}={\cal{DS}}_{r}$ on ${\cal T}_h^1$ meshes}\\ \hline \08 & 1.285e-02 & 2.02 & 1.537e-03 & 3.05 & 1.141e-04 & 3.99 & 5.201e-06 & 4.99 \\ 12 & 5.690e-03 & 2.01 & 4.507e-04 & 3.03 & 2.261e-05 & 3.99 & 6.856e-07 & 5.00 \\ 16 & 3.197e-03 & 2.00 & 1.894e-04 & 3.01 & 7.164e-06 & 4.00 & 1.628e-07 & 5.00 \\ 24 & 1.420e-03 & 2.00 & 5.597e-05 & 3.01 & 1.416e-06 & 4.00 & 2.144e-08 & 5.00 \\ \hline \end{tabular} \end{center} \end{table} Tables~\ref{tab:t2-l2}--\ref{tab:t2-h1} show the errors (in the $L^2$ and $H^1$-seminorms, respectively) and the orders of convergence for the trapezoidal mesh sequence ${\cal T}_h^2$. The tensor product space $\mathbb P_{r,r}$ achieves the expected optimal convergence rates. The direct serendipity space ${\cal{DS}}_r$ retains an optimal $(r+1)$-st order of convergence in the $L^2$ norm and an optimal $r$-th order convergence in the $H^1$-seminorm, as Theorem~\ref{thm:convergence} predicts. The regular serendipity spaces ${\cal S}_r$ have worse than optimal convergence rates in both norms (as was also observed in~\cite{ABF_2002}). The errors and convergence rates for $\mathbb P_{r,r}$, ${\cal{DS}}_r$, and ${\cal S}_r$ on mesh sequence ${\cal T}_h^3$ are similar to those on ${\cal T}_h^2$, so we omit showing them. \begin{table}[ht] \caption{$L^2$-errors and convergence rates for $\mathbb P_{r,r}$, ${\cal{DS}}_r$, and ${\cal S}_r$ spaces on trapezoidal meshes.}\label{tab:t2-l2} \begin{center} \begin{tabular}{c|cc|cc|cc|cc} \hline & \multicolumn{2}{c|}{$r=2$} & \multicolumn{2}{c|}{$r=3$} & \multicolumn{2}{c|}{$r=4$} & \multicolumn{2}{c}{$r=5$} \\ $n$ & error & rate & error & rate & error & rate & error & rate \\ \hline \multicolumn{9}{c}{$\vphantom{\int^H}\mathbb P_{r,r}$ on ${\cal T}_h^2$ meshes}\\ \hline \08 & 3.329e-04 & 2.99 & 9.740e-06 & 3.99 & 2.382e-07 & 4.99 & 5.076e-09 & 5.99 \\ 12 & 9.888e-05 & 2.99 & 1.928e-06 & 3.99 & 3.142e-08 & 5.00 & 4.462e-10 & 6.00 \\ 16 & 4.176e-05 & 3.00 & 6.107e-07 & 4.00 & 7.459e-09 & 5.00 & 7.946e-11 & 6.00 \\ 24 & 1.238e-05 & 3.00 & 1.207e-07 & 4.00 & 9.827e-10 & 5.00 & 6.979e-12 & 6.00 \\ \hline \multicolumn{9}{c}{$\vphantom{\int^H} {\cal S}_{r}$ on ${\cal T}_h^2$ meshes}\\ \hline \08 & 5.714e-04 & 2.92 & 4.844e-04 & 2.89 & 2.612e-05 & 3.72 & 2.005e-06 & 4.13 \\ 12 & 1.731e-04 & 2.94 & 1.482e-04 & 2.92 & 6.084e-06 & 3.59 & 3.884e-07 & 4.05 \\ 16 & 7.409e-05 & 2.95 & 6.383e-05 & 2.93 & 2.265e-06 & 3.43 & 1.234e-07 & 3.99 \\ 24 & 2.254e-05 & 2.94 & 1.963e-05 & 2.91 & 5.984e-07 & 3.28 & 2.516e-08 & 3.92 \\ 32 & 9.799e-06 & 2.90 & 8.635e-06 & 2.85 & 2.408e-07 & 3.16 & 8.342e-09 & 3.84 \\ 64 & 1.440e-06 & 2.70 & 1.332e-06 & 2.61 & 2.862e-08 & 3.05 & 6.644e-10 & 3.56 \\ \hline \multicolumn{9}{c}{$\vphantom{\int^H} {\cal{DS}}_{r}$ on ${\cal T}_h^2$ meshes}\\ \hline \08 & 3.492e-04 & 3.00 & 3.897e-05 & 4.07 & 2.187e-06 & 5.00 & 8.896e-08 & 5.96 \\ 12 & 1.036e-04 & 3.00 & 7.457e-06 & 4.08 & 2.889e-07 & 4.99 & 7.870e-09 & 5.98 \\ 16 & 4.373e-05 & 3.00 & 2.313e-06 & 4.07 & 6.868e-08 & 4.99 & 1.404e-09 & 5.99 \\ 24 & 1.296e-05 & 3.00 & 4.469e-07 & 4.05 & 9.058e-09 & 5.00 & 1.235e-10 & 6.00 \\ \hline \end{tabular} \end{center} \end{table} \begin{table}[ht] \caption{$H^1$-seminorm errors and convergence rates for $\mathbb P_{r,r}$, ${\cal{DS}}_r$, and ${\cal S}_r$ spaces on trapezoidal meshes.}\label{tab:t2-h1} \begin{center} \begin{tabular}{c|cc|cc|cc|cc} \hline & \multicolumn{2}{c|}{$r=2$} & \multicolumn{2}{c|}{$r=3$} & \multicolumn{2}{c|}{$r=4$} & \multicolumn{2}{c}{$r=5$} \\ $n$ & error & rate & error & rate & error & rate & error & rate \\ \hline \multicolumn{9}{c}{$\vphantom{\int^H}\mathbb P_{r,r}$ on ${\cal T}_h^2$ meshes}\\ \hline \08 & 1.734e-02 & 2.00 & 7.206e-04 & 2.99 & 2.310e-05 & 3.99 & 6.083e-07 & 4.99 \\ 12 & 7.710e-03 & 2.00 & 2.139e-04 & 3.00 & 4.570e-06 & 4.00 & 8.021e-08 & 5.00 \\ 16 & 4.337e-03 & 2.00 & 9.027e-05 & 3.00 & 1.447e-06 & 4.00 & 1.904e-08 & 5.00 \\ 24 & 1.928e-03 & 2.00 & 2.676e-05 & 3.00 & 2.859e-07 & 4.00 & 2.509e-09 & 5.00 \\ \hline \multicolumn{9}{c}{$\vphantom{\int^H} {\cal S}_{r}$ on ${\cal T}_h^2$ meshes}\\ \hline \08 & 2.413e-02 & 1.94 & 1.834e-02 & 1.90 & 1.818e-03 & 2.65 & 1.537e-04 & 3.18 \\ 12 & 1.105e-02 & 1.93 & 8.572e-03 & 1.88 & 6.582e-04 & 2.51 & 4.483e-05 & 3.04 \\ 16 & 6.432e-03 & 1.88 & 5.091e-03 & 1.81 & 3.345e-04 & 2.35 & 1.945e-05 & 2.90 \\ 24 & 3.104e-03 & 1.80 & 2.560e-03 & 1.70 & 1.360e-04 & 2.22 & 6.370e-06 & 2.75 \\ 32 & 1.920e-03 & 1.67 & 1.643e-03 & 1.54 & 7.378e-05 & 2.12 & 3.029e-06 & 2.58 \\ 64 & 7.097e-04 & 1.34 & 6.602e-04 & 1.23 & 1.776e-05 & 2.03 & 5.953e-07 & 2.26 \\ \hline \multicolumn{9}{c}{$\vphantom{\int^H} {\cal{DS}}_{r}$ on ${\cal T}_h^2$ meshes}\\ \hline \08 & 1.836e-02 & 2.01 & 2.517e-03 & 3.02 & 1.625e-04 & 3.99 & 7.384e-06 & 4.99 \\ 12 & 8.143e-03 & 2.00 & 7.400e-04 & 3.02 & 3.216e-05 & 4.00 & 9.757e-07 & 4.99 \\ 16 & 4.577e-03 & 2.00 & 3.109e-04 & 3.01 & 1.018e-05 & 4.00 & 2.318e-07 & 5.00 \\ 24 & 2.033e-03 & 2.00 & 9.170e-05 & 3.01 & 2.012e-06 & 4.00 & 3.056e-08 & 5.00 \\ \hline \end{tabular} \end{center} \end{table} We remark that the time cost for the assembly routine can be scaled nearly perfectly in parallel, since it basically involves only local computations. Therefore, reducing the global number of degrees of freedom in a serendipity space versus a tensor product space, even perhaps at the expense of a slightly more expensive assembly, is worthwhile~\cite{Tao_2017_phd, Arbogast_Tao_2017_serendipity}. \subsection{Fully direct mixed finite elements on quadrilaterals} In this section, we verify the convergence rate for the new fully direct mixed finite elements derived in Section~\ref{sec:newMixed}. These are implemented without the use of any mapping from the reference element. We take $\xi_V=\xi_H=\eta_V=\eta_H=1$, although taking the values in \eqref{def:xi-eta-vh} provides similar results. We apply the hybrid form of the the mixed finite element method~\cite{Arnold_Brezzi_1985}. The errors and the orders of convergence for the reduced and full $H(\textrm{div})$-approximation spaces when $r=1,2$ on mesh ${\cal T}_h^2$ are presented in Table~\ref{tab:at1-h2}. Again, results are similar on ${\cal T}_h^3$ meshes. As the theory predicts, the scalar $p$, the vector ${\mathbf u}$, and the divergence $\nabla\cdot{\mathbf u}$ retain $r$-th, $(r+1)$-st, and $r$-th order approximation, respectively, for the reduced $H(\textrm{div})$-approximation spaces, and all three quantities show $r$-th order approximation for the full $H(\textrm{div})$-approximation spaces. \begin{table}[ht] \caption{Errors and convergence rates for fully direct mixed spaces on trapezoidal meshes ${\cal T}_h^2$.}\label{tab:at1-h2} \begin{center} \begin{tabular}{c|cc|cc|cc} \hline & \multicolumn{2}{c|}{$||p-p_h||$} & \multicolumn{2}{c|}{$\|{\mathbf u}-{\mathbf u}_h\|$} & \multicolumn{2}{c}{$\|\nabla\cdot({\mathbf u}-{\mathbf u}_h)\|$}\\ $n$ & error & rate & error & rate & error & rate\\ \hline \multicolumn{7}{c}{$r=1$, reduced $H(\textrm{div})$-approximation}\\ \hline \04 & 1.670e-01 & --- & 2.609e-01 & --- & 3.163e-00 & --- \\ \08 & 8.271e-02 & 1.01 & 6.803e-02 & 1.96 & 1.612e-00 & 0.98 \\ 16 & 4.117e-02 & 1.00 & 1.719e-02 & 1.99 & 8.099e-01 & 1.00 \\ 32 & 2.056e-02 & 1.00 & 4.309e-03 & 2.00 & 4.054e-01 & 1.00 \\ \hline \multicolumn{7}{c}{$r=2$, reduced $H(\textrm{div})$-approximation}\\ \hline \04 & 3.079e-02 & --- & 2.319e-02 & --- & 6.067e-01 & --- \\ \08 & 7.847e-03 & 1.98 & 2.906e-03 & 3.00 & 1.549e-01 & 1.98 \\ 16 & 1.972e-03 & 2.00 & 3.633e-04 & 3.00 & 3.892e-02 & 2.00 \\ 32 & 4.936e-04 & 2.00 & 4.543e-05 & 3.00 & 9.742e-03 & 2.00 \\ \hline \multicolumn{7}{c}{$r=1$, full $H(\textrm{div})$-approximation}\\ \hline \04 & 3.079e-02 & - & 5.562e-02 & - & 6.067e-01 & - \\ \08 & 7.847e-03 & 1.98 & 1.350e-02 & 2.02 & 1.549e-01 & 1.98 \\ 16 & 1.972e-03 & 2.00 & 3.355e-03 & 2.01 & 3.892e-02 & 2.00 \\ 32 & 4.936e-04 & 2.00 & 8.378e-04 & 2.00 & 9.742e-03 & 2.00 \\ \hline \multicolumn{7}{c}{$r=2$, full $H(\textrm{div})$-approximation}\\ \hline \04 & 4.081e-03 & - & 7.198e-03 & - & 8.050e-02 & - \\ \08 & 5.201e-04 & 2.98 & 9.105e-04 & 2.99 & 1.026e-02 & 2.98 \\ 16 & 6.533e-05 & 3.00 & 1.141e-04 & 3.00 & 1.289e-03 & 3.00 \\ 32 & 8.176e-06 & 3.00 & 1.428e-05 & 3.00 & 1.614e-04 & 3.00 \\ \hline \end{tabular} \end{center} \end{table} Our numerical test agrees with that taken in~\cite{Arbogast_Correa_2016}, where results for the full and reduced AC spaces and the mapped BDM spaces appear. Results for our fully direct mixed spaces agree very closely with the results for the AC spaces, and these far exceed the performance of the mapped BDM spaces. \subsection{Serendipity space based on mapped supplements} In this section, we present the errors and convergence rates for the serendipity spaces ${\cal{DS}}_r^{\text{map}}$ using elements defined in \eqref{eq:DSmap}, which has supplements mapped from the reference element. The results for $r=2,3,4$ on mesh ${\cal T}_h^2$ are shown in Table~\ref{tab:ds-mapped}. As predicted by the theory, this new family of spaces shows an $(r+1)$-st order convergence in the $L^2$-norm and an $r$-th order convergence in the $H^1$-seminorm. The results compare favorably with those for the fully direct spaces in Tables~\ref{tab:t2-l2}--\ref{tab:t2-h1}, although the latter are perhaps slightly better. \begin{table}[ht] \caption{Errors and convergence rates for ${\cal{DS}}_r^{\textrm{map}}$ spaces on trapezoidal meshes ${\cal T}_h^2$.}\label{tab:ds-mapped} \begin{center} \begin{tabular}{c|cc|cc|cc|cc} \hline & \multicolumn{2}{c|}{$r=2$} & \multicolumn{2}{c|}{$r=3$} & \multicolumn{2}{c|}{$r=4$} & \multicolumn{2}{c}{$r=5$}\\ $n$ & error & rate & error & rate & error & rate & error & rate \\ \hline \multicolumn{9}{c}{$\vphantom{\int^H} L^2$-errors and convergence rates}\\ \hline \08 & 5.737e-04 & 2.92 & 4.128e-05 & 4.09 & 2.344e-06 & 5.04 & 9.134e-08 & 6.00 \\ 12 & 1.727e-04 & 2.96 & 7.968e-06 & 4.06 & 3.048e-07 & 5.03 & 8.023e-09 & 6.00 \\ 16 & 7.329e-05 & 2.98 & 2.493e-06 & 4.04 & 7.182e-08 & 5.03 & 1.428e-09 & 6.00 \\ 24 & 2.180e-05 & 2.99 & 4.869e-07 & 4.03 & 9.380e-09 & 5.02 & 1.252e-10 & 6.00 \\ \hline \multicolumn{9}{c}{$\vphantom{\int^H} H^1$-seminorm errors and convergence rates}\\ \hline \08 & 2.410e-02 & 1.99 & 2.851e-03 & 3.05 & 1.730e-04 & 4.03 & 7.609e-06 & 5.01 \\ 12 & 1.074e-02 & 1.99 & 8.333e-04 & 3.03 & 3.385e-05 & 4.02 & 9.979e-07 & 5.01 \\ 16 & 6.047e-03 & 2.00 & 3.491e-04 & 3.02 & 1.065e-05 & 4.02 & 2.362e-07 & 5.01 \\ 24 & 2.690e-03 & 2.00 & 1.027e-04 & 3.02 & 2.091e-06 & 4.02 & 3.102e-08 & 5.01 \\ \hline \end{tabular} \end{center} \end{table} \section{Summary and Conclusions}\label{sec:conc} It is possible to define a wide variety of direct serendipity elements on a nondegenerate, convex quadrilateral $E$. Most or perhaps all of these elements appear to be new, and they have the form \begin{equation}\label{eq:serendipityFormConc} {\cal{DS}}_r(E) = \mathbb P_r(E)\oplus\mathbb S_r^{\cal{DS}}(E),\quad r\geq2. \end{equation} The supplemental space $\mathbb S_r^{\cal{DS}}(E)$ can be defined by four functions. Referring to Figure~\ref{fig:numbering}, the linear functions $\lambda_H$ and $\lambda_V$ are arbitrary except that the zero line of $\lambda_H$ must intersect the lines containing $e_1$ and $e_2$ above or below the intersection point ${\mathbf x}_{12}$, if it exists, and $\lambda_V$ must intersect the lines containing $e_3$ and $e_4$ to the left or right of the intersection point ${\mathbf x}_{34}$, if it exists. The bounded, (most likely) nonlinear functions $R_V$ and $R_H$ can be chosen arbitrarily as long as they are negative constants on $e_1$ and $e_3$, respectively, and positive constants on $e_2$ and $e_4$, respectively. For example, one can take the simple choices \begin{equation} \lambda_H = \lambda_3 - \lambda_4,\quad R_V = \frac{\lambda_1 - \lambda_2}{\lambda_1 + \lambda_2},\quad \lambda_V = \lambda_1 - \lambda_2,\quad R_H = \frac{\lambda_3 - \lambda_4}{\lambda_3 + \lambda_4}, \end{equation} or the choices given in Lemma~\ref{lem:equ-space}, in which case the explicit basis \eqref{eq:ds-baseE1}--\eqref{eq:ds-baseV13} can be constructed easily. The fully direct supplemental space is \begin{equation} \label{eq:supplementSpaceConc} \mathbb S_r^{{\cal{DS}}}(E) = \textrm{span}\{\lambda_3\lambda_4\lambda_H^{r-2}R_V,\lambda_1\lambda_2\lambda_V^{r-2}R_H\}, \end{equation} but a supplemental space can also be defined using the bilinear map between $\hat E=[-1,1]^2$ and $E$ as \begin{equation} \label{eq:supplementSpaceConcMapped} \mathbb S_{r}^{{\cal{DS}},\text{map}}(E) = \textrm{span}\big\{F_{\!E}^0\big((1-\hat x_2^2)\hat x_1\hat x_2^{r-2}\big), F_{\!E}^0\big((1-\hat x_1^2)\hat x_2\hat x_1^{r-2}\big)\big\}. \end{equation} It is possible to define a wide variety of direct mixed elements on $E$. The de Rham theory is useful in this regard, and the elements take the form \begin{align} {\mathbf V}_r^{\textrm{red}}(E) &= \textrm{curl}\,{\cal{DS}}_{r+1}(E)\oplus{\mathbf x}\mathbb P_{r-1}(E) = \mathbb P_r^2(E)\oplus\mathbb S_r^{\mathbf V}(E)\quad r\geq1,\\ {\mathbf V}_r^{\textrm{full}}(E) &= {\mathbf V}_r^{\textrm{red}}(E)\oplus{\mathbf x}\tilde\mathbb P_r(E),\quad r\geq1, \end{align} for reduced and full $H(\textrm{div})$-approximation spaces, where \begin{equation} \mathbb S_r^{\mathbf V}(E) = \textrm{curl}\,\mathbb S_r^{\cal{DS}}(E). \end{equation} If \eqref{eq:supplementSpaceConcMapped} is used, the AC spaces \cite{Arbogast_Correa_2016} result. Otherwise, the elements appear to be new, and they are the first families of fully direct mixed spaces defined on quadrilaterals. The direct serendipity and mixed elements can be merged to create $H^1(\Omega)$ and $H(\textrm{div};\Omega)$ conforming spaces, respectively, on a mesh ${\cal T}_h$ of nondegenerate, convex quadrilaterals. If the meshes are shape regular as $h\to0$, the spaces have both optimal approximation properties and minimal local dimension. Numerical results were presented to illustrate their performance. We close with a simple observation. Another well-known de Rham complex is \begin{equation}\label{eq:deRhamCurl} \mathbb R \DOTSB\lhook\joinrel\longrightarrow H^1 \overset{\textrm{grad}\,}{\longlongrightarrow} H(\textrm{curl}) \overset{\textrm{curl}\,}{\longlongrightarrow} L^2 \longrightarrow 0, \end{equation} where the curl of a vector function $\pmb\psi=\big(\psi_1,\psi_2\big)$ is the scalar $\textrm{curl}\,\pmb\psi({\mathbf x}) = \dfrac{\partial\psi_1}{\partial x_2} - \dfrac{\partial\psi_2}{\partial x_1}$. This complex is essentially just a rotation of \eqref{eq:deRham}. It gives us full ($s=r$) and reduced ($s=r-1$) direct $H(\textrm{curl})$-approximating elements \begin{equation} \label{eq:generalHCurl} {\mathbf V}_{r,\textrm{curl}}(E) = \nabla\,{\cal{DS}}_{r+1}(E)\oplus(x_2,-x_1)\mathbb P_{s}. \end{equation} These then satisfy the de Rham complexes \begin{align} \label{eq:deRhamVCurl} \mathbb R \DOTSB\lhook\joinrel\longrightarrow {\cal{DS}}_{r+1}(E) \overset{\textrm{grad}\,}{\longlongrightarrow} &{\mathbf V}_{r,\textrm{curl}}^{\textrm{full}}(E) \overset{\textrm{curl}\,}{\longlongrightarrow} \mathbb P_{r}(E) \longrightarrow 0,\\ \label{eq:deRhamVredCurl} \mathbb R \DOTSB\lhook\joinrel\longrightarrow {\cal{DS}}_{r+1}(E) \overset{\textrm{grad}\,}{\longlongrightarrow} &{\mathbf V}_{r,\textrm{curl}}^{\textrm{red}}(E) \overset{\textrm{curl}\,}{\longlongrightarrow} \mathbb P_{r-1}(E) \longrightarrow 0, \end{align} for any variant of our new direct serendipity spaces. We can merge these elements globally, since tangential derivatives of $\phi$, say, map to tangential components of $\nabla\phi$, which are both $\nabla\phi\cdot\tau$. So we have full and reduced direct $H(\textrm{curl})$-approximating spaces over meshes of quadrilaterals.

6942db8c7b504b9c0e68df9f7903aa88f7c73217

\section{Introduction} In most markets, sellers have to deal with competition. On modern e-commerce platforms, it has become easy to observe and to adjust prices. On such platforms, sales do not only depend on a sellers' prices but also on competitors' prices. Moreover, customers take further product features into account, e.g., product quality, ratings, shipping time, shipping costs, etc. \subsection{Motivation} Modern market platforms such as Amazon Marketplace or eBay are highly dynamic as sellers can observe the current market situation at any time and adjust their prices instantly. Sellers typically rival against dozens of competitors, decide on prices for thousands of products, and face steadily changing market situations, cf. Profitero (2014). As electronic markets are constantly growing and the costs of price changes are close to free, intelligent automated repricing systems become increasingly important for practitioners, see Popescu (2016). Since demand information is limited and market dynamics are uncertain, deriving intelligent pricing strategies is highly complex. Pricing strategies are required to take a multitude of dimensions for each competitor into account (e.g., price, quality, shipping time, shipping costs, ratings). Further, aspects such as discounting, cash-flows, and inventory holding costs also have to be included. In this paper, we study oligopoly pricing models in a stochastic dynamic finite horizon framework. We consider single-product problems. We seek to deal with the following real-life aspects: (i) \hspace{0.3cm}demand probabilities are unknown, (ii) \hspace{0.20cm}products have multiple offer dimensions (price, quality, ratings, shipping time, etc.), (iii) \hspace{0.11cm}products are offered by multiple competitors, (iv) \hspace{0.11cm}competitors' sales, inventories, and strategies are not observable. \medskip The setting of limited and asymmetric information, multiple offer dimensions and multiple competitors is highly challenging. To compute viable pricing strategies in competitive settings, we introduce a dynamic programming model that circumvents the curse of dimensionality. We show how to calibrate the model using data-driven demand estimations. To illustrate the applicability of our approach, we use reproducible numerical examples. Finally, we present the performance result of our strategy applied on Amazon. \subsection{Literature Review} The best way to sell products is employing a classical application of revenue management theory. The problem is closely related to the field of dynamic pricing, which is summarized in the books by Talluri, van Ryzin (2004), Phillips (2005), and Yeoman, McMahon-Beattie (2011). The surveys by Chen, Chen (2015) and Den Boer (2015a) provide an excellent overview of recent pricing models under competition. In the article by Gallego, Wang (2014) the authors consider a continuous time multi-product oligopoly for differentiated perishable goods. They use optimality conditions to reduce the multi-dimensional dynamic pure pricing problem to a one-dimensional one. Gallego, Hu (2014) analyze structural properties of equilibrium strategies in more general oligopoly models for the sale of perishable products. The solution of their model is based on a deterministic version of the model. Martinez-de-Albeniz, Talluri (2011) consider duopoly and oligopoly pricing models for identical products. They use a general stochastic counting process to model customer's demand. Further related models are studied by Yang, Xia (2013), Wu, Wu (2015), and Schlosser (2017). Dynamic pricing models under competition that also include strategic customers are analyzed by Levin et al. (2009) and Liu, Zhang (2013). In most existing models, the demand intensity is assumed to be known. Dynamic pricing competition models with \textit{limited} demand information are analyzed by Tsai, Hung (2009), Adida, Perakis (2010), Chung et al. (2012), and Den Boer (2015b) using robust optimization and learning approaches. For a more comprehensive review, we refer to Chen, Chen (2015). In contrast to the assumptions of most publications, in real-life applications, specific information is not observable, demand and price reactions are typically unknown, and customers as well as sellers might not act rational. Moreover, when dealing with dynamic pricing competition models, the most critical problem is their high complexity and the enormous number of potential market situations (cf. curse of dimensionality). Thus, most solution approaches are only applicable if specific assumptions can be verified and the number of competitors is small. In general, dynamic pricing competition problems with incomplete information are \textit{not tractable} and cannot be optimally solved. Hence, in real-life applications, heuristic strategies have to be used. In practice, rule-based strategies as well as data-driven strategies are used, cf. Popescu (2016). As the evaluation of complex competing pricing strategies is also not analytically tractable, it is impossible to give performance guarantees of a specific strategy -- even if the underlying customer behavior as well as the competing strategies are known. Hence, simulation studies have to be used to compare the performance of dynamic pricing strategies and to analyze their strategic interaction in different oligopoly settings, see, e.g., Kephart et al. (2000) or Serth et al. (2017). \subsection{Contribution} We tackle the problem of identifying efficient pricing strategies that are applicable in real-life scenarios. In this paper, we present the following contributions: (i) \hspace{0.15cm We show how to compute dynamic price adjustments to be applied in practice (Sec. 3). (ii) \hspace{0.05cm}We demonstrate that our heuristic is applicable even if the number of competitors is large and the \hspace{0.7cm}competitors' strategies are unknown (Sec. 4). (iii) We verify the performance of our pricing strategy by comparing it to upper bounds, which are \hspace{0.7cm}obtained by optimal strategies that take advantage of price anticipations (Sec. 5). (iv) We find that higher price adjustment frequencies can easily overcompensate a loss in expected profits \hspace{0.7cm}due to the lack of price anticipations. (v)\hspace{0.11cm} We successfully applied our strategy on the Amazon Marketplace (Sec. 6). \medskip We present a dynamic programming model that allows for high dimensional market situations characterized by offers of several competitors with multiple features. We show how real-life market data consisting of competitors' offers, our offers, and our sales can be used to estimate sales probabilities for arbitrary market situations. This way, the impact of market dynamics and competitors' price reactions are implicitly taken into account without anticipating future market situations explicitly. Further, we use a decomposition approach to compute pricing decisions for single situations separately. This makes it possible to efficiently compute price adjustments that are based on current market situations. Finally, we compensate the lack of perfect price anticipations by frequent price adjustments, which in turn are possible as the model allows for rapid re-computations of optimized prices. Further, we present an extended model which takes price adjustment costs into account. This allows to combine fast reaction times and a limited use of price updates. Our approach is successfully applied in practice. The model was calibrated using historical market data from the Amazon Marketplace. Results show that our strategy outperforms the established rule-based strategy of an experienced seller regarding profitability \textit{and} speed of sales. \section{Model Description} We consider the situation where a firm wants to sell a finite number of perishable goods on a digital market platform (e.g., Amazon or eBay). For the products, there are several competitors. In our model, we allow that customers compare prices. They also might take additional product features, such as qualities or ratings of different competitors into account. We assume that items cannot be reproduced or reordered. The initial number of items to sell is $N$, $N \le \infty$, and the time horizon is finite. If a sale takes place, shipping costs $c$ have to be paid, $c \ge 0$. A sale of one item at price $a$ leads to a net revenue of $a-c$. Moreover, we consider inventory holding costs. We assume that each unsold item leads to inventory costs of $l$ per period (e.g., one hour or one day), $l \ge 0$. Furthermore, we use the discount factor $\delta$, $0 < \delta \le 1$, for a period of length one. In our discrete time model, we consider sales probabilities denoted by $P$. Due to customer choice, sales probabilities will particularly depend on our offer price $a$ and the competitors' prices. Moreover, we allow sales probabilities to depend on time. In our model, $P$ is a general function of our offer price $a$ and a market situation denoted by $\vec s$. The market situation $\vec s$ is a vector which includes all relevant observable quantities of interest, such as time $t$, the competitors' prices $\vec p$, customer ratings $\vec v$, product conditions, etc. (see, e.g., Kachani, Shmatov (2010)). The probability to sell exactly $i$ items within the time span $(t,t + 1)$ in a \textit{stable} market situation $\vec s = (t,\vec p,\vec v,...)$ is denoted by, $t = 0,1,2,...,T-1$, $a \ge 0$, $i=0,1,2,...$, \begin{equation} \label{2.1} {P_t}(i,a,\vec s). \end{equation} In real-life scenarios the sales probabilities ${P_t}(i,a,\vec s)$ are typically not known and have to be estimated using data-driven approaches. In simulation models, the sales probabilities ${P_t}(i,a,\vec s)$ can be specifically defined and allow to verify the quality of data-driven estimations as well as the performance of pricing strategies. For instance, assuming a given sales intensity $\lambda = {\lambda _t}(a,\vec s)$ and a scaling factor $d$, we can define Poisson distributed sales probabilities for various intervals, $t = 0,1,2,...,T-1$, $i = 0,1,2,...$, $a \ge 0$, $d>0$, \[\blau{{P_t}(i,a,\vec s): = Pois\left( {d \cdot {\lambda _t}(a,\vec s)} \right). }\] \smallskip The random inventory level at time $t$ -- the beginning of period $t$ -- is denoted by ${X_t}$, $t = 0,1,...,T$. The end of sale (for our firm) is either time $T$ or the random time $\tau$, when all $N$ products are sold. For each period $t$, a new offer price $a$ can be chosen. We call pricing strategies admissible if they are non-anticipating (Markov policies); pricing decisions ${a_t} \ge 0$ may depend on time $t$, the own inventory level ${X_t}$, and the current market situation (denoted by ${\vec S_t}$), which particularly contains the current prices of the competitors. A list of variables and parameters is given in the Appendix, cf. Table A.7. The firm's profits are defined by each period's realized sales (cf. $X_{t} - X_{t+1}$) and inventory holding costs (cf. $l \cdot X_{t}$). For chosen prices $a_t$, the random accumulated profit $G_t$ from time $t$ on (discounted on time $t$) amounts to, $t = 0,1,...,T$, \begin{equation} \label{2.3} {G_t}: = \sum\limits_{s = t}^{T-1} {{\delta ^{s - t}} \cdot \left( {({a_s} - c) \cdot \left( {{X_{s}} - {X_{s+1}}} \right) - l \cdot {X_{s}}} \right)}. \medskip \end{equation} We look for a pricing policy that maximizes expected (discounted) future profits $E({G_t}\left| {{X_t} = n,} \right.{\vec S_t} = {\vec s_t})$ for all times $t$, inventories $n$, and occurring market situations ${\vec s_t}$. In the following sections, we will solve dynamic pricing problems that are related to \eqref{2.1} - \eqref{2.3}. In the next section, we propose our heuristic pricing strategy for scenarios with many competitors and unknown strategies. The applicability of our strategy is demonstrated in Section 4. In Section 5, we measure the performance of our pricing strategy by studying various settings in which competitors frequently and strategically adjust their prices. In Section 6, our approach is used in a live production setting on Amazon. \section{Circumventing the Curse of Dimensionality: A Heuristic Approach In this section, we derive viable pricing strategies for markets with many competitors. We use an efficient algorithm to circumvent the curse of dimensionality and propose an effective heuristic pricing strategy. There are two major problems to derive applicable dynamic programming strategies in competitive markets: (i) as demand is affected by many parameters (e.g., dozens of competitors' prices) a model's state space explodes, and (ii) in general, competitors' strategies are not known, and competitors' price adjustments and other market evolutions cannot be fully anticipated. Our approach deals with both problems. Most importantly, we subsequently compute prices for \textit{one period} only based on the current market situations that occur during a sales process. To compute \blue{a price for the next time period}, in general, the current market situation (current state) as well as potential evolutions of the market (future states) have to be taken into account. As price reactions of competitors occur with a certain delay, the \textit{short-term} evolution of the market can be well approximated by the current market situation. The \textit{long-term} evolution of the market, however, can hardly be predicted. In our use-case, the optimal price for one period mostly depends on the current state and is less affected by possible market evolutions in the far future. Note, while the expected profit of the next period highly depends on the current price decision and the current market situation, the expected future profits are typically much less influenced by the current price decision. This is due to the following: First, future profits are determined by various upcoming decisions that account for future states. Second, our price update typically hardly affects the long-term evolution of the market (especially if many firms are involved). Third, in markets in which price patterns are cyclic or fluctuate around certain reference prices, it can be assumed that the best expected long-term profits are similar for different market situations. Hence, instead of trying to predict possible market evolutions in the future, in our approach, we compute price updates based on an accurate calculation of short-term profits and expected long-term profits that are approximated based on current market conditions. For a current state, we manage problem (i) as follows: We roughly approximate future market situations by using sticky prices. While the degree of inaccuracy is acceptable, we gain a \textit{structure} that makes it possible to circumvent the curse of dimensionality, cf. problem (i), as the states of our dynamic system (i.e., the market situation) are not coupled and can be easily decomposed. Thus, decisions for single market situation can be computed independently from others. The second key idea is to compensate the ``sticky'' model's inaccuracy as well as the lack of correct price anticipations, cf. problem (i), by \textit{frequent} price adjustments, which in turn, are possible as the model's simplicity allows for fast re-computations. Due to price adjustments, the exit, or entry of firms, in general, market situations are not stable. In our model, we differentiate between sales probabilities ${P_t}(i,a,\vec s)$ for \textit{stable} market situations, cf. \eqref{2.1}, and \textit{conditional} sales probabilities, $t = 0,1,...,T-1$, $a \ge 0$, $i = 0,1,2,...$, \begin{equation} \label{2.2} {\tilde P_t}(i,a|\vec s) \end{equation} for selling $i$ items within \blue{the time interval} $(t,t+1)$ at price $a$ under the condition that at time $t$ the market situation is $\vec s$ \blue{(but \textit{may change} within the period)}. Note, the probabilities ${\tilde P_t}(i,a|\vec s)$, cf. \eqref{2.2}, can be estimated from real-life data \blue{where market situations are typically not stable}. As our approach to compute price adjustments is designed to be applied in unstable markets, it will be based on (estimated) conditional probabilities ${\tilde P_t}(i,a|\vec s)$\blue{, cf. (3)}. As described in the beginning of the section, we use a simplified dynamic programming approach in which expected future profits $E({G_t}|{X_t} = n{,_{}}{\vec S_t} = \vec s)$, cf. \eqref{2.3}, are described by the value function ${V_t}(n,\vec s)$, $t = 0,1,...,T$, $n = 0,1,...,N$. If all items are sold or the time horizon is over, no future profits can be made, i.e., for any \blue{market situation} $\vec s$ the natural boundary conditions of the value function are given by, $n = 0,1,...,N$, $t = 0,1,...,T$, \begin{equation} \label{4.1} V_T^{}(n,\vec s) = 0, \end{equation} \begin{equation} \label{4.2} V_t^{}(0,\vec s) = 0. \end{equation} The remaining values ${V_t}(n,\vec s)$ \blue{are given by the Hamilton-Jacobi-Bellman (HJB) equation}, $n = 1,...,N$, $t = 0,1,...,T-1$, \begin{equation} \label{4.3} {V_t}(n,\vec s) = \mathop {\max }\limits_{a \in A} \left\{ {\sum\limits_{i \ge 0} {{{\tilde P}_t}(i,a|\vec s)} \cdot \left( {(a - c) \cdot \min (n,i) - n \cdot l + \delta \cdot {V_{t + 1}}\left( {{{(n - i)}^ + },\vec s} \right)} \right)} \right\} \end{equation} and can be computed recursively. The set of admissible prices $A$ can be continuous or discrete. Finally, the pricing strategy is determined by the arg max of \eqref{4.3}, $n = 1,...,N$, $t = 0,1,...,T-1$, \begin{equation} \label{4.4} {a_t}(n,\vec s) = \mathop {\arg \max }\limits_{a \in A} \left\{ {\sum\limits_{i \ge 0} {{{\tilde P}_t}(i,a|\vec s)} \cdot \left( {(a - c) \cdot \min (n,i) - n \cdot l + \delta \cdot V_{t + 1}^{}\left( {{{(n - i)}^ + },\vec s} \right)} \right)} \right\}. \end{equation} \noindent In case the prices determined by \eqref{4.4} are not unique, we choose the largest one. Note, due to the size of the state space it is not possible to compute prices ${a_t}(n,\vec s)$ for \textit{all} states $\vec s$ in advance. The following algorithm, however, circumvents the curse of dimensionality \blue{by considering just single states and computing feedback strategies multiple times. This allows to derive} viable heuristic pricing strategies in competitive markets with a large number of competitors. \bigskip \noindent \textbf{\blue{Algorithm 3.1}} We propose the following pricing heuristic: \smallskip (Step 1) \hspace{0.19cm} For the current point in time $t$, observe the new state, i.e., the current inventory level ${X_t}$ \hspace{1.7cm}and the current market situation ${\vec S_t}$. (Step 2) \hspace{0.18cm} Use $T - t$ recursion steps and the probabilities ${\tilde P_s}(i,a|{\vec S_t})$, $s=t,...,T-1$, $i = 0,1,...,{X_t}$, \hspace{1.72cm}$a \in A$, to compute the specific value ${V_t}({X_t},{\vec S_t})$, cf. \eqref{4.3}. (Step 3) \hspace{0.22cm} Choose price $a_t^{}({X_t},{\vec S_t})$, which is associated to the last step of the recursion, cf. \eqref{4.4}. \hspace{1.72cm}At the point in time of the next price adjustment, i.e., $t+1$, go to Step (1) and re-solve the \hspace{1.72cm}system \blue{(6) - (7) based on the new inventory $X_{t+1}$ and} the new state $\vec S_{t+1}$. \medskip \blue{We just have to} compute prices for single (current) market situations and to regularly refresh prices in response to changing market situations. Note, a single recomputation \blue{of (6) - (7)} is very fast and does neither increase with the number of competitors nor increase with the number of dimensions of the market situation. \section{Application of the Heuristic Strategy} \subsection{Sales Probabilities} In this section, we assume that market situations are \blue{only} characterized by the point in time $t$ and the prices of $K$ competitors. The vector of prices $\vec p := (p_1,...,p_K)$ does not necessarily have to be ordered. To be able to demonstrate the applicability and the performance of pricing strategies, cf. next Section 5, we need to quantify realistic sales probabilities in competitive markets. Instead of using made-up or highly stylized demand probabilities (e.g., of linear, exponential, or isoelastic type), we seek to use a more general (data-driven) demand setting. As an example, we define such probabilities based on a logistic regression model using linear combinations of state-dependent features/regressors $\vec x = \vec x(a,\vec s)$ and given coefficients $\vec \beta $; i.e., we consider binary probabilities of the form, $t = 0,1,...,T$, $a \ge 0$, $i \in \{0,1\}$, \[{\tilde P_t}(i,a|\vec s):= i \cdot \hat P(a,\vec s) + (1-i) \cdot (1-\hat P(a,\vec s)),\] where the logit probability $\hat P(a,\vec s)$ is given by \begin{equation} \label{4.5} \hat P(a,\vec s) := {\frac{{{e^{\vec x(a,\vec s)'\vec \beta }}}}{{1 + {e^{\vec x(a,\vec s)'\vec \beta }}}}}. \end{equation} \blue{The dependent variable is the number of sales within a certain time frame}. In the following definition, we give simple examples of explanatory variables. \blue{In this framework}, further explanatory variables can be easily defined to capture the impact of various effects, such as customer ratings, product quality, shipping time, the type of product (categories/clusters), etc. \smallskip \begin{definition} \blue{Let market situations be described by $\vec s = (t,\vec p)$. For potential offer prices $a$}, we define the following regressors $\vec x = \vec x(a,\vec s)$: (i)\hspace{0.4cm} $x_1(a,\vec s) := 1$ \hspace{3.2cm} \blue{constant/intercept} (ii)\hspace{0.3cm} $x_2(a,\vec s):= r(a,\vec p)$ \hspace{2.4cm} rank of price $a_{}$ within the prices $\vec p_{}$, where \[ \blue{ r(a,\vec p):= 1 + card ({\left\{ {k = 1,...,K_{}\left| {p_k < a} \right.} \right\}}) + 0.5 \cdot card ({\left\{ {k = 1,...,K\left| {p_k = a} \right.} \right\}}) } \] (iii)\hspace{0.2cm} $x_3(a,\vec s):= a_{} - \mathop {\min }\limits_{k = 1,...,K_{}} \{ p_k\} $ \hspace{0.87cm} price gap between \blue{price} $a_{}$ and the best competitor's price (iv)\hspace{0.23cm} $x_4(a,\vec s):= K_{}$ \hspace{3.1cm} total number of competitors for product $i$ in period $t$ (v)\hspace{0.32cm} $x_5(a,\vec s):= (a_{} + \sum\nolimits_k {p_k} )/(1 + {K})$ \hspace{0.2cm} average price level for product $i$ in period $t$ (vi)\hspace{0.25cm} $x_6^{}(a,\vec s): = {\alpha _1} \cdot {t^{{\alpha _2}}}$ \hspace{2.3cm} time dependence, e.g., via scaling parameters $\alpha_1$, $\alpha_2$ \label{def1} \end{definition} \begin{example} \blue{Let market situations} be characterized by $\vec s = \vec p$. We consider the five variables $x_m(a,\vec p)$, $m = 1,...,5$, see Definition \ref{def1} (i)-(v). Inspired by a production data set obtained from Amazon, see Section 6, we let \[\vec \beta = (\beta _1^{},\beta _2^{},\beta _3^{},\beta _4^{},\beta _5^{}) = ({\rm{ - 3}}{\rm{.89}}{,_{}}{\rm{ - 0}}{\rm{.56}}{{\rm{,}}_{}}{\rm{ - 0}}{\rm{.01}}{{\rm{,}}_{}}{\rm{0}}{\rm{.07}}{{\rm{,}}_{}}{\rm{ - 0}}{\rm{.05}}).\] \blue{In this context}, we define binary sales probabilities for a given market situation $\vec p$ \blue{and offer prices $a$} via $\hat P(a,\vec p):= {e^{\vec x(a,\vec p)'\vec \beta }}/(1 + {e^{\vec x(a,\vec p)'\vec \beta }})$, cf. \eqref{4.5}. In this example, we consider the following market situation \blue{with $K=10$ competitor prices}: \[\vec p = \left( {{{5.18, }_{}}{{5.96, }_{}}{{6.31, }_{}}{{8.28, }_{}}{{9.48, }_{}}{{9.88, }_{}}{{10.33, }_{}}{{10.98, }_{}}{{11.67, }_{}}13.52} \right).\] \label{ex1} \end{example} \vspace{-0.5cm} The demand coefficients defined in Example \ref{ex1} are based on a large data set ($>$20 M observations/month) from the Amazon market for used books including $100\,000$ different products and up to $K=20$ competitors, cf. Schlosser et al. (2016). The average length of a period is 2.4 hours. Given the market situation $\vec p$ described in Example \ref{ex1}, Figure 1a illustrates sales probabilities $\hat P(a,\vec p)$ for different potential offer prices $a$. \blue{As expected, we observe that the sales probability decreases with the price $a$, cf. Definition 4.1 (ii) and $\beta_2:=-0.56<0$}. The jumps in Figure 1a occur whenever the price rank changes. Note, undercutting a \blue{competitor's price} yields a better price rank, which can easily double the sales probability. The corresponding expected profits $(a - c) \cdot \hat P(a,\vec p)$ for one period of time are shown in Figure 1b. Expected profits are positive as long as $a$ exceeds the shipping costs $c$. The peaks in Figure 1b are caused by the price rank related jumps of $\hat P(a,\vec p)$. In this example, expected profits are maximized if one slightly undercuts the \blue{best competitors' price, i.e., $p_1 = 5.18$}. \begin{figure}[ht] \begin{center} \includegraphics[scale=0.5]{fig1a}\qquad \includegraphics[scale=0.5]{fig1b} \caption{Predicted sales probabilities (left window 1a) and expected profits for one period (right window 1b) \blue{for different potential offer prices $a \in A:= \left\{ {0.01,0.02,...,20} \right\}$, shipping costs $c = 3$}; Example \ref{ex1}.} \label{fig1} \end{center} \end{figure} \begin{remark} (Model calibration and data-driven demand estimations) Our approach to model sales probabilities is simple yet reasonable, especially for practitioners. Logistic regression is a well-understood method and allows to handle hundreds of features and millions of observations. Regression results are robust and can be directly interpreted. Nevertheless, our model also allows to calibrate sales probabilities using other approaches (e.g., using decision trees, see Quinlan (1986), or gradient-boosted decision trees, see Chen, Guestrin (2016), etc.). \label{rem3} \end{remark} \subsection{Numerical Examples} In the following, we demonstrate the applicability of our repricing approach in competitive markets, cf. Algorithm 3.1. We illustrate the computation of the heuristic strategy using the sales probabilities defined in Section 4.1. \smallskip \begin{example} We consider the setting of Example \ref{ex1}. We let $T=100$, $c=3$, $\delta=0.9995$, $N=25$, $l=0.01$, and assume the exemplary state $\vec s = \vec p = \left( {{{5.18,}_{}}{{5.96,}_{}}{{6.31,}_{}}{{8.28,}_{}}{{9.48,}_{}}{{9.88,}_{}}{{10.33,}_{}}{{10.98,}_{}}{{11.67,}_{}}13.52} \right)$. We illustrate state-dependent prices \blue{of our heuristic}, cf. \eqref{4.3}-\eqref{4.4}, in case of the Poisson probabilities ${\tilde P_t}(i,a|\vec s):= Pois( {d \cdot \hat P(a,\vec p)} )$, $i=0,1,...$, $d=10$, $a \in A:= \left\{ {0.01,0.02,...,20} \right\}$. \label{ex2} \end{example} \vspace{0.3cm} \begin{figure}[ht] \begin{center} \includegraphics[scale=0.45, trim = 22mm 0mm 0mm 0mm, clip]{fig2a}\qquad \includegraphics[scale=0.45, trim = 22mm 0mm 0mm 0mm, clip]{fig2b} \caption{\blue{Illustration of the value function for different points in time $t$ (left window 2a) and for different inventory levels $n$ (right window 2b) given a specific market situation, $t=0,...,100$, $n=1,...,25$}; Example \ref{ex2}.} \label{fig2} \end{center} \end{figure} \vspace{-0.1cm} The value function (Figure 2) and the associated heuristic feedback prices (Figure 3) are illustrated for the setting of Example \ref{ex2}. \blue{Figure 2a and 2b summarize (approximated) expected profits at a \textit{current} time $t$ in state $n$ given the market situation $\vec s$}. As inventory holding costs are included in the model it is not advantageous to store too many items. \blue{In this example, we observe that at time $0$ one should not store more than 15 items. If 20 periods of time are left ($t=80$) the best expected profit can be gained by offering 5 items. Having more items in stock is unfavourable due to inventory holding costs. } Figure 3 illustrates prices that are recommended by our heuristic at a \textit{current} time $t$ in state $n$ given the market situation $\vec s = \vec p$. We observe that prices are higher the less items are left to sell and the more time is left. For instance, if more than 50 periods of time are left and there is only one item left to sell the price is \blue{$9.47 = p_5 - 0.01$ (rank 5)}. If more than 60 periods of time are left and there are 2 or 3 items left to sell the price is 8.27 (rank 4). If more than 3 items are left to sell the price is either 5.95 (rank 2) or 5.17 (rank 1) depending on the time-to-go. If more than 7 items are left in stock the price ${a_t}(n,\vec s)$ will occupy price rank 1 at every point in time. Note, price rank 3 (6.30) is not chosen at all since the difference between price rank 2 and 3 is too small. \begin{figure}[ht] \begin{center} \includegraphics[scale=0.4, trim = 22mm 0mm 0mm 0mm, clip]{fig3} \caption{\blue{Heuristic feedback pricing policy for a specific market situation, $t=0,...,100$, $n=1,...,10$}; Example \ref{ex2}.} \label{fig3} \end{center} \end{figure} \vspace{-0.15cm} Next, we demonstrate the \textit{applicability} of our approach in competitive markets with many active competitors and unknown strategies. We consider a numerical example that can be reproduced by the reader. \smallskip \begin{example} We consider the setting of Example \ref{ex2}, i.e., we let $T=100$, $c=3$, $\delta=0.9995$, $N=10$, $d=10$, $l=0.01$. We simulate competitors' price trajectories over time as follows. We let competitors adjust their prices using random reaction times and randomized price changes. The probability that a single firm $k$ adjusts its price $p_t^{(k)}$, $k=1,...,K$, at $t = 0,h,...,T - h$, $h$=0.1, is denoted by $\pi$, $\pi = 0.3 \cdot h$. \blue{The random amplitude of price jumps are normalized via $h$, $T$, and $\pi$ such that average trends over time are the same even if the frequency of price jumps (cf. $\pi$) or the period length $h$ are different.} Initial prices ${\vec p_0}$ are chosen as in Example \ref{ex2}. We simulate price trajectories using four scenarios of randomized price adjustments, cf. case (i) -- (iv), $K=10$, $t = 0,h,...,T - h$: \label{ex3} \medskip (i)\hspace{0.38cm} no price trend: \hspace{0.95cm} $p_{t + h}^{(k)} = \max \left( {c + 0.01,p_t^{(k)} + {1_{\{ U(0,1) < \pi \} }} \cdot U( - 20,20) \cdot h/\pi /T} \right)$ (ii)\hspace{0.28cm} positive price trend: \hspace{0.18cm} $p_{t + h}^{(k)} = \max \left( {c + 0.01,p_t^{(k)} + {1_{\{ U(0,1) < \pi \} }} \cdot U( - 15,25) \cdot h/\pi /T} \right)$ (iii)\hspace{0.16cm} negative price trend: \hspace{0.11cm} $p_{t + h}^{(k)} = \max \left( {c + 0.01,p_t^{(k)} + {1_{\{ U(0,1) < \pi \} }} \cdot U( - 25,15) \cdot h/\pi /T} \right)$ \smallskip (iv)\hspace{0.23cm} strategic competitor:\hspace{0.22cm}Consider setting (i), in which the other competitors' price adjustments are \hspace{0.9cm}mutual independent and have no trend. One single firm, however, plays a strategic price adjustment \hspace{0.9cm}strategy by steadily undercutting our price by $\varepsilon$=0.2. \medskip In this example, we also consider stock-outs of firms by simulating competitors to exit the market at random times. Similarly, we allow new competitors to enter the market with a random starting price. Our realized sales are simulated using sales probabilities for time intervals $(t,t+h)$, $t = 0,h,...,T - h$ of size $h$ defined by $P_t^{(h)}(i,a,\vec p):= Pois( {h \cdot d \cdot \hat P(a,\vec p)})$, cf. \eqref{4.5}. Our firm adjusts its price at $t = 0,1,...,T - 1$ according to Algorithm 3.1; for ease of simplicity, we use ${\tilde P_t}(i,a|\vec s):= Pois( {d \cdot \hat P(a,\vec p)})$. \end{example} \smallskip \begin{figure}[ht] \begin{center} \includegraphics[scale=0.32]{fig4a}\qquad \includegraphics[scale=0.32]{fig4b} \smallskip \includegraphics[scale=0.32]{fig4c}\qquad \includegraphics[scale=0.32]{fig4d} \caption{Application of our heuristic dynamic pricing strategy under competition; scenario (i) (upper left window 4a), scenario (ii) (upper right window 4b), scenario (iii) (lower left window 4c) and scenario (iv) (lower right window 4d); Example \ref{ex3}.} \label{fig4} \end{center} \end{figure} For case (i) - (iv) of Example 4.3, the four subfigures of Figure 4 show price trajectories of the competing firms over time. The price trajectory of our firm can be identified by little blue bullet points indicating the price adjustments at times $t$, $t=0,1,...,T-1$, according to the strategy defined in Algorithm 3.1 for the observed states $\vec S_t$ and remaining inventory levels $X_t$. In this context, the black decreasing step function illustrates the remaining inventory level of our firm. The evaluation of each scenario takes less than a second. In scenario (i)-(iii), we observe that the heuristic strategy slightly undercuts other competitors, cf., Figure 4a, 4b, and 4c. Furthermore, we observe two effects. First, if the prices of the competitors are high, then the strategy occupies leading price ranks, see Figure 4b; if competitors' prices are low, it is more profitable to use moderate price ranks, see Figure 4c. Second, at the end of the time horizon the strategy occupies leading price ranks. Scenario (iv) illustrates the case of an aggressive strategic competitor. We observe that the interplay of the strategic competitor's and the price adjustments of our heuristic lead to specific cyclical price patterns, cf. Figure 4d, which have similarities to Edgeworth-price cycles, see, e.g., Maskin, Tirole (1988) or Noel (2007). Whenever the price level is too low, our firm raises the price up to a certain level. Note, the heuristic strategy just responds to market situations, prices are not anticipated. Example \ref{ex3} demonstrates that our heuristic strategy is applicable for any scenario of the competitors' price trajectories -- even if the number of competitors is large and competitors enter or exit the market. The price adjustments of the heuristic depend on time, the inventory level as well as the current market situation. Note, the number of computation steps of Algorithm 3.1 does not increase with the number of competitors. The approach remains also applicable if the competitors' price adjustments are mutually dependent. The repeating price adjustments of our firm allow to respond to any change of the market environment. Moreover, the approach allows to take additional factors, such as ratings, product conditions, or shipping times into account. These extensions will only affect the sales probabilities of the model, e.g., by using additional explanatory variables, cf. Definition 4.1. The simple optimization of the model, however, is not affected, since the number of computation steps of Algorithm 3.1 does not increase. Our approach allows to directly include dozens of features and dozens of competitors without increasing the model's complexity. The following remark summarizes important managerial recommendations. \begin{remark} (Managerial recommendations) (i)\hspace{0.35cm} If market conditions can be assumed to be overall stable, a reasonable initial inventory can be \hspace{0.9cm}determined based on $\mathop {\max }\nolimits_{n \ge 0} \left\{ {{V_0}(n,\vec s)} \right\}$, cf. economic order quantities. (ii)\hspace{0.27cm} It is mostly optimal to undercut one of the competitors' prices by the smallest unit (e.g., 1 cent). \hspace{0.9cm}In this context, Algorithm 3.1 can be accelerated dramatically by considering just a small subset of \hspace{0.9cm}admissible prices, i.e., \[ \tilde A(\vec p):= \bigcup\limits_{k = 1,...,K} {\left\{ {{{({p_k} - \varepsilon )}^ + }} \right\}}. \] (iii)\hspace{0.25cm}As expected, discounting motivates short-term profits and leads to more aggressive, i.e., lower \hspace{0.9cm}feedback prices. Hence, a lower (higher) discount factor $\delta$ can be used to increase (or decrease) \hspace{0.9cm}the overall aggressiveness of the strategy. This "instrument" can also be used to actively manage \hspace{0.9cm}the total inventory of a firm as well as its cash flow. If negative (or positive) price trends can be \hspace{0.9cm}anticipated, cf. Fig. 4b (and 4c), the aggressiveness of the strategy can also be used to further \hspace{0.9cm}improve the heuristic strategy. \label{rem4} \end{remark} \section{Measuring the Performance of the Heuristic Strategy} In this section, we want to measure the \textit{performance} of our strategy derived in Section 3. We consider the case of many active competitors, cf. Section 5.1, as well as the case of strategic competitors that strategically respond to our price adjustments, cf. Section 5.2. In Section 5.3, we analyze the case in which our strategy \blue{is played against itself}. \subsection{Heuristic vs. Optimal Forward-Looking Strategies in Dynamic Oligopoly Competition} \blue{We measure the performance} of our heuristic by comparing its expected profits to \textit{upper bounds} determined by optimal expected profits for the case that the competitors' future prices are \textit{\blue{completely} known} in advance. In this subsection, we assume that the competitors' prices are not affected by our price decisions. Let $H$ \blue{denote the number of subperiods} of one period, and let $h:= 1/H$, e.g., $H=10$ and $h=0.1$. We assume a realization of competitors price trajectories over time via step functions ${\vec p_t}$, $t = 0,h,2h,...,T - h$. \subsubsection{Deriving Upper Bounds: Evaluation of Forward Looking Strategies} As we look for upper bounds, in this subsection, we consider the true probabilities $P_t^{(h)}(i,a,\vec p)$, for the subperiods $(t, t+h)$, $t = 0,h,2h,...,T - h$. If price adjustments are allowed at all points in time $t = 0,h,...,T - h$, the optimal \blue{forward-looking (OFL) pricing policy denoted by $a_t^{OFL(h)}(n;{\vec p_t})$} -- which takes advantage of perfect price anticipations -- is determined by the arg max of, $t = 0,h,...,T-h$, $n = 1,...,N$, $V_T^{OFL(h)}(n;{\vec p_T}):= 0$, \begin{equation} \label{5.1} V_t^{OFL(h)}(n;{\vec p_t}) = \mathop {\max }\limits_{a \in A} \left\{ {\sum\limits_{i \ge 0} {P_t^{(h)}(i,a,{{\vec p}_t})} \cdot \left( {(a - c) \cdot \min (n,i) - n \cdot l \cdot h + {\delta ^h} \cdot V_{t + h}^{OFL(h)}\left( {{{(n - i)}^ + };{{\vec p}_{t + h}}} \right)} \right)} \right\}. \end{equation} Next, we consider a second version of the anticipating strategy above, which adjusts prices \textit{less often}. \blue{Price updates are only allowed at $t = 0,1,...,T - 1$ instead of $t = 0,h,...,T-h$. To still be able to use a recursive computation (backward induction), we use $a^-$ to store our most recent offer price in an extended state space.} \blue{The optimal (forward-looking) pricing policy $a_t^{OFL(1)}(n, a^-;{\vec p_t})$} is determined by the arg max of, $t = 0,h,...,T-h$, ${a^ - } \in A$, $V_T^{OFL(1)}(n,{a^ - };{\vec p_T}):= 0$, \[V_t^{OFL(1)}(n,{a^ - };{\vec p_t}) = \mathop {\max }\limits_{a \in \left\{ {\begin{array}{*{20}{c}} A&{,t = 0 \vee t\bmod 1 = 0}\\ {\{ {a^ - }\} }&{,else} \end{array}} \right.}\] \begin{equation} \label{5.2} \left\{ {\sum\limits_{i \ge 0} {P_t^{(h)}(i,a,{{\vec p}_t})} \cdot \left( {(a - c) \cdot \min (n,i) - n \cdot l \cdot h + {\delta ^h} \cdot V_{t + h}^{OFL(1)}\left( {{{(n - i)}^ + },a;{{\vec p}_{t + h}}} \right)} \right)} \right\}. \end{equation} \smallskip \blue{Note, the action space in \eqref{5.2} is time-dependent and guarantees that price adjustments are only admissible at $t = 0,1,...,T - 1$ while at all other times the fixed previous price $a^-$ is used.} In \eqref{5.2} also perfect price anticipations are used. \blue{Further, using the property} that it is advisable to undercut one of the competitors' prices by the smallest unit $\varepsilon$, cf. Remark \ref{rem4} (ii), we can accelerate the computation of \eqref{5.2} significantly: \blue{(i) instead of $A$, we use the smaller action set} ${\tilde A_t}:= \bigcup\nolimits_{k = 1,...,{K_t}} {\left\{ {{{(p_t^{(k)} - \varepsilon )}^ + }} \right\}} $, $\varepsilon>0$, $t = 0,1,...,T - 1$, and \blue{(ii) we use a smaller state space} by considering only ${a^ - } \in A_t^{( - )}:= {\tilde A_{\left\lfloor {{{(t - h)}^ + }} \right\rfloor }}$, for all points in time $t$, $t = 0,h,...,T$, where $\left\lfloor \cdot \right\rfloor $ is the floor operator. \subsubsection{Evaluation of Non-Anticipating Heuristics} \blue{Next, we show how to analytically evaluate the non-anticipating policy described in Section 3, cf. Algorithm 3.1}. First, we allow prices to be adjusted \blue{only at time} $t = 0,1,...,T - 1$. The corresponding heuristic strategy, cf. Algorithm 3.1., is denoted by ${a^{H(1)}}$. We consider given \blue{(conditional)} probabilities $\tilde P_t^{(1)}(i,a|\vec p)$ for periods $(t, t+1)$, $t = 0,1,...,T - 1$. The prices ${a^{H(1)}}(n,{\vec p_t})$ are determined by the arg max of, $t = 0,h,...,T$, $n = 0,1,...,N$, $V_T^{H(1)}(n;{\vec p_T}):= 0$, \begin{equation} \label{5.3} V_t^{H(1)}(n;{\vec p_t}) = \mathop {\max }\limits_{a \in A} \left\{ {\sum\limits_{i \ge 0} {\tilde P_t^{(1)}(i,a|{{\vec p}_t})} \cdot \left( {(a - c) \cdot \min (n,i) - n \cdot l + \delta \cdot V_{t + 1}^{H(1)}\left( {{{(n - i)}^ + };{{\vec p}_t}} \right)} \right)} \right\}. \end{equation} \blue{Note, \eqref{5.3} defines single prices $a_t^{H(1)}(n,{\vec p_t})$ for time $t$, $t=0,1,...,T-1$, which depend on the current market situation $\vec p_t$ and the current inventory level $n:=X_t$.} The expected profit of this non-anticipating heuristic (NAH) policy $a^{H(1)}$ can be \textit{analytically evaluated} via, $t = 0,h,...,T$, $n = 0,1,...,N$, $V_T^{NAH(1)}(n;{\vec p_T}): = 0$, \[V_t^{NAH(1)}(n;{\vec p_t}) \] \begin{equation} \label{5.4} = \mathop {\max }\limits_{a \in \left\{ {a_{\left\lfloor t \right\rfloor }^{H(1)}(n,{{\vec p}_{\left\lfloor t \right\rfloor }})} \right\}} \left\{ {\sum\limits_{i \ge 0} {P_t^{(h)}(i,a,{{\vec p}_t})} \cdot \left( {(a - c) \cdot \min (n,i) - n \cdot l \cdot h + {\delta ^h} \cdot V_{t + h}^{NAH(1)}\left( {{{(n - i)}^ + };{{\vec p}_{t + h}}} \right)} \right)} \right\}. \end{equation} The recursion \eqref{5.4} is a simple evaluation of the heuristic prices $a_t^{H(1)}(n,{\vec p_t})$, $t = 0,1,...,T - 1$, applied in all subperiods $t = 0,h,...,T$ using the formulation ${a_{\left\lfloor t \right\rfloor }^{H(1)}}(n,{\vec p_{\left\lfloor t \right\rfloor }})$ and the correct probabilities $P_t^{(h)}(i,a,{\vec p_t})$. \blue{Note, in (12) no maximization is used as in each state the max operator only evaluates the prices $a = {a_{\left\lfloor t \right\rfloor }^{H(1)}}(n,{\vec p_{\left\lfloor t \right\rfloor }})$ which are uniquely determined by \eqref{5.3}.} \bigskip Next, we also consider a second version of the heuristic above, which adjusts prices \textit{more often}. Allowing the heuristic, cf. Algorithm 3.1., to adjust prices at all $t = 0,h,...,T - h$, we define the corresponding policy $a_t^{H(h)}(n,{\vec p_t})$ by the arg max of, $t = 0,h,...,T$, $n = 0,1,...,N$, $V_T^{H(h)}(n;{\vec p_T}):= 0$, \begin{equation} \label{5.5} V_t^{H(h)}(n;{\vec p_t}) = \mathop {\max }\limits_{a \in \left\{ {a_t^{H(h)}(n,{{\vec p}_t})} \right\}} \left\{ {\sum\limits_{i \ge 0} {\tilde P_t^{(h)}(i,a|{{\vec p}_t})} \cdot \left( {(a - c) \cdot \min (n,i) - n \cdot l \cdot h + {\delta ^h} \cdot V_{t + h}^{H(h)}\left( {{{(n - i)}^ + };{{\vec p}_t}} \right)} \right)} \right\}. \end{equation} which is a generalization of \eqref{5.3} using \blue{periods of length $h$ instaed of 1 and using $\tilde P_t^{(h)}(i,a|{{\vec p}_t})$ instead of $\tilde P_t^{(1)}(i,a|{{\vec p}_t})$}. An evaluation of the second \blue{fast adjusting} non-anticipating heuristic $a_{}^{H(h)}$ yields the expected profits, $t = 0,h,...,T$, $V_T^{NAH(h)}(n;{\vec p_T}):= 0$, \[ V_t^{NAH(h)}(n;{\vec p_t})\] \begin{equation} \label{5.6} = \mathop {\max }\limits_{a \in \left\{ {a_t^{H(h)}(n,{{\vec p}_t})} \right\}} \left\{ {\sum\limits_{i \ge 0} {P_t^{(h)}(i,a,{{\vec p}_t})} \cdot \left( {(a - c) \cdot \min (n,i) - n \cdot l \cdot h + {\delta ^h} \cdot V_{t + h}^{NAH(h)}\left( {{{(n - i)}^ + };{{\vec p}_{t + h}}} \right)} \right)} \right\}. \end{equation} \subsubsection{Comparison of Strategies} Comparing adjustment frequencies and information structures, we obtain relations between the expected profits of \blue{forward-looking strategies, cf. \eqref{5.1}, \eqref{5.2}, and non-anticipating heuristics}, cf. \eqref{5.4}, \eqref{5.6}. \begin{lemma} For any given scenario of price trajectories ${\vec p_t}$, $t = 0,h,...,T$, for all $n = 1,...,N$, $t = 0,h,...,T - h$, ${a^ - } \in A$, we have \medskip (i)\hspace{0.35cm} $V_t^{OFL(h)}(n) \ge V_t^{OFL(1)}(n,{a^ - }) \ge V_t^{NAH(1)}(n)$, \medskip (ii)\hspace{0.29cm} $V_t^{OFL(h)}(n) \ge V_t^{NAH(h)}(n,{a^ - }) \ge V_t^{NAH(1)}(n)$. \medskip Proof. The first inequality of (i) holds since in \eqref{5.1} prices can be adjusted at additional points in time compared to \eqref{5.2}. The second inequality of (i) holds since in \eqref{5.2} the set of admissible prices includes those of \eqref{5.4}. Assertion (ii) follows from similar arguments. \label{lem1} \end{lemma} \blue{Finally, we quantify the performance of our heuristic strategies} compared to the upper bounds derived. \begin{example} Consider the setting of Example \ref{ex3} (i)-(iii), $T=100$, $c=3$, $\delta=0.9995$, $N=10$, $d=10$, $l=0.01$, \blue{cf. Figure 4a, 4b, and 4c}. Now, the initial competitors' prices are randomized using ${\vec p_0}: = U(5,15)$. The probability that a single firm $k$ adjusts its \blue{price $p_t^{(k)}$} at $t = 0,h,...,T - h$, is denoted by $\pi \in (0,1)$. We simulate price trajectories using three families of randomized price adjustments, \blue{$k = 1,...,K$, $K=10$}, $t = 0,h,...,T - h$, $h=0.1$, and different \blue{adjustment probabilities} $\pi=0.01,0.03,0.1,0.3$: \medskip (i)\hspace{0.35cm} no price trend: \hspace{0.9cm} $p_{t + h}^{(k)} = \max \left( {c + 0.01,p_t^{(k)} + {1_{\{ U(0,1) < \pi \} }} \cdot U( - 20,20) \cdot h/\pi /T} \right)$ (ii)\hspace{0.24cm} positive price trend: \hspace{0.15cm} $p_{t + h}^{(k)} = \max \left( {c + 0.01,p_t^{(k)} + {1_{\{ U(0,1) < \pi \} }} \cdot U( - 15,25) \cdot h/\pi /T} \right)$ (iii)\hspace{0.12cm} negative price trend: \hspace{0.07cm} $p_{t + h}^{(k)} = \max \left( {c + 0.01,p_t^{(k)} + {1_{\{ U(0,1) < \pi \} }} \cdot U( - 25,15) \cdot h/\pi /T} \right)$ \medskip In the three settings of Example \ref{ex4}, we apply the following five different strategies, cf. \eqref{5.1}-\eqref{5.6}: \smallskip (A)\hspace{0.17cm} Frequent informed:\hspace{0.4cm} adjustments in $t = 0,h,...,T - h$, full anticipation, cf. ${V^{OFL(h)}}$ (B)\hspace{0.2cm} Relaxed informed:\hspace{0.52cm} adjustments in $t = 0,1,...,T - 1$, full anticipation, cf. ${V^{OFL(1)}}$ (C)\hspace{0.2cm} Frequent heuristic:\hspace{0.4cm} adjustments in $t = 0,h,...,T - h$, no anticipation, cf. ${V^{NAH(h)}}$ (D)\hspace{0.2cm} Relaxed heuristic:\hspace{0.52cm} adjustments in $t = 0,1,...,T - 1$, no anticipation, cf. ${V^{NAH(1)}}$ (E)\hspace{0.2cm} Optimal fix price:\hspace{0.6cm} adjustment only in $t = 0$, full anticipation \medskip \noindent We use $P_t^{(h)}(i,a,\vec p):= Pois( {h \cdot d \cdot \hat P(a,\vec p)})$ and $\tilde P_t^{(\Delta)}(i,a|\vec p):= Pois( {\Delta \cdot d \cdot \hat P(a,\vec p)}$, $\Delta \in \{h,1\}$. \label{ex4} \end{example} \smallskip A comparison of the performance of the different strategies (A) - (E) applied in the \blue{three} settings (i) - (iii) of Example \ref{ex4} are summarized in Table 1. \blue{For different competitors' price adjustment frequencies $\pi$} the table contains the average value of the expected profits $V_0^{(A)}(N;{\vec p_0})$ of the benchmark strategy A \blue{for various random scenarios ${\vec p_t}$}. The expected profit $V_0^{(A)}(N;{\vec p_0})$ represents an upper bound as full information of future prices is used and the possibility of price adjustments is the largest. The remaining columns of the table contain the average ratio of the expected profits of the strategies (B) - (E) compared to strategy (A). For each simulated competitive scenario the expected profits of all five strategies have been evaluated analytically. For each setting (i) - (iii) 1000 scenarios of \blue{trajectories $\vec p_t$, $t = 0,h,...,T - h$, were evaluated.} \begin{table}[ht] \begin{center} \setlength{\tabcolsep}{4mm} \begin{tabular}{c ccc ccc} \toprule ${\vec p_t}$ & $\pi$ &$V_0^{(A)}(N;{\vec p_0})$ &$\frac{{V_0^{(B)}(N;{{\vec p}_0})}}{{V_0^{(A)}(N;{{\vec p}_0})}}$ &$\frac{{V_0^{(C)}(N;{{\vec p}_0})}}{{V_0^{(A)}(N;{{\vec p}_0})}}$ &$\frac{{V_0^{(D)}(N;{{\vec p}_0})}}{{V_0^{(A)}(N;{{\vec p}_0})}}$ &$\frac{{V_0^{(E)}(N;{{\vec p}_0})}}{{V_0^{(A)}(N;{{\vec p}_0})}}$ \\ \midrule (i) &0.01 &22.53 &0.983 &0.986 &0.964 &0.706 \\ (i) &0.03 &25.31 &0.973 &0.985 &0.953 &0.760 \\ (i) &0.10 &26.73 &0.948 &0.987 &0.926 &0.802 \\ (i) &0.30 &26.87 &0.911 &0.990 &0.884 &0.836 \\ \midrule (ii) &0.01 &36.11 &0.990 &0.964 &0.954 &0.781 \\ (ii) &0.03 &41.50 &0.983 &0.943 &0.930 &0.786 \\ (ii) &0.10 &43.96 &0.968 &0.932 &0.910 &0.780 \\ (ii) &0.30 &45.08 &0.953 &0.926 &0.898 &0.772 \\ \midrule (iii) &0.01 &10.83 &0.959 &0.976 &0.920 &0.345 \\ (iii) &0.03 &12.04 &0.930 &0.984 &0.891 &0.440 \\ (iii) &0.10 &12.51 &0.844 &0.986 &0.793 &0.469 \\ (iii) &0.30 &12.37 &0.686 &0.984 &0.629 &0.474 \\ \bottomrule \end{tabular} \caption{Expected profits of the heuristic strategy compared to upper bounds for simulated competitive scenarios (i)-(iii), $T=100$, $N=10$, $d=10$, $h=0.1$, and $\pi=0.01,0.03,0.1,0.3$; Example \ref{ex4}. \label{tab1}} \end{center} \end{table} The relaxed heuristic strategy (D) yields 63-96\% of the optimal results; the frequent heuristic strategy (C) even obtains 93-99\%. The performance of the strategies (B) and (C) is in between those of strategy (A) and (D), cf. Lemma \ref{lem1}. We observe that the more volatile the market, the more important become adjustment frequencies. Our example shows that the frequent heuristic strategy (C) can even beat the relaxed informed strategy (B). \blue{The best \textit{fix price} strategy (E)} only yields 35-84\% of the optimal profits, although future prices are fully known. A boxplot of Table 1 is given in the Appendix, see Figure A.10. Note, while \blue{in Example 5.1 $P_t^{(h)}$ and $\tilde P_t^{(h)}$ coincide (which in real-life applications can be justified for small intervals $h$)}, the probabilities $\tilde P_t^{(1)}$ used in \eqref{5.3} might strongly over- or underestimate the correct conditional probabilities in the specific settings of Example 5.1 (i) - (iii). Hence, results of strategy (D) could be even improved by estimating $\tilde P_t^{(1)}$ accurately. In further simulations, we also varied other parameters, such as time horizon $T$, initial items $N$, number of competitors $K$, as well as the demand parameters $\beta$ or the adjustment dynamics of the competitors. Overall, the results were similar. We summarize the most \blue{important findings} in the following remark. \begin{remark} (Impact of adjustment frequencies and price anticipations) (i)\hspace{0.41cm} In general, adjustment frequencies as well as price anticipations significantly raise profits. (ii)\hspace{0.29cm} Higher adjustment frequencies, cf. heuristic (C), can overcompensate the value of price \hspace{0.95cm}anticipations, cf. strategy (B). (iii)\hspace{0.2cm} The performance of fix price strategies is low, even if they are optimally chosen and perfect price \hspace{0.95cm}anticipations are taken into account. (iv)\hspace{0.3cm} If market prices have a positive trend, cf. scenario (ii), then price anticipations are more important \hspace{0.95cm}than frequent price adjustments. (v)\hspace{0.38cm} If market prices have a negative trend, cf. scenario (iii), then frequent price adjustments are more \hspace{0.95cm}important than price anticipations. (vi)\hspace{0.26cm} If prices are adjusted frequently, then price anticipations are less important. (vii)\hspace{0.14cm} If market volatility is high, then it is important to adjust prices frequently. \label{rem5} \end{remark} \subsection{Heuristic vs. Optimal Response Strategies in Strategic Duopoly Competition} \blue{In this subsection, we consider strategic competitors, who choose his/her prices in response to our current price. The goal is to} compare our heuristic approach to optimal response strategies. Since optimal response strategies cannot be computed for complex competition settings (due to the curse of dimensionality), we consider a duopoly setting. \subsubsection{Optimal Response Strategies \blue{in a Duopoly}} We want to compare the performance of our heuristic strategy derived in Section 3 to upper bounds, which are determined by optimal response strategies that make use of full information. We assume that the price reaction of the competitor as well as the reaction times can be fully anticipated. Note, our heuristic strategy does not use this additional information, which corresponds to real-life scenarios. \blue{We assume that the market situation (state) is one-dimensional} and simply characterized by the competitor's price $p$, i.e., we let $\vec s:= p$. We assume that the competitor adjusts its price $p$ in response to \blue{our price $a$} with a fixed delay of $\Delta$ periods, $0 < \Delta < 1$. Choosing a price $a$ at time $t$ is followed by the competitor's price reaction $F$, which can depend on $p$ and $a$. I.e., after an interval of size $\Delta$ the competitor adjusts its price from $p$ to $F(a,p)$. \blue{Consequently, our reaction time is $1-\Delta$}. Following the examples of the previous sections, in period $t$ the probability to sell exactly $i$ items during the first interval of size $\Delta$ (Phase 1) is $P_t^{(\Delta )}(i,a,p):= Pois( {\Delta \cdot d \cdot \hat P(a,p)} )$. For the rest of the period (Phase 2) the sales probability changes to $P_{t + \Delta }^{(1 - \Delta )}\left( {i,a,F(a,p)} \right) = Pois( {(1 - \Delta ) \cdot d \cdot \hat P\left( {a,F(a,p)} \right)} )$. In case the competitors' strategy is known the Hamilton-Jacobi-Bellman (HJB) equation of \blue{the duopoly problem} can be written as, $t = 0,1,...,T - 1$, $n = 1,...,N$, $p \in A$, $0 < \Delta < 1$, \[{V_t}^{(\Delta)*}(n,p) = \mathop {\max }\limits_{a \in A} \left\{ {\sum\limits_{{i_1} \ge 0} {P_t^{(\Delta )}({i_1},a,p)} \cdot \sum\limits_{{i_2} \ge 0} {P_{t + \Delta }^{(1 - \Delta )}\left( {{i_2},a,F(a,p)} \right)} } \right.\] \begin{equation} \label{5.7} \left. { \cdot \left( {(a - c) \cdot \min (n,{i_1} + {i_2}) - n \cdot l + \delta \cdot V_{t + 1}^{(\Delta )*}\left( {{{(n - {i_1} - {i_2})}^ + },F(a,p)} \right)} \right)} \right\}, \end{equation} where $V_T^{(\Delta )*}(n,p) = 0$ for all $n,p$. The associated optimal pricing strategy $a_t^{(\Delta )*}(n,p)$, $t = 0,1,...,T - 1$, $n = 1,...,N$, $p \in A$, is determined by the arg max of \eqref{5.7}. \subsubsection{Heuristic Response Strategies \blue{in a Duopoly}} \blue{In case neither the competitor's strategy nor his/her reaction time $\Delta$ is known, our heuristic strategy, cf. Algorithm 3.1, can be applied using given conditional probabilities ${\tilde P_t^{(1)}(i,a|p)}$.} The corresponding strategy that adjusts prices at $t=0,1,...,T-1$ \blue{denoted by} $\tilde a_t^{(\Delta )}(n,p)$, $n = 1,...,N$, $p \in A$, is defined by the arg max of, $t = 0,1,...,T - 1$, $n = 1,...,N$, $p \in A$, $0 < \Delta < 1$, $\tilde V_T^{(\Delta )}(n,p):= 0$ for all $n,p$, \begin{equation} \label{5.8} \tilde V_t^{(\Delta )}(n,p) = \mathop {\max }\limits_{a \in A} \left\{ {\sum\limits_{i \ge 0}{\tilde P_t^{(1)}(i,a|p)} \cdot \left( {(a - c) \cdot \min (n,i) - n \cdot l + \delta \cdot \tilde V_{t + 1}^{(\Delta )}\left( {{{(n - i)}^ + },p} \right)} \right)} \right\}. \end{equation} The performance of the heuristic strategy $\tilde a_t^{(\Delta )}$, cf. \eqref{5.8}, can \blue{again} be evaluated analytically and yields the (suboptimal) expected profits, $t = 0,1,...,T - 1$, $n = 0,1,...,N$, $p \in A$, \blue{$0 < \Delta < 1$}, ${\bar V_T}^{(\Delta )}(n,p) = 0$, \[{\bar V_t}^{(\Delta )}(n,p) = \mathop {\max }\limits_{a \in \{ \tilde a_t^{(\Delta )}(n,p)\} } \left\{ {\sum\limits_{{i_1} \ge 0} {P_t^{(\Delta )}\left( {{i_1},a,p} \right)} \cdot \sum\limits_{{i_2} \ge 0} {P_{t + \Delta }^{(1 - \Delta )}\left( {{i_2},a,F\left( {a,p} \right)} \right)} } \right.\] \begin{equation} \label{5.9} \left. { \cdot \left( {(a - c) \cdot \min (n,{i_1} + {i_2}) - n \cdot l + \delta \cdot \bar V_{t + 1}^{(\Delta )}\left( {{{(n - {i_1} - {i_2})}^ + },F\left( {a,p} \right)} \right)} \right)} \right\}. \end{equation} Note, due to the reaction of the competitor, the probabilities $P_t^{(1)}(i,a,p)$ and $\tilde P_t^{(1)}(i,a|p)$ can be quite \textit{different}, e.g., if the competitor quickly undercuts our price and our price rank changes, \blue{cf. Figure 1a}. Hence, assuming stable prices and using $\tilde P_t^{(1)}(i,a|p):= P_t^{(1)}(i,a,p)$ might (strongly) overestimate the conditional sales probabilities. Hence, \blue{in such a case, observed sales will indicate that sales probabilities $\tilde P_t^{(1)}(i,a|p)$} significantly differ from $P_t^{(1)}(i,a,p)$. In general, the characteristics of a specific competitive setup will be reflected in sales data. Analyzing the relation between realized sales, offer prices, and the underlying market situations that have been observed at the time of the price adjustment, cf. Section 4.1, makes it possible to estimate the conditional sales probabilities $\tilde P_t^{(1)}(i,a|p)$, see also, e.g., Vulcano et al. (2012), Abdallah, Vulcano (2016), and Fisher et al. (2017). Note, neither the strategy nor the reaction time of the competitor has to be discovered. An accurate estimation of $\tilde P_t^{(1)}(i,a|p)$ may provide \begin{equation} \label{5.10} \tilde P_t^{(1)}(i,a|p) \approx \sum\limits_{{i_1},{i_2} \ge 0{:_{}}{i_1} + {i_2} = i} {P_t^{(\Delta )}\left( {{i_1},a,p} \right) \cdot P_{t + \Delta }^{(1 - \Delta )}\left( {{i_2},a,F(a,p)} \right)}. \end{equation} \subsubsection{Comparison of Strategies \blue{in a Duopoly}} Next, we want to measure the performance of the heuristic compared to the \blue{optimal (informed) strategy}. The following example studies one of the most common strategies \blue{which always slightly undercuts} the competitor's price, \blue{cf. Example 4.3 (iv), Fig. 4d}. We also study impact of the reaction time $\Delta$ as well as the accuracy of the estimation of $\tilde P_t^{(1)}$. \smallskip \begin{example} We assume a duopoly, i.e., $K=1$. Let $T=100$, $N=5$, $l=0.01$, $c=3$, $\delta = 0.9995$, $0<\Delta<1$, $d=10$. The demand is defined as in Example \ref{ex3}, i.e., $P_t^{(\Delta )}(i,a,p):= Pois( {\Delta \cdot d \cdot \hat P(a,p)} )$. We let our competitor play the response strategy $F(a,p):= \max (a - \varepsilon ,c)$, $\varepsilon=1$, $A:= \left\{ {1,2,...,120} \right\}$. The competitor adjusts prices in $t = \Delta ,1 + \Delta ,...,T - 1 + \Delta$. We adjust prices at $t = 0,1,...,T - 1$. We simulate three different \blue{price adjustment} strategies: \smallskip (i)\hspace{0.4cm} the optimal response strategy $^{(i)}a_{}^{(\Delta )*}$, cf. \eqref{5.7}, (ii)\hspace{0.3cm} the heuristic strategy ${^{(ii)}\tilde a^{(\Delta)}}$, cf. \eqref{5.8}, based on $\tilde P_t^{(1)}(i,a|p):= P_t^{(1)}(i,a,p)$, and (iii)\hspace{0.2cm} the heuristic strategy ${^{(iii)}\tilde a^{(\Delta)}}$, cf. \eqref{5.8}, based on \blue{(accurate)} $\tilde P_t^{(1)}(i,a|p)$ defined as in \eqref{5.10}. \label{ex5} \end{example} \smallskip Figure 5 shows optimal response strategies \blue{for Example 5.2 (i) - (iii) at time} $t=0$ for different inventory levels $n$ and competitor's prices $p$ \blue{in case of} $\Delta=0.5$. Figure 5a illustrates the optimal policy; Figure 5b and 5c depict the heuristic policy in case (ii) and (iii), respectively. \blue{The optimal response strategy (i) is of similar shape for different inventory levels:} If the competitor's price is very low, it is optimal to raise the price up to a certain price level. Almost the same price level has to be chosen if the competitor's price is very high. If the competitor's price is somewhere in between (intermediate range), it is best to undercut that price by one price unit $\varepsilon$. If $n$ increases, then the upper price level decreases; the intermediate range is of the same size but at a lower level. \begin{figure}[ht] \begin{center} \includegraphics[scale=0.45]{fig5a}\quad \includegraphics[scale=0.45]{fig5b}\quad \includegraphics[scale=0.45]{fig5c} \caption{Comparison of feedback prices \blue{for different inventory levels $n$}: Optimal response policy ${^{(i)}a_0^{(\Delta)*}}(n,p)$ (left window 5a) and heuristic policy in case (ii) $^{(ii)}\tilde a_0^{(\Delta )}(n,p)$ (middle window 5b), and case (iii) $^{(iii)}\tilde a_0^{(\Delta )}(n,p)$ (right window 5c) for $n=1, 2, 3, 5$, $\Delta=0.5$; Example \ref{ex5}.} \label{fig5} \end{center} \end{figure} \begin{figure}[ht] \begin{center} \includegraphics[scale=0.45]{fig6a}\quad \includegraphics[scale=0.45]{fig6b}\quad \includegraphics[scale=0.45]{fig6c} \caption{Comparison of simulated price paths over time: Optimal policy ${^{(i)}a_0^{(\Delta)*}}(n,p)$ (left window 6a) and heuristic policy in case (ii) $^{(ii)}\tilde a_0^{(\Delta )}(n,p)$ (middle window 6b), and case (iii) $^{(iii)}\tilde a_0^{(\Delta )}(n,p)$ (right window 6c) for the case $N=5$, $T=100$, $\Delta=0.5$; Example \ref{ex5}.} \label{fig6} \end{center} \end{figure} \blue{Overall, the heuristic policies have similar characteristics.} Compared to the optimal response policy and heuristic (iii), the prices of heuristic (ii) are lower and the intermediate range is wider. \blue{The corresponding figures of case (i) and (iii) for $\Delta=0.1$ and $\Delta=0.9$ are shown in the appendix, cf. Figure A.11 and A.12}. Note, the policy of case (ii) is independent of $\Delta$ \blue{as the reaction time is neither explicitly nor implicitly taken into account}. We observe that the intermediate range of the optimal policy is increasing in $\Delta$. Policy (i) and (iii) are quite similar for all $\Delta$ values. If $\Delta$ is large then policy (ii) is similar to the other two. If $\Delta$ is small then policy (ii) differs significantly. The application of \blue{the optimal and both heuristic} strategies leads to cyclic price patterns over time, cf. Edgeworth cycles, see, e.g., Maskin, Tirole (1988) or Noel (2007). The resulting price paths for the optimal response strategy and the heuristic strategies are shown in Figure 6. We observe that compared to heuristic (ii) the cycle length and the amplitude of the price patterns are smaller for the optimal policy (i) and the heuristic (iii). \blue{Finally, we obtain that heuristic (iii) is very similar to the optimal strategy (i).} The cyclic \blue{price patterns} are basically independent of the initial price. Hence, if the time horizon and the discounting factor are not too "small" then the expected future long-term profits ${V^{(\Delta)*}}(n,p)$ and ${\bar V_0^{(\Delta)}}(n,p)$, respectively, are (almost) the same for different initial prices $p$ (in our example the maximum deviation is 0.1). The number of items left to sell and particularly the reaction time $\Delta$ plays a more prominent role. \medskip \begin{table}[ht] \begin{center} \setlength{\tabcolsep}{1mm} \begin{tabular}{c ccc ccc} \toprule $n$ &$^{(i)}{V_0}^{(0.1)*}(n,50)$ &$\frac{{{}^{(ii)}{{\bar V}_0}^{(0.1)}(n,50)}}{{^{(i)}{V_0}^{(0.1)*}(n,50)}}$ &$\frac{{{}^{(iii)}{{\bar V}_0}^{(0.1)}(n,50)}}{{^{(i)}{V_0}^{(0.1)*}(n,50)}}$ &$^{(i)}{V_0}^{(0.9)*}(n,50)$ &$\frac{{{}^{(ii)}{{\bar V}_0}^{(0.9)}(n,50)}}{{^{(i)}{V_0}^{(0.9)*}(n,50)}}$ &$\frac{{{}^{(iii)}{{\bar V}_0}^{(0.9)}(n,50)}}{{^{(i)}{V_0}^{(0.9)*}(n,50)}}$\\ \midrule 1 &23.3637 &0.9801 &0.9949 &29.0480 &0.9881 &0.9852 \\ 2 &34.5616 &0.9766 &0.9942 &45.2496 &0.9867 &0.9841 \\ 3 &39.7475 &0.9716 &0.9925 &54.4413 &0.9801 &0.9803 \\ 5 &41.9375 &0.9584 &0.9910 &61.5614 &0.9731 &0.9761 \\ 7 &40.6005 &0.9473 &0.9890 &61.9205 &0.9690 &0.9774 \\ 10 &37.7302 &0.9413 &0.9879 &59.4264 &0.9675 &0.9795 \\ \bottomrule \end{tabular} \caption{Comparison of expected profits for \blue{different estimations of sales probabilities: Optimal policy ${^{(i)}a^{(\Delta)*}}$, cf. (i), vs. heuristic policies} ${^{(ii)}\tilde a^{(\Delta)}}$, cf. (ii), and ${^{(iii)}\tilde a^{(\Delta)}}$, cf. (iii), $\Delta=0.1,0.9$, $n=1,2,3,5,7,10$; Example \ref{ex5}. \label{tab2}} \end{center} \end{table} In Table 2, the expected profits $^{(i)}{V^{(\Delta )*}}$, $^{(ii)}{\bar V^{(\Delta )}}$, and $^{(iii)}{\bar V^{(\Delta )}}$ of Example \ref{ex5} are compared for $\Delta=0.1$ and $\Delta=0.9$. In our example, the expected profits $V^{(\Delta )*}$ and $\bar V^{(\Delta )}$ are maximized for an inventory level of 5-7 items. A higher number of items is not advantageous which is due to inventory holding costs. \blue{Finally, we observe} that the performance of our heuristic is quite promising. Although the heuristic does not anticipate the competitor's \blue{blue} reaction, in case (ii), 94-98\% of the theoretically optimal expected profits are obtained; in case (iii), we obtain even 98-99\%. \blue{If the competitor's reaction time $\Delta$ is small, the results of heuristic (iii) outperform} those of heuristic (ii) by up to 5\%. The reason is the following: If $\Delta$ is large then the market situation hardly changes during one period of time and, \blue{thus}, $P_{}^{(1)}$ and $\tilde P_{}^{(1)}$ are almost identical. If $\Delta$ is \blue{small} then the market situation quickly changes, i.e., $P_{}^{(1)}$ and $\tilde P_{}^{(1)}$ differ significantly, and in turn, the performance of the heuristic suffers. \subsubsection{Comparison of Reaction Times \blue{in a Duopoly}} Due to the significant impact of reaction times, firms will try to minimize their reaction times by anticipating their competitors' time of adjustment and to time their price adjustments optimally. In order not to act predictably, in turn, strategic firms might randomize their reaction times. \blue{Finally, the} percentage of time a firm has the most recent price will be determined by the ratio of \blue{the firms'} adjustment frequencies. \blue{Refreshing the prices $x$ times as often as the competitor can be compared to our model with $\Delta:= x/(1 + x)$, or $x = \Delta/(1 - \Delta)$, $x > 0$.} For instance, the ratio $x=2.33$ corresponds to a value of $\Delta=0.7$. Note, while some firms adjust their prices once a day, other firms adjust their prices every one or two hours. In the following, we study to which extent expected profits of our strategies are affected by $\Delta$. Following Example \ref{ex5}, we compare the expected profits ${V^{(\Delta)*}}$ of the optimal policy and ${\bar V^{(\Delta)}}$ \blue{of the heuristic policy with the benchmark profit} ${V^{(0.5)*}}$ (for the case $\Delta=0.5$ or $x=1$, respectively). \begin{table}[ht] \begin{center} \setlength{\tabcolsep}{2mm} \begin{tabular}{cccccccc} \toprule Strategy & \blue{Competitor's reaction time $\Delta:$} &0.1 &0.3 &0.5 &0.55 &0.7 &0.9 \\ \hline (i) &${V_0}^{(\Delta )*}(1,50)/{V_0}^{(0.5)*}(1,50)$ &0.8873 &0.9444 &1.0000 &1.0135 &1.0529 &1.1032 \\ (i) &${V_0}^{(\Delta )*}(5,50)/{V_0}^{(0.5)*}(5,50)$ &0.8101 &0.9041 &1.0000 &1.0239 &1.0954 &1.1892 \\ (i) &${V_0}^{(\Delta )*}(10,50)/{V_0}^{(0.5)*}(10,50)$ &0.7799 &0.8878 &1.0000 &1.0284 &1.1138 &1.2284 \\ \midrule (ii) &${\bar V_0}^{(\Delta )}(1,50)/{V_0}^{(0.5)*}(1,50)$ &0.8697 &0.9333 &0.9908 &1.0043 &1.0429 &1.0900 \\ (ii) &${\bar V_0}^{(\Delta )}(5,50)/{V_0}^{(0.5)*}(5,50)$ &0.7765 &0.8762 &0.9730 &0.9968 &1.0669 &1.1573 \\ (ii) &${\bar V_0}^{(\Delta )}(10,50)/{V_0}^{(0.5)*}(10,50)$ &0.7341 &0.8478 &0.9614 &0.9898 &1.0750 &1.1884 \\ \midrule (iii) &${\bar V_0}^{(\Delta )}(1,50)/{V_0}^{(0.5)*}(1,50)$ &0.8828 &0.9331 &0.9882 &1.0005 &1.0370 &1.0868 \\ (iii) &${\bar V_0}^{(\Delta )}(5,50)/{V_0}^{(0.5)*}(5,50)$ &0.8028 &0.8858 &0.9710 &0.9988 &1.0650 &1.1601 \\ (iii) &${\bar V_0}^{(\Delta )}(10,50)/{V_0}^{(0.5)*}(10,50)$ &0.7705 &0.8697 &0.9722 &1.0024 &1.0838 &1.2032 \\ \bottomrule \end{tabular} \caption{Additional information vs. reaction time: Optimal policy ${a^{(\Delta)*}}$, cf. (i), and heuristic policies ${\tilde a^{(\Delta )}}$, cf. (ii) and (iii), for different \blue{competitor's} reaction times $0.1 \le \Delta \le 0.9$ compared to the \blue{expected benchmark profits ${V_0^{(0.5)*}}$ of policy ${a^{(0.5)*}}$ in case $\Delta =0.5$}, $n=1,5,10$; Example \ref{ex5}. \label{tab3}} \end{center} \end{table} Table 3 illustrates the ratio ${V^{(\Delta)*}}/{V^{(0.5)*}}$ as well as the ratio ${\bar V^{(\Delta)}}/{V^{(0.5)*}}$ for different $n$ and $\Delta$ in case (i) - (iii) of Example \ref{ex5}. We observe that the reaction time $\Delta$ has a significant impact on expected profits. \blue{On the one hand,} our example shows that small reaction frequencies can shrink expected profits by more than 20\%, cf. $\Delta$=0.1. \blue{On the other hand,} heuristics using shorter reaction times or more frequent price adjustments, respectively, can easily overcompensate a lack of information, cf. $\Delta$=0.55, and increase expected profits by more than 20\%, cf. $\Delta$=0.9. We conclude that the more volatile the market is, the more important is it to use an accurate estimation of conditional probabilities $\tilde P$, which are also characterized by the own adjustment frequency. Note, it is not necessary to know the competitors' reaction \blue{strategy -- the information} of the competitors' behavior and the market dynamics are implicitly contained in $\tilde P$. \smallskip \begin{remark} (Main insights) (i)\hspace{0.4cm} For successful strategies regular price adjustments as well as anticipation of market dynamics are important. Both effects are challenging: on the one hand, computations of dynamic systems are time-consuming; on the other hand, real-life market dynamics are usually complex and characterized by a huge number of factors. (ii)\hspace{0.3cm} To circumvent the curse of dimensionality, we \blue{simplify the problem: While the reduced state space} of "time and inventory level" allows for fast optimization, in our model the anticipation of market evolutions is transferred to the model's sales probabilities. The inaccuracy of our simplifying "stability assumption" -- which is the key for the decomposition -- can be compensated as frequent \blue{observations of the current market situation and corresponding price adjustments keep the system} up to date. (iii)\hspace{0.2cm} In real-life applications, the conditional sales probabilities of the model can be estimated from market data using state-of-the-art regression models (logistic regression, decision trees, neural networks, etc.). \label{rem6} \end{remark} \medskip Our examples verify that the performance of our heuristic compared to optimal policies (that use full information) is excellent even if strategic competitors are involved. While optimal solutions are limited to very few competitors, the heuristic strategy instead can easily be applied in the presence of even very large numbers of competitors. Since the state space of the heuristic approach just consists of two dimensions, the computation of prices is constantly fast. Hence, it is ensured that it is possible to compute prices for complex situations and to adjust prices with high frequencies. \subsection{Heuristic vs. Heuristic Response Strategies in Strategic Oligopoly Competition} In this subsection, we study the case in which \blue{our heuristic strategy} is played by multiple firms against each other. In the next example, we consider such a scenario. \smallskip \begin{example} Consider the setting of Example \ref{ex4} (i). Let $K=3$, $T=100$, $N=10$, $l=0.01$, $c=3$, $\delta = 0.9995$, $d=40$, and $A:= \left\{ {1,2,...,100} \right\}$. All competitors play the heuristic strategy, cf. Algorithm 3.1. All firms use randomized reaction times: In every point in time $t = 0,h,...,T - h$ prices are adjusted by each player with probability $\pi = h$, $h=0.1$. Initial prices are given by $p_0^{(k)}:= 20$, $k = 1,...,K$. For evaluation we use the sales probabilities $P_t^{(h)}(i,a,\vec p):= Pois( {h \cdot d \cdot \hat P(a,\vec p)} )$. For computation of the heuristics, all firms use $\tilde P_t^{(1)}(i,a|\vec p): = P_t^{(1)}(i,a,\vec p)/K$ as a simple estimation for the conditional probabilities. \label{ex6} \end{example} \medskip \begin{figure}[ht] \begin{center} \includegraphics[scale=0.63]{fig7a}\qquad \includegraphics[scale=0.63]{fig7b} \caption{Heuristic strategies played against each other: Simulated price trajectories (left window 7a) and inventory levels over time (right window 7b), $K=3$, $N=10$, $T=100$; Example \ref{ex6}.} \label{fig7} \end{center} \end{figure} Figure 7a illustrates a scenario in which the heuristic strategy, cf. Algorithm 3.1, is played by multiple firms. We observe that firms undercut each other until a certain price level is obtained. Then firms - one after another - raise their prices up to a certain maximum. This evolution is similar to \blue{the price patterns of Figure 6, cf. Example 5.2}. We observe, that the length of price cycles is longer if less firms are involved. \blue{The evolution of the inventory levels of all firms is shown in Figure 7b. Firms exits the market due to stock-outs ($t=88$ and $t=96$, cf. Figure 7b)}. Overall, \blue{we observe that} the heuristic feedback strategy is "reasonably aggressive" but not "destructive" and effectively avoids a decline in price (race to the bottom). Note, this character has similarities to equilibrium strategies. In case reaction times and inventory levels are observable and firms are rational, best response strategies could be derived using backward induction. This assumption, however, appears not to be a realistic one. Pricing strategies of firms will depend on time and their inventory level, which, in general, is not observable for competitors. Hence, we have competition with asymmetric information. In this setting, equilibrium pricing strategies will be hard to determine. Mutual best response strategies might use observations of offer prices to estimate probability distributions for the competitors' remaining inventory (Hidden Markov Model), \blue{cf. Schlosser, Richly (2018). However, the applicability of such models is again limited} due to the curse of dimensionality. Finally, our heuristic is a viable and successful strategy to be applied in competitive real-life markets. Moreover, our heuristic is also a suitable strategy to be played as a response against itself. \subsection{Price Adjustment Costs} \blue{In the previous sections, we have assumed that price changes are free. Hence, frequent updates can be used to compensate for the lack of market anticipation. However, in some markets the number of price updates are limited or may create even (hidden) costs as frequent price adjustments may confuse customers and reduce their trust. Hence, it can be reasonable to put penalties on price changes in order to implement only necessary price adjustments. } \blue{In this context, we present a model with a fixed penalty $z$ for each price adjustment. To be able to include price adjustment costs in our dynamic programming framework, cf. (4) - (7), we extend the state space by one dimension -- the offer price ${a^ - }$ of the previous period, ${a^ - } \in A$. The costs $z$, $z > 0$, of a price change occur if the price update $a$ differs from the previous price ${a^ - }$. To model the case in which no previous price exists, we let ${a^ - }:= 0$.} \blue{For any market situation $\vec s$ the natural boundary condition for the value function is given by, cf. (4) - (5), $n = 0,...,N$, $t = 0,...,T$, ${a^ - } \in A \cup \{ 0\}$,} \blue{ \begin{equation} \label{19} V_T^{}(n,\vec s,{a^ - }) = V_t^{}(0,\vec s,{a^ - }) = 0. \end{equation}} \blue{The remaining values ${V_t}(n,\vec s,{a^ - })$ are defined by the HJB equation, $n = 1,...,N$, $t = 0,...,T-1$, ${a^ - } \in A \cup \{ 0\}$,} \blue{ \[{V_t}(n,\vec s,{a^ - })\] \begin{equation} \label{20} = \mathop {\max }\limits_{a \in A} \left\{ {\sum\limits_{i \ge 0} {{{\tilde P}_t}(i,a|\vec s)} \cdot \left( {(a - c) \cdot \min (n,i) - n \cdot l - z \cdot {\mathds{1}_{\{ a \ne {a^ - } \wedge {a^ - } > 0\} }} + \delta \cdot {V_{t + 1}}\left( {{{(n - i)}^ + },\vec s,a} \right)} \right)} \right\}, \end{equation}} \noindent \blue{where the indicator function ${\mathds{1}_{\{ a \ne {a^ - } \wedge {a^ - } > 0\} }}$ is 1 if $a \ne {a^ - }$ and ${a^ - } > 0$, it is 0 if $a = {a^ - }$ or ${a^ - } = 0$. The transition of the new state component is straightforward -- the new price $a$ becomes the old price $a^-$ of the next period. Again, the associated pricing strategy is determined by the arg max of (20), $n = 1,...,N$, $t = 0,...,T-1$, ${a^ - } \in A \cup \{ 0\}$.} \blue{Algorithm 3.1 can still be used in the same way. The numerical complexity is slightly higher but tolerable. Due to adjustment costs, prices are only updated if opportunity costs (of not updating the price) are sufficiently large. The larger $z$ is chosen, the less price updates will be made. Note, compared to our previous models, in general, larger delays between two price adjustments (i.e., lower adjustment frequencies) also lead to less price updates, but they sacrifice the advantage of fast reactions to changing markets. Hence, the model above allows for fast reactions \textit{and} a reasonable limitation of the number of price updates.} \section{Evaluation in Practice} To evaluate \blue{the real-life performance of the presented approach}, we applied our heuristic strategy on Amazon Marketplace. Online market platforms such as Amazon or eBay are highly dynamic as sellers can observe the current market situation at any point in time and adjust their prices instantly. This dynamic is hard to manage as pricing decision requires handling a multitude of dimensions for each competitor \blue{(price, quality, seller ratings, etc.).} In this experiment, we \blue{have partnered} with a German bookseller. The seller is among the top 10 sellers for used books on Amazon in Germany and has an inventory of over 100\,000 distinct books (ISBN), each with multiple items (1-20). Our seller can decide -- to some extent -- on the replenishment of used books (by adjusting purchase prices). However, supply is limited and it is not possible to directly reorder specific \blue{(used)} books. Hence, the challenge is to extract as much profit as possible from a given number of books (cf. inventory level) in a reasonable amount of time. The pricing strategy of our project partner is characterized by a rule-based system that has been developed over the past years by carefully adjusting rules to lessons learned from selling books on Amazon. As our project partner has more than ten years of experience in the market, we consider his strategy to be effective and accurate. However, market dynamics are increasingly sophisticated making rule-based strategies increasingly hard to handle and maintain. \subsection{Implementation and Architecture} As our approach is designed to be applied in \blue{practice, we needed to calibrate} the model, \blue{i.e., holding costs, shipping costs, discount rate, and particularly the sales probabilities, which are based on multiple offer dimensions. Requested market situations for a certain product from Amazon Marketplace include product-specific features (e.g., Amazon sales rank) as well as merchant-specific features (price, quality, seller ratings, feedback count, etc.). The most important offer dimensions are summarized in Table 4.} \smallskip \begin{table}[ht] \centering \begin{tabular}{l l l} \toprule &Offer Dimension & Range/Unit \\ \midrule \multirow{3}{*}{\rotatebox[origin=c]{90}{\parbox[c]{2.15cm}{\centering Product-Specific\\Parameters}}} &time & seconds\\ &Amazon sales rank of ISBN & 1 -- 5\,000\,000 by 1\\ &weight & gram\\ &original price & 0.01 -- 500 Euro by 0.01\\ &number of used offers & 0 -- 20 by 1\\ \cmidrule{1-2} \multirow{3}{*}{\rotatebox[origin=c]{90}{\parbox[c]{3.0cm}{\centering Competitor-Specific Parameters}}} &price & 0.01 -- 500 Euro by 0.01\\ &condition/quality & new -- acceptable (6 levels)\\ &rating & 0\% -- 100\%\\ &feedback count & 0 -- 5\,000\,000 by 1\\ &shipping time & 0 -- 30 days by 1\\ &shipping costs & 0 -- 10 Euro by 0.01\\ &domestic shipping & yes / no \\ \bottomrule \end{tabular} \caption{\blue{Product- and} merchant-specific offer dimensions.} \label{tab4} \end{table} \blue{Figure 8 illustrates real-life price patterns on the Amazon Marketplace where heterogeneous firms compete for one market. We observe the following typical characteristics: (i) active and passive competitors, (ii) volatile prices with wide price ranges, (iii) entries (cf. Firm 6) and exits of firms (cf. Firm 4), and (iv) Edgeworth-like price cycles. Those characteristics have similarities with our examples and verify the suitability of our models, cf. Section 4 and 5.} \begin{figure}[ht] \begin{center} \includegraphics[width=0.95\textwidth]{fig8} \caption{\blue{Example of real-life price trajectories from Amazon Marketplace over time for the sale of used books in an oligopoly setting (based on observations in discrete points in time); the legend contains (i) product quality Q, (ii) merchant rating R, and (iii) feedback count FC of all competing firms}.} \label{fig_prices} \end{center} \end{figure} \blue{ \Cref{fig_process} depicts the seller's communication with the Amazon Marketplace and the data flow. The decision when to update which product is determined by product queues, which separate the seller's products by an adjustable priority (Step 1), cf. Inventory Update Queue. To update the price of a given product of the inventory, the seller requests the current market situation via the Marketplace Web Service API (Step~2).} \blue{ The main component integrating our heuristic strategy is the Data-Driven Pricing Component (Step 3). Based on predicted sales probabilities for a product's market situation (Step 4), cf. Demand Learning, the Price Optimization determines a price and the price update is sent back to Amazon (Step 5).} \blue{ As the rate with which sellers can request market situations and update prices is limited, products of higher importance are stored in the most frequently updated queue. On the other hand, products which do not face strong competition or are expected to be accessed less often by customers, are stored in the least frequently updated queue. The queuing mechanism updates products every 2-12 hours (i.e., 2-10 updates per book per day), which results in $>$20 M observed market situations per month. Considering on average seven competitors per market situation corresponds to $>$140 M single competitor observations per month. The majority of requests also lead to a price update.} \bigskip \begin{figure}[ht] \begin{center} \includegraphics[width=\textwidth]{fig9 \caption{\blue{Illustration of the data-driven price update process of a merchant on Amazon Marketplace}.} \label{fig_process} \end{center} \end{figure} \blue{ Another used service is the Push Notification Service, cf. \Cref{fig_process}, which informs sellers of products sold on the Marketplace platform. Sellers can store both data streams (requested market situations and pushed sales) locally in a database. We join the market data with our partner's price updates, placed orders, and stock data to create the required observations and the corresponding features (cf. Definition 4.1). Working directly on the raw time-series data provides us with more flexibility, e.g., when regressing only a subset of comparable market situations. The components regularly train a regression model that allows to predict sales probabilities for any market situation.} Our estimations of sales probabilities for a specific book in a certain time interval are based on market situations, which \blue{ are characterized by product-specific features as well as multiple competitor-specific offer dimensions for each present competitor, cf. Table 4.} \blue{We defined 30 customized features (explanatory variables, cf.~Section 4.1) to describe the relative competitiveness (i.e., sales probabilities) of a single offer in a particular market situation. We used features similar to those described in Definition \ref{def1} such as, e.g., the price rank of our offer price within the present competitors' prices, etc. Besides price also other offer dimensions were taken into account.} \blue{ To reduce estimation biases, in an exploration phase, also a certain share of randomized prices updates were used by the seller.} \subsection{Performance Results} We tested different demand learning techniques to quantify how \blue{sales probabilities of specific products are affected by our offer prices in specific market situations. We decided for a weighted logistic regression approach to estimate probabilities to sell a certain book within a certain time frame at potential offer prices given a particular market situation}. We calibrated our dynamic programming \blue{model, cf. Section 3,} based on the estimated (conditional) sales probabilities $(\tilde P)$. Using our heuristic approach, \blue{we computed price adjustments for exemplary product offers in specific market situations and checked their plausibility compared to the merchant's strategy and experience.} The application of our dynamic pricing strategy works as described in the previous sections, \blue{cf. Algoritm 3.1. Note, as discussed earlier,} the extension of the dimensionality of the market situation does not affect the algorithm's complexity. Finally, we used the calibrated model to determine heuristic pricing strategies to be applied on the Amazon Marketplace. In our experiment, we used four different discount factors $(\delta^{(1)}>\delta^{(2)}>\delta^{(3)}>\delta^{(4)})$ to vary the strategies' aggressiveness, cf. strategy $S^{(1)}$ -- $S^{(4)}$ in Table 5. $S^{(4)}$ is the most aggressive strategy; \blue{$S^{(1)}$ is the most tame strategy, cf. Remark 4.2 (iii)}. Over two months, we compared our data-driven strategies with the seller's rule-based benchmark strategy. To each strategy, we randomly assigned a test group of around 3\,300 books. \blue{To guarantee a fair comparison of the \textit{quality} of a strategies' pricing decisions, the price adjustment frequencies were the same for the merchant's benchmark strategy and our four data-driven strategies.} \bigskip \begin{table}[ht] \begin{center} \setlength{\tabcolsep}{2mm} \begin{tabular}{c c c c c} \toprule Strategy & Test group size & \% Sold & Revenue per sale (EUR) & Profit per sale (EUR) \\ \midrule \\[-5pt] Benchmark &3\,399 &36.33\quad(100\%) &6.71 \quad(\hphantom{,}100\%) &1.90 \quad(\hphantom{,}100\%) \\[2pt] S$^{(1)}$ &3\,210 &23.24\quad(--36\%) &8.19 \quad(+22\%) &3.07 \quad(+61\%) \\ S$^{(2)}$ &3\,339 &31.57\quad(--13\%) &7.56 \quad(+13\%) &2.57 \quad(+35\%) \\ S$^{(3)}$ &3\,185 &35.35\quad(--\hphantom{2}3\%) &7.37 \quad(+10\%) &2.42 \quad(+27\%) \\ S$^{(4)}$ &3\,155 &37.27\quad(+\hphantom{,}3\%) &7.00 \quad(+\hphantom{2}4\%) &2.13 \quad(+12\%) \\ \bottomrule \end{tabular} \caption{\blue{Performance comparison} of our data-driven strategies and a \blue{merchant's} rule-based benchmark strategy. \label{tab5}} \end{center} \end{table} Table 5 summarizes a comparison of sales, revenues per sale, and profits per sale of the different strategies. Profits are defined as revenue minus costs, i.e., shipping, Amazon provision, tax (7\%), packing, additional costs (warehouse rent, electricity costs, staff costs), and the average purchase price per item. As expected, the speed of sales increases and the profitability decreases if our heuristic strategies \blue{are} more aggressive. Hence, the aggressiveness of \blue{data-driven strategies} can be used to actively control the tradeoff between profitability and speed of sales. Moreover, strategy $S^{(4)}$ reveals that our approach can sell faster (+3\%) and at the same time more profitable (+12\%) as the seller's benchmark strategy. \begin{table}[ht] \centering \begin{tabular}{cc} \toprule Strategy & Relative accumulated profit \\ \midrule \\[-5pt] Benchmark & \hphantom{2}100\% \\[2pt] S$^{(1)}$ & +\hphantom{ 1}3\% \\ S$^{(2)}$ & +\hphantom{ }17\% \\ S$^{(3)}$ &+\hphantom{ }24\% \\ S$^{(4)}$ & +\hphantom{ }15\% \\ \bottomrule \end{tabular} \caption{Comparison of accumulated profits. \label{tab6}} \end{table} In Table 6, we compare the accumulated profits of all strategies. The relative accumulated profit denotes the quantity "profit per sale (EUR) $\times$ \% of items sold" compared to the corresponding value of the benchmark strategy. Results show that with our strategy applied, profits can be increased by more than 20\%, cf. S$^{(3)}$. Moreover, the model's discount factor can be used as a management instrument to smoothly balance profits, revenues, and the speed of sales according to a seller's needs. \blue{The individual price patterns of the merchant's benchmark strategy and our data-driven strategy highly depend on (i) the specific article involved, (ii) the number of competitors, (iii) the competitors' offers, (iv) the kind of competitors (commercial, private, strategic) involved, and (v) the price reaction frequencies of competitors. In general, it is advisable not to exclude higher price ranks in advance. Instead, multiple offer dimensions of competitors' offers have to be taken into account. Price adjustment frequencies and reaction times play a prominent role, particularly, when the customer arrival intensity is high. Further, a successful strategy has to balance two effects: a decline in price by undercutting competitors and moderate price decisions that retain/restore a profitable price level.} \section{Conclusion} Dynamic pricing under strong competition and incomplete information is a major open problem in revenue management. Practical relevance is enormous, but the problem appears intrinsically hard due to the curse of dimensionality. The challenge is to derive viable heuristic approaches that (i) can be applied in various settings, (ii) allow for data-driven demand estimations, (iii) achieve near-optimal performance results, and (iv) have minimal computation times. \blue{We have shown how to compute} dynamic pricing strategies under competition. We have demonstrated that our heuristic approach is applicable even if the number of competitors is large and the competitors' strategies are unknown. Due to the intractibility of the problem, in general, neither optimal policies nor performance guarantees can be derived. To measure the performance of our strategy, we computed upper bounds for expected profits, which are obtained by optimal strategies that take advantage of price anticipations. We observed that reaction times have a major effect on the performance of dynamic pricing strategies. Moreover, we found that higher adjustment frequencies can easily overcompensate a loss in expected profits due to the lack of price anticipations. Adjusting the price more frequently can be more advantageous than anticipating future market dynamics. Our heuristic approach allows for frequent price adjustments as the computation of prices is efficient and fast. \blue{Furthermore, in order to choose the number of price adjustments with care without sacrifizing a short reaction time, we present an extended version of the model which allows to take price adjustment costs into account.} Our approach is based on conditional sales probabilities that are estimated from observable market data. Market situations are allowed to be \blue{high-dimensional and might also include multiple product features}, such as quality, ratings, shipping time, etc. \blue{Estimated (conditional) sales probabilities make it possible to capture the essence of market dynamics as the impact of (even unobservable) market evolutions} on sales probabilities is implicitly included. Thus, it is not necessary to know competitors' strategies explicitly or their reaction times, but we can still take their characteristic behaviour into account. Note, on average, this also includes the impact of exits or entries of new competitors. \blue{Finally, changes of market situations are captured within the framework of regular market observations and price adjustments.} Using numerical examples that are based on real-life data, we verified the performance of our heuristic pricing strategy. In a real-life experiment, we outperformed the rule-based strategy of an experienced seller by more than 20\%. Finally, our approach combines key features that are important for real-life applications. First, the approach is applicable if many competitors are involved and offers have \blue{multiple features}. Second, market dynamics do not have to be explicitly known, but they can be indirectly taken into account using data-driven demand estimations. Third, \blue{the computation} of prices is efficient, easy to implement, and allows for frequent price adjustments.

762440049cc20fe567fb7561b8704b520e25a8d3

\section{Introduction} The late time acceleration in the standard Einstein gravity is propelled by a mysterious energy component which consist of a huge negative pressure that expands the Universe. This is called dark energy \cite{SN,SN2,SN3,per}. Keeping in mind the standard cosmological model, the DE undertake the simplest form of cosmological constant $\Lambda$, which has absolutely no spatial fluctuations but a negative pressure and constant energy density which covers the entire expansion history of the Universe. That leads to $\Lambda$CDM model. Cosmic Microwave Background (CMB) \cite{ade} , Supernova Type-Ia (SnIa) (SnIa) \cite{jla}, Baryon Acoustic oscillation (BAO) measurement in galaxy surveys \cite{bao} : can be demonstrated via $\Lambda$CDM model . However, it gets into serious conceptual problems like " fine tuning problem " \cite{win} and "coincidence problem" \cite{hu,de}. Recent observational results also indicate inconsistency with $\Lambda$CDM model \cite{discrep,riess,kids,valen}. In order to solve these problems, a number of scalar field models including quintessence \cite{Ratra:1987rm,Peebles:1987ek,Copeland:2006wr,Sahni:1999gb, Frieman:2008sn,Padmanabhan:2002ji,Padmanabhan:2006ag, Sahni:2006pa,Peebles:2002gy,Perivolaropoulos:2006ce, Sami:2009dk,Sami:2009jx,Sami:2013ssa}, phantom field \cite{Caldwell:1999ew,Caldwell:2003vq,Carroll:2003st,Singh:2003vx,Hao}, rolling tachyon \cite{Sen:2002nu,Sen:2002in,Sen:2003mv,Mazumdar:2001mm} and others have been proposed. Another way to look at this problem is through interaction between dark energy and other matter species in the Universe as suggested by \cite{Wetterich,Amendola,Farrar,Gubser,Farrar2}. In relation to this, minimally coupled dynamic scalar field can be extended to interacting Quientessence models (from now IQ). In these categories of models, DE is coupled to matter (dark matter and baryons), but coupling of DE with baryonic particles will result in time variation of constants of nature and hence, is tightly bound by observations \cite{Hagiwara}. However, these constraints does not apply to interaction between dark sector. This interaction between DM and DE is still permitted by observations \cite{Damour,Casas}. In this work, we will focus on coupling between dark sector only. Various types of interaction between DE and CDM have been proposed and investigated in the literature, like, Linear coupling of scalar field with matter (DM and baryon) suggested by Amendola \cite{Amendola}, Coupling with dark matter only \cite{Damour,Casas}, Nonlinear couplings \cite{Nonlinear1,Nonlinear2}, Variational approach \cite{Boehmer1,Boehmer2}. In relation to aforementioned models, the coupling describes the exchange of energy-momentum between dark energy and dark matter. Interaction between dark sector could change the expansion history of the Universe . Interaction can change source term of Poisson equation through scalar field perturbation. Additionally it can produce a fifth force between matter particles which results into stronger clustering of matter \cite{NL4,NL5}. All these factors could affect structure formation on both small and large scales. In recent years, the impact of interaction on linear growth of structure \cite{L1,L2,L3,L4,L5,L6,L7} , non-linear structure formation \cite{NL1,NL2,NL3,NL4,NL5,NL6} has been studied in detail. Observational data like CMB , BAO, LSS, Weak lensing \cite{C1,C2,C3,C4,C5}, local gravity tests \cite{LG1,LG2} and high redshifted intergalactic medium \cite{IGM1} have constrained the interaction, but none of them have ruled out the interaction. In order to advance the work by Bikash et al.\cite{Bikash} we shall include coupling between Quintessence and CDM and investigate its effect on linear structure formation.We consider a thawing class of model. We include red shift space distortion term and GR effects in power spectrum. We vary the interaction strength to show the percentage deviation wrt $\Lambda$CDM model at different redshifts. We normalize IQ model to give it parameters ($\Omega_{m0}$ , $H_{0}$) that are similar to that of ordinary non-interacting Quintessence at present redshift. The outline of the paper is : Sec II deals with background IDE models while Sec section III discusses perturbed DE models . Linear structure formation is discussed in section IV. Result and conclusion are discussed in section V. \\ \section{Background evolution with scalar field coupled with DM} We have considered a model which allow coupling between dark sector. Scalar field is coupled to dark matter but not to baryons. Here, we follow prescription discussed many times in literature \citep{Amendola,Amendola2004,Amendola2008,L2,NL2,Piloyan,Sumit}. We also follow the same formalism here. The important equations are as follows: \begin{eqnarray} \ddot{\phi}+\frac{dV}{d\phi}+3H\dot{\phi}=C(\phi)\rho_{d} \nonumber \\ \dot{\rho}_{d}+3H(\rho_{d})=-C(\phi)\rho_{d}\dot{\phi} \\ \dot{\rho_b}+3H(\rho_b)=0 \nonumber\\ H^2=\frac{\kappa^2}{3}(\rho_b+\rho_{d}+\rho_\phi) \nonumber \\ \end{eqnarray} \begin{equation} 1 = \frac{\kappa^2\rho_{b}}{3H^2}+\frac{\kappa^2\rho_{d}}{3H^2}+\frac{\kappa^2\dot{\phi}^2}{6H^2}+ \frac{\kappa^2V(\phi)}{3H^2} \end{equation} Herein, $C(\phi)$ comprises an interaction between dark sector. Due to lack of detail of the nature of interaction we consider it to be constant \citep{Amendola,Amendola2004} . We can even study uncoupled case by putting $C=0$. Now, we introduce these dimensionless parameters: \begin{align} x = \frac{\kappa\dot{\phi}}{\sqrt{6}H}, \hspace{1mm} y = \frac{\kappa\sqrt{V(\phi)}}{\sqrt{3}H} \nonumber \\ s =\frac{\kappa\sqrt{\rho_b}}{\sqrt{3}H}, \hspace{1mm} \lambda = \frac{-1}{\kappa{V}}\frac{dV}{d\phi} \hspace{1mm} \Gamma = \frac{V\frac{d^2V}{d\phi^2}}{\left(\frac{dV}{d\phi}\right)^2} \end{align} \noindent Here, the variable $\Gamma$ shows potential. In terms of $x$ and $y$, $\Omega_{\phi}$ and equation of state $w_{\phi}$ are : \begin{align} \Omega_\phi =& x^2+y^2 \\ \gamma =& 1+w_\phi = \frac{2x^2}{x^2+y^2} \end{align} \noindent Using aforementioned dimensionless variables, eq (1) and (2) can be converted in following autonomous systems: \begin{align} \Omega_\phi' =& W\sqrt{3\gamma\Omega_\phi}(1-\Omega_\phi-s^2)+3\Omega_\phi(1-\Omega_\phi)(1-\gamma) \nonumber \\ \gamma' =& W\sqrt{\frac{3\gamma}{\Omega_\phi}}(1-\Omega_\phi-s^2)(2-\gamma)+\lambda\sqrt{3\gamma\Omega_\phi}(2-\gamma)\nonumber\\ - &3\gamma(2-\gamma) \nonumber \\ s' =& -\frac{3}{2}s\Omega_\phi(1-\gamma) \nonumber \\ \lambda' =& \sqrt{3\gamma\Omega_\phi}\lambda^2(1-\Gamma), \end{align} here, $W=\frac{C}{\kappa}$. \subsection{Initial conditions to solve Background equations} To solve this autonomous system (7) we would need initial conditions for ($\gamma$, $\Omega_{\phi}$,s,$\lambda$). We settled our initial conditions at ($z = 1000$). At ($z = 1000$) for thawing class of models, scalar field is frozen thus $\gamma_{i} \approx 0$. We have taken initial value of $\lambda_{i}$ as a model parameter. We settles the initial condition for $\Omega_{\phi }$ by fine tuning it so as to obtain appropriate value of $\Omega_{\phi }$ today at $z=0$ . In the same way, we set initial value of $s$ to get appropriate value of the $\Omega_{b}$ today. Subsequently, we set initial condition for IQ model to get same $\Omega_{m0}$ and $H_{0}$ as in non-interacting case. \subsection{Behaviour of Background cosmological parameter} Set of equations (7) will be solved using the earlier mentioned initial conditions and the cosmological parameters for different values of interacting parameter W will be studied. We focus on linear potential. Figure 1 displays the variation of equation of state as function of redshift for different values of W. Since we are considering thawing model, $w_{\phi}$ starts from -1 at $z= 1000$ for non interacting case. On adding interaction, $w_{\phi}$ increases which shows strong dark energy effect. Figure 2 displays the percentage change in matter density parameter wrt LCDM. Here percentage change is negative which implies suppression wrt LCDM. The background dark matter density ($ \Omega_{m} $) decreases wrt LCDM. Figure 3 displays the percentage change in normalized hubble parameter wrt LCDM. Here, percentage change is positive which implies enhancement wrt LCDM. On increasing Interacting parameter (W), percentage change increases which entails stronger dark energy effect. \begin{center} \begin{figure*} \begin{tabular}{c@{\quad}c} \epsfig{file=compare_eos.eps,width=9.0 cm} \end{tabular} \caption{Conduct of the eos for different values of interacting parameter W. } \end{figure*} \end{center} \begin{center} \begin{figure*} \begin{tabular}{c@{\quad}c} \epsfig{file=compare_omegam.eps,width=7.5 cm} \end{tabular} \caption{Percentage diversion of $ \Omega_{m} $ in IQ as compared to $ \Lambda$CDM model for different values of interacting parameter W: negative values in y-axis means they are less than that that of $ \Lambda $CDM. $\Omega_{m0} = 0.28$ and $\lambda_{i} = 0.7$ in these plots. Subsequently , $\% \Delta X = (X^{de} / X^{\Lambda} - 1)\times 100$. } \end{figure*} \end{center} \begin{center} \begin{figure*} \begin{tabular}{c@{\quad}c} \epsfig{file=compare_hubble.eps,width=7.5 cm} \end{tabular} \caption{Percentage deviation in normalized hubble h from $ \Lambda$CDM model for different values of interacting parameter W.. } \end{figure*} \end{center} \section{GROWTH OF LINEAR PERTURBATION WITH INTERACTING Quintessence} Herein effect of the interaction on matter and scalar field perturbation in linear regime will be reflected upon. Matter include both dark matter and baryons. Since dark matter perturbation is dominant and baryon follows dark matter perturbation, hence baryonic perturbations in our study can be excluded . Even inclusion of that will not effect our result. \\ The perturbed (FRW) metric in conformal and Newtonian gauge is given by\\ \begin{equation} ds^{2}=a^{2}[-(1+2\Phi)d\tau^{2}+(1-2\Psi)dx^{i}dx_{j}] \end{equation} where, $\Phi$ and $\Psi$ are gravitational potential and $\tau$ is conformal time. Without anisotropic stress the two can be related as $ \Phi = \Psi $.\\ Consider two components, a scalar field and cold dark matter described by the energy- momentum tensors $T_{\mu\nu(\phi)}$ and $T_{\mu\nu(d)}$. Conservation equations with interacting terms for scalar field and cold dark matter as discussed by Amendola \cite{Amendola} are. \begin{equation} \nabla_{\mu}T^{\mu}_{(\phi) \nu}= C T_{d} \nabla_{\nu}\phi \end{equation} \begin{equation} \nabla_{\mu}T^{\mu}_{(c) \nu}= -C T_{d} \nabla_{\nu}\phi \end{equation} here, C defines interaction between scalar field and dark matter. We assume phenomenologically, C to be constant.\\ For perturbation equations we follow the set up as provided by Bikash et al .We generalize the results in \cite{Bikash} for the interacting scenario. We mention only relevent equations here , for detailed calculation see eg \cite{Bikash}. \\ Relativistic poisson equation is. \begin{equation} \bigtriangledown^{2}\Phi-3 {\cal {H}}(\Phi^{'}+{\cal {H}}\Phi)= 4\pi G a^{2} \Sigma\delta\rho_{i} \end{equation} The scalar field influences a change upon source term of Poisson equation . \cite{NL4,NL5}.Here prime is derivative wrt conformal time. The perturbation equations for scalar field coupled to dark matter is \begin{equation} \delta\phi^{..}+3H\delta\phi^{.} - 4\Phi^{.}\phi^{.} + 2\Phi V_{\phi} +V_{\phi \phi} \delta\phi - \frac{1}{a^{2}} \nabla^{2} \delta\phi = C(\delta\rho_{\phi} + 2\Phi\rho_{\phi}) \end{equation} \\ where, $V_{\phi}=\dfrac{dV}{d\phi}$, $V_{\phi \phi}=\dfrac{d^{2}V}{d^{2}\phi}$ and $\delta\phi$ is scalar field fluctuation. Then we construct following dimensionless variables. \begin{eqnarray} && g = -\frac{V_{\phi}}{H \phi \dot{•}}\\ && q = \frac{\delta\phi H}{\phi \dot{•}}\\ && B = 3 + \frac{H\dot{•}}{H^{2}}\\ && B_{\phi} = 6g - \frac{1}{H}\frac{dB}{dt} - 2Bg + \frac{k^{2}}{a^{2}H^{2}}\\ && W = \frac{C}{k}\\ \end{eqnarray} In terms of these dimensionless parameters eq (11) and (12) can be written as \begin{eqnarray} && \frac{d^{2}\Phi}{dN^{2}}+(1+B)\frac{d\Phi}{dN} + (2B-3+3x^{2})\Phi = 3x^{2}[\frac{dq}{dN} + \frac{3W\Omega_{d}q}{\sqrt{6}x} + (2g-B)q]\\ &&\frac{d^{2}q}{dN^{2}}+ (\frac{\sqrt{6}W\Omega_{d}}x + 2g - B)\frac{dq}{dN} + (\frac{9W\Omega_{d}}{\sqrt{6}x} - 3\Omega_{d}W^{2} + B_{q})q = 4\frac{d\Phi}{dN} + 2\Phi g + \frac{3W\Omega_{d}}{\sqrt{6}x}(\delta_{d}+2\Phi)\nonumber\\ \end{eqnarray} Matter density contrast is given by \begin{equation} \delta_{d} = -\frac{2}{\Omega_{d}}[\frac{d\Phi}{dN} + (1-x^{2}+\frac{k^{2}}{3 {\cal {H}} ^{2}})\Phi + x^{2}(\frac{dq}{dN}-B_{q}+\frac{3W\Omega_{d}q}{\sqrt{6}x})]\\ \end{equation} where, $N = lna$ \\ Here, comoving density contrast \begin{equation} \Delta_{d} = \delta_{d} + y_{d} \end{equation} where, \begin{equation} y_{d} = 3 {\cal {H}}v_{d} = \frac{2}{\Omega_{d}}(\frac{d\Phi}{dN} + \Phi -3x^{2}q) \end{equation} On subhorizon scale $ \Delta_{d} \simeq \delta_{d} $ , that is usually used in study of small-scale structure (Newtonian limit). On large scales, $\Delta_{d}$ should be used instead of $\delta_{d}$ \cite{Duniya2013}. For growth function, we use equations given by Duniya et al in \cite{L7} and probe the large - scale structure of the universe at different redshifts. In the following equations $\Phi_{d}$ represents gravitational potential at decoupling.\\ Gravitational potential growth function $D_{\Phi}$ is defined by \begin{equation} D_{\phi}(k,z)=\frac{1}{(1+z)}\dfrac{\Phi(k,z)}{\Phi_{d}(k)} \end{equation} $D_{d}$ is the growth function of the comoving matter overdensity.\\ \begin{equation} D_{d}(k,z)=-\dfrac{3\Delta_{d}(k,a)\Omega_{d0}H_{0}^{2}}{2k^{2}\Phi_{d}(k)} \end{equation} Matter velocity growth function \begin{equation} Dv_{d}(k,z)=-\dfrac{3v_{d}(k,z)\Omega_{d0}H_{0}^{2}}{2\Phi_{d}(k)} \end{equation} Dark energy velocity growth function $D_{\phi}(k,z)$ is associated to that of dark matter by \begin{equation} \frac{Dv_{\phi}(k,z)}{Dv_{d}(k,z)}= -\frac{\Omega_{d}}{(1-\Omega_{d})(1+w)}(\frac{HD_{\Phi}^{'}}{Dv_{d}}+1) \end{equation} Ratio of comoving matter density $\Delta_{d}$ and gravitational potential $\Phi(0,k)$ can be defined as \begin{equation} D_{d\Phi}(z,k)=\dfrac{\Delta_{d}(z,k)}{\Phi(0,k)} \end{equation} here, prime is derivative with respect to redshift z.\\ Quantity f which is related to velocity perturbation and hence redshift-space distortion \cite{Duniya2016} term can be defined as \begin{equation} f=\frac{Dv_{d}}{{\cal {H}}D_{d}} \end{equation} which reduces to the growth rate of dark matter in standard uncoupled DE m0dels.\\ Standard matter power spectrum \begin{equation} P(z,k) = A k^{n_s-4} T^2(k)(1+z)^{2}(\frac{\Delta_{d}(z,k)}{\Phi(0,k)})^{2} \end{equation} Here $A$ represents the normalization constant the value of this is determined by $\sigma_{8}$ normalization whereas $n_{s}$ is defined as spectral index , $T(k)$ as given by Einstein and Hu is the transfer function \cite{Eisenstein1998}. \subsection{Initial conditions} To solve perturbation equations (19), (20), we need initial conditions for ($\Phi, \dfrac{d\Phi}{dN}, q, \dfrac{d q}{dN} $) .At the time when matter dominated Universe there was negligible dark energy contribution. We have settled set our initial condition at decoupling ($z=1000$). Due to negligible dark matter initial conditions for interacting and non interacting model is same. Thus we have used initial conditions discussed by Bikash et al \cite{Bikash} for non-interacting model.. There is no contribution from DE at $z=1000$ hence scalar field perturbation is insignificant and $q= \dfrac{d q}{dN} = 0$ . Since $\Phi$ is constant during matter domination, hence $\dfrac{d \Phi}{dN}=0$ . Using Poisson equation (9) and the fact that $\Delta_{m} \sim a$ , its initial condition is \begin{equation} \Phi_{i} = \frac{-3}{2}\frac{{\cal {H}}^{2}_{in}}{k^{2}}a_{in} \end{equation} \begin{center} \begin{figure*} \begin{tabular}{c@{\quad}c} \epsfig{file=compare_potentialz0.eps,width=7.5 cm} \epsfig{file=compare_potentialz05.eps,width=7.5 cm}\\ \epsfig{file=compare_potentialz1.eps,width=7.5 cm} \epsfig{file=compare_potentialz35.eps,width=7.5 cm} \end{tabular} \caption{Percentage diversion of $ \Phi $ in IQ model as compared to $ \Lambda$CDM model for different values of interacting parameter W. } \end{figure*} \end{center} \begin{center} \begin{figure*} \begin{tabular}{c@{\quad}c} \epsfig{file=compare_deltamz0.eps,width=7.5 cm} \epsfig{file=compare_deltamz05.eps,width=7.5 cm}\\ \epsfig{file=compare_deltamz1.eps,width=7.5 cm} \epsfig{file=compare_deltamz35.eps,width=7.5 cm} \end{tabular} \caption{Percentage diversion of comoving density contrast $ \Delta_{m} $ in IQ model as compared to $ \Lambda$CDM model . } \end{figure*} \end{center} \begin{center} \begin{figure*} \begin{tabular}{c@{\quad}c} \epsfig{file=compare_potential_deltamz0.eps,width=7.5 cm} \epsfig{file=compare_potential_deltamz05.eps,width=7.5 cm}\\ \epsfig{file=compare_potential_deltamz1.eps,width=7.5 cm} \epsfig{file=compare_potential_deltamz35.eps,width=7.5 cm} \end{tabular} \caption{Ratio of $D_{d\Phi}(k,z)$ IQ model and $\Lambda$CDM defined in equation (28) . } \end{figure*} \end{center} \begin{center} \begin{figure} \begin{tabular}{c@{\quad}c} \epsfig{file=compare_growthz0.eps,width=7.5 cm} \epsfig{file=compare_growthz05.eps,width=7.5 cm}\\ \epsfig{file=compare_growthz1.eps,width=7.5 cm} \epsfig{file=compare_growthz35.eps,width=7.5 cm} \end{tabular} \caption{Percentage deviation in $f$ from $ \Lambda$CDM model.} \end{figure} \end{center} \subsection{Behaviour of cosmological parameters} Using aforementioned initial conditions we solve perturbed equation (19) and (20 ) and study various pertubation parameters, for different sets of interacting parameter W. Figure 4 displays deviation in gravitational potential from LCDM model for different values of interacting parameter. For low redshift on sub-hubble scale there is a suppression from LCDM , which is due to different background evolutions as there is no contribution from DE perturbation at smaller scales. However on larger scale the enhancement in $\Phi$ is due to contribution from dark energy perturbtion . On increasing the value of interacting parameter W there is overall enhancement on subhubble and super hubble scale respectively.This is due to transfer of energy and momenutm from DM to DE in both background as well as perturbed Universe. Figure 5 displays variation in comoving matter density contrast $\Delta_{m}$. For non-interacting case variation is very small and is almost scale independent. At low redshift ($z=0,0.5,1$) on large scale an increasing in interaction (W) results in an enhancement in $\Delta_{m}$ wrt LCD ($ 0-8.5\%$) which decrease with redshift. But on smaller scales $\Delta_{m}$ is slightly suppressed less than $ 1\%$ . At redshift z=3.5 on large scale there is slight suppression wrt lcdm less than$ 1\%$ and on small scale slight enhancement less than$ 1\%$. Figure 6 displays variation of f defined in equation (31) which is related to velocity perturbation and hence redshift space distortion. For non interactig case just like $\Delta_{m}$ variation is small and scale independent. On adding interaction its behaviour is exactly opposite to that of matter density contrast $\Delta_{m}$. On large scale with increasing interaction a suppression can be seen at red shift 0,0.5 and 1 ($4-8\%$). But at higher redshift at $z=3.5$ there is slight enhancement less than $ 1\%$ . Hence interaction effect growth of structure at all scales.But this effect is smaller at higher redshift. \section{ Influence of Interaction on power spectrums} The growth of large scale structure in the Universe is ascertain by matter power spectrum. Forthcoming surveys of galaxies can probe distribution of dark matter on large scales .These surveys can provide strong bound on dark energy models including interaction in dark sector. We must incorporate observed galaxy distribution effects like redshift space distortion , GR effects like weak lensing convergence , SW, ISW, time delay effect in our analysis to recognize potentiality of these surveys \cite{Yoo2010ni,Challinor2011bk,Jeong2011as,Bertacca2012tp,Hu2001yq}. Certain astrophysical processes such as gas cooling, star formation and feedback from supernovae, in conjunction with the gravitational effect of the dark matter has a bearing upon formation of galaxy. This can further cause a contrast between the spatial distribution of baryons and dark matter. The association between the spatial distribution of galaxies and ubiquitous dark matter must be understood to employ galaxies as cosmolgical probes known as galaxy bias. Matter power spectrum can be related to galaxies distribution through bias $b$ defined as \cite{Challinor2011bk,Bruni2011ta,Jeong2011as} \begin{equation}\label{bias} \Delta_{\rm g}(k,z) = b(z)\, \Delta_d(k,z). \end{equation} To study effect of interaction on matter and galaxy power spectrum we use prescription discussed by Duniya et al in \cite{L7}. In a galaxy redshift survey, the observers measure the number of galaxies in direction n at redshift z.The number overdensity of Galaxy () $\Delta^{obs}$) is shown as follows \begin{equation} \Delta^{obs} = \left[{b + f \mu^2} + \mathcal{A} (\frac{\mathcal{H}}{k})^2 + i\mu\mathcal{B} (\frac{\mathcal{H}}{k})\right]\Delta_{m}, \end{equation} Here $\mu = -\frac{\vec{n}.\vec{k}}{k}$ , $\vec{n}$ expressinng the direction of observation, $f$ signifies the redshift space distortion and b stands for galaxy bias. Variables A and B which are considered in connection to GR corrections are delineated as follows: \begin{equation} A=(3-b_{e})f-\frac{3\Omega_{d0}H_{0}^{2}}{2{\cal H}^{2} D_{d}}(4Q-b_{e}-1-(1+z)\frac{{\cal H}^{'}}{\cal H}+2\frac{(1-Q)}{r{\cal H}} + \frac{1}{D_{\Phi}(1+z)^{2}}(-(1+z)^{2}(D_{\Phi}+(1+z)D_{\Phi}^{'})D_{\Phi}(1+z) \end{equation} \begin{equation} B=[b_{e}-2Q+(1+z)\frac{{\cal H}^{'}}{\cal H}-2\frac{(1-Q)}{r{\cal H}}-w\sqrt{6}x(1-\frac{Dv_{\phi}}{Dv_{d}})]f \end{equation} Full general relativistic power spectrum that includes all GR effects \cite{L7}. \\ \begin{equation} P_{g}^{obs}(k,z)=((b+f\mu^{2})^{2}+2(b+f\mu^{2})\frac{A{\cal {H}}^{2}}{k^{2}}+\frac{A^{2}{\cal {H}}^{4}}{k^{4}}+\mu^{2}\frac{B^{2}{\cal {H}}^{4}}{k^{4}})P(k,z) \end{equation} Matter power spectrum with kaiser term is given by \begin{equation} P_{k}(z,k) = \big[b(z) + f(z,k)\mu^2\big]P(z,k) \end{equation} Here prime is derivative wrt redshift.\\ A constant comoving galaxy number density is presumed hence $b_e =0$ , galaxy bias $b=1$ and the magnification bias ${Q} = 1,$. Here x is defined in equation (4). A is in connection with peculiar velocity potential and the gravitation potential, on the other hand B is connected to the Doppler effect. The latter comprises an interaction explicitly. However no momentum is transported in the Dark energy rest frame which leads to last term in equation (35) would be zero \cite{L7} . \begin{center} \begin{figure*} \begin{tabular}{c@{\quad}c} \epsfig{file=comparepowrspectrm_stdz0.eps,width=5.5 cm} \epsfig{file=comparepowrspectrm_kaisz0.eps,width=5.5 cm} \epsfig{file=comparepowrspectrm_genz0.eps,width=5.5 cm}\\ \epsfig{file=comparepowrspectrm_stdz05.eps,width=5.5 cm} \epsfig{file=comparepowrspectrm_kaisz05.eps,width=5.5 cm} \epsfig{file=comparepowrspectrm_genz05.eps,width=5.5 cm}\\ \epsfig{file=comparepowrspectrm_stdz1.eps,width=5.5 cm} \epsfig{file=comparepowrspectrm_kaisz1.eps,width=5.5 cm} \epsfig{file=comparepowrspectrm_genz1.eps,width=5.5 cm}\\ \epsfig{file=comparepowrspectrm_stdz35.eps,width=5.5 cm} \epsfig{file=comparepowrspectrm_kaisz35.eps,width=5.5 cm} \epsfig{file=comparepowrspectrm_genz35.eps,width=5.5 cm}\\ \end{tabular} \caption{Percentage of diversion in $ P (k) $ in IQ model as compared to $ \Lambda$CDM model for different values of interacting parameter W as a function of $K$ for different redshifts. First column is standard matter power spectra $P$ which is specified by eqn. (30), Second column present Kaiser power spectra $P_{k}$ specified by eqn (37), third column portrays galaxy power spectrum specified by eq (36). } \end{figure*} \end{center} Figure 8 (left column) shows percentage change in standard matter power spectrum wrt $\Lambda$CDM. It depends on ratio $\frac{\Delta_{d}}{\Phi}$ which is shown in figure (6). At $z=0$ and on large scale percentage change for non-interacting case ($W=0$) is -ve which shows suppression in power.But on addding interaction ($2-8\%$) enhancement is observed.Which shows transfer of energy from scalar field to dark matter , rate of transfer increase with increase in interaction parameter W.On smaller scale all models converge to$\Lambda$CDM due to our Normalization.On larger scale and higher red shift dark energy gives negative contribution as compared to $z=0$ .Thus suppression is seen wrt $\Lambda$CDM. This suppression contribution increases with redshift.But on smaller scales slight enhancement of power can be seen which is due to difference in background evolution. Figure 8 (middle column) displays percentage change in matter power spectrum with kaiser term wrt $\Lambda$CDM. At $z=0$ on large scale enhancement in kaiser spectrum is ($0-2\%$) which is less than standard matter power spectrum discussed earlier. Which is due to contribution from kaiser term eq (37) that give negative contribution to power spectrum. It depends on parameter f eq (33) . Behaviour of f is shown shown in fig 7 which is suppressed wrt $\Lambda$CDM. On higher redshift suppression is more as compared to standard power spectrum . But on smaller scale all models converge to$\Lambda$CDM for all redshift. Figure 8 (right column) displays percentage change in galaxy power spectrum wrt $\Lambda$CDM. At $z=0$ for large scales percentage change is negative ($12-23\%$) which shows suppression wrt $\Lambda$CDM. On increasing interaction this suppression increases and it decreases with redhsift. The reason of suppression is general relativistic terms A and B in equation (34) which give negative contribution. On small scales and low redshift percentage change is small but on high red shift $z=3.5$ percentage change is positive ($2-6\%$).The reason for this is weaker dark energy effect at higher z but GR effects are comparative stronger even on small scales. \section{Conclusion} We have generalised non-interacting quientessence model as discussed by Bikas et al \cite{Bikash} for interacting scenerio. We studied effect of interaction between Quientessence and dark matter on both background and perturbed universe. At background level interaction effect bacground energy density and hubble parameter .Suppression in dark matter density $\Omega_{d}$ wrt $\Lambda$CDM is ($0-8\%$) . While enhancement in hubble is ($0-8\%$). Thus interaction makes dark energy effect stronger. A careful analysis of growth of strucure characterises that interaction effects matter, kaiser and galaxy power spectrum on sub and super horizon scales. The former is because of dark energy perturbation and GR effect and the latter is the result of background evolution. We find that Standard matter power spectrum at $z= 0$ is enhanced wrt $\Lambda$CDM on large scales and on small scale no deviation due to our normalization. This enhancement on large scale increases with increase in interaction. At higher redshift matter power spectrum suppressed wrt $\Lambda$CDM. This suppression also increases with interaction. On adding kaiser redshift distortion term enhancement is less as compared to standard matter power spectrum. Thus it can be concluded that that kaisr term gives negative contribution to power spectrum.Which suppress power on large scale.On higher redshift it is further suppressed . We also found that Galaxy power spectrum is suppressed ($12-24\%$) wrt $\Lambda$CDM on large scales .This suppression increases with interaction. But on higher redshift effect of dark energy is weaker hence suppression decreases with increase in z. But on small scales enhancement is observed ($2-6\%$) which is due to difference in background evolution and GR effects .Thus on higher redshifts interaction effects galaxy power spectrum even on smaller scale. This deviation can be probed by future surveys like SKA.

8284c93ccf6291d7bed310265d710ebd0ebc1ac4

\section{Introduction} Energy level splitting is a very well-known phenomenon in physics, which occurs whenever a set of degenerate levels is split in an external field. One of the best known examples is the Zeeman effect in which rotational states are split in an external magnetic field according to their magnetic substates (\Cref{fig:Zeeman effect}) \cite{bohr1998nuclear}. \begin{figure}[h] \centering{} {\centering{}\includegraphics[scale=0.55]{Zeeman_front_page} }\hspace{2pc} \begin{minipage}[b]{16pc}\caption{Zeeman effect: splitting of levels due to an external magnetic field. \label{fig:Zeeman effect}} \end{minipage} \end{figure} \\ Another example is the Nilsson model of deformed shell model where the single-particle levels are split in a deformed quadrupole field \cite{casten2000nuclear}. In this case the energy levels are not only split but also strongly mixed by the quadrupole field. The amount of mixing depends on the strength of the deformed field (\Cref{fig:Nilsson Model}). Level crossing and repulsion occur due to this deformation. In the case of the Zeeman effect, the levels are split but not mixed, and can therefore be labeled by the same quantum numbers as before the splitting, i.e the angular momentum $l$ and its projection m$_{l}$. In the case of the Nilsson model, the levels are strongly mixed and are labeled by the quantum numbers in the intrinsic or body-fixed frame $\left[\mathrm{N}n_{z}\varLambda\mathrm{\Omega}\right]^\pi$ where $\Omega$ denotes the projection of the single-particle angular momentum on the symmetry axis (z-axis), $\pi$ is the parity, $N$ the principal quantum number, $n_{z}$ the number of nodes in the wave function in the z-direction and $\varLambda$ the projection of the orbital angular momentum on the symmetry axis. \begin{figure}[h] \centering{} {\centering{}\includegraphics[scale=0.55]{NilssonDeformation}}\hspace{2pc} \begin{minipage}[b]{15pc}\caption{Splitting of single-particle levels as a function of the quadrupole deformation in the deformed shell model. Levels with even and odd parity are drawn with solid and dashed lines respectively (taken from \cite{bohr1998nuclear}). \label{fig:Nilsson Model}} \end{minipage} \end{figure} \\ In our work, we use a deformed field, that of cluster potentials, where the clusters are alpha particles. The deformation is taken to be the relative distance to the center of mass of the whole $k\alpha+x$ nucleon structure, where is $x$ a nucleon. The splitting of the single-particle levels is studied as a function of the relative distance for the cases of a dumbbell, equilateral triangle and a regular tetrahedron, where a symmetry factor is crucial for the level splitting and configuration mixing. Relabeling is also necessary in this case specially for the cases of an equilateral triangle and regular tetrahedron. The main goal for this study is to obtain patterns that gain insight into how the single-particle levels evolve with deformation. That is done by studying the mixing between states, how these are related with the geometrical configuration and what kind of symmetry is present. \section{Cluster Shell Model} We start by reviewing the Cluster Shell Model \cite{DELLAROCCA2017158}. In this model nuclei with $Z=N=2k$ are treated as a cluster of $k$ $\alpha$-particles whose matter and charge density are given by a gaussian form \begin{align} \rho\left(\vec{r}\right)=\left(\frac{\alpha}{\pi}\right)^{\frac{3}{2}}e^{-\alpha\left(r^{2}+\beta^{2}\right)}4\pi\sum_{\lambda\mu}i_{\lambda}\left(2\alpha\beta r\right)Y_{\lambda\mu}\left(\theta,\phi\right)\sum_{i=1}^{k}Y_{\lambda\mu}^{*}\left(\theta_{i},\phi_{i}\right).\label{eq:distributionnucleus} \end{align} Here $\vec{r}_{i}=\left(\beta,\theta_{i},\phi_{i}\right)$ where $\beta$ denotes the relative distance of the $\alpha$ particles to the center of mass, and $\theta_{i}$ and $\phi_{i}$ are the angles. An important factor is the cluster factor, $\sum_{i=1}^{k}Y_{\lambda\mu}^{*}\left(\theta_{i},\phi_{i}\right)$, which contains the information about the geometrical configuration of the alpha particles. \\ The potential is obtained by convoluting the density, \cref{eq:distributionnucleus}, with a Volkov potential \cite{VOLKOV196533} to obtain \begin{equation} V\left(\vec{r}\right)=-V_{0}\sum_{\lambda\mu}4\pi e^{-\alpha\left(r^{2}+\beta^{2}\right)} i_{\lambda}\left(2\alpha\beta r\right)Y_{\lambda\mu}\left(\theta,\phi\right)\sum_{i=1}^{k}Y_{\lambda\mu}^{*}\left(\theta_{i},\phi_{i}\right)\label{eq:potencialcentral}. \end{equation} \\ \begin{center} \begin{figure}[h] \centering{}\includegraphics[scale=0.5]{pasted1}\hspace{2pc}\begin{minipage}[b]{15pc}\caption{\label{fig:Geometrical-configuration}Geometrical configuration for $k=2,3,4$ clusters (taken from \cite{DELLAROCCA2017158}).} \end{minipage} \end{figure} \par\end{center} The spin-orbit interaction is taken as \begin{equation} V_{so}\left(\vec{r}\right)=V_{0,so}\frac{1}{2}\left[\frac{1}{r}\frac{\partial V\left(\vec{r}\right)}{\partial r}\left(\vec{s}\cdot\vec{l}\right)+\left(\vec{s}\cdot\vec{l}\right)\frac{1}{r}\frac{\partial V\left(\vec{r}\right)}{\partial r}\right]. \end{equation} Finally, the coulomb potential is found by convoluting the carge density $\frac{Ze^2}{k}\rho\left(\vec{r}\right)$ with Green's function giving \begin{eqnarray} V_{C}\left(\vec{r}\right) & = & \frac{Ze^{2}}{k}\left(\frac{\alpha}{\pi}\right)^{\frac{3}{2}}\sum_{\lambda\mu}\frac{4\pi}{2\lambda+1}Y_{\lambda\mu}\left(\theta,\phi\right)\sum_{i=1}^{k}Y_{\lambda\mu}^{*}\left(\theta_{i},\phi_{i}\right)\times \\ & & \times\left[\frac{1}{r^{\lambda+1}}\intop_{0}^{r}f_{\lambda}\left(r'\right)r'^{\lambda+2}dr'+r^{\lambda}\intop_{r}^{\infty}f_{\lambda}\left(r'\right)\frac{1}{r'^{\lambda-1}}dr'\right]. \end{eqnarray} We then obtain the single-particle energy levels and intrinsic states as a function of $\beta$ for each configuration $\left(k=2,3,4\right)$ of \cref{fig:Geometrical-configuration} by solving the single-particle Schr$\mathrm{\ddot{o}}$dinger equation \begin{equation} H=\frac{\vec{p}^{2}}{2m}+V\left(\vec{r}\right)+V_{so}\left(\vec{r}\right)+V_{C}\left(\vec{r}\right)\label{eq:hamiltonian free particle}. \end{equation} In the case of neutrons $V_{C}\left(\vec{r}\right)=0$. \section{Results} The Hamiltonian of \cref{eq:hamiltonian free particle} is diagonalized in the harmonic oscillator basis. The correlation diagrams are shown as a function of $\beta$ in \Cref{Correlation-Diagrams_pvsn,CorrTrianPirOvlp}. The states for the spherical case correspond to $\beta=0$. For the two-body cluster configuration the levels with projection $\pm \Omega$ are degenerate. As a consequence, a single-particle level $\left(l,\nicefrac{1}{2}\right)j$ is split into a series of doublets with $\Omega=\nicefrac{1}{2},\ \nicefrac{3}{2},\ \ldots,\ j $ and parity $P=\left(-\right)^l$. For the triangular configuration neither the angular momentum nor parity is conserved. The spherical single-particle levels are split into a series of doublets characterized by $E'_{2}$ and $E'_{1}$, the double valued representations of $D_{3}$, the rotational subgroup of $D_{3h}.$ For the tetrahedral configuration, the single-particle levels are split into a series of doublets and quadruplets characterized by $E'$ and $G'$, respectively, the double valued representations of $T$, the rotational subgroup of $T_{d}$. The results are summarized in \cref{EnergyLvlsplit}. \begin{table}[h] \centering \caption[]{Resolution of single-particle levels into the double-valued representations of the corresponding point group. Notation from Hamermesh \cite{hamermesh1962group}.} \label{EnergyLvlsplit} \begin{tabular}{cccc} \br \noalign{\smallskip} $j$ & Dumbbell & Triangle & Tetrahedron \\ \noalign{\smallskip} \hline \noalign{\smallskip} $\nicefrac{1}{2}$ & $(\nicefrac{1}{2})^P$ & $E'_{2}$ & $E'$ \\ $\nicefrac{3}{2}$ &$(\nicefrac{1}{2}+ \nicefrac{3}{2})^P$ & $E'_{1}+E'_{2}$ & $G'$ \\ $\nicefrac{5}{2}$ &$(\nicefrac{1}{2}+\nicefrac{3}{2}+ \nicefrac{5}{2})^P$ & $E'_{1}+2E'_{2}$ & $E'+G'$ \\ $\nicefrac{7}{2}$ &$(\nicefrac{1}{2}+\nicefrac{3}{2}+ \nicefrac{5}{2}+\nicefrac{7}{2})^P$ & $E'_{1}+3E'_{2}$ & $2E'+G'$ \\ \noalign{\smallskip} \br \end{tabular} \end{table} \\ The first result that we can see by comparing \Cref{Correlation-Diagrams_pvsn} is that the correlation diagrams for protons and neutrons behave in the same manner. This result shows that it does not matter much whether the particle is a neutron or a proton, there is just a small shift in energy. \\ In order to adress the issue of mixing between the levels we study the overlap between intrinsic states of different values of deformation, $\left\langle \chi_{\Omega}\left(\beta\right)\left|\chi_{\Omega'}\left(\beta'\right)\right.\right\rangle $. We can then have the overlap with a fixed value of $\beta'$, namely $\beta'=0$. In that way we can see the mixing between states as we vary $\beta$. This can be seen in \Cref{CorrTrianPirOvlp}, where from comparing the graphs of overlap of the triangle and tetrahedron, there is present more mixing in the case of triangle than the case of tetrahedron. A main reason why this occurs is because the density of states of same symmetry is higher in the configuration of triangle, as a result those levels are then closer energetically to each other and can mix more strongly in comparison to the tetrahedron. \begin{center} \begin{figure}[h] \noindent \begin{centering} \subfloat {\includegraphics[scale=0.45]{PlotNiveles_de_energiaprotonesDumbell} \label{fig:Energy-level-splitting DumbbellProton}} \qquad \subfloat {\includegraphics[scale=0.45]{PlotNiveles_de_energiaDumbellneutrones} \label{fig:Energy-level-splitting DumbbellNeutron}} \caption{\label{Correlation-Diagrams_pvsn}Correlation diagrams for the case of dumbbell, obtained from \cref{eq:hamiltonian free particle} with $V_{0}=32$ MeV, $\alpha=0.0511\ \mathrm{fm}^{-2}$ and $V_{0,so}=27.5$ MeV fm$^2$ \cite{DELLAROCCA2017158} for a proton (left) and a neutron (right).} \par\end{centering} \end{figure} \end{center} \begin{figure}[!h] \subfloat {\includegraphics[scale=0.50]{PlotNiveles_de_energiaTriangle} \label{Correlationtriangle}} \quad \subfloat{\includegraphics[scale=0.50]{PlotNiveles_de_energiaPiramid} \label{Correlationtetrahedron}} \\ \subfloat{\includegraphics[scale=0.50]{PlotOvlpTriangel} \label{overlaptriangle}} \quad \subfloat{\includegraphics[scale=0.50]{PlotOvlpPiramid} \label{overlaptetrahedron}} \caption{\label{CorrTrianPirOvlp}Correlation diagrams (top) and overlaps between states of the same symmetry (bottom) for the case of a triangle (left) and a tetrahedron (right), and a single-particle neutron, obtained from \cref{eq:hamiltonian free particle} with $V_{0}=32$ MeV, $\alpha=0.0511\ \mathrm{fm}^{-2}$ and $V_{0,so}=17$ MeV $\mathrm{fm^{2}}$ for the triangle, and $V_{0,so}=12.1$ MeV $\mathrm{fm^{2}}$ for the tetrahedron \cite{DELLAROCCA2017158}.} \end{figure} For the case of triangle there exist strong repulsion between states, which causes crossings of wave function from one level to another. These cases can be seen in the graph of overlap of the triangle in the region 3.5 to 4.5 fm for the states belonging to the $E'_{2}$ symmetry. In states belonging to the $E'_{1}$ that doesn't happen, and the mixing occurs smoothly with no wave function crossings. \\ In the case of the tetrahedron, although there occurs a wave function crossing (between 3.5 to 4 fm) this doesn't happens as much as in the case of triangle. There is strong mixing but states are sufficiently separated so that wave function crossings don't occur as often as in the case of triangle. \\ In both cases the difference between states of the same symmetry group arise from how much energetically distanced they are from other states. A simple example are the ground levels, which in comparison to higher levels present much less mixing. That can be seen by comparing the curves $\left\langle \chi_{\Omega}\left(\beta\right)\left|\chi_{\Omega}\left(\beta'=0\right)\right.\right\rangle$ where $\Omega$ can be any of the double-valued representations for triangle and tetrahedron shown in \Cref{EnergyLvlsplit}. The faster those curves decrease the more the state has mixed with others. Again, the mayor difference between the states is seen in the case of the triangle. Also the higher the state of the same group of symmetry the more it mixes. \\ From both correlation diagrams of \Cref{CorrTrianPirOvlp} it can be seen that the single-particle levels asymptotically tend to a certain degeneracy, which corresponds to the case where the alpha particles are too far apart and thus interact as separated alpha particles. This is something that is not obtained in the Nilsson model, where the degeneracy shown is that of an elongated cigar. As this occurs mixing between states decrease. This is shown in the curves of $\left\langle \chi_{\Omega}\left(\beta\right)\left|\chi_{\Omega'}\left(\beta'=0\right)\right.\right\rangle$. Here $\Omega$ and $\Omega'$ are states of the same symmetry, but not the same state. The decrease starts in most cases at 5 fm, hinting that the frontier between the moment where the alphas can still be seen as to be part of the same nucleus and not separate particles is around that value. In the correlation diagrams it is also shown that convergence towards the degeneracy of separate alphas occur after 5 fm. \section{Conclusions} From the obtained patterns it has been illustrated how single-particle levels evolve with deformation and which states interact most and why. From the overlap it can also be thought which interactions could be the predominant ones (quadrupole, octupole, etc). This and the interpretation of the behaviour due to the symmetry will be further analysed. All the information from this study can later be used to extend the Algebraic Cluster Model (ACM)\cite{Bijker2000}\cite{Bijker2014} to odd-cluster nuclei like $^{9}$Be, $^{9}$B, $^{13}$C, $^{13}$N, $^{17}$O and $^{17}$F, by developing the Algebraic Cluster-Fermion Model (ACFM)\cite{Adrian2018}. \ack This work was supported in part by grant IN109017 from DGAPA-UNAM, Mexico. \section*{References}

b1402371d2c628a1af8ee314faec282268317f29

\section*{Abstract} \textbf{The surface of the Martian moon Phobos exhibits two distinct geologic units, known as the red and blue units. The provenance of these regions is uncertain yet crucial to understanding the origin of the Martian moon and its interaction with the space environment. Here we show that Phobos' orbital eccentricity can cause sufficient grain motion to refresh its surface, suggesting that space weathering is the likely driver of the dichotomy on the moon's surface. In particular, we predict that blue regions are made up of pristine endogenic material that can be uncovered in steep terrain subject to large variations in the tidal forcing from Mars. The predictions of our model are consistent with current spacecraft observations which show that blue units are found near these regions.} \section*{Acknowledgements} R.L.B acknowledges support from JAXA's Aerospace Project Research Associate Program. N.B. conducted this work as a JSPS International Research Fellow. S.T.C. was supported by the JAXA International Top Young Fellowship Program. The authors also thank Patrick Michel for constructive feedback on the results and implications of this work. Grain dynamics simulations were calculated on YORP cluster run by the Center for Theory and Computation at the Department of Astronomy at the University of Maryland. For data visualization, the authors made use of the freeware, multi-platform, ray-tracing package, Persistence of Vision Raytracer. \section*{Author Contributions} R.L.B. conceptualized the study, designed and performed the local simulations of granular dynamics, and led the research. N.B. initiated the project through a study of the three body elliptical problem on Phobos, performed the gravitational dynamics calculations, and contributed to the analyses. S.T.C. provided geophysical and geomorphological expertise, remote sensing analysis, and constructed the model for regolith development on Phobos. Y.K. provided guidance and advice on the formulation and scope of the research. M.F. provided guidance and discourse on the implications of the results. Y.K. and M.F. provided expertise in small body exploration and contextualized the research in the frame of the MMX mission. All authors contributed to the interpretation of the results and preparation of the manuscript. \section*{Competing interests} The authors declare no competing interests. \section*{Data Availability} The datasets generated and analyzed during the current study are available from the corresponding author on reasonable request. \section*{Code Availability} The code used to generate the datasets are available from the corresponding author on reasonable request. \section{Introduction} \label{S:1} Previous dynamical analyses of Phobos' decaying orbit have shown that changes to its tidal environment lead to surface mobility and evolution, such as the formation of mass-wasting features \cite{Shi2016} and linear grooves \cite{Hurford2016}. These studies ignore the eccentricity of the moon's orbit, as this would have a small effect over the dynamical timescales set by the orbital decay rate of the planetary satellite ($10^8$ years). Here, we model the effect of Phobos' eccentric orbit at its current distance, and demonstrate that time-varying forces can lead to substantial mass motion and surface erosion over much shorter timescales ($10^3$-$10^6$ years). We find that librations induced by the eccentricity of the Martian moon can vary dynamical slopes by up to 2$^{\circ}$ per orbital period (7hr 39 min). These variations are sufficient to trigger a slow erosion process in high-slope regions that experience the largest changes in the tidal environment. Using direct numerical simulations of particle dynamics, we demonstrate this new mechanism for surface mobility on Phobos for the first time. In addition, we find that these regions of high surface mobility coincide with one of two distinct "color" units revealed by multispectral and hyperspectral imaging \cite{Fraeman2014}: the blue unit. The blue unit is characterized by a relatively high albedo and increasing visible to infrared continuum slope. In contrast, the red unit is defined by a slightly lower albedo and relatively steeper spectral slope. It has been unclear whether these two geologic units are compositionally different, or if the spectral differences are due to space weathering, as there is an observed lack of strong absorption features in spectral observations of both units\cite{Basilevsky2014, Pieters2014, Murchie2015}. Through a new eccentricity-driven mechanism for surface mobility, we develop a new model for regolith development on Phobos that can explain the relationship of the two color units. We show that this process can transform "red" space-weathered \cite{Hapke2001,Vernazza2009, Kaluna2016} regolith by exposing "blue" sub-surface material in a process analogous to the tidally-induced refreshing of asteroid surfaces \cite{Binzel2010}. Our finding provides new perspectives on the space weathering process for airless bodies in the Solar System. For example, since the proposed mechanism makes predictions on the rate at which fresh material can be exposed, it is possible to place new constraints on the space weathering timescales of Phobos by comparing current and future observations of its surface. Additionally, a relatively fast surface-refreshing process on Phobos has implications on the origin of the Martian moons. An accretion scenario (following a giant-impact or during the formation of Mars) or a capture of an inner solar system body implies the presence of mafic materials on Phobos \cite{Fraeman2014, Thomas2011}. Since our analysis suggests that the blue unit represents pristine endogenic material, the observed lack of mafic mineral absorptions in this unit \cite{Fraeman2012} raises challenges for a giant impact scenario. Most of these challenges are likely to remain unsolved until dedicated missions to Phobos and Deimos will visit these remote objects. The Mars Moons eXploration mission\cite{Kawakatsu2017} (MMX), which will visit Phobos and return samples from its surface, will conclusively determine the origin of this enigmatic body and shed light on the history of the Martian system. \section{Results} \label{S:2} We perform a two-step analysis to investigate the dynamical stability of grain particles on the surface of Phobos. In the first step, we model the gravitational attraction of the planetary satellite via a constant density polyhedron model \cite{WernerScheeres1996} and calculate the evolution of its dynamical slopes (Fig.\ 1a) in the framework of the elliptical restricted three body problem (ERTBP, \cite{Broucke1969}). Despite the tiny eccentricity of Phobos' orbit ($e \simeq 0.0151$), this model produces time-varying effects that change the local acceleration across the whole surface of Phobos. In particular, these fluctuations result in dynamical slope changes up to 2$^{\circ}$ per orbit that may trigger mass-wasting events near Stickney and the anti-Mars point (Fig.\ 1b). In the second step, we attempt to quantify the erosion rate in these regions. \begin{figure}[h] \label{fig:GravityVar} \centering\includegraphics[width=1\linewidth]{Fig1_SlopeChangeEccentricity.png} \caption{\textbf{Changes in the local surface slopes over 1 Phobos orbit.} \textbf{a,} Using a polyhedron shape model of Phobos, we can calculate the surface slopes at its current orbit. \textbf{b,} We evaluate the variations in the surface slope over 1 orbital period. Starting from periaerion, i.e., when the true anomaly, $\nu$, is equal to $0^\circ$, we find that the dynamical slopes of Phobos can vary by up to $\sim$ 2$^{\circ}$ over one orbital period. The largest variations are found in regions along the equator and, especially, in the Mars and anti-Mars directions when $\nu \simeq 180^\circ$.} \end{figure} In order to evaluate the magnitude of this effect on transforming Phobos' surface features, we consider the analytical theory of erosion for transport-limited downslope flow\cite{Culling1960}. For a planetary surface with loose regolith (with little to no cohesion) where the flow of grains is controlled by the transportation rate rather than the regolith supply or production rate, the flow rate, $q$, can be defined as a function of the local surface slope, $\theta)$, as \begin{equation}\label{Eq:qtheory} q = K \tan(\theta), \end{equation} where $K$ is the downslope flow constant in units of volume flux per unit time. In previous studies of regolith mobility on small bodies\cite{Richardson2014}, the value of $K$ was estimated for impact-induced disturbances using a Newmark slide-block model\cite{Richardson2005}, and the value of the flow rate was shown to be linear with the slope only when $\theta < 10^{\circ}$ circa \cite{Richardson2014}. For higher slopes, fast granular flows are better modeled as a non-linear function that takes the critical slope, $\theta_{\mathrm{C}}$, into consideration \cite{Roering1999}. Here, we perform local simulations of grain dynamics in order to estimate the value of $K$ for a variety of surface slopes and periodic fluctuations, i.e., $\delta \theta$. Using a soft-sphere discrete element code, \texttt{PKDGRAV}\cite{ Richardson2011,Schwartz2012}, we set up simulations of a 1.5 m by 1.2 m regolith bed with a porosity of 45\% settled in Phobos gravity ($\simeq 5$ mm/s$^2$) and made up of grains with radii between 1.4 and 1.6 cm. The material properties of the grains are similar to sand of medium hardness\cite{Jiang2015}. The simulated grains have a measured critical angle of repose of $34^{\circ}$. In order to allow the grains to flow freely in the event of an avalanche trigger or creep-induced motion, we impose periodic boundaries in the directions perpendicular to the initial local gravity vector. The simulation setup is illustrated in Fig.\ 2a. With the goal of determining a conservative estimate for expected volume flux across different regions of Phobos, we perform simulations for initial surface slopes between 5$^{\circ}$ and 32$^{\circ}$ with $\delta \theta \in [0.5^\circ, 1.5^\circ]$. The highest surface slope achieved across all of the simulations is 34.5$^{\circ}$. The output of the simulations display creeping motion of small grain displacements over the course of a Phobos orbital period. In particular, we find that a granular bed that initially starts at a high slope with significant slope changes will experience the most mass flow. However, even granular beds that are in a sub-critical state and are subject to relatively small variations can lead to surface mobility. We term this a \emph{cold flow} process as opposed to the ``fast" flow of an actual landslide. The cold flow process is illustrated in Fig.\ 2b with an example of local simulations for the case of a regolith bed that has a mean slope of 31 deg and $\delta \theta$ between $0.625^\circ$ and $1.5^{\circ}$. Here, we highlight the largest particle displacements (anything greater than 3 cm) by showing the initial (black) and final (yellow) positions of the particles along with their tracks (dotted red lines). For each simulation, we measure the total displacement of particles in order to calculate the volume flux across the simulation area over one orbit. For the phase space studied here, which consists of regolith resting at sub-critical angles, we calculate volume flux rates between 10$^{-5}$ to 4 particles/m$^2$/orbit. We show the results of all simulations in Fig.\ 2c. \begin{figure}[h] \label{fig:LocalSims} \centering\includegraphics[width=1\linewidth]{Fig2_LocalSims.jpg} \caption{\textbf{Local simulations of grain motion on the surface.} \textbf{a,} We set up our simulation with periodic boundary conditions in order to better measure the volume flux induced by Phobos’ eccentric orbit. Arrows show exaggerated variations in the surface slope that are determined by the dynamical calculations demonstrated in Fig.\ 1. \textbf{b,} Comparison of grain motion across three different simulations that have a mean slope of 31$^{\circ}$ and have variable slope change amplitudes as shown in the bottom right corner of each panel. In dark gray, we show the top-most layer of grains. We highlight the largest particle displacements ($> 3$ particle radii) by showing the initial position (black circle) and final position (yellow circle) of these particles, connected by their paths (dashed red lines). \textbf{c,} Aggregating our results across 47 simulations, we formulate a prescription for the expected volume flux $Q$ for a given gravity slope and its variation over 1 orbit, see equations (\ref{Eq:qtheory}) \& (\ref{Eq:Qdata}).} \end{figure} We find that across low and high slopes, the expected volume flux due to eccentricity-driven surface slope changes, $Q$, is best described by, \begin{equation}\label{Eq:Qdata} Q = K \tan(\theta) \frac{\tan(\delta \theta)}{\tan(\theta_{\mathrm{C}})}. \end{equation} Based on the simulations conducted in this study, we find that the magnitude of the particle flow can be described by two linear regimes with different values of the flow constant $K$. The cold flow prescription depends on the proximity of the initial slope to the critical angle of repose and the magnitude of the surface slope-change. We find that the value of $K$ switches from $5.7$ particles/m$^2$/orbit to $143.2$ particles/m$^2$/orbit when $Q/K = 7\times10^{-3}$, which corresponds to a value of $\delta \theta = 0.4^{\circ}$ when $\theta = \theta_{\mathrm{C}}$. Therefore, equation (\ref{Eq:Qdata}) provides a description for the cold-flow mechanism on the surface of Phobos induced by its eccentric orbit. We use this prescription to compare the expected erosion rates in different regions of Phobos with observational data and identify areas where the cold flow mechanism may be dominating the geomorphology of the Martian moon. Observations by the HiRISE camera on NASA's Mars Reconnaissance Orbiter (MRO) spacecraft and the High Resolution Stereo Camera (HRSC) on ESA's Mars Express spacecraft have produced detailed images showing the heterogeneity of the area in and around Stickney crater (Fig.\ 3a,d). While high albedo features are typically indicative of mass-wasting motion, the area east of Stickney has a relatively smooth surface\cite{Basilevsky2014}, free of streaks that typify granular flow on Phobos \cite{Shi2016}. There have been some suggestions that the color variation in this area is a result of a catastrophic landslide that began in Stickney’s western wall and transitioned to the sub-Martian point by motion of a thin surface layer of material in a long run-out landslide \cite{ShingarevaKuzmin2001, CollinsMelosh2003}. However, these studies either assume small and unrealistic values for the friction properties of the Phobos surface (friction coefficient of 0.01-0.09\cite{ShingarevaKuzmin2001}), or require the development of acoustic fluidization \cite{CollinsMelosh2003}. Another interpretation of this region is that it is asymmetric ejecta deposit from Stickney\cite{Thomas1998}. Here, we offer a novel interpretation. \begin{figure}[h] \label{fig:Maps} \centering\includegraphics[width=1\linewidth]{Fig3_SlopeChangeQuartet.jpg} \caption{\textbf{High-slope regions on Phobos that undergo large variations coincide with blue surface units.} \textbf{a,} Color composite image of Stickney crater derived from Mars Express HRSC data \cite{Patsyn2012}. \textbf{b,} Dynamical slopes in and around Stickney when $\nu = 0^\circ$. \textbf{c,} Highlighting regions of possible mass-wasting based on our estimate of $Q$. \textbf{d,} Image \texttt{HiRISE\_PSP\_007769\_9010\_IRB} synthesized from three infrared and visible channels showing Stickney crater, highlighting the blue regions east of the crater.} \end{figure} By analyzing the spatial distribution of the surface slopes in the region east of Stickney (Fig.\ 3b) and the predicted variation due to Phobos' eccentric orbit, we predict a relatively large amount of regolith flow compared to nearby regions (Fig.\ 3c). This area, labeled A in Fig.\ 3a, should have the most mobile regolith due to the unique combination of steep terrain and high surface-slope changes. In contrast, the relatively steep southern wall of Stickney (labeled B) does not experience large variations in the tidal environment of Mars and--in agreement with our model--shows little indication of blue material (Fig.\ 3d). Through this qualitative comparison of Phobos color data with our predicted volume flux, we propose that blue geologic units are composed of sub-surface material that is being continually exposed by the eccentricity-driven librations of Phobos. This suggests that the blue units are the footprints of ``fresh", pristine material, as opposed to the ``old", space weathered regolith that is found in the red unit of the Martian moon. To evaluate this possibility, we compare our model's predicted excavation rate to the likely space weathering rate at Phobos. The results of this analysis are discussed in the next section. \section{Discussion} \label{S:3} In order to determine the efficiency of this process in uncovering fresh material, we consider the possible timescale and depth of space weathering on Phobos. Since we have a poor understanding of Phobos' composition \cite{Murchie2015} and the specific mechanisms that may influence the maturation of its regolith \cite{PietersNoble2016}, we consider a range of space weathering rates to provide some perspective on the possible surface refreshing rate. We use estimates for space-weathering based on previous spectral analysis of main-belt asteroids. Studies on space weathering of S-complex asteroid dynamical families \cite{Vernazza2009, Binzel2010} reveal that space-weathering rate for these bodies is relatively rapid ($< 10^6$ years), suggesting that solar wind ion implantation dominates the process. In contrast, analysis of C-complex asteroids \cite{Kaluna2016} suggests that space weathering is a long term process, with spectral slopes being modified on timescales closer to $10^9$ years. Using the proposed space-weathering rates \cite{Kaluna2016} and slope changes in CRISM reflectance measurements between blue and red geologic units \cite{Fraeman2012}, we estimate that space weathering on Phobos can occur on timescales with a lower limit of 10$^6$ years. Due to the lack of in-situ exploration of small-body subsurface properties, we rely on analysis of lunar core samples from the Apollo missions to obtain an estimate on the depth of space-weathered material on Phobos. Measurements of the regolith maturity parameter ($I_s/FeO$ ratio) of 12 Apollo core samples show that the regolith typically transitions between mature weathered material to relatively fresh material around 20-50 cm below the surface \cite{Lucey2006}. Combining this estimate of the depth of weathered material with our assumptions on the space-weathering rate and the predicted excavation rates derived from our numerical simulations, we analyze the efficiency of the eccentricity-driven cold-flow process in exposing fresh material. We find that, while the calculated excavation rates are relatively small (only a few grains typically move from a 1 m$^2$ high-slope region every Phobos orbit), the accumulated effect over the course of Phobos' geologic time can be quite substantial. In Fig.\ 4, we show the excavation rates for the cold-flow of mm-size particles required to expose material at certain depths (represented by the solid curves) for different timescales. It is important to note that we make a steady-state assumption for regolith as it becomes exposed. In reality, the regolith can mature as it is uncovered. Nevertheless, this analysis provides an order of magnitude estimate of the necessary excavation rates required to expose fresh material. From our numerical simulations of the cold-flow mechanism, we show that the simulation-derived mean excavation rates (dashed lines) for the labeled regions in Fig.\ 3a. The values of $\theta$ and $\delta \theta$ for these regions are determined based on our dynamical calculation (Fig.\ 1), and are fed into the prescription of mass flux, equation (\ref{Eq:Qdata}). We find that the regions with the highest spatial distribution of blue material in the HRSC data (region A) has the highest mean excavation rate, followed by the western wall of Stickney (region D), and both the southern wall (region B) and the crater floor (region C) have very low predicted excavation rates. Overall, we find that blue regions represent areas where grain-mobility is high enough to expose fresh material down to depths of $>$ 20 cm at a rate faster than the expected space weathering time-scale (grayed-out region). Red regions have predicted mean excavation rates that may only expose material a few cm deep before $10^6$ years, which is the age where our steady-state assumption may fail, as some asteroids\cite{Binzel2010} are found to rapidly redden in this time-scale. Therefore, we conclude that eccentricity-driven mass-wasting is an efficient process for uncovering pristine un-weathered material on Phobos, represented by blue geologic units found near Phobos’ sub-Mars point. We propose a model for regolith development where i) Phobos is inherently blue, ii) space weathering reddens the surface material, iii) impacts and Phobos' decaying orbit can expose fresh or weathered material depending on crater depth and the characteristic space weathering depth $d_{\mathrm{SW}}$, iv) this freshly exposed material is reddened over time, and v) blue regions are continuously refreshed by mass-wasting that occurs in high-slope regions that experience significant variations in tidal forces due to Phobos’ eccentric orbit. \begin{figure}[h] \label{fig:EXR} \centering\includegraphics[width=1\linewidth]{Fig4_ExcavationWeathering.png} \caption{\textbf{Required volume fluxes necessary to uncover subsurface material before a certain weathering timescale.} We show the excavation rate required to uncover fresh material buried at different depths ($d_{\mathrm{SW}}$) ranging from 1 to 100 cm that requires a corresponding number of years (Space Weathering Age) to mature. This may be interpreted as the number of mm-size particles, $N_{mm}$, in the top layer of regolith that need to flow out of a 1 m$^2$ region every Phobos orbit (vertical axis) in order to excavate un-weathered material, which takes a given number of years to weather, $t_{\mathrm{SW}}$, down to a depth, $d_{\mathrm{SW}}$, described by the solid curves. The gray area represents the estimated timescale for material on Phobos to mature based on analysis of spectral slope reddening in CRISM data. The dashed lines represent the mean excavation rates of the regions highlighted in Fig.\ 3a: region A (cyan), region B (magenta), region C (yellow), region D (black). This demonstrates the agreement between blue regions in the HRSC color data and the efficiency of the eccentricity-driven excavation process on Phobos.} \end{figure} \section{Perspectives} \label{S:4} We have shown that Phobos' small eccentricity can drive changes in the surface slope that mobilize grains on the top-layers of high-slope regions. While the rate of mass flow is slow, the accumulated effect over time is enough to expose fresh subsurface material, giving a plausible explanation for the dichotomy between red and blue geologic units on the surface. The analysis described here relies on a simplistic model of space weathering based on other airless bodies, and does not take into account Phobos' unique environment, which is not well understood. However, since the predicted surface refreshing is tied to Phobos' orbital motion, we can place new constraints on the space weathering timescale for Phobos itself, which may then allow us to better understand this process for other airless bodies in our Solar System. With a better understanding of the surface and subsurface conditions on the Martian moon, the Phobos surface can become a ``Rosetta-stone'' for understanding space weathering elsewhere in the solar system. MMX\cite{Kawakatsu2017} (led by JAXA, with contributing instruments from CNES and NASA) will explore Phobos and obtain a core sample from its surface at a depth of up to 10 cm. The combination of a returned sample, new remote sensing observations of the Phobos surface, and a better understanding of surface-refreshing process on Phobos will revolutionize our understanding of the interaction between the surface of an airless body and the space environment. Finally, one of MMX's main goals is to provide an answer to the question of Phobos' enigmatic origin. It is unclear whether Phobos is a captured Solar System body \cite{Burns1978,Hansen2018}, or if it formed in an accretion disk following a giant impact with Mars \cite{Craddock1994,Hyodo2017}. Current theories of a giant impact or in-situ formation origin predict that Phobos should be made up of a relatively large proportion of Martian material and imply the presence of mafic minerals such as olivine and pyroxene. The observed lack of diagnostic absorption features in spectral data suggests that either: a) all of Phobos' surface is heavily space weathered, removing spectral features, b) the surface grains are extremely fine, or c) the regolith is substantially mixed with opaque materials such as carbon. Our analysis suggests that the blue units on Phobos represent pristine endogenic material. Therefore, the lack of mafic absorption features in the spectral data of blue units \cite{Fraeman2012}, which are relatively featureless, would suggest that the first scenario (a) is unlikely, and that it is possible that Phobos has a sparse abundance of Martian material. Therefore, we find that this cold-flow mechanism which can refresh the Phobos surface indicates that Phobos is not spectrally like Mars; or, the observed lack of mafic absorptions must be explained in some other scenario, such as those listed above. \input{methods} \input{references} \input{acknowledgments_contributions} \end{document} \section*{Methods} \subsection*{The dynamical environment on Phobos' surface} A point mass in the vicinity of Phobos is subject to the gravitational attraction of both the planetary satellite itself and Mars. Accordingly, the acceleration felt by particle grains on the surface of the Martian moon can be approximated by the equations of the ERTBP \cite{Broucke1969}. Since the distance between the grain and the barycenter of Phobos is much smaller than the distance between the grain and the barycenter of Mars, the equations of motion can be further simplified by expanding the gravitational term of Mars in a Taylor series up to the first order. The resulting system of equations of motion is known as the Hill approximation of the ERTBP and is better understood in a pulsating rotating reference frame centered on Phobos and such that $\versor{x}$ is constantly align with an imaginary line connecting the two primaries, $\versor{z}$ is parallel to the orbital angular momentum of the moon, and $\versor{y} = \versor{z} \times \versor{z}$. Following the length and time scaling described in \cite{Scheeres2002}, the equations read as \begin{equation} \left\{ \begin{array}{ccl} x'' & = & \dfrac{1}{1+e\,\cos{\nu}}\,(g_x + 3\,x) + 2\,y',\\ y'' & = & \dfrac{1}{1+e\,\cos{\nu}}\,g_y - 2\,x',\\ z'' & = & \dfrac{1}{1+e\,\cos{\nu}}\,g_z - z,\end{array} \right. \end{equation} where $(\cdot)'$ denotes differentiation with respect to $\nu$, i.e., the true anomaly of Phobos, $X = \begin{bmatrix} x, y, z, x', y', z'\end{bmatrix}^T$ represents the state of the mass particle, $e = 0.0151$ is the eccentricity of Phobos, and $\vect{g} = \begin{bmatrix} g_x, g_y, g_z\end{bmatrix}^T$ is the gravitational attraction of the Martian moon. The acceleration is obtained in closed form via a constant density polyhedron model consisting of $40962$ vertices, $81920$ facets, and density equal to $1860$ km/m$^3$\cite{Werner1996, Willner2014, Murchie2015}. In this paper, we are interested in monitoring the variation of the surface slopes over one orbital period of Phobos around Mars. Accordingly, let $\versor{n}_i$ and $\vect{r}_i$ denote the normal vector and barycenter position vector of the i-th facet, respectively. The surface slope is obtained as the supplement of the angle between $\versor{n}_i$ and the acceleration vector $\vect{a}_i = \begin{bmatrix} x’’, y’’ , z’’\end{bmatrix}^T$ evaluated in $\vect{X}_i = \begin{bmatrix} x_i, y_i, z_i, 0, 0, 0\end{bmatrix}^T$ at $\nu = \nu^*$, i.e., \begin{equation} \theta(\nu^*) = \pi - \arccos{[(\versor{n}_i \cdot \vect{a}_i) / \| \vect{a}_i \| ]}. \end{equation} While Phobos rotates at a constant rate of $\omega \simeq 3.13$ rev/day, the pulsating reference frame rotates about the $\versor{z}$-axis with a time-changing angular velocity of $\dot{\nu} = \omega\,\dfrac{(1 + e\,\cos{\nu})^2}{(1-e^2)^{3/2}}$. Consequently, as seen from the rotating reference frame, the polyhedral shape of Phobos describes librations of approximately $\pm 1$ deg about the $\versor{x}$ axis. These librations cause periodic changes in the tidal acceleration due to Mars that may be resposible for triggering the cold flow mechanism in high-surface slopes regions of Phobos. \subsection*{Granular Dynamics Simulation} For the local simulations of granular flow, we used the \textit{N}-body code \texttt{PKDGRAV} \cite{Richardson2011} which is capable of accurately simulating the complexity of grain-grain and grain-boundary interactions through a soft-sphere discrete element method \cite{Schwartz2012} (SSDEM). In SSDEM, collisions of spherical grains are resolved by allowing them to slightly overlap (typically $< 1$\% of their radii) and then applying multi-contact and multi-frictional forces, including static, rolling and twisting friction. Modeling grain friction accurately is a critical component for high-fidelity granular physics simulations. For \texttt{PKDGRAV}, a new rolling friction model has recently been implemented \cite{Zhang2017} that allows for a more accurate modeling of grain shape and angularity, through a shape parameter $\beta$. The mechanical properties of Phobos' surface are poorly known; however, the range of plausible material types are limited, and some remote sensing observations can help constrain some important parameters that are required for accurate simulations. Radar albedo measurements \cite{Busch2007} provide an estimate of the near surface porosity ($40 \pm 10$\%). Local slopes on the surface of Phobos can be as high as 45$^{\circ}$; however, these are likely regions that are in a super-critical state and may not accurately reflect the material's true angle of repose. We chose a set of friction parameters for the grains such that they emulate the static behavior of sand of medium hardness as determined by laboratory experiments\cite{Jiang2015}, which have a corresponding angle of friction of 34$^{\circ}$. The grain size of Phobos regolith is highly uncertain. Early measurements from the Viking spacecraft suggested that the typical grain size of Phobos regolith is 50-100 $\mu$m\cite{Sasaki1995}. More recent work that used thermal inertia measurements combined with a thermal conductivity model for regolith suggests a larger average grain size of 1.1 mm\cite{GundlachBlum2013} ; however, there are large uncertainties ($+0.9/-0.7$ mm). We performed a few high-resolution simulations with grains that have radii of $R_{\mathrm{grain}}$ = 0.9-1.1 mm; however, the combination of particle number (50,000 particles), simulation time (7 h 39.2 m), and small timestep (0.05 s) make the computation very expensive. Therefore, to speed up the calculation, we used a few high-resolution simulations as a baseline, and the majority of our simulations were lower-resolution cases ($R_{\mathrm{grain}}$ = 1.4-1.6 cm, N=7,000). This was necessary to sweep a larger parameter space in starting slope, $\theta$, and slope change amplitude, $\delta \theta$. Table 1 summarizes the properties of grains and regolith bed for our simulations. \begin{table}[h] \centering \begin{tabular}{l l l} \hline \textbf{Property} & \textbf{Low Resolution} & \textbf{High Resolution}\\ \hline $R_{\mathrm{grain}}$ & 1.4-1.6 cm & 0.09-0.11 cm \\ $\phi$ & 34$^{\circ}$ & 34$^{\circ}$ \\ $\rho_{\mathrm{grain}}$ & 3.1 g cm$^{-3}$ & 3.1 g cm$^{-3}$ \\ $\rho_{\mathrm{bulk}}$ & 1.7 g cm$^{-3}$ & 1.7 g cm$^{-3}$ \\ $N$ & 7,000 & 50,000 \\ Sim. Area & 1.5 $\times$ 1.2 m & 0.60 $\times$ 0.4 m \\ \hline \end{tabular} \caption{Properties of simulated regolith for granular dynamics simulations. $R_{\mathrm{grain}}$: grain radius, $\phi$: critical angle, $\rho_{\mathrm{grain}}$: grain density, $\rho_{\mathrm{bulk}}$: grain bed bulk density, $N$: number of particles, Sim. Area: surface area of grain bed. } \end{table} Since we are interested in measuring the possibility of surface grain motion due to these small variations in the surface slope, and not complete hillslope-failure, we restrict the vertical scale of our simulations to 5 grain diameters. We set up simulations with initial slopes ranging from 5$^{\circ}$ to 32$^{\circ}$ and amplitudes of variation between 0.5$^{\circ}$ and 1.5$^{\circ}$. This is done by performing an initial setup simulation where the direction of the gravity vector is changed till the desired gravity slope is achieved. The change in the gravity vector is done gradually (over 1 hr in simulated time) as a precaution so as to prevent granular flow from triggering before the time-varying slope change is simulated. The grains are then allowed to settle till all small motions are dissipated. Only after this setup is complete do we perform a full simulation, varying the gravity slope of the inclined granular bed over 1 Phobos orbit (7hr 39 minutes). For the limited number of high-resolution cases, we found that, for the same $\theta$ and $\delta \theta$, the particle flux (particle/m$^2$/orbit) was similar to low resolution cases. Therefore, we expect the volume flux (m$^3$/m$^2$/orbit) to depend on the actual grain size on the surface of Phobos. For the same period of time, a surface made up of small particles will excavate material at a given depth at a slower rate than a surface with larger particles. Future work will focus on testing this finding in order to develop scaling laws for the volume flux as a function of grain size. While the grains we simulate here are cohesion-less, the actual grains on Phobos are likely to have some non-zero cohesion that will influence the grain dynamics \cite{Scheeres2010}. Cohesive bonding between grains may act as an inhibitor to particle motion, increasing an individual particle's shear strength; however, it is not obvious how grains may actually bond in the Phobos environment. In particular, a size distribution of grains may lead to small grains being bonded to larger ones, and being carried away by the eccentricity-driven mass-wasting process, enhancing the expected excavation rates. Furthermore, as the sizes of the grains become smaller, the relative importance of cohesion is magnified, and processes such as grain lofting\cite{HartzellScheeres2011}, driven by charging from the solar wind or from passing coronal mass ejections\cite{Farrel2017}, may be present on Phobos' surface. Therefore, the contribution of cohesion to this process may be non-trivial, and we will investigate its influence in the future with a recently implemented cohesion-model in \texttt{PKDGRAV}\cite{Zhang2018}. \subsection*{Comparing excavation rates to space weathering rates} We calculate the necessary volume flux required to uncover fresh material (Fig.\ 4) by equating an excavation rate with a space weathering rate to a certain depth. We define the excavation rate, $EXR$, to be the volume, $V$, out of an area, $A$, for a given time period, $\delta t$, \begin{equation}\label{Eq:EXR} EXR = \frac{V}{A \delta t}, \end{equation} and we define the space weathering rate, $SWR$, as the timescale, $t_{\mathrm{SW}}$ necessary to mature regolith up to a depth of $d_{\mathrm{SW}}$. \begin{equation}\label{Eq:SWR} SWR = d_{\mathrm{SW}}/t_{\mathrm{SW}} \end{equation} We calculate the equilibrium values for excavation and weathering to different depths by equating equations (\ref{Eq:EXR}) and (\ref{Eq:SWR}), setting the area to units of 1m$^2$, $\delta t$ to one Phobos orbit, and the grain radius, $R_{\mathrm{grain}}$ to the volume of a mm-size particle, such that the number of particles that must flow out of a 1m$^2$ region in 1 Phobos orbital period, $T_{\mathrm{Phobos}}$, to excavate that region to a depth of $d_{\mathrm{SW}}$ in $t_{\mathrm{SW}}$ years is given by: \begin{equation}\label{Eq:QRAte} N_{mm} = \frac{d_{\mathrm{SW}}}{t_{\mathrm{SW}}} \frac{1\mathrm{m}^2}{A} \frac{\delta t}{T_{\mathrm{Phobos}}} \left(\frac{10^{-3}\mathrm{m}}{R_{\mathrm{grain}}}\right)^3 \end{equation} \subsection*{Space weathering timescales for Phobos from spectral slopes} We derive spectral slopes from CRISM spectra of representative blue and red regions on Phobos corrected to I/F with incidence and phase angles equal to 30 degrees and emission angle of 0 degrees \cite{Fraeman2012}. We measure spectral slopes following \cite{Kaluna2016} using the normalized reflectivity gradient S' defined by \cite{JewittMeech1986}: \begin{equation} S'(\lambda_1,\lambda_2) = \dfrac{dS/d\lambda}{S_{0.55}}\, \end{equation} where $dS/d\lambda$ is the slope of the reflectance between wavelengths $\lambda_1 = 0.49 \mu$m and $\lambda_2 = 0.91 \mu$m and slopes are normalized to reflectance at $0.55 \mu$m. We find $S'_{red}=0.69\%\mu$m$^{-1}$ and $S'_{blue}=0.29\%\mu$m$^{-1}$, for an overall slope change from blue to red of $0.4\%\mu$m$^{-1}$. Slope change rates due to space weathering have been estimated for several asteroid families of different compositions: \cite{Kaluna2016} observed a slope change of $0.08\%\mu$m$^{-1}$ over 2.3 Gyr for C-complex asteroids, while \cite{Lazzarin2006} determined a rate of $8.8\times 10^{-5}$AU$^{2}\%\mu$m$^{-1}$Myr$^{-1}$ for C-complex and $24.9\times 10^{-5}$AU$^{2}\%\mu$m$^{-1}$Myr$^{-1}$ for S-complex asteroids. \cite{Vernazza2009} determined that asteroids may weather relatively quickly, with a slope change of $0.4\%\mu$m$^{-1}$ on the order of a million years. Although the space weathering environment may be very different at Phobos than for asteroids, and Phobos’ composition is not well constrained, these space weathering rates can be used to obtain an order of magnitude of the effect expected for Phobos. Using these slope change rates due to space weathering of asteroid families and the slope difference of $0.4\%\mu$m$^{-1}$ between blue and red regions, we estimate a space weathering age for Phobos’ red regions on the order of $10^{6}-10^{10}$ yr. While the upper time range based on slow C-type asteroid reddening is physically infeasible, it is the lower time limit that is relevant to our study, in order to determine whether the cold flow mechanism is efficient enough to compete with rapid space weathering. \section*{References} \section*{Tables} \begin{table}[h] \centering \begin{tabular}{l l l} \hline \textbf{Property} & \textbf{Low Resolution} & \textbf{High Resolution}\\ \hline $R_{\mathrm{grain}}$ & 1.4-1.6 cm & 0.09-0.11 cm \\ $\phi$ & 32$^{\circ}$ & 32$^{\circ}$ \\ $\rho_{\mathrm{grain}}$ & 3.1 g cm$^{-3}$ & 3.1 g cm$^{-3}$ \\ $\rho_{\mathrm{bulk}}$ & 1.7 g cm$^{-3}$ & 1.7 g cm$^{-3}$ \\ $N$ & 7,000 & 50,000 \\ Sim. Area & 1.5 $\times$ 1.2 m & 0.60 $\times$ 0.4 m \\ \hline \end{tabular} \caption{Properties of simulated regolith for granular dynamics simulations. $R_{\mathrm{grain}}$: grain radius, $\phi$: critical angle, $\rho_{\mathrm{grain}}$: grain density, $\rho_{\mathrm{bulk}}$: grain bed bulk density, $N$: number of particles, Sim. Area: surface area of grain bed. } \end{table}

d0276b295e11d94f0ec1aa77647932fa378e2b54

\section{Introduction} The device independent approach in quantum information theory allows one to infer physical properties of systems and to implement protocols based solely on the observed statistics, i.e.\ without making any assumption on the underlying states and how they interact with the measurement apparatuses. Different tasks that can be implemented in this way include quantum key distribution, randomness generation and amplification, genuine multipartite entanglement certification and self-testing of states and measurements \cite{reviewnl}. It has been also observed \cite{bellwit,gallego} that the underlying dimension of an uncharacterized (classical or quantum) physical system can be tested in this way, i.e.\ using only the observed probabilities of obtaining certain outcomes conditioned on implementing different uncharacterized measurements. These procedures are referred to as device-independent dimension witnessing (DIDW) and are the object of active current investigation. On the one hand, from a foundational perspective, DIDW allows to estimate the degrees of freedom of a system without a priori including this information in the physical model. On the other hand, from the point of view of applications, quantum information tasks can be more efficiently implemented the larger the dimensions of the quantum systems one can prepare and control. Thus, dimension is regarded as a valuable resource in this context and DIDW provides experimental means to test it. In fact, DIDW can be regarded as a primitive for semi-device-independent protocols, which make no assumption on the inner functioning of devices and physical systems except for bounds on the underlying dimension \cite{semi}. In general, DIDW is deeply rooted in the field of quantum communication complexity, which studies the necessary amount of communication different parties must exchange (as measured by the dimensionality of the physical systems being sent) in order to implement distributed computations \cite{reviewcc}. DIDW was introduced in the Bell scenario in which two parties share an entangled state \cite{bellwit}. Soon after, Ref.\ \cite{gallego} presented an alternative scenario, the so-called prepare-and-measure, which consists of two devices: one that prepares and sends states and one that measures them. This setting is simpler in the sense that it does not require entanglement nor multicomponent systems and this proposal has been already verified in experiments \cite{experiments}. Reference \cite{gallego} analyzed the mathematical structure of the set of possible behaviours to be observed in this scenario depending on the dimension and provided explicit constructions of functionals acting there, known as dimension witnesses, whose values provide lower bounds on the classical and quantum dimension. Although other constructions of dimension witnesses have appeared in subsequent works \cite{dallarno,didwpm,didwseveral}, this approach suffers from two difficulties. First, these functionals are usually tailor-made to detect specific behaviours one targets at. Second, the explicit corresponding bounds are very case-dependent and difficult to find in general. Thus, general bounds to constrain the dimension of arbitrary quantum behaviours are of great use in this context. In this sense, it is particularly worth mentioning the work of Ref.\ \cite{num}, which offers a powerful and general numerical approach based on semidefinite programming that makes it possible to obtain such bounds for a given dimension witness. However, this approach is bound to problems that can be tackled numerically. Analytical results in this direction not only enable a better understanding of the mathematical structure of dimension-constrained behaviours but also to obtain results for problems that go beyond computational efficiency such as the asymptotic scenario, to devise new dimension witnesses or to consider nonlinear constraints. In this sense, general lower bounds on the dimension can be found in \cite{brunner,sikora,yo}. Notwithstanding, all these analytical approaches assume that the preparer and measurer devices are not correlated, i.e.\ that they do not share a random variable. Although this is justified in certain scenarios, there are others in which this assumption is not admissible. This is particularly the case when the devices are not trusted. Suppose, for instance, that the parties want to verify that devices provided by a manufacturer, which are regarded by them as black boxes, operate on quantum systems of a given dimension. Then, malicious providers could fake higher-dimensional behaviours by mixing lower-dimensional preparations using shared randomness if this is not taken into account. Another example to consider this scenario is when the parties have to use their devices to implement a particular task and the constraints of the problem allow them to use shared randomness as a resource. In fact, it is known that the availability or not of shared randomness can have drastic consequences in what comes to the necessary dimension underlying a given observed behaviour; for instance, without this resource at disposal almost all behaviours are high-dimensional while, when it is given, low-dimension behaviours are no longer negligible \cite{yo}. The main goal of the present work is to close this gap and provide general analytical conditions to test the dimension of quantum behaviours in the prepare-and-measure scenario even when devices might share randomness. After presenting our notation and definitions in Sec.\ II, in Sec.\ III we provide a general lower bound on the dimension of a quantum behaviour based on simple matrix analysis techniques, which is valid when preparations can be mixed either because shared randomness is available as a resource or because it is not safe to assume that the devices are uncorrelated. To further illustrate its usefulness, we provide two applications of this result in Sec.\ IV: we obtain bounds on the efficiency of a communication-complexity protocol known as random access coding and we provide improved constructions of dimension witnesses. We finish in Sec.\ V with some concluding remarks. \section{Notation and definitions} The prepare-and-measure scenario for DIDW \cite{gallego,dallarno} is composed by two parties: the preparer, Alice (or A), and the measurer, Bob (or B). A and B receive respectively inputs $x$ and $y$ from finite alphabets $\x$ and $\y$. They can only communicate by A sending a classical or quantum physical system to B depending on her input $x$. B can then perform a measurement on the system he receives depending on his input $y$. Using the outcome of this measurement together with all previous information held by him, B produces then an output $b$, which also takes values from a finite alphabet $\b$. Moreover, as explained in the introduction, we assume that the devices held by A and B may be correlated. That is, both parties have access to a common random variable, whose value determines the strategy to be followed from a pre-established list available to them. The main object in this scenario is the conditional probabilities with which each output occurs for any given pair of inputs: $P(b|xy)$. We will refer to this object as behaviour and it will be denoted by $\P$. Behaviours are lists of real numbers characterized by $P(b|xy)\geq0$ $\forall b,x,y$ and $\sum_bP(b|xy)=1$ $\forall x,y$ due to the fact that they are a collection of conditional probability distributions. Suppose now that an observer can monitor sufficient repetitions of this process so as to infer the corresponding behaviour but has no information about the systems sent by Alice and the measurements implemented by Bob nor about any details of the strategy the parties use to determine the output corresponding to the different possible inputs. The task DIDW aims at is to determine the minimal amount of classical or quantum communication (as quantified by the dimension of the systems sent from A to B) that is compatible with the observed behaviour. In order to give a rigorous definition of this quantity we distinguish between the cases in which A sends classical or quantum states. In the first case, A will send a message $m(x)\in[d]=\{1,\ldots,d\}$, and the number of dits $d$ necessary to construct it quantifies the amount of classical communication. The availability of shared randomness boils down to the fact that the parties can prepare any convex combination of strategies using messages of dimension less than or equal to $d$. We denote the set of all such behaviours by $\c_d$ (this set and the analogous for the quantum case to be defined below depend on $|\x|$, $|\y|$ and $|\b|$, but we do not make this explicit in order to ease the notation as these quantities should be in general clear from the context). Notice that the availability of shared randomness imposes that the set $\c_d$ is convex. Furthermore, it can be seen that this set is actually a convex polytope \cite{gallego,dallarno}: it is the convex hull of a finite number of behaviours $\{P^D_i\}$, which we call deterministic. These have the structure \begin{equation}\label{setc} P^D(b|xy)=\sum_{m=1}^{d}s(m|x)t(b|my), \end{equation} where $s(m|x)$ codifies the conditional probability with which A sends the message $m$ given $x$, and $t(b|ym)$ the conditional probability with which B outputs $b$ given $y$ and the reception of $m$. The deterministic condition amounts to the fact that $s(m|x)=\delta_{m,f(x)}$ and $t(b|my)=\delta_{b,g(m,y)}$ with arbitrary functions $f:\x\to[d]$ and $g:[d]\times\y\to\b$. Considering all possible choices for these functions gives rise to the finite list $\{P^D_i\}$. In the quantum case A sends quantum states $\rho_x$. The dimension of her message is thus \begin{equation}\label{quantumdim} d=\dim \sum_x\supp\rho_x, \end{equation} where $\supp$ stands for the support of an operator. In order to produce his output, B interacts with the state he receives by choosing an arbitrary quantum measurement conditioned on his input. Thus, the set $\q_d$ of behaviours achievable by sending quantum states of dimension at most $d$ is given by the convex hull of all behaviours $\P$ that take this form: there exists measurements $\{\Pi_b^y\geq0\}$ with $\sum_b\Pi_b^y=\one$ $\forall y$, such that \begin{equation}\label{setq} P(b|xy)=\tr(\rho_x\Pi_b^y) \end{equation} where the $\{\rho_x\}$ are of dimension less than or equal to $d$ as given by Eq.\ (\ref{quantumdim}). One can readily find that $\c_{|\x|}=\q_{|\x|}$, which constitute the set of all behaviours in a given setting. This is because if $d=|\x|$, A can transmit to B the value of her input through her message. Therefore, given any observed behaviour $\P$ there always exist minimal values of $1\leq d\leq|\x|$ and $1\leq d'\leq|\x|$ such that $\P\in\c_d$ and $\P\in\q_{d'}$ (notice that in general $d'\leq d$ since it is straightforward to see that for any fixed value of $d$, it holds that $\c_d\subseteq\q_d$). To determine them is precisely the goal of DIDW. The fact that the sets $\c_d$ are polytopes provides techniques to bound the classical dimension necessary to observe a given behaviour \cite{gallego}. However, the quantum case is much harder to deal with. In the following we provide such a bound in terms of a simple function of the behaviour. \section{Main result} We will arrange the array of numbers given by $\P$ into a matrix $P\in\mathbb{R}^{|\x|\times|\y||\b|}$ according to the rule \begin{equation}\label{behavior} P=\sum_{bxy}P(b|xy)|x\rangle\langle yb|, \end{equation} where in the standard notation of quantum mechanics $|yb\rangle=|y\rangle\otimes|b\rangle$ and $\{|y\rangle\}$ denotes the computational basis of $\mathbb{R}^{|\y|}$ and similarly for the other alphabet elements. We will consider different Schatten norms for matrices: \begin{equation}\label{schatten} ||A||_p=\left(\sum_i\sigma_i^p(A)\right)^{1/p}\quad(1\leq p\leq\infty), \end{equation} where $\{\sigma_i(A)\}$ are the singular values of the matrix $A$. Finally, we will denote the standard Hilbert-Schmidt inner product of matrices by \begin{equation} \langle P,G\rangle=\tr(PG^T)=\sum_{bxy}P(b|xy)G(b|xy), \end{equation} where for an arbitrary collection of $|\x||\y||\b|$ real numbers $G(b|xy)$, we define the matrix $G$ following the same prescription as in Eq.\ (\ref{behavior}). \begin{theorem} In any prepare-and-measure scenario $(|\x|,|\y|,|\b|)$, if $\P\in\q_d$ then \begin{equation} d\geq\frac{||P||_1^2}{|\x||\y|}. \end{equation} \end{theorem} \begin{proof} Due to the triangle inequality, the maximal value of $||P||_1$ in $\q_d$ must correspond to behaviours of the form given by Eq.\ (\ref{setq}) and $\rho_x\in\mathbb{C}^{d\times d}$ $\forall x$, i.e.\ shared randomness can be ignored. Defining the matrix \begin{equation} Z=\sum_{bxy}|x\rangle\langle yb|\otimes\rho_x\Pi_b^y\in\mathbb{R}^{|\x|\times|\y||\b|}\otimes\mathbb{C}^{d\times d}, \end{equation} we have that $P=\tr_{\mathbb{C}^{d\times d}} Z$. Thus, since the trace norm cannot increase by partial tracing, one arrives at \begin{equation} ||P||_1\leq||Z||_1\leq\left|\left|\sum_x|x\rangle\otimes\rho_x\right|\right|_2\left|\left|\sum_{by}\langle yb|\otimes\Pi_b^y\right|\right|_2, \end{equation} where in the last step we have used a particular case of H\"{o}lder's inequality for Schatten norms (see e.g.\ \cite{bhatia}). The result follows by noticing that \begin{equation} \left|\left|\sum_x|x\rangle\otimes\rho_x\right|\right|_2\leq\sum_x\left|\left||x\rangle\otimes\rho_x\right|\right|_2=\sum_x\tr(\rho_x^2)\leq|\x|, \end{equation} where we have used that $\tr(\rho_x^2)\leq1$ $\forall x$, and \begin{align} \left|\left|\sum_{by}\langle yb|\otimes\Pi_b^y\right|\right|_2&\leq\sum_{by}\left|\left|\langle yb|\otimes\Pi_b^y\right|\right|_2=\sum_{by}\tr[(\Pi_b^y)^2]\nonumber\\ &\leq\sum_{by}\tr(\Pi_b^y)=d|\y|, \end{align} where we have used that $0\leq\Pi_b^y\leq\one$ $\forall b,y$. \end{proof} \begin{theorem} In any prepare-and-measure scenario $(|\x|,|\y|,|\b|)$ and for every matrix $G\in\mathbb{R}^{|\x|\times|\y||\b|}$, if $\P\in\q_d$ then \begin{equation} \langle P,G\rangle\leq||G||_\infty\sqrt{d|\x||\y|}. \end{equation} \end{theorem} \begin{proof} This follows readily from Theorem 1 by another particular case of H\"older's inequality for Schatten norms ($\langle A,B\rangle\leq||A||_\infty||B||_1$). \end{proof} It should be noticed that Theorems 1 and 2 are equivalent since the former can also be deduced from the latter by noticing that $||P||_1=\max_U\tr(PU)$ where the maximization is over all partial isometries in $\mathbb{R}^{|\y||\b|\times|\x|}$ (and, hence, $||U||_\infty=||U^T||_\infty=1$) \cite{HJ2}. Theorem 1 provides a directly checkable condition that allows to lower bound the quantum dimension necessary to observe any given behaviour without the need of a clever choice for the matrix $G$. However, any linear functional acting on the set of behaviours takes the form $\langle P,G\rangle$ for some $G$ and, thus, Theorem 2 is also of interest since, among other applications that we shall discuss in the next section, it provides upper bounds within $\q_d$ for the so-called dimension witnesses. Before considering these applications, let us first discuss the attainability of the bound given in Theorem 1. It turns out that the bound cannot be improved in general since there exist scenarios in which it is sharp $\forall d$. In particular, it suffices to consider deterministic behaviours in $\c_d$. For this, let $|\x|=dn$, $|\y|=m$ and $|\b|=d$ for any $d,m,n\in\mathbb{N}$ and let us introduce the notation $\textbf{1}_n$ and $\textbf{0}_n$ for the vectors in $\mathbb{R}^n$ that have all entries equal to 1 and 0 respectively and $e_i^{(n)}$ for the vector of $\mathbb{R}^n$ that has zeroes everywhere except a 1 in the $i$th entry. Take then the behaviour $\P$ with matrix \begin{align} P&=\left( \begin{array}{c} \textbf{1}_n \\ \textbf{0}_n \\ \textbf{0}_n \\ \vdots \\ \textbf{0}_n \\ \end{array} \right)\left( \begin{array}{cccc} (e_1^{(d)})^T & (e_1^{(d)})^T & \cdots & (e_1^{(d)})^T \\ \end{array} \right)\nonumber\\&+\left( \begin{array}{c} \textbf{0}_n \\ \textbf{1}_n \\ \textbf{0}_n \\ \vdots \\ \textbf{0}_n \\ \end{array} \right)\left( \begin{array}{cccc} (e_2^{(d)})^T & (e_2^{(d)})^T & \cdots & (e_2^{(d)})^T \\ \end{array} \right)\nonumber\\&+\cdots+\left( \begin{array}{c} \textbf{0}_n \\ \textbf{0}_n \\ \vdots \\ \textbf{0}_n \\ \textbf{1}_n \\ \end{array} \right)\left( \begin{array}{cccc} (e_d^{(d)})^T & (e_d^{(d)})^T & \cdots & (e_d^{(d)})^T \\ \end{array} \right), \end{align} where the column (row) vectors belong to $\mathbb{R}^{dn}$ ($\mathbb{R}^{dm}$). It then follows that $||P||_1=d\sqrt{nm}=\sqrt{d|\x||\y|}$ and that $\P\in\c_d$ (and, hence, $\P\in\q_d$). To see the first claim notice that $P=\sum_{i=1}^d\sqrt{nm}|u_i\rangle\langle v_i|$ where the $\{|u_i\rangle\}$ and $\{|v_i\rangle\}$ are sets of orthonormal vectors. To see the second claim, notice that $P(b|xy)$ takes the form (\ref{setc}) with $s(m|x)=\delta_{m,\lceil x/d\rceil}$ and $t(b|my)=\delta_{bm}$. It is worth remarking that, despite the above example, not all deterministic behaviours in $\c_d$ attain the bound, i.e.\ it can be easily checked that there exist such instances where $||P||_1<\sqrt{d|\x||\y|}$. Notice, however, that this does not imply that the estimation given by Theorem 1 is not optimal as it may happen that $d\neq||P||_1^2/(|\x||\y|)$ but $d=\lceil||P||_1^2/(|\x||\y|)\rceil$. More interestingly, as we show in the next section, we can also prove that there exist $\P\in\q_d$ such that $\P\notin\c_d$ for which $||P||_1=\sqrt{d|\x||\y|}$. Certain quantum random access codes or the behaviours considered in \cite{didwpm} provide such examples. \section{Applications} \subsection{Quantum random access codes} As mentioned in the introduction, DIDW is closely related to the field of communication complexity in the setting of one-way communication complexity. Here, one asks what the minimal dimension of the (classical or quantum) messages from A to B must be in order for B to compute a given function $f(x,y): \x\times\y\to\b$ with a certain degree of success. The figure of merit which is usually considered here is the worst-case probability \begin{equation} p_w=\min\{P(b|xy):f(x,y)=b\}. \end{equation} A particular instance of this problem that has received quite some attention in the literature is random access coding \cite{rac}. Although several particular versions of this protocol have been considered, here we take the most general form in which A receives a string $x=x_1\cdots x_n$ where $x_i\in\{1,\ldots,m\}$ and B receives an input $y\in\{1,\ldots,n\}$ with the goal that $f(x,y)=x_y$ (thus $|\x|=m^n$, $|\y|=n$ and $|\b|=m$). It has been shown in \cite{ozols} that if A and B have access to shared randomness, then for the corresponding optimal strategies $p_w$ equals the average success probability \begin{equation} p=\frac{1}{nm^n}\sum_{f(x,y)=b}P(b|xy). \end{equation} Thus, the probability of success of any $(m,n)$ quantum random access code (QRAC) with communication cost $d$ can be written as $\langle P,G(m,n)\rangle$ with $\P\in\q_d$ and Theorem 2 can be applied to upper bound $p$ as a function of $d$. Here, $G(m,n)=F(m,n)/(nm^n)\in\mathbb{R}^{m^n\times mn}$, where \begin{equation} F(m,n)=\sum_{b,x,y}\delta_{b,x_y}|x\rangle\langle yb|. \end{equation} It might also be helpful to have in mind the following inductive construction of this matrix, \begin{equation} F(m,n)=\left( \begin{array}{ccccc} \textbf{1}_{m^{n-1}} & \textbf{0}_{m^{n-1}} & \cdots & \textbf{0}_{m^{n-1}} & F(m,n-1) \\ \textbf{0}_{m^{n-1}} & \textbf{1}_{m^{n-1}} & \cdots & \textbf{0}_{m^{n-1}} & F(m,n-1) \\ \vdots & \vdots & \ddots & \vdots & \vdots \\ \textbf{0}_{m^{n-1}} & \cdots & \textbf{0}_{m^{n-1}} & \textbf{1}_{m^{n-1}} & F(m,n-1) \\ \end{array} \right) \end{equation} with $F(m,1)=\one_m$. Before proceeding to establish the bound it should be stressed that the application of Theorem 2 is not completely straightforward as it allows for a certain form of optimization. This is because there exist different choices of matrix $G$ to codify the same function $f$ due to the fact that $\sum_bP(b|xy)=1$ $\forall x,y$. Indeed, denoting by $\{A_{xy}\}$ the matrices \begin{equation} A_{xy}=\sum_b|x\rangle\langle yb|, \end{equation} we have that $\langle P,A_{xy}\rangle=1$ $\forall x,y$ and for every behaviour $\P$. Thus, $\forall\P\in\q_d$ it holds that \begin{equation} \langle P,G\rangle\leq||G+\sum_{xy}\alpha_{xy}A_{xy}||_\infty\sqrt{d|\x||\y|}-\sum_{xy}\alpha_{xy} \end{equation} for any choice of real numbers $\{\alpha_{xy}\}$. Interestingly, these norms are sensitive to the constraint fulfilled by behaviours and they can lead to different bounds. In our case, it seems that best results are obtained when the matrix $G+\sum_{xy}\alpha_{xy}A_{xy}$ is chosen to be a partial isometry. In the following we use the notation $\textbf{1}(m,n)$ for the $m\times n$ matrix with all entries equal to one. \begin{lemma} Let \begin{equation}\label{eqlemma3} H=\frac{1}{\sqrt{m^{n-1}}}(F(m,n)-a_{mn}\textbf{1}(m^n,mn)), \end{equation} with \begin{equation} a_{mn}=\frac{1}{m}-\frac{1}{m\sqrt{n}}. \end{equation} Then, $H$ is a partial isometry (and, hence, $||H||_\infty=1$). \end{lemma} \begin{proof} In order to prove the claim we show that all eigenvalues of $H^T H$ are either 1 or 0. We will denote by $h_i$ the columns of the matrix $H$, which means that the index takes values $i=(y,b)\in\y\times\b$. Consequently, we will say that $i$ and $j$ belong to the same input if $i=(y,b)$ and $j=(y,b')$ for some $y\in\y$. Notice that all entries of $H$ are either $(1-a_{mn})/\sqrt{m^{n-1}}$ or $-a_{mn}/\sqrt{m^{n-1}}$ and, therefore, \begin{widetext} \begin{equation} h_i^Th_i=\frac{1}{m^{n-1}}\left[m^{n-1}(1-a_{mn})^2+(m-1)m^{n-1}a_{mn}^2\right]=1-\frac{1}{m}+\frac{1}{mn}:=a, \end{equation} and, if $i\neq j$, \begin{equation} h_i^Th_j=\frac{1}{m^{n-1}}\left[-2m^{n-1}(1-a_{mn})a_{mn}+(m-2)m^{n-1}a_{mn}^2\right]=-\frac{1}{m}+\frac{1}{mn}:=b \end{equation} if $i$ and $j$ belong to the same input while otherwise we have that \begin{equation} h_i^Th_j=\frac{1}{m^{n-1}}\left[m^{n-2}(1-a_{mn})^2+(m-1)^2m^{n-2}a_{mn}^2-2(m-1)m^{n-2}(1-a_{mn})a_{mn}\right]=\frac{1}{mn}. \end{equation} \end{widetext} Thus, our $mn\times mn$ matrix is given by \begin{equation} H^TH=\left( \begin{array}{cccc} A & B & \cdots & B \\ B & A & \ddots & \vdots \\ \vdots & \ddots & \ddots & B \\ B & \cdots & B & A \\ \end{array} \right) \end{equation} with $m\times m$ blocks $B=\textbf{1}(m,m)/(mn)$ and \begin{equation} A=\left( \begin{array}{cccc} a & b & \cdots & b \\ b & a & \ddots & \vdots \\ \vdots & \ddots & \ddots & b \\ b & \cdots & b & a \\ \end{array} \right). \end{equation} Since $H^TH$ happens to be a row stochastic matrix, it follows that $\textbf{1}_{mn}$ is an eigenvector with eigenvalue 1. On the other hand, $A$ is a circulant matrix, so it is easily found that its eigenvalues are $1/n$ and $a-b=1$, the last one having degeneracy equal to $m-1$ and eigenvectors $\{v_i\}$ in the orthogonal complement of the span of $\textbf{1}_m$. This last property implies that the $\{v_i\}$ are in the kernel of $B$ and, therefore, the $\mathbb{R}^{mn}$ vectors \begin{equation} \left\{\left( \begin{array}{c} v_i \\ \textbf{0}_m \\ \vdots \\ \textbf{0}_m \\ \end{array} \right),\left( \begin{array}{c} \textbf{0}_m \\ v_i \\ \vdots \\ \textbf{0}_m \\ \end{array} \right),\ldots,\left( \begin{array}{c} \textbf{0}_m \\ \vdots \\ \textbf{0}_m \\ v_i \\ \end{array} \right)\right\} \end{equation} are all eigenvectors of $H^TH$ with eigenvalue equal to 1. Thus, altogether, we have seen that that this eigenvalue has degeneracy at least $1+n(m-1)$. However, $\tr(H^TH)=1+n(m-1)$, which implies that all the remaining eigenvalues necessarily must be equal to 0. \end{proof} Now, using Eq.\ (\ref{eqlemma3}) we can write $\langle P,G(m,n)\rangle$ in terms of $\langle P,H\rangle$ and $\langle P,\textbf{1}(m^n,mn)\rangle$. Bounding the former with Theorem 2 and using that the latter equals $nm^n$ for every behaviour $\P$, we immediately arrive at the following result. \begin{theorem} Every $(m,n)$ QRAC with $\P\in \q_d$ fulfills \begin{equation} p\leq\frac{1}{m}+\frac{\sqrt{md}-1}{m\sqrt{n}}. \end{equation} \end{theorem} This bound reduces to that of \cite{ozols} in the case $m=d=2$. This happens to be tight when $n=2,3$ and the corresponding $\q_2$ behaviours (which cannot be in $\c_2$) have the property that $||P||_1$ fulfills Theorem 1 with equality. One should notice, however, that the bounds are in general not sharp, which can be seen in the particular instance $n=2$ and $d=m>2$ since the exact value of $p$ in this case is known \cite{tavakoli,qracmub,qracmub2}. This can also be seen by comparing with the numerical techniques of \cite{num}. \subsection{Sharpening witnesses based on state discrimination} Reference \cite{didwpm} has considered the following scenario: A receives $x\in\x=\{1,\ldots,N\}$ and B $(y,z)\in\x\times\x$ ($y<z$) with the promise that either $x=y$ or $x=z$. The goal of B is to identify which of the two possibilities actually occurred with his output $b\in\{-1,1\}$ (thus $|\x|=N$, $|\y|=N(N-1)/2$ and $|\b|=2$). The aforementioned reference has provided optimal bounds for the performance in this game both in $\c_d$ and $\q_d$ through non-linear ($W_N=\sum_{y<z}(P(1|x=y)-P(1|x=z))^2$) and linear ($V_N=\sum_{y<z}P(1|x=y)-P(1|x=z)$) witnesses. Interestingly, for a fixed value of $d$ there can exist gaps between the classical and quantum bounds; however, this is not the case if $N$ is a multiple of $d$. The optimal quantum strategy consists in sending the states \begin{equation}\label{states} |\psi_x\rangle=\frac{1}{\sqrt{d}}\sum_{k=0}^{d-1}\exp\left(i\frac{2\pi kx}{N}\right)|k\rangle \end{equation} and B implementing the Helstrom measurement that optimally discriminates between $|\psi_y\rangle$ and $|\psi_z\rangle$ \cite{didwpm,helmstrom}. In the following we show that for $d=2$, the corresponding behaviour is such that $||P||_1=\sqrt{2|\x||\y|}$ $\forall N$. This not only provides another example where Theorem 1 is tight on quantum but non-classical behaviours but, more interestingly, one can then use this insight to construct an alternative witness to $W_N$ and $V_N$ that allows to amplify the gap between the classical and quantum bounds. The general idea on how to construct the witness is very simple. If $P$ has singular value decomposition (SVD) $P=U\Sigma V^T$, then $\langle P,G\rangle=||P||_1$ if $G=UV^T$. Thus, if $\P\in\q_d$ is such that $||P||_1=\sqrt{d|\x||\y|}$, then, by Theorem 2, the witness $G$ constructed following the SVD prescription is optimal for this behaviour. We illustrate this for the behaviour discussed above in the extreme case of $N$ even where the witnesses $W_N$ and $V_N$ cannot discriminate between $\c_2$ and $\q_2$. After some algebra one finds that in the case $d=2$ the above quantum strategy leads to the behaviour \begin{equation}\label{optqbeh} P(b=\pm1|xyz)=\frac{1}{2}\left(1\mp\sin\left[\frac{\pi}{N}(2x-y-z)\right]\right). \end{equation} Notice that here we are ignoring the promise that either $x=y$ or $x=z$ and $P\in\mathbb{R}^{N\times N(N-1)}$. In order to compute $||P||_1$, we compute the eigenvalues of $PP^T\in\mathbb{R}^{N\times N}$. Using the above equation we find that the entries of this matrix are given by \begin{align}\label{calc} (PP^T)_{xx'}&=\frac{1}{2}\sum_{y<z}\left[1+\frac{1}{2}(\cos\theta_{x-x'}-\cos\theta_{x+x'-y-z})\right]\nonumber\\ &=\frac{N(N-1)}{4}\left(1+\frac{\cos\theta_{x-x'}}{2}\right), \end{align} where we have used the shorthand $\theta_j=2\pi j/N$. Since $(PP^T)_{xx'}$ only depends on $|x-x'|$, the matrix is circulant and its eigenvalues $j=0,\ldots,N-1$ are given by \begin{align} \lambda_j&=\sum_{k=0}^{N-1}(PP^T)_{1k}\exp\left(i\frac{2\pi jk}{N}\right)=\frac{N(N-1)}{4}\nonumber\\ &\times\left[\sum_{k=0}^{N-1}\exp\left(i\frac{2\pi jk}{N}\right +\frac{1}{2}\sum_{k=0}^{N-1}\cos\theta_k\exp\left(i\frac{2\pi jk}{N}\right)\right]\nonumber\\ &=\frac{N(N-1)}{4}\left[N\delta_{j0}+\frac{N}{4}(\delta_{j1}+\delta_{j,N-1})\right]. \end{align} Thus, $P$ has rank 3 $\forall N$ and its non-zero singular values are $N\sqrt{N-1}/2$, $N\sqrt{N-1}/4$ and $N\sqrt{N-1}/4$ amounting to $||P||_1=N\sqrt{N-1}=\sqrt{2|\x||\y|}$. \begin{theorem} In the above scenario the witness \begin{equation}\label{witness} G(b=\pm1|xyz)=\frac{2}{N\sqrt{N-1}}\left(\frac{1}{2}\mp\sin\left[\frac{\pi}{N}(2x-y-z)\right]\right) \end{equation} is such that $\forall\P\in\q_2$ \begin{equation} \langle P,G\rangle\leq B_Q=N\sqrt{N-1} \end{equation} with equality attained by the behaviour given in Eq.\ (\ref{optqbeh}). On the other hand, $\forall\P\in\c_2$ and for even $N$ it holds that \begin{align} &\langle P,G\rangle\leq B_C=\frac{2}{N\sqrt{N-1}}\nonumber\\ &\times\left(\frac{N^2(N-1)}{4}+\frac{2}{\sin(\pi/N)}\sum_{y<z}\left|\cos\left[\frac{\pi}{N}(1+y+z)\right]\right|\right),\label{bc} \end{align} with equality attained by a deterministic behaviour. \end{theorem} Figure 1 plots the ratio $B_C/B_Q$ for different values of $N$. It can be readily seen therein that this is always smaller than 1, quickly approaching its asymptotic value $1/2+4/\pi^2\simeq0.9053$. \begin{figure}[h] \includegraphics[scale=0.4]{ratioBcBqb} \caption{Ratio $B_C/B_Q$ for the witness given in Eq.\ (\ref{witness}).} \end{figure} \begin{proof} Proceeding as in Eq.\ (\ref{calc}) on, one finds that $GG^T$ is also a rank 3 matrix with all non-zero eigenvalues equal to 1. Thus, $||G||_\infty=1$ and Theorem 2 gives that $\langle P,G\rangle\leq B_Q$ $\forall\P\in\q_2$. An analogous calculation shows that $\langle P,G\rangle=\tr(PG^T)=B_Q$ for the $\q_2$ behaviour given by Eq.\ (\ref{optqbeh}). Alternatively, the quantum part of the theorem can also be proven by noticing that $G=UV^T$ if the behaviour (\ref{optqbeh}) has the reduced SVD $P=U\Sigma V^T$, i.e.\ $\Sigma=diag(N\sqrt{N-1}/2, N\sqrt{N-1}/4, N\sqrt{N-1}/4)$ and $U$ and $V$ respectively the corresponding $N\times3$ and $N(N-1)\times3$ partial isometries. It remains to obtain the classical bound $B_C$ for behaviours in $\c_2$. Due to linearity, it must be attained by a deterministic behaviour, i.e.\ such that for every $x,y,z$, $P(b|x,y,z)$ equals 0 or 1 depending on whether $b=\pm1$. Obviously, the best possible strategy is to assign $P(1|xyz)=1$ if $\sin\left[\frac{\pi}{N}(2x-y-z)\right]<0$ and $P(1|xyz)=0$ otherwise. However, Bob does not know $x$ but $m(x)$, which can only take two values (say 0 and 1) given that $\P\in\c_2$. Thus, the best Bob can do is to compute $\sum_{x:m(x)=0}\sin\left[\frac{\pi}{N}(2x-y-z)\right]$ and check the sign of this expression for his inputs $(y,z)$. Since $\sum_{x=1}^N\sin\left[\frac{\pi}{N}(2x-y-z)\right]=0$, the optimal value for a given coding function $m$ is then given by \begin{align} &\langle P,G\rangle=\frac{2}{N\sqrt{N-1}}\nonumber\\ &\times\left(\sum_{x,y<z}\frac{1}{2}+2\sum_{y<z}\left|\sum_{x:m(x)=0}\sin\left[\frac{\pi}{N}(2x-y-z)\right]\right|\right). \end{align} Noticing now that the best coding function A and B can agree on is that for which most $x$ with the same image lead to $\sin\left[\frac{\pi}{N}(2x-y-z)\right]$ having the same sign for most pairs $(y,z)$, it follows that the optimal strategy corresponds to assigning the same value under $m$ to a consecutive set of elements in $\x$. Thus, \begin{equation} B_C=\frac{2}{N\sqrt{N-1}}\left(\frac{N^2(N-1)}{4}+2\max_{j,k}\sum_{y<z}|S_{j,k}|\right), \end{equation} where \begin{align} S_{j,k}&=\sum_{x=j}^{j+k}\sin\left[\frac{\pi}{N}(2x-y-z)\right]\nonumber\\ &=\frac{\sin\left[\frac{\pi}{N}(2j+k-y-z)\right]\sin\left[\frac{\pi}{N}(k+1)\right]}{\sin(\pi/N)}.\label{maximization} \end{align} It is not difficult to show that $\forall j\in\mathbb{Z}$ it holds that \begin{align} &\sum_{y=1}^{N-1}\sum_{z=y+1}^N\left|\sin\left[\frac{\pi}{N}(2j+k-y-z)\right]\right|\nonumber\\ &=\sum_{y=1}^{N-1}\sum_{z=y+1}^N\left|\sin\left[\frac{\pi}{N}(k-y-z)\right]\right|,\label{eqsin}\\ &\sum_{y=1}^{N-1}\sum_{z=y+1}^N\left|\cos\left[\frac{\pi}{N}(2j+k-y-z)\right]\right|\nonumber\\ &=\sum_{y=1}^{N-1}\sum_{z=y+1}^N\left|\cos\left[\frac{\pi}{N}(k-y-z)\right]\right|,\label{eqcos} \end{align} which we will use repeatedly in the following. Equation (\ref{eqsin}) implies that the value of $j$ is irrelevant in our maximization and, thus, we can write \begin{equation} \max_{j,k}\sum_{y<z}|S_{j,k}|=\max_{k}\sum_{y<z}|S_{0,k}|. \end{equation} Furthermore, using again Eq.\ (\ref{eqsin}) and Eq.\ (\ref{maximization}) (and taking into account that we are considering $N$ to be even) we can conclude that the above maximum must occur at either $k=N/2-1$ or $k=N/2$. However, \begin{align} &\sin(\pi/N)\sum_{y<z}|S_{0,N/2}|\nonumber\\ &=\frac{1}{2}\sum_{y<z}\left|\cos\left[\frac{\pi}{N}(1+y+z)\right]+\cos\left[\frac{\pi}{N}(1-y-z)\right]\right|\nonumber\\ &\leq\frac{1}{2}\sum_{y<z}\left(\left|\cos\left[\frac{\pi}{N}(1+y+z)\right]\right|+\left|\cos\left[\frac{\pi}{N}(1-y-z)\right]\right|\right)\nonumber\\ &=\sum_{y<z}\left|\cos\left[\frac{\pi}{N}(1+y+z)\right]\right|=\sin(\pi/N)\sum_{y<z}|S_{0,N/2-1}|, \end{align} where to arrive at the last line we have used Eq.\ (\ref{eqcos}). Hence, the maximum occurs when $k=N/2-1$, i.e.\ the optimal coding function A and B can agree on is one that assigns the same value to a subset of $N/2$ consecutive elements in $\x$. This proves Eq.\ (\ref{bc}). \end{proof} As discussed before, the witness $G$ of Theorem 5 can be changed to a witness of the form $G+\sum_{xyz}\alpha_{xyz}A_{xyz}$ keeping track of the corresponding classical and quantum bounds. It might be interesting to notice that this allows to reinterpret its value as the probability of computing some distributed function $f(x,y,z)$. One then has that with probability \begin{equation} \pi(x,y,z)=\frac{|\sin\left[\frac{\pi}{N}(2x-y-z)\right]|}{\sum_{x,y<z}|\sin\left[\frac{\pi}{N}(2x-y-z)\right]|} \end{equation} A and B receive the inputs $x$ and $(y,z)$ respectively with the goal that Bob answers $f(x,y,z)=1$ ($f(x,y,z)=-1$) whenever $\sin\left[\frac{\pi}{N}(2x-y-z)\right]<0$ ($\sin\left[\frac{\pi}{N}(2x-y-z)\right]>0$). The bounds $B_C$ and $B_Q$ can be changed correspondingly to obtain the maximal average success probability in each setting, which will then be larger in the quantum case. Here, we have illustrated how the witnesses of \cite{didwpm} can be improved to distinguish classical and quantum bidimensional behaviours. However, different dimensions can be assessed. First, one can directly apply Lemma 2 to conclude that for the witness of Eq.\ (\ref{witness}) it holds \begin{equation} \langle P,G\rangle\leq N\sqrt{\frac{d(N-1)}{2}} \end{equation} for every $\P\in\q_d$. Another strategy is to consider a different witness taking the one corresponding to the (reduced) SVD of the $\q_d$ behaviour given by Helstrom measurements on the states (\ref{states}) as we did in the $d=2$ case, for which the same bound as above holds. \section{Conclusion} Previous works obtaining bounds for DIDW in the prepare-and-measure scenario had relied on particular constructions of dimension witnesses or considered arbitrary behaviours under the condition that shared randomness among parties is not available. In this work we have obtained general and explicit lower bounds on the dimension of arbitrary quantum behaviours dropping this assumption, which are based on standard techniques from matrix theory. Although the bounds are not always tight and in general perform worse than the numerical techniques of Ref.\ \cite{num}, we expect that their simple and easy-to-use form make them helpful for further investigations in this context, particularly when a numerical approach is not feasible due to the analytical nature of the problem at hand or because it is computationally too demanding. In fact, we have provided two applications of our result. First, we have proved that using our techniques the probability of success of distributed computational tasks can be upper bounded in general as a function of the dimension of the message in a paradigmatic example such as random access codes. Second, we have shown that our construction allows one to derive powerful dimension witnesses for given behaviours. In particular, when the behaviour $\P$ is at the boundary of $\q_d$ and fulfills the condition of Theorem 1 with equality, the SVD of the matrix associated to $\P$ yields immediately an optimal dimension witness for it. As an example, we have used this to improve the constructions of dimension witnesses given in \cite{didwpm}. For the future we hope that this insight makes it possible to find systematically adequate dimension witnesses in relevant physical situations, to improve known semi-device-independent protocols and, in general, to understand better the mathematical structure of fixed-dimensional behaviours. It might be also interesting to apply these techniques to bound the success probability as a function of the allowed quantum communication for other distributed tasks of interest. This research was funded by the Spanish MINECO through grants MTM2017-84098-P and MTM2017-88385-P and by the Comunidad de Madrid through grant QUITEMAD+CM S2013/ICE-2801.

2044b67a1ec256af9e5e177d1d550f72b5a50650

\section{Motivation and Background} \label{sec:intro} Weakly supervised training of neural networks is often based on {\em regularized losses} combining an empirical loss with some regularization term, which compensates for lack of supervision \cite{weston2012deep,goodfellow2016deep}. Regularized losses are also useful for CNN segmentation \cite{NCloss:CVPR18,Rloss:ECCV18} where full supervision is often infeasible, particularly in biomedical applications. Such losses are motivated by regularization energies in {\em shallow}\footnote{In this paper, ``shallow'' refers to methods unrelated to deep learning.} segmentation, where multi-decade research went into designing robust regularization models based on geometry \cite{Mumford-Shah-89,Sapiro:97,BK:ICCV03}, physics \cite{Kass:88,BlakeZis:87}, or robust statistics \cite{Geman:84}. Such models should represent realistic shape priors compensating for image ambiguities, yet be amenable to efficient solvers. Many robust regularizers commonly used in vision \cite{SzeliskiMRFCompare:08,kappes2015comparative} are non-convex and require powerful optimizers to avoid many weak local minima. Basic local optimizers typically fail to produce practically useful results with such models. Effective weakly-supervised CNN methods for vision should incorporate priors compensating for image data ambiguities and lack of supervision just as in shallow vision methods. For example, recent work \cite{weston2012deep,Rloss:ECCV18} formulated the problems of semi-supervised classification and weakly-supervised segmentation as minimization of regularized losses. This principled approach outperforms common `'proposal generation'' methods \cite{scribblesup,kolesnikov2016seed} computing ``fake'' ground truths to mimic standard fully-supervised training. However, we show that the use of regularization models as losses in deep learning is limited by GD, the backbone optimizer in current training methods. It is well-known that GD leads to poor local minima for many regularizers in shallow segmentation and many stronger algorithms were proposed~\cite{BJ:01,BVZ:PAMI01,trws,SzeliskiMRFCompare:08,lsa-tr}. Similarly, we show better optimization beyond GD for regularized losses in deep segmentation. One popular general approach applicable to regularized losses is ADMM \cite{Boyd2011} that splits optimization into two efficiently solvable sub-problems separately focusing on the empirical loss and regularizer. We advocate similar {\em splitting} to improve optimization of regularized losses in CNN training. In contrast, ADMM-like splitting of network parameters in different layers was used in \cite{taylor2016training} to improve parallelism. In our work weakly-supervised CNN segmentation is a context for discussing regularized loss optimization. As a regularizer, we use the common Potts model \cite{BVZ:PAMI01} and consider its nearest- and large-neighborhood variants, a.k.a. sparse {\em grid CRF} and {\em dense CRF} models. We show effectiveness of ADMM-like splitting for grid CRF losses due to availability of powerful sub-problem solvers, \eg graph cuts \cite{BK:ICCV03}. As detailed in \cite[Sec.3]{Rloss:ECCV18}, an earlier iterative proposal-generation technique by \cite{kolesnikov2016seed} can be related to regularized loss splitting, but their method is limited to dense CRF and its approximate {\em mean-field} solver~\cite{koltun:NIPS11}. In fact, given such weak sub-problem solvers, splitting is inferior to basic GD over the regularized loss \cite{Rloss:ECCV18}. More insights on grid and dense CRF are below. \subsection{Pairwise CRF for Shallow Segmentation} \label{sec:shallow} \begin{figure*}[t] \centering \begin{tabular}{ccc} \includegraphics[width=1.7in]{fig/1D_image.png} & \includegraphics[width=1.7in]{fig/1D_sparseCRF.png} & \includegraphics[width=1.7in]{fig/1D_denseCRF.png} \\ (a) 1D image & (b) grid CRF \cite{BJ:01} & (c) dense CRF \cite{koltun:NIPS11} \end{tabular} \caption{{\em Synthetic segmentation example} for grid and dense CRF (Potts) models: (a) intensities $I(x)$ on 1D image. The cost of segments $S^t = \{x\;|\;x<t\}$ with different discontinuity points $t$ according to (b) nearest-neighbor (grid) Potts and (c) larger-neighborhood (dense) Potts. The latter gives smoother cost function, but its flatter minimum may complicate discontinuity localization. } \label{fig:synthetic} \end{figure*} \begin{figure*}[h] \centering \begin{tabular}{ccc} \includegraphics[width=1.8in,trim={5cm 0.6cm 2cm 4cm},clip]{fig/croc_seeds.png} & \includegraphics[width=1.8in,trim={5cm 0.6cm 2cm 4cm},clip]{fig/croc_sparseCRF.png} & \includegraphics[width=1.8in,trim={5cm 0.6cm 2cm 4cm},clip]{fig/croc_denseCRF.png} \\ (a) image + seeds & (b) grid CRF \cite{BJ:01} & (c) dense CRF \cite{koltun:NIPS11} \end{tabular} \caption{{\em Real "shallow" segmentation example} for sparse (b) and dense (c) CRF (Potts) models for image with seeds (a). Sparse Potts gives smoother segment boundary with better edge alignment, while dense CRF inference often gives noisy boundary.} \label{fig:shallow} \end{figure*} Robust pairwise Potts model and its binary version (Ising model) are used in many application such as stereo, reconstruction, and segmentation. One can define this model as a cost functional over integer-valued labeling $S:=(S_p \in Z^+ \,|\,p\in\Omega)$ of image pixels $p\in\Omega$ as follows \begin{equation} \label{eq:potts} E_P (S) \;\;=\;\; \sum_{pq\in{\cal N}} w_{pq} \cdot [S_p \neq S_q] \end{equation} where ${\cal N}$ is a given neighborhood system, $w_{pq}$ is a discontinuity penalty between neighboring pixels $\{p,q\}$, and $[\cdot]$ is {\em Iverson bracket}. The nearest-neighbor version over $k$-connected grid ${\cal N}_k$, as well as its popular variational analogues, \eg geodesic active contours \cite{Sapiro:97}, convex relaxations \cite{pock:CVPR09,chambolle2011first}, or continuous max-flow \cite{Yuan:CVPR2010}, are particularly well-researched. It is common to use contrast-weighted discontinuity penalties \cite{BVZ:PAMI01,BJ:01} between the neighboring points, as emphasized by the condition $\{pq\}\in{\cal N}_k$ below \begin{equation} \label{eq:NNweights} w_{pq} \;\;\;\; = \;\;\;\; \lambda\cdot\exp{\frac{-||I_p-I_q||^2}{2\sigma^2}}\;\;\cdot\;\;[\{pq\}\in{\cal N}_k]. \end{equation} Nearest neighbor Potts models minimize the contrast-weighted length of the segmentation boundary preferring shorter perimeter aligned with image edges, \eg see Fig.~\ref{fig:shallow}(b). The popularity of this model can be explained by generality, robustness, well-established foundations in geometry, and a large number of efficient discrete or continuous solvers that guarantee global optimum in binary problems \cite{BJ:01} or some quality bound in multi-label settings, \eg $\alpha$-expansion~\cite{BVZ:PAMI01}. Dense CRF \cite{koltun:NIPS11} is a Potts model where pairwise interactions are active over significantly bigger neighborhoods defined by a Gaussian kernel with a relatively large bandwidth $\Delta$ over pixel locations \begin{equation} \label{eq:NNweightsColt} w_{pq} \;\; = \;\; \lambda\cdot\exp{\frac{-||I_p-I_q||^2}{\sigma^2}}\;\cdot\;\exp{\frac{-\|p-q\|^2}{\Delta^2}}. \end{equation} Its use in shallow vision is limited as it often produces noisy boundaries \cite{koltun:NIPS11}, see also Fig.~\ref{fig:shallow}(c). Also, global optimization methods mentioned above do not scale to dense neighborhoods. Yet, dense CRF model is popular in the context of CNNs where it can be used as a differentiable regularization layer \cite{zheng2015conditional,schwing2015fully}. Larger bandwidth yields smoother objective \eqref{eq:potts}, see Fig. \ref{fig:synthetic}(c), amenable to gradient descent or other local linearization methods like mean-field inference that are easy to parallelize. Note that existing efficient inference methods for dense CRF require {\em bilateral filtering} \cite{koltun:NIPS11}, which is restricted to Gaussian weights as in \eqref{eq:NNweightsColt}. This is in contrast to global Potts solvers, \eg $\alpha$-expansion, that can use arbitrary weights, but become inefficient for dense neighborhoods. Noisier dense CRF results, \eg in Fig.~\ref{fig:shallow}(c), imply weaker regularization. Indeed, as discussed in \cite{veksler2018efficient}, for larger neighborhoods the Potts model gets closer to {\em cardinality potentials}. Bandwidth $\Delta$ in \eqref{eq:NNweightsColt} is a resolution scale at which the model sees the segmentation boundary. Weaker regularization in dense CRF may preserve some thin structures smoothed out by fine-resolution boundary regularizers, \eg nearest-neighbor Potts. However, this is essentially the same ``noise preservation'' effect shown in Fig.~\ref{fig:shallow}(c). For consistency, the rest of the paper refers to the nearest-neighbor Potts model as {\em grid CRF}, and large-neighborhood Potts as {\em dense CRF}. \subsection{Summary of Contributions} \label{sec:contributions} Any motivation for standard regularization models in shallow image segmentation, as in the previous section, directly translates into their motivation as regularized loss functions in weakly supervised CNN segmentation \cite{NCloss:CVPR18,Rloss:ECCV18}. The main issue is how to optimize these losses. Standard training techniques based on gradient descent may not be appropriate for many powerful regularization models, which may have many local minima. Below is the list of our main contributions: \begin{itemize} \item As an alternative to gradient descent (GD), we propose a splitting technique, \emph{alternating direction method} (ADM)\footnote{Standard ADM\textit{\textbf{M}}~\cite{Boyd2011} casts a problem $\min_{x} f(x) + g(x)$ into $\min_{x,y}\max_\lambda f(x) + g(y) + \lambda^\mathrm{T}(x-y) + \rho\|x-y\|^2$ and alternates updates over $x$, $y$ and $\lambda$ optimizing $f$ and $g$ in parallel. Our ADM uses a different form of splitting and can be seen as a {\em penalty method}, see Sec.~\ref{sec:ADM}.}, for minimizing {\em regularized losses} during network training. ADM can directly employ efficient regularization solvers known in shallow segmentation. \item Compared to GD, our ADM approach with $\alpha$-expansion solver significantly improves optimization quality for the {\em grid CRF} (nearest-neighbor Potts) loss in weakly supervised CNN segmentation. While each iteration of ADM is slower than GD, the loss function decreases at a significantly larger rate with ADM. In one step it can reach lower loss values than those where GD converges. Grid CRF has never been investigated as loss for CNN segmentation and is largely overlooked. \item The training quality with grid CRF loss achieves the-state-of-the-art in weakly supervised CNN segmentation. We compare dense CRF and grid CRF losses. \end{itemize} Our results may inspire more research on regularized segmentation losses and their optimization. \section{ADM for Regularized Loss Optimization} \label{sec:ADM} Assume there is a dataset of pairs of images and partial ground truth labelings. For simplicity of notation, we implicitly assume summation over all pairs in dataset for all expressions of loss functions. For each pixel $p \in \Omega$ of each image $I$ there is an associated color or intensity $I_p$ of that pixel. The labeling $Y=(Y_p|p \in \Omega_{\cal L})$ where $\Omega_{\cal L} \subset \Omega$ is a set of labeled pixels, each $Y_p \in \{0,1\}^K$ is a one-hot distribution and $K$ is the number of labels. We consider a regularized loss for network $\phi_\theta$ of the following form \begin{equation} \label{CNN-regularized-loss} \ell (S_\theta,Y) \;\;+ \;\;\lambda\cdot R(S_\theta) \;\;\;\; \to \;\;\;\; \min_\theta \end{equation} where $S_\theta \in [0,1]^{|\Omega|\times K}$ is a $K$-way softmax segmentation generated by the network $S_\theta := \phi_\theta(I)$, and $R(\cdot)$ is a regularization term, \eg \emph{relaxed} sparse Potts or dense CRF, and $\ell (\cdot,\cdot)$ is a partial ground-truth loss, for instance: $$ \ell (S_\theta,Y) \; = \sum_{p\in\Omega_{\cal L}} {H}(Y_p,S_{p, \theta}), $$ where $H(Y_p,S_{p, \theta})=-\sum_kY_p^k \log S_{p, \theta}^k$ is the cross entropy between predicted segmentation $S_{p, \theta}$ (a row of matrix $S_{\theta}$ corresponding to pixel $p$) and ground truth labeling $Y_p$. We present a general alternating direction method (ADM) to optimize neural network regularized losses of the general form in \eqref{CNN-regularized-loss} using the following splitting of the problem: \begin{equation} \label{} \begin{aligned} & \underset{\theta,X}{\text{minimize}} & & \ell (S_\theta,Y) \; + \; \lambda R(X) \\ & \text{subject to} & & \sum_{p\in\Omega_{{\cal U}}} D(X_p|S_{p,\theta}) = 0, \\ & & & X_p = Y_p & \forall p \in \Omega_{\cal L} \end{aligned} \end{equation} where we introduced one-hot distributions $X_p \in \{ 0,1 \}^K$ and some divergence measure $D$, \eg the Kullback-Leibler divergence. $R(X)$ can now be a \emph{discrete} classic MRF regularization, \eg \eqref{eq:potts}. This equates to the following Lagrangian \begin{equation} \label{Alternating-direction-KL} \begin{aligned} & \min_{\theta,X}\max_\gamma & & \ell (S_\theta,Y) + \lambda R(X) + \gamma \!\sum_{p\in\Omega_{{\cal U}}} \!D(X_p|S_{p,\theta})\\ & \text{subject to} & & X_p = Y_p \;\;\; \forall p \in \Omega_{\cal L}. \end{aligned} \end{equation} We alternate optimization over $X$ and $\theta$ in \eqref{Alternating-direction-KL}. The maximization over $\gamma$ increases its value at every update resulting in a variant of simulated annealing. We have experimented with variable multiplier $\gamma$ but found no advantage compared to fixed $\gamma$. So, we fix $\gamma$ and do not optimize for it. In summary, instead of optimizing the regularization term with gradient descent, our approach splits regularized-loss problem \eqref{CNN-regularized-loss} into two sub-problems. We replace the softmax outputs $S_{p,\theta}$ in the regularization term by latent discrete variables $X_p$ and ensure consistency between both variables (\ie, $S_\theta$ and $X$) by minimizing divergence $D$. This is similar conceptually to the general principles of ADMM~\cite{Boyd2011,NIPS2014_5612}. Our ADM splitting accommodates the use of powerful and well-established discrete solvers for the regularization loss. As we show in Sec.~\ref{sec:experiments}, the popular $\alpha$-expansion solver \cite{BVZ:PAMI01} significantly improves optimization of grid CRF losses yielding state-of-the-art training quality. Such efficient solvers guarantee global optimum in binary problems \cite{BJ:01} or a quality bound in multi-label case~\cite{BVZ:PAMI01}. Our discrete-continuous ADM method\footnote{Combining continuous and discrete sub-problem solvers is not uncommon in tailored ADMM-inspired splitting algorithms \cite{Torr:BMVC14,dtADMM:ICIP16,Ismail:CVPR17,Ismail:miccai17,Manu:CVPR18}.} alternates two steps, each decreasing \eqref{Alternating-direction-KL}, until convergence. Given fixed discrete latent variables $X_p$ computed at the previous iteration, the first step learns the network parameters $\theta$ by minimizing the following loss via standard back-propagation and a variant of stochastic gradient descent (SGD): \begin{equation} \label{SGD} \underset{\theta}{\text{minimize}} \;\; \ell (S_\theta,Y) \; + \; \gamma \sum_{p\in\Omega_{{\cal U}}} D(X_p|S_{p,\theta}) \end{equation} The second step fixes the network output $S_\theta$ and finds the next latent binary variables $X$ by minimizing the following objective over $X$ via any suitable discrete solver: \begin{equation} \label{Alpha-expansion} \begin{aligned} & \underset{{X \in \{ 0,1 \}^{|\Omega|\times K}}}{\text{minimize}} & &\lambda R(X) \; + \; \gamma \sum_{p\in\Omega_{{\cal U}}} D(X_p|S_{p,\theta})\\ & \text{\;\;\,subject to} & & X_p = Y_p \;\;\;\;\; \forall p \in \Omega_{\cal L}. \end{aligned} \end{equation} Because $X_p$ is a discrete variable with only $K$ possible values, the second term in \eqref{Alpha-expansion} is a basic unary term. Similarly, the equality constraints could be implemented as unary terms using prohibitive values of unary potentials. Unary terms are simplest possible energy potentials that can be handled by any general discrete solver. On the other hand, the regularization term $R(X)$ usually involves interactions of two or more variables introducing new properties of solution together with optimization complexity. In case of grid CRF one can use graph cut \cite{BJ:01}, $\alpha$-expansion \cite{BVZ:PAMI01}, QPBO \cite{qpbo,qpbo-i}, TRWS \cite{trws}, LBP \cite{lbp}, LSA-TR \cite{lsa-tr} \etc. In summary, our approach alternates the two steps described above. For each minibatch we compute network prediction, then compute hidden variables $X$ optimizing \eqref{Alpha-expansion}, then compute gradients of loss \eqref{SGD} and update the parameters of the network using a variant of SGD. The outline of our ADM scheme is shown in Alg.~\ref{alg:adm}. \begin{algorithm} \begin{algorithmic} \REQUIRE sequence of minibatches \STATE $i \leftarrow 0$; \STATE initialize network parameters $\theta^{(0)}$; \FOR{each minibatch $B$} \FOR{each image-labeling pair $(I,Y)\in B$} \STATE Compute segmentation prediction $S_{\theta} \leftarrow \phi_{\theta^{(i)}}(I)$; \STATE Solve energy \eqref{Alpha-expansion} for $X$ with \eg $\alpha$-expansion; \STATE Compute gradients $g$ w.r.t. parameters $\theta$ of \eqref{SGD}; \ENDFOR \STATE Compute average over the batch $g^{(i)} \leftarrow \frac1{|B|}\sum g$; \STATE Update network parameters $\theta^{(i+1)}$ using gradient $g^{(i)}$; \STATE $i \leftarrow i + 1$; \ENDFOR \end{algorithmic} \caption{ADM for regularized loss~\eqref{CNN-regularized-loss}.} \label{alg:adm} \end{algorithm} \input{experiments.tex} \section{Conclusion} Gradient descent (GD) is the default method for training neural networks. Often, loss functions and network architectures are designed to be amenable to GD. The top-performing weakly-supervised CNN segmentation \cite{NCloss:CVPR18,Rloss:ECCV18} is trained via regularized losses, as common in weakly-supervised deep learning \cite{weston2012deep,goodfellow2016deep}. In general, GD allows any differentiable regularizers. However, in shallow image segmentation it is known that generic GD is a substandard optimizer for (relaxations of) standard robust regularizers, \eg grid CRF. Here we propose a general splitting technique, ADM, for optimizing regularized losses. It can take advantage of many existing efficient regularization solvers known in shallow segmentation. In particular, for grid CRF our ADM approach using $\alpha$-expansion solver achieves significantly better optimization quality compared to GD. With such ADM optimization, training with grid CRF loss achieves the-state-of-the-art in weakly supervised CNN segmentation. We systematically compare grid CRF and dense CRF losses from modeling and optimization perspectives. Using ADM optimization, the grid CRF loss achieves CNN training favourably comparable to the best results with the dense CRF loss. Our work suggests that in the context of network training more attention should be paid to optimization methods beyond GD. \begin{figure*}[h!] \setlength{\tabcolsep}{1pt} \newcommand{0.156\textwidth}{0.16\textwidth} \centering \begin{tabular}{@{}ccccccc} \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2009_000037.jpg} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2009_000037_DenseCRFGD.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2009_000037_GridCRFGD.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2009_000037_GridCRFADM.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2009_000037_gt.png} \\ \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2009_000704.jpg} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2009_000704_DenseCRFGD.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2009_000704_GridCRFGD.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2009_000704_GridCRFADM.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2009_000704_gt.png} \\ \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2009_004021.jpg} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2009_004021_DenseCRFGD.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2009_004021_GridCRFGD.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2009_004021_GridCRFADM.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2009_004021_gt.png} \\ \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2010_005118.jpg} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2010_005118_DenseCRFGD.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2010_005118_GridCRFGD.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2010_005118_GridCRFADM.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2010_005118_gt.png} \\ \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2010_005428.jpg} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2010_005428_DenseCRFGD.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2010_005428_GridCRFGD.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2010_005428_GridCRFADM.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2010_005428_gt.png} \\ \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2010_005888.jpg} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2010_005888_DenseCRFGD.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2010_005888_GridCRFGD.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2010_005888_GridCRFADM.png} & \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/exam/2010_005888_gt.png} \\ (a) input & (b) Dense GD & (c) Grid GD & (d) Grid ADM & (e) ground truth \end{tabular} \caption{Example segmentations (Deeplab-MSc-largeFOV) by variants of regularized loss approaches. Gradient descent (GD) for grid CRF gives segmentation of poor boundary alignment though grid CRF is part of the regularized loss. ADM for grid CRF significantly improves edge alignment and compares favorably to dense CRF based method.} \label{fig:segexamples} \end{figure*} \begin{figure*}[h!] \centering \includegraphics[width=.45\textwidth]{fig/shorterscribbles_nopost.png} \includegraphics[width=.475\textwidth]{fig/pixelaccuracy_yscaled.png} \caption{Experiment results of training with shorter scribbles with variants of regularized loss approaches. The results are for Deeplab-MSc-largeFOV. We report mIOU (left) and pixel-wise accuracy (right).} \label{fig:shortscribbles} \end{figure*} In general, our ADM approach applies to many regularized losses, as long as there are efficient solvers for the corresponding regularizers. This work is focused on ADM in the context of common pairwise regularizers. Interesting future work is to investigate losses with non-Gaussian pairwise CRF potentials and higher-order segmentation regularizers, \eg $P^n$ Potts model \cite{kohli2009robust}, curvature \cite{Nieuwenhuis_2014_CVPR}, and kernel clustering \cite{Shi2000,Tang2019kernel}. Also with ADM framework, we can explore other optimization methods \cite{kappes2015comparative} besides $\alpha$-expansion for various kinds of regularized losses in segmentation. Our work bridges optimization method in "shallow" segmentation and loss minimization in deep CNN segmentation. \small \bibliographystyle{ieee} \section{Experimental Results} \label{sec:experiments} We conduct experiments for weakly supervised CNN segmentation with scribbles as supervision~\cite{scribblesup}. The focus is on regularized loss approaches~\cite{NCloss:CVPR18,Rloss:ECCV18} yet we also compare our results to proposal generation based method, \eg ScribbleSup \cite{scribblesup}. We test both the grid CRF and dense CRF as regularized losses. Such regularized loss can be optimized by stochastic gradient descent (GD) or alternative direction method (ADM), as discussed in Sec.~\ref{sec:ADM}. We compare three training schemes, namely dense CRF with GD \cite{Rloss:ECCV18}, grid CRF with GD and grid CRF with ADM for weakly supervised CNN segmentation. Before comparing segmentations, in Sec.~\ref{sec:ADMvsGD} we test if using ADM gives better regularization losses than that using standard GD. Our plots of training losses (CRF energy) vs training iterations show how fast the losses converge when minimized by ADM or GD. Our experiment confirms that first order approach like GD leads to a poor local minimum for the grid CRF loss. There are clear improvements of ADM over GD for minimization of the grid CRF loss. In Sec.~\ref{sec:mIOU}, rather than comparing in terms of optimization, we compare ADM and GD in terms of segmentation quality. We report both mIOU (mean intersection over union) and accuracy in particular for boundary regions. In Sec.~\ref{sec:shortenedscribbles}, we also study these variants of regularized loss method in a more challenging setting of shorter scribbles~\cite{scribblesup} or clicks in the extreme case. With ADM as the optimizer, our approach of grid CRF regularized loss compares favorably to dense CRF based approach \cite{Rloss:ECCV18}. \paragraph{Dataset and implementation details} Following recent work \cite{deeplab,scribblesup,kolesnikov2016seed,NCloss:CVPR18} on CNN semantic segmentation, we report our results on PASCAL VOC 2012 segmentation dataset. We train with scribbles from~\cite{scribblesup} on the augmented datasets of 10,582 images and test on the \textit{val} set of 1,449 images. We report mIOU (mean intersection over union) and pixel-wise accuracy. In particular, we are interested in how good is the segmentation in the boundary region. So we compute accuracy for those pixels close to the boundary, for example within 8 or 16 pixels from semantic boundaries. Besides mIOU and accuracy, we also measure the regularization losses, \ie the grid CRF. Our implementation is based on DeepLabv2\footnote{https://bitbucket.org/aquariusjay/deeplab-public-ver2} and we show results on different networks including deeplab-largeFOV, deeplab-msc-largeFOV, deeplab-vgg16 and resnet-101. We do not apply any post-processing to network output segmentation. The networks are trained in two phases. Firstly, we train the network to minimize partial cross entropy (pCE) loss w.r.t scribbles. Then we train with a grid CRF or dense CRF regularization term. To implement gradient descent for the discrete grid CRF loss, we first take its quadratic relaxation, \begin{equation} \langle\mathbf{1},S_{p,\theta}\rangle+\langle\mathbf{1},S_{q,\theta}\rangle-2\langle S_{p,\theta},S_{q,\theta}\rangle. \end{equation} where $S_{p,\theta},S_{q,\theta}\in [0,1]^K$ and $\langle\cdot,\cdot\rangle$ is the dot product. Then we differentiate \wrt $S_{\theta}$ during back-propagation. While there are ways, \eg \cite{grady2011,chambolle2012convex}, to relax discrete Pott's model, we focus on this simple and standard relaxation~\cite{yuille2010belief,koltun:NIPS11,Rloss:ECCV18}. For our proposed ADM algorithm, which requires inference in the grid CRF, we use a public implementation of $\alpha$-expansion\footnote{http://mouse.cs.uwaterloo.ca/code/gco-v3.0.zip}. The CRF inference and loss are implemented and integrated as Caffe \cite{caffe} layers. We run $\alpha$-expansion for five iterations, which in most cases gives convergence. Our dense CRF loss does not include the Gaussian kernel on locations $XY$, since ignoring this term does not change the mIOU measure \cite{koltun:NIPS11}. The bandwidth for the dense Gaussian kernel on $RGBXY$ is validated to give the best mIOU. For the grid CRF, the kernel bandwidth selection in \eqref{eq:NNweights} follows standard Boykov-Jolly \cite{BJ:01} $$\sigma^2 \;\; = \;\; \frac1{|N|} \sum_{pq \in N}{ \|I_p - I_q\|^2}.$$ In general, our ADM optimization for regularized loss is slower than GD due to the inference of grid CRF. However, for inference algorithms, \eg $\alpha$-expansion, that cannot be easily parallelized, we utilize simple multi-core parallelization for all images in a batch to accelerate training. Note that we do not use CRF inference during testing. \subsection{Loss Minimization} \label{sec:ADMvsGD} In this section, we show that for grid CRF losses the ADM approach employing $\alpha$-expansion~\cite{BVZ:PAMI01}, a powerful discrete optimization method, outperforms common gradient descend methods for regularized losses \cite{NCloss:CVPR18,Rloss:ECCV18} in terms of finding a lower minimum of regularization loss. Tab.~\ref{tb:gridCRFloss} shows the grid CRF losses on both training and validation sets for different network architectures. Fig.~\ref{fig:iterations}(a) shows the evolution of the grid CRF loss over the number of iterations of training. ADM requires fewer iterations to achieve the same CRF loss. The networks trained using ADM scheme give lower CRF losses for both training and validation sets. \begin{table}[bt] \renewcommand{\arraystretch}{1.3} \centering \setlength\tabcolsep{5pt} \begin{tabular}{| l | c | c | c | c|} \hline \multirow{ 2}{*}{network} & \multicolumn{2}{|c|}{training set$^\dag$} & \multicolumn{2}{|c|}{validation set} \\ \cline{2-5} & GD & ADM & GD & ADM \\ \hline Deeplab-LargeFOV & 2.52 & 2.41 & 2.51 & 2.33 \\ \hline Deeplab-MSc-largeFOV & 2.51 & 2.40 & 2.49 & 2.33 \\ \hline Deeplab-VGG16 & 2.37 & 2.10 & 2.42 & 2.14 \\ \hline Resnet-101 & 2.66 & 2.49 & 2.61 & 2.42 \\ \hline \end{tabular} \vspace{1ex} \caption{ADM gives better grid CRF losses than gradient descend (GD). \dag We randomly selected 1,000 training examples.}\label{tb:gridCRFloss} \end{table} \begin{figure}[h] \centering \begin{tabular}{ccc} \includegraphics[width=0.45\textwidth]{fig/gridcrfloss.png} \\ \end{tabular} \caption{Training progress of ADM and gradient descend (GD) on Deeplab-MSc-largeFOV. Our ADM for the grid CRF loss with $\alpha$-expansion significantly improves convergence and achieves lower training loss. For example, first 1,000 iterations of ADM give grid CRF loss lower than GD's best result } \label{fig:iterations} \end{figure} \vspace{-2ex} \paragraph{The gradients} with respect to the soft-max layer's input of the network are visualized in Fig.~\ref{fig:gradients}. Clearly, our ADM approach with the grid CRF enforces better edge alignment. Despite different formulations of regularized losses and their optimization, the gradients of either~\eqref{CNN-regularized-loss} or~\eqref{SGD} \wrt network output $S_\theta$ are the driving force for training. In most of the cases, GD produces significant gradient values only in the vicinity of the current model prediction boundary as in Fig.~\ref{fig:gradients}(c,d). If the actual object boundary is sufficiently distant the gradient methods fail to detect it due to the sparsity of the grid CRF model, see Fig.~\ref{fig:synthetic} for an illustrative ``toy'' example. On the other hand, the ADM method is able to predict a good latent segmentation allowing gradients leading to a good solution more effectively, see Fig.~\ref{fig:gradients}(e). \begin{figure*}[t] \setlength{\tabcolsep}{1pt} \newcommand{0.156\textwidth}{0.156\textwidth} \centering \begin{tabular}{@{}ccccccc} \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/gradients/train/input.png}& \includegraphics[width=0.156\textwidth,trim=0 0 0 0,clip]{fig/gradients/train/prediction.png}& \includegraphics[width=0.156\textwidth,trim=2.75cm 1.35cm 4.2cm 1.51cm,clip]{fig/gradients/train/19scorediff_dense_grad} & \includegraphics[width=0.156\textwidth,trim=2.75cm 1.35cm 4.2cm 1.51cm,clip]{fig/gradients/train/19scorediff_sparse_grad} & \includegraphics[width=0.156\textwidth,trim=2.75cm 1.35cm 4.2cm 1.51cm,clip]{fig/gradients/train/19scorediff_sparse_adm} \\ \includegraphics[width=0.156\textwidth,clip,trim=0 0 0 5cm]{fig/gradients/sofa/input.png} & \includegraphics[width=0.156\textwidth,clip,trim=0 0 0 5cm]{fig/gradients/sofa/prediction.png} & \includegraphics[width=0.156\textwidth,clip,trim=0 0 0 5cm]{fig/gradients/sofa/18scorediff_dense_grad} & \includegraphics[width=0.156\textwidth,clip,trim=0 0 0 5cm]{fig/gradients/sofa/18scorediff_sparse_grad} & \includegraphics[width=0.156\textwidth,clip,trim=0 0 0 5cm]{fig/gradients/sofa/18scorediff_sparse_adm} \\ \includegraphics[width=0.156\textwidth,clip,trim=0 6cm 0 0]{fig/gradients/office/input.png} & \includegraphics[width=0.156\textwidth,clip,trim=0 6cm 0 0]{fig/gradients/office/prediction.png} & \includegraphics[width=0.156\textwidth,clip,trim=0 6cm 0 0]{fig/gradients/office/20scorediff_dense_grad} & \includegraphics[width=0.156\textwidth,clip,trim=0 6cm 0 0]{fig/gradients/office/20scorediff_sparse_grad} & \includegraphics[width=0.156\textwidth,clip,trim=0 6cm 0 0]{fig/gradients/office/20scorediff_sparse_adm} \\ \includegraphics[width=0.156\textwidth,clip,trim=0 0 0 5cm]{fig/gradients/boat/input.png} & \includegraphics[width=0.156\textwidth,clip,trim=0 0 0 5cm]{fig/gradients/boat/prediction.png} & \includegraphics[width=0.156\textwidth,clip,trim=0 0 0 5cm]{fig/gradients/boat/4scorediff_dense_grad} & \includegraphics[width=0.156\textwidth,clip,trim=0 0 0 5cm]{fig/gradients/boat/4scorediff_sparse_grad} & \includegraphics[width=0.156\textwidth,clip,trim=0 0 0 5cm]{fig/gradients/boat/4scorediff_sparse_adm} \\ \includegraphics[width=0.156\textwidth,clip,trim=0 0 0 2cm]{fig/gradients/car/input.png} & \includegraphics[width=0.156\textwidth,clip,trim=0 0 0 2cm]{fig/gradients/car/prediction.png} & \includegraphics[width=0.156\textwidth,clip,trim=0 0 0 2cm]{fig/gradients/car/7scorediff_dense_grad} & \includegraphics[width=0.156\textwidth,clip,trim=0 0 0 2cm]{fig/gradients/car/7scorediff_sparse_grad} & \includegraphics[width=0.156\textwidth,clip,trim=0 0 0 2cm]{fig/gradients/car/7scorediff_sparse_adm} \\ \includegraphics[width=0.156\textwidth]{fig/gradients/bike/input.png} & \includegraphics[width=0.156\textwidth]{fig/gradients/bike/prediction.png} & \includegraphics[width=0.156\textwidth]{fig/gradients/bike/2scorediff_dense_grad} & \includegraphics[width=0.156\textwidth]{fig/gradients/bike/2scorediff_sparse_grad} & \includegraphics[width=0.156\textwidth]{fig/gradients/bike/2scorediff_sparse_adm} \\ (a) input & (b) prediction &(c) Dense GD\cite{Rloss:ECCV18} & (d) Grid GD & (e) Grid ADM \end{tabular} \caption{The gradients with respect to scores of the deeplab\_largeFOV network with the dense CRF (c) and grid CRF (d~and~e for using either the plain stochastic gradient descent or our ADM scheme). Latent segmentation in ADM with the grid CRF loss produces gradients more directly pointing to a good solution (e). Note, the object boundaries are more prominent in (e).} \label{fig:gradients} \end{figure*} Thus, in the context of grid CRFs, the ADM approach coupled with $\alpha$-expansion shows drastic improvement in the optimization quality. In the next section, we further compare ADM with GD to see which gives better segmentation. \subsection{Segmentation Quality} \label{sec:mIOU} The quantitative measures for segmentation by different methods are summarized in Tab.~\ref{tab:mainresult} and Tab.~\ref{tab:accuracy}. The mIOU and segmentation accuracy on the \textit{val} set of PASCAL 2012~\cite{pascal-voc-2012} are reported for various networks. The supervision is scribbles~\cite{scribblesup}. The quality of weakly supervised segmentation is bounded by that with full supervision and we are interested in the gap for different weakly supervised approaches. The baseline approach is to train the network using proposals generated by GrabCut style interactive segmentation with such scribbles. Besides the baseline (train w/ proposals), here we compare variants of regularized losses optimized by gradient descent or ADM. The regularized loss is comprised of the partial cross entropy (pCE) \wrt scribbles and grid/dense CRF. Other losses \eg normalized cut \cite{Shi2000,NCloss:CVPR18} may give better segmentation, but the focus is to compare gradient descent vs ADM optimization for the grid CRF. It is common to apply dense CRF post-processing~\cite{deeplab} to the network's output during testing. However, for the sake of clear comparison, we show results without it. As shown in Tab.~\ref{tab:mainresult}, all regularized approaches work better than the non-regularized approach that only minimizes the partial cross entropy. Also, the regularized loss approaches are much better than proposal generation based method since erroneous proposals may mislead training. Among regularized loss approaches, grid CRF with GD performs the worst due to the fact that a first-order method like gradient descent leads to the poor local minimum for the grid CRF in the context of energy minimization. Our ADM for the grid CRF gives much better segmentation competitive with the dense CRF with GD. The alternative grid CRF based method gives good quality segmentation approaching that for full supervision. Tab.~\ref{tab:accuracy} shows \emph{accuracy} of different methods for pixels close to the semantic boundaries. Such measure tells the quality of segmentation in boundary regions. Fig.~\ref{fig:segexamples} shows a few qualitative segmentation results. \begin{table*}[t] \renewcommand{\arraystretch}{1.3} \centering \begin{tabular}{| l | c| c | c |c| c |c|} \hline & &\multicolumn{5}{|c|}{Weak supervision} \\ \cline{3-7} {Network}& Full sup. &\multirow{ 2}{12mm}{train w/ proposals} &\multirow{ 2}{10mm}{\centering \centering pCE loss} & \multicolumn{1}{|c|}{+dense CRF loss} & \multicolumn{2}{|c|}{\cellcolor{gray!25}+grid CRF loss} \\ \cline{5-7} & & & & GD \cite{Rloss:ECCV18} & {\cellcolor{gray!25} GD} & {\cellcolor{gray!25} ADM} \\ \hline Deeplab-largeFOV & 63.0 & 54.8&55.8 & {\bf62.2}&{\cellcolor{gray!25} 60.4}&{\cellcolor{gray!25} 61.7} \\ \hline Deeplab-MSc-largeFOV & 64.1 & 55.5 & 56 & {\bf63.1} & {\cellcolor{gray!25} 61.2} & {\cellcolor{gray!25}62.9} \\ \hline Deeplab-VGG16 & 68.8 &59.0& 60.4& 64.4 & {\cellcolor{gray!25} 63.3} & {\cellcolor{gray!25} {\bf65.2}} \\ \hline ResNet-101 & 75.6 &64.0& 69.5& {\bf72.9} & {\cellcolor{gray!25} 71.7 } & {\cellcolor{gray!25} 72.8} \\ \hline \end{tabular} \vspace{1ex} \caption{Weakly supervised segmentation results for different choices of network architecture, regularized losses and optimization via gradient descent or ADM. We show mIOU on \textit{val} set of PASCAL 2012. ADM consistently improves over GD for different networks for grid CRF. Our grid CRF with ADM is competitive to previous state-of-the-art dense CRF (with GD) \cite{Rloss:ECCV18}.} \label{tab:mainresult} \end{table*} \begin{table*}[h] \renewcommand{\arraystretch}{1.3} \centering \begin{tabular}{l| l | c| c | c |c| c |c|} \cline{2-8} & & &\multicolumn{5}{|c|}{Weak supervision} \\ \cline{4-8} & {Network}& Full sup. &\multirow{ 2}{12mm}{train w/ proposals} &\multirow{ 2}{10mm}{\centering \centering pCE loss} & \multicolumn{1}{|c|}{+dense CRF loss} & \multicolumn{2}{|c|}{\cellcolor{gray!25}+grid CRF loss} \\ \cline{6-8} & & & & & GD \cite{Rloss:ECCV18} & \cellcolor{gray!25}GD & \cellcolor{gray!25}ADM \\ \cline{2-8} \multirow{ 3}{*}{\begin{sideways}\parbox{10ex}{\centering all\\pixels}\end{sideways}} & Deeplab-MSc-largeFOV & 90.9 & 86.4 & 86.5 & {\bf90.6} & \cellcolor{gray!25}89.9 & \cellcolor{gray!25}90.5 \\ \cline{2-8} & Deeplab-VGG16 & 91.6 & 88.6&88.9 & 91.1 & \cellcolor{gray!25}90.5 & \cellcolor{gray!25}{\bf91.3} \\ \cline{2-8} & ResNet-101 & 94.5 & 90.2&92 &93.1 & \cellcolor{gray!25}92.9 & \cellcolor{gray!25}{\bf93.4} \\ \hline\hline \multirow{ 3}{*}{\begin{sideways}\parbox{10ex}{\centering trimap\\16 pixels}\end{sideways}} & Deeplab-MSc-largeFOV & 80.1 & 73.9 & 66.7 & {\bf77.8} & \cellcolor{gray!25}74.8 & \cellcolor{gray!25}76.7 \\ \cline{2-8} & Deeplab-VGG16 & 81.9 & 75.5&70.9 & 77.8 & \cellcolor{gray!25}75.6 & \cellcolor{gray!25}{\bf78.1} \\ \cline{2-8} & ResNet-101 & 85.7 & 78.4&77.7 &82.0 & \cellcolor{gray!25}80.6 & \cellcolor{gray!25}{\bf82.2} \\ \hline\hline \multirow{ 3}{*}{\begin{sideways}\parbox{10ex}{\centering trimap\\8 pixels}\end{sideways}} & Deeplab-MSc-largeFOV & 75.0 & 69.5 & 60.3 & {\bf72.5} & \cellcolor{gray!25}68.4&\cellcolor{gray!25}71.4 \\ \cline{2-8} & Deeplab-VGG16 & 76.9 & 70.4&64.1 & 72.0 & \cellcolor{gray!25}69.0 & \cellcolor{gray!25}{\bf72.4} \\ \cline{2-8} & ResNet-101 & 81.5 & 73.8&71.2 &76.7 & \cellcolor{gray!25}74.6 & \cellcolor{gray!25}{\bf77.0} \\ \cline{2-8} \end{tabular} \vspace{1ex} \caption{Pixel-wise accuracy on \textit{val} set of PASCAL 2012. Top 3 rows: accuracy over all pixels. Middle 3 rows: accuracy for pixels within 16 pixels away from semantic boundaries. Bottom 3 rows: accuracy for pixels within 8 pixels aways from semantic boundaries. Pixels closer to boundaries are more likely to be mislabeled. Our ADM scheme improves over GD for grid CRF loss consistently for different networks. Note that weak supervision with our approach is almost as good as full supervision.} \label{tab:accuracy} \end{table*} \subsection{Shortened Scribbles} \label{sec:shortenedscribbles} \vspace{-1ex} Following the evaluation protocol in ScribbleSup \cite{scribblesup}, we also test our regularized loss approaches training with shortened scribbles. We shorten the scribbles from the two ends at certain ratios of length. In the extreme case, scribbles degenerate to clicks for semantic objects. We are interested in how the weakly-supervised segmentation methods degrade as we reduce the length of the scribbles. We report both mIOU and pixel-wise accuracy. As shown in Fig.~\ref{fig:shortscribbles}, our ADM for the grid CRF loss outperforms all competitors giving significantly better mIOU and accuracy than GD for the grid CRF loss. ADM degrades more gracefully than the dense CRF as the supervision weakens. The grid CRF has been overlooked in regularized CNN segmentation currently dominated by the dense CRF as either post-processing or trainable layers. We show that for weakly supervised CNN segmentation, the grid CRF as the regularized loss can give segmentation at least as good as that with the dense CRF. The key to minimizing the grid CRF loss is better optimization via ADM rather than gradient descent. Such competitive results for the grid CRF loss confirm that it has been underestimated as a loss regularizer for neural network training, as discussed in Sec.~\ref{sec:intro}. It has not been obvious whether the grid CRF as a loss is beneficial for CNN segmentation. We show that straightforward gradient descent for the grid CRF does not work well. Our technical contribution on optimization helps to reveal the limitation and advantage of the grid CRF vs dense CRF models. The weaker regularization properties, as discussed in Sec.~\ref{sec:shallow}, of the dense CRF and our experiments favors the grid CRF regularizer compared to the dense CRF.

8d6e09356a128107ba0f0cc752921974694f1f4c

\section{Introduction} Symbolic powers of ideals have been studied intensely over the last two decades (see \cite{SurveySymbPowers} for a recent survey). Particular attention has been given to square-free monomial ideals, as in this setting methods from combinatorics, convex geometry, and linear optimization can be utilized to investigate properties of symbolic powers. For instance, the Cohen-Macaulay property of all symbolic powers of a square-free monomial ideal is characterized in terms of the combinatoric structure of its underlying simplicial complex \cite{TT,TT2,Varbaro}. In addition, the symbolic Rees algebra of monomial ideals is Noetherian \cite{LyubeznikAriRank, herzoga2007symbolic}. We note that this phenomenon does not hold for an arbitrary ideal in a polynomial ring \cite{roberts1985prime}. In this article, we propose a technique to deal with questions about symbolic powers of square-free monomial ideals. Specifically, we study the symbolic Rees algebras of an square-free monomial ideal via methods in prime characteristic. This combination, to the best of our knowledge, has not been previously used in combinatorial commutative algebra. In order for our results to hold over fields of arbitrary characteristic, we consider the map that raises every monomial to a power and resembles the Frobenius map. Considering this map, we obtain that the symbolic Rees algebra and the symbolic associated graded algebra are split in this general context (see Theorem \ref{ThmGradedSpl}). In particular, these symbolic algebras are $F$-pure in prime characteristic (see Corollary \ref{Fpure}). Motivated by the behavior of the Castelnuovo-Mumford regularity for powers of ideals \cite{CHT,KV}, Herzog, Hoa, and Trung \cite{HHT} asked whether the limit $$\lim\limits_{n\to\infty}\frac{\reg(R/I^{(n)})}{n}$$ exists for every homogeneous ideal $I$ in a polynomial ring $R$. It is known that the function $\reg(R/I^{(n)})$ is bounded by a linear function on $n$ if $I$ is a monomial ideal. This follows because the regularity of a monomial ideal is bounded by the degree of the least common multiple of the generators \cite{BHMultRes,HTMathZ} Hoa and Trung showed that the limit above exists for square-free monomial ideals. In fact, they showed a stronger version for the $a$-invariants \cite[Theorems 4.7 and 4.9]{LT10}. As a first consequence of our methods, we recover this result, providing an alternative proof. \begin{theoremx}[{see Theorem \ref{limAinv} and Corollary \ref{limReg}}]\label{MainLim} Let $I$ be a square-free monomial ideal. Then $$ \lim\limits_{n\to\infty}\frac{a_i(R/I^{(n)})}{n} $$ exists for every $0\leqslant i\leqslant \dim R/I$. In particular, $$ \lim\limits_{n\to\infty}\frac{\reg(R/I^{(n)})}{n} $$ exists. \end{theoremx} The function $\reg(R/I^{(n)})$ is in fact a linear quasi-polynomial if $I$ is a monomial ideal \cite{HHT}. In addition, as a consequence of our previous result, we recover properties of this quasi-polynomial in Corollary \ref{CorQP}. Along the way towards proving Theorem \ref{MainLim}, we showed that $$a_i(R/I^{(n)})\geqslant m a_i(R/I^{(\lceil\frac{n}{m}\rceil)})$$ for every $n,\, m\in \mathbb Z_{>0}$. In addition, we showed that $$\depth(R/I^{(n)})\leqslant \depth(R/I^{(\lceil\frac{n}{m}\rceil)})$$ for every $n,\, m\in \mathbb Z_{>0}$ (see Theorem \ref{ThmDivision}). Similar results were previously obtained for the Stanley depth of symbolic powers \cite{StanleyDepth}. Conforti and Cornu\'ejols \cite{CC} made a conjecture in the context of linear optimization. This conjecture was translated as a characterization of the set of square-free monomials ideals whose symbolic and ordinary powers are equal \cite {GRV,GVV}. The following definition is needed to state this conjecture. \begin{definitionx} \rm A square-free monomial ideal $I$ of height $c$ is \it K\"onig \rm if there exists a regular sequence of monomials in $I$ of length $c$. The ideal $I$ is said to be \it packed \rm if every ideal obtained from $I$ by setting any number of variables equal to $0$ or $1$ is K\"onig. \end{definitionx} The Conforti-Cornu\'ejols conjecture can be stated as follows. \begin{conjecturex}[\cite{CC}]\label{ConjPackingProb} A square-free monomial ideal $I$ is packed if and only if $I^n=I^{(n)}$ for every $n\in\mathbb Z_{>0}.$ \end{conjecturex} We point out that one direction of the conjecture is already known. Explicitly, if $I^n=I^{(n)}$ for every $n\in\mathbb Z_{>0}$, then $I$ is packed. A related question due to Huneke, and brought to our attention by H\`a\footnote{ BIRS-CMO workshop on {\it Ordinary and Symbolic Powers of Ideals} Summer of 2017, Casa Matem\'atica Oaxaca, Mexico.}, asks whether there exists a number $N$, in terms of $I$, such that if $I^n=I^{(n)}$ for $n\leqslant N$, then the equality holds for every $n\geqslant 1$. If the answer to this question is $\height(I)$, then Conforti-Cornu\'ejols holds. This is expected as a similar property is known for integral closures of ordinary powers. Specifically, it suffices to verify the equality up to the analytic spread of $I$ minus one \cite{Singla} (see also \cite{RRV}). A uniform positive answer to Huneke's question can be deduced from \cite[Theorem 5.6]{herzoga2007symbolic}; if $d=\dim R$, this result guarantees that $N=\lfloor\frac{(d+1)^{(d+3)/2}}{2^d}\rfloor$ works for any $I$. In our next main result, we provide a sharper $N$. \begin{theoremx}[{see Theorem \ref{allPowMon}}]\label{MainAllPowMon} Let $I$ be a square-free monomial ideal and $\mu(I)$ its minimal number of generators. Then $I^n=I^{(n)}$ for every $n\leqslant \lceil\frac{ \mu(I)}{2}\rceil$ if and only if $I^n=I^{(n)}$ for every $n\in\mathbb Z_{>0}$. \end{theoremx} In Example \ref{ExSharp} we show that the bound given in Theorem \ref{MainAllPowMon} is sharp. The previous result gives a finite algorithm to verify Conjecture \ref{ConjPackingProb} for a specific square-free monomial ideal. We refer to the work of Gitler, Valencia, and Villarreal \cite[Remark 3.5]{GVV} for a different algorithm to verify Conjecture \ref{ConjPackingProb}. The referee pointed out that another finite algorithm can be obtained from an equivalent conjecture to Conjecture \ref{ConjPackingProb} \cite[Conjecture 5.7]{HaTrung}. Indeed, this conjecture states that if a clutter satisfies the {\it packing property} then it satisfies the {\it integer round-down property}. On the other hand, the integer round-down property can be checked in a finite number of steps \cite[Corollary 2.3]{BT82}. In Section \ref{SecOptimization}, we recall the ideas from linear optimization that originally gave rise to Conjecture \ref{ConjPackingProb}. In particular, we translate Theorem \ref{MainAllPowMon} to this context in Theorem \ref{MFMC}, showing that the Max-Flow-Min-Cut property of clutters can be verified with a finite process (see \cite[Corollary 2.3]{BT82} for a related result). \section{Notation}\label{prelim} In this section we set up the notation used throughout the entire manuscript. We assume $R=k[x_1,\ldots, x_d]$ is a standard graded polynomial ring over the field $k$ and $\mathfrak{m}=(x_1,\ldots,x_d)$. The ideal $I\subseteq R$ is assumed to be a monomial ideal. For a fixed $m\in\mathbb Z_{>0}$ we set $R^{1/m}=k[x^{1/m}_1,\ldots, x^{1/m}_d]$. We denote by $I^{1/m}$ the ideal of $R^{1/m}$ generated by $\{f^{1/m}\mid f\in I \hbox{ monomial} \} $. We note that $R$ and $R^{1/m}$ are isomorphic as rings. Then the category of $R$-modules is naturally equivalent to the category of $R^{1/m}$-modules. In addition, $I$ corresponds to $I^{1/m}$ under this isomorphism. For an $R$-module, $M$, we denote by $M^{1/m}$ the corresponding $R^{1/m}$-module. Given that $R\subseteq R^{1/m}$, $M^{1/m}$ obtains a structure of $R$-module via restriction of scalars. If $A=k[x_1,\ldots,x_d,t]$, where $t$ is an extra variable, we also use the notation $A^{1/m}=k[x^{1/m}_1,\ldots, x^{1/m}_d,t^{1/m}]$. We often refer to the containment $R\subseteq R^{1/m}$ and $A\subseteq A^{1/m}$. For a vector $\mathbf{a}=(a_1,\ldots, a_d)\in \mathbb Q^d_{\geqslant 0}$ we denote by $\mathbf{x}^\mathbf{a}$ the monomial $x_1^{a_1}\cdots x_d^{a_d}$. \begin{definition} Given $n\in \mathbb Z_{>0}$, we denote by $I^{(n)}$ the {\it {$n$th}-symbolic power of $I$}: $$I^{(n)}=\bigcap_{\mathfrak{p}\in \Min(R/I)}I^nR_\mathfrak{p}\cap R.$$ \end{definition} We now consider algebras associated to ordinary and symbolic powers of ideals. \begin{definition} \label{blowup} We consider the following graded algebras. \begin{enumerate} \item[(i)] The {\it Rees algebra} of $I$: $\mathcal R(I)=R[It]=\oplus_{n\geqslant 0}I^nt^n\subseteq A$. \item[(ii)] The {\it associated graded algebra} of $I$: $\gr_I(R)=\oplus_{n\geqslant 0}I^n/I^{n+1}$. \item[(iii)] The {\it symbolic Rees algebra} of $I$: $\mathcal R^s(I)=\oplus_{n \geqslant 0}I^{(n)}t^n\subseteq A$. \item[(iv)] The {\it symbolic associated graded algebra} of $I$: $\gr^s_I(R)=\oplus_{n\geqslant 0}I^{(n)}/I^{(n+1)}$. \end{enumerate} \end{definition} \section{Castelnuovo-Mumford regularity and $a$-invaratians} In this section we study the graded structure of symbolic powers. In particular, we prove Theorem \ref{MainLim}. The techniques here are inspired by methods in prime characteristic used to bound $a$-invariants of $F$-pure graded rings \cite{HRFpurity,DSNBFpurity}. In particular, the results proved in this section are motivated by the fact that the symbolic Rees and associated graded algebra are $F$-pure for every prime characteristic (see Corollary \ref{Fpure}). We now recall the definition of $a$-invariants and Castelnuovo-Mumford regularity in terms of local cohomology. We refer to Broadmann and Sharp's book \cite{BroSharp} on local cohomology for more details about this subject. If $M$ is a graded $R$-module, we denote by $$a_i(M)=\max\{j\mid \HH{i}{\mathfrak{m}}{M}_j\neq 0\}$$ the {\it $i$-th $a$-invariant} of $M$ for $0\leqslant i\leqslant \dim M$. The {\it Castelnuovo-Mumford regularity} of $M$ is defined as $$\reg(M)=\max\{a_i(M)+i\}.$$ We now define a splitting from $R^{1/m}$ to $R$ which is inspired by the trace map in prime characteristic. In fact, these maps are the same if $k=\mathbb F_p$ and $m=p$. \begin{definition}\label{defSplitting} For $m\in \mathbb Z_{>0}$, we define the $R$-homomorphism $\Phi^R_m:R^{1/m}\to R$ induced by $$ \Phi^R_m(\mathbf{x}^{\mathbf{a}/m})= \begin{cases} \mathbf{x}^{\mathbf{a}/m} & \mathbf{a} \equiv \mathbf{0}\, (\bmod\, m);\\ 0 & \hbox{otherwise,} \end{cases}. $$ \end{definition} \noindent where $\mathbf{a}\in \mathbb Z_{\geqslant 0}^d$ and $\mathbf{0}=(0,\ldots, 0)\in \mathbb Z_{\geqslant 0}^d$. We note that $\Phi^R_m$ restricted to $R$ is the identity. Then $R$ is isomorphic to a direct summand of $R^{1/m}$. We also consider the analogous map $\Phi^A_m:A^{1/m}\to A$. We now show that our splitting is compatible with symbolic powers; for this we need the following remark. \begin{remark}\label{inclusionSymbol} Let $I\subseteq R$ be a square-free monomial and $Q_1,\ldots, Q_s$ its minimal primes. We have \begin{align*} I^{(n)}=Q^n_1\cap \ldots \cap Q^n_s &\subseteq (Q^{nm}_1)^{1/m}\cap \ldots \cap (Q^{nm}_s)^{1/m} \\ &=(Q^{nm}_1\cap \ldots \cap Q^{nm}_s)^{1/m}=(I^{(nm)})^{1/m}. \end{align*} \end{remark} \begin{lemma}\label{LemmaStable} Let $I$ be a square-free monomial ideal. Then $$\Phi_m^R\big(( I^{(n m+j)})^{1/m}\big)= I^{(n+1)}$$ for every $n\in \mathbb Z_{\geqslant 0}$, $m\in \mathbb Z_{>0}$, and $1\leqslant j\leqslant m$. \end{lemma} \begin{proof} From Remark \ref{inclusionSymbol} it follows that $ I^{(n+1)}\subseteq ( I^{((n+1) m)})^{1/m} \subseteq (I^{(n m+j)})^{1/m}$. Since $\Phi_m^R$ is $R$-linear, we conclude that $I^{(n+1)}\subseteq \Phi_m^R\big(( I^{(n m+j)})^{1/m}\big)$. We now focus on the other containment. First we prove our claim when $I$ is a prime monomial ideal $Q=(x_{i_1},\ldots,x_{i_\ell})$ for some $1\leqslant i_1<\cdots <i_\ell\leqslant d$. In this case, we have that $( Q^{(n m+j)})^{1/m}$ is generated as a $k$-vector space by $$\{ (\mathbf{x}^{\mathbf{a}})^{1/m}\mid a_{i_1}+\ldots+a_{i_\ell}\geqslant nm+j \} $$ Let $(\mathbf{x}^{\mathbf{a}})^{1/m} \in ( Q^{(n m+j)})^{1/m}$ such that $\Phi_m^R( (\mathbf{x}^{\mathbf{a}})^{1/m})\neq 0$, then $\mathbf{a}/m\in \mathbb Z_{\geqslant 0}^d$ and $a_{i_1}+\ldots+a_{i_\ell}\geqslant nm+j$. Set $\mathbf{b}=(b_{i_1},\ldots, b_{i_\ell})=\mathbf{a}/m$, since $ b_{i_1}+\ldots+b_{i_\ell}\geqslant n+\frac{j}{m}$, we have $ b_{i_1}+\ldots+b_{i_\ell}\geqslant n+1$ and hence $\Phi_m^R((\mathbf{x}^{\mathbf{a}})^{1/m} )\in Q^{(n+1)}$ as desired. Now, let $I$ be an arbitrary square-free monomial ideal and let $Q_1,\ldots,Q_s$ be its minimal primes. Then, $ ( I^{(n m+j)})^{1/m} =( Q_1^{(n m+j)})^{1/m} \cap \cdots\cap ( Q_s^{(n m+j)})^{1/m}. $ Therefore, \begin{align*} \Phi_m^R\big( (I^{(n m+j)})^{1/m}\big) &\subseteq \Phi_m^R\big( ( Q_1^{(n m+j)}\big)^{1/m} \big)\cap \cdots\cap \Phi_m^R\big(( Q_s^{(n m+j)})^{1/m} \big)\\ &\subseteq Q^{(n+1)}_1\cap\cdots\cap Q^{(n+1)}_s=I^{(n+1)}, \end{align*} hence the result follows. \end{proof} As a consequence of the previous lemma we obtain the following relations on depths and $a$-invariants of symbolic powers; these relations are key ingredients in the proof of Theorem \ref{MainLim}. \begin{thm}\label{ThmDivision} Let $I$ be a square-free monomial ideal and $n,\, m\in \mathbb Z_{>0}$. Then \begin{enumerate} \item $\depth(R/I^{(n)})\leqslant \depth(R/I^{(\lceil\frac{n}{m}\rceil)})$. \item $a_i(R/I^{(n)})\geqslant m a_i(R/I^{(\lceil\frac{n}{m}\rceil)})$ for every $0\leqslant i\leqslant \dim R/I$. \end{enumerate} \end{thm} \begin{proof} By Lemma \ref{LemmaStable} the natural map $\iota :R/I^{(n+1)}\to R^{1/m}/(I^{(nm+j)})^{1/m}$ splits for every $n,\, m\in \mathbb Z_{>0} $ and $1\leqslant j\leqslant m$ with the splitting $\Phi_m^R$. Therefore, the module $\HH{i}{\mathfrak{m}}{R/I^{(n+1)}}$ is a direct summand of $\HH{i}{\mathfrak{m}}{R^{1/m}/(I^{(nm+j)})^{1/m}}$ for every $1\leqslant i\leqslant\dim R/I$. We note that \begin{equation}\label{eqds} \left( \HH{i}{\mathfrak{m}}{ R/I^{(nm+j)}\right)}^{1/m}= \HH{i}{\mathfrak{m}^{1/m}}{R^{1/m}/(I^{(nm+j)})^{1/m}} =\HH{i}{\mathfrak{m}}{R^{1/m}/(I^{(nm+j)})^{1/m}}, \end{equation} because the ring isomorphism $R\cong R^{1/m}$ gives an equivalence of categories and $\mathfrak{m} R^{1/m}$ is $\mathfrak{m}^{1/m}$. We conclude that $\HH{i}{\mathfrak{m}}{R/(I^{(nm+j)})}=0$ implies $\HH{i}{\mathfrak{m}}{R^{1/m}/(I^{(nm+j)})^{1/m}}=0$, and then $\HH{i}{\mathfrak{m}}{R/I^{(n+1)}}=0$. Therefore, $$\depth(R/I^{(n+1)})\geqslant \depth( R/(I^{(nm+j)})),$$ which proves the first part. From the splitting $\iota$ and Equation \eqref{eqds}, we have $$ a_i(R/I^{(n+1)})\leqslant a_i \left( R^{1/m}/(I^{(nm+j)})^{1/m}\right)=\frac{1}{m} a_i \left( R/I^{(nm+j)}\right), $$ and the second part follows. \end{proof} As a consequence of Theorem \ref{ThmDivision} we recover the following limits for $a$-invariants of symbolic powers. \begin{thm}[{\cite[Theorem 4.7]{LT10}}]\label{limAinv} Let $I$ be a square-free monomial ideal. Then $$ \lim\limits_{n\to\infty}\frac{a_i(R/I^{(n)})}{n} $$ exists for every $0\leqslant i\leqslant \dim R/I$. \end{thm} \begin{proof} Fix $i$. The sequence $\{\frac{\reg(R/I^{(n)})}{n}\}_{n\geqslant 0}$ has an upper bound \cite[Corollary 3.3]{herzoga2007symbolic}, then so does $\{\frac{a_i(R/I^{(n)})}{n}\}_{n\geqslant 0}$. Set $\alpha_n = \frac{a_i (R/I^{(n)})}{n}$ for every $n\geqslant 0$ and $\alpha=\sup\{ \alpha_n\}$. If $\alpha=-\infty$, we have that $\alpha_n=-\infty$ for every $n$ and the claim follows. We now assume that $\alpha\neq -\infty$ and show that $\lim\limits_{n\to\infty}\alpha_n=\alpha.$ We note that Theorem \ref{ThmDivision}(2) implies \begin{equation}\label{useIneq} \alpha_n\geqslant \frac{m}{n}\left\lceil\frac{n}{m}\right\rceil\alpha_{\lceil\frac{n}{m}\rceil}\qquad \text{ for every }n,m\in \mathbb Z_{>0}. \end{equation} Fix $\epsilon\in \mathbb R_{>0}$ and let $t\in \mathbb Z_{\geqslant 0}$ be such that $\alpha-\alpha_t<\epsilon.$ It suffices to show $\alpha_n> \alpha_t$ for every $n\geqslant t^2$ as this implies that $\alpha-\alpha_n<\epsilon$. Let $n = t q + r$ for some $q, r\in \mathbb Z_{\geqslant 0}$ and $r\in [0,t-1]$, then $n=(q+1)t-(t-r)$. Since $n\geqslant t^2$, we obtain $q\geqslant t$ and then, $0< t-r\leqslant q-r<q+1,$ and so $\lceil\frac{n}{q+1}\rceil=t.$ Applying inequality \eqref{useIneq} with $m=q+1$ we have $\alpha_n\geqslant \frac{tq+t}{tq+r}\alpha _t>\alpha_t$, which finishes the proof. \end{proof} As a corollary we obtain that the related limit for the Castelnuovo-Mumford regularity of symbolic powers exists. \begin{cor}[{\cite[Theorem 4.9]{LT10}}]\label{limReg} Let $I$ be a square-free monomial ideal. Then $$ \lim\limits_{n\to\infty}\frac{\reg(R/I^{(n)})}{n}=\max\left\{\lim\limits_{n\to\infty}\frac{a_i(R/I^{(n)}) }{n}\right\}. $$ \end{cor} \begin{proof} The result follows from Theorem \ref{limAinv} and the following equalities \begin{align*} \lim\limits_{n\to\infty}\frac{\reg(R/I^{(n)})}{n} &= \lim\limits_{n\to\infty}\frac{ \max\{a_i(R/I^{(n)}) + i\}}{n} \\ &= \lim\limits_{n\to\infty}\frac{ \max\{a_i(R/I^{(n)}) \}}{n} = \max\left\{\lim\limits_{n\to\infty}\frac{a_i(R/I^{(n)}) }{n}\right\}. \end{align*} \end{proof} \begin{remark}\label{RemWC} If $I$ is a square-free monomial ideal, then $$ \lim\limits_{n\to\infty}\frac{\reg(I^{(n)})}{n}=\lim\limits_{n\to\infty}\frac{\reg(R/I^{(n)})+1}{n}=\lim\limits_{n\to\infty}\frac{\reg(R/I^{(n)})}{n}.$$ Let $\alpha(I^{(n)})$ denote the smallest degree of a nonzero element of $I^{(n)}$. The {\it Waldschmidt constant} is defined by $\widehat{\alpha}(I)=\lim\limits_{n\to\infty}\frac{\alpha(I^{(n)}) }{n}$ \cite{BocciEtAl}. Then $$ \widehat{\alpha}(I)\leqslant \lim\limits_{n\to\infty}\frac{\reg(I^{(n)})}{n}=\lim\limits_{n\to\infty}\frac{\reg(R/I^{(n)})}{n}. $$ \end{remark} We recall that $\mathcal R^s(I)$ is a Noetherian algebra \cite[Theorem 3.2]{herzoga2007symbolic}. Therefore, $\reg(I^{(n)})$ agrees with a linear quasi-polynomial $\gamma(n) n + \theta(n)$ for $n\gg 0$. As a consequence of Corollary \ref{limReg} we obtain that the leading coefficient of this quasi-polynomial is constant for square-free monomial ideals and we obtain a bound for $\theta(n)$. \begin{cor}\label{CorQP}\cite[Theorem 4.9]{LT10} Let $I$ be a square-free monomial ideal. Then $\gamma(n) $ is equal to a constant $\gamma$ and $\theta(n)\leqslant \dim(R/I)+1$ for every $n\gg 0$. \end{cor} \begin{proof} From Corollary \ref{limReg} it follows that $\gamma(n)$ must be equal to $\gamma:=\sup\{\frac{a_i(R/I^{(n)})}{n}\}$ for $n\gg 0$. For the second claim, we observe that $$ \gamma n + \theta(n)=\reg(R/I^{(n)})+1= \max\{a_i(R/I^{(n)})+i)\}+1\leqslant \gamma n+\dim(R/I)+1. $$ for $n\gg 0$. We conclude that $\theta(n)\leqslant \dim(R/I)+1$ for every $n\gg 0$. \end{proof} The previous results, together with results for matroids \cite{RegSymbMatroit} and low dimension \cite{RegSymbDim2}, motivated Minh and Trung \cite{RegSymbMatroit} to ask wheather $\reg(R/I^{(n)})$ is a linear polynomial for $n\gg 0$ for square-free monomial ideals. We note that very recently a counter-example to this question was found by Dung, Hien, Nguyen, and Trung \cite[Remark 5.16]{DHNT}. It is a classical result that for any homogeneous ideal $I$, $\reg(I^n)$ is a linear function $b(I)n+c(I)$ for $n\gg 0$ \cite{CHT,KV}. In general, not much is known about the invariant $c(I)$ besides the fact that it is non-negative \cite[Corollary 3.3]{trung2005asymptotic}. Corollary \ref{CorQP} provides an upper bound for $c(I)$ for a wide family of square-free monomial ideals. \begin{cor} Let $I$ be a square-free monomial ideal such that $I^n=I^{(n)}$ for every $n\gg 0$, then $c(I)\leqslant \dim (R/I)+1$. In particular, this holds for bipartite edge ideals. \end{cor} \begin{proof} The result follows by the assumption and Corollary \ref{CorQP}. The case of edge ideals follows because they satisfy $I^n=I^{(n)}$ for every $n\gg 0$ \cite[Theorem 5.9]{SVV}. \end{proof} \section{Associated Graded Algebras and Equality of Symbolic and Ordinary Powers} In this section, we study the graded algebras defined in Definition \ref{blowup}. This is in order to prove Theorem \ref{MainAllPowMon}. Our strategy is the following. We first show that the symbolic Rees and associated graded algebras split from their rings of $m$-roots in Theorem \ref{ThmGradedSpl}. Then in Theorem \ref{MainSQfree}, we characterize the equality of symbolic and ordinary powers in terms of this splitting. Finally, we use this characterization to prove Theorem \ref{MainAllPowMon}. We start with introducing rings of $m$-roots for the Rees and associated graded algebras. \begin{notation}\label{notBlw} For $m\in \mathbb Z_{>0}$, we set $$\mathcal R(I)^{1/m}:= \oplus_{n\geqslant 0} I^{n/m} t^{n/m}\subseteq A^{1/m}$$ and $$\mathcal R^s(I)^{1/m}:= \oplus_{n\geqslant 0} (I^{(n)})^{1/m} t^{n/m}\subseteq A^{1/m}.$$ We consider the ideals $$\mathcal J(I)=\oplus_{n\geqslant 0} I^{n+1}t^n\subseteq \mathcal R(I)$$ and $$\mathcal J^s(I)=\oplus_{n\geqslant 0} I^{(n+1)}t^n\subseteq \mathcal R^s(I).$$ \end{notation} A classical result states that $\gr_I(R)$ is reduced if and only if $I^n=I^{(n)}$ for every $n\in\mathbb Z_{>0}$ \cite[Corollary 1.6]{herzog2008standard}. As a consequence of this result, and its proof, one has that $\mathcal J^s(I)$ is a radical ideal. We set $$\gr^s_{I} (R)^{1/m} = (\mathcal R^s(I)/\mathcal J^s(I))^{1/m}.$$ \begin{remark} \ \begin{enumerate} \item By Remark \ref{inclusionSymbol} we have the inclusion $$\mathcal R^s(I)=\oplus_{n\geqslant 0} I^{(n)} t^n \subseteq \oplus_{n\geqslant 0} \left( I^{(n)}\right)^{1/m} t^{n/m}= \mathcal R^s(I)^{1/m}.$$ \item Since $I\subseteq R$ is monomial, we have that $I^{n}\subseteq (I^{nm})^{1/m}$. Then $$\mathcal R(I)=\oplus_{n\geqslant 0} I^n t^n \subseteq \oplus_{n\geqslant 0} I^{n/m} t^{n/m}=\mathcal R(I)^{1/m}.$$ \end{enumerate} \end{remark} \begin{thm}\label{ThmGradedSpl} Let $I$ be a square-free monomial ideal. Then the maps induced by the observation in Remark \ref{inclusionSymbol} $$\mathcal R^s(I)\to \mathcal R^s(I)^{1/m}\quad\text{ and }\quad\gr^s_{I} (R)\to \gr^s_{I} (R)^{1/m}$$ split for every $m\in \mathbb Z_{>0}$. \end{thm} \begin{proof} Fix $m\in\mathbb Z_{>0}$ and let $\Phi^R_m$ be the splitting in Definition \ref{defSplitting}. We define $$\varphi:\mathcal R^s(I)^{1/m}\to \mathcal R^s(I)$$ to be the homogeneous morphism of $\mathcal R^s(I)$-modules induced by $\varphi(r^{1/m} t^{n/m}) =\Phi^R_m(r^{1/m})t^{n/m}$ if $m$ divides $n$, and $\varphi(r^{1/m} t^{n/m}) =0$ otherwise. The map $\varphi$ is well-defined because $$ \Phi^R_m \big(( I^{((n+1)m)} )^{1/m}\big) \subseteq I^{(n+1)} $$ for every $n\geqslant 0$ by Lemma \ref{LemmaStable}, and it is $\mathcal R^s(I)$-linear since $\Phi^R_m$ is $R$-linear. If $r\in I^{(n)}\subseteq ( I^{(nm)})^{1/m}$, then $\varphi(rt^{nm/m})=\Phi^R_m(r) t^n= rt^n$ because $\Phi^R_m$ is a splitting. We conclude that $\varphi$ is also a splitting. Consider the ideal $\mathcal J=\mathcal J^s(I)$ as in Notation \ref{notBlw}. By Lemma \ref{LemmaStable} we obtain $$\varphi\big(( I^{(nm+1)})^{1/m} t^{nm/m}\big)\subseteq I^{(n+1)} t^n$$ for every $n\geqslant 0$. Therefore, $\varphi(\mathcal J^{1/m})\subseteq \mathcal J$, i.e., $\mathcal J$ is compatible with $\varphi$. This induces a splitting $\overline{\varphi}: (\mathcal R^s(I)/\mathcal J)^{1/m} \to \mathcal R^s(I)/\mathcal J$. The conclusion follows. \end{proof} Theorem \ref{ThmGradedSpl} has the following consequence if the field $k$ has positive characteristic. \begin{cor}\label{Fpure} Let $I$ be a square-free monomial ideal. If $k$ is a perfect field of prime characteristic $p$, then $\mathcal R^s(I)$ and $\gr_I^s(R)$ are $F$-pure. \end{cor} \begin{proof} We note that the rings involved are $F$-finite. Then, they are $F$-pure if and only if they are $F$-split \cite[Corollary 5.3]{HoRo}. Since $k$ is perfect, we have that $R^{1/p}$ and $\mathcal R^s(I)^{1/p}$ correspond to the rings of $p$-roots of $R$ and $\mathcal R^s(I)$ respectively. In addition, $(\mathcal J^s(I))^{1/p}$ is the ideal of $\mathcal R^s(I)^{1/p}$ that corresponds to $\mathcal J^s(I)$. Since $\mathcal J^s(I)$ is radical, $\gr_I^s(R)$ is a reduced ring. Then, $\gr^s_{I} (R)^{1/p}$ corresponds to the ring of $p$-roots of $\gr^s_{I} (R)$. Then, the result follows from Theorem \ref{ThmGradedSpl} with $m=p$. \end{proof} The following Theorem provides necessary and sufficient conditions for the equality of ordinary and symbolic powers of square-free monomial ideals. \begin{thm}\label{MainSQfree} Let $I$ be a square-free monomial ideal. Then the following are equivalent. \begin{enumerate} \item $\Phi_m^R\big(( I^{nm+1})^{1/m}\big)\subseteq I^{n+1}$ for every $m\in\mathbb Z_{>0}$ and $n\in \mathbb Z_{\geqslant 0}$. \item $\Phi_m^R\big(( I^{nm+1})^{1/m}\big)\subseteq I^{n+1}$ for some $m\in\mathbb Z_{>1}$ and every $n\in \mathbb Z_{\geqslant 0}$. \item $I^n=I^{(n)}$ for every $n\in \mathbb Z_{>0}.$ \end{enumerate} \end{thm} \begin{proof} Clearly $(1)$ implies $(2)$. We now assume $(2)$ and prove $(3)$. We first prove that $\mathcal J(I)$ is a radical ideal. Let $f\in I^n$ be a monomial such that $ft^n\in \sqrt{\mathcal J(I)}$ then there exists and integer $e\geqslant 0$ such that $(f t^{n})^{m^{e}}\in\mathcal J(I)$. We observe that $f^{m^e}\in I^{nm^e+1}$. We note that the assumption in $(2)$ implies $\Phi_{m}^R\big(( I^{nm^e+1})^{1/m}\big)\subseteq I^{nm^{e-1}+1}$. Therefore, \begin{align*} f^{m^{e-1}}t^{nm^{e-1}} =\Phi_{m}^A((f^{m^e}t^{nm^e})^{1/m})) &\in \Phi_{m}^A(( I^{nm^e+1} t^{nm^e})^{1/m})\\ & = \Phi_{m}^R(( I^{nm^e+1} )^{1/m})t^{nm^{e-1}}\subseteq I^{nm^{e-1}+1}t^{nm^{e-1}}\subseteq \mathcal J(I). \end{align*} A decreasing induction on $e$ shows $ft^n\in \mathcal J(I)$. Then, $\gr_I(R)=\mathcal R(I)/\mathcal J(I)$ is reduced. As a consequence, $I^n=I^{(n)}$ for every $n\in\mathbb Z_{>0}$ \cite[Corollary 1.6]{herzog2008standard}. Now, we assume that $(3)$. By Lemma \ref{LemmaStable}, we have $\Phi_m^R\big(( I^{nm+1})^{1/m}\big)\subseteq I^{n+1}$ for every $n,\, m\in \mathbb Z_{>0}$, therefore $(1)$ follows. \end{proof} We now state a problem due to Huneke which was brought to our attention by H\`a at the BIRS-CMO workshop on {\it Ordinary and Symbolic Powers of Ideals} at Casa Matem\'atica Oaxaca in the summer of 2017. \begin{problem}[Huneke]\label{ProbHa} Let $I$ be a square-free monomial ideal. Find a number $N\in\mathbb Z_{> 0}$, in terms of $I$, such that $I^n=I^{(n)}$ for every $n\leqslant N$ implies $I^n=I^{(n)}$ for every $n\in\mathbb Z_{>0}$ \end{problem} This problem is strongly related to Conforti-Cornu\'ejols (Conjecture \ref{ConjPackingProb}). In fact, if $N=\height(I)$ satisfies the conclusion of Problem \ref{ProbHa}, then Conjecture \ref{ConjPackingProb} follows \cite[Remark 4.19]{SurveySymbPowers}. H\`a also asked for an optimal value for $N$. In Example \ref{ExSharp}, we prove that our bound is sharp. In general $N=\lfloor\frac{(d+1)^{(d+3)/2}}{2^d}\rfloor$ works for any $I$ \cite[Theorem 5.6]{herzoga2007symbolic}; and if one assumes Conjecture \ref{ConjPackingProb}, then $N=\lceil \frac{d+1}{2}\rceil$ would also work \cite[Remark 4.8]{HaMorey}. As a consequence of our methods, in Theorem \ref{allPowMon} we provide a sharper $N$ in terms of the number of generators of $I$. For the proof of this result, we need the following well-known lemma. We include its proof for the sake of completeness. We denote by $\mu(I)$ the minimal number of generators of $I$. If $I$ is generated by the monomials $\mathbf{x}^{\mathbf{a}_1},\ldots, \mathbf{x}^{\mathbf{a}_u}$, we denote by $I^{[m]}$ the ideal generated by $\mathbf{x}^{m\mathbf{a}_1},\ldots, \mathbf{x}^{m\mathbf{a}_u}$. \begin{lemma}\label{Lemma Obs pe u(I)} Let $I$ be a monomial ideal. If $r\geqslant \mu(I)( m-1)+1$, then $I^r= I^{r- m} I^{[m]}$. \end{lemma} \begin{proof} Let $u=\mu(I)$ and $\mathbf{x}^{\mathbf{a}_1},\ldots, \mathbf{x}^{\mathbf{a}_u}$ a minimal set of generators of $I.$ Let $\alpha_1,\ldots,\alpha_u$ be natural numbers such that $\alpha_1+\ldots+\alpha_u=r,$ then by assumption there must exist $\alpha_i$ such that $\alpha_i\geqslant m.$ Therefore, $$ \mathbf{x}^{\alpha_1\mathbf{a}_1}\cdots \mathbf{x}^{\alpha_u\mathbf{a}_u}= \mathbf{x}^{\alpha_1\mathbf{a}_1}\cdots \mathbf{x}^{(\alpha_i-m)\mathbf{a}_i} \cdots \mathbf{x}^{\alpha_u\mathbf{a}_u}\mathbf{x}^{\alpha_im} \in I^{r-m}I^{[m]}. $$ This shows that $I^r\subseteq I^{r- m} I^{[m]}$. To obtain the other containment, we observe that $I^{[m]}\subseteq I^m$. \end{proof} \begin{thm}\label{allPowMon} Let $I$ be a square-free monomial ideal. If $I^n=I^{(n)}$ for every $n\leqslant \lceil\frac{ \mu(I)}{2}\rceil$, then $I^n=I^{(n)}$ for every $n\in\mathbb Z_{>0}$. \end{thm} \begin{proof} We show that $$\Phi_{2}^R\big(( I^{2n+1})^{1/2}\big)\subseteq I^{n+1}$$ for every $n\geqslant 0$ and then the result follows by Theorem \ref{MainSQfree}, $(2)\Rightarrow (3)$. By assumption and Lemma \ref{LemmaStable} this inclusion holds for $n< \lceil\frac{\mu(I)}{2}\rceil$, as for these values $I^{(n+1)}=I^{n+1}$. We fix $n\geqslant \lceil\frac{\mu(I)}{2}\rceil$. Then $$2n+1\geqslant \mu(I)+1$$ and hence $( I^{2n+1})^{1/2}=( I^{2(n-1)+1})^{1/2}I$ by Lemma \ref{Lemma Obs pe u(I)}. Therefore, by induction $$ \Phi_{2}^R\big( ( I^{2n+1})^{1/2} \big) =\Phi_{2}^R\big(( I^{2(n-1)+1})^{1/2} I \big) =\Phi_{2}^R\big(( I^{2(n-1)+1})^{1/2} \big)I \subseteq I^{n}I=I^{n+1}, $$ finishing the proof. \end{proof} We point out that Theorem \ref{allPowMon} relates to an open problem stated by Francisco, H\`a, and Mermin \cite[Problem 5.14(a)]{francisco2013powers}. The following example shows that Theorem \ref{allPowMon} is sharp. \begin{example}\label{ExSharp} Let $R=k[x_1,\ldots, x_{2t-1}]$ for some $t\geqslant 2$, and let $$I=(x_1x_2,x_2x_3,\ldots, x_{2t-2}x_{2t-1}, x_{2t-1}x_{1}).$$ Then $I^{n}=I^{(n)}$ for every $n<t=\lceil\frac{\mu(I)}{2}\rceil$, whereas $I^{t}\neq I^{(t)}$ \cite[Corollary 4.5]{LamTrung2015}. \end{example} \begin{cor}\label{CorFrob} Let $I$ be a square-free monomial ideal. Then $I^n=I^{(n)}$ for every $n\in\mathbb Z_{>0}$ if and only if $x_1\cdots x_d I^{2n+1}\subseteq \left( I^{n+1}\right)^{[2]}$ for $n< \frac{\mu(I)}{2}$. \end{cor} \begin{proof} Let $m\in \mathbb Z_{>0}$. Since $R^{1/m}$ and $R$ are regular of the same dimension, we have that $\Hom_R(R^{1/m}, R)\cong R^{1/m}$ as $R^{1/m}$-modules \cite[Lemma 1.6 (1)]{FedderFputityFsing}. Furthermore, standard computations show that the map $\Psi_m :R^{1/m}\to R$ induced by $$ \Psi_m(\mathbf{x}^{\mathbf{a}/m})= \begin{cases} \mathbf{x}^{(\mathbf{a} - (m-1)\mathbf{1})/m} & \mathbf{a} \equiv -\mathbf{1}\, (\bmod\, m);\\ 0 & \hbox{otherwise,} \end{cases} $$ where $\mathbf{1} = (1,\ldots, 1)\in \mathbb Z^d$, is a generator of $\Hom_R(R^{1/m}, R)$ as an $R^{1/m}$-module (this is a standard computation in prime characteristic $p$ when $m=p$). We now focus on the case $m=2$. We stress that we are not making any assumption on the characteristic of the field. We note that $x_1\cdots x_d I^{2n+1}\subseteq \left( I^{n+1}\right)^{[2]}$ if and only if $$(x_1\cdots x_d)^{1/2} (I^{2n+1})^{1/2}\subseteq I^{n+1}R^{1/2},$$ as $R^{1/2}$ is isomorphic to $R$. In addition, $(x_1\cdots x_d)^{1/2} (I^{2n+1})^{1/2}\subseteq I^{n+1}R^{1/2}$ is equivalent to $\Psi_2\big((x_1\cdots x_d)^{1/2} (I^{2n+1})^{1/2}\big)\subseteq I^{n+1}$ for every $n\geqslant 0$ \cite[Lemma 1.6 (2)]{FedderFputityFsing}. Since $\Phi_2^R(-)=\Psi_2\big( (x_1\cdots x_d)^{1/2}-\big)$, the result follows from Theorem \ref{MainSQfree} and the proof of Theorem \ref{allPowMon}. \end{proof} We note that Theorem \ref{allPowMon} and Corollary \ref{CorFrob} give an algorithm to verify Conjecture \ref{ConjPackingProb} for a specific ideal. \section{Applications to linear optimization}\label{SecOptimization} In this brief section we translate our result to the context of linear programming. For more on this topic, we refer to \cite{HaTrung}. A {\it clutter} $\mathcal C=(V, E)$ is a collection of subsets $E$ of $V=\{v_1,\ldots, v_n\}$ such that every two elements of $E$ are incomparable with respect to inclusion. We denote by $M:=M(\mathcal C)$ the $n\times m$ matrix with entries equal to 0 or 1, such that its columns are the incidence vectors of the sets in $E$. Given $\mathbf{c}\in \mathbb Z^n_{\geqslant 0}$, by the Strong Duality Theorem we have the following equality of dual linear programs. \begin{equation}\label{dualSystem} \min\{\mathbf{c}\cdot \mathbf{x}\mid \mathbf{x}\in \mathbb R^n_{\geqslant 0},\, M^T\mathbf{x}\geqslant \mathbf{1}_m\}= \max\{\mathbf{1}_m\cdot \mathbf{y}\mid \mathbf{y}\in \mathbb R^m_{\geqslant 0},\, M\mathbf{y}\leqslant \mathbf{c}\}, \end{equation} where $\mathbf{1}_m=(1,\ldots, 1)\in \mathbb Z^m$. We say that $\mathcal C$ {\it packs for $\mathbf{c}$} if Equation \eqref{dualSystem} holds for integer vectors $\mathbf{x} \in \mathbb Z_{\geqslant 0}^n$ and $\mathbf{y}\in \mathbb Z_{\geqslant 0}^m$. The clutter $\mathcal C$ satisfies the {\it Max-Flow-Min-Cut (MFMC) property} if it packs for every $\mathbf{c}\in \mathbb Z_{\geqslant 0}^n$. We now recall a lemma that allows us to rephrase Theorem \ref{allPowMon} in this context. We first set up some notation. Let $\{\mathbf{b}_1,\ldots, \mathbf{b}_m\}$ be the column vectors of $M$. We denote by $I$ the ideal of the polynomial ring $R=k[v_1,\cdots, v_n]$ generated by the monomials $\{\mathbf{v}^{\mathbf{b}_1},\ldots, \mathbf{v}^{\mathbf{b}_m}\}$. \begin{lemma}[{\cite[Proposition 3.1]{Trung06}}]\label{translation} Set $$\gamma(\mathbf{c}) := \min\{\mathbf{c}\cdot \mathbf{x}\mid \mathbf{x}\in \mathbb Z^n_{\geqslant 0},\, M^T\mathbf{x}\geqslant \mathbf{1}_m\}, \quad \textnormal{ and }$$ $$ \sigma(\mathbf{c}) := \max\{\mathbf{1}_m\cdot \mathbf{y}\mid \mathbf{y}\in \mathbb Z^m_{\geqslant 0},\, M\mathbf{y}\leqslant \mathbf{c}\}.$$ Then \begin{enumerate} \item[(i)] $\mathbf{v}^\mathbf{c}\in I^{(t)}$ if and only if $t\leqslant \gamma(\mathbf{c})$. \item[(ii)] $\mathbf{v}^\mathbf{c}\in I^{t}$ if and only if $t\leqslant \sigma(\mathbf{c})$. \end{enumerate} \end{lemma} We are now ready to present the main theorem of this section which is related to previous results in integer programming \cite[Corollary 2.3]{BT82}. \begin{thm}\label{MFMC} The clutter $\mathcal C$ packs for every $\mathbf{c}\leqslant \mathbf{1}_n\lceil \frac{m}{2}\rceil$ if and only if $\mathcal C$ satisfies the MFMC property. \end{thm} \begin{proof} By definitions the MFMC property implies the $\mathcal C$ packs for every $\mathbf{c}\in \mathbb Z_{\geqslant 0}^m$. We focus on the other direction. We assume $\mathcal C$ packs for every $\mathbf{c}\leqslant \mathbf{1}_n\lceil \frac{m}{2}\rceil$. Let $\mathbf{v}^\mathbf{c}$ be a minimal monomial generator of $I^{(t)}$ for some $1\leqslant t\leqslant \lceil\frac{m}{2}\rceil$, then it clearly divides $\mathbf{v}^{t\mathbf{1}_n}$. Moreover, by Lemma \ref{translation}, we have $\gamma(\mathbf{c})=t$; this equality holds since reducing a component of $\mathbf{c}$ by 1, reduces $\gamma(\mathbf{c})$ by at most 1. Therefore, by assumption we have $\sigma(\mathbf{c}) = \gamma(\mathbf{c})= t$ and then $\mathbf{v}^\mathbf{c}\in I^t$ by Lemma \ref{translation}. We conclude that $I^{(t)}=I^t$ for $t=1,\ldots, \lceil\frac{m}{2}\rceil=\lceil\frac{\mu(I)}{2}\rceil$. By Theorem \ref{allPowMon} it follows that $I^{(t)}=I^t$ for every $t\geqslant 1$. Now, let $\mathbf{c}\in \mathbb Z_{\geqslant 0}^n$ be arbitrary, then $\mathbf{v}^\mathbf{c}\in I^{(\gamma(\mathbf{c}))}=I^{\gamma(\mathbf{c})}$ and by Lemma \ref{translation} we have $\sigma(\mathbf{c})\geqslant \gamma(\mathbf{c})$. Since $\sigma(\mathbf{c})\leqslant \gamma(\mathbf{c})$ always holds, it follows that $\sigma(\mathbf{c})=\gamma(\mathbf{c})$ and this finishes the proof. \end{proof} \section*{Acknowledgments} The authors started this project after participating in the BIRS-CMO workshop on {\it Ordinary and Symbolic Powers of Ideals} Summer of 2017, Casa Matem\'atica Oaxaca, Mexico, where they learned new open problems on the subject. The authors thank the organizers: Chris Francisco, Tai H\`a, and Adam Van Tuyl. They also thank Craig Huneke for suggestions that improved Theorem \ref{allPowMon} and Rafael H. Villarreal for suggesting Remark \ref{RemWC}. The authors would like to thank Hailong Dao, Elo\'isa Grifo, Tai H\`a, and Robert Walker for suggestions on an earlier draft. Part of this project was completed in the Mathematisches Forschungsinstitut Oberwolfach (MFO) while the first author was in residence at the institute under the program {\it Oberwolfach Leibniz Fellows}. The first author thanks MFO for their hospitality and excellent conditions for conducting research. The second author thanks Jack Jeffries for inspiring conversations. The authors would like to thank the referees for their helpful comments and suggestions which improved this paper. \bibliographystyle{plain}

090bb6ff2f4ffc9d6e980e2dc3d1ff350f991519

\section{Introduction} In recent years considerable progress was made with overcoming the complex action problem at finite density for several lattice field theories. It was possible to exactly map the partition sum of these systems to a representation in terms of worldlines (and/or worldsheets) where all contributions to the partition sum are real and positive, such that a Monte Carlo simulation can be done directly in terms of the worldlines (see, e.g., \cite{Endres,Weisz,phi4_1,phi4_2} for work on the $\phi^4$ theory studied here). With the worldline approach it is possible to address new physics questions related to finite density. An example is condensation of particles at low temperatures, which is the topic of this contribution. To illustrate the condensation phenomenon we study here, in Fig.~\ref{condexample} we show the results for the expectation value of the particle number $\langle N \rangle$ versus the chemical potential $\mu$. The results are for $\phi^4$ theory in 2d on a $L \times N_t$ lattice with $N_t = 400$ and three different values of $L$. The temperature is very low ($T = 1/N_t = 0.0025$ in lattice units) and indeed we observe condensation as a function of $\mu$ (details see below). For each of the three values of $L$ the particle number $\langle N \rangle$ quickly rises from $\langle N \rangle = 0$ to $\langle N \rangle = 1$ at some critical chemical potential value $\mu_1(L)$, then further to $\langle N \rangle = 2$ at a second critical value $\mu_2(L)$ and similar for higher particle number sectors. Note that at zero temperature ($N_t = \infty$) one expects discontinuous jumps of $\langle N \rangle$ which are here rounded by temperature effects. Nevertheless we can identify the critical values $\mu_n(L)$ and determine their values as a function of $L$. The values $\mu_n(L)$ correspond to the values of the chemical potential where we observe condensation of another particle visible in the step from $\langle N \rangle = n-1$ to $\langle N \rangle = n$. \begin{figure}[b!] \centering \hspace*{-3mm} \includegraphics[width=90mm,clip]{cond_example.pdf} \caption{The expectation value of the particle number $\langle N \rangle$ as a function of the chemical potential $\mu$ (in lattice units). We show the results for the 2d case at $N_t = 400$ and different values of $L$.} \label{condexample} \end{figure} In \cite{Bruckmann} it was shown that at very low temperature the condensation thresholds $\mu_n$ are related to the physical mass $m(L)$ and the $n$-particle energies $W_n(L)$ via the relations \begin{equation} m(L) \; = \; \mu_1(L) \; , \qquad W_n(L) \; = \; \sum_{k=1}^n \mu_k(L) \; , \label{mu_energy} \end{equation} where we now made explicit, that not only the $\mu_n(L)$, but also the physical mass $m(L)$ and the $n$-particle energies $W_n(L)$ depend on the spatial extent $L$. It has been known since the pioneering paper \cite{huang_yang} that the dependence of the 2- and 3-particle energies $W_2(L)$ and $W_3(L)$ on the spatial extent $L$ can be parameterized in terms of scattering data of the underlying theory. Thus we conclude from (\ref{mu_energy}) that the condensation thresholds $\mu_n(L)$ are governed by the scattering data. Demonstrating and analyzing the connection between low temperature condensation and scattering data in $\phi^4$ theory in two and four dimensions is the topic of this contribution (see also \cite{prl}). \section{Worldline representation and Monte Carlo simulation} The system where we explore the relation between condensation and scattering data is the complex $\phi^4$ field in $d = 2$ and $d = 4$ dimensions. The lattice action is given by \begin{equation} S[\phi] \; = \; \sum_{x \in \Lambda} \bigg( \eta \, |\phi _{x}|^2 \; + \; \lambda \, |\phi _{x}|^4 \; - \; \sum_{\nu = 1}^{d} \left[ e^{\,\mu \delta_{\nu,d}}\phi _{x}^{\ast} \phi _{x+\hat{\nu}} + e^{\,-\mu \delta_{\nu,d}}\phi _{x+\hat{\nu}}^{\ast} \phi _{x} \right]\bigg) \; , \label{action_conventional} \end{equation} where $\eta \equiv 2d + m^{2}_b$ with $m_b$ the bare mass parameter. $\lambda$ is the quartic coupling and $\mu$ the chemical potential. The fields $\phi_x$ are assigned to the sites $x$ of a lattice of size $L^{d-1} \times N_t$. At $\mu \neq 0$ the action is complex and the Boltzmann factor $e^{-S[\phi]}$ cannot be used as a probability in a Monte Carlo simulation. This complex action problem of the conventional representation (\ref{action_conventional}) can be solved by exactly mapping the system to a worldline representation (see, e.g., \cite{phi4_1} for a derivation of the form we use here). In the worldline representation the partition sum reads ($\beta \equiv N_t$) \begin{equation} Z \; = \; \sum_{\{k\}} \left[ \prod_{x} \delta \left(\vec{\nabla} \cdot \vec{k}_{x} \right) \right] \, e^{\, \mu \, \beta \, \omega[k]} \; B[k] \; . \label{worldlineZ} \end{equation} $Z$ is a sum over configurations of the worldline variables $k_{x,\nu} \in \mathbb{Z}$ assigned to the links of the lattice. They have to obey constraints which have the form of a product over Kronecker deltas $\delta(j) \equiv \delta_{j,0}$ at all sites $x$. At each $x$ the Kronecker deltas enforce $\vec{\nabla} \cdot \vec{k}_{x} \equiv \sum_{\nu}(k_{x,\nu} - k_{x-\hat{\nu},\nu}) = 0$, i.e., zero divergence for $k_{x,\nu}$, and as a consequence the worldline variables $k_{x,\nu}$ must form closed loops of conserved flux. By $\omega[k]$ we denote the total winding number of the $k$-flux around the compact time direction and the chemical potential couples to $\omega[k]$ in the form $e^{\mu \beta \omega[k]}$. The observable we need for our analysis is the expectation value of the particle number $\langle N \rangle = \partial \ln Z / \partial \beta \mu = \langle \omega[k] \rangle_{wl}$, where $\langle .. \rangle_{wl}$ denotes the vacuum expectation in the worldline representation. The configurations of the worldline variables $k_{x,\nu}$ come with a real and positive weight factor \begin{equation} B[k] \; = \; \sum_{\{a\}} \, \prod_{x,\nu}\frac{1}{(a_{x,\nu}+|k_{x,\nu}|)! \, a_{x,\nu}!} \, \prod_{x} I(s_x) \; \; \quad \mbox{with} \quad \; \; I(s_x) \; = \; \int_{0}^{\infty} \!\! d r \; r^{\, s_x + 1} \, e^{\, -\eta \, r^2 \, - \, \lambda \, r^4} \; . \label{weights} \end{equation} $B[k]$ is a sum over configurations $\sum_{\{a\}}$ of auxiliary link variables $a_{x,\nu} \in \mathbb{N}_0$, and by $s_x$ we denote the non-negative integer combination $s_x \; = \; \sum_{\nu}\big[|k_{x,\nu}| + |k_{x-\hat{\nu}}| + 2(a_{x,\nu} + a_{x-\hat{\nu}})\big] $, which appears as an argument in the integrals $I(s_x)$ that come from integrating out the radial degrees of freedom of the original field variables at site $x$. They are pre-calculated and stored for the simulations. All weight factors in the worldline representation are real and positive such that the complex action problem is solved. Concerning the details of the updates for the worldline variables $k_{x,\nu}$ and the auxiliary variables $a_{x,\nu}$ we refer to \cite{algorithm,preliminary}. In 4d we use lattices with $N_t = 320$ and 640, and $L$ between 3 and 10 at coupling values of $\eta = 7.44$ and $\lambda = 1.0$, with a statistics of $2 \times 10^5$ configurations. In 2d the corresponding parameters are $N_t = 400$, $L$ between 2 and 16 with $\eta = 2.6$, $\lambda = 1.0$ and a statistics of $4 \times 10^5$. \section{Analysis of the 4d case} After computing $\langle N \rangle$ as a function of $\mu$ we identify the steps where $\langle N \rangle$ transits from $\langle N \rangle = n-1$ to $\langle N \rangle = n$ (compare Fig.~\ref{condexample}). To determine the corresponding critical values $\mu_n$ we fit the data for $\langle N \rangle$ in the vicinity of the steps with the logistic function $\langle N \rangle \!=\! [1\!\,+\!\, \exp(-a_{n}[\mu\!-\!\mu_n])]^{-1}\!+\!n\!-\!1$. Using (\ref{mu_energy}) we then compute $m(L)$, $W_2(L)$ and $W_3(L)$ from the critical values $\mu_n(L)$. \begin{figure}[t!] \centering \hspace*{-3mm} \includegraphics[width=80mm,clip]{energies.pdf} \caption{The physical mass $m(L)$ and the 2- and 3-particle energies $W_2(L)$ and $W_3(L)$ for the 4d case as a function of the lattice extent $L$ (figure from \cite{prl}). We show the results determined from the condensation steps (blue squares) and compare them to the results from spectroscopy (black diamonds). The full red curves are the fits of $m$ and $W_2$ with (\ref{mL}) and (\ref{W2}). The dashed maroon curve is the function $W_3$ from Eq.~(\ref{W3}) when using the scattering length $a$ from the fit of $W_2$ as input.} \label{energy_plot} \end{figure} In Fig.~\ref{energy_plot} we show the results for $m(L)$, $W_2(L)$ and $W_3(L)$ determined from the critical chemical potential values $\mu_n(L)$ as squares. To test the relations (\ref{mu_energy}) and the reliability of our determination of the critical values $\mu_n(L)$, we computed $m(L)$, $W_2(L)$ and $W_3(L)$ also in a spectroscopy analysis based on $2n$-point functions calculated at $\mu = 0$ in the conventional representation (\ref{action_conventional}). The corresponding results are shown as diamonds in Fig.~\ref{energy_plot} and coincide almost perfectly with the data from the condensation steps. This cross check confirms the interpretation of the critical chemical potential values as combinations of multi-particle energies. The next step is to invoke the finite volume relations for $m(L)$ \cite{kari}, the result \cite{huang_yang,Luscher_w2} for the 2-particle energy $W_2(L)$ (using the notation of \cite{sharpe}) and the results \cite{beane,sharpe,sharpe1,sharpe2,sharpe3} for the 3-particle energy $W_3(L)$ (the numerical constants ${\cal I}$ and ${\cal J}$ are given by ${\cal I} = -8.914, {\cal J} = 16.532$): \begin{eqnarray} m(L) & = & m_\infty + \frac{A}{L^{\frac{3}{2}}} \, e^{ - L \; m_\infty } , \label{mL} \\ W_2(L) & = & 2m + \frac{4\pi a}{m L^3} \! \Bigg[ 1 - \frac{a}{L} \frac{{\cal I}}{\pi} + \bigg(\!\frac{a}{L}\! \bigg)^{\!\!2} \, \frac{ {\cal I}^{\,2} \!-\! {\cal J}}{\pi^2} + {\cal O} \! \bigg(\!\frac{a}{L}\! \bigg)^{\!\!3}\Bigg]\! , \label{W2} \\ W_3(L) & = & 3m + \frac{12\pi a}{m L^3} \! \Bigg[ 1 - \frac{a}{L} \frac{{\cal I}}{\pi} + \bigg(\!\frac{a}{L}\! \bigg)^{\!\!2} \, \frac{ {\cal I}^{\,2} \!+ \!{\cal J}}{\pi^2} + {\cal O} \! \bigg(\!\frac{a}{L}\! \bigg)^{\!\!3}\Bigg]\! . \label{W3} \end{eqnarray} Up to order $1/L^5$ only three parameters are needed to describe the data: the infinite volume mass $m_\infty$, the amplitude $A$ and the scattering length $a$. Fitting the data for $m(L)$ with the relation (\ref{mL}) we find a value of $m_\infty = 0.168(1)$ in lattice units. For fitting $W_2(L)$ we use (\ref{W2}), with the mass parameter $m$ on the rhs.~replaced by the corresponding values $m(L)$. This gives rise to a value of $a = - 0.078(7)$ for the scattering length in lattice units and a value of $a \, m_\infty = - 0.013(1)$ for the dimensionless product of $a$ and $m_\infty$. The functions (\ref{mL}) and (\ref{W2}) with the fit values for $m_\infty$, $A$ and $a$ are shown as full red curves in Fig.~\ref{energy_plot} and describe the data for $m(L)$ and $W_2(L)$ very well (with the exception of the smallest $L$ where higher corrections in $1/L$ would be necessary). Having determined the mass and the scattering length, no further parameters are necessary to describe $W_3(L)$ with (\ref{W3}). Inserting the fit value for $a$ and again using $m(L)$ in the rhs.\ of (\ref{W3}), we thus get a ''prediction'' for the data $W_3(L)$. This prediction is shown as a dashed curve in Fig.~\ref{energy_plot} and obviously describes the data for $W_3(L)$ very well (again with the exception of the smallest $L$). This concludes the discussion of the 4d case and our results confirm the relations (\ref{mu_energy}) of the condensation thresholds to multi-particle energies, which in turn are described by scattering data. Thus we have quantitatively established the connection of condensation and scattering data. \section{Analysis of the 2d case} Also in the 2d case we determined the critical values $\mu_n(L)$ from fitting the steps of $\langle N \rangle$ and then computed $m(L)$, $W_2(L)$ and $W_3(L)$ using the relations (\ref{mu_energy}). We cross-checked these results with a spectroscopy calculation in the conventional representation and again found very good agreement between the condensation and the spectroscopy results. The next step is the finite volume analysis of $m(L)$, $W_2(L)$ and $W_3(L)$. As before the mass $m(L)$ can be described with a 2-parameter ansatz, which in 2d reads $m(L) = m_\infty + A \, e^{-m_\infty \, L} / \sqrt{L}\, $. For analyzing the 2-particle energy $W_2(L)$ we follow the approach \cite{Luscher_2d} that is applicable to short range potentials. Outside the interaction range the wave function is a 2-particle plane wave $\psi = e^{- i x_1 p_1} \, e^{- i x_2 p_2}$ with momenta $p_1$ and $p_2$. The corresponding energy is $W_2(L) = \sum_{j=1}^2 \sqrt{ m(L)^2 + p_j^2}$. We rewrite the wave function $\psi$ by using the center of mass coordinate $(x_1 + x_2)/2$ and the relative coordinate $r = x_1 - x_2$. The energy values $W_2(L)$ determined from the condensation steps correspond to vanishing total momentum $p_1 + p_2 = 0$, and we set $p_1 = p = -p_2$. For vanishing total momentum the wave function then has the form $\psi = e^{-i p r}$ and the 2-particle energy is given by \begin{equation} W_2(L) \; = \; 2 \sqrt{ m(L)^2 + p^2} \; . \label{2dW2} \end{equation} For finalizing the connection between the 2-particle energy and the scattering data we need to invoke the quantization for the momenta $p$ in a finite box of size $L$. This condition is obtained from the boundary condition for $\psi$ which connects the wave function at $r = 0$ to its value at $r = L$ and reads $e^{-i \,p L} = e^{\, i \,2 \delta(p)}$. It expresses the fact that the plane wave solution is correct only outside the interaction range, and that when connecting $r = 0$ with $r = L$ one has to take into account the phase shift $\delta(p)$ that is picked up when the two particles interact. Thus we obtain \begin{equation} \delta(p) \; = \; - \frac{p L}{2} \; . \label{2ddelta} \end{equation} The two equations (\ref{2dW2}) and (\ref{2ddelta}) constitute the connection between scattering data and $W_2(L)$. We can use the numerically determined values for $W_2(L)$ to compute from (\ref{2dW2}) the corresponding relative momenta $p$, and then use these to compute the scattering phase shift $\delta(p)$. We show the corresponding results in the lhs.\ plot of Fig.~\ref{plot2D} and compare the data from a determination based on the condensation thresholds to those from a determination based on standard spectroscopy. Before we discuss $W_3(L)$, we stress that at fixed couplings the phase shift is only a function of the lattice size $L$: the momentum $p$ determined from (\ref{2dW2}) depends only on the lattice size $L$, i.e., we have $p = p(L)$. Thus the phase shift from (\ref{2ddelta}) is given by $\delta(p(L)) = - \frac{p(L) \, L}{2} \equiv \delta(L)$. \begin{figure}[t!] \centering \hspace*{-3mm} \includegraphics[height=62mm,clip]{delta_vs_k_2to16_conv_wl_v2.pdf} \hspace{4mm} \includegraphics[height=62mm,clip]{comp_W3_cond_deltaW2_v1.pdf} \caption{Lhs.: The scattering phase shift $\delta(p)$ versus $p$. We compare the results from the condensation threshold (blue squares) to reference data from spectroscopy (black diamonds). Rhs.: The 3-particle energy $W_3$ as a function of $L$. We show the results of the direct determination from the condensation thresholds (blue squares) and compare it to a prediction based on $\delta(p)$ determined from $W_2$.} \label{plot2D} \end{figure} Similar to the 4d case we now use the scattering data determined from $W_2(L)$ to ''predict'' $W_3(L)$ and thus the third critical chemical potential value $\mu_3(L)$. The approach is a generalization of the strategy \cite{Luscher_2d} we have followed for analyzing $W_2(L)$. Again we make a plane wave ansatz $\psi = e^{- i x_1 p_1} \, e^{- i x_2 p_2} \, e^{- i x_3 p_3}$ for three particles which describes the system when all three particles are sufficiently remote from each other. The corresponding energy is $W_3(L) = \sum_{j=1}^3 \sqrt{ m(L)^2 + p_j^2}$. As before we introduce the center of mass coordinate $(x_1 + x_2 + x_3)/3$, as well as the relative coordinates $r_{2} = x_2 - x_1$ and $r_{3} = x_3 - x_1$. Using these to parameterize $\psi$ and demanding total vanishing momentum $p_1 + p_2 + p_3 = 0$, we find $\psi = e^{- i r_2 p_2} \, e^{- i r_3 p_3}$ and $p_1 = - p_2 - p_3$. This 3-particle wave function has to obey two quantization conditions of the form (\ref{2ddelta}) that contain $p_2$ and $p_3$. Using the fact that the phase shift is only a function of $L$ we can determine $p_2$ and $p_3$ as $p_2 = p_3 = - 2 \delta(L) / L$. Inserting these values and $p_1 = - p_2 - p_3$ into $W_3(L) = \sum_{j=1}^3 \sqrt{ m(L)^2 + p_j^2}$ we obtain our prediction for $W_3(L)$. The corresponding values are shown as red circles in the rhs.\ plot of Fig.~\ref{plot2D}. We compare them to the results of a direct determination from all three condensation thresholds. The results agree very well and we conclude that the structure of the condensation thresholds $\mu_n(L)$ can indeed be correctly described with the scattering data of the theory. We remark that the 2- and 3-particle energies $W_2(L)$ and $W_3(L)$ can be analyzed with a different approach \cite{Guo1,Guo2}, where one uses the exact solution for the scattering phase shift which depends on a single parameter, the amplitude $V_0$ of the point-like interaction. This parameter can be determined from $W_2(L)$ and subsequently used for $\delta$ in the 3-particle quantization conditions to determine the two independent momenta $p_2$ and $p_3$ needed to compute $W_3(L)$. \section{Concluding remarks} In this contribution we have shown for a simple scalar field theory in two and four dimensions that low temperature condensation is governed by the scattering data of the theory. This relation is expected to be a general non-perturbative feature, but in order to study it on the lattice usually a complex action problem has to be solved, which so far has been achieved for only a few systems. However, there exist interesting theories which are already free of the complex action problem. Examples are lattice field theories based on the gauge group SU(2), and more interestingly, QCD with isospin chemical potential where the condensation of pions is expected to be related to pion scattering data. For these systems an analysis along the lines sketched here should be possible. \vskip3mm \noindent {\bf Acknowledgements:} We thank F.\ Bruckmann, P.\ Guo, M. Hansen, T.\ Kloiber, A.\ Maas, C.B.\ Lang, C.\ Marchis, S.\ Sharpe and T.\ Sulejmanpasic for discussions. This work is supported by the Austrian Science Fund FWF, grant I 2886-N27 and the FWF DK W 1203, ''Hadrons in Vacuum, Nuclei and Stars".

1a5fe0990519ffb40ae54655a188b7e1ac12456f

\section{Sample fabrication and experimental condition details} For this study, we have used a bilayer graphene (BLG) device presented in \cite{Kraft2018} and added an overall top gate (TG). The original device is an edge-connected hBN-BLG-hBN heterostructure \cite{Wang2013} (top and bottom hBN of about $38\,$nm and $35\,$nm thick respectively) placed onto a pre-patterned back gate (BG) designed on a sapphire substrate, covered by an additional dielectric layer Al$_2$O$_3$ ($20\,$nm) deposited by atomic layer deposition (ALD). The BLG is edge-connected \cite{Wang2013} with Ti/Al electrodes and the QPC is defined by a split gate (SG) designed on top of the heterostructure (see \cite{Kraft2018} for details). The entire sample was then covered by an extra layer of Al$_2$O$_3$ ($30\,$nm) deposited by ALD and an overall Ti/Cu top gate (TG). The electrical measurements have been performed in a $^3$He/$^4$He dilution refrigerator BF-LD250 from BlueFors at a base temperature below 20mK unless otherwise mentioned. The sample is probed in a two-terminal configuration using standard low-frequency ($\sim 13\,$Hz) lock-in technique with an AC excitation ranging from 1 to 20\,$\mu$V. Finally, all magnetic fields used in these experiments were applied perpendicular to the BLG plane. \section{Effect of split and top gate} \begin{figure*} \centering \includegraphics{figS1a} \includegraphics{figS1b} \caption{(a), (b) Color maps of the differential resistance $R$ as a function of BG voltage $V_\mathrm{BG}$ and SG voltage $V_\mathrm{SG}$ (a) or TG voltage $V_\mathrm{TG}$ (b). The diagonal high resistance line corresponds to the condition of charge neutrality in the split-gated or top-gated region respectively, while the vertical line only tuned by the BG corresponds to charge neutrality in the regions that are independent of either SG or TG respectively. The displacement field $D$ increases along the diagonal line, resulting in the opening of a band gap while the Fermi level is kept constant at zero energy in the band gap center. (c), (d) Linecuts of the maps showing the differential resistance $R$ as a function of BG voltage $V_\mathrm{BG}$ for various SG voltages (c) or TG voltages (d).} \label{fig:S1} \end{figure*} Here we analyze the effect on transport of the SG and TG in the presence of a BG voltage $V_\mathrm{BG}$. The differential resistance $R$ of the device is shown in Fig.\,\ref{fig:S1}a,\,b as a function of BG voltage $V_\mathrm{BG}$ and either SG voltage $V_\mathrm{SG}$ or TG voltage $V_\mathrm{TG}$. During each measurement, the unused gate was kept grounded. Corresponding horizontal linecuts are presented in Fig.\,\ref{fig:S1}c,\,d. Two lines of resistance maxima are visible in the color maps. In case of the SG, the vertical line corresponds to charge neutrality of the reservoirs while the diagonal line corresponds to charge neutrality of the SG region which is tuned by both BG and SG. As the gap develops with increasing displacement field $D$ along the diagonal resistance line, conductance is fully suppressed underneath the SG. However, as the SG does not cover entirely the width of the device, the resistance maximum does not go beyond $2\,$k$\Omega$ due to the remaining conducting channel between the SG electrodes (see \cite{Kraft2018} for detailed explanations). In contrast, the TG fully covers the device, \textit{i.e.} the entire width of the BLG layer. Then the diagonal resistance maximum rises up to resistance values in the order of $10\,$M$\Omega$, \textit{i.e.}, maximum limit of our lock-in detection technique (similarly as in \cite{Oostinga2007}). We note that a double maximum in the resistance along the displacement field line is observed, which might be explained by the two different top-gated regions of our sample, \textit{i.e.} reservoirs and QPC, where a partial screening by the SG may lead to slightly different capacitive coupling of the TG. \section{Quantized conductance and magnetic field dependence under different gate conditions} We have measured our QPC under various gate conditions. Figure\,\ref{fig:S2} shows the differential conductance $G$ and differentiated differential conductance $\mathrm{d}G/\mathrm{d}V_\mathrm{TG}$ as functions of TG voltage $V_\mathrm{TG}$ and SG voltage $V_\mathrm{SG}$ at BG voltage $V_\mathrm{BG}=10\,$V. While the confinement is changed with respect to the data presented in the main text, the main features are conserved. Quantized conductance plateaus are observed as large stripes, whereas superimposed vertical conductance oscillations are due to Fabry-P\'{e}rot resonances of the reservoirs. As shown in the main text, the magnetic depopulation \cite{Vanwees1988a} of the 1D subbands shows a complex pattern of splitting and bunching of magnetoelectric subbands (see Fig.\,\ref{fig:S3}). \begin{figure} \centering \includegraphics{figS2} \caption{(a) Differential conductance $G$ and (b) differentiated differential conductance $\mathrm{d}G/\mathrm{d}V_\mathrm{TG}$ as functions of TG voltage $V_\mathrm{TG}$ and SG voltage $V_\mathrm{SG}$ at constant BG voltage $V_\mathrm{BG}=10\,$V.} \label{fig:S2} \end{figure} \begin{figure} \centering \includegraphics{figS3} \caption{(a) Differential conductance $G$ and (b) differentiated differential conductance $\mathrm{d}G/\mathrm{d}V_\mathrm{TG}$ as functions of TG voltage $V_\mathrm{TG}$ and magnetic field $B$ at constant BG voltage $V_\mathrm{BG}=10\,$V and SG voltage $V_\mathrm{SG}=-11.6\,$V. The first subband is reached just at the limit of the maximum applied TG voltage $V_\mathrm{TG}=-12\,$V. (c) Differential conductance $G$ as a function of TG voltage $V_\mathrm{TG}$ for different $B$. Curves correspond to linecuts of the above panels as marked by the dashed lines in (b).} \label{fig:S3} \end{figure} \section{Ambipolar QPC} We have probed our device in the opposite gate polarity, \textit{e.g.}, with BG voltage $V_\mathrm{BG}=-9.0\,$V and positive SG voltage $V_\mathrm{SG}=8.8\,$V. Therefore, while the gap opens, the Fermi level is positioned in the valence band in the 1D constriction and the 2D reservoirs. Figure\,\ref{fig:S4} displays the differential conductance $G$ as a function of TG voltage $V_\mathrm{TG}$ at $B=6\,$T. By applying an increasing positive TG voltage $V_\mathrm{TG}$ the hole subbands are depopulated, resulting in a stepwise decrease of the conductance. The spin- and valley-degeneracy of the magnetoelectric subbands is fully lifted below $G=12\,e^2/h$ and plateaus appear in steps of $e^2/h$. When the Fermi level is finally tuned into the gap, the conduction through the channel is pinched-off. However, further increase of the TG voltage $V_\mathrm{TG}$ results in tuning the Fermi level into the conduction band and thus the population of the system with electrons. Then, we observe a stepwise increase of the conductance with a full splitting of the Landau levels in steps of $e^2/h$ below the first plateau. This highlights the ambipolarity of our system where the splitting of the Landau levels in steps of $e^2/h$ can be probed continuously for both types of charge carriers. \begin{figure} \centering \includegraphics{figS4} \caption{Differential conductance $G$ as a function of TG voltage $V_\mathrm{TG}$ at constant BG voltage $V_\mathrm{BG}=-9\,$V and SG voltage $V_\mathrm{SG}=8.8\,$V for $B=6\,$T.} \label{fig:S4} \end{figure} \begin{figure} \centering \includegraphics{figS5} \caption{Differentiated conductance $\mathrm{d}G/\mathrm{d}V_\mathrm{BG}$ as a function of the BG voltage $V_\mathrm{BG}$ and magnetic field $B$ at SG and TG voltage $V_\mathrm{SG/TG}=0$, measured at a temperature $T=4.1\,$K. The gate configuration resembles the scenario of the 2D device without constriction, though it is not measured exactly at charge neutrality of SG and TG because of small residual doping due to slightly shifted Fermi levels.} \label{fig:S5} \end{figure} \section{Landau level fan of the 2D system} The magnetic field dependence at zero SG and TG voltage of the BLG is displayed in Fig.\,\ref{fig:S5}. Under these conditions the differentiated differential conductance $\mathrm{d}G/\mathrm{d}V_\mathrm{TG}$ as a function of BG voltage $V_\mathrm{BG}$ and magnetic field $B$ shows a regular Landau level fan diagram. We note that no valley splitting can be seen. \section{Fabry-P\'{e}rot interferences} \begin{figure*} \centering \includegraphics{figS6} \caption{(a) Differential conductance $G$ and (b) oscillating part of the conductance $\delta G$ as functions of the TG voltage $V_\mathrm{TG}$ (upper abscissa) and Fermi wave vector $k_\mathrm{F}$ (lower abscissa) at constant BG voltage $V_\mathrm{BG}=9\,$V and SG voltage $V_\mathrm{SG}=-9.1\,$V. The oscillating part is obtained by subtracting the smooth background conductance (orange) from the raw conductance (blue). (c) Normalized Fourier transform of $\delta G$ as a function of length $L$, yielding the frequency spectrum of the oscillations corresponding to the size of the cavity at the resonance condition. The smooth curve (black) is obtained by convolving the more noisy raw signal (blue) with a gaussian filter.} \label{fig:S6} \end{figure*} Here, we analyze the Fabry-P\'{e}rot interferences observed as conductance oscillations as aforementioned. As we see in Fig.\,2b of the main text and Fig.\,\ref{fig:S2} in this Supplemental Material (SM), the interferences are mainly tuned by the TG voltage. Therefore, the resonances may occur in a cavity formed by the non-splitgated part of the device. We estimate the size of the cavity associated with the observed Fabry-P\'{e}rot interferences. Figure\,\ref{fig:S6}a shows the differential conductance $G$ as a function of TG voltage $V_\mathrm{TG}$ and Fermi wave vector $k_\mathrm{F}$ respectively (at $V_\mathrm{BG}=9.0\,$V and $V_\mathrm{SG}=-9.1\,$V), corresponding to Fig.\,2b of the main text. The oscillating part of the conductance $\delta G$, plotted in Fig.\,\ref{fig:S6}b, is obtained by subtracting the smooth background from the conductance. The size of the cavity is directly linked to the oscillation frequency at resonance condition $L=j\cdot\frac{\mathrm{\pi}}{k_\mathrm{F}}$, with $j$ an integer number and thus can be extracted by performing a Fourier transform. The resulting frequency spectrum is shown in Fig.\,\ref{fig:S6}c. A pronounced peak is observed at about $230\,$nm, being in good agreement with the physical distance between contacts and SG.\\ \section{Stability of the electrostatically induced QPC} \begin{figure} \centering \includegraphics{figS7} \caption{Differential conductance $G$ as a function of TG voltage $V_\mathrm{TG}$ at constant BG voltage $V_\mathrm{BG}=9.0\,$V and SG voltage $V_\mathrm{SG}=-10.6\,$V. The plot showing four curves each up and down for consecutive measurements.} \label{fig:S7} \end{figure} Here, we test the stability of the conductance quantization. Multiple TG sweeps (up and down) are recorded under the same confinement condition as presented data in the main text ($V_\mathrm{BG}=9.0\,$V and $V_\mathrm{SG}=-10.6$V). In total four curves each up ($V_\mathrm{TG}=-12\,$V$\rightarrow 0$) and down ($V_\mathrm{TG}=0\rightarrow -12\,$V) are measured and plotted in Fig.\,\ref{fig:S7}. We note that all four curves perfectly sit on top of each other making them indistinguishable and both features conductance plateaus and Fabry-P\'{e}rot interferences are fully reproduced. We note that a very small hysteresis between up- and down-sweeps is visible. \section{Source-drain bias spectroscopy and 1D subband spacing} \begin{figure*} \centering \includegraphics[width=0.95\textwidth]{figS8} \caption{Transconductance as a function of source-drain bias voltage $V_\mathrm{bias}$ and TG voltage $V_\mathrm{TG}$ at (a) BG voltage $V_\mathrm{BG}=9.0\,$V and SG voltage $V_\mathrm{SG}=-10.6\,$V and (b) $V_\mathrm{BG}=10.0\,$V and $V_\mathrm{SG}=-11.6\,$V. The displayed numbers in the plots correspond to the quantized conductance value of the plateaus in units of $4\,e^2/h$ and cyan lines trace the transitions between plateaus.} \label{fig:S8} \end{figure*} Source-drain bias spectroscopy is commonly used to probe the energy level spacing of the 1D subbands formed in QPCs \cite{Patel1990,Patel1991,Martin-Moreno1992}. We have used this measurement technique not only to extract the 1D subband spacing as described in the main text, but we also utilized it to extract the coupling factor $\alpha_\mathrm{TG}$. This parameter allows the conversion between $V_\mathrm{TG}$ that is applied to depopulate the 1D constriction and the energy levels of the QPC. $\alpha_\mathrm{TG}$ is then used to plot valley subband dispersion as a function of magnetic field obtained by our model. In Fig.\,\ref{fig:S8} color maps of the transconductance as functions of source-drain bias voltage $V_\mathrm{bias}$ and TG voltage $V_\mathrm{TG}$ are shown under two similar confinement conditions: Figure\,\ref{fig:S8}a at BG voltage $V_\mathrm{BG}=9.0\,$V and SG voltage $V_\mathrm{SG}=-10.6\,$V (see Fig.\,3 of the main text) and Fig.\,\ref{fig:S8}b at BG voltage $V_\mathrm{BG}=10.0\,$V and SG voltage $V_\mathrm{SG}=-11.6\,$V (corresponding to a confinement condition as presented in Fig.\,3 of the SM). The quantized conductance plateaus are labeled with the associated conductance values in units of $4\,e^2/h$. In both cases, as shown in the main text, the subband spacing increases as the confinement is weakened as well as the effect of Fabry-P\'{e}rot interferences (see main text). Additionally, we plot an overlaying set of cyan lines, marking transitions across subband edges, which are described by the expression ${ \alpha_\mathrm{TG}e(V_\mathrm{TG}-V_\mathrm{TG}^0) = E_N^\mathrm{QPC} \pm eV_\mathrm{bias}/2 }$\,. To fit the energy levels $E_N^\mathrm{QPC}$ we considered a linearly increasing energy level spacing of the size-quantized subbands $ \Delta_{N,N+1} \propto N $ (see Fig.\,3b of the main text). The resulting set of lines are in qualitative good agreement with the energy levels of the subbands in the transconductance pattern. From the slope of these lines we find the proportionality factor $\alpha_\mathrm{TG}=3.8\cdot 10^{-3}$ converting TG voltage $V_\mathrm{TG}$ into energy ${E=\alpha_\mathrm{TG}e(V_\mathrm{TG}-V_\mathrm{TG}^0)}$. The extracted gate coupling parameter from source-drain bias spectroscopy is used for plotting, together with the fitted energy levels, the energy levels of the magnetoelectic subbands derived from our model in terms of TG voltage $V_\mathrm{TG}$. \section{Weak confinement regime} \begin{figure} \centering \includegraphics[width=0.41\textwidth]{figS9} \caption{(a) Greyscale map of the differentiated differential conductance $\mathrm{d}G/\mathrm{d}V_\mathrm{TG}$ as a function of TG voltage $V_\mathrm{TG}$ and magnetic field $B$ at constant BG voltage $V_\mathrm{BG}=6.0\,$V and SG voltage $V_\mathrm{SG}=-6.7\,$V. Magnetoelectric subbands are visible as dark lines. (b) Valley subband dispersion as a function of magnetic field $B$ calculated with our model. Two colors distinguish between the two valleys. Lines here are understood as double lines, reflecting the twofold spin degeneracy in the model, except for the first thick cyan line , reflecting a fourfold (spin and ``touching-band'') degeneracy. In strong magnetic field, the Zeeman splitting (neglected in the model), enhanced at low densities by the Coulomb interaction, lifts the spin degeneracy. A similar effect lifts the ``accidental'' mexican-hat degeneracy.} \label{fig:S9} \end{figure} Here we present data and modelling of the 1D subband splitting in a weak confinement regime as defined in \cite{Knothe2018}. An additional ``accidental'' degeneracy appears featuring an $8\,e^2/h$ first quantized step due to the mexican-hat shape of the gapped band structure in BLG. The peculiar eightfold degeneracy is also observed in the depopulation of magnetoelectric subbands (see Fig.\,\ref{fig:S9}). Unlike the data presented in the main text, here the first Landau level fully emerges from the first size-quantized energy level of the QPC. This trend can be captured within our model by shifting the size-quantized energy levels $E_N^\mathrm{QPC}\rightarrow\, E_{N-1}^\mathrm{QPC}$ for $N\geq 1$.\\ \section{Model: Landau levels in BLG QPC} \subsection{Simplest model} Here, we present details of the simplified model used in the main text to describe qualitatively the magnetoelectric subbands. This model disregards the effects related to the fine structure of the BLG spectrum (trigonal warping and mexican-hat features) \cite{Knothe2018} and the Zeeman splitting. In this approximation, the $2\times 2$ Hamiltonian \cite{McCann2006} for the lowest conductance and valence bands in valleys $K$ and $K'$ can be written in the following form: \begin{gather} \hat{H}_K = \left( \begin{matrix} \Delta && \frac{1}{2m}(\hat{p}_x - i \hat{p}_y)^2 \\ \frac{1}{2m}(\hat{p}_x + i \hat{p}_y)^2 && -\Delta \end{matrix} \right),\\ \hat{H}_{K'} = \left( \begin{matrix} \Delta && \frac{1}{2m}(\hat{p}_x + i \hat{p}_y)^2 \\ \frac{1}{2m}(\hat{p}_x - i \hat{p}_y)^2 && -\Delta \end{matrix} \right). \end{gather} Here $m$ is the effective mass in BLG, $\Delta$ is half of the gap introduced by the displacement field, \textit{i.e.}, the difference between on-site energies of upper and lower graphene layers. This Hamiltonian acts in the space of sublattices A1 and B2 of the two layers. To describe the main features of the QPC conductance in a magnetic field, we consider an infinite BLG strip of width $2W$ in the $x$--$y$ plane, with the $y$ axis oriented across the strip. This model corresponds to an infinite gap outside the strip, and we use below the zero boundary conditions for the eigenfunctions at $y=\pm W$. The magnetic field $B$ is included in the Hamiltonian via the minimal coupling $\hat{\bf{p}} \rightarrow \hat{\bf{p}} - e {\bf{A}}/c$, where for the vector potential we use the gauge ${\bf{A}} = (-By,\, 0,\, 0)$. It is convenient to introduce the following notation \begin{gather} \hat{p}_+\!=\hat{p}_x-\frac{e}{c}A_x+i\hat{p}_y-i\frac{e}{c}A_y= -i\hbar\frac{\partial}{\partial x}-\hbar \left(\frac{y}{l_B^2}-\frac{\partial}{\partial y}\right), \notag \\ \hat{p}_-\!=\hat{p}_x-\frac{e}{c}A_x-i\hat{p}_y+i\frac{e}{c}A_y= -i\hbar\frac{\partial}{\partial x}-\hbar\left(\frac{y}{l_B^2}+\frac{\partial}{\partial y}\right), \notag \end{gather} where $l_B = \sqrt{\hbar c / e B}$ is the magnetic length and the commutator of $p_\pm$ is nonzero in a finite magnetic field: \begin{equation} \left[\hat{p}_-,\, \hat{p}_+ \right] = \frac{2 \hbar^2}{l_B^2}. \end{equation} In the absence of magnetic field, the conductance steps are determined by the size-quantization levels at zero momentum along the strip. In strong magnetic fields, $l_B \ll W$ , the conductance is determined by the edge states corresponding to the bulk Landau levels, and we again can set $k_x = 0$ to obtain the steps in conductance. The spectrum for $k_x=0$ is found from the following equations for the two valleys: \begin{gather} \label{K1eq} K: \quad \begin{cases} \left(E^2 - \Delta^2 - \dfrac{\hat{p}_-^2 \hat{p}_+^2}{4m^2} \right)\!\psi_{A1}(y)=0, \\[0.2cm] \psi_{A1}(\pm W) = 0, \quad \psi_{B2}(\pm W) = 0, \end{cases} \\[0.5cm] \label{K2eq} K': \quad \begin{cases} \left(E^2 - \Delta^2 - \dfrac{\hat{p}_+^2 \hat{p}_-^2}{4m^2} \right)\!\psi_{A1}(y)=0, \\[0.2cm] \psi_{A1}(\pm W) = 0, \quad \psi_{B2}(\pm W) = 0. \end{cases} \end{gather} Introducing the operator \begin{equation} \hat{A} = \frac{1}{2m} \hat{p}_- \hat{p}_+ = \frac{1}{2m} \hbar^2 \left( - \frac{\partial^2}{\partial y^2} + \frac{y^2}{l_B^4} + \frac{1}{l_B^2} \right), \end{equation} we re-write equations for $\psi_{A1}$ as follows: \\ \begin{gather} K: \, \left[ E^2 - \Delta^2 - \left(\hat{A} + \frac{\hbar^2 }{2m l_B^2}\right)^2\! +\left(\frac{\hbar^2}{2m l_B^2}\right)^2 \right]\psi_{A1}\!=\!0\,, \label{K1eqSimple}\\ K': \, \left[ E^2 - \Delta^2 - \left(\hat{A} - \frac{3\hbar^2}{2m l_B^2}\right)^2\! +\left(\frac{\hbar^2 }{2m l_B^2}\right)^2 \right]\psi_{A1}\!=\!0\,. \label{K2eqSimple} \end{gather} It is convenient to introduce the auxiliary eigenfunctions and eigenenergies of the operator $\hat{A}$ with zero boundary conditions, \begin{gather} \label{Aeq} \hat{A} \varphi_n = \varepsilon_n^A \varphi_n\,, \quad \varphi_n(\pm W) = 0\,. \end{gather} The functions $\varphi_n$ form a complete basis and every function satisfying the zero boundary conditions at $\pm W$ can be decomposed over this basis: \begin{equation} \psi_{A1}^{(n)} = \sum_m a_m^{(n)} \varphi_m\,, \qquad \psi_{B2}^{(n)} = \sum_m b_m^{(n)} \varphi_m\,. \end{equation} One then substitutes $\psi_{A1}^{(n)}$ in Eqs.\ (\ref{K1eqSimple}) and (\ref{K2eqSimple}), which yields the eigenenergies (we consider here only positive energies) \begin{gather} E_n^K = \sqrt{\Delta^2 + \left( \varepsilon_n^A + \hbar \omega_B/2 \right)^2 - (\hbar \omega_B/2)^2 }\,, \\ E_n^{K'} = \sqrt{\Delta^2 + \left( \varepsilon_n^A -3 \hbar \omega_B/2 \right)^2 - (\hbar \omega_B/2)^2 }\,, \end{gather} where $\hbar \omega_B = \hbar^2/m l_B^2$ is the cyclotron energy. Note that the shifts of $\varepsilon_n^A$ for $K$ and $K'$ are different. This leads to the valley splitting by the magnetic field. The coefficients in the expansion of the eigenfunctions $\psi_{A1}$ and $\psi_{B2}^{(n)}$ in $\varphi_m$ are given by \begin{equation} a_m^{(n)} = \delta_{nm}, \quad b_m^{(n)}=\frac{\langle m |\hat{p}_+^2|n\rangle}{E_n^{K,K'}+\Delta}, \end{equation} where $\delta_{nm}$ is the Kronecker delta symbol.\\ In the absence of magnetic field, $B=0$, one expresses the size-quantization levels of the QPC, \begin{gather} \label{Ek1k2B0} E_n^{K,K'} = \sqrt{\Delta^2 + \left[E_n^{(0)}\right]^2}, \end{gather} through the energy levels in a quantum well with infinitly high walls, $E_n^{(0)}=\varepsilon_n^A(B=0)$. The zero-$B$ levels $E_n^{K,K'}$ of the QPC are degenerate in the valleys. In strong magnetic fields, $l_B \ll W$, the boundary conditions are not important (for $k_x = 0$) and one can use the same procedure to obtain the Landau levels in BLG: \begin{gather} \hat{A} \tilde{\psi}_{A1}^n = \frac{\hbar^2}{m l_B^2} (n + 1) \tilde{\psi}_{A1}^n, \\ E_n^{K} = \sqrt{\Delta^2 + \left(\frac{\hbar^2}{m l_B^2}\right)^2 (n+1)(n+2)}, \\ E_n^{K'} = \sqrt{\Delta^2 + \left(\frac{\hbar^2}{m l_B^2}\right)^2 (n-1)n}. \end{gather} To find the energy levels at intermediate magnetic fields, one needs to solve exactly the problem defined by Eqs.\ (\ref{K1eq}) and (\ref{K2eq}), which reduces to Eq. (\ref{Aeq}). However, even without finding the exact energy levels, one sees that the size-quantized wave functions with the given number $n$ transforms into the harmonic oscillator wave functions with the same number, i.e., $\psi_{A1}^n \xrightarrow[B \rightarrow \infty]{} \tilde{\psi}_{A1}^{n}$. In order to describe the experimental data, we used the following simplest interpolation formulas: \begin{gather} \varepsilon_n^A = \sqrt{\varepsilon_n^2 + (\hbar \omega_B)^2(n+1)^2}, \\ \varepsilon_n = \sqrt{\left(E_n^\text{QPC}\right)^2 + 2 \Delta E_n^\text{QPC}}. \end{gather} Here, $\Delta$ is used as a free fitting parameter, whereas $E_n^\mathrm{QPC}$ is extracted from source-drain bias spectroscopy as described above. For $\Delta$ we have used $35\,$meV in Fig.\,4d of the main text and $18\,$meV in Fig.\ 9b of the SM. This is why the model is considered as semi-phenomenological. Indeed, in order to describe the energy levels in a fully analytical manner, one should find the actual profile of the constriction, which is determined by the electrostatic properties of the setup. \subsection{From four- to eight-fold degeneracy} \begin{figure*} \centering \includegraphics[width=0.41\textwidth]{Delta003w18}\hspace*{0.5cm} \includegraphics[width=0.41\textwidth]{Delta003w30}\vspace*{0.5cm} \includegraphics[width=0.41\textwidth]{Delta006w12}\hspace*{0.5cm} \includegraphics[width=0.41\textwidth]{Delta006w18} \caption{Evolution of zero-$B$ energy bands in a QPC with increasing width of the QPC for for the two values of $\Delta$.} \label{fig:S10} \end{figure*} Here, we briefly discuss the additional ``accidental'' degeneracy of the conductance related to the mexican-hat feature of the BLG spectrum \cite{Knothe2018}. The starting point for the effective two-band Hamiltonian \cite{McCann2006} is the tight-binding Hamiltonian, \begin{align*} \hat{\mathcal{H}}=\begin{pmatrix} \epsilon_{A1} & -\gamma_0f(\vec{k}) & \gamma_4f(\vec{k}) & -\gamma_4 f^\star(\vec{k})\\ \gamma_0f^\star(\vec{k}) & \epsilon_{B1} & \gamma_1 & \gamma_4f(\vec{k})\\ \gamma_4 f^\star(\vec{k}) & \gamma_1 & \epsilon_{A2} & -\gamma_0 f(\vec{k})\\ -\gamma_3 f/\vec{k}) & \gamma_4 f^\star(\vec{k}) & -\gamma_0 f^\star(\vec{k}) & \epsilon_{B2} \end{pmatrix}, \end{align*} which acts on the orbital states $A1, B1, A2, B2$. Expanding $f(\vec{k})=\exp(ik_ya/\sqrt{3}) + 2 \exp(-ik_ya/2\sqrt{3})\cos(k_xa/2)$ around $\vec{K}_\pm=\pm(+4\pi/3a,0)$ and omitting both the $\gamma_4$ and the trigonal warping induced by $\gamma_3$, one obtains the effective four-band Hamiltonian \begin{align} \hat{\mathcal{H}}_4= \begin{pmatrix} \epsilon_{A1} & v\pi^\dagger & 0 & 0\\ v\pi & \epsilon_{B1} & \gamma_1 & 0\\ 0 & \gamma_1 & \epsilon_{A2} & v\pi^\dagger\\ 0 & 0 & v\pi & \epsilon_{B2} \end{pmatrix},\qquad \pi:=\xi p_x+i p_y, \end{align} where $v=\sqrt{3} a \gamma_0/2\hbar$ and $\xi=\pm 1$ labels the two valleys. The QPC channel is modelled by the $y$-dependence of the sublattice energies: \begin{align} \epsilon_{A1}=\epsilon_{B1}=U(y)+\Delta(y),\\ \epsilon_{A2}=\epsilon_{B2}=U(y)-\Delta(y). \end{align} In order to simplify the calculations, we further expand the $4\times 4$ Hamiltonian to get an effective two-band Hamiltonian for the low energy components $(\psi_{A1}, \psi_{B2})$. Assuming the step-like change of $U(y)$ and $\Delta(y)$ forming the channel, and $U(y)=U$ and $\Delta(y)=\Delta$ for $|y|<W$, the $2\times 2$ Hamiltonian in the channel region is given by \begin{gather} \hat{\mathcal{H}}_2 =\begin{pmatrix} U+\Delta & \ -\dfrac{\pi^{\dagger 2}}{2m} \\ -\dfrac{\pi^2}{2m} & \ U-\Delta \end{pmatrix}+ 2\Delta \frac{v^2}{\gamma_1^2}\begin{pmatrix} \pi^\dagger\pi & 0\\ 0 & -\pi\pi^\dagger \end{pmatrix}. \end{gather} In the simplest model above, we have neglected the second term proportional to $\Delta v^2/\gamma_1^2$. Below, we analyze the effect of this term. With the same boundary conditions as above, we solve the Schr\"odinger equation \begin{gather} \mathcal{H}_2 \begin{pmatrix} \psi_{A1}\\\psi_{B2} \end{pmatrix} = E\begin{pmatrix} \psi_{A1}\\\psi_{B2} \end{pmatrix}, \notag\\ \psi_{A1}(y= \pm W)=\psi_{B2}(y= \pm W)=0. \end{gather} Expressing $\psi_{B2}$ through $\psi_{A1}$, we get \begin{widetext} \begin{align} \left[ (E-U)^2 - \Delta^2-\frac{1}{(2m)^2}(\pi^\dagger\pi)^2+4\Delta^2 \frac{v^2}{\gamma^2} \pi^\dagger\pi -4\Delta^2 \frac{v^4}{\gamma_1^4}(\pi^\dagger\pi)^2\right]\psi_{A1}=0. \end{align} Without magnetic fields, all momentum operators commute and $\pi^\dagger\pi=p_x^2+p_y^2$. Thus this differential equation can be solved by the general ansatz \begin{align} \psi_{A1}(x,y)\propto \exp(\pm ik_x x)\exp(\pm ik_y y). \end{align} In principle, one should consider linear combinations with all four possibilities for the signs, but since there is no restriction along the $x$ direction, it is reasonable to chose even waves with one sign along this direction. Using this ansatz we get the following condition on the momenta \begin{align} (E-U)^2 - \Delta^2-\frac{\hbar^4}{(2m)^2}(k_x^2+k_y^2)^2+4\Delta^2 \hbar^2\frac{v^2}{\gamma^2} (k_x^2+k_y^2)-4\Delta^2\hbar^4 \frac{v^4}{\gamma_1^4}(k_x^2+k_y^2)^2=0. \end{align} To account for the boundary conditions along the $y$ direction, we introduce quantized momenta and the corresponding discrete energies \begin{align} k_y = \frac{n \pi}{2 W},\quad E_n^{(0)} =\frac{\hbar^2}{2m} \frac{n^2\pi^2}{4W^2},\quad n\in \mathbb{Z}, \end{align} yielding the continuous energy spectrum $E_n(k_x)$ in the form \begin{align} (E_n-U)^2 = \Delta^2+4\Delta^2\frac{v^4}{\gamma_1^4}(2mE_n^{(0)} +\hbar^2 k_x^2)^2 - 4\Delta^2 \frac{v^2}{\gamma_1^2}\left(2mE_n^{(0)}+\hbar^2k_x^2\right)+\frac{1}{(2m)^2}\left(2mE_n^{(0)} +\hbar^2k_x^2\right)^2 \end{align} In sufficiently wide channels, when the two conditions \begin{align} W>\frac{\pi \hbar v}{2 \gamma_1}\quad \text{and} \quad \Delta >\frac{\pi \hbar \gamma_1^2}{4 m v} \frac{1}{ \sqrt{4 W^2 \gamma _1^2-\pi ^2 \hbar^2 v^2}}, \end{align} are simultaneously satisfied, the two lowest bands ($n=0$ and $n=1$) touch at \begin{align} k_x=\pm\frac{\sqrt{16\Delta ^2 m^2 v^2 \left(4 \gamma _1^2 W^2 -\pi ^2\hbar^2 v^2\right)-\pi ^2\hbar^2 \gamma _1^4}}{2\sqrt{2} W \hbar \sqrt{16 \Delta ^2 m^2 v^4+\gamma _1^4}}. \end{align} \end{widetext} It is this touching of the bands \cite{Knothe2018}, see Fig. \ref{fig:S10}, that leads to the additional degeneracy of the lowest conductance steps, replacing the generic four-fold degeneracy by the eight-fold degeneracy in wide constrictions.

a837ca6238c6cf49a37ef45e8013d6e8f93bf8aa

\subsection{The events class --- \texorpdfstring{\pyth{EventsAtVelocity}}{EventsAtVelocity}} This class is defined in \pyth{events.py}. This class builds an events function, $w(\vec v)$, for DM of a particular mass. By default, it uses XENON1T. E.g., \begin{python} mass = 100. # GeV w = EventsAtVelocity(mass) \end{python} The main methods are the events functions in the galactic frame, e.g., \begin{python} velocity = 200. # km/s w(velocity) # Angle-averaged cos_theta = 0. w(velocity, cos_theta) # Not angle-averaged w.plot() # Plots events function \end{python} There are further methods for the events function in e.g., the earth frame. We can convolute with a velocity distributions by e.g., \begin{python} from relax import MB velocity_dist = MB() # Maxwell-Boltzmann with default parameters n_events = velocity_dist * w # This performs integral over velocity \end{python} \subsection{Poisson likelihood --- \texorpdfstring{\pyth{Poisson}}{Poisson}} This class is defined in \pyth{poisson.py} and calculates the likelihood $\mathcal{L}$, as a function of the cross section, assuming a Maxwell-Boltzmann velocity distribution by default. We use it by, e.g., \begin{python} mass = 100. # GeV w = EventsAtVelocity(mass) poisson = Poisson(w) sigma = 1E-42 # cm^2 poisson.loglike(sigma) \end{python} This requires an \pyth{EventsAtVelocity} instance. We can optionally specify a velocity distribution, e.g., \pyth{Poisson(w, velocity_dist=my_dist)}. There are further properties for inspecting results, e.g., \begin{python} relax.chi_squared_limit() # 9 relax.best_fit_sigma # Best-fit cross section \end{python} return the $90\%$ upper limit and best-fit cross section. \subsection{The relaxed Maxwellian class --- \texorpdfstring{\pyth{Relax}}{Relax}} This class is defined in \pyth{relax.py} and calculates the likelihood averaged upon an entropic prior, $\mean{\mathcal{L}}$, as a function of the cross section. We use it by, e.g., \begin{python} beta = 10 mass = 100. # GeV w = EventsAtVelocity(mass) relax = Relax(beta , w, isotropic=True) sigma = 1E-42 # cm^2 relax.loglike(sigma) \end{python} This requires the hyperparameter \pyth{beta } and an \pyth{EventsAtVelocity} instance. We can optionally specify a velocity distribution and whether we wish to assume isotropy. The further methods are similar to that for \pyth{Poisson}. \subsection{Averaging upon shape parameters --- \texorpdfstring{\pyth{Shape}}{Shape}} This class is defined in \pyth{shape.py} and calculates the likelihood averaged upon parametric uncertainties in the modal and escape velocities. The usage is similar to the \pyth{Relax} and \pyth{Poisson} classes, e.g., \begin{python} beta = 10 mass = 100. # GeV w = EventsAtVelocity(mass) relax = Relax(beta , w, isotropic=True) shape = Shape(relax) sigma = 1E-42 # cm^2 shape.loglike(sigma) \end{python} Note that we initialise a \pyth{Shape} instance with a \pyth{Relax} or \pyth{Poisson} instance. \section{Introduction} There is evidence from gravitational interactions for the existence of dark matter (DM) throughout our Universe (see e.g., Ref.~\cite{Bertone:2004pz}). Weakly interacting massive particles (WIMPs) are a popular candidate for DM, since they naturally arise in many well-motivated extensions of the Standard Model (SM), e.g., supersymmetry, and correctly predict the relic abundance of DM by the so-called WIMP miracle~\cite{Jungman:1995df}. As WIMPs must annihilate to SM particles in the early Universe, by crossing symmetry, we expect that WIMPs must scatter elastically with SM particles. No evidence of such scattering was found in direct detection (DD) experiments by, inter alia, XENON1T\xspace~\cite{Aprile:2018dbl}, LUX~\cite{Akerib:2016lao} or PandaX~\cite{Cui:2017nnn}, resulting in upper limits on the DM scattering cross section with nucleons, e.g., for a $35\,\text{GeV}$ DM particle the cross section must be less than about $10^{-46}\,\text{cm}^2$~\cite{Aprile:2018dbl}. The limits depend upon assumptions about the velocity distribution of DM. The correct treatment and impact of uncertainties in the velocity distribution are the subjects of this work. From only theoretical considerations we anticipate that the velocity distribution could be similar to a Maxwell-Boltzmann distribution, \begin{equation}\label{eq:mb} m(v, \cos\theta, \phi) \propto \begin{cases} v^2 e^{-\left(v / v_0\right)^2} & v < v_\text{esc}\\ 0 & v \ge v_\text{esc} \end{cases}, \end{equation} where $v_\text{esc}$ and $v_0$ are the escape and modal velocities, respectively, and $\int_{\vec v} m(\vec v) d^3v = 1$. Indeed, this is the distribution that is assumed by DD experiments, including XENON1T\xspace. There are parametric uncertainties in the escape and modal velocities. There are, furthermore, non-parametric uncertainties as we know that departures from Maxwell-Boltzmann are plausible (and in fact may be preferred; see e.g., Ref.~\cite{Necib:2018iwb,Mandal:2018efq}). A Maxwell-Boltzmann follows from assuming a spherically-symmetric, isothermal halo of collisionless DM particles with density $\rho(r) \propto 1/r^2$; each assumption is questionable (see e.g.,~Ref.~\cite{Ibarra:2017mzt}). We recently proposed a non-parametric treatment of this state of knowledge~\cite{Fowlie:2017ufs} using the formalism of quantified maximum entropy (QME). We did not assume any particular parametric distribution for the velocity distribution; instead, we constructed an entropic prior for the velocity distribution that peaked at a Maxwell-Boltzmann and penalised departures from Maxwell-Boltzmann according to the relative entropy, \begin{equation}\label{eq:entropy} S[f, m] = - \int_{\vec v} f(\vec v) \ln \left(\frac{f(\vec v)}{m(\vec v)}\right) \,d^3v. \end{equation} The techniques in Ref.~\cite{Fowlie:2017ufs} relied on a Laplace approximation and were difficult to apply. A further drawback of QME is that the results depend upon details of the discretization of the velocity and that in the continuum limit it suffers from the law of large numbers, such that it overwhelming favours a Maxwell-Boltzmann (see Sec.~\ref{sec:recap}). In this work we present a similar entropic prior that overcomes this drawback. In Sec.~\ref{sec:dd}, we review our treatment of the expected number of signal events in a DD experiment. In Sec.~\ref{sec:recap}, we recapitulate the pertinent aspects of Ref.~\cite{Fowlie:2017ufs} and the merits of our use of the relative entropy, before presenting a formula for the marginalisation of an entropic prior based on a multinomial process. In Sec.~\ref{sec:numeric} we apply it to recent results from XENON1T\xspace (2018). Lastly, we conclude in Sec.~\ref{sec:conclusions}. We furthermore motivate and discuss our new formula in App.~\ref{app:average} and App.~\ref{app:proof}, respectively, and present our code that implements it in App.~\ref{app:code}. \section{Events function}\label{sec:dd} The number of expected events, $\lambda$, in a DD experiment such as XENON1T\xspace may be expressed as an expectation of the velocity distribution in the galactic frame, $f$, \begin{equation}\label{eq:lambda} \lambda = \mean{w}_f \equiv \int f(\vec v) \cdot w(\vec v) \,d^3 v. \end{equation} where $\mean{y}_f \equiv \int y(\vec v) \cdot f(\vec v) \,d^3 v$ indicates an average over the velocity distribution, $f$, and the function $w(\vec v)$ defines the number of expected events as a function of the DM velocity in the galactic frame, $\vec v$. We define it in the laboratory frame and transform it to the galactic frame by a Galilean boost. In the laboratory frame it is isotropic and may be written as \begin{equation} w_\text{lab}(v) = \frac{2 MT \rho}{m_\chi} \cdot v \cdot \int \frac{\,d\sigma}{\,d q^2} \cdot \Phi(E) \,d E + b, \end{equation} where $MT$ is the exposure; $\Phi(E)$ is the detector efficiency at recoil energy $E$; $b$ is the expected number of background events; $m_\chi$ and $\rho$ are the mass and local density of DM, respectively; and ${\,d\sigma}/{\,d q^2}$ is the differential cross section. We assume that the interactions are velocity and momentum independent such that the differential cross section may be written as \begin{equation} \frac{\,d\sigma}{\,d q^2} = \frac{\sigma}{4\mu^2 v^2} \cdot F^2(q) \cdot \theta(q_\text{max} - q), \end{equation} where $\sigma$ is the scattering cross section at zero momentum; the momentum $q^2 = 2 m_n E$; by kinematics $q_\text{max} = 2 \mu v$; $\mu$ is the reduced mass of the DM and nucleon; $m_n$ is the nucleon mass; $F$ is a nuclear form-factor; and $\theta$ denotes a stepfunction. Our treatment of the expected number of events differs from the canonical one (see e.g., Ref.~\cite{Workgroup:2017lvb}) only in our presentation; we reversed the order of the energy and velocity integrals and boosted $w(\vec v)$ to the galactic frame rather than $f(\vec v)$ to the laboratory frame. This approach was introduced in Ref.~\cite{DelNobile:2013cva,Gondolo:2017jro}. \section{Entropic prior}\label{sec:recap} We recently proposed treating uncertainties in the velocity distribution with quantified maximum entropy~\cite{Fowlie:2017ufs}. Rather than assuming any particular velocity distribution, we constructed a prior upon possible velocity distributions and averaged upon it. The prior penalised departures from a default distribution by the relative entropy, \begin{equation}\label{eq:QME} \pg{\vec f}{\vec m} \propto \frac{e^{\beta S[f, m]}}{\prod_i \sqrt{f_i}} \cdot \delta\left(\sum f_i - 1\right). \end{equation} where $S[f, m]$, defined in Eq.~\ref{eq:entropy}, is the entropy of the velocity distribution, $f$, relative to a Maxwellian, $m$, and we denote a discrete distribution across $r$ bins by $\vec f = \{f_1, f_2, \dots, f_r\}$, and similarly for the default distribution $\vec m$. When $f = m$, the entropy vanishes, and it is otherwise negative. The hyperparameter $\beta$ represented the strength of our conviction that the velocity distribution is Maxwellian; as $\beta\to \infty$, our uncertainty vanished and the prior selected $f = m$, and as $\beta\to0$, the penalty for departures from the default model, $m$, vanished. Thus by varying $\beta$, we interpolated between a halo-independent ($\beta\to 0$) and halo-dependent ($\beta\to\infty$) approach. Following Ref.~\cite{Fowlie:2017ufs}, a frequentist treatment was proposed~\cite{Ibarra:2018yxq} in which departures were measured by \begin{equation} \Delta[f, m] = \max_{\vec v} \left|\frac{f(\vec v) - m(\vec v)}{m(\vec v)}\right|. \end{equation} Rather than averaging upon a set of velocity distributions, a distribution was found that maximised the likelihood subject to an upper bound on the discrepancy, $\Delta[f, m]$. There were two main advantages to our approach. First, there is an information theoretic meaning to the relative entropy and the entropic prior may be derived as a unique choice subject to modest axioms (see e.g., Ref.~\cite{Skilling1988,Skilling1989}), whereas $\Delta[f, m]$ is ad hoc. Second, we coherently incorporated uncertainty by marginalising rather than profiling. Only the former respects the fact that the plausibilities of disjoint propositions should sum. We note, however, a subtle drawback in the QME prior: the results are sensitive to the discretization of the velocity and are counter-intuitive in the continuum limit~\cite{doi:10.1111/1467-9868.00065}. In that limit there are an infinite number of contributions to the probability in a macroscopic interval, $\Delta v$, i.e., in the continuum limit, the sum, \begin{equation} f_i \equiv \P{v \le v^\prime \le v + \Delta v} = \sum_{i=1}^n \P{v + \tfrac{(i - 1)\Delta v}{n} \le v^\prime \le v + \tfrac{i \Delta v}{n}}, \end{equation} contains an infinite number of terms as $n\to\infty$. By the law of large numbers, for the QME prior the probability in such an interval equals its expected value, $f_i \to m_i$. This means that the QME prior overwhelming favours the default distribution on macroscopic scales. There is thus a delicate interplay between $\beta$ and the discretization of the velocity; although $\beta$ penalises departures from the default distribution, in the continuum limit it operates at the microscopic scale, $dv$. On macroscopic scales, by the law of large numbers, departures average away. We could avoid this problem by specifying a default distribution (e.g., a Maxwellian) and a finite scale $\Delta v$ below which we wish to penalise departures from it. We instead avoid it by generating velocity distributions by scattering $\beta$ quanta of probability on possible velocities. We initially discretize the velocity distribution, $f_i \approx f(\vec v_i) \Delta v^3$, by dividing the velocity into $r$ bins of volume $\Delta v^3$, but ultimately we take a continuum limit. We assume that the quanta fall into particular bins with probabilities from the default model, $m_i$. This is a multinomial process. The law of large numbers strikes only in the limit $\beta\to\infty$, forcing the velocity distribution to the default one, as desired. This is detailed in App.~\ref{app:average}. This choice is motivated by the fact that just like the QME prior, we find that our prior penalizes departures from a parametric model according to the relative entropy, \begin{equation} \Pg{\vec f}{\vec m} \propto e^{\beta S[\vec f, \vec m]}. \end{equation} In fact, our prior approximates the QME one when the number of bins in the QME prior $r \lesssim \beta$, which implies a bin width $\Delta v \gtrsim v_\text{esc} / \beta$. It differs from the QME prior in that it requires $f_i$ to be quantized in multiples of $1/\beta$. To incorporate uncertainty in the velocity distribution, we begin from a Poisson probability for observing $q$ events given that $\lambda$ events were expected, \begin{equation}\label{eq:poisson} \mathcal{L} \equiv \Pg{q}{\lambda} = \frac{e^{-\lambda} \lambda^q}{q!}. \end{equation} We note that the expected number of events is a function of the DM mass, scattering cross section with nucleons and velocity distribution, i.e, $\lambda \equiv \lambda(m_\chi, \sigma, \vec f)$. We want to marginalise upon the velocity distribution, i.e., calculate the sum, \begin{equation}\label{eq:integral} \mean{\mathcal{L}} \equiv \Pg{q}{m_\chi, \sigma} = \sum \Pg{q}{m_\chi, \sigma, \vec f} \cdot \Pg{\vec f}{\vec m} = \sum \frac{e^{-\lambda} \lambda^q}{q!} \cdot \Pg{\vec f}{\vec m}, \end{equation} where we marginalised the velocity distribution over our prior, $\Pg{\vec f}{\vec m}$. We compute the sum exactly in App.~\ref{app:proof}. For the experiment that we investigate, XENON1T\xspace (2018), the number of observed events was $q=2$ such that using Eq.~\ref{eq:result} we find, \begin{equation}\label{eq:average_like} \mean{\mathcal{L}} = \frac{1}{2} \mean{e^{-w / \beta}}_m^\beta \left( \frac{\beta - 1}{\beta} \frac{\largemean{w e^{-w / \beta}}_m^2}{\largemean{e^{-w / \beta}}_m^2} + \frac{1}{\beta} \frac{\largemean{w^2 e^{-w / \beta}}_m}{\largemean{e^{-w / \beta}}_m} \right), \end{equation} for integer $\beta\ge 1$ and where $\mean{y}_m \equiv \int y(\vec v) \cdot m(\vec v) \,d^3 v$ indicates an average over the default model, i.e., a Maxwell-Boltzmann. For the common case in which no events were observed, $q=0$, we find, \begin{equation} \mean{\mathcal{L}} = \mean{e^{-w / \beta}}_m^\beta. \end{equation} The marginalized likelihoods resemble our original Poisson likelihood in Eq.~\ref{eq:poisson}, which in a similar notation for $q=2$ reads \begin{equation}\label{eq:beta_infinity} \mathcal{L} = \frac12 e^{-\mean{w}_m} \mean{w}_m^2. \end{equation} The changes result from our incorporation of the uncertainty in the velocity distribution. In the limit in which our uncertainty vanishes, $\beta\to\infty$, we indeed recover Eq.~\ref{eq:beta_infinity}, \begin{equation} \lim_{\beta\to\infty} \mean{\mathcal{L}} = \mathcal{L}. \end{equation} We thus interpret our treatment as a non-parametric relaxation of a default distribution. We cannot, however, throw away all information about the default model as our multinomial process requires $\beta \ge 1$. \subsection{Isotropy}\label{sec:isotropy} Our averaged likelihood in Eq.~\ref{eq:average_like} makes no assumptions about isotropy --- it averages over anisotropic and isotropic velocity distributions weighted by an entropic prior. If we wish to assume isotropy, we must pick an isotropic default model (such as the Maxwell-Boltzmann), \begin{equation} m(v, \cos\theta, \phi) = \frac{1}{4\pi} m(v) , \end{equation} and omit an entropic prior for the angular variables, $m(\cos\theta, \phi)$, such that there are no deviations from isotropy. In our formalism, this is mathematically equivalent to replacing the event function by \begin{equation} w(\vec v) \to w(v) = {1}/{4\pi}\int w(\vec v) \,d\cos\theta \,d\phi \end{equation} throughout, i.e., using an angle-averaged event function. \section{Impact on XENON1T\xspace (2018) limits}\label{sec:numeric} \subsection{Events function}\label{sec:events} \begin{figure}[t] \centering % \begin{subfigure}[t]{0.48\textwidth} \centering \includegraphics[height=0.26\textheight]{maxwellian.pdf} \caption{The XENON1T\xspace $90\%$ bound} \label{fig:limit} \end{subfigure} % \begin{subfigure}[t]{0.48\textwidth} \centering \includegraphics[height=0.26\textheight]{events.pdf} \caption{Angle-averaged events function} \label{fig:events} \end{subfigure} \caption{Validation of our events function, $w(\vec v)$. In \refsubfig{fig:limit} we compare the XENON1T\xspace $90\%$ bound (solid violet); our reproductions from $\Delta\chi^2$ (solid green) and Poisson statistics (solid blue); and reproductions from Ref.~\cite{Workgroup:2017lvb,Athron:2018hpc} (dashed green) and Ref.~\cite{github} (dashed blue). In \refsubfig{fig:events} we show the angle-averaged number of events per velocity, $w(v)$, normalised to the number of expected events for five different DM masses (solid lines). For reference we show a Maxwell-Boltzmann distribution (dashed brown) with a separate vertical axis.}\label{fig:reproduction} \end{figure} To validate our treatment of the XENON1T\xspace (2018) experiment, we first reproduced the XENON1T\xspace (2018) $90\%$ upper bound on the spin-independent scattering cross section with nucleons assuming a Maxwell-Boltzmann velocity distribution. With Eq.~\ref{eq:poisson} we calculated the likelihood of $2$ events in $278.8$ days, given a DM signal with a particular mass and cross section, and $1.62$ expected background events. We considered only events in the reference region of the $900\,\text{kg}$ inner detector, $M = 0.475 \times 900\,\text{kg}$.\footnote{See Tab.~1 of Ref.~\cite{Aprile:2018dbl}.} We show our results in Fig.~\ref{fig:limit}. We calculated a $90\%$ limit at $3.7$ signal events from Poisson statistics, which closely matches a similar reproduction~\cite{github}. Our $90\%$ limit from Wilks' theorem is similar to that from \texttt{DDCalc}~\cite{Workgroup:2017lvb,Athron:2018hpc}, which used additional binning information but in only the $650\,\text{kg}$ inner detector. The minor differences between our limit and the XENON1T\xspace one were expected since XENON1T\xspace used spectral information and an unbinned analysis. Thus we are satisfied that we successfully computed the events function for the XENON1T\xspace experiment, $w(\vec v)$, which appears in our treatment of the uncertainty in the velocity distribution. We plot our angle-averaged events function for five different DM masses in Fig.~\ref{fig:events}. We find that for lower DM masses, as expected from kinematics, the events function peaks at higher velocity. This explains the loss in sensitivity for light DM masses: for light DM masses, the signal vanishes as the events functions favours high velocities but the velocity distribution is zero beyond the escape velocity. For higher DM masses, the sensitivity deteriorates as the number density, $\rho / m_\chi$, shrinks as the DM mass increases. We note that for $m_\chi \simeq 60 \,\text{GeV}$ the events function is particularly flat. \subsection{Isotropic velocity distribution} \begin{figure}[t] \centering \begin{subfigure}[t]{0.48\textwidth} \centering \includegraphics[height=0.26\textheight]{relaxed_100_isotropic.pdf} \caption{$\beta = 100$} \label{fig:isotropic_a} \end{subfigure} \hspace{-0.75cm} \begin{subfigure}[t]{0.48\textwidth} \centering \includegraphics[height=0.26\textheight]{relaxed_10_isotropic.pdf} \caption{$\beta = 10$} \label{fig:isotropic_b} \end{subfigure} \begin{subfigure}[t]{0.48\textwidth} \centering \includegraphics[height=0.26\textheight]{relaxed_5_isotropic.pdf} \caption{$\beta = 5$} \label{fig:isotropic_c} \end{subfigure} \hspace{-0.75cm} \begin{subfigure}[t]{0.48\textwidth} \centering \includegraphics[height=0.26\textheight]{relaxed_1_isotropic.pdf} \caption{$\beta = 1$} \label{fig:isotropic_d} \end{subfigure} \caption{Chi-squared from XENON1T\xspace (2018) on the $(m_\chi, \sigma)$ plane for four values of the parameter, $\beta$, assuming an isotropic velocity distribution. We show the $90\%$ limit assuming an entropic prior (solid green) and that from a Maxwell-Boltzmann (dashed blue).} \label{fig:isotropic} \end{figure} We begin by assuming an isotropic velocity distribution in the galactic frame. To investigate the dependence of DD searches on the velocity distribution, we marginalise possible departures from a Maxwell-Boltzmann. The hyperparameter $\beta$ governs the strength of our belief in a Maxwellian distribution. As we relax the Maxwellian by decreasing $\beta$, the limit could, a priori, weaken or strengthen. In Fig.~\ref{fig:isotropic} we show maps of \begin{equation} \Delta\chi^2 \equiv -2 \ln \frac{\mean{\mathcal{L}}}{\max\limits_{m_\chi, \sigma} \mean{\mathcal{L}}} \end{equation} where the average likelihood, $\mean{\mathcal{L}}$, is a function of the hyperparameter, $\beta$, and the DM mass and cross section. We calculate $90\%$ limits from a hybrid approach~\cite{COUSINS1992331} at $\Delta\chi^2 \simeq 1.64$, following Ref.~\cite{Workgroup:2017lvb,Athron:2018hpc}. Although we could calculate credible regions in a completely Bayesian approach, we note that hybrid approaches are common in experimental searches and could be adopted by DD experiments themselves. For $\beta = 100$ in Fig.~\ref{fig:isotropic_a}, we see that the $90\%$ limit approximately matches that from a Maxwell-Boltzmann, i.e., at $\beta=100$ we find that we are not sensitive to departures from a Maxwellian. As we decrease our belief in a Maxwellian distribution to $\beta=10$ and $\beta=5$ in Fig.~\ref{fig:isotropic_b} and Fig.~\ref{fig:isotropic_c}, we see that the XENON1T\xspace limit becomes slightly weaker than that from a Maxwell-Boltzmann, particularly for DM masses $m_\chi \lesssim 60\,\text{GeV}$. As shown in Fig.~\ref{fig:events}, the angle-averaged events function for light DM exhibits a sharp increase in the tail of the Maxwellian distribution and thus XENON1T\xspace is particularly sensitive to the velocity distribution for light DM. Once we relax to $\beta=1$, Fig.~\ref{fig:isotropic_d}, we find pronounced differences for light DM, $m_\chi \lesssim 60\,\text{GeV}$. For heavy DM, $m_\chi \gtrsim 60\,\text{GeV}$, however, the limit stubbornly remains close to that from a Maxwellian distribution. We can understand this by noting that the angle-averaged events function in Fig.~\ref{fig:events} is quite flat for DM masses greater than about $60\,\text{GeV}$. As the events function is approximately constant in velocity, we are not sensitive to the velocity distribution. It is somewhat inevitable that there exists a DM mass at which the angle-averaged events function is approximately flat. For light DM by kinematics we see a sharp rise in the events function near the escape velocity. For heavy DM, high velocities are mildly suppressed. Thus, we find an approximately flat distribution in the transition between these regimes. The fact that this occurs at about $60\,\text{GeV}$ is interesting, as it is close to peak sensitivity and approximately corresponds to $m_\chi \simeq m_h / 2$. Thus, DD limits on WIMPs that annihilate through an on-shell Higgs boson are particularly robust with respect to the velocity distribution. \subsection{Anisotropic velocity distribution} \begin{figure}[t] \centering \begin{subfigure}[t]{0.48\textwidth} \centering \includegraphics[height=0.26\textheight]{relaxed_100_anisotropic.pdf} \caption{$\beta = 100$} \label{fig:anisotropic_a} \end{subfigure} \hspace{-0.75cm} \begin{subfigure}[t]{0.48\textwidth} \centering \includegraphics[height=0.26\textheight]{relaxed_10_anisotropic.pdf} \caption{$\beta = 10$} \label{fig:anisotropic_b} \end{subfigure} \begin{subfigure}[t]{0.48\textwidth} \centering \includegraphics[height=0.26\textheight]{relaxed_5_anisotropic.pdf} \caption{$\beta = 5$} \label{fig:anisotropic_c} \end{subfigure} \hspace{-0.75cm} \begin{subfigure}[t]{0.48\textwidth} \centering \includegraphics[height=0.26\textheight]{relaxed_1_anisotropic.pdf} \caption{$\beta = 1$} \label{fig:anisotropic_d} \end{subfigure} \caption{Chi-squared from XENON1T\xspace (2018) on the $(m_\chi, \sigma)$ plane for four values of the parameter, $\beta$, permitting anisotropic departures from a Maxwellian. We show the $90\%$ limit assuming an entropic prior (solid green) and that from a Maxwell-Boltzmann (dashed blue).} \label{fig:anisotropic} \end{figure} We now relax our assumption of isotropy and place our prior on the magnitude and angular components of the velocity distribution, $\cos\theta$ and $\phi$. This permits anisotropic departures from a Maxwellian distribution. This is important since anisotropy in the galactic frame could be tuned such that the flux of DM particles is zero in the laboratory frame, lifting the limit altogether. In Fig.~\ref{fig:anisotropic} we show the $90\%$ limit for four choices of $\beta$. As expected and as in the isotropic case, for $\beta=100$ in Fig.~\ref{fig:anisotropic_a} we find that the limit is approximately that from an isotropic Maxwellian. As we decrease to $\beta=10$ and $\beta=5$ in Fig.~\ref{fig:anisotropic_b} and Fig.~\ref{fig:anisotropic_c}, we see that the limit weakens. The weakening, although more pronounced than in the isotropic case, remains limited. Even once we relax to $\beta=1$, the weakening is modest, and for DM masses $m_\chi \gtrsim 60\,\text{GeV}$, stubbornly remains close to the Maxwellian limit, as in the isotropic case. \begin{figure}[t] \begin{subfigure}[t]{0.48\textwidth} \centering \includegraphics[height=0.26\textheight]{impact_isotropic.pdf} \caption{Isotropic} \label{fig:impact_isotropic} \end{subfigure} % \begin{subfigure}[t]{0.48\textwidth} \centering \includegraphics[height=0.26\textheight]{impact_anisotropic.pdf} \caption{Anisotropic} \label{fig:impact_anisotropic} \end{subfigure} \caption{Change in the $90\%$ limit, relative to the limit from a Maxwellian, from non-parametric uncertainties in the velocity distribution. The hyperparameter $\beta$ governs the strength of our belief in a Maxwell-Boltzmann. We \refsubfig{fig:impact_isotropic} assume an isotropic velocity distribution and \refsubfig{fig:impact_anisotropic} include non-parametric uncertainties in the angular dependence of the velocity distribution.} \label{fig:impact} \end{figure} In Fig.~\ref{fig:impact} we compare our isotropic and anisotropic limits by showing the changes in the limit as we relax the Maxwellian side by side. We divide the limit with that from a Maxwellian. We see that the impact of uncertainty in the velocity distribution in the isotropic case, Fig.~\ref{fig:impact_isotropic}, is mild, as it only substantially weakens the limit for light DM and once almost all information about the distribution is disregarded, $\beta \simeq 1$. Even in the most extreme cases, the limit is weakened by less than an order of magnitude. The anisotropic case, in Fig.~\ref{fig:impact_anisotropic}, on the other hand, is slightly more dramatic, with noticeable weakening by up to two orders of magnitude for $\beta = 1$. Nevertheless, for DM masses greater than about $60\,\text{GeV}$ the limit stays similar to that from a Maxwell-Boltzmann. We do not investigate DM masses less than $10\,\text{GeV}$, as the limit is acutely sensitive to precision in the tiny efficiency at low recoil energies. We note, though, that in all cases the limit weakens; a priori, it could have strengthened. \subsection{Parametric uncertainties} Finally, we consider the impact of parametric uncertainties in the shape parameters of the default distribution. We suspect that the modal and escape velocities are approximately $v_0 = 235 \pm 20 \,\text{km/s}$ and $v_\text{esc} = 550 \pm 35 \,\text{km/s}$~\cite{Workgroup:2017lvb}. We treat them in four ways: we fix them to their central values; marginalise Gaussian uncertainties in them; permit them to vary by as much as $3\sigma$; and profile Gaussian uncertainties in them. We find, as expected, that their impact is extremely limited. In Fig.~\ref{fig:shape} we show the $90\%$ limits from our four treatments. For a Maxwellian, Fig.~\ref{fig:shape_maxwellian}, the limits from fixing, marginalising and profiling are extremely similar. The impact of parametric uncertainties is noticeable only when they are permitted to vary by $3\sigma$ without any penalty. The story for the relaxed Maxwellian, Fig.~\ref{fig:shape_relaxed}, is similar; the limit cannot be significantly changed by parametric uncertainties. \section{Discussion and conclusions}\label{sec:conclusions} We presented a new technique for treating non-parametric uncertainties that applies to any counting experiment for which the expected number of events may be written as an expectation, e.g., a counting experiment at a collider experiment where the number of events depends upon integrating over a parton distribution function. We treated non-parametric uncertainties with an multinomial prior that contained a hyperparameter, $\beta$, which governed the strength of our conviction in a particular parametric model. We detail our result in App.~\ref{app:average} and App.~\ref{app:proof}. We briefly mentioned in App.~\ref{app:multiple} that our result could generalise to multiple independent counting experiments, though leave a detailed discussion and example to a future work. Our prior was motivated by quantified maximum entropy; but unlike it, it did not suffer from problems with the law of large numbers in the continuum limit. The prior, however, quantized probabilities in multiples of $1/\beta$. It may be desirable to marginalize a prior that permits arbitrary probabilities. We applied our technique to limits on the scattering cross section of DM from the XENON1T\xspace experiment. We validated our model of XENON1T\xspace by reproducing the limit with a Maxwell-Boltzmann distribution. Once we relaxed that distribution, we found only a mild impact from non-parametric uncertainties in the velocity distribution of DM. The impact was greatest when non-parametric uncertainty was included in the angular dependence of the velocity distribution, i.e., in the anisotropic case. For $\beta = 1$ and DM masses less than about $60\,\text{GeV}$, non-parametric uncertainties weakened the upper limit by about two orders of magnitude. Assuming isotropy, however, it weakened by less than one order of magnitude. For DM masses greater than about $60\,\text{GeV}$ the weakening was always mild and the limit at about $60\,\text{GeV}$ was particularly robust with respect to the velocity distribution, as we found that for that mass the events function was approximately flat. The non-parametric uncertainties were, however, significantly greater than the parametric ones. Indeed, after marginalising or profiling them, the impact from uncertainties in the modal and escape velocities was negligible. Our approach is somewhat in contrast with Ref.~\cite{Ibarra:2018yxq}; whereas we constructed a multinomial prior upon velocity distributions and marginalised it, Ref.~\cite{Ibarra:2018yxq} selected the most extreme distributions from a set. The former reflects our Bayesian treatment of uncertainty; the latter a frequentist approach. Whilst our results appear to be consistent with Ref.~\cite{Ibarra:2018yxq}, in that limits from our marginalised likelihood appear to lie between the extremes found in Ref.~\cite{Ibarra:2018yxq}, our results suggest that the impact of uncertainty in the velocity distribution is mild.\footnote{Note, however, that Ref.~\cite{Ibarra:2018yxq} used 2017 XENON1T\xspace results \cite{Aprile:2017iyp}, in which no events were observed.} The formalism itself should lend itself to inclusion in global fits of DM models, as it is not especially computationally demanding. We briefly describe our publicly available implementation in App.~\ref{app:code}. The ordinary treatment of DD experiments requires a single integral upon the velocity distribution; this one requires an integral for every observed event. Previously global fits of DM models, e.g., Ref.~\cite{Balazs:2017ple,Athron:2018ipf}, included at most parametric uncertainties. We can now, however, incorporate coherently all major sources of uncertainty in DD experiments in official limits and global fits. \begin{figure}[t] \begin{subfigure}[t]{0.48\textwidth} \centering \includegraphics[height=0.26\textheight]{shape_maxwellian.pdf} \caption{Maxwellian} \label{fig:shape_maxwellian} \end{subfigure} % \begin{subfigure}[t]{0.48\textwidth} \centering \includegraphics[height=0.26\textheight]{shape_relaxed.pdf} \caption{$\beta=1$} \label{fig:shape_relaxed} \end{subfigure} \caption{The $90\%$ limit from XENON1T\xspace with parametric uncertainties in the velocity distribution. We show $90\%$ limits with the modal and escape velocities fixed (solid blue), averaged (dashed green) and profiled (dotted red). We show the impact on \refsubfig{fig:shape_maxwellian} a Maxwellian distribution and \refsubfig{fig:shape_relaxed} with parametric uncertainties from our entropic prior.} \label{fig:shape} \end{figure} \bibliographystyle{JHEP} \subsection{Generalized result}\label{app:multiple} We may in fact generalise our result in Eq.~\ref{eq:general_result} to e.g., the expectation of \begin{equation} \lambda_1^{q_1} \lambda_2^{q_2} = \left(\sum (w_1)_i n_i / \beta \right)^{q_1} \left(\sum (w_2)_i n_i / \beta \right)^{q_2}, \end{equation} where $n_i$ follow a multinomial distribution with $\beta$ trials and event probabilities $m_i$. Eq.~\ref{eq:general_result} applies to the special case in which $w_1 = w_2$. Such a result would be useful for e.g., marginalizing the uncertainty in a likelihood from two or more independent counting experiments. In this case, we may in fact apply Eq.~\ref{eq:general_result} by using $q = q_1 + q_2$ and replacing e.g., $\mean{w^n}_m$ terms with appropriate generalisations involving $\mean{w_1}_m$ and $\mean{w_2}_m$ etc. For example, \begin{equation} \begin{split} \mean{\lambda_1^2 \lambda_2^1} &= \frac{(\beta - 1)(\beta - 2)}{\beta^2} \largemean{w_1}_m^2 \largemean{w_2}_m + \frac{(\beta - 1)}{\beta^2} \left(\largemean{w_1^2}_m \largemean{w_2}_m + 2 \largemean{w_1 w_2}_m \largemean{w_1}_m \right)\\ & + \frac{1}{\beta^2} \largemean{w_1^2 w_2}_m, \end{split} \end{equation} which bears a close resemblence to Eq.~\ref{eq:q_3_case}. We leave a detailed discussion of this generalised case and an application to the likelihood from multiple independent counting experiments to a future work.

5f8641327c519a2f18bd03a67ad74a04778bc572

\section{Introduction} An {\em operator algebra} for us is a closed subalgebra of $B(H),$ for a complex Hilbert space $H.$ Here we study operator algebras with involution. Examples include the operator $*$-algebras occurring in noncommutative differential geometry studied recently by Mesland, Kaad, Lesch, and others (see e.g.\ \cite{Mes14,KL,BKM} and references therein), (complexifications) of real operator algebras, and an operator algebraic version of the {\em complex symmetric} operators studied by Garcia, Putinar, Wogen, Zhu, and many others (see \cite{GP} for a survey, or e.g.\ \cite{GW}). By an {\em operator $\ast$-algebra} we mean an operator algebra with an involution $\dagger$ making it a $\ast$-algebra with $\Vert [a_{ji}^{\dagger}]\Vert=\Vert [a_{ij}]\Vert$ for $[a_{ij}] \in M_n(A)$ and $n \in \Ndb$. Here we are using the {\em matrix norms} of operator space theory (see e.g.\ \cite{Pisbk}). This notion was first introduced by Mesland in the setting of noncommutative differential geometry \cite{Mes14}, who was soon joined by Kaad and Lesch \cite{KL}. In several recent papers by these authors and coauthors they exploit operator $\ast$-algebras and involutive modules in geometric situations. Subsequently we noticed very many other examples of operator $\ast$-algebras, and other involutive operator algebras, occurring naturally in general operator algebra theory which seem to have not been studied hitherto. It is thus natural to investigate the general theory of involutive operator algebras, and this is the focus of the present paper. We are able to include a rather large number of results since many proofs are similar to their operator algebra counterparts in the literature (see e.g.\ \cite{BLM}). Thus we often need only discuss the new points that arise. However to follow some of the arguments the reader will need to have the operator algebra variant from the original papers in hand. It is worth saying that some of the arguments we are following are complicated, and so it is not clear in advance whether they have `involutive variants'. In fact some of the main theorems about operator algebras do not have operator $*$-algebra variants, so some work is needed to disentangle the items that do work. We make no attempt to be comprehensive for the sake of avoiding tedium. We will simply illustrate the main techniques and features, indicating what can be done. Many of the results are focused around `real positivity' in the sense of several recent papers of the first author and collaborators referenced in our bibliography. Some related theory and several complementary results can be found in the second authors PhD thesis \cite{Wang}. \subsection{Structure of our paper} In the rest of this section we give some background, perspective, and notations. In Section 2 we give several general results. For example we prove some facts about involutions on nonselfadjoint operator algebras and their relationship to the $C^*$-algebras they generate. As an application of some ideas in the theory of complex symmetric operators we characterize the symmetric operator algebras introduced in \cite{BComm}. This is a problem outstanding from the early years of operator space theory. Section 3 is devoted to examples of involutive operator algebras, for instance examples coming from operator space theory, subdiagonal algebras, model theory for contractions on a Hilbert space, and complex symmetric operators. In the remaining sections we restrict our focus, for specificity, to operator $\ast$-algebras. Many of our results involve {\em real positivity} in the sense of several recent papers of the first author and collaborators referenced in our bibliography, where the cones $\Rdb_+ {\mathfrak F}_A$ and ${\mathfrak r}_A$ (defined below) act as a substitute for the usual positive cone in a $C^*$-algebra. In an operator $*$-algebra $A$, ${\mathfrak F}_A$ and ${\mathfrak r}_A$ are closed under the involution, and taking $n$th roots commutes with the involution. Thus the theory of real positivity studied in many of the first authors recent papers will have good involutive variants. In Section 4 we discuss contractive approximate identities, Cohen factorization for operator $\ast$-algebras, multiplier operator $\ast$-algebras, dual operator $\ast$-algebras (by which we mean an operator $*$-algebra which is a dual operator space with weak* continuous involution), and involutive $M$-ideals. Section 5 has a common theme of hereditary subalgebras and ideals, noncommutative topology (e.g.\ open projections, support projections, and compact and peak projections), and peak interpolation, in the involutive setting. Thus we are finding the involutive variants of the operator algebra theory of these topics from e.g.\ the papers \cite{BRI, BRII, BRord, BNII, BHN, Bnpi}. \subsection{Involutions, and notation} \label{inv} By an involution we mean at least a bijection $\tau : A \to A$ which is of period 2: $\tau^2(a) = a$ for $a \in A$. A $C^*$-algebra $B$ may have two kinds of extra involution: a period 2 conjugate linear $*$-antiautomorphism or a period 2 linear $*$-antiautomorphism. The former is just the usual involution $*$ composed with a period 2 $*$-automorphism of $B$. The latter is essentially the same as a `real structure', that is if $\theta$ is the antiautomorphism then $B$ is just the complexification of a real $C^*$-algebra $D = \{ x \in B : x = \bar{x} \}$, where $\bar{x} = \theta(x)^*$. We may characterize $x \mapsto \bar{x}$ on $B$ very simply as the map $a + i b \mapsto a - ib$ for $a, b \in D$. By way of contrast, there are four distinct natural kinds of `completely isometric involution' on a general operator algebra $A$. Namely, period 2 bijections which are \begin{itemize} \item [(1)] conjugate linear antiautomorphisms $\dagger : A \to A$ satisfying $\| [a_{ji}^\dagger ] \| = \| [ a_{ij}] \| ,$ \item [(2)] linear antiautomorphisms $\theta : A \to A$ satisfying $\| [a_{ji}^\theta ] \| = \| [ a_{ij}] \| ,$ \item [(3)] conjugate linear automorphisms $\textendash : A \to A$ satisfying $\| [\overline{a_{ij}} ] \| = \| [ a_{ij}] \| ,$ \item [(4)] linear automorphisms $\pi : A \to A$ satisfying $\| [a_{ij}^\pi ] \| = \| [ a_{ij}] \| .$ \end{itemize} Here $[a_{ij}]$ is a generic element in $M_n(A)$, the $n \times n$ matrices with entries in $A$, for all $n \in \Ndb$. Class (1) is just the operator $*$-algebras mentioned earlier. In this paper we will call the algebras in class (2) {\em operator algebras with linear involution} $\theta$, and write $\theta(a)$ as $a^\theta$. We will not discuss (4) in this paper, these are well studied and are only mentioned here because most of the results in the present paper apply to all four classes. We will just say that this class is in bijective correspondence with the unital {\em completely symmetric projections} on $A$ in the sense of \cite{BNpac}, this correspondence is essentially Corollary 4.2 there. Similarly, for the same reasons we will not discuss class (3) in this paper, except in the final theorem. By \cite[Theorem 3.3]{Sharma}, class (3) is essentially the same as `real operator algebra structure', that is $A$ is just the complexification of a real operator algebra $D = \{ x \in B : x = \bar{x} \}$, and we may rewrite $\bar{x} = a - i b$ if $x = a + ib$ for $a, b \in D$. Thus the variant of the main aspects of our paper in case (3) seem best treated within the theory of real operator algebras. However it is worth saying that the theory in our paper in case (3) may be viewed as a transliteration of a chapter in the theory of real operator algebras. We also remark that if $A$ is unital or approximately unital then one can easily show using the Banach-Stone theorem for operator algebras (see e.g.\ \cite[Theorem 4.5.13]{BLM}), and the `opposite and adjoint algebras' discussed at the start of Section 2, that the matrix norm equality in (3) and (4) (resp.\ (1) and (2)) plus a `unital condition' force the involution to be multiplicative (resp.\ anti-multiplicative). If $A$ is a $C^*$-algebra then classes (1) and (4) are essentially the same after applying the $C^*$-algebra involution $*$. (Note that in this case the matrix norm equality in (1) or (4) follows from the same equality for $1 \times 1$ matrices, that is that the involution is isometric. Indeed it is well known that $*$-isomorphisms of $C^*$-algebras are completely isometric.) Similarly classes (2) and (3) essentially coincide if $A$ is a $C^*$-algebra. We will mostly focus on class (1) for specificity. In fact most of the results in the present paper apply to all four classes, however it would be too tedious to state several cases of each result. Instead we leave it to the reader to state the matching results in cases (2)--(4). For example to get from case (1) to case (2) of results below one replaces $a^\dagger$ by $a^\theta$, and $\dagger$-{\em selfadjoint elements}, that is elements satisfying $a^\dagger = a$, by elements with $a^\theta = a$. We remark that if $A$ is an operator algebra with linear involution $\theta$, then $\{ a \in A : a = a^\theta \}$ is a Jordan operator algebra in the sense of \cite{BWj}. (We remark that these `$\theta$-{\em selfadjoint} elements' need not generate $A$ as an operator algebra, unlike for involutions of type (1).) Most of our discussion of class (2) involves finding interesting examples of such involutions. Indeed although classes (1)--(4) have similar theory from the viewpoint of our paper, the {\em examples} of algebras in these classes are quite different in general. Because of the ubiquity of the asterisk symbol in our area of study, we usually write the involution on an operator $*$-algebra as $\dagger$, and refer to, for example, $\dagger$-selfadjoint elements or subalgebras, and $\dagger$-homomorphisms (the natural morphisms for $*$-algebras). A little more background and notation: A {\em unital} operator algebra has an identity of norm $1$, and an {\em approximately unital} operator algebra has a contractive approximate identity (cai). For background on operator spaces and operator algebras from an operator space point of view we refer the reader to \cite{BLM,Pisbk}. Meyer's theorem states that any operator algebra $A$ has a unitization $A^1$ that is unique up to completely isometric isomorphism \cite[Corollary 2.1.15]{BLM}. If $A$ is nonunital then $A$ is of codimension 1 in the unital operator algebra $A^1$; otherwise set $A^1 = A$. In this paper, all projections $p \in A$ are orthogonal projections. If $X$ and $Y$ are sets then we write $XY$ for the {\em norm closure} of the span of terms of the form $xy,$ for $x\in X, y\in Y.$ The second dual $A^{\ast\ast}$ of an operator algebra $A$ is again an operator algebra, which is unital if $A$ is approximately unital. We recall that a $C^{\ast}$-cover $(B, j)$ of an operator algebra $A$ is a $C^*$-algebra $B$ and a completely isometric homomorphism $j : A \to B$ such that $j(A)$ generates $B$ as a $C^*$-algebra. Sometimes we simply call this a $C^*$-algebra generated by $A$. There is a `biggest' and `smallest' $C^{\ast}$-cover, $C^*_{\rm max}(A)$ and $C^*_e(A)$ (see \cite[Propositions 4.3.5 and 2.4.2]{BLM}). For example $C^*_{\rm max}(A)$ has the universal property that any completely contractive representation $\pi : A \to B(H)$ extends to a $*$-representation of $C^*_{\rm max}(A)$ on $H$. Any completely isometric homomorphism $j : A \to B$ into a $C^*$-cover $B$ of $A$ generated by the copy of $A$, gives rise to a $*$-homomorphism $B \to C^*_e(A)$ which is `the identity' on the copy of $A$. Because of the uniqueness of unitization, for an operator algebra $A$ we can define unambiguously ${\mathfrak F}_A = \{ a \in A : \Vert 1 - a \Vert \leq 1 \}$. Then $\frac{1}{2} {\mathfrak F}_A = \{ a \in A : \Vert 1 - 2 a \Vert \leq 1 \} \subset {\rm Ball}(A)$. Here and throughout ${\rm Ball}(X)$ denotes $\{ x \in X : \| x \| \leq 1 \}$. Similarly, ${\mathfrak r}_A$, the {\em real positive} or {\em accretive} elements in $A$, is $\{ a \in A : a + a^* \geq 0 \}$, where the adjoint $a^*$ is taken in any $C^*$-cover of $A$. We write oa$(x)$ for the operator algebra generated by an operator $x$. By a {\em symmetry} we mean either a selfadjoint unitary operator, or a period 2 $*$-automorphism of a $C^*$-algebra, depending on the context. Recall that a projection in $A^{\ast\ast}$ is {\em open} in $A^{**},$ or $A$-open for short, if $p\in(pA^{\ast\ast}p\, \cap A)^{\perp\perp}$. That is, if and only if there is a net $(x_t)$ in $A$ with $$x_t=px_t=x_tp=px_tp\to p,\, {\rm weak^*}.$$ This is a generalization of Akemann's notion of open projections for $C^*$-algebras. If $p$ is open in $A^{**}$ then clearly $$D=pA^{\ast\ast}p\cap A=\{a\in A: a=ap=pa=pap\}$$ is a closed subalgebra of $A,$ and the subalgebra $D^{\perp\perp}$ of $A^{\ast\ast}$ has identity $p.$ By \cite[Proposition 2.5.8]{BLM} $D$ has a cai. If $A$ is also approximately unital then a projection $p$ in $A^{\ast\ast}$ is closed if $p^{\perp}$ is open. We call such a subalgebra $D$ a {\em hereditary subalgebra} of $A$ (or HSA) and we say that $p$ is the {\em support projection} of the HSA $pA^{\ast\ast}p\cap A.$. It follows from the above that the support projection of a HSA is the weak* limit of any cai from the HSA. The above correspondence between hereditary subalgebras and open projections is bijective and order preserving. \section{Involutive operator algebras} We recall for any operator space $X$ the {\em opposite} and {\em adjoint} operator spaces $X^\circ$ and $X^\star$ from e.g.\ 1.2.25 in \cite{BLM}. Here $X^\circ$ is $X$ but with `transposed matrix norms' $|||[a_{ij} ]||| = \| [a_{ji} ] \|$. Similarly $X^\star$ is the set of formal symbols $x^\star$ for $x \in X$, but with the same operator space structure as $\{ x^* \in B : x \in X \}$, if $X$ is (completely isometrically) a subspace of a $C^*$-algebra $B$. If $A$ is an operator algebra we write $A^1$ for the unitization of $A$. If $X$ is an operator space then $I(X)$ and ${\mathcal T}(X)$ are respectively the {\em injective and ternary envelopes} of $X$ from e.g.\ \cite[Chapter 4]{BLM} (the word `triple' is used in place of `ternary' in \cite{BLM}). We will not apply the injective and ternary envelope in this paper, but include them in the following result since they will be important in future work on involutive operator algebras and spaces. We recall from e.g.\ the last reference that $I(X)$ is an injective operator space containing $X$, and it is the unique such with the `rigidity' (or with the `essential') property of 4.2.3 in \cite{BLM}. Then ${\mathcal T}(X)$, sometimes called the {\em noncommutative Shilov boundary} of $X$, is the smallest closed subspace of $I(X)$ containing $X$ which is closed under the ternary product $x y^* z$. It may also be characterized by the universal property in \cite[Theorem 8.3.9]{BLM}; and if $X$ is unital or is an approximately unital operator algebra then ${\mathcal T}(X)$ is the $C^*$-envelope $C^*_e(X)$ mentioned earlier. \begin{proposition} \label{ciop} If $X$ is an operator space and $A$ is an operator algebra then \begin{itemize} \item [(1)] $(A^1)^\circ = (A^\circ)^1$, $I(X^\circ) = I(X)^\circ, {\mathcal T}(X^\circ) = {\mathcal T}(X)^\circ, C^*_e(A^\circ) = C^*_e(A)^\circ,$ and $C^*_{\rm max}(A^\circ) = C^*_{\rm max}(A)^\circ$. \item [(2)] $(A^1)^\star= (A^\star)^1$, $I(X^\star) = I(X)^\star, {\mathcal T}(X^\star) = {\mathcal T}(X)^\star, C^*_e(A^\star) = C^*_e(A)^\star,$ and $C^*_{\rm max}(A^\star) = C^*_{\rm max}(A)^\star$. \end{itemize} \end{proposition} \begin{proof} Note that $(A^\circ)^1$ is a unital operator algebra containing $A^\circ$ as a codimension 1 subalgebra, so by Meyer's theorem \cite[Corollary 2.1.15]{BLM} it must be the unitization. Similarly for $(A^1)^{\star}$. The rest all follow by the universal properties defining these objects, and a diagram chase applying $\circ$ or $\star$ to the maps in the diagrams. For example, such a strategy shows that $I(X)^\circ$ is injective. It contains $X^\circ$, and a similar strategy shows that it has the `rigidity' property (or `essential' property) characterizing the injective envelope. \end{proof} {\bf Remark.} There is a similar result for $\bar{X}$ and $\bar{A}$, which would be useful in treating class (3) from the start of Section {\rm \ref{inv}}. Here $\bar{X} = (X^{\star})^\circ$, and from this formula the proof of the result in this case is clear. \medskip The following result, the involutive variant of Meyer's theorem \cite[Corollary 2.1.15]{BLM}, is useful in treating involutions on operator algebras with no identity or approximate identity. \begin{lemma} \label{mey} Let $A$ be a nonunital operator algebra with an involution of one of the types {\rm (1)--(4)} at the start of Section {\rm \ref{inv}}. Then the involution on $A$ has a unique extension to an involution of the same type on the unitization $A^1$, with the involution of $1$ being $1$. \end{lemma} \begin{proof} For operator $*$-algebras this is \cite[Lemma 1.15]{BKM}. If $\theta : A \to A$ is a linear involution (type (2) in the list at the start of Section \ref{inv}), then by Meyer's theorem $a \mapsto \theta(a)^\circ$ extends to a unital completely isometric homomorphism $A^1 \to (A^\circ)^1$. Composing this with $\circ$, and using the fact that $(A^1)^\circ = (A^\circ)^1$ from Proposition \ref{ciop}, we obtain our result. The other cases are similar or easier. \end{proof} \begin{proposition}\label{dmeyer1} Let $A$ and $B$ be operator algebras with $A$ nonunital. Also suppose that there exists involutions on $A$ and $B$ of one of the types {\rm (1)--(4)} at the start of Section {\rm \ref{inv}}. Let $\pi: A\to B$ be a completely contractive (resp.\ completely isometric) involution preserving homomorphism. Then there is a unital completely contractive (resp.\ completely isometric) involution preserving homomorphism extending $\pi:$ from $A^1$ to $B^1$ (for the completely isometric case we also need $B$ nonunital). \end{proposition} \begin{proof} By Lemma \ref{mey}, we know that both $A^1$ and $B^1$ are operator algebras with the same type of involution. The unital extension of $\pi$ to $A^1$ is completely contractive (resp.\ completely isometric) by \cite[Theorem 2.1.13 and Corollary 2.1.15]{BLM}. It is easy to check that it is also involution preserving. \end{proof} {\bf Remark.} One may replace completely contractive (resp.\ completely isometric) by contractive (resp.\ isometric) in the last result. Thus the unitization $A^1$ of an involutive operator algebra is unique up to (completely) isometric involutive isomorphism. \medskip We now characterize class (2) from the start of Section {\rm \ref{inv}}. A {\em conjugation} on a complex Hilbert space $H$ is a conjugate linear period $2$ isometry $u : H \to H$. If $j : H \to \bar{H}$ is the canonical conjugate linear map into the conjugate Hilbert space then $ju$ is unitary, so $\langle uy , ux \rangle = \langle ju x, ju y \rangle_{\bar{H}} = \langle x, y \rangle$ for $x, y \in H$. An operator $T$ on $H$ is called $c$-{\em symmetric} if $c T c = T^*$, and is called {\em complex symmetric} if it is $c$-symmetric for a conjugation $c$ on $H$. The class of complex symmetric operators is very large and significant (see e.g.\ \cite{GP}). \begin{theorem} \label{lininvch} Let $A$ be an operator algebra. The following are equivalent: \begin{itemize} \item [(i)] $A$ is an operator algebra with linear involution (that is, type {\rm (2)} in the list at the start of Section {\rm \ref{inv}}) $\theta$. \item [(ii)] There exists a $C^*$-algebra $B$ generated by $A$ (or by $A^1$), and a period 2 $*$-anti-isomorphism $\rho : B \to B$ with $\rho(A) = A$. \item [(iii)] There exists a conjugation $c$ on a complex Hilbert space $H$ on which $A$ may be completely isometrically represented as an operator algebra such that $c A^* c \subset A$ (here we are identifying $A$ with its image in $B(H)$). \end{itemize} We may take $\theta$ in {\rm (i)} to be the restriction to $A$ of the $\rho$ in {\rm (ii)}, or of the map $T \mapsto c T^* c$ in {\rm (iii)}. We may take $B$ in {\rm (ii)} to be $C^*_e(A)$ (or $C^*_e(A^1)$), or $C^*_{\rm max}(A)$ (or $C^*_{\rm max}(A^1)$). \end{theorem} \begin{proof} (i) $\Rightarrow$ (ii) \ We may assume that $A$ is unital by Proposition \ref{dmeyer1}. Given (i), the map $A \to A^\circ : a \mapsto \theta(a)^\circ$ is a completely isometric isomorphism, so extends to a $*$-isomorphism $B = C^*_e(A) \to C^*_e(A^\circ)$. The latter algebra equals $C^*_e(A)^\circ$ by Proposition \ref{ciop}. This gives a $*$-anti-isomorphism on $B$ taking $A$ onto $A$, which is easily checked to be period 2. Similarly with $B = C^*_{\rm max}(A)$ using the universal property of these $C^*$-algebras and the appropriate item in Proposition \ref{ciop}. (ii) $\Rightarrow$ (iii) \ The map $\pi(b) = \rho(b)^*$ is a period 2 conjugate linear $*$-isomorphism on $B$. Then $D = \{ b \in B : b = \rho(b)^* \}$ is a real $C^*$-algebra, which we can represent on a real Hilbert space $K$. Let $H = K_c$ and $c : H \to H$ be the canonical conjugation. Then $B$ may be viewed as the $C^*$-algebra $D + i D$ acting on $H$ by $(a + ib)(\xi + i \eta) = a \xi - b \eta + i(a \eta + b \xi)$. And $\pi(a+ib) = a - ib$ by definition of $D$. Then $$c (a + ib) c (\xi + i \eta) = c( a \xi + b \eta + i(- a \eta + b \xi)) = a \xi + b \eta + i (a \eta - b \xi) = \pi(a+ib) (\xi + i \eta).$$ So $\pi(b) = c b c$, and $\rho(b) = c b^* c$. (iii) $\Rightarrow$ (i) \ The map $\theta(a) = c a^* c$ is a linear period 2 anti-isomorphism from $A$ to $A$, and $$\| [ a_{ij} ] \| = \| [ a_{ji}^* ] \| = \| [ c a_{ji}^* c ] \|= \| [ a_{ji}^\theta ] \| $$ for $[a_{ij}] \in M_n(A)$. \end{proof} A {\em symmetric operator algebra} is an operator algebra $A$ with $A = A^\circ$ completely isometrically via the identity map. These were introduced in \cite{BComm} where it was observed that such algebras were commutative, etc. They are characterized by the following result, which says that such algebras must be operator algebras of matrices that equal their transpose (i.e.\ which are symmetric with respect to the transpose as matrices). By an operator algebra of matrices we mean a subalgebra of $\Mdb_I$ for a cardinal $I$, where $\Mdb_I = B(\ell^2_I)$ thought of as $I \times I$ matrices. Any operator algebra $A$ of symmetric matrices is commutative, since $ab = (ab)^T = ba$ for $a, b \in A$. Hence $A$ is the algebra generated by a set of commuting symmetric matrices in $\Mdb_I$. \begin{corollary} \label{chsym} An operator algebra $A$ is symmetric if and only if there exists a conjugation $c$ on a complex Hilbert space $H$ on which $A$ may be completely isometrically represented as an operator algebra, such that $c a c = a^*$ for all $a \in A$ (here we are identifying $A$ with its image in $B(H)$). That is, if and only if there exists a completely isometric representation of $A$ as $c$-symmetric operators for a conjugation $c$. Equivalently, $A$ is symmetric if and only if there exists an orthonormal basis ${\mathfrak B}$ for a complex Hilbert space $H$ on which $A$ has been completely isometrically represented as an operator algebra, such that the matrix with respect to ${\mathfrak B}$ of every element in $A$ equals its transpose. \end{corollary} \begin{proof} Set $\theta$ equal the identity map and apply the last theorem and its proof to see that $A$ is symmetric if and only if the desired conjugation $c$ on $H$ exists with $c a^* c = a$ for $a \in A$. From this the first two if and only if's follow. The last statement follows from the fact alluded to earlier that the $c$-symmetric operators are precisely the operators whose matrix with respect to some appropriate orthonormal basis ${\mathfrak B}$ is symmetric. This follows from the well known trick (certainly well known in the theory of complex symmetric operators) that if $K = \{ \xi = c \xi \in H \}$, then $K^\perp = (0)$, so a (real) orthonormal basis ${\mathfrak B}$ for $K$ is a (complex) basis for $H$. Also $c T c = T^*$ implies the matrix with respect to ${\mathfrak B}$ is symmetric. \end{proof} \begin{corollary} \label{chsym2} An operator algebra $A$ is commutative if and only if it is isometrically isomorphic to an operator algebra of matrices that equal their transpose. \end{corollary} \begin{proof} If $A$ is commutative then it is isometrically isomorphic to $\{ (a, a^\circ) \in A \oplus A^{\circ} : a \in A \}$, which is a symmetric operator algebra \cite{BComm}. The rest follows from Corollary \ref{chsym} (and the observation above that an algebra of symmetric matrices is commutative). \end{proof} \begin{corollary} \label{chsym3} The algebra generated by any operator on a Hilbert space is isometrically isomorphic to the algebra generated by a complex symmetric operator on another Hilbert space. \end{corollary} \begin{proof} As in the proof of Corollary \ref{chsym2}, this algebra is isometrically isomorphic to a symmetric operator algebra, and by Corollary \ref{chsym} its generator may be taken to be $c$-symmetric for a conjugation $c$. \end{proof} There are similar characterizations for the other three classes of `involutions' considered at the start of Section {\rm \ref{inv}}. Indeed the result matching Theorem \ref{lininvch} for operator $*$-algebras is the following, mostly from \cite[Section 1]{BKM}. \begin{theorem} \label{opinvch} Let $A$ be an operator algebra. The following are equivalent: \begin{itemize} \item [(i)] $A$ is an operator $*$-algebra. \item [(ii)] There exists a $C^*$-algebra $B$ generated by $A$ (or of $A^1$), and a period 2 $*$-automorphism $\rho : B \to B$ with $\rho(A) = A^*$. \item [(iii)] There exists a symmetry $u$ on a complex Hilbert space $H$ on which $A$ may be completely isometrically represented as an operator algebra such that $u A^* u \subset A$ (here we are identifying $A$ with its image in $B(H)$). \end{itemize} We may take $\theta$ in {\rm (i)} to be the restriction to $A$ of $\rho(\cdot)^*$ for $\rho$ as in {\rm (ii)}, or to be the map $T \mapsto u T^* u$ in {\rm (iii)}. We may take $B$ in {\rm (ii)} to be $C^*_e(A)$ (or $C^*_e(A^1)$), or $C^*_{\rm max}(A)$ (or $C^*_{\rm max}(A^1)$). \end{theorem} \begin{proof} This is proved in \cite[Section 1]{BKM}, except for the assertion about $C^*_{\rm max}(A)$. If $\rho : A \to C^*_{\rm max}(A)$ is the canonical `inclusion' let $\pi : A\to C^{\ast}_{\max}(A)$ be the completely isometric homomorphism defined by $\pi(a)= \rho(a^{\dagger})^{\ast}.$ By the universal property of $C^{\ast}_{\max}(A),$ there exists a unique $\ast$-homomorphism $\sigma: C^{\ast}_{\max}(A)\to C^{\ast}_{\max}(A)$ such that $\sigma(\rho(a))=\pi(a)=\rho(a^{\dagger})^{\ast}$ for any $a\in A.$ Moreover, $\sigma$ has order 2 since $$\sigma^2(\rho(a))=\sigma(\rho(a^{\dagger})^{\ast})=\sigma(\rho(a^{\dagger}))^{\ast}=\rho(a),$$ and since $\rho(A)$ generates $C^{\ast}_{\max}(A)$ as a $C^{\ast}$-algebra. The final assertion follows by extending to the unitization and using $C^{\ast}_{\max}(A^1) = C^{\ast}_{\max}(A)^1$ from Proposition \ref{ciop}. \end{proof} It is natural to ask if in item (ii) in Theorems \ref{lininvch} or \ref{opinvch} (ii), or in matching results for the other types of involutions, one may use {\em any} $C^*$-algebra generated by a completely isometric copy of $A$. The answer is in the negative, as one sees in the following result and the example following it. \begin{definition} \label{ivltcmptb} Suppose that an operator algebra $A$ has an involution $\nu$ of one of the types {\rm (1)--(4)} at the start of Section \ref{inv}. If a $C^{\ast}$-cover $(B, j)$ of $A$ has an involution $\omega$ of the same type, and if $j(a)^{\omega}=j(a^{\nu}),$ for any $a\in A,$ we say that the involution on $B$ is {\em compatible with} $A.$ \end{definition} \begin{lemma} Suppose that $A$ is an operator algebra (possibly not approximately unital) with involution $\nu$ of type {\rm (1)} (resp.\ type {\rm (2)}) at the start of Section {\rm \ref{inv}}, and $(B,j)$ is a $C^{\ast}$-cover of $A$. Then $B$ has an involution compatible with $A$ if and only if there exists an order $2$ $\ast$-automorphism (resp.\ $\ast$-antiautomorphism) $\sigma: B\to B$ such that $\sigma(j(a^{\nu}))=j(a)^{*}$ (resp.\ $\sigma(j(a^{\nu}))=j(a)$) for any $a\in A$. \end{lemma} \begin{proof} ($\Rightarrow$) If $B$ has an involution $\omega$ compatible with $A,$ then $j(a)^{\omega}=j(a^{\nu}),$ for all $a\in A.$ Define $\sigma: B\to B$ by $\sigma(b)=(b^{\ast})^{\omega}$ (resp.\ $b^\omega$) for any $b\in B.$ Then it is easy to see that $\sigma$ is an order $2$ $\ast$-automorphism (resp.\ $\ast$-antiautomorphism). ($\Leftarrow$) The involution on $B$ is defined by $b^{\omega}=\sigma(b)^{\ast}$ (resp.\ $b^{\omega}= \sigma(b)$) for any $b\in B.$ Then $B$ is a $C^{\ast}$-algebra with involution which is compatible with $A.$ \end{proof} \begin{example} Let $A(\Ddb)$ be the disk algebra. By the above $C^*_{\rm max}(A(\Ddb))$ and $C(\Tdb) = C^*_e(A(\Ddb))$ are $C^*$-covers of $A(\Ddb)$ which are compatible with the involution $\overline{f(\bar{z})}$ on $A(\Ddb)$. So too is $C(\overline{\Ddb})$ clearly, with the same involution $\overline{f(\bar{z})}$. The Toeplitz $C^*$-algebra is a well known $C^*$-cover of the disk algebra $A(\Ddb)$, however we show that it is not compatible with the involution $\overline{f(\bar{z})}$ on $A(\Ddb)$. Let $S$ be the unilateral shift on $l^2(\Ndb_0)$ and $\oa{S}$ be the operator algebra generated by $S.$ Then $\oa{S}$ is an operator $\ast$-algebra with trivial involution induced by $S^{\dagger}=S.$ Suppose that the Toeplitz $C^*$-algebra $C^{\ast}(S)$ has an involution compatible with $\oa{S}.$ Then there exists an order-2 $\ast$-isomorphism $C^{\ast}(S)$ such that $\sigma(S^{\dagger})=S^{\ast}.$ Moreover, we have $$I=\sigma(I)=\sigma(S^{\ast}S)=\sigma(S)^{\ast}\sigma(S)=SS^{\ast}\neq I,$$ which is a contradiction. It is similarly not hard to find (using \ref{Shb} below) commutative $C^*$-algebras $C(K)$ generated by a function algebra $A$ with linear involution $\theta$ (such as $A = A(\Ddb)$), such that $\theta$ does not extend to a linear involution on $C(K)$. \end{example} \begin{lemma} \label{bidu} Let $A$ be an operator algebra with an involution of one of the types {\rm (1)--(4)} at the start of Section {\rm \ref{inv}}. Then the involution on $A$ has a unique extension to a weak* continuous involution of the same type on the bidual $A^{**}$. \end{lemma} \begin{proof} We will just prove this in the case of a linear involution $\theta$; the others are similar. The associated completely isometric homomorphism $A \to A^\circ : a \mapsto \theta(a)^\circ$ extends to a weak* continuous completely isometric homomorphism $A^{**} \to (A^\circ)^{**}$. However it is an easy exercise to see that $(A^\circ)^{**} \cong (A^{**})^\circ$. Composing with $\circ$ we obtain a weak* continuous linear involution on $A^{**}$ extending $\theta$. \end{proof} We mention that the Cayley transform $\kappa$ and the ${\mathfrak F}$ transform of \cite[Section 2.2]{BRord}, important tools in the area of the later sections of our paper, do work well with respect to involutions. For example suppose that $\dagger$ is the involution on an operator $*$-algebra, and $\sigma$ the associated $*$-automorphism on a (compatible) $C^*$-cover. If $x$ is real positive then $\sigma(x^*) = x^\dagger$ is real positive, and $$\sigma \kappa(x^\dagger) = \sigma((x^\dagger - 1)( x^\dagger + 1)^{-1}) = (x^* - 1)( x^* + 1)^{-1} = \kappa(x)^*.$$ So $\kappa(x^\dagger) = \kappa(x)^\dagger$. Similarly, if $x$ is a contraction with $1-x$ invertible then the same is true for $x^\dagger$ and the inverse Cayley transform $\kappa^{-1}(x^\dagger) = (1+x^\dagger)(1- x^\dagger)^{-1}$ is real positive, and must equal $\kappa^{-1}(x)^\dagger$. The ${\mathfrak F}$-transform is ${\mathfrak F}(x) = \frac{1}{2}(1 + \kappa(x)) = x(1+x)^{-1}$. Following the proof in Lemma 2.5 in \cite{BRord}, it is easy to see that for any operator $\ast$-algebra $A,$ the ${\mathfrak F}$-transform maps the $\dagger$-selfadjoint elements in ${\mathfrak r}_A$ bijectively onto the set of $\dagger$-selfadjoint elements in $\frac{1}{2}{\mathfrak F}_A$ of norm $<1.$ \section{Examples} \label{Exs} We give many examples of operator $*$-algebras and operator algebras with linear involution here. Of course any real operator algebra at all gives an example of the third type of involution mentioned at the start of Section \ref{inv}, namely the complexification. We will not consider these here. \subsection{Examples from noncommutative differential geometry} Several examples of operator $*$-algebras were given in \cite{BKM}, most of them examples from noncommutative differential geometry (historically the first such example being due to Mesland). Other examples from noncommutative differential geometry may be found in other recent papers of Kaad, Mesland, and their coauthors. \subsection{Function algebra examples} \label{Shb} Let $A$ be a uniform algebra (with minimal operator space structure, see 1.2.21 in \cite{BLM}). Then $A\subset C^{\ast}_e(A)=C(\partial A),$ where $\partial A$ is the Shilov boundary of $A$ (see e.g.\ \cite[Section 4.1]{BLM}). If $A$ is an operator $*$-algebra (resp.\ has linear involution $\theta$), then there exists a period $2$ homeomorphism $\tau : \partial A\to \partial A$ such that $f^{\dagger}(\omega)=\overline{f(\tau(\omega))}$ (resp.\ $f^\theta = f \circ \tau$) for any $f \in A$. From this formula it is easy to write down function algebra examples. For example, the disk algebra $A(\Ddb)$ is an operator $*$-algebra with $f^\dagger(z) = \overline{f(\bar{z})}$, and so are its closed $\dagger$-ideals of functions e.g.\ vanishing at 0, or at 1. The latter ideal is interesting from the perspective of approximate identities: it is nonunital, is a $\dagger$-ideal, and has a real positive $\dagger$-selfadjoint cai (see Lemma \ref{dcai} below), etc). Similarly $H^\infty(\Ddb)$ is a dual operator $*$-algebra with the same involution. This involution is weak* continuous, and extends to an involution on the von Neumann algebra $L^\infty(\Tdb)$. These two algebras also have linear involution $f^\theta(z) = f(-z)$. This and the identity map are the only linear involutions on $A(\Ddb)$ and $H^\infty$, by the well known theory of automorphisms of these algebras. We recall that a {\em Q-algebra} is an operator algebra quotient of a function algebra (with minimal operator space structure) by a closed ideal. Q-algebras are symmetric operator algebras, and in particular have a linear involution. If the function algebra has an involution making it an operator $*$-algebra, and the ideal is involutive, then we call the quotient an {\em involutive Q-algebra}. We will see later that for example the algebra generated by the Volterra operator is an involutive Q-algebra. \subsection{Examples from complex symmetric and $*$-exchangeable operators} An operator $T$ in a $C^*$-algebra $B$ will be called $*$-{\em exchangeable} if $\| p(T, T^*) \| = \| p(T^*, T) \|$ for any polynomial $p$ in two noncommuting variables. One may use polynomials without constant term here if one wishes. Indeed if the equality holds for such polynomials $p$ then it follows that the map $p(T, T^*) \mapsto p(T^*, T)$ is well defined on a dense subset of $C^*(T)$, hence extends to a $*$-homomorphism $\sigma$ on $C^*(T)$ taking $T$ to $T^*$, and extends further to $C^*(1,T)$. This shows that the norm equality holds for polynomials with constant terms too. It is easy to see that $\sigma$ is a period 2 $*$-automorphism. For a polynomial $p$ of one variable we write $p^\dagger$ for the same polynomial but with coefficients replaced by their complex conjugate (that is, $p^\dagger(z) = \overline{p(\bar{z})}$). \begin{theorem} \label{oneg} Let $A$ be an operator algebra with a single generator $T$. The following are equivalent: \begin{itemize} \item [(i)] For $n \in \Ndb$ and polynomials $p_{ij}$ for $1 \leq i, j \leq n$ we have $$\| [ p_{ji}^\dagger (T) ] \| = \| [ p_{ij}(T) ] \| .$$ \item [(ii)] $T$ is $*$-exchangeable in some $C^*$-algebra generated by (a completely isometrically homomorphic copy of) $A$. \item [(iii)] $A$ is an operator $*$-algebra with $T$ $\dagger$-selfadjoint. \item [(iv)] There exists a symmetry $u$ on a Hilbert space $H$ on which $A$ may be completely isometrically represented as an operator algebra such that $T^* = u T u$ (here we are identifying $T$ with its image in $B(H)$). \end{itemize} In {\em (i)} one may if one wishes use only polynomials with no constant term. \end{theorem} \begin{proof} (iv) $\Rightarrow$ (i) \ Given linear symmetry $u : H \to H$ with $T^* = uTu$ then $p(T)^* = p^\dagger (T^*) = u p^\dagger (T) u$, so that $$\| [ p_{ji}^\dagger (T) ] \| = \| [ u \, p_{ji}(T)^* \, u] \| = \| [ p_{ij}(T) ] \| .$$ (i) $\Rightarrow$ (iii) \ If for polynomials $p_{ij}$ with no constant term we have $\| [ p_{ij}(T) ] \| = \| [ p_{ji}^\dagger (T) ] \| = \| [ p_{ij} (T^*) ] \|$, then the map $p(T) \mapsto p(T^*)$ is well defined and completely isometric. Here $*$ is the involution on a $C^*$-algebra containing $A$. It extends to a completely isometric surjective homomorphism oa$(T) \to {\rm oa}(T^*)$. Composing this with the involution $*$ we obtain an involution on oa$(T)$ making it an operator $*$-algebra with $T$ $\dagger$-selfadjoint. (If one wishes then we can extend the involution to the unitization by Proposition \ref{dmeyer1}, which implies the equality in (i) for polynomials with constant term.) (iii) $\Rightarrow$ (ii) \ If oa$(T)$ has such involution then by the characterization of operator $*$-algebras in Theorem \ref{opinvch} there exists a $*$-isomorphism $C^*_e({\rm oa}(T)) \to C^*_e({\rm oa}(T))$ taking $T^\dagger = T$ to $T^*$. Equivalently (as in the discussion above the theorem), $p(T, T^*) \mapsto p(T^*, T)$ is a well-defined isometry. (ii) $\Rightarrow$ (iii) \ If $T$ is $*$-exchangeable in some $C^*$-algebra $B$ generated by $A$, then as explained above the theorem we have a period 2 $*$-automorphism $\sigma : B \to B$ with $\sigma(T) = T^*$. The restriction of $\sigma$ to $A$ maps onto $A^*$. So $A$ is an operator $*$-algebra with $T$ $\dagger$-selfadjoint if we define $a^\dagger = \sigma(a)^*$. (iii) $\Rightarrow$ (iv) \ By the characterization of operator $*$-algebras in Theorem \ref{opinvch}, there exists a symmetry $u$ on a Hilbert space $H$ on which $A$ may be completely isometrically represented as an operator algebra such that $a^\dagger = u a^* u$ for all $a \in A$. Setting $a = T$ we obtain $T^* = u T u$. \end{proof} There is a similar result for operator algebras with linear involution $\theta$ with $\theta(T) = T$. The analogue of condition (ii) in Theorem \ref{oneg} is the condition called $g$-normality in \cite{GuoJiZhu}, namely that $\| p(T, T^*) \| = \| p^\dagger (T^*, T) \|$ for any polynomial $p$ in two free variables. Here $p^\dagger$ is obtained from $p$ by conjugating each coefficient. The equivalence of (ii) and (iv) is known: after our paper was written we found this equivalence in \cite{SZ} with a quite different proof. The paper \cite{ZZ} also contains some other very interesting related results. \begin{theorem} \label{cosym} Let $A$ be an operator algebra with a single generator $T$. The following are equivalent: \begin{itemize} \item [(i)] For $n \in \Ndb$ and polynomials $p_{ij}$ for $1 \leq i, j \leq n$ we have $$\| [ p_{ji} (T) ] \| = \| [ p_{ij}(T) ] \| .$$ \item [(ii)] $T$ is $g$-normal in some $C^*$-algebra generated by $A$. \item [(iii)] $A$ is a symmetric operator algebra (that is $I_A$ is a linear involution). \item [(iv)] There exists a Hilbert space $H$ on which $A$ may be completely isometrically represented as an operator algebra such that $T$ becomes a complex symmetry on $H$ (in the sense defined above {\rm Theorem \ref{lininvch}}). \end{itemize} In {\rm (i)} one may if one wishes use only polynomials with no constant term. \end{theorem} \begin{proof} (iv) $\Rightarrow$ (i) \ Given conjugation $c : H \to H$ with $T^* = cTc$ then $p(T)^* = p^\dagger (T^*) = c p (T) c$, so that $$\| [ p_{ij} (T) ] \| = \| [ p_{ji}(T)^* ] \| = \| [ c \, p_{ji}(T) \, c] \| = \| [ p_{ji}(T) ] \| .$$ (i) $\Leftrightarrow$ (iii) \ Obvious. (iii) $\Rightarrow$ (ii) \ If oa$(T)$ has such involution then by Theorem \ref{lininvch} there exists a $*$-antiautomorphism $C^*_e({\rm oa}(I,T)) \to C^*_e({\rm oa}(I,T))$ taking $T$ to $T$. Composing with $*$, we get a conjugate linear $*$-automorphism of $C^*_e({\rm oa}(I,T))$ taking $T$ to $T^*$. Equivalently, $p(T, T^*) \mapsto p^\dagger(T^*, T)$ is a well-defined isometry. (ii) $\Rightarrow$ (iii) \ If $T$ is $g$-normal in some $C^*$-algebra $B$ generated by $A$, then we have a period 2 $*$-antiautomorphism $\sigma : B \to B$ with $\sigma(T) = T$. The restriction of $\sigma$ to $A$ maps onto $A$. So $A$ is an operator algebra with linear involution $\theta$ with $T^\theta = T$. (iii) $\Rightarrow$ (iv) \ Immediate from Corollary \ref{chsym}. \end{proof} \begin{example} One may ask if all operators $T$ satisfy the conditions in the last theorem, or in the one before it. However Halmos' example $x = 2E_{12} + E_{23}$ in $M_3$ may be shown to be a counterexample. Since $x^\intercal= x^*$ (we write $\intercal$ for the transpose) the same example will work for both. Indeed one can show that $x$ generates $M_3$ as a $C^*$-algebra, and since this is simple we have $C^*_e({\rm oa}(x)) = M_3$. Any $*$-automorphism of $M_3$ is inner, and also $*$-antiautomorphisms of $M_3$ are of form $u^* a^\intercal u$ for a unitary $u \in M_3$. An easy matrix computation show that there are no unitary solutions to $u^* x u = x^\intercal = x^*$. Thus $x$ is not $*$-exchangeable or g-normal in $M_3$, hence oa$(x)$ is not symmetric nor is an operator $*$-algebra with $x$ $\dagger$-selfadjoint. On the other hand, the matrix $x \oplus x^\intercal$ in $M_6$ does satisfy the conditions in the last two theorems (this may be seen similarly to the idea in the proof of Corollary \ref{chsym2}). \end{example} Many `truncated Toeplitz operators' are complex symmetric, and some are $*$-exchangeable, giving by the theorems above examples of operator $*$-algebras, and operator algebras with linear involution. To see these assertions it is helpful to recall the Sz.\ Nagy-Foias model theory for contractions \cite{NFK,Berc}. For many contractions $T$ it is known that $T$ is unitarily isomorphic to a truncated Toeplitz operator, a so-called {\em Jordan block} \cite[Chapter 3]{Berc}, namely the compression $S(u) = P_K S_{|K}$ of the unilateral shift $S$, viewed as multiplication by $z$ on $H^2$, to the subspace $K = H^2 \ominus u H^2$, for a (nonconstant) inner function $u$ on the disk. Thus the weak* closed algebra $A_T$ generated by $T$ (and $I$) is completely isometrically and weak* homeomorphically isomorphic to the weak* closed algebra generated by $S(u)$ (and $I$). On the other hand, the last weak* closed algebra is known to be equal to the commutant $\{ S(u) \}'$, and is isometrically weak* homeomorphic to the quotient $H^\infty/ u H^\infty$ (see \cite[Corollary 1.20]{Berc}). \begin{lemma} \label{symm} The weak* closed operator algebra generated by a Jordan block $S(u)$ is symmetric, indeed is a Q-algebra, for every inner function $u$. Thus $S(u)$ satisfies the conditions of the last theorem. \end{lemma} \begin{proof} In \cite[Corollary 1.20]{Berc}, it is shown that $A_{S(u)} = \{ S(u) \}'$ is isometrically weak* homeomorphic to the quotient $H^\infty/ u H^\infty$. Following the ideas in the proof of \cite[Corollary 1.20]{Berc} one can see that this isometry is a complete isometry. The functional calculus $H^\infty \to A_{S(u)}$ for $S(u)$ is a complete contraction since it has a positive unital, hence completely positive and completely contractive, extension to $L^\infty$. Thus we have an isometric complete contraction $H^\infty/ u H^\infty \to A_T$. The unilateral shift $S$ is a minimal isometric dilation of $S(u)$. Suppose that $x = [ x_{ij} ] \in {\rm Ball}(M_n(A_{S(u)}))$. Then $S^{(n)}$ is a minimal isometric dilation of $S(u)^{(n)}$. By the commutant lifting theorem (e.g.\ \cite[Theorem 1.10]{Berc}) there exists $y = [y_{ij} ] \in {\rm Ball}(M_n(B(H^2))$ such that $y \in \{ S^{(n)} \}'$, so that $y_{ij} \in \{ S \}'$, and $P_{K^{(n)}} y_{|K^{(n)}} = x$. Thus $P_K (y_{ij})_{|K} = x_{ij}$. Since $y_{ij} \in \{ S \}'$, and the $H^\infty$ functional calculus is a complete isometry $H^\infty \to \{ S \}'$, we see that $y_{ij} = f_{ij}(S)$ for $[f_{ij}] \in {\rm Ball}(M_n(H^\infty))$. Thus $[f_{ij} + u H^\infty] \in {\rm Ball}(M_n(H^\infty/ u H^\infty))$ is a preimage of $x$. It follows that $A_{S(u)} \cong H^\infty/ u H^\infty$ completely isometrically (and weak* homeomorphically). Now $H^\infty/ u H^\infty$ is a $Q$-algebra. Hence $A_{S(u)}$ is a $Q$-algebra and symmetric operator algebra. Its subalgebra oa$(S(u))$ is thus also symmetric, so $S(u)$ satisfies the conditions of the last theorem. \end{proof} As a consequence, the large class of contractions $T$ unitarily equivalent to a Jordan block $S(u)$ for some (nonconstant) inner function $u$ on the disk, all generate symmetric operator algebras, in particular operator algebras having linear involution. Turning to more specific examples, the Volterra operator $Vf(x) = \int_0^x \, f(t) \, dt$ on $L^2([0,1])$ is both $*$-exchangeable and complex symmetric (the latter via the conjugation $cf(t) = \overline{f(1-t)}$). Thus the operator algebra generated by $V$ is both an operator $*$-algebra and has linear involution. The same is true for the weak* closed algebra generated by $V$. These may be viewed as infinite dimensional versions of the upper triangular matrices. Indeed the Volterra operator $V$ is unitarily equivalent to $S(u)$ with $u(z) = \exp ((z+1)(z-1)^{-1})$, by e.g.\ \cite[Lemma 3.18 on p.\ 97]{Berc}, and this $u$ is invariant under the involution $f^\dagger(z) = \overline{f(\bar{z})}$. Thus $H^\infty/ u H^\infty$ is an operator $*$-algebra, indeed is an involutive Q-algebra, and also is a dual operator $*$-algebra. This is because $\dagger$ is a weak* continuous involution on $H^\infty$. Hence $A_V$ is a dual operator $*$-algebra. Similarly, the norm closed algebra oa$(V)$ generated by the Volterra algebra is completely isometrically isomorphic to $A_1(\Ddb)/u A_1(\Ddb)$ where $A_1(\Ddb)$ are the disk algebra functions vanishing at $1$ (the isomorphism $H^\infty/ u H^\infty \to A_V$ restricts to an isomorphism $A_1(\Ddb)/u A_1(\Ddb) \to {\rm oa}(V)$, see \cite{PW}). The latter quotient again is an operator $*$-algebra (since $A_1(\Ddb)$ and its ideal $u A_1(\Ddb)$ are invariant under the involution $\dagger$). So oa$(V)$ is an involutive Q-algebra, with $\dagger$-selfadjoint generator. The associated $\dagger$-selfadjoint contractive generator is $1 - (1+V)^{-1} = V(I + V)^{-1}$ (see the discussion just above Section \ref{Exs}), which corresponds to the image of $(1-z)/2 \in A_1(\Ddb)$. It is known to be a radical Banach algebra \cite{PW} so the spectrum of every element is $(0)$. Thus this is an example of an operator $*$-algebra such that every $\dagger$-selfadjoint element has real spectrum, but which is not a $C^*$-algebra. Indeed in this algebra for every $a \in A$ we have Sp$(a^\dagger a) \subset [0,\infty)$. Slightly more generally a contraction operator unitarily equivalent to Jordan block $S(\theta)$ for an inner function $\theta$, generates an operator $*$-algebra with $T^\dagger = T$ if $\theta(\bar{z}) = \overline{\theta(z)}$. Such inner functions include Blaschke products with real zeroes and the function $u$ in the previous paragraph. \subsection{Examples based on upper triangular matrices} The upper triangular $n$ by $n$ matrix algebra is an example which has all four types of involutions mentioned at the start of Section \ref{inv}. The $*$-algebra involution is given by $x^\dagger = u_n x^* u_n$ where $u_n$ is the order reversing $n \times n$ permutation matrix. Similarly $ u_n x^\intercal u_n$ is a linear involution, where $\intercal$ is the transpose. Similarly, the infinite dimensional version of the upper triangulars acting on $l^2(\Zdb)$ is an operator $*$-algebra and operator algebra with linear involution. Here $u((a_n)) = (a_{-n})$ for $(a_n) \in l^2(\Zdb)$, and $x^\dagger = u x^* u$, etc. These algebras have as one `involutive ideal' the strictly upper triangular subalgebra. The following is an example of an operator $*$-algebra which is a maximal subdiagonal algebra in the sense of Arveson \cite{Arv2} within the hyperfinite II$_1$ factor $R$. One could call this the {\em hyperfinite upper triangulars}. In $M_{2^n}$ consider conjugation by the order reversing permutation matrix which for convenience we write as $u_n$ (in the notation above it is $u_{2^n}$). Then we have $u_{n+1} (x \oplus x) u_{n+1} = (u_n x u_n) \oplus (u_n x u_n)$ for $x \in M_{2^n}$. It follows that $(u_n x^* u_n)$ gives rise to a well defined period 2 $*$-automorphism on the union ${\mathcal C}$ of the copies on $M_{2^n}$ . This extends by density to a period 2 $*$-automorphism $\theta$ of the CAR algebra $B$, and this gives an involution on the closure $A$ of the union of the upper triangular matrices in $M_{2^n}$ for all $n \in \Ndb$, since $\theta(A) \subset A^*$. So $A$ is an operator $*$-algebra. A similar construction using the transpose in place of $*$ gives a linear involution on $A$. Note that for the normalized trace $$\tau_n(\theta(x) y^*) = \tau_n(u_n x u_n y^*) = \tau_n(x \theta(y)^*), \qquad x, y \in M_{2^n},$$ so that $\theta$ extends to a symmetry $U$ on the Hilbert space of the GNS representation of the trace of $B$. Since $U A U \subset A^*$, it is easy to argue that the weak* closure $N$ of $A$ is a dual operator $*$-algebra inside ${\mathcal R}$, the hyperfinite II$_1$ factor. Similarly one has a linear involution on $N$. We claim that $N$ is a subdiagonal algebra in the sense of Arveson. If $\Phi_n : M_n \to D_n \subset M_n$ is the canonical projection onto the matrices supported on main diagonal in $M_n$, then $\Phi_{n+1}(x \oplus x) = \Phi_n(x) \oplus \Phi_n(x)$ for $x \in M_{2^n}$. Thus we obtain a trace preserving projection $\Psi$ from the CAR algebra $B$ onto the `main diagonal' part $D_0$ of $B$. Indeed $\tau(\Psi(x) y ) = \tau(\Psi(xy)) = \tau (xy)$ for $x \in B, y \in D_0$. On the other hand the canonical trace preserving conditional expectation $\Phi$ from $R$ onto the `main diagonal' part $D$ of $R$ restricts to a trace preserving conditional expectation from $B$ onto $D_0$, so by the unicity of the trace preserving normal conditional expectation we get that $\Phi$ extends $\Psi$. Since $\Psi$ is multiplicative on $B$, by density $\Phi$ is a homomorphism onto $D$. Also since $A + A^*$ is clearly dense in $B$, by density we have $N + N^*$ is weak* dense in $R$, so $N$ is a maximal subdiagonal algebra in the sense of Arveson. \subsection{The algebra of an involutive operator space} \label{uxy} Let $X$ be an operator system or selfadjoint subspace of $B(H)$, and consider, as in 2.2.10 in \cite{BLM}, $${\mathcal U}(X) = \biggl\{ \left[ \begin{array}{ccl} \lambda_1 & x\\ 0 &\lambda_2 \end{array} \right] \, :\, x\in X,\ \lambda_1 ,\lambda_2 \in\Cdb\biggr\} \subset B(H \oplus H),$$ where $\lambda_1$ and $\lambda_2$ stand for the operators $\lambda_1 I_H$ and $\lambda_2 I_H$ respectively. Note that ${\mathcal U}(X)$ may be regarded as a subspace of the Paulsen system ${\mathcal U}(X) + {\mathcal U}(X)^*$. Give ${\mathcal U}(X)$ the involution that we gave the upper triangular matrices, namely $(u_2 \otimes I_H) a^* (u_2 \otimes I_H)$, where $u_2$ is the usual permutation matrix on $\Cdb^2$. Then ${\mathcal U}(X)$ is an operator $*$-algebra. More abstractly we define an {\em operator} $*$-{\em space} to be an operator space $X$ with a period 2 conjugate linear bijection $* : X \to X$ satisfying $\| [a_{ji}^* ] \| = \| [ a_{ij}] \|.$ As shown in the introduction to \cite{BKNW} if $u : X \to B(K)$ is a linear complete isometry, then $$\Theta(x) \; = \; \left[ \begin{array}{ccl} 0 & u(x) \\ u(x^*)^* & 0 \end{array} \right] \in B(K \oplus K), \qquad x \in X, $$ is a $*$-linear complete isometry. So $X$ `is' a selfadjoint subspace of $B(H)$ for a Hilbert space $H$, and gives rise to an operator $*$-algebra ${\mathcal U}(X)$ as above. A special case that is sometimes used is when $X$ is an `involutive Banach space'. Then $X$ with its Min or Max operator space structure (see 1.2.21 and 1.2.22 in \cite{BLM}) will be an operator $*$-space. There is a similar construction for the other types of `involution' mentioned at the start of Section \ref{inv}. Namely, if $X$ is an operator space with an involution satisfying the conditions at the start of Section \ref{inv} of one of these four types, but with no multiplicativity or anti-multiplicativity condition assumed, then the operator algebra ${\mathcal U}(X)$ may be given an operator algebra involution of the matching type. \subsection{The algebra of an involutive bimodule} There is an operator module version of the last example. Since we plan to study operator modules in an involutive setting later we will be brief here. In \cite{BKM} a kind of involutive operator module is studied that is quite different to, and much more interesting than, the ones below, although the representation in the next result has a superficial similarity inspired by the `standard forms' considered there. \begin{theorem} Let $A$ be an approximately unital operator $\ast$-algebra and let $X$ be an operator $\ast$-space in the sense of Subsection {\rm \ref{uxy}} above. Suppose that $X$ is a nondegenerate operator $A$-$A$-bimodule in the sense of {\rm \cite[Chapter 3]{BLM}} such that $(ax)^{\dagger}=x^{\dagger}a^{\dagger},$ for any $a\in A$ and $x\in X.$ Then there exist a Hilbert space $H,$ a completely isometric linear map $\sigma: X\to B(H),$ and a nondegenerate completely isometric homomorphism $\pi$ of $A$ on $H,$ and selfadjoint unitary $u$ on $H,$ such that \begin{align*} &\sigma(x^{\dagger})=u\sigma(x)^{\ast}u \,\,\quad {\rm and} \quad \pi(a^{\dagger})=u\pi(a)^{\ast}u,\\ &\pi(a)\sigma(x)=\sigma(ax)\quad {\rm and} \quad \sigma(x)\pi(a)=\sigma(xa), \end{align*} for all $a\in A$ and $x\in X.$ \end{theorem} \begin{proof} By \cite[Theorem 3.3.1, Lemma 3.3.5]{BLM}, there exist a Hilbert space $H_0,$ a completely isometric linear map $\phi:X\to B(H_0),$ and nondegenerate completely completely isometric homomorphism $\Theta$ of $A$ such that \begin{align*} \Theta(a)\phi(x)=\phi(ax)\quad \mbox{and}\quad \phi(x)\Theta(b)=\phi(xb), \end{align*} for all $a, b\in A$ and $x\in X.$ We consider the Hilbert space $H :=H_0\oplus H_0$ and the completely isometric homomorphism $\pi: A\to B(H)$ given by \[\pi(a)=\begin{pmatrix} \Theta(a) & 0 \\ 0&\Theta(a^{\dagger})^{\ast} \end{pmatrix}\] and the complete isometry $\sigma: X\to B(H)$ given by \[\sigma(x)=\begin{pmatrix} \phi(x)& 0 \\ 0& \phi(x^{\dagger})^* \end{pmatrix}.\] The self-adjoint unitary $u=\begin{pmatrix} 0 & 1 \\ 1& 0 \end{pmatrix}$ implements the relations $$\sigma(x^{\dagger})=u\sigma(x)^{\ast}u \,\,\quad {\rm and} \quad \pi(a^{\dagger})=u\pi(a)^{\ast}u.$$ Moreover, it is easy to see that $\pi(a)\sigma(x)=\sigma(ax)$ and $\sigma(x)\pi(a)=\sigma(xa).$ \end{proof} Let $X$ be a nondegenerate operator $A$-$A$-bimodule over an approximately unital operator $*$-algebra $A$, of the type characterized in the last theorem. For $H, \pi, \sigma$ as in that theorem consider $${\mathcal U}_A(X) = \biggl\{ \left[ \begin{array}{ccl} \pi( a_1) & \sigma(x) \\ 0 & \pi(a_2) \end{array} \right] \, :\, x\in X,\ a_1, a_2 \in A \biggr\} \subset B(H \oplus H).$$ This is an operator $*$-algebra, with involution $a \mapsto (u_2 \otimes I_H) a^* (u_2 \otimes I_H)$, where $u_2$ is the usual permutation matrix on $\Cdb^2$, similarly to Example \ref{uxy}. There are similar constructions for some of the other type of involutions listed at the start of Section \ref{inv}. \medskip We mention a few final examples. If $A$ is an operator $*$-algebra then so are $A^\circ$ and $A^\star$. For any operator algebra $A$ we have that $A \oplus A^\star$ is an operator $*$-algebra with obvious involution. Example 1.9 (2) of \cite{BKM} is a natural example of an operator $*$-algebra inside the reduced free group $C^*$-algebra. This may be modified to give examples in the group von Neumann algebra or full group $C^*$-algebra. Finally, there are many examples of period 2 automorphisms in the literature of operator algebras, although they are usually not very explicit. For example they sometimes occur as a special case of finite group actions on operator algebras, or the $\Zdb_2$-action case of crossed product operator algebras. \section{Operator $\ast$-algebras} Henceforth in our paper for specificity our involutive algebras will be operator $\ast$-algebras. As said earlier, we leave the case of the remaining material for the other kinds of involutions to the reader. We remark that the $C^{\ast}$-algebras which are operator $\ast$-algebras are exactly the $\Zdb_2$-{\em graded $C^*$-algebras}. An {\em involutive ideal} or $\dagger${\em -ideal} in an operator algebra with involution $\dagger$ is an ideal $J$ with $J^\dagger \subset J$. \begin{proposition}\label{qdoai} Let $A$ be an operator $\ast$-algebra. Suppose $J$ is a closed $\dagger$-ideal, then $J$ and $A/J$ are operator $\ast$-algebras. \end{proposition} \begin{proof} This follows from the matching fact for operator algebras \cite[Proposition 2.3.4]{BLM}, and the computation $$\Vert [a_{ji}^\dagger + J ] \Vert \leq \Vert [a_{ji}^\dagger + x_{ji}^\dagger] \Vert \leq \Vert [a_{ij}+ x_{ij}] \Vert, \qquad x_{ij} \in J,$$ so that $\Vert [a_{ji}^\dagger + J ] \Vert \leq \Vert [a_{ij}+ J] \Vert$ for $a_{ij} \in A$. Similarly, we have $\Vert [a_{ij}+ J] \Vert\leq \Vert [a_{ji}^\dagger + J ] \Vert.$ \end{proof} {\bf Remark.} There are $*$-algebra variants of the usual consequences of the matching fact in operator algebra theory. For example one may deduce easily from Proposition \ref{qdoai} following the method in e.g.\ 1.2.30, 2.3.6, 2.3.7 in \cite{BLM}, that one may {\em interpolate} between operator $*$-algebras. Indeed suppose that $(A_0, A_1)$ is a compatible couple of Banach $*$-algebras which happen to be operator $*$-algebras. Just like in the general operator space case \cite[1.2.30]{BLM}, let $\Sl$ be the strip of all complex numbers $z$ with $0\leq \re z\leq 1$ and let $\Fl=\Fl(A_0, A_1)$ be the space of all bounded and continuous functions $f:\Sl \to A_0+A_1$ such that the restriction of $f$ to the interior of $\Sl$ is analytic, and such that the maps $t\to f(it)$ and $t\to f(1+it)$ belong to $C_0(\Rdb; A_0)$ and $C_0(\Rdb, A_1)$ respectively. For any $f\in \Fl,$ the function $f^{\dagger}$ is defined by $f^{\dagger}(z)=f(\bar{z})^* \in \Fl.$ The last asterisk here is the involution on $A_0+A_1$. Then $\Fl(A_0, A_1)$ with the operator space considered in 1.2.30 in \cite{BLM} is an operator $\ast$-algebra with the involution $\dagger$. For any $0\leq \theta \leq 1,$ let $\Fl_{\theta}(A_0, A_1)$ be the two-sided closed ideal of all $f\in\Fl$ for which $f(\theta)=0.$ This is $\dagger$-selfadjoint. The interpolation space $A_{\theta}=[A_0, A_1]_{\theta}$ is the subspace of $A_0+ A_1$ formed by all $x=f(\theta)$ for some $f\in\Fl$. As operator spaces, the interpolation space $A_{\theta} \cong \Fl(A_0, A_1)/\Fl_{\theta}(A_0, A_1)$ through the map $\pi: f\mapsto f(\theta).$ It is easy to see that $\pi$ is $\dagger$-linear. By Proposition \ref{qdoai}, the quotient $A_{\theta}\cong \Fl(A_0, A_1)/\Fl_{\theta}(A_0, A_1)$ is an operator $\ast$-algebra. \subsection{Contractive approximate identities} \begin{lemma}\label{dcai} Let $A$ be an operator $\ast$-algebra. Then the following are equivalent: \begin{itemize} \item[(i)] $A$ has a cai. \item[(ii)]$A$ has a $\dagger$-selfadjoint cai. \item[(iii)] $A$ has a left cai. \item[(iv)] $A$ has a right cai. \item[(v)] $A^{\ast\ast}$ has an identity of norm $1.$ \end{itemize} \end{lemma} \begin{proof} (i) $\Rightarrow$ (ii) \ If $(e_t)$ is a cai for $A,$ then $(e_t^{\dagger})$ is also a cai for $A.$ Let $f_t=(e_t+e_t^{\dagger})/2,$ then $(f_t)$ is a $\dagger$-selfadjoint cai for $A.$ (iii) $\Rightarrow$ (iv) \ If $(e_t)$ is a left cai for $A,$ then $(e_t^{\dagger})$ is a right cai. Analogously, it is easy to see that (iv) $\Rightarrow$ (iii.) (iv) $\Rightarrow$ (v) \ By a well-known fact in operator algebra that if $A$ has a left cai and right cai, then $A^{\ast\ast}$ has an identity of norm $1$ (see e.g. \cite[Proposition 2.5.8]{BLM}). That (ii) $\Rightarrow$ (i), and (i) $\Rightarrow$ (iii), are obvious. That (v) $\Rightarrow$ (i) follows from Proposition 2.5.8 in \cite{BLM}. \end{proof} \begin{corollary}\label{cdgcai12} If $A$ is an operator $\ast$-algebra with a countable cai $(f_n),$ then $A$ has a countable $\dagger$-selfadjoint cai in $\frac{1}{2}{\mathfrak F}_A.$ \end{corollary} \begin{proof} By \cite[Theorem 1.1]{BRI}, $A$ has a cai $(e_t)$ in $\frac{1}{2}{\mathfrak F}_A.$ Denote $e_t'=\frac{e_t+e_t^{\dagger}}{2},$ then $(e_t')$ is also a cai in $\frac{1}{2}{\mathfrak F}_A.$ Choosing $t_n$ with $\Vert f_n e_{t_n}'-f_n\Vert +\Vert e_{t_n}'f_n-f_n\Vert <2^{-n},$ it is easy to see that $(e_{t_n}')$ is a countable $\dagger$-selfadjoint cai in $\frac{1}{2}{\mathfrak F}_A.$ \end{proof} \begin{corollary} If $J$ is a closed two-sided $\dagger$-ideal in an operator $\ast$-algebra $A$ and if $J$ has a cai, then $J$ has a $\dagger$-selfadjoint cai $(e_t)$ with $\Vert 1-2e_t\Vert\leq 1$ for all $t,$ which is also quasicentral in $A$. \end{corollary} \begin{proof} By the proof of Corollary \ref{cdgcai12}, we know that $J$ has a $\dagger$-selfadjoint cai, denoted $(e_t),$ in $\frac{1}{2}{\mathfrak F}_A.$ The rest is as in the proof of \cite[Corollary 1.5]{BRI}. \end{proof} Let $A$ be an operator algebra (possibly not unital). Then the {\em left (resp.\, right) support projection} of an element $x$ in $A$ is the smallest projection $p\in A^{\ast\ast}$ such that $px=x$ (resp.\, $xp=x$), if such a projection exists (it always exists if $A$ has a cai, see e.g. \cite{BRI}). If the left and right support projection exist, and are equal, then we call it the {\em support projection}, written $s(x).$ \begin{theorem}\cite[Corollary 3.4]{BRII} For any operator algebra $A,$ if $x\in {\mathfrak r}_A$ and $x\neq 0,$ then the left support projection of $x$ equals the right support projection, and equals the weak* limit of $(x^{1/n})$. It also equals $s(y),$ where $y=x(1+x)^{-1}\in \frac{1}{2}{\mathfrak F}_A$. Also, $s(x)$ is open in $A^{\ast\ast}$. \end{theorem} \begin{proposition} \label{sdaggi} In an operator $*$-algebra $A$, ${\mathfrak F}_A$ and ${\mathfrak r}_A$ are $\dagger$-closed, and if $x \in {\mathfrak r}_A$ we have $(x^\dagger)^{1/n} = (x^{1/n})^{\dagger}$ for $n \in \Ndb$, and $s(x)^\dagger = s(x^{\dagger})$. If $x \in {\mathfrak F}_A$ then $s(x) \vee s(x^\dagger) = s(x + x^\dagger)$. In particular if $x$ is $\dagger$-selfadjoint then so is $s(x)$. \end{proposition} \begin{proof} Applying $\dagger$ we see that $\| 1 - x \| \leq 1$ implies $\| 1 - x^\dagger \| \leq 1$. For the $\dagger$-invariance of ${\mathfrak r}_A$ note that this is easy to see for a $C^*$-cover $B$ with compatible involution (Definition \ref{ivltcmptb}), and then one may use the fact that ${\mathfrak r}_A = A \cap {\mathfrak r}_B$. Similarly if $x \in {\mathfrak r}_A$ and $\sigma$ is the $*$-automorphism on $B$ inducing $\dagger$ then $\sigma(x^{1/n}) = \sigma(x)^{1/n}$ clearly. Applying $*$, it follows that $(x^{\dagger})^{1/n} = (x^{1/n})^\dagger.$ Note that if $x \in {\mathfrak F}_A$ then this may be seen more explicitly since $x^{1/n}$ may be written as a power series in $1-x$ with real coefficients. Then $s(x)^\dagger = (w^{\ast}\lim_n x^{1/n})^\dagger = s(x^{\dagger})$. The $s(x + x^\dagger)$ assertion follows from e.g.\ the proof of \cite[Proposition 2.14]{BRI}. \end{proof} Thus the theory of real positivity studied in many of the first authors recent papers will have good involutive variants. See the end of Section \ref{inv} for notation used in the following result and proof. \begin{corollary} \label{fbab} For any operator $\ast$-algebra $A,$ if $x\in {\mathfrak r}_A$ is $\dagger$-selfadjoint, then $a={\mathfrak F}(x)=x(1+x)^{-1}\in \frac{1}{2}{\mathfrak F}_A$ is $\dagger$-selfadjoint, and $\overline{xA}=\overline{aA}=s(x)A^{\ast\ast}\cap A$ is a right ideal in $A$ with a $\dagger$-selfadjoint left cai. Also, $\overline{xAx}=\overline{aAa}$ is a $\dagger$-selfadjoint HSA whose support projection is $\dagger$-selfadjoint. \end{corollary} \begin{proof} It is an exercise that $a=x(1+x)^{-1}$ is $\dagger$-selfadjoint, and is in $\frac{1}{2}{\mathfrak F}_A$ by the previous result. Since $(a^{1/n})$ is $\dagger$-selfadjoint by a fact in the last proof, $(a^{1/n})$ serves as a $\dagger$-selfadjoint left cai for $\overline{aA}.$ Besides, $\overline{aAa}$ is $\dagger$-selfadjoint and the weak* limit of $(a^{1/n})$ is $s(a).$ The rest follows from \cite[Corollary 3.5]{BRII}. \end{proof} We remark that by the operator algebra theory, $\overline{xAx}$ is the HSA matching the right ideal $\overline{xA}$ under the bijective correspondence between these objects (see e.g. \cite{BHN,BRI,BRII}). See also Section 5.1 below for the involutive variant of this correspondence. \begin{lemma} If $x\in {\mathfrak F}_A,$ with $x\neq 0,$ then the operator $\ast$-algebra generated by $x$, denoted $\osa{x},$ has a cai. Indeed, the operator $\ast$-algebra $\osa{x}$ has a $\dagger$-selfadjoint sequential cai belonging to $\frac{1}{2}{\mathfrak F}_A.$ \end{lemma} \begin{proof} If $x\in {\mathfrak F}_A,$ then $x^{\dagger}\in {\mathfrak F}_A$ as we proved above. Denote $B=C^{\ast}_e(A),$ then $p = s(x)\vee s(x^{\dagger}) = s(x + x^\dagger)$ in $B^{\ast\ast}$ is in $\osa{x}^{\ast\ast}.$ Clearly $p x=x p=x$ and $p x^{\dagger}=x^{\dagger} p=x^{\dagger}.$ Therefore, $p$ is an identity in $\osa{x}^{\ast\ast}.$ By \cite[Theorem 2.5.8]{BLM}, $\osa{x}$ has a cai. Moreover, since $\osa{x}$ is separable, by \cite [Corollary 2.17]{BRI}, there exists $a\in{\mathfrak F}_A$ such that $s(a)=1_{\osa{x}^{\ast\ast}}$. Therefore $\osa{x}$ has a countable $\dagger$-selfadjoint cai by applying to \cite[Theorem 2.19]{BRI} and Corollary \ref{cdgcai12}. \end{proof} We write $x \preccurlyeq y$ if $y - x \in {\mathfrak r}_A$. The ensuing `order theory' in the involutive case is largely similar to the operator algebra case from \cite{BRord}. For example: \begin{theorem} \label{havin} Let $A$ be an operator $*$-algebra which generates a $C^*$-algebra $B$ with compatible involution $\dagger$, and let ${\mathcal U}_A = \{ a \in A : \Vert a \Vert < 1 \}$. The following are equivalent: \begin{itemize} \item [(1)] $A$ is approximately unital. \item [(2)] For any $\dagger$-selfadjoint positive $b \in {\mathcal U}_B$ there exists $\dagger$-selfadjoint $a \in {\mathfrak c}_A$ with $b \preccurlyeq a$. \item [(2')] Same as {\rm (2)}, but also $a \in \frac{1}{2} {\mathfrak F}_A$ and `nearly positive' in the sense of the introduction to {\rm \cite{BRord}}: we can make it as close in norm as we like to an actual positive element. \item [(3)] For any pair of $\dagger$-selfadjoint elements $x, y \in {\mathcal U}_A$ there exist nearly positive $\dagger$-selfadjoint $a \in \frac{1}{2} {\mathfrak F}_A$ with $x \preccurlyeq a$ and $y \preccurlyeq a$. \item [(4)] For any $\dagger$-selfadjoint $b \in {\mathcal U}_A$ there exist nearly positive $\dagger$-selfadjoint $a \in \frac{1}{2} {\mathfrak F}_A$ with $-a \preccurlyeq b \preccurlyeq a$. \item [(5)] For any $\dagger$-selfadjoint $b \in {\mathcal U}_A$ there exist $\dagger$-selfadjoint $x, y \in \frac{1}{2} {\mathfrak F}_A$ with $b = x-y$. \item [(6)] ${\mathfrak r}_A$ is a generating cone, indeed any $\dagger$-selfadjoint element in $A$ is a difference of two $\dagger$-selfadjoint elements in ${\mathfrak r}_A$. \item [(7)] Same as {\rm (6)} but with ${\mathfrak r}_A$ replaced by $\Rdb_+ \, {\mathfrak F}_A$. \end{itemize} \end{theorem} \begin{proof} (1) $\Rightarrow$ (2') \ By the proof in \cite[Theorem 2.1]{BRord} for any $\dagger$-selfadjoint positive $b \in {\mathcal U}_B$ there exists $c \in \frac{1}{2} {\mathfrak F}_A$ and nearly positive with $b \leq {\rm Re} \ c$. Hence it is easy to see that $b \leq {\rm Re} \, (c^\dagger)$ and $b \leq {\rm Re} \, a$ where $a = (c + c^\dagger)/2$. (2') $\Rightarrow$ (3) \ By $C^*$-algebra theory there exists positive $b \in {\mathcal U}_B$ with ${\rm Re} \, x$ and ${\rm Re} \, y$ both $\leq b$. It is easy to see that $b^\dagger = \sigma(b) \geq 0$. Then ${\rm Re} \, x \leq b^\dagger$, so that ${\rm Re} \, x \leq (b + b^\dagger)/2$. Similarly for $y$. Then apply (2') to obtain $a$ from $(b + b^\dagger)/2$. The remaining implications follow the proof in \cite[Theorem 2.1]{BRord} but using tricks similar to the ones we have used so far in this proof. We leave the details to the reader. \end{proof} {\bf Remark.} Similarly as in Proposition 2.6 in \cite{BRord}, but using our Theorem \ref{havin} (3) in the proof in place of the matching result referenced there, one can show that for an approximately unital operator $\ast$-algebra $A$, the $\dagger$-selfadjoint elements of norm $< 1$ in $\frac{1}{2}{\mathfrak F}_A$ is a directed set in the $\preccurlyeq$ ordering, and is a cai for $A$ which is increasing in this ordering. \medskip The following is a version of the Aarnes-Kadison Theorem for operator $\ast$-algebras. \begin{theorem}[Aarnes-Kadison type Theorem] If $A$ is an operator $\ast$-algebra then the following are equivalent: \begin{itemize} \item[(i)] There exists a $\dagger$-selfadjoint $x\in{\mathfrak r}_A$ with $A$ the norm closure of $xAx.$ \item[(ii)] There exists a $\dagger$-selfadjoint $x\in{\mathfrak r}_A$ with $A$ the norm closure of $xA$, which equals the norm closure of $Ax.$ \item[(iii)] There exists a $\dagger$-selfadjoint $x\in{\mathfrak r}_A$ with $s(x)=1_{A^{\ast\ast}}.$ \item[(iv)] $A$ has a countable $\dagger$-selfadjoint cai. \item[(v)] $A$ has a $\dagger$-selfadjoint and strictly real positive element. \end{itemize} Indeed these are all equivalent to the same conditions with `$\dagger$-selfadjoint' removed. \end{theorem} \begin{proof} In (i)--(iii) we can assume that $x\in{\mathfrak F}_A$ by replacing it with the $\dagger$-selfadjoint element $x(1+x)^{-1} \in \frac{1}{2} {\mathfrak F}_A$ (see \cite[Section 2.2]{BRord}). Then the equivalence of (i)--(iv) follows as in \cite[Lemma 2.10 and Theorem 2.19]{BRI}, for (iv) using that $x^{\frac{1}{n}}$ is $\dagger$-selfadjoint as we said in the proof of Proposition \ref{sdaggi}. Similarly (v) follows from these by \cite[Lemma 2.10]{BRI}, and the converse follows since strictly real positive elements have support projection $1$ (see \cite[Section 3]{BRord}). The final assertion follows since if $A$ has a countable cai, then $A$ has a $\dagger$-selfadjoint countable cai (Lemma \ref{dcai}). \end{proof} \subsection{Cohen factorization for operator $\ast$-algebras and their modules} The Cohen factorization theorem is a crucial tool for Banach algebras, operator algebras and their modules. In this section we will give a variant that works for operator $\ast$-algebras and their modules. Recall that if $X$ is a Banach space and $A$ is a Banach algebra then $X$ is called a {Banach $A$-module} if there is a module action $A\times X\to X$ which is a contractive linear map. If $A$ has a bounded approximate identity $(e_t)$ then we say that $X$ is nondegenerate if $e_t x\to x$ for $x\in X.$ A {\em Banach $A$-bimodule} is both a left and a right Banach $A$-module such that $a(xb)=(ax)b.$ The following is an operator $\ast$-algebra version of the Cohen factorization theorem: \begin{theorem} If $A$ is approximately unital operator $\ast$-algebra, and if $X$ is a nondegenerate Banach $A$-module(resp.\ $A$-bimodule), if $b\in X$ then there exists an element $b_0\in X$ and a $\dagger$-selfadjoint $a\in {\mathfrak F}_A$ with $b=ab_0$ (resp.\ $b=ab_0a$). Moreover if $\Vert b\Vert<1$ then $b_0$ and $a$ may be chosen of norm $<1.$ \end{theorem} \begin{proof} In \cite[Theorem 4.1]{PED}, the $a$ is constructed from convex combinations of elements in a cai, and in our case the cai may be chosen $\dagger$-selfadjoint by Lemma \ref{dcai}. The details are left as an exercise to the reader. \end{proof} \subsection{Multiplier algebras} \begin{theorem}\label{lefmult} Let $A$ be an approximately unital operator $\ast$-algebra. Then the following algebras are completely isometrically isomorphic: \begin{itemize} \item[(i)] $LM(A)=\{\eta\in A^{\ast\ast}: \eta A\subset A\},$ \item[(ii)] $LM(\pi)=\{T \in B(H): T\pi(A)\subset\pi(A)\},$ where $\pi$ is a nondegenerate completely isometric representation of $A$ on a Hilbert space $H$ such that there exists an order 2 $\ast$-automorphism $\sigma: B(H)\to B(H)$ satisfying $\sigma(\pi(a))^{\ast}=\pi(a^{\dagger})$ for any $a\in A,$ \item[(iii)] the set of completely bounded right $A$-module maps $CB_A(A).$ \end{itemize} \end{theorem} \begin{proof} See \cite[Theorem 2.6.3]{BLM}. \end{proof} \begin{definition} Let $A$ be an approximately unital operator $\ast$-algebra. Then we define \begin{itemize} \item[(i)] $RM(A)=\{\xi\in A^{\ast\ast}: A\xi\subset A\};$ \item[(ii)] $RM(\pi)=\{S\in B(H): \pi(A)S\subset\pi(A)\},$ for any nondegenerate completely isometric representation $\pi$ of $A$ on a Hilbert space $H$ and there exists order-2 $\ast$-automorphism $\sigma: B(H)\to B(H)$ satisfies $\sigma(\pi(a))^{\ast}=\pi(a^{\dagger})$ for any $a\in A;$ \item[(iii)] the set of completely bounded left $A$-module maps, which we denote as $_A{CB}(A).$ \end{itemize} \end{definition} \begin{corollary} Let $A$ be an approximately unital operator $\ast$-algebra. Then \begin{itemize} \item[(a)] $\eta\in LM(A)$ if and only if $\eta^{\dagger}\in RM(A),$ where $\eta, \eta^{\dagger}\in A^{\ast\ast}$ and $\dagger$ is the involution in $A^{**};$ \item[(b)] $T\in LM(\pi)$ if and only if $T^{\dagger}\in RM(\pi),$ where $T^{\dagger}=\sigma(T)^{\ast};$ \item[(c)] $L\in CB_A(A)$ if and only if $L^{\dagger}\in$ $_A{CB}(A),$ where the map $L^{\dagger}$ is defined by $L^{\dagger}(a)=L(a^{\dagger})^{\dagger}.$ \end{itemize} \end{corollary} \begin{proof} We just give the proof of (b). Suppose that $T\in LM(\pi),$ then \begin{align*} \pi(a)T^{\dagger}=\sigma(\pi(a^{\dagger}))^{\ast}\sigma(T)^{\ast}=(\sigma(T\pi(a^{\dagger})))^{\ast}\in \sigma(\pi(A))^{\ast}\subset \pi(A). \end{align*} Thus, $T^{\dagger}\in RM(\pi).$ Similarly, if $T^{\dagger}\in RM(\pi)$ then $T\in LM(\pi).$ \end{proof} We consider pairs $(D,\mu)$ consisting of a unital operator $\ast$-algebra $D$ and a completely isometric $\dagger$-homomorphism $\mu:A\to D,$ such that $D\mu(A)\subset\mu(A),\ \mu(A)D\subset\mu(A)$. We use the phrase {\em multiplier operator $\ast$-algebra} of $A,$ and write $M(A),$ for any pair $(D,\mu)$ which is completely $\dagger$-isometrically $A$-isomorphic to $\Ml(A)=\{x\in A^{\ast\ast}: xA\subset A\ \mbox{and}\ Ax\subset A\}.$ Note that by Lemma \ref{bidu} the inclusion of $A$ in $A^{**}$ is a $\dagger$-homomorphism, hence the canonical map $i:A\to \Ml(A)$, is a $\dagger$-homomorphism. From this it follows that there is a unique involution on $M(A)$ for which $i$ is a $\dagger$-homomorphism. \begin{proposition} Suppose that $A$ is an approximately unital operator $\ast$-algebra. If $(D,\mu)$ is a left multiplier operator algebra of $A,$ then the closed subalgebra $$\{d\in D:\mu(A)d\subset \mu(A)\}$$ of $D,$ together with the map $\mu$, is a multiplier operator $\ast$-algebra of $A.$ \end{proposition} \begin{proof} Let $E$ denote the set $\{d\in D:\mu(A)d\subset \mu(A)\}$. By \cite[Proposition 2.6.8]{BLM}, we know that $E$ is a multiplier operator algebra of $A. $ Thus, there exists a completely isometric surjective homomorphism $\theta:\Ml(A)\to E$ such that $\theta\circ i_A=\mu.$ Now we may define an involution on $E$ by $d^{\dagger}=\theta(\eta^{\dagger})$ if $d=\theta(\eta).$ Then it is easy to check that $E$ is an operator $\ast$-algebra which is completely $\dagger$-isometrically $A$-isomorphic to $\Ml(A).$ \end{proof} \begin{example} Let $A= A_1(\Ddb)$, the functions in the disk algebra vanishing at $1$, which is the norm closure of $(z-1) A(\Ddb)$, and let $B = \{ f \in C(\Tdb) : f (1) = 0 \}$. By the nonunital variant of the Stone-Weierstrass theorem, $B$ is generated as a $C^*$-algebra by $A$. Indeed $B = C^*_e(A)$, since any closed ideal of $B$ is the set of functions that vanish on a closed set in the circle containing $1$. Also for any $z_0 \in \Tdb$, $z_0 \neq 1$, there is a function in $A$ that peaks at $z_0$, if necessary by the noncommutative Urysohn lemma for approximately unital operator algebras \cite{BNII}. So the involution on $A$ descends from the natural involution on $B$. It is easy to see, for example by examining the bidual of $B^{**}$ and noticing that $A$ and $B$ have a common cai, that $M(A) = \{ T \in M(B) : TA \subset A \} = \{ g\in C_b(\Tdb \setminus \{1 \}) : g (z-1) \in A(\Ddb) \}$. For such $g$, since the negative Fourier coefficients of $k=g(1-z)$ are zero, the negative Fourier coefficients of $g$ are constant, hence zero by the Riemann-Lebesgue lemma. Thus $g$ is in $H^\infty$, and has an analytic extension to the open disk. Viewing $g$ as a function $h$ on $\bar{\Ddb} \setminus \{1 \}$ we have $h = k/(z-1)$ for some $k \in A(\Ddb)$. So $M(A)$ consists of the bounded continuous functions on $\bar{\Ddb} \setminus \{1 \}$ that are analytic in the open disk, with involution $\overline{f(\bar{z})}.$ \end{example} Let $A, B$ be approximately unital operator $\ast$-algebras. A completely contractive $\dagger$-homomorphism $\pi: A\to M(B)$ will be called a {\em multiplier-nondegenerate $\dagger$-morphism}, if $B$ is a nondegenerate bimodule with respect to the natural module action of $A$ on $B$ via $\pi.$ This is equivalent to saying that for any cai $(e_t)$ of $A,$ we have $\pi(e_t)b\to b$ and $b\pi(e_t)\to b$ for $b\in B.$ \begin{proposition}\label{multext} If $A, B$ are approximately operator $\ast$-algebras, and if $\pi:A\to M(B)$ is a multiplier-nondegenerate $\dagger$-morphism then $\pi$ extends uniquely to a unital completely contractive $\dagger$-homomorphism $\hat{\pi}: M(A)\to M(B).$ Moreover $\hat{\pi}$ is completely isometric if and only if $\pi$ is completely isometric. \end{proposition} \begin{proof} Regard $M(A)$ and $M(B)$ as $\dagger$-subalgebras of $A^{\ast\ast}$ and $B^{\ast\ast}$ respectively. Let $\tilde{\pi}:A^{\ast\ast}\to B^{\ast\ast}$ be the unique $w^{\ast}$-continuous $\dagger$-homomorphism extending $\pi.$ From \cite[Proposition 2.6.12]{BLM}, we know that $\hat{\pi}=\tilde{\pi}(\cdot)_{\vert_{M(A)}}$ is the unique bounded homomorphism on $M(A)$ extending $\pi,$ and $\hat{\pi}(M(A))\subset M(B).$ Let $(e_t)$ be a $\dagger$-selfadjoint cai for $A.$ Then for any $\eta\in M(A),$ $\eta e_t \in A$ and $\eta e_t\xrightarrow{w^{\ast}} \eta$ . Hence $$\hat{\pi}(\eta)=w^{\ast}-\lim_{t}\pi(\eta e_t).$$ On the other hand, $(\eta e_t)^{\dagger}\to \eta^{\dagger},$ which implies $$\hat{\pi}(\eta^{\dagger})=w^{\ast}-\lim_{t}\pi((\eta e_t)^{\dagger})=w^{\ast}-\lim_{t}\pi((\eta e_t))^{\dagger}.$$ Since the involution on $A^{\ast\ast}$ is $w^{\ast}$-continuous, we get that $$\hat{\pi}(\eta^{\dagger})=\hat{\pi}(\eta)^{\dagger}.$$ The rest follows from \cite[Proposition 2.6.12]{BLM}. \end{proof} \subsection{Dual operator $\ast$-algebras} \begin{definition} Let $M$ be a dual operator algebra (that is, an operator algebra with an operator space predual--see \cite[Section 2.7]{BLM}) and operator $\ast$-algebra such that the involution on $M$ is weak* continuous. Then $M$ is called a {\em dual operator $\ast$-algebra}. \end{definition} We will identify any two dual operator $\ast$-algebras $M$ and $N$ which are $w^*$-homeomorphically and completely $\dagger$-isometrically isometric. \begin{proposition} Let $M$ be a dual (possibly nonunital) operator $\ast$-algebra. \begin{itemize} \item[(1)] The $w^*$-closure of a $\ast$-subalgebra of $M$ is a dual operator $\ast$-algebra. \item[(2)] The unitization of $M$ is also a dual operator $\ast$-algebra. \end{itemize} \end{proposition} \begin{proof} For (1), the weak* -closure of $\ast$-subalgebra of $M$ is a dual operator algebra by \cite[Proposition 2.7.4 (4)]{BLM}. For (2), suppose that $M$ is a nonunital operator $\ast$-algebra and write $I$ for the identity in $M^1.$ Suppose that $(x_t)_t$ and $(\lambda_t)_t$ are nets in $M$ and $\Cdb$ respectively, with $(x_t+\lambda_t I)$ converging in $w^*$-topology. By applying a nonzero weak* continuous functional annihilating $M$, it is easy to see that $(\lambda_t)_t$ converges in $\Cdb.$ It follows that $(x_t)_t$ converges in $M$ in the $w^{*}$-topology. Thus, $(x_t+\lambda_t I)^{\dagger}$ converges in $M^1,$ in the $w^{*}$-topology. The rest follows immediately from \cite[Proposition 2.7.4 (5)]{BLM}.\end{proof} \begin{proposition} Let $A$ be an operator $\ast$-algebra, and $I$ any cardinal. Then $\Kdb_I(A)^{\ast\ast}\cong \Mdb_I(A^{\ast\ast})$ as dual operator $\ast$-algebras. \end{proposition} \begin{proof} The canonical embedding $A\subset A^{\ast\ast}$ induces a completely isometric $\dagger$-homomorphism $\theta: \Kdb_I(A)\to \Kdb_I(A^{\ast\ast})\subset \Mdb_{I}(A^{\ast\ast}).$ Notice that the involutions on $\Kdb_I(A)^{\ast\ast},$ $\Mdb_I(A^{\ast\ast})$ are $w^*$-continuous and $\Kdb_I(A)^{\ast\ast}\cong \Mdb_I(A^{\ast\ast})$ as operator algebras. Thus, $\Kdb_I(A)^{\ast\ast}\cong \Mdb_I(A^{\ast\ast})$ as dual operator $\ast$-algebras. \end{proof} \begin{lemma} \label{uxd} If $X$ is a weak* closed selfadjoint subspace of $B(H)$ for a Hilbert space $H$, then $\Ul(X)$ as defined as in Example \ref{uxy} is a dual operator $\ast$-algebra. \end{lemma} \begin{proof} We leave this as an exercise to the reader. \end{proof} In connection with the last result we note that any operator $*$-space $X$ in the sense of \ref{uxy}, which is a dual operator space, and whose involution is weak* continuous, may be embedded weak* homeomorphically, via a $*$-linear complete isometry, as a weak* closed selfadjoint subspace of $B(H)$ for a Hilbert space $H$. So $\Ul(X)$ is a dual operator $\ast$-algebra, again by Lemma \ref{uxd}. To see this simply use the proof in \ref{uxy}, taking $u$ there to be a weak* homeomorphic complete isometry from $X$ into $B(H)$. The last result can be used to produce counterexamples concerning dual operator $*$-algebras, such as algebras with two distinct preduals, etc. Similarly one may use the $\Ul(X)$ construction to easily obtain an example of a dual operator algebra which is an operator $\ast$-algebra, but the involution is not weak*-continuous. We omit the details. Recall that in \cite{BS} the maximal $W^*$-algebra $W^*_{\rm max}(M)$ was defined for unital dual operator algebras $M$. If $M$ is a dual operator algebra but is not unital we define $W^*_{\rm max}(M)$ to be the von Neumann subalgebra of $W^*_{\rm max}(M^1)$ generated by the copy of $M$. Note that it has the desired universal property: if $\pi : M \to N$ is a weak* continuous completely contractive homomorphism into a von Neumann algebra $N$, then by the normal version of Meyer's theorem (see \cite[Proposition 2.7.4 (6)]{BLM}) we may extend to a weak* continuous completely contractive unital homomorphism $\pi^1 : M^1 \to N$. Hence by the universal property of $W^*_{\rm max}(M^1)$, we may extend further to a normal unital $*$-homomorphism from $W^*_{\rm max}(M^1)$ into $N$. Restricting to $W^*_{\rm max}(M)$ we have shown that there exists a normal $*$-homomorphism $\tilde{\pi} : W^*_{\rm max}(M) \to N$ extending $\pi$. \begin{proposition} Let $B = W^*_{\rm max}(M)$. Then $M$ is a dual operator $*$-algebra if and only if there exists an order two $*$-automorphism $\sigma : B \to B$ such that $\sigma(M) = M^*$. In this case the involution on $M$ is $a^{\dagger} = \sigma(a)^*$. \end{proposition} \begin{proof} This follows from a simple variant of the part of the proof of Theorem \ref{opinvch} that we did prove above, where one ensures that all maps there are weak* continuous. \end{proof} \begin{proposition} For any dual operator $*$-algebra $M,$ there is a Hilbert space $H$ (which may be taken to be $K \oplus K$ if $M \subset B(K)$ as a dual operator algebra completely isometrically), and a symmetry (that is, a selfadjoint unitary) $u$ on $H$, and a weak* continuous completely isometric homomorphism $\pi : M \to B(H)$ such that $\pi(a)^* = u \pi(a^\dagger) u$ for $a \in M$. \end{proposition} \begin{proof} This is a tiny modification of the proof in \cite[Proposition 1.12]{BKM}, beginning with a weak* continuous completely isometric homomorphism $\rho : M \to B(K)$ and checking that the $\pi : M \to B(H)$ produced in that proof is weak* continuous. \end{proof} Similarly one has variants of the last two propositions valid for a dual operator algebra with a weak* continuous linear involution. These will look like Theorem \ref{lininvch} but with the representations and isomorphisms also being weak* homeomorphisms. Thanks to this, one will also have dual operator algebra variants of Corollaries \ref{chsym}, \ref{chsym2}, and \ref{chsym3}, basically characterizing symmetric dual operator algebras. In these we are assuming also that $A$ is a dual operator algebra, and in the conclusions the representations and isomorphisms will also be weak* homeomorphisms. For example: \begin{corollary} \label{chsym4} The weak* closed algebra generated by any operator on a Hilbert space is isometrically isomorphic via a weak* homeomorphism to the weak* closed algebra generated by a complex symmetric operator on another Hilbert space. \end{corollary} \begin{proposition} Let $M$ be a dual operator $\ast$-algebra, and let $I$ be a $w^*$-closed $\dagger$-ideal. Then $M/I$ is a dual operator $\ast$-algebra. \end{proposition} \begin{proof} From \cite[Proposition 2.7.11]{BLM}, we know that $M/I$ is a dual operator algebra. As dual operator spaces, $M/I\cong (I_{\perp})^{\ast},$ from which it is easy to see that the involution on $M/I$ is $w^*$-continuous. \end{proof} \begin{lemma}\label{delta} If $A$ is an operator $*$-algebra then $\Delta(A) = A \cap A^*$ (adjoint taken in any containing $C^*$-algebra; see {\rm 2.1.2} in {\rm \cite{BLM}}) is a $C^{\ast}$-algebra and $\Delta(A)^\dagger = \Delta(A)$. \end{lemma} \begin{proof} That $\Delta(A)$ does not depend on the particular containing $C^*$-algebra may be found in e.g.\ 2.1.2 in \cite{BLM}. as is the fact that it is spanned by its selfadjoint (with respect to the usual involution) elements. If $A$ is also an operator $*$-algebra then $\Delta(A)$ is invariant under $\dagger$. Indeed suppose that $B$ is a $C^{\ast}$-cover of $A$ with compatible involution coming from a $*$-automorphism $\sigma$ as usual. If $x = x^* \in \Delta(A)$ then $\sigma(x)$ is also selfadjoint, so is in $\Delta(A)$. This holds by linearity for any $x \in \Delta(A)$. So $\Delta(A)^\dagger = \Delta(A)$. \end{proof} If $M$ is a dual operator algebra then $\Delta(M) = M \cap M^*$, is a $W^*$-algebra (see e.g.\ 2.1.2 in \cite{BLM}). If $M$ is a dual operator $*$-algebra then $\Delta(M)$ is a dual operator $\ast$-algebra, indeed it is a $W^*$-algebra with an extra involution $\dagger$ inherited from $M$. \begin{proposition}\label{jmdg} Suppose that $M$ is a dual operator $*$-algebra. Suppose that $(p_i)_{i\in I}$ is a collection of projections in $M$. Then $(\wedge_{i\in I}\, p_i)^{\dagger}=\wedge_{i\in I}\, p_i^{\dagger}$ and $(\vee_{i\in I}\, p_i)^{\dagger}=\vee_{i\in I}\, p_i^{\dagger}.$ \end{proposition} \begin{proof} By the analysis above the proposition we may assume that $M$ is a $W^*$-algebra with an extra involution $\dagger$, which is weak* continuous and is of the form $x^\dagger = \sigma(x)^*$ for a weak* continuous period 2 $*$-automorphism $\sigma$ of $M$. If $p_i$ and $p_j$ are two projections in $M$, then $p_i\wedge p_j=\lim_n (p_ip_j)^n=\lim_n (p_jp_i)^n.$ By the weak$^{\ast}$-continuity of involution on $M$ we have $$(p_i\wedge p_j)^{\dagger}=\lim_n [(p_jp_i)^n]^{\dagger}=\lim_n (p_i^{\dagger}p_j^{\dagger})^n=p_i^{\dagger}\wedge p_j^{\dagger}.$$ Thus for any finite subset $F$ of $I,$ we have $(\wedge_{i\in F}\, p_i)^{\dagger}=\wedge_{i\in F}\, p_i^{\dagger}.$ Note that the net $(\wedge_{i\in F}\, p_i)_F$ indexed by the directed set of finite subsets $F$ of $I,$ is a decreasing net with limit $\wedge_{i\in I}\, p_i.$ We have $$(\wedge_{i\in I}\, p_i)^{\dagger}=\lim_F \, (\wedge_{i\in F}\, p_i)^{\dagger}=\lim_F\, (\wedge_{i\in F}\, p_i^{\dagger})=\wedge_{i\in I} \, p_i^{\dagger}.$$ by weak* continuity of involution. The statement about suprema of projections follows by taking orthocomplements. \end{proof} \subsection{Involutive M-ideals} An $M$-projection $P$ on a Banach space with involution $X$ is called a {\em $\dagger$-$M$-projection} if $P$ is $\dagger$-preserving. A subspace $Y$ of $X$ is called a {\em $\dagger$-$M$-summand} if $Y$ is the range of a $\dagger$-$M$-projection. Such range is $\dagger$-closed. Indeed, if $y\in Y,$ then $y=P(x)$ for some $x\in X.$ Thus, $y^{\dagger}=P(x)^{\dagger}=P(x^{\dagger})\in Y.$ A subspace $Y$ of $E$ is called an {\em involutive M-ideal} or a {\em $\dagger$-$M$-ideal} in $E$ if $Y^{\perp\perp}$ is a $\dagger$-$M$-summand in $E^{\ast\ast}.$ If $X$ is an operator $\ast$-space, then an $M$-projection is called a {\em complete $\dagger$-$M$-projection} if the amplification $P_n$ is a $\dagger$-$M$-projection on $M_n(X)$ for every $n\in \Ndb.$ Similarly, we could define {\em complete $\dagger$-$M$-summand}, {\em complete $\dagger$-$M$-ideal}, {\em left $\dagger$-$M$-projection}, {\em right $\dagger$-$M$-summand} and {\em right $\dagger$-$M$-ideal}. \begin{proposition}\label{dmsmi} Let $X$ be an operator $\ast$-space. \begin{itemize} \item[(1)] A linear idempotent $\dagger$-linear map $P: X\to X.$ $P$ is a left $\dagger$-$M$-projection if and only if it is a right $\dagger$-$M$-projection, and these imply $P$ is a complete $\dagger$-$M$-projection. \item[(2)] A subspace $Y$ of $X$ is a complete $\dagger$-$M$-summand if and only if it is a left $\dagger$-$M$-summand if and only if it is a right $\dagger$-$M$-summand. \item[(3)] A subspace $Y$ of $X$ is a complete $\dagger$-$M$-ideal if and only if it is a left $M$-$\dagger$-ideal if and only if it is a right $\dagger$-$M$-ideal. \end{itemize} \end{proposition} \begin{proof} (1) \ If $P$ is a left $\dagger$-$M$-projection, then the map \[\sigma_p(x)=\begin{pmatrix} P(x) \\ x-P(x) \end{pmatrix}\] is a completely isometry from $X$ to $C_2(X).$ Also, \begin{align*} \Vert x^{\dagger}\Vert&=\Vert \sigma_P(x^{\dagger})\Vert=\biggl\Vert\begin{pmatrix} P(x^{\dagger}) \\ x^{\dagger}-P(x^{\dagger}) \end{pmatrix}\biggr\Vert= \biggl\Vert \begin{pmatrix} P(x^{\dagger})&0 \\ x^{\dagger}-P(x^{\dagger})&0 \end{pmatrix}\biggr\Vert\\ &=\biggl\Vert \begin{pmatrix} P(x)^{\dagger})&0 \\ x^{\dagger}-P(x)^{\dagger}&0 \end{pmatrix} \biggr\Vert=\biggl\Vert \begin{pmatrix} P(x)&x-P(x) \\ 0&0 \end{pmatrix}^{\dagger} \biggr\Vert\\ &=\biggl\Vert \begin{pmatrix} P(x)&x-P(x) \\ 0&0 \end{pmatrix}\biggr\Vert =\biggl\Vert \begin{pmatrix} P(x), x-P(x)\end{pmatrix}\biggr\Vert=\Vert x\Vert. \end{align*} One can easily generalize this to matrices, so that $P$ is a right $\dagger$-$M$-projection. Similarly, if $P$ is is a right $\dagger$-$M$-projection then $P$ is a left $\dagger$-$M$-projection. By Proposition 4.8.4 (1) in \cite{BLM}, we know that $P$ is a complete $\dagger$-$M$-projection. (2) \ It follows from (1) and \cite[Proposition 4.8.4 (2)]{BLM}. Now (3) is also clear. \end{proof} \begin{theorem} Let $A$ be an approximately unital operator $\ast$-algebra. \begin{itemize} \item[(i)] The right $\dagger$-$M$-ideals are the $\dagger$-$M$-ideals in $A$, which are also the complete $\dagger$-$M$-ideals. These are exactly the approximately unital closed $\dagger$-ideals in $A.$ \item[(ii)] The right $\dagger$-$M$-summands are the $\dagger$-$M$-summands in $A,$ which are also the complete $\dagger$-$M$-summands. These are exactly the principal ideals $Ae$ for a $\dagger$-selfadjoint central projection $e\in M(A).$ \end{itemize} \end{theorem} \begin{proof} (ii) \ By Proposition \ref{dmsmi} (2), the right $\dagger$-$M$-summands are exactly the complete $\dagger$-$M$-summands. Moreover, by \cite[Theorem 4.8.5 (3)]{BLM}, the $M$-summands in $A$ are exactly the complete $M$-summands. If $D$ is a $\dagger$-$M$-summand, then $D$ is a complete $\dagger$-$M$-summand and there exists a central projection $e\in M(A)$ such that $D=eA.$ Then $D^{\perp\perp}=eA^{\ast\ast}$ and $e$ is an identity for $D^{\perp\perp}$. Also, $e^{\dagger}$ serves as an identity in $D^{\perp\perp}$, so that $e=e^{\dagger}.$ (i) \ By a routine argument, the results follow as in \cite[Theorem 4.8.5 (1)]{BLM} and Proposition \ref{dmsmi} (3). \end{proof} \section{Involutive hereditary subalgebras, ideals, and $\dagger$-projections} \subsection{Involutive hereditary subalgebras} Throughout this section $A$ is an operator $\ast$-algebra (possibly not approximately unital). Then $A^{\ast\ast}$ is an operator $\ast$-algebra. We recall from the end of Section \ref{inv} the definition of an {\em open} (or $A$-open) and {\em closed} projection in $A^{**},$ {\em hereditary subalgebras} of $A$ (or HSA's), and {\em support projections} of HSA's (which are the same as $A$-open projections). If $p$ is $A$-open, then $p^{\dagger}$ is also $A$-open. Indeed, if $x_t \in A$ with $x_t=px_t=x_tp=px_tp\to p\, \, {\rm weak^*},$ then $x_t^{\dagger}=p^{\dagger}x_t^{\dagger}=x_t^{\dagger}p^{\dagger}=p^{\dagger}x_t^{\dagger}p^{\dagger}\to p^{\dagger}\, \, {\rm weak^*},$ which means $p^{\dagger}$ is also open. If $p$ is $\dagger$-selfadjoint and open, then we say $p$ is {\em $\dagger$-open} in $A^{**}$. This happens if and only if there exists a $\dagger$-selfadjoint net $(x_t)$ in $A$ with $$x_t=px_t=x_t p=p x_t p\to p \,\, {\rm weak^*}.$$ To see this replace $(x_t)$ in the last paragraphs by $((x_t+x_t^\dagger)/2)$. If also $A$ is approximately unital then we say that $p^\perp = 1-p$ is $\dagger$-closed. If $p$ is $\dagger$-open in $A^{**}$ then clearly $$D=pA^{\ast\ast}p \cap A=\{a\in A: a=ap=pa=pap\}$$ is a closed $\dagger$-subalgebra of $A.$ We call such a $\dagger$-subalgebra $D$ is an {\em involutive hereditary subalgebra} or a {\em $\dagger$-hereditary subalgebra} of $A$ (or,\ $\dagger$-HSA). HSA's are inner ideals: that is $DAD \subset D$. In the following statements, we often omit the proof details where are similar to usual operator algebras case (see e.g. \cite{BLM, BHN,BRI,BRII}) \begin{proposition}\label{dhsa1} A subalgebra $D$ of an operator $*$-algebra $A$ is a HSA and $D^{\dagger}\subset D$ if and only if $D$ is a $\dagger$-HSA. \end{proposition} \begin{proof} One direction is trivial. Conversely, if $D$ is a HSA, then $D=pA^{\ast\ast}p\cap A,$ for some open projection $p\in A^{\ast\ast}.$ Here, $p\in D^{\perp\perp}$ and $p$ is an identity for $D^{\perp\perp}.$ If also $D$ is $\dagger$-selfadjoint, then $p^{\dagger}\in D^{\perp\perp}$ also serves as identity. By uniqueness of identity for $D^{\perp\perp},$ then $p=p^{\dagger}.$ \end{proof} \begin{proposition}\label{dhsainner} A subspace of an operator $\ast$-algebra $A$ is a $\dagger$-HSA if and only if it is an approximately unital $\dagger$-selfadjoint inner ideal. \end{proposition} \begin{proof} If $J$ is a $\dagger$-HSA, then $J$ is an approximately unital $\dagger$-selfadjoint inner ideal. If $J$ is an approximately unital $\dagger$-selfadjoint inner ideal, then by Proposition \ref{dhsa1} $J$ is a HSA and $\dagger$-selfadjoint which means that J is a $\dagger$-HSA. \end{proof} {\bf Remark.} If $J$ is an approximately unital ideal or inner ideal of operator $\ast$-algebra, we cannot necessarily expect $J$ to be $\dagger$-selfadjoint. For example, let $A(\Ddb)$ be the disk algebra and $$A_i(\Ddb)=\{f: f\in A(\Ddb), f(i)=0\}.$$ Then $A_i(\Ddb)$ is an approximately unital ideal but obviously it is not $\dagger$-selfadjoint. \medskip The following is another characterization of $\dagger$-HSA's. \begin{corollary}\label{dghsa} Let $A$ be an operator $\ast$-algebra and suppose that $(e_t)$ is a $\dagger$-selfadjoint net in $\ball(A)$ such that $e_te_s\to e_s$ with $t.$ Then $$\{x\in A: xe_t\to x, e_t x\to x\}$$ is a $\dagger$-HSA of $A.$ Conversely, every $\dagger$-HSA of $A$ arises in this way. \end{corollary} \begin{proof} Let $J=\{x\in A: xe_t\to x, e_t x\to x\}.$ Then it is easy to see that $J$ is an inner ideal and $J^{\dagger}\subset J.$ By Proposition \ref{dhsainner}, $J$ is a $\dagger$-HSA. Conversely, if $D$ is a $\dagger$-HSA and $(e_t)$ is a $\dagger$-selfadjoint cai for $D,$ then $$D=pA^{**}p\cap A=\{x\in A: xe_t\to x, e_t x\to x\},$$ where $p$ is the weak* limit of $(e_t).$ \end{proof} Closed right ideals $J$ of an operator $\ast$-algebra $A$ possessing a $\dagger$-selfadjoint left cai will be called {\em r-$\dagger$-ideals}. Similarly, closed left ideals $J$ of an operator $\ast$-algebra $A$ possessing a $\dagger$-selfadjoint right cai will be called l-$\dagger$-{\em ideals}. Note that there is a bijective correspondence between r-$\dagger$-ideals and l-$\dagger$-ideals, namely $J\to J^{\dagger}.$ For $C^*$-algebras r-$\dagger$-ideals are precisely the closed right ideals, and there is an obvious bijective correspondence between r-$\dagger$-ideals and l-$\dagger$-ideals, namely $J\to J^*$. \begin{theorem}\label{dgrlhsa} Suppose that $A$ is an operator $\ast$-algebra (possibly not approximately unital), and $p$ is a $\dagger$-projection in $A^{\ast\ast}.$ Then the following are equivalent: \begin{itemize} \item[(i)] $p$ is $\dagger$-open in $A^{\ast\ast}.$ \item[(ii)] $p$ is the left support projection of an r-$\dagger$-ideal of $A.$ \item[(iii)] $p$ is the right support projection of an l-$\dagger$-ideal of $A.$ \item[(iv)] $p$ is the support projection of a $\dagger$-hereditary algebra of $A.$ \end{itemize} \end{theorem} \begin{proof} The equivalence of (i) and (iv) is just the definition of being $\dagger$-open in $A^{\ast\ast}.$ Suppose (i). If $p$ is $\dagger$-open then $p$ is the support projection for some $\dagger$-HSA $D.$ Let $(e_t)$ be a $\dagger$-selfadjoint cai for $D,$ then $p=w^{\ast}$-$\lim_{t}e_t.$ Let $$J=\{x\in A: e_tx\to x\},$$ then $J$ is a right ideal of $A$ with $\dagger$-selfadjoint left cai $(e_t)$ and $p$ is the left support projection of $J.$ Suppose (ii). If $p$ is the left support projection of an r-$\dagger$-ideal $J$ of $A$ with $\dagger$-selfadjoint left cai $(e_t),$ then $J=pA^{\ast\ast}\cap A.$ Therefore $J^{\dagger}=A^{\ast\ast}p\cap A,$ which is an l-$\dagger$-ideal and $p$ is the right support projection of $J^{\dagger}.$ Suppose (iii). If $p$ is the right support projection of an l-$\dagger$-ideal of $A$ with $\dagger$-selfadjoint right cai $(e_t),$ then $p=$weak*-$\lim_t e_t=p^{\dagger},$ which means that $p$ is $\dagger$-open. Similarly we can get the equivalence between (i) and (iii). \end{proof} If $J$ is an operator $\ast$-algebra with an $\dagger$-selfadjoint left cai $(e_t),$ then we set $$\Ll(J)=\{a\in J: ae_t\to a\}.$$ \begin{corollary} A subalgebra of an operator $\ast$-algebra $A$ is $\dagger$-hereditary if and only if it equals $\Ll(J)$ for an r-$\dagger$-ideal $J$. Moreover the correspondence $J\mapsto \Ll(J)$ is a bijection from the set of r-$\dagger$-ideals of $A$ onto the set of $\dagger$-HSA's of $A.$ The inverse of this bijection is the map $D\to DA.$ Similar results hold for the l-$\dagger$-ideals of $A.$ \end{corollary} \begin{proof} If $D$ is a $\dagger$-HSA, then by Corollary \ref{dghsa}, we have $$D=\{x\in A: xe_t\to x, e_tx\to x\},$$ where $(e_t)$ is a $\dagger$-selfadjoint cai for $D.$ Set $J=\{x\in A: e_tx\to x\},$ then $J$ is an r-$\dagger$-ideal with $D=\Ll(J).$ Conversely, if $J$ is an r-$\dagger$-ideal and $(e_t)$ is a $\dagger$-selfadjoint left cai for $J,$ then $$D=\{x\in A: xe_t \to x, e_tx \to x\}$$ is a $\dagger$-HSA by Corollary \ref{dghsa}, and $D=\Ll(J).$ The remainder is as in \cite[Corollary 2.7]{BHN}. \end{proof} As in the operator algebra case \cite[Corollary 2.8]{BHN}, if $D$ is a $\dagger$-hereditary subalgebra of an operator $\ast$-algebra $A,$ and if $J=DA$ and $K=AD,$ then $JK=J\cap K=D.$ Also as in the operator algebra case \cite[Theorem 2.10]{BHN}, any $\dagger$-linear functional on a HSA $D$ of an approximately unital operator $\ast$-algebra $A$ has a unique $\dagger$-linear Hahn-Banach extension to $A$. This is because if $\varphi$ is any Hahn-Banach extension to $A$, then $\overline{\varphi(x^\dagger)}$ is another, so these must be equal by \cite[Theorem 2.10]{BHN}. \begin{proposition} Let $D$ be an approximately unital $\dagger$-subalgebra of an approximately unital operator $\ast$-algebra $A.$ The following are equivalent: \begin{itemize} \item[(i)] $D$ is a $\dagger$-hereditary subalgebra of $A.$ \item[(ii)] Every completely contractive unital $\dagger$-linear map from $D^{\ast\ast}$ into a unital operator $\ast$-algebra $B$, has a unique completely contractive unital $\dagger$-extension from $A^{\ast\ast}$ into $B.$ \item[(iii)] Every completely contractive $\dagger$-linear map $T$ from $D$ into a unital weak* closed operator $\ast$-algebra $B$ such that $T(e_t)\to 1_B$ weak* for some cai $(e_t)$ for $D$ has a unique completely contractive weakly $\dagger$-extension $\tilde{T}$ from $A$ into $B$ with $\tilde{T}(f_s)\to 1_B$ weak* for some(or all) cai $(f_s)$ for $A.$ \end{itemize} \end{proposition} \begin{proof} Let $e$ be the identity of $D^{\ast\ast}.$ Obviously, $e$ is $\dagger$-selfadjoint. If (iii) holds, then the inclusion from $D$ to $D^{\perp\perp}$ extends to a unital complete $\dagger$-contraction $T: A\to D^{\ast\ast}\subset eA^{\ast\ast}e.$ The map $x\to exe$ on $A^{\ast\ast}$ is also a completely contractive unital $\dagger$-extension of the inclusion map $D^{\ast\ast}\to eD^{\ast\ast}e.$ It follows from the hypothesis that these maps coincide, and so $eA^{\ast\ast}e=D^{\ast\ast},$ which implies that $D$ is a $\dagger$-HSA. The rest is left as an exercise to the reader, being very similar to the proof of \cite[Proposition 2.11]{BHN}. \end{proof} \subsection{Support projections and $\dagger$-HSA's} \begin{lemma}\label{sdhsa} If $(J_i)$ is a family of r-$\dagger$-ideals in an operator $\ast$-algebra $A,$ with matching family of $\dagger$-HSA's $(D_i),$ and if $J$ is the norm closure of $\sum_i \, J_i$ then the $\dagger$-HSA matching $J$ is the $\dagger$-HSA $D$ generated by the $(D_i)$. \end{lemma} \begin{proof} This follows from the matching operator algebra result, since e.g.\ any $\dagger$-HSA is a HSA. \end{proof} In the next result, e.g.\ $\overline{zA}$ denotes the {\em norm closure} of $zA$. \begin{proposition}\label{csrdihsa} Let $A$ be an operator $\ast$-algebra (not necessarily with an identity or approximate identity). Suppose that $(x_k)$ is a sequence of $\dagger$-selfadjoint elements in $\mathfrak{F}_A,$ and $\alpha_k\in (0,1]$ add to $1.$ Then the closure of the sum of the r-$\dagger$-ideals $\overline{x_k A},$ is the r-$\dagger$-ideal $\overline{zA},$ where $z=\sum_{k=1}^{\infty} \alpha_kx_k\in \mathfrak{F}_A.$ Similarly, the $\dagger$-HSA generated by all the $\overline{x_kAx_k}$ equals $\overline{zAz}.$ \end{proposition} \begin{proof} By the matching operator algebra result, the right ideal $\overline{zA}$ is the closure of the sum of the right ideals $\overline{x_kA}.$ If $z\in \mathfrak{F}_A$ is $\dagger$-selfadjoint then $\overline{zA}$ is an r-$\dagger$-ideal. \end{proof} If $S\subset A,$ define $S_{\dagger}$ to be the set of $\dagger$-selfadjoint elements in $S.$ \begin{lemma}\label{spjoin} Let $A$ be an operator $\ast$-algebra, a subalgebra of a $C^{\ast}$-algebra $B.$ \begin{itemize} \item[(i)] The support projection of a $\dagger$-HSA $D$ in $A$ equals $\vee_{a\in ({\mathfrak F}_D)_{\dagger}}s(a)$ (which equals $\vee_{a\in ({\mathfrak r}_D)_{\dagger}}s(a)$). \item[(ii)] The support projection of an r-$\dagger$-ideal $J$ in $A$ equals $\vee_{a\in ({\mathfrak F}_J)_{\dagger}}s(a)$ (which equals $\vee_{a\in ({\mathfrak r}_J)_{\dagger}}s(a)$). \end{itemize} \end{lemma} \begin{proof} (i) \ Suppose $p$ is the support projection of $D,$ then $p=\vee_{b\in {\mathfrak F}_D}s(b)=\vee_{b\in {\mathfrak r}_D}s(b)$ by the operator algebra case we are generalizing. Thus, $$p\geq \vee_{a\in ({\mathfrak r}_D)_{\dagger}} \, s(a)\geq \vee_{a\in ({\mathfrak F}_D)_{\dagger}} \, s(a).$$ For any $b\in {\mathfrak F}_D,$ we have $b^{\dagger}\in {\mathfrak F}_D$ and $s(b)\vee s(b^{\dagger})=s(\frac{b+b^{\dagger}}{2})$, by Proposition \ref{sdaggi}. Hence, $$p=\vee_{b\in {\mathfrak F}_D}s(b)\leq \vee_{a\in ({\mathfrak F}_D)_{\dagger}}s(a).$$ Therefore, $p\leq \vee_{a\in ({\mathfrak F}_D)_{\dagger}}s(a)\leq \vee_{a\in ({\mathfrak r}_D)_{\dagger}}s(a).$ (ii) \ This is similar. \end{proof} \begin{lemma}\label{mindgshsa} For any operator $\ast$-algebra $A,$ if $E\subset ({\mathfrak r}_A)_{\dagger},$ then the smallest $\dagger$-hereditary subalgebra of $A$ containing $E$ is $pA^{\ast\ast}p\cap A,$ where $p=\vee_{x\in E}\, s(x).$ \end{lemma} \begin{proof} By the operator algebra case we are generalizing, $pA^{\ast\ast}p\cap A$ is the smallest $\dagger$-hereditary subalgebra of $A$ containing $E$. Conversely, if $D$ is a $\dagger$-HSA of $A$ containing $E$ then $D^{\perp\perp}$ contains $p$ by a routine argument, so $pA^{\perp\perp}p\subset D^{\perp\perp}$ and $pA^{\perp\perp}p\cap A\subset D^{\perp\perp}\cap A=D.$ \end{proof} \begin{corollary} For any operator $\ast$-algebra $A,$ suppose that a convex set $E\subset {\mathfrak r}_A$ and $E^{\dagger}\subset E.$ Then the smallest hereditary subalgebra of $A$ containing $E$ is $pA^{\ast\ast}p\cap A,$ where $p=\vee_{x\in E_{\dagger}} s(x).$ Indeed, this is the smallest $\dagger$-HSA of $A$ containing $E.$ \end{corollary} \begin{proof} By the lemma the smallest HSA (and smallest $\dagger$-HSA) containing $E$ is $pA^{\ast\ast}p\cap A,$ where $p=\vee_{a\in E}\, s(a).$ For any $a\in E$, $\frac{a+a^{\dagger}}{2}\in E$ by convexity of $E$. Notice that $s(\frac{a+a^{\dagger}}{2})\leq p$ and $s(\frac{a+a^{\dagger}}{2})\geq s(a).$ So $p=\vee_{x\in E_{\dagger}}s(x)$. \end{proof} Again in the next several results in this section, `overline' denotes the {\em norm closure}. \begin{theorem} If $A$ is an operator $\ast$-algebra then $\dagger$-HSA's (resp.\ r-$\dagger$-ideals) in $A$ are precisely the sets of form $\overline{EAE}$ (resp.\ $\overline{EA}$) for some $E\subset({\mathfrak r}_A)_{\dagger}.$ The latter set is the smallest $\dagger$-HSA (resp.\ r-$\dagger$-ideal) of $A$ containing $E.$ \end{theorem} \begin{proof} If $D$ is a $\dagger$-HSA (resp.\ r-$\dagger$-ideal) and taking $E$ to be a $\dagger$-selfadjoint cai for the $\dagger$-HSA $D$ (resp.\ a $\dagger$-selfadjoint left cai for the r-$\dagger$-ideal), then the results follows immediately. Conversely for any $x\in ({\mathfrak r}_A)_{\dagger},$ we have $x(1+x)^{-1}\in (\frac{1}{2}{\mathfrak F}_A)_{\dagger}$ as we said in Corollary \ref{fbab}. Then as in \cite[Theorem 3.18]{BWj} we may assume that $E\subset(\frac{1}{2}{\mathfrak F}_A)_{\dagger}.$ Note that $D=\overline{EAE}$ is the smallest HSA containing $E$ by \cite[Theorem 3.18]{BWj} and $D$ is $\dagger$-selfadjoint, so that $D$ is the smallest $\dagger$-HSA containing $E.$ Similarly, $\overline{EA}$ is the smallest right ideal with a $\dagger$-selfadjoint left contractive identity of $A$ containing $E$. Moreover, for any finite subset $F\subset E$ if $a_F$ is the average of the elements in $F,$ then $(a_F^{1/n})$ will serve as a $\dagger$-selfadjoint left cai for $\overline{EA}.$ \end{proof} In particular, the largest $\dagger$-HSA in an operator $\ast$-algebra $A$ is the largest HSA in $A$, and the largest approximately unital subalgebra in $A$ (see \cite[Section 4]{BRII}), namely $A_H = \overline{{\mathfrak r}_A A {\mathfrak r}_A} = \overline{({\mathfrak r}_A)_{\dagger} A ({\mathfrak r}_A)_{\dagger}}$. The latter equality follows because $A_H$ has a cai in ${\mathfrak r}_A$, hence has a cai in $({\mathfrak r}_A)_{\dagger}$. \begin{theorem}\label{chdghsar} Let $A$ be an operator $\ast$-algebra (not necessarily with an identity or approximate identity.) The $\dagger$-HSA's (resp.\ r-$\dagger$-ideals) in $A$ are precisely the closures of unions of an increasing net of $\dagger$-HSA's (resp.\ r-$\dagger$-ideals) of the form $\overline{xAx}$ (resp.\ $\overline{xA}$) for $x\in ({\mathfrak r}_A)_{\dagger}.$ \end{theorem} \begin{proof} Suppose that $D$ is a $\dagger$-HSA (resp.\ an r-$\dagger$-ideal). The set of $\dagger$-HSA's (resp.\ r-$\dagger$-ideals) $\overline{a_FAa_F}$ (resp.\ $\overline{a_F A}$) as in the last proof, indexed by finite subsets $F$ of $({\mathfrak F}_D)_{\dagger},$ is an increasing net. Lemma \ref{spjoin} can be used to show, as in \cite{BWj}, that the closure of the union of these $\dagger$-HSA's (resp.\ r-$\dagger$-ideals) is $D.$ \end{proof} As in the theory we are following, it follows that $\dagger$-open projections are just the sup's of a collection (an increasing net if desired) of $\dagger$-selfadjoint support projections $s(x)$ for $\dagger$-selfadjoint $x \in {\mathfrak r}_A$. \begin{theorem}\label{strdghsa} Let $A$ be any operator $\ast$-algebra (not necessarily with an identity or approximate identity). Every separable $\dagger$-HSA or $\dagger$-HSA with a countable cai (resp.\ separable r-$\dagger$-ideal or r-$\dagger$-ideal with a countable cai) is equal to $\overline{xAx}$ (resp.\, $\overline{xA}$) for some $x\in ({\mathfrak F}_A)_{\dagger}.$ \end{theorem} \begin{proof} If $D$ is a $\dagger$-HSA with a countable cai, then $D$ has a countable $\dagger$-selfadjoint cai $(e_n)$ in $\frac{1}{2}{\mathfrak F}_D.$ Also, $D$ is generated by the $\dagger$-HSA's $\overline{e_n A e_n}$ so $D=\overline{xAx},$ where $x=\sum_{n=1}^{\infty} \frac{e_n}{2^n}.$ For the separable case, note that any separable approximately unital operator $\ast$-algebra has a countable cai. For r-$\dagger$-ideals, the result follows from the same argument. \end{proof} \begin{corollary} If $A$ is a separable operator $\ast$-algebra, then the $\dagger$-open projections in $A^{\ast\ast}$ are precisely the $s(x)$ for $x\in({\mathfrak r}_A)_{\dagger}.$ \end{corollary} \begin{proof} If $A$ is separable, then so is any $\dagger$-HSA. So the result follows from Theorem \ref{strdghsa}. \end{proof} \begin{corollary} If $A$ is a separable operator $\ast$-algebra with cai, then there exists an $x\in({\mathfrak F}_A)_{\dagger}$ with $A=\overline{xA}=\overline{Ax}=\overline{xAx}.$ \end{corollary} \subsection{Involutive compact projections} Throughout this section, $A$ is an operator $\ast$-algebra. We will say that a projection $q\in A^{\ast\ast}$ is {\em compact relative to} $A$ if it is closed and $q=qx$ for some $x\in \ball(A).$ Furthermore, if $q$ is $\dagger$-selfadjoint, we say that such $q$ is an {\em involutive compact projection}, or is {\em $\dagger$-compact} in $A^{\ast\ast}.$ \begin{proposition} A $\dagger$-projection $q$ is compact if only if there exists a $\dagger$-selfadjoint element $a\in {\ball(A)}$ such that $q=qa.$ \end{proposition} \begin{proof} One direction is trivial. Conversely if $q$ is compact, then there exists $a\in\ball(A)$ such that $q=qa.$ It is easy to argue from elementary operator theory that we have $aq=q.$ Thus, $q=q(\frac{a+a^{\dagger}}{2}).$ \end{proof} \begin{theorem}\label{eqrdcp} Let $A$ be an approximately unital operator $\ast$-algebra. If $q$ is a projection in $A^{\ast\ast}$ then the following are equivalent: \begin{itemize} \item[(i)] $q$ is a $\dagger$-closed projection in $(A^1)^{\ast\ast},$ \item[(ii)] $q$ is $\dagger$-compact in $A^{\ast\ast}$ \item[(iii)] $q$ is closed in $A^{**}$ and there exists a $\dagger$-selfadjoint element $x\in\frac{1}{2}{\mathfrak F}_A$ such that $q=qx.$ \end{itemize} \end{theorem} \begin{proof} This follows from a variant of the proof of \cite[Theorem 2.2]{BNII}: one just needs to go carefully through the proof noting that all elements may be chosen to be $\dagger$-selfadjoint. \end{proof} \begin{corollary} Let $A$ be an approximately unital operator $\ast$-algebra. Then the infimum of any family of $\dagger$-compact projections in $A^{\ast\ast}$ is a $\dagger$-compact projection in $A^{\ast\ast}.$ Also, the supremum of two commuting $\dagger$-compact projections in $A^{\ast\ast}$ is a $\dagger$-compact projection in $A^{\ast\ast}.$ \end{corollary} \begin{proof} Note that the infimum and supremum of $\dagger$-projections are still $\dagger$-projections. Then the results follow immediately from \cite[Corollary 2.3]{BNII}. \end{proof} \begin{corollary} Let $A$ be an approximately unital operator $\ast$-algebra, with an approximately unital closed $\dagger$-subalgebra $D.$ A projection $q\in D^{\perp\perp}$ is $\dagger$-compact in $D^{\ast\ast}$ if and only if $q$ is $\dagger$-compact in $A^{\ast\ast}.$ \end{corollary} \begin{corollary} Let $A$ be an approximately unital operator $\ast$-algebra. If a $\dagger$-projection $q$ in $A^{\ast\ast}$ is dominated by an open projection $p$ in $A^{\ast\ast},$ then $q$ is $\dagger$-compact in $pA^{\ast\ast}p.$ \end{corollary} In much of what follows we use the peak projections $u(a)$ defined and studied in e.g.\ \cite{BNII,BRII}. These may be defined to be projections $q$ in $A^{**}$ which are the weak* limits of $a^n$ for some $a \in {\rm Ball}(A)$, in the case such weak* limit exists. We will not take the time to review the properties of $u(a)$ here. We will however several times below use silently the following fact: \begin{lemma} \label{ulemm} If $a \in {\rm Ball}(A)$ for an operator $\ast$-algebra $A$, and if $u(a)$ is a peak projection, with $a^n \to u(a)$ weak*, then $u((a + a^\dagger)/2) = u(a) \wedge u(a)^\dagger$ in $A^{**}$ and this is a peak projection. Indeed $((a + a^\dagger)/2)^n \to u((a + a^\dagger)/2)$ weak*. \end{lemma} \begin{proof} Clearly $(a^\dagger)^n \to u(a)^\dagger$ weak*, so that $u(a^\dagger) = u(a)^\dagger$ is a peak projection. Then $u((a + a^\dagger)/2) = u(a) \wedge u(a)^\dagger$ by \cite[Proposition 1.1]{BNII}, and since this is a projection it is by \cite[Section 3]{BNII} a peak projection, is $\dagger$-selfadjoint, and $((a + a^\dagger)/2)^n \to u((a + a^\dagger)/2)$ weak*. \end{proof} The following is the involutive variant of the version of the Urysohn lemma for approximately unital operator $\ast$-algebras in \cite[Theorem 2.6]{BNII}. \begin{theorem} \label{ulosacai} Let $A$ be an approximately unital operator $\ast$-algebra. If a $\dagger$-compact projection $q$ in $A^{**}$ is dominated by a $\dagger$-open projection $p$ in $A^{\ast\ast}$, then there exists $b\in(\frac{1}{2}{\mathfrak F}_A)_{\dagger}$ with $q=qb, b=pb.$ Moreover, $q\leq u(b)\leq s(b)\leq p,$ and $b$ may also be chosen to be `nearly positive' in the sense of the introduction to {\rm \cite{BRord}}: we can make it as close in norm as we like to an actual positive element. \end{theorem} \begin{proof} If $q\leq p$ as stated, then by the last corollary we know $q$ is $\dagger$-compact in $D^{\ast\ast}=pA^{\ast\ast}p,$ where $D$ is a $\dagger$-HSA supported by $p.$ By Theorem \ref{eqrdcp}, there exists a $\dagger$-selfadjoint $b\in \frac{1}{2}{\mathfrak F}_D$ such that $q=qb$ and $b=bp.$ The rest follows as in \cite[Theorem 2.6]{BNII}. \end{proof} \begin{theorem} Suppose that $A$ is an operator $\ast$-algebra (not necessarily approximately unital), and that $q\in A^{\ast\ast}$ is a projection. The following are equivalent: \begin{itemize} \item[(1)] $q$ is $\dagger$-compact with respect to $A.$ \item[(2)] $q$ is $\dagger$-closed with respect to $A^1$ and there exists $a\in \ball(A)_{\dagger}$ with $aq=qa=q.$ \item[(3)] $q$ is a decreasing weak$^{\ast}$ limit of $u(a)$ for $\dagger$-selfadjoint element $a\in \ball(A).$ \end{itemize} \end{theorem} \begin{proof} (2) $\Rightarrow$ (3) \ Given (2) we certainly have $q$ compact with respect to $A$ by \cite[Theorem 6.2]{BRII}. By \cite[Theorem 3.4]{BNII}, $q=\lim_t u(z_t),$ where $z_t\in \ball(A)$ and $u(z_t)$ is decreasing. We have $q=q^\dagger = \lim_ t u(z_t^{\dagger}).$ Moreover, $u(z_t)\wedge u(z_t^{\dagger})=u(\frac{z_t+z_t^{\dagger}}{2}).$ Hence, $q$ is a decreasing weak* limit of $u(\frac{z_t+z_t^{\dagger}}{2})$ since the involution preserves order. The rest follows from \cite[Theorem 6.2]{BRII}. \end{proof} \begin{corollary} Let $A$ be a (not necessarily approximately unital) operator $\ast$-algebra. If $q$ is $\dagger$-compact then $q$ is a weak* limit of a net of $\dagger$-selfadjoint elements $(a_t)$ in $\ball(A)$ with $a_tq=q$ for all $t.$ \end{corollary} \subsection{Involutive peak projections} Let $A$ be an operator $\ast$-algebra. A $\dagger$-projection $q\in A^{\ast\ast}$ is called an {\em involutive peak projection} or a {\em $\dagger$-peak projection} if it is a peak projection. \begin{proposition}\label{scpd} Suppose $A$ is a separable operator $\ast$-algebra (not necessarily approximately unital), then the $\dagger$-compact projections in $A^{\ast\ast}$ are precisely the peak projections $u(a),$ for some $\dagger$-selfadjoint $a\in \ball(A).$ \end{proposition} \begin{proof} If $A$ is separable then a projection in $A^{\ast\ast}$ is compact if and only if $q=u(a),$ for some $a\in \ball(A)$ (see \cite[Proposition 6.4]{BRII}). If $q$ is $\dagger$-selfadjoint, then $$q=u(a^{\dagger}) = u(a)\wedge u(a^{\dagger}) = u((a^{\dagger}+a)/2),$$ using e.g.\ Lemma \ref{ulemm}. \end{proof} \begin{proposition}\label{dpeak} If $a\in\frac{1}{2}{\mathfrak F}_A$ with $a^{\dagger}=a,$ then $u(a)$ is a $\dagger$-peak projection and it is a peak for $a.$ \end{proposition} \begin{proof} Since $u(a)=\lim a^n$ weak* in this case, we see that $u(a)$ is $\dagger$-selfadjoint. From \cite[Lemma 3.1, Corollary 3.3]{BNII}, we know that $u(a)$ is a peak projection and is a peak for $a.$ \end{proof} \begin{theorem} If $A$ is an approximately unital operator $\ast$-algebra, then \begin{itemize} \item[(i)] A projection $q\in A^{\ast\ast}$ is $\dagger$-compact if only if it is a decreasing limit of $\dagger$-peak projections. \item[(ii)] If $A$ is a separable approximately unital operator $\ast$-algebra, then the $\dagger$-compact projections in $A^{\ast\ast}$ are precisely the $\dagger$-peak projections. \item[(iii)] A projection in $A^{\ast\ast}$ is a $\dagger$-peak projection in $A^{\ast\ast}$ if and only if it is of form $u(a)$ for some $a\in (\frac{1}{2}{\mathfrak F}_A)_{\dagger}.$ \end{itemize} \end{theorem} \begin{proof} (ii) \ Follows from Proposition \ref{scpd} and Proposition \ref{dpeak}. (i) \ One direction is obvious. Conversely, let $q\in A^{\ast\ast}$ be a $\dagger$-compact projection with $q=qx$ for some $\dagger$-selfadjoint element $x\in \ball(A).$ Then $q\leq u(x).$ Now $1-q$ is an increasing limit of $s(x_t)$ for $\dagger$-selfadjoint elements $x_t\in \frac{1}{2}{\mathfrak F}_{A^1}$, by Theorem \ref{chdghsar} and the remark after it. Therefore, $q$ is a decreasing weak* limit of the $q_t=s(x_t)^{\perp}=u(1-x_t).$ Let $z_t=\frac{1-x_t+x}{2},$ then $u(z_t)$ is a projection. Since $q\leq q_t$ and $q\leq u(x),$ then $q\leq u(z_t)$. Note that $z_t$ is $\dagger$-selfadjoint, so $u(z_t)=u(z_t)^{\dagger}.$ Let $a_t=z_tx\in \ball(A),$ then $u(a_t)=u(z_t)$ by the argument in \cite[Lemma 3.1]{BNII}. Thus, $u(a_t)=u(z_t)\searrow q$ as in that proof. Moreover, $u(a_t^{\dagger}) = u(a_t)^{\dagger} \searrow q,$ which implies by an argument above that $u(\frac{a_t+a_t^{\dagger}}{2})\searrow q.$ (iii) \ One direction is trivial. For the other, if $q$ is a $\dagger$-peak projection, then by the operator algebra case there exists $a\in \frac{1}{2}{\mathfrak F}_A$ such that $q=u(a).$ Let $b=(a+a^{\dagger})/2,$ then $q=u(b)$ by e.g.\ Lemma \ref{ulemm}. \end{proof} \begin{corollary} Let $A$ be an operator $\ast$-algebra. The supremum of two commuting $\dagger$-peak projections in $A^{\ast\ast}$ is a $\dagger$-peak projection in $A^{\ast\ast}.$ \end{corollary} \begin{lemma} For any operator $\ast$-algebra $A,$ the $\dagger$-peak projections for $A$ are exactly the weak* limits of $a^n$ for $\dagger$-selfadjoint element $a\in \ball(A)$ if such limit exists. \end{lemma} \begin{proof} If $q$ is a $\dagger$-peak projection, then there exists $a\in \ball(A)$ such that $q=u(a)$ which is also the weak* limit of $a^n.$ Since $q$ is $\dagger$-selfadjoint, by Lemma \ref{ulemm} we have $q=u(a^{\dagger}) =u(\frac{a+a^{\dagger}}{2})$, which is the weak* limit of $((a+a^{\dagger})/2)^n.$ The converse follows from \cite[Lemma 1.3]{BRII}. \end{proof} {\bf Remark.} In the theory of peak projections for operator algebras $A$ which are not necessarily approximately unital in \cite[Section 6]{BRII}, there are two notions of peak projection, one stronger than the other. If $A$ is an operator $\ast$-algebra one would say that a projection is a {\em $\dagger$-${\mathfrak F}$-peak projection} for $A$ if it is $\dagger$-selfadjoint and ${\mathfrak F}$-peak, where the latter means that it equals $u(a)$ for an $a \in \frac{1}{2}{\mathfrak F}_A$. A projection in $A^{\ast\ast}$ is {\em $\dagger$-${\mathfrak F}$-compact} if it is a decreasing limit of $\dagger$-${\mathfrak F}$-peak projections. We recall that $A_H$ was discussed above Theorem \ref{chdghsar}. One may then prove similarly to the development in \cite[Proposition 6.5]{BRII} (with appropriate tweaks in the proofs): \begin{itemize} \item[(i)] A projection $q$ in $A^{\ast\ast}$ is $\dagger$-${\mathfrak F}$-compact if only if $q$ is a $\dagger$-compact projection for $A_H.$ \item[(ii)] A projection in $A^{\ast\ast}$ is a $\dagger$-${\mathfrak F}$-peak projection if and only if it is a $\dagger$-peak projection for $A_H.$ \item[(iii)] If $A$ is separable then every $\dagger$-${\mathfrak F}$-compact projection in $A^{\ast\ast}$ is a $\dagger$-${\mathfrak F}$-peak projection. \end{itemize} \subsection{Some interpolation results} Item (ii) in the following should be compared with Theorem \ref{ulosacai} which gets a slightly better result in the case that $A$ is approximately unital. \begin{theorem}[Noncommutative Urysohn lemma for operator $\ast$-algebras] Let $A$ be a (not necessarily approximately unital) operator $\dagger$-subalgebra of $C^{\ast}$-algebra $B$ with a second involution $\dagger$. Let $q$ be a $\dagger$-compact projection in $A^{\ast\ast}.$ \begin{itemize} \item[(i)] For any $\dagger$-open projection $p\in B^{\ast\ast}$ with $p\geq q,$ and any $\varepsilon>0,$ there exists an $a\in \ball(A)_{\dagger}$ with $aq=q$ and $\Vert a(1-p)\Vert <\varepsilon.$ \item[(ii)] For any $\dagger$-open projection $p\in A^{\ast\ast}$ with $p\geq q,$ there exists a $\dagger$-selfadjoint element $a\in \ball(A)$ with $q=qa$ and $a=pa.$ \end{itemize} \end{theorem} \begin{proof} (i) \ By \cite[Theorem 6.6]{BRII} there exists $b\in \ball(A)$ such that $$bq=q,\,\Vert b(1-p)\Vert <\varepsilon~\mbox{and}~\Vert (1-p)b\Vert <\varepsilon.$$ Then $a=\frac{b+b^{\dagger}}{2}\in \ball(A)_{\dagger}$ does the trick, since $$\Vert (\frac{b+b^{\dagger}}{2})(1-p) \Vert\leq \frac{1}{2}\Vert b(1-p)\Vert+\frac{1}{2}\Vert ((1-p)b)^{\dagger}\Vert<\varepsilon.$$ (ii) \ Apply Theorem \ref{ulosacai} in $A^1$ to obtain a $\dagger$-selfadjoint element $a\in \ball(A^1)$, $p\in A^{\perp\perp}$ and $ap=q.$ Then $a\in A^{\perp\perp}\cap A^1=A.$ \end{proof} The following is an involutive variant of the noncommutative peak interpolation type result in \cite[Theorem 5.1]{BRII}. \begin{theorem} \label{peakthang} Suppose that $A$ is an operator $*$-algebra and that $q$ is a $\dagger$-closed projection in $(A^1)^{**}$. If $b = b^\dagger \in A$ with $b q= qb$, then $b$ achieves its distance to the right ideal $J = \{ a \in A : q a = 0 \}$ (this is a r-$\dagger$-ideal if $1-q \in A^{**}$), and also achieves its distance to $\{ x \in A : x q = qx = 0 \}$ (this is a $\dagger$-HSA if $1-q \in A^{**}$). If further $\Vert b q \Vert \leq 1$, then there exists a $\dagger$-selfadjoint element $g \in {\rm Ball}(A)$ with $g q = q g = b q$. \end{theorem} \begin{proof} Proceed as in the proof of \cite[Theorem 5.1]{BRII}. The algebra $\tilde{D}$ is a $\dagger$-HSA in $A^1$. Thus if $C$ is as in that proof, $C$ is $\dagger$-selfadjoint and $\tilde{D}$ is a $\dagger$-ideal in $C$. Also $I = C \cap A$ and $D = I \cap \tilde{D}$ are $\dagger$-selfadjoint in $C$. Note that if $x \in A$ with $x q = qx = 0$ then $x \in \tilde{D} \cap A \subset C \cap A = I$, so $x \in \tilde{D} \cap A \subset \tilde{D} \cap I = D$. So $D = \{ x \in A : x q = qx = 0 \}$. By the proof we are following, there exists $y \in D \subset J$ such that $$\| b - y \| = \| b - y^\dagger \| = d(b,D) = \| b q \| = d(b,J) \geq \| b - z \|,$$ where $z = (y + y^\dagger)/2$. Set $g = b - z$, then $g$ is $\dagger$-selfadjoint with $g q = q g = b q$ (since $D$ is $\dagger$-selfadjoint), and $\| g \| = \| b q \|$. \end{proof} Theorem 4.10 in \cite{BRord} is the (noninvolutive) operator algebra version of the last result (and \cite[Theorem 5.1]{BRII}), but with the additional feature that the `interpolating element' $g$ in the last result is also in $\frac{1}{2} {\mathfrak F}_A$. Whence after replacing $g$ by $g^{\frac{1}{n}}$, it is `nearly positive' in the sense of the introduction to \cite{BRord}: we can make it as close in norm as we like to an actual positive element. As we have pointed out elsewhere, there seems to be a mistake in Theorem 4.10 (and hence also in 4.12) in \cite{BRord}. It is claimed there (and used at the end of the proof) that $D$ is approximately unital. However this error disappears in what is perhaps the most important case, namely that $q$ is the `perp' of a (open) projection in $A^{**}$. Then $D$ is certainly a HSA in $A$, and is approximately unital. Hence we have (in the involutive case by averaging the element produced by the original proof, with its `dagger'): \begin{theorem} \label{peakth2} Suppose that $A$ is an operator $*$-algebra, $p$ is a $\dagger$-open projection in $A^{**}$, and $b = b^\dagger \in A$ with $bp= p b$ and $\Vert b (1-p) \Vert \leq 1$ (where $1$ is the identity of the unitization of $A$ if $A$ is nonunital). Suppose also that $\| (1-2b) (1-p) \| \leq 1$. Then there exists a $\dagger$-selfadjoint element $g \in \frac{1}{2} {\mathfrak F}_A \subset {\rm Ball}(A)$ with $g (1-p) = (1-p) g = b(1-p)$. Such $g$ may be chosen `nearly positive' in the sense of the introduction to {\rm \cite{BRord}}, indeed it may be chosen to be as close as we like to an actual positive element. \end{theorem} \begin{theorem} \label{peakthang2} {\rm (A noncommutative Tietze theorem)} \ Suppose that $A$ is an operator $*$-algebra (not necessarily approximately unital), and that $p$ is a $\dagger$-open projection in $A^{**}$. Set $q = 1-p \in (A^1)^{**}$. Suppose that $b = b^\dagger \in A$ with $b p= pb$ and $\Vert b q \Vert \leq 1$, and that the numerical range of $bq$ (in $q (A^1)^{**}q$ or $(A^1)^{**}$) is contained in a compact convex set $E$ in the plane satisfying $E = \bar{E}$, where the latter is the reflection of $E$ with respect to the real axis. We also suppose, by fattening it slightly if necessary, that $E \not\subset \Rdb$. Then there exists a $\dagger$-selfadjoint element $g \in {\rm Ball}(A)$ with $g q = q g = b q$, such that the numerical range of $g$ with respect to $A^1$ is contained in $E$. \end{theorem} \begin{proof} By \cite[Theorem 4.12]{BRord}, there exists $a \in {\rm Ball}(A)$ with $aq = q a = b q$, such that the numerical range of $a$ with respect to $A^1$ is contained in $E$. Then $g = (a+ a^\dagger)/2$ is $\dagger$-selfadjoint. Let $B$ be a unital $C^*$-cover of $A^1$ with compatible involution $\sigma(b)^*$ as usual. If $\varphi$ is a state of $B$ then $\varphi \circ \sigma$ is a state too, and so $$\varphi(a^\dagger) = \overline{\varphi(\sigma(a))} \in \bar{E} = E.$$ From this it is clear that $\varphi(g) \in E$. \end{proof} \begin{corollary} Suppose that $A$ is an operator $*$-algebra (not necessarily approximately unital), and that $J$ is an approximately unital closed $\dagger$-ideal in $A$. Suppose that $b = b^\dagger$ is an element in ${\mathfrak F}_{A/J}$ (resp.\ in ${\mathfrak r}_{A/J}$). Then there exists a $\dagger$-selfadjoint element $a$ in ${\mathfrak F}_{A}$ (resp.\ in ${\mathfrak r}_{A}$) with $a + J = b$.\end{corollary} \begin{proof} Indeed the operator $*$-algebra variant of \cite[Proposition 6.1]{BRI} and \cite[Corollary 6.1]{BRI} hold. The ${\mathfrak r}_{A/J}$ lifting follows from the last theorem with $E= [0,K] \times [-K,K]$ and $K = \| b \|$ say. However both results also follow by the usual $(a+ a^\dagger)/2$ trick. \end{proof} The following is the `nearly positive' case of a simple noncommutative peak interpolation result which has implications for the unitization of an operator $\ast$-algebra. \begin{proposition} \label{interpnp} Suppose that $A$ is an approximately unital operator $*$-algebra, and $B$ is a $C^*$-algebra generated by $A$ with compatible involution $\dagger$. If $c = c^\dagger \in B_+$ with $\Vert c \Vert < 1$ then there exists a $\dagger$-selfadjoint $a \in \frac{1}{2} {\mathfrak F}_A$ with $|1 - a|^2 \leq 1-c$. Indeed such $a$ can be chosen to also be nearly positive. \end{proposition} \begin{proof} As in \cite[Proposition 4.9]{BRord}, but using our Theorem \ref{havin} (2), there exists nearly positive $\dagger$-selfadjoint $a \in \frac{1}{2} {\mathfrak F}_A$ with $$c \leq {\rm Re}(a) \leq 2 {\rm Re}(a) - a^* a,$$ and $|1-a|^2 \leq 1-c$. \end{proof} We end with an involutive case of the best noncommutative peak interpolation result (from \cite{Bnpi}), a noncommutative generalization of a famous interpolation result of Bishop. See \cite{Bnpi} for more context and an explanation of the classical variant, and the significance of the noncommutative variant. Unfortunately we cannot prove this result for operator $\ast$-algebras without imposing a further strong condition ($d$ commutes with $b$ and $q^\perp (A^1)^{**} q^\perp \cap A^1$). This is a good example of a complicated result which is not clear in advance whether it has `involutive variants'. In this case it is valid, without the strong condition just mentioned, for involutions of types (3) and (4) at the start of Section {\rm \ref{inv}}. We treat the type (3) case. For an operator algebra $A$, let $\bar{A}$ be as in the Remark after Proposition \ref{ciop}. In this case a conjugate linear completely isometric involution $\pi$ on $A$ of type {\rm (3) at the start of Section {\rm \ref{inv}}, gives rise after composition with the canonical map $- : A \to \bar{A}$, to a linear completely isometric isomorphism $A \to \bar{A}$. This map extends to a $*$-isomorphism $C^*_{\rm max}(A) \to C^*_{\rm max}(\bar{A}) = (C^*_{\rm max}(A)^\circ)^\star$. Composing this with the canonical map $-$, we obtain a conjugate linear $*$-automorphism on $B = C^*_{\rm max}(A)$ which we will also write as $\pi$. This is the compatible conjugate linear involution on $B$. \begin{theorem} Suppose that $A$ is an operator algebra, with a conjugate linear completely isometric involution $\pi$ of type {\rm (3)} at the start of Section {\rm \ref{inv}}. Suppose that $A$ is a subalgebra of a unital $C^*$ -algebra $B$ with compatible conjugate linear $*$-automorphism $\pi$ on $B$. Suppose that $q$ is a closed projection in $B^{**}$ which lies in $(A^1)^{\perp\perp}$ and satisfies $\pi^{**}(q) = q$. If $b$ is an element in $A$ with $bq=qb$ and $b = \pi(b),$ and if $qb^*bq\leq qd$ for an invertible positive $d\in B$ with $d=\pi(d)$ which commutes with $q,$ then there exists a $g\in\ball(A)$ with $gq=qg=bq, g = \pi(g),$ and $g^* g\leq d.$ \end{theorem} \begin{proof} By the proof of \cite[Theorem 3.4]{Bnpi}, there exists $h\in A$ with $qh=hq=bq,$ and $h^*h\leq d$. (We remark that $f = d^{-{\frac{1}{2}}}$ in that proof.) Thus also $\pi(h^*h) \leq \pi(d) = d$. Let $g=\frac{h+\pi(h)}{2}.$ Then $g = \pi(g)$ and $qg=gq=bq$. Also $$g^* g \leq (\frac{h+\pi(h)}{2})^*(\frac{h+\pi(h)}{2})+(\frac{h-\pi(h)}{2})^*(\frac{h-\pi(h)}{2}) .$$ Thus $$g^* g \leq \frac{h^*h}{2}+\frac{\pi(h)^*\pi(h)}{2}= \frac{h^*h}{2}+\pi (\frac{h^*h}{2}) \leq d,$$ as desired. \end{proof} As we said in the introduction, a theorem such as the last one about conjugate linear completely isometric involutions is often simply a result about real operator algebras in the sense of e.g.\ \cite{Sharma}. In this case the result is: Suppose that $A$ is a real operator algebra, a subalgebra of a unital real $C^*$ -algebra $B$. Suppose that $q$ is a closed projection in $B^{**}$ which lies in $(A^1)^{\perp\perp}$. If $b$ is an element in $A$ with $bq=qb$ and $qb^*bq\leq qd$ for an invertible positive $d\in B$ which commutes with $q,$ then there exists a $g\in\ball(A)$ with $gq=qg=bq,$ and $g^* g\leq d.$ We end by listing a couple of other results in our paper that do not seem to extend to all of the types of involutions discussed at the start of Section {\rm \ref{inv}}. Namely, the parts of Lemma \ref{dcai} involving left and right do not work for involutions of types (3) and (4). Also the correspondence between left and right multipliers, left and right $M$-ideals, and left and right $\dagger$-ideals, do not work for these types of involutions. \medskip {\bf Acknowledgments.} This project grew out of \cite{BKM}, and we thank Jens Kaad and Bram Mesland for several ideas and perspectives learned there. We also thank Stephan Garcia--whose work on complex symmetric operators has influenced some results in our paper--for helpful conversations, and also Elias Katsoulis.

3b4b67341bec0627316665c38952984a8588197f

\section{Introduction} \emph{Balls and bins models} are classical probabilistic models, which find numerous applications in Economics and Computer Science. They involve a number of bins, to which balls are added according to a certain rule. A simple example of a balls and bins model is equivalent to the P\'olya urn process introduced by Eggenberger and P\'olya in 1923, \cite{EP}. The model has two bins to which balls are added one by one, with the probability of a ball landing in a particular bin being proportional to the number of balls already in the bin. This is a prime example of a \emph{reinforced} random process, \cite{Ps}. \smallskip The balls and bins model \emph{with feedback}, or a \emph{non-linear P\'olya urn}, is a non-linear generalisation of the P\'olya-Eggenberger model, where the probability of a new ball choosing a bin with $m$ balls is proportional to $f(m)$, with some feedback function $f$. A commonly studied scenario is when there are two bins and $f(m)=m^{\alpha}$, with some positive exponent $\alpha$. The case $\alpha=1$ corresponds to the original P\'olya-Eggenberger model, while the cases $\alpha>1$ and $\alpha\in (0,1)$ are referred to as \emph{positive} and~\emph{negative} feedback, respectively. In the model with positive feedback the bin with a larger number of balls keeps getting more balls, reflecting the principle ``the strong gets stronger'', which leads to its eventual dominance. If the feedback is negative ``the strong gets weaker'', which pushes the bins towards an equilibrium. \smallskip The balls and bins model with feedback was introduced in~\cite{DFM} and was motivated by the problems of economic competition. As noted in~\cite{SV}, the old industrial economies were more prone to a negative feedback, while the market dynamics of new information economy is typically governed by positive feedback. This is due to its network structure and to the fundamental economic effect: it is better to be connected to a bigger network rather than to a smaller one; or, in other words, the more popular a company is the more likely it is to get a new customer. This basic mechanism of the network evolution is called \emph{preferential attachment}, \cite{BA}, and the balls and bins process with feedback is a natural example of a preferential attachment process. \smallskip It is well-known that if there are two bins and $\alpha=1$ then the proportion of the balls in each bin in the long run converges to a beta-distributed random variable (in particular, to a uniform distribution if both bins initially contain the same number of balls). \smallskip For the positive feedback scenario $\alpha>1$, it was shown in~\cite{DFM} that the proportion of balls in each bin converges almost surely to a $\{0,1\}$-valued random variable, an event that we call \emph{dominance}. A much stronger result was obtained in~\cite{KK}, where an equivalent model was considered in the context of neuron growth. It was proved that almost surely one of the bins gets all but finitely many balls, an event that we call \emph{monopoly}. The onset of monopoly was characterised in more detail in~\cite{O2}. A scaling limit for the probability of a certain bin to win, given that the initial number of balls is large, was described in~\cite{MOS}. A generalised model with a growing number of bins was studied in~\cite{CCL}. \smallskip If the feedback is negative the system tends to converge towards an equal number of balls in each bin, and neither monopoly nor dominance occurs, see~\cite{DFM, KK, O}. \smallskip We consider a \emph{time-dependent} balls and bins model with feedback, where the number of balls added at time $n$ is no longer one but is a function of $n$. The balls added at time $n$ choose between the bins independently, and for each ball the probability to land in a bin with $m$ balls is proportional to $m^{\alpha}$ --- just like in the original model. We assume that there are two bins, and that the feedback is either positive ($\alpha>1$) or non-existent ($\alpha=1$). The aim of the paper is to analyse dominance and monopoly in that time-inhomogeneous scenario. This requires a new approach as methods used in~\cite{CCL, KK} rely on Rubin's construction, which fails if the number of added balls is not equal to one. \smallskip Let $(\sigma_n)$ be a positive sequence representing the number of added balls at times $n\in\mathbb N$. Denote by $\tau_0$ the initial number of balls and, for each $n$, let \begin{align*} \tau_n=\tau_0+\sum_{i=1}^n \sigma_i \end{align*} be the total number of balls at time $n$. \smallskip Denote by $0<T_0<\tau_0$ the initial deterministic number of the balls in the first bin. Given that the bin contains $T_{n}$ balls at time $n$, we denote by \begin{align} \label{def1th} \Theta_n=\frac{T_n}{\tau_n} \end{align} the proportion of balls in the first bin and define \begin{align} \label{def1t} T_{n+1}=T_{n}+B_{n+1}, \end{align} where $B_{n+1}$ is a Binomial random variable with size $\sigma_{n+1}$ and parameter \begin{align} \label{defp5} P_n=\frac{\Theta_n^{\alpha}}{\Theta_n^{\alpha}+(1-\Theta_n)^{\alpha}}, \end{align} otherwise independent of $\mathcal{F}_n=\sigma(B_1,\dots,B_n)$. Denote the corresponding probability and expectation by $\P$ and $\mathbb E$, and the conditional probabilities and expectations by $\P_{\mathcal{F}_n}$ and $\mathbb E_{\mathcal{F}_n}$, respectively. \smallskip We denote by \begin{align*} \mathcal{D}=\big\{\exists\lim_{n\to\infty}\Theta_n\in \{0,1\}\big\} \end{align*} the event that eventually the number of balls in one of the bins is negligible, and call this event \emph{dominance}. Further, we denote by \begin{align*} \mathcal{M}=\big\{B_n=0\text{ eventually for all }n\big\} \cup \big\{B_n=\sigma_n\text{ eventually for all }n\big\} \end{align*} the event that eventually all balls are added to one of the bins, and call this event \emph{monopoly}. It is easy to see that \begin{align*} \mathcal{M}\subset \mathcal{D}. \end{align*} \bigskip Our first result is about the no feedback scenario. \begin{theorem} \label{polya} Suppose $\alpha=1$. Then $\Theta_n$ converges almost surely to a random variable $\Theta$, and $\P(\mathcal{D})=0$. \end{theorem} \smallskip \begin{remark} Convergence of $\Theta_n$ follows from a simple martingale argument so the non-trivial part of this theorem is to show that almost surely there is no dominance. \end{remark} \begin{remark} A similar problem was considered in~\cite{P, S}, where all $\sigma_n$ new balls were added to the same bin in a bulk (rather than independently to both bins). This is equivalent to replacing the $(\Theta_{n-1},\sigma_n)$-binomial random variables $B_n$ by $\sigma_n I_n$, where $I_n$ are Bernoulli random variables with parameters $\Theta_{n-1}$, otherwise independent of the past. It was shown in~\cite{P} that $$\sum\limits_{n=0}^{\infty}\Big(\frac{\sigma_{n+1}}{\tau_n}\Big)^2=\infty \qquad\Leftrightarrow\qquad \P(\mathcal{D})=1. $$ Further, it was proved in~\cite{S} that even if the above series converges, we have \begin{align*} \sum_{n=1}^{\infty}\frac{1}{\tau_n}<\infty \qquad\Leftrightarrow\qquad \P(\mathcal{D})>0, \end{align*} provided that $(\sigma_n)$ satisfies some regularity conditions. In other words, there is a phase transition from dominance to no dominance depending on the growth rate of $(\sigma_n)$. This is in a striking contrast with our result, where dominance does not occur for any $(\sigma_n)$. This effect can be heuristically explained by higher step-by-step fluctuations of the second model, i.e.\ $\text{var} (B_n)=\sigma_n\Theta_{n-1}(1-\Theta_{n-1})\ll \sigma_n^2\Theta_{n-1}(1-\Theta_{n-1})=\text{var} (\sigma_nI_n)$, leading to more extreme behaviour of $(\Theta_n)$ for faster-growing $(\sigma_n)$. \end{remark} Now we turn our attention to the positive feedback scenario and assume that $\alpha>1$. Denote \begin{align*} \rho_n=\frac{\sigma_{n+1}}{\tau_n},\quad n\in\mathbb N_0. \end{align*} We will often (but not always) impose the following regularity conditions on the sequence $(\sigma_n)$: \begin{itemize} \item[(S)] $(\sigma_n)$ is either bounded or tends to infinity; \item[(R)] $(\rho_n)$ is either bounded or tends to infinity. \end{itemize} \begin{theorem} \label{main0} Suppose $\alpha>1$, and {\rm (S)} and {\rm (R)} are satisfied. Then $\P(\mathcal{D})=1$. \end{theorem} This theorem means that a time-dependent balls and bins model with positive feedback always exhibits dominance regardless of the growth of $(\sigma_n)$. Monopoly is more delicate, and whether or not it occurs is determined by the growth parameter \begin{align*} \theta=\lim_{n\to\infty}\alpha^{-n}\log\tau_n\in [0,\infty]. \end{align*} In the sequel we assume that this limit exists. We will distinguish between three regimes: supercritical ($\theta=\infty$), subcritical ($\theta=0$), and critical ($0<\theta<\infty$). We will see that monopoly occurs with probability zero and one in the supercritical and subcritical regimes, respectively. In the critical regime the probability of monopoly depends on the finer details of the growth of $(\tau_n)$ and can be strictly between zero and one. \begin{theorem}[Supercritical regime] \label{main2} Suppose $\alpha>1$ and $\theta=\infty$. Then $\P(\mathcal{M})=0$. \end{theorem} \begin{theorem}[Subcritical regime] \label{main1} Suppose $\alpha>1$, $\theta=0$, and {\rm (S)} is satisfied. \smallskip \begin{itemize} \item If $(\rho_n)$ is bounded then $\P(\mathcal{M})=1$. \smallskip \item If $\rho_n\to\infty$ then \begin{align*} \P(\mathcal{M})= \left\{\begin{array}{ll} 1 & \text{ \rm if }\lambda<1,\\ 0 & \text{ \rm if }\lambda>1,\end{array}\right. \end{align*} where \begin{align*} \lambda= \limsup_{n\to\infty} \frac{\sigma_{n+1}\sigma_{n-1}^{\alpha}}{\sigma_n^{\alpha+1}}. \end{align*} In particular, if $\lambda$ is a limit (finite or infinite) then $\lambda=0$ and $\P(\mathcal{M})=1$. \end{itemize} \end{theorem} \medskip \begin{remark} In the subcritical regime one has monopoly almost surely unless the sequence $(\sigma_n)$ is rather irregular. \end{remark} \begin{remark} It is easy to see that if $(\rho_n)$ is bounded then $\theta=0$, see Lemma~\ref{omegazero}, so there is no need to assume the latter. We do it only to make the distinction between the regimes according to the value of $\theta$ more transparent. \end{remark} \begin{theorem}[Critical regime] \label{main3} Suppose $\alpha>1$ and $\theta\in (0,\infty)$. Then \begin{align*} \sum_{n=0}^{\infty}\frac{\tau_{n+1}}{\tau_n^{\alpha}}=\infty &\qquad\Rightarrow \qquad \P(\mathcal{M})=0,\\ \sum_{n=0}^{\infty}\frac{\tau_{n+1}}{\tau_n^{\alpha}}<\infty &\qquad\Rightarrow \qquad \P(\mathcal{M})\in (0,1). \end{align*} \end{theorem} \smallskip \begin{remark} Theorems~\ref{main2} and~\ref{main1} suggest that the phase transition from monopoly to no monopoly happens when $(\tau_n)$ changes from growing slowly to growing fast. Surprisingly, this is no longer true in the critical regime. For example, the sequence $\tau_n=\lfloor b^n e^{\alpha^n}\rfloor$ exhibits monopoly with positive probability if $b>1$ and almost surely no monopoly if $b\le 1$ . \end{remark} The heuristics of the positive feedback scenario can be described as follows. By~\eqref{defp5} we have for all $n\in\mathbb N_0$ \begin{align} \label{defp} P_n=\psi(\Theta_n), \end{align} where \begin{align*} \psi(x)=\frac{x^{\alpha}}{x^{\alpha}+(1-x)^{\alpha}},\qquad x\in [0,1]. \end{align*} It follows from~\eqref{def1th} and~\eqref{def1t} that \begin{align} \label{itit11} \Theta_{n+1}=\frac{\tau_n}{\tau_{n+1}}\Theta_n+\frac{1}{\tau_{n+1}}B_{n+1}, \end{align} and therefore by~\eqref{defp} we have approximately \begin{align} \label{itit111} \Theta_{n+1} \approx\frac{\tau_n}{\tau_{n+1}}\Theta_n+\frac{\sigma_{n+1}}{\tau_{n+1}}P_{n} =\frac{\tau_n}{\tau_{n+1}}\Theta_n+\frac{\sigma_{n+1}}{\tau_{n+1}}\psi(\Theta_n) =\psi_n(\Theta_n), \end{align} where $\psi_n$ is defined as a convex combination of the identity function and the function $\psi$ with the coefficients $\frac{\tau_n}{\tau_{n+1}}$ and $\frac{\sigma_{n+1}}{\tau_{n+1}}$, respectively. Hence the graph of $\psi_n$ is squeezed in the grey area between those two functions on Picture~\ref{figure}, and each $\psi_n$ has three fixed points $0$, $1/2$, and $1$. Since $\psi_n'(1/2)>1$ the equilibrium $1/2$ is not an attractor, so the only natural candidates for the limit of $(\Theta_n)$ are the end points $0$ and $1$. This makes the conjecture of dominance plausible, unless the fluctuations of $(B_n)$ overpower the bias caused by the feedback, especially around the equilibrium. \smallskip \begin{figure} \label{figure} \input{graph.pdf_tex} \caption{The graph of $\psi$ and the location of $\psi_n$} \end{figure} In order to understand the phase transition from the monopolistic to non-monopolistic behaviour, due to domination and symmetry it suffices to consider the event $\{\Theta_n\to 0\}$. In that case we have $\psi(\Theta_n)\approx \Theta_n^{\alpha}$, and the iteration~\eqref{itit111} turns into \begin{align} \label{itit222} \Theta_{n+1} \approx\frac{\tau_n}{\tau_{n+1}}\Theta_n+\frac{\sigma_{n+1}}{\tau_{n+1}}\Theta_n^{\alpha}. \end{align} Next we make a natural assumption that one of terms of~\eqref{itit222} is more important than the other in the long run. If the second term is negligible then we have \begin{align} \label{ass1} \Theta_{n}\approx \frac{\tau_{n-1}}{\tau_n}\Theta_{n-1} \approx\frac{\tau_{n-2}}{\tau_n}\Theta_{n-2}\approx \cdots\approx \frac{1}{\tau_n}. \end{align} Otherwise, if the first term is negligible, then $\sigma_{n+1}\approx\tau_{n+1}$ due to $\Theta_n\gg \Theta_n^{\alpha}$, and we obtain \begin{align*} \Theta_{n}\approx \Theta_{n-1}^{\alpha}\approx \Theta_{n-2}^{\alpha^2} \approx \cdots\approx e^{\alpha^n}. \end{align*} Equating \begin{align*} \frac{1}{\tau_n}\approx e^{\alpha^n} \end{align*} we observe that the phase transition is likely to occur when \begin{align*} \alpha^n \log\tau_n\approx\text{const}, \end{align*} which indeed corresponds to our critical regime. \smallskip This analysis suggests, in particular, that non-monopolistic regimes should be much easier to analyse than monopolistic ones. Indeed, to show that no monopoly occurs we need to bound $\Theta_n$ from below, which can be done by simply dropping the negligible part of~\eqref{itit222}. On the contrary, upper bounds for $\Theta_n$ required for monopolistic regimes prove to be quite challenging. Another difficulty we face is dealing with random fluctuations of $B_n$, which can be quite big and may begin to interfere with the averages discussed so far. This may be particularly damaging around the equilibrium, where the averages have less power due to it being a fixed point of each $\psi_n$. \smallskip The paper is organised as follows. In Section~\ref{s:nomo} we find a sufficient condition for non-occurrence of monopoly (Lemma~\ref{diverges}). This allows us to handle the supercritical case in the positive feedback scenario (Theorem~\ref{main2}). We also show that almost surely monopoly does not occur if there is no feedback (Lemma~\ref{l:lll}). In Section~\ref{s:nofeedback} we consider the feedback-less case in detail and strengthen the latter result from no monopoly to no dominance (Theorem~\ref{polya}). Then we move to the positive feedback scenario. In Section~\ref{s:away} we show that the proportion $\Theta_n$ does not get stuck at the equilibrium and deviates from it significantly (although still infinitesimally) infinitely often. In Section~\ref{s:dom} we prove almost sure dominance (Theorem~\ref{main0}) by first deviating far enough from the equilibrium and then showing that the feedback drags $\Theta_n$ even further away. Section~\ref{s:pro} contains some technical lemmas required later for the subcritical regime. In Section~\ref{s:sub} we prove monopoly in the two cases corresponding to a bounded $(\rho_n)$ and $(\rho_n)$ tending to infinity with $\lambda<1$ (Propositions~\ref{p:mod1} and \ref{p:fast}, respectively). In the third case of $(\rho_n)$ tending to infinity with $\lambda>1$ we use the same idea as in Section~\ref{s:nomo} to show no monopoly. This completes the proof of Theorem~\ref{main1}. Finally, Section~\ref{s:cri} is devoted to the critical regime. There we prove that the probability of monopoly is always less then one, and provide a sufficient condition for it to be zero (Proposition~\ref{p:cri1}). Then we show that that condition is also a necessary one (Proposition~\ref{p:cri2}), completing the proof of Theorem~\ref{main3}. \smallskip Finally, we would like to introduce notation that will be used throughout the paper. Let \begin{align*} \varepsilon_n=\frac{B_n-\sigma_{n}P_{n-1}}{\sqrt{\sigma_{n}P_{n-1}(1-P_{n-1})}}, \quad n\in \mathbb N, \end{align*} be the normalised fluctuation of $B_n$. It is easy to see that \begin{align} \label{norm} \mathbb E_{\mathcal{F}_{n-1}}\varepsilon_{n}=0 \qquad\text{and}\qquad \mathbb E_{\mathcal{F}_{n-1}}\varepsilon_{n}^2=1. \end{align} It follows from~\eqref{def1t} and~\eqref{defp} that for all $n\in\mathbb N_0$ \begin{align} \label{iteration1} T_{n+1}=T_n+\sigma_{n+1}\psi(\Theta_n)+\varepsilon_{n+1}\sqrt{\sigma_{n+1}P_n(1-P_n)}. \end{align} At last, the following obvious bound on $\psi$ will be used repeatedly: \begin{align} \label{psibounds} x^{\alpha}\le \psi(x)\le 2^{\alpha-1}x^{\alpha} \qquad\text{for all }x\in [0,1]. \end{align} \bigskip \section{No monopoly} \label{s:nomo} This section is devoted to non-monopolistic regimes. As we mentioned earlier, showing no monopoly is not hard, and a sufficient condition for non-occurrence of monopoly is established in the lemma below. In fact, it will follow from further results that this condition is a necessary one provided the sequence $(\sigma_n)$ satisfies the regularity conditions (S) and (R) and has a well-defined parameter $\theta$. \begin{lemma} \label{diverges} If \begin{align} \label{div7} \sum_{n=0}^{\infty}\frac{\sigma_{n+1}}{\tau_n^{\alpha}}=\infty, \end{align} then $\P(\mathcal{M})=0$. \end{lemma} \begin{proof} First, let us show that \begin{align} \label{BC} \big\{T_n\to\infty\big\}\supset\Big\{\sum_{n=1}^{\infty}\sigma_nP_{n-1}=\infty\Big\}. \end{align} Observe that for all $n\in\mathbb N_0$ \begin{align} \label{defm} T_n=T_0+\sum_{i=1}^{n}B_i=T_0+M_n+Y_n, \end{align} where \begin{align*} M_n=\sum_{i=1}^{n}\big(B_i-\sigma_iP_{i-1}) \qquad\text{and}\qquad Y_n=\sum_{i=1}^{n}\sigma_iP_{i-1} \end{align*} We have \begin{align} \langle M\rangle_n =\sum_{i=1}^{n}\sigma_i P_{i-1}(1-P_{i-1})\le Y_n. \label{my} \end{align} Suppose $Y_n\to\infty$. If $\langle M\rangle_n$ converges then $M_n$ converges, see~\cite[\S 12.13]{DW}, and $T_n\to\infty$ by~\eqref{defm}. If $\langle M\rangle_n\to\infty$ then $M_n/\langle M\rangle_n\to 0$, see~\cite[\S 12.14]{DW}. By~\eqref{my} we have $M_n/Y_n\to 0$, which implies $T_n\to\infty$ by~\eqref{defm}. This completes the proof of~\eqref{BC}. \smallskip Second, observe that by~\eqref{psibounds} almost surely \begin{align*} \sum_{n=1}^{\infty}\sigma_{n}P_{n-1} =\sum_{n=1}^{\infty}\sigma_{n}\psi(\Theta_{n-1}) \ge \sum_{n=1}^{\infty}\sigma_{n}\Theta_{n-1}^{\alpha} =\sum_{n=1}^{\infty}\frac{\sigma_{n}}{\tau_{n-1}^{\alpha}}T_{n-1}^{\alpha} \ge \sum_{n=1}^{\infty}\frac{\sigma_{n}}{\tau_{n-1}^{\alpha}}=\infty. \end{align*} Hence by~\eqref{BC} we have $\P(T_n\to\infty)=1$. By symmetry this implies $\P(\mathcal{M})=0$. \end{proof} \smallskip \begin{proof}[Proof of Theorem~\ref{main2}] By Lemma~\ref{diverges} it suffices to show that $\theta=\infty$ implies~\eqref{div7}. Suppose the series converges. Then $\frac{\sigma_{n}}{\tau_{n-1}^{\alpha}}\to 0$. Let $k$ be such that $\sigma_{n}<\tau_{n-1}^{\alpha}$ for all $n\ge k$. Then for all $n\ge k$ \begin{align*} \tau_n\le \tau_{n-1}+\tau_{n-1}^{\alpha}\le 2\tau_{n-1}^{\alpha}\le\cdots\le 2^{1+\alpha+\cdots+\alpha^{n-k-1}}\tau_k^{\alpha^{n-k}} \le (2\tau_k)^{\alpha^{n-k}}. \end{align*} This implies \begin{align*} \theta\le \alpha^{-k}\log(2\tau_k)<\infty \end{align*} leading to a contradiction. \end{proof} \smallskip \begin{lemma} \label{l:lll} Suppose $\alpha=1$. Then $\P(\mathcal{M})=0$. \end{lemma} \begin{proof} This follows from Lemmas~\ref{diverges} and Lemma~\ref{1101series} below. \end{proof} \smallskip \begin{lemma} \label{1101series} We have \begin{align*} \sum_{n=0}^{\infty}\frac{\sigma_n}{\tau_n}=\infty. \end{align*} \end{lemma} \begin{proof} We have \begin{align*} \tau_n=\tau_{n-1}\Big(1-\frac{\sigma_{n}}{\tau_{n}}\Big)^{-1}=\tau_0\prod_{k=1}^{n}\Big(1-\frac{\sigma_k}{\tau_k}\Big)^{-1} =\tau_0\exp\Big\{-\sum_{k=1}^n\log \Big(1-\frac{\sigma_k}{\tau_k}\Big)\Big\}. \end{align*} This implies \begin{align*} -\sum_{k=1}^n\log \Big(1-\frac{\sigma_k}{\tau_k}\Big)=\log\frac{\tau_n}{\tau_0}\to \infty, \end{align*} which is equivalent to the statement of the lemma. \end{proof} \bigskip \section{No feedback scenario} \label{s:nofeedback} In this section we use Laplace transforms to prove that no dominance occurs in the case $\alpha=1$. \begin{proof}[Proof of Theorem~\ref{polya}] Observe that $(\Theta_n)$ is a bounded martingale as for all $n$ we have by~\eqref{itit11} \begin{align*} \mathbb E_{\mathcal{F}_{n-1}}\Theta_n=\frac{\tau_{n-1}}{\tau_n}\Theta_{n-1}+\frac{\sigma_n}{\tau_n}\psi(\Theta_{n-1})=\Theta_{n-1}. \end{align*} Hence it converges to a random variable $\Theta$ bounded between zero and one. By symmetry it suffices to prove that $\P(\Theta=0)=0$. \smallskip Denote by \begin{align*} f_n(\lambda)=\mathbb E e^{-\lambda \Theta_n} \qquad\text{and}\qquad f(\lambda)=\mathbb E e^{-\lambda \Theta}, \qquad \lambda\in\mathbb R, \end{align*} the Laplace transforms of $\Theta_n$ and $\Theta$, respectively. Since \begin{align*} f(\lambda)\ge \P(\Theta=0) \end{align*} for all $\lambda$, it suffices to show that there is a sequence $(\lambda_m)$ such that \begin{align} \label{bbb2} \lim_{m\to\infty}f(\lambda_m)=0. \end{align} To do so, observe that there is $c\in (0,1)$ such that \begin{align} \label{exp0} e^{-x}\le 1-x+\frac{x^2}{2} \end{align} for all $x\in [0,c]$, and then define $\lambda_m=c\tau_m$. \smallskip Let us prove by induction over $k$ that \begin{align} \label{indu} f_n(\lambda_m)\le f_{n-k}\Big(\lambda_m-\lambda_m^2\sum_{i=n-k+1}^{n}\frac{\sigma_i}{\tau_i^2}\Big) \end{align} for all $m$, $n>m$, and $1\le k\le n-m$. \smallskip First, observe that for all $\lambda\in [0,c\tau_m]$ we have by~\eqref{itit11} and~\eqref{exp0} that \begin{align*} f_n(\lambda) &=\mathbb E \Big[\exp\Big\{-\frac{\tau_{n-1}}{\tau_n}\lambda\Theta_{n-1}+\sigma_n\log \Big(1-\Theta_{n-1}+\Theta_{n-1} \exp\Big\{-\frac{\lambda}{\tau_n}\Big\}\Big)\Big]\\ &\le \mathbb E \Big[\exp\Big\{-\frac{\tau_{n-1}}{\tau_n}\lambda\Theta_{n-1} +\sigma_n\log\Big(1- \frac{\lambda}{\tau_n}\Theta_{n-1}+ \frac{\lambda^2}{2\tau_n^2}\Theta_{n-1}\Big)\Big\}\Big] \end{align*} for all $m$ and $n>m$ since \begin{align*} \frac{\lambda}{\tau_n}\le \frac{\lambda}{\tau_m}=c. \end{align*} Further, using $\log(1+x)\le x$ and dropping the factor $1/2$ we obtain \begin{align} f_n(\lambda) &\le \mathbb E \Big[\exp\Big\{-\lambda\Theta_{n-1} + \frac{\sigma_n}{\tau_n^2}\lambda^2\Theta_{n-1}\Big\}\Big] \le f_{n-1}\Big(\lambda-\frac{\sigma_n}{\tau_n^2}\lambda^2\Big). \label{it7} \end{align} This proves the statement~\eqref{indu} for $k=1$. Further, suppose~\eqref{indu} it is true for some $k$. Observe that \begin{align} \label{bbb1} \lambda_m-\lambda_m^2\sum_{i=n-k+1}^{n}\frac{\sigma_i}{\tau_i^2}\le \lambda_m=c\tau_m \end{align} and, since $\tau_m\le \tau_{n-k}$ and $c<1$, \begin{align} \label{bbb} \lambda_m-\lambda_m^2\sum_{i=n-k+1}^{n}\frac{\sigma_i}{\tau_i^2}\ge \lambda_m\Big[1-c\tau_m\int_{\tau_{n-k}}^{\infty}\frac{dx}{x^2}\Big] =\lambda_m\Big[1-c\frac{\tau_m}{\tau_{n-k}}\Big]\ge \lambda_m(1-c)\ge 0. \end{align} By~\eqref{bbb1} and~\eqref{bbb} we can use~\eqref{it7} together with monotonicity of $f_{n-k-1}$ to obtain \begin{align*} f_{n-k}\Big(\lambda_m-\lambda_m^2\sum_{i=n-k+1}^{n}\frac{\sigma_i}{\tau_i^2}\Big) & \le f_{n-k-1}\Big(\lambda_m-\lambda_m^2\sum_{i=n-k+1}^{n}\frac{\sigma_i}{\tau_i^2} -\frac{\sigma_{n-k}}{\tau_{n-k}^2}\Big(\lambda_m-\lambda_m^2\sum_{i=n-k+1}^{n}\frac{\sigma_i}{\tau_i^2}\Big)^2\Big)\\ & \le f_{n-k-1}\Big(\lambda_m-\lambda_m^2\sum_{i=n-k}^{n}\frac{\sigma_i}{\tau_i^2}\Big) \end{align*} by replacing the squared term by $\lambda_m^2$ only. This, together with the induction hypothesis~\eqref{indu} for $k$, completes the induction step from $k$ to $k+1$. \smallskip Substituting $k=n-m$ into~\eqref{indu} and using~\eqref{bbb} together with monotonicity of $f_m$ we obtain \begin{align} f_n(\lambda_m)\le f_{m}\Big(\lambda_m-\lambda_m^2\sum_{i=m+1}^{n}\frac{\sigma_i}{\tau_i^2}\Big) \le f_m \big(\lambda_m(1-c)\big) =\mathbb E e^{-c(1-c)\tau_m\Theta_m}=\mathbb E e^{-c(1-c)T_m} \label{lili7} \end{align} for all $m$ and $n>m$. \smallskip By the dominated convergence theorem we have $f_n(\lambda)\to f(\lambda)$ for all $\lambda>0$. Hence we can take the limit in~\eqref{lili7} to obtain \begin{align*} f(\lambda_m)\le \mathbb E e^{-c(1-c)T_m} \end{align*} for all $m$. To prove~\eqref{bbb2}, it remains to notice that \begin{align*} \lim_{m\to\infty}\mathbb E e^{-c(1-c)T_m}= 0 \end{align*} by the dominated convergence theorem as there is no monopoly, that is, $T_m\to\infty$ almost surely by Lemma~\ref{l:lll}. \end{proof} \bigskip \section{Getting away from the equilibrium} In this section we show that in the positive feedback scenario $\Theta_n$ deviates from the equilibrium far enough infinitely often. \label{s:away} \begin{prop} \label{devp} Suppose $\alpha>1$ and {\rm (S)} is satisfied. Let $(\delta_n)$ be a positive sequence converging to zero and such that \begin{align} \label{ser1} \sum_{n=0}^{\infty}\delta_n\, \frac{\sigma_{n+1}}{\tau_{n+1}}<\infty. \end{align} Then \begin{align*} \P\big(|\Theta_n-1/2|>\delta_n \text{ infinitely often}\big)=1. \end{align*} \end{prop} The idea of the proof can be vaguely described as follows. We assume that $\Theta_n$ does not deviate from the equilibrium enough and aim to show that it is very unlikely. If $\Theta_n\approx 1/2$ then $P_n\approx 1/2$ and $\psi(\Theta_n)\approx \alpha (\Theta_n-1/2)+1/2$ as $\psi'(1/2)=\alpha$. Substituting this into~\eqref{iteration1} we obtain \begin{align*} \Theta_{n}-\frac 1 2 &\approx\Big(\frac{\tau_{n-1}}{\tau_{n}} +\alpha\frac{\sigma_n}{\tau_n}\Big)\Big(\Theta_{n-1}-\frac 1 2 \Big)+\varepsilon_{n}\frac{\sqrt{\sigma_{n}}}{2\tau_{n}}. \end{align*} Carefully iterating this from a fixed large $m$ to $n\to\infty$, we obtain \begin{align} \label{app1} \Theta_{n}-\frac 1 2 \approx \pi_{m,n}\Big(\Theta_{m}-\frac 1 2 +\frac{1}{2}\mu_{m,n}N_{m,n}\Big), \end{align} where $\pi_{m,n}$ and $\mu_{m,n}$ are some deterministic scales, and $N_{m,n}$ is a random variable arisen from the noises $(\varepsilon_i)$ between the times $m$ and $n$. Then we observe by a CLT argument that the distribution of $N_{m,n}$ is asymptotically close to normal. Finally, we observe that $\pi_{m,n}\mu_{m,n}\to \infty$ which, together with $\Theta_n\approx 1/2$, $\Theta_m\approx 1/2$ and $N_{m,n}$ being of a finite non-negligible order, makes~\eqref{app1} impossible. \begin{proof}[Proof of Proposition~\ref{devp}] Denote \begin{align*} \mathcal{H}_m=\big\{|\Theta_n-1/2|\le \delta_n\text{ for all }n\ge m\big\}. \end{align*} Since the events $\mathcal{H}_m$ are increasing it suffices to show that \begin{align*} \P(\mathcal{H}_m)\to 0. \end{align*} By the mean value theorem we have \begin{align*} \psi(x)=\psi'(\xi_x)\big(x-\frac{1}{2}\big)+\frac 1 2, \end{align*} where $\xi_x$ lies between $x$ and $1/2$. Hence \begin{align} \label{ka1p} \frac{\tau_n}{\tau_{n+1}}x+\frac{\sigma_{n+1}}{\tau_{n+1}}\psi(x)-\frac 1 2 &=\frac{\tau_n}{\tau_{n+1}} x+\frac{\sigma_{n+1}}{\tau_{n+1}} \Big[\psi'(\xi_x)\big(x-\frac{1}{2}\big)+\frac 1 2\Big]-\frac 1 2 =\kappa_n(x) \big(x-\frac 1 2\big), \end{align} where \begin{align*} \kappa_n(x)=\frac{\tau_n}{\tau_{n+1}}+\frac{\sigma_{n+1}}{\tau_{n+1}}\psi'(\xi_x). \end{align*} It follows from~\eqref{iteration1} and~\eqref{def1th} that for all $m$ and $n\ge m$ \begin{align*} \Theta_{n}-\frac 1 2 &=\frac{\tau_{n-1}}{\tau_{n}} \Theta_{n-1}+\frac{\sigma_{n}}{\tau_{n}}\psi(\Theta_{n-1})-\frac 1 2+\varepsilon_{n}\frac{\sqrt{\sigma_{n}P_{n-1}(1-P_{n-1})}}{\tau_{n}} \\ &=\kappa_{n-1}(\Theta_{n-1})\Big(\Theta_{n-1}-\frac 1 2\Big)+\varepsilon_{n}\frac{\sqrt{\sigma_{n}P_{n-1}(1-P_{n-1})}}{\tau_{n}}. \end{align*} Iterating this procedure we get \begin{align} \Theta_{n}-\frac 1 2 &=\Big[\prod_{j=m}^{n-1}\kappa_j(\Theta_{j})\Big] \Big(\Theta_{m}-\frac 1 2\Big) +\sum_{k=m+1}^{n}\Big[\prod_{j=k}^{n-1}\kappa_j(\Theta_{j})\Big] \varepsilon_{k}\frac{\sqrt{\sigma_{k}P_{k-1}(1-P_{k-1})}}{\tau_{k}}\notag\\ &=\Big[\prod_{j=m}^{n-1}\kappa_j(\Theta_{j})\Big] \Big(\Theta_{m}-\frac 1 2 +\sum_{k=m+1}^n\Big[\prod_{j=m}^{k-1}\frac{1}{\kappa_j(\Theta_{j})}\Big] \varepsilon_{k}\frac{\sqrt{\sigma_{k}P_{k-1}(1-P_{k-1})}}{\tau_{k}}\Big). \label{ka4p} \end{align} Observe that on the event $\mathcal{H}_{m}$ \begin{align} \label{bobo} \kappa_j(\Theta_j)=\frac{\tau_j}{\tau_{j+1}}+\frac{\sigma_{j+1}}{\tau_{j+1}}\Big[\psi'(1/2)+O(\delta_j)\Big] =\frac{\tau_j}{\tau_{j+1}}+\frac{\sigma_{j+1}}{\tau_{j+1}}\alpha+\frac{\sigma_{j+1}}{\tau_{j+1}}O(\delta_j) \end{align} as $j\to\infty$ uniformly on the probability space. Denote \begin{align} \label{rhorho} \pi_{m,k}=\prod_{j=m}^{k-1}\frac{\tau_j+\alpha\sigma_{j+1}}{\tau_{j+1}}=\exp\Big\{\sum_{j=m+1}^{k}\log\Big(1+(\alpha-1)\frac{\sigma_j}{\tau_j}\Big)\Big\}. \end{align} It follows from~\eqref{bobo} that, as $m\to\infty$, uniformly in $\omega$ and $k\ge m$ \begin{align} \prod_{j=m}^{k-1}\kappa_j(\Theta_{j}) =\pi_{m,k} \exp\Big\{\sum_{j=m}^{k-1}\log\Big(1+\frac{\sigma_{j+1}}{\tau_{j}+\alpha\sigma_{j+1}}O(\delta_j)\Big)\Big\ &=\pi_{m,k} \big(1+o(1)\big), \label{ka2p} \end{align} since by~\eqref{ser1} we have \begin{align*} \sum_{j=m}^{k-1}\frac{\delta_j\sigma_{j+1}}{\tau_{j}+\alpha\sigma_{j+1}} \le \sum_{j=m}^{\infty}\delta_j\, \frac{\sigma_{j+1}}{\tau_{j+1}}=o(1). \end{align*} Further, on $\mathcal{H}_{m}$ we have, as $m\to\infty$, uniformly in $\omega$ and $k\ge m$ \begin{align} \label{ka3p} \sqrt{P_{k-1}(1-P_{k-1})}=\sqrt{\psi(\Theta_{k-1})(1-\psi(\Theta_{k-1}))}=1/2+O(\delta_{k-1})=1/2+o(1). \end{align} Substituting~\eqref{ka2p} and~\eqref{ka3p} into~\eqref{ka4p} we obtain, as $m\to\infty$, uniformly in $\omega$ and $n\ge m$ \begin{align*} &\Theta_{n}-\frac 1 2 =(1+o(1))\pi_{m,n} \Big(\Theta_{m}-\frac 1 2 +\frac{1+o(1)}{2}\sum_{k=m+1}^n \varepsilon_{k}\frac{\sqrt{\sigma_{k}}}{\pi_{m,k}\tau_{k}}\Big). \end{align*} Let $(B^{\ssup 0}_{\ell})_{\ell\in\mathbb N}$ be a sequence of independent Binomial random variables with parameter $\frac 1 2$ and size~$\sigma_{\ell}$, and let, for all $\ell\in\mathbb N$, \begin{align*} \varepsilon_{\ell}^{\ssup 0}=\frac{B^{\ssup 0}_{\ell}-\sigma_{\ell}/2}{\sqrt{\sigma_{\ell}/4}}. \end{align*} Denote \begin{align*} \hat{\varepsilon}_{\ell}=\varepsilon_{\ell}{\mathbf 1}\big\{|\Theta_{{\ell}-1}-1/2|<1/4\big\}+\varepsilon_{\ell}^{\ssup 0}{\mathbf 1}\big\{|\Theta_{{\ell}-1}-1/2|\ge 1/4\big\}. \end{align*} On the event $\mathcal{H}_{m}$ we have, as $m\to\infty$, uniformly in $\omega$ and $n\ge m$ \begin{align} \Theta_{n}-\frac 1 2 &=(1+o(1))\pi_{m,n} \Big(\Theta_{m}-\frac 1 2 +\frac{1+o(1)}{2}\sum_{k=m+1}^n\hat{\varepsilon}_{k}\frac{\sqrt{\sigma_{k}}}{\pi_{m,k}\tau_{k}} \Big). \label{ka7p} \end{align} \smallskip It is easy to see that for a Binomial random variable $\text{Bin}(p,n)$ we have \begin{align*} \mathbb E \exp\Big\{it\frac{\text{\rm Bin}(p,n)-np}{\sqrt{np(1-p)}}\Big\} &=\exp\Big\{n\log\Big(1-p+p\exp\Big\{\frac{it}{\sqrt{np(1-p)}}\Big\}\Big)-\frac{it\sqrt{np}}{\sqrt{1-p}}\Big\}\\ &=\exp\Big\{n\log\Big(1+\frac{it\sqrt{p}}{\sqrt{n(1-p)}}- \frac{t^2 }{2n(1-p)}+O\big(\frac{t^3}{n\sqrt n}\big)\Big)-\frac{it\sqrt{np}}{\sqrt{1-p}}\Big\}\\ &=\exp\Big\{- \frac{t^2 }{2}+nO\big(\frac{t^3}{n\sqrt n}\big)\Big\} \end{align*} as $n\to\infty$, uniformly in ${\frac 1 4<p<\frac 3 4}$ and $t$. Hence \begin{align} \label{ka9p} \mathbb E_{\mathcal{F}_{\ell-1}}e^{it\hat{\varepsilon}_{\ell}}&=\exp\Big\{- \frac{t^2 }{2}+\sigma_{\ell}O\big(\frac{t^3}{\sigma_{\ell}\sqrt{\sigma_{\ell}}}\big)\Big\} \end{align} as $\ell\to\infty$, uniformly in $\omega$ and $t$. Denote \begin{align} \label{mumu} \mu_{m,n}=\Big[\sum_{k=m+1}^n\frac{\sigma_{k}}{\pi^2_{m,k}\tau^2_{k}}\Big]^{1/2}. \end{align} Using~\eqref{ka9p} we have for each fixed $t$, as $m\to\infty$, uniformly in $\omega$ and $n\ge m$, \begin{align} \label{ka6p} \mathbb E_{\mathcal{F}_m}\exp\Big\{\frac{it}{\mu_{m,n}}\sum_{k=m+1}^n\hat{\varepsilon}_{k}\frac{\sqrt{\sigma_{k}}}{\pi_{m,k}\tau_{k}}\Big\} &=\exp\Big\{-\frac{t^2}{2}+O(1)\frac{1}{\mu_{m,n}^3}\sum_{k=m+1}^n\frac{\sigma_{k}}{\pi^3_{m,k}\tau^3_{k}}\Big\}. \end{align} Let us show that the term next to $O(1)$ tends to zero as $m\to\infty$. We need to consider two cases: when $(\sigma_i)$ tends to infinity and when it is bounded. \smallskip First, assume that $\sigma_i\to \infty$. Observe that since $(x_1+\cdots+x_m)^{\frac 3 2}\ge x_1^{\frac 3 2}+\cdots+x_m^{\frac 3 2}$ for all $m$ and all non-negative $x_1,\dots,x_m$, we have \begin{align*} \mu^3_{m,n}\ge \sum_{k=m+1}^n\frac{\sigma^{3/2}_{k}}{\pi^3_{m,k}\tau^3_{k}} \ge \min_{m<k\le n}\sqrt{\sigma_{k}}\,\sum_{k=m+1}^n\frac{\sigma_{k}}{\pi^3_{m,k}\tau^3_{k}} \end{align*} and hence \begin{align*} \frac{1}{\mu_{m,n}^3}\sum_{k=m+1}^n\frac{\sigma_{k}}{\pi^3_{m,k}\tau^3_{k}}\le \max_{m<k\le n}\frac{1}{\sqrt{\sigma_{k}}}\to 0 \end{align*} as $m\to\infty$ uniformly in $n$. \smallskip Second, assume that $\sigma_i$ is bounded by a constant $\sigma$. Hence $i\le \tau_i\le \sigma (i+1)\le 2\sigma i$ for all $i$. We have \begin{align} \pi_{m,k} &\le \exp\Big\{\sum_{j=m+1}^{k}\log\Big(1+(\alpha-1)\frac{\sigma}{j}\Big)\Big\} = \exp\Big\{\big[(\alpha-1)\sigma+o(1)\big]\sum_{j=m+1}^k\frac 1 j\Big\}\notag\\ &= \exp\Big\{\big[(\alpha-1)\sigma+o(1)\big]\big(\log k-\log m+o(1)\big)\Big\}=\frac{k^{(\alpha-1)\sigma+o(1)}}{m^{(\alpha-1)\sigma+o(1)}}. \label{upbo} \end{align} and, similarly, \begin{align} \label{lobo} \pi_{m,k} &\ge\frac{k^{\frac{\alpha-1}{\sigma}+o(1)}}{m^{\frac{\alpha-1}{\sigma}+o(1)}}. \end{align} as $m\to\infty$ uniformly in $k\ge m$. Observe that for all $\gamma>1$ \begin{align*} \sum_{k=m+1}^n\frac{1}{k^{\gamma+o(1)}}=\frac{1}{m^{\gamma-1+o(1)}} \end{align*} as $m\to\infty$ uniformly in $n\ge m^2$. Using this we obtain by~\eqref{upbo} \begin{align} \label{mumum} \mu_{m,n}^3\ge\frac{m^{3(\alpha-1)\sigma+o(1)}}{8\sigma^3}\Big[\sum_{k=m+1}^n \frac{1}{k^{2(\alpha-1)\sigma+2+o(1)}}\Big]^{3/2} =m^{3(\alpha-1)\sigma-\frac 3 2 (2(\alpha-1)\sigma+1)+o(1)}=m^{-\frac 3 2 +o(1)} \end{align} and by~\eqref{lobo} \begin{align} \label{mumumu} \sum_{k=m+1}^n\frac{\sigma_{k}}{\pi^3_{m,k}\tau^3_{k}} &\le \sigma m^{\frac{3(\alpha-1)}{\sigma}+o(1)}\sum_{k=m+1}^n \frac{1}{k^{\frac{3(\alpha-1)}{\sigma}+3+o(1)}} =m^{-2+o(1)} \end{align} as $m\to\infty$ uniformly in $n\ge m^2$. Combining~\eqref{mumum} and~\eqref{mumumu} we obtain \begin{align*} \frac{1}{\mu_{m,n}^3}\sum_{k=m+1}^n\frac{\sigma_{k}}{\pi^3_{m,k}\tau^3_{k}}=m^{-1/2+o(1)}\to 0 \end{align*} as $m\to\infty$ uniformly in $n\ge m^2$. \smallskip Let $(n_m)$ be an $\mathbb N$-valued sequence satisfying $n_m\ge m^2$. We will need $(n_m)$ to grow sufficiently fast but will specify this condition later. It follows from~\eqref{ka7p} and~\eqref{ka6p} that on the event $\mathcal{H}_m$ \begin{align*} \Theta_{n_m}-\frac 1 2 &=(1+o(1))\pi_{m,n_m} \Big(\Theta_{m}-\frac 1 2 +\frac{1+o(1)}{2}\,\mu_{m,n_m} N_{m,n_m}\Big), \end{align*} where $N_{m,n_m}$ is a random variable, which conditionally on $\Theta_m$ converges weakly to a standard normal random variable $N$. Hence \begin{align*} \P(\mathcal{H}_m)\le \P\Big((1+o(1))\pi_{m,n_m} \Big(\Theta_{m}-\frac 1 2 +\frac{1+o(1)}{2}\,\mu_{m,n_m} N_{m,n_m}\Big)\in [-\delta_{n_m},\delta_{n_m}]\Big)\to 0 \end{align*} as required if \begin{align} \label{fin} \frac{\pi_{m,n_m}\mu_{m,n_m}}{\delta_{n_m}}\to \infty. \end{align} Finally, to prove~\eqref{fin}, we estimate the sum in~\eqref{mumu} by the first term and obtain using $\sigma_{m+1}\ge 1$, $\pi_{m.m+1}\le \alpha$, and~\eqref{rhorho} \begin{align*} \pi_{m,n_m}\mu_{m,n_m} \ge \pi_{m,n_m}\,\frac{\sqrt{\sigma_{m+1}}}{\pi_{m,m+1}\tau_{m+1}} \ge\frac{1}{\alpha\tau_{m+1}}\exp\Big\{\sum_{j=m+1}^{n_m}\log\Big(1+(\alpha-1)\frac{\sigma_j}{\tau_j}\Big)\Big\}. \end{align*} The series \begin{align*} \sum_{j=0}^{\infty} \log\Big(1+(\alpha-1)\frac{\sigma_j}{\tau_j}\Big)=\infty \end{align*} diverges by Lemma~\ref{1101series} and hence we can choose $(n_m)$ to grow sufficiently fast to guarantee \begin{align*} \pi_{m,n_m}\mu_{m,n_m}\to \infty, \end{align*} which implies~\eqref{fin}. \end{proof} \bigskip \section{Dominance} \label{s:dom} The aim of this section is to prove almost sure dominance in the positive feedback scenario. We will rely on Proposition~\ref{devp} for showing that the proportion $\Theta_n$ of the balls in the first bin does not get stuck at the equilibrium. To do so we need to pick the sequence of deviations $(\delta_n)$ satisfying the assumption~\eqref{ser1}. Let \begin{align} \label{d:deltan1} \delta_n=\frac{1}{\log^2\tau_n},\quad n\in\mathbb N_0. \end{align} \begin{lemma} \label{1101m} Suppose {\rm (R)} is satisfied. Then \begin{align} \label{tmp1} \sum_{n=0}^{\infty}\delta_n\frac{\sigma_{n+1}}{\tau_{n+1}}<\infty. \end{align} \end{lemma} \begin{proof} First, suppose $(\rho_n)$ is bounded. We have \begin{align*} \log \tau_{n+1}=\log\tau_{n}+\log (1+\rho_{n})\sim \log\tau_{n} \end{align*} as $n\to\infty$, and the convergence of the above series follows from \begin{align*} \sum_{n=0}^{\infty}\frac{\sigma_{n+1}}{\tau_{n+1}\log^2\tau_{n+1}}\le \int_{\tau_0}^{\infty}\frac{dx}{x\log^2x}<\infty. \end{align*} Second, suppose $\rho_n\to\infty$. Then there exists $m\in \mathbb N$ such that for all $n\ge m$ we have $\rho_n\ge 2$ and hence $\tau_{n}\ge \sigma_{n}\ge 2\tau_{n-1}\ge\cdots\ge 2^{n-m}\tau_m$. Together with $\sigma_{n+1}\le \tau_{n+1}$ this implies \begin{align*} \delta_n\frac{\sigma_{n+1}}{\tau_{n+1}}\le \frac{1}{\log^2 \tau_{n}}\le \frac{1}{((n-m)\log 2+\log\tau_m)^2} \sim \frac{1}{n^2\log^22} \end{align*} as $n\to\infty$ implying convergence of the series~\eqref{tmp1}. \end{proof} \smallskip Now we are ready to prove dominance. Our strategy will be as follows. First we pick a time when $\Theta_n$ deviates from the equilibrium. Then we observe that if that time is large enough then the future fluctuations of the martingale part of $\Theta_n$ will be small with high probability, and will keep $\Theta_n$ away from the equilibrium. At the same time, the bias caused by the positive feedback will move $\Theta_n$ away from the equilibrium, and its power will be sufficient to bring it to zero or one, respectively. \begin{proof}[Proof of Theorem~\ref{main0}] \smallskip Fix $r\ge 3$ and let \begin{align*} \eta=\inf\big\{n\ge r: |\Theta_n-1/2|>\delta_n\big\} \end{align*} be the first time $\Theta_n$ significantly deviates from the equilibrium after time $r$. Observe that $\eta$ is finite almost surely by Proposition~\ref{devp} and Lemma~\ref{1101m}. Due to symmetry it suffices to consider the event \begin{align} \label{lo3} \mathcal{E}=\big\{\Theta_{\eta}<1/2-\delta_{\eta}\big\}. \end{align} and show that on $\Theta_n\to 0$ on $\mathcal{E}$. \smallskip For each $n\in\mathbb N_0$ we have by~\eqref{def1th}, \eqref{def1t}, and~\eqref{defp} \begin{align*} \Theta_{n+1}=\frac{\tau_n}{\tau_{n+1}}\Theta_n+\frac{1}{\tau_{n+1}}B_{n+1}=\Theta_n+\frac{B_{n+1}-\sigma_{n+1}P_n}{\tau_{n+1}} -\frac{\sigma_{n+1}}{\tau_{n+1}}\big(\Theta_n-\psi(\Theta_n)\big). \end{align*} Hence for each $k\ge \eta$ we have \begin{align} \label{lo2} \Theta_{\eta+n} =\Theta_{\eta}+M_n-R_n, \end{align} where, for each $n\in\mathbb N_0$, \begin{align*} M_n=\sum_{k=\eta+1}^{\eta+n}\frac{B_{k}-\sigma_{k}P_{k-1}}{\tau_k} \qquad\text{and}\qquad R_n=\sum_{k=\eta+1}^{\eta+n}\frac{\sigma_{k}}{\tau_{k}}\big(\Theta_{k-1}-\psi(\Theta_{k-1})\big). \end{align*} It is easy to see that $(M_n)$ is a martingale with respect to the filtration $(\mathcal{F}_{\eta+n})$. Moreover, it is bounded in $L^2$ as for all $n$ \begin{align} \label{bb1} \mathbb E_{\mathcal{F}(\eta)} M_n^2 &=\mathbb E_{\mathcal{F}(\eta)} M_{n-1}^2+ \mathbb E_{\mathcal{F}(\eta)}\Big[\frac{\sigma_{\eta+n}}{\tau_{\eta+n}^2} P_{\eta+n-1}(1-P_{\eta+n-1})\Big]\notag\\ &\le \mathbb E_{\mathcal{F}(\eta)} M_{n-1}^2+\frac{\sigma_{\eta+n}}{\tau_{\eta+n}^2} \le\cdots\le \sum_{k=\eta+1}^{\eta+n}\frac{\sigma_{k}}{\tau^2_{k}}\le\int_{\tau_{\eta}}^{\infty}\frac{dx}{x^2}=1/\tau_{\eta}. \end{align} Hence $(M_n)$ converges almost surely conditionally on $\mathcal{F}_{\eta}$. Denote \begin{align*} \mathcal{S}=\Big\{\sup_{n\in\mathbb N}M_n\le \frac{\delta_{\eta}}{2}\Big\}. \end{align*} By Doob's submartingale inequality we have using~\eqref{bb1} \begin{align*} \P_{\mathcal{F}(\eta)}\Big(\max_{1\le k\le n}M_k> \frac{\delta_{\eta}}{2}\Big) &\le \P_{\mathcal{F}(\rho)}\Big(\max_{1\le k\le n} M^2_k> \frac{\delta_{\eta}^2}{4}\Big) \le \frac{4}{\delta_{\eta}^2} \mathbb E_{\mathcal{F}(\eta)} M_n^2 \le \frac{4\log^2\tau_{\eta}}{\tau_{\eta}} \le\frac{4\log^2\tau_{r}}{\tau_{r}}, \end{align*} as the function $x\mapsto \frac{\log^2x}{x}$ is increasing on $[3,\infty)$ and $r\ge 3$. This implies \begin{align} \label{lo5} \P(\mathcal{S}^c) \le\frac{4\log^2\tau_{r}}{\tau_{r}}. \end{align} Let us prove by induction that \begin{align} \label{bb2} \Theta_{\eta+n}<\frac 1 2-\frac{\delta_{\eta}}{2} \end{align} for all $n\in\mathbb N_0$ on the event $\mathcal{S}\cap \mathcal{E}$. Indeed, for $n=0$ it follows from~\eqref{lo3}. Suppose it is true for all indices between $0$ and $n-1$. Since $x\ge \psi(x)$ for all $x\in [0,1/2]$ we have $\Theta_{k-1}\ge \psi(\Theta_{k-1})$ for all $\eta+1\le k\le \eta+n$ and hence $R_n\ge 0$. By~\eqref{lo2} this implies \begin{align*} \Theta_{\eta+n}\le \Theta_{\eta}+M_n<\frac 1 2-\delta_{\eta}+\frac{\delta_{\eta}}{2}=\frac 1 2-\frac{\delta_{\eta}}{2} \end{align*} as required. \smallskip Observe that~\eqref{bb2} and $x\ge \psi(x)$ for all $x\in [0,1/2]$ imply that $(R_n)$ is increasing on $\mathcal{S}\cap \mathcal{E}$. Since $(M_n)$ converges we obtain by~\eqref{lo2} that $(\Theta_n)$ converges on $\mathcal{S}\cap \mathcal{E}$. Let us show that on $\mathcal{S}\cap \mathcal{E}$ \begin{align} \label{lili} \lim_{n\to\infty} \Theta_{n}=0. \end{align} Indeed, if the limit $\Theta$ is positive for some $\omega\in\mathcal{S}\cap \mathcal{E}$ then $\Theta>\psi(\Theta)$ and \begin{align*} \frac{\sigma_{k}}{\tau_{k}}\big[\Theta_{k-1}-\psi(\Theta_{k-1})\big]\sim \frac{\sigma_{k}}{\tau_{k}}\big[\Theta-\psi(\Theta)\big], \end{align*} implying \begin{align*} \sum_{k=\eta+1}^{\infty}\frac{\sigma_{k}}{\tau_{k}}\big[\Theta_{k-1}-\psi(\Theta_{k-1})\big]=\infty \end{align*} by Lemma~\ref{1101series}. This means that $R_n\to\infty$, which by~\eqref{lo2} implies $\Theta= -\infty$, which is clearly impossible. \smallskip Since $r$ is arbitrary, it follows from~\eqref{lo5} that~\eqref{lili} holds on $\mathcal{E}$ almost surely. \end{proof} \bigskip \section{Some properties of $(\sigma_n)$ and $(\varepsilon_n)$} \label{s:pro} In this section we collect some elementary results about the sequence of sample sizes $(\sigma_n)$ and the sequence of random noises $(\varepsilon_n)$ that will be later used for the subcritical regime. \smallskip For each $n\in\mathbb N$, denote \begin{align} \label{d:lambdan} \lambda_n=\frac{\sigma_{n+1}\sigma_{n-1}^{\alpha}}{\sigma_n^{\alpha+1}}. \end{align} Observe that if $\rho_n\to\infty$ we have \begin{align} \label{st} \lim_{n\to\infty}\frac{\tau_n}{\sigma_n}=\lim_{n\to\infty}\Big(\frac{\tau_{n-1}}{\sigma_n}+1\Big)=1. \end{align} The following lemma explains the nature of $\lambda$ and shows that for regular enough sequences it is always equal to zero in the subcritical regime. \begin{lemma} \label{l:regular} If $\rho_n\to\infty$, $\theta=0$, and $\lambda_n\to \lambda$ then $\lambda=0$. \end{lemma} \begin{proof} Suppose $\lambda\in (0,\infty]$. It follows from~\eqref{st} that \begin{align} \label{r1} \lim_{n\to\infty}\frac{\rho_n}{\rho_{n-1}^{\alpha}} =\lim_{n\to\infty}\frac{\sigma_{n+1}\tau_{n-1}^{\alpha}}{\tau_n\sigma_n^{\alpha}} =\lim_{n\to\infty}\lambda_n=\lambda. \end{align} Let $\varepsilon\in (0,\lambda)$ and let $k\in \mathbb N$ be such that $\varepsilon^{\frac{1}{\alpha-1}}\rho_k>1$ and, for all $n> k$, $$\rho_n>\varepsilon\rho_{n-1}^{\alpha},$$ which is possible by~\eqref{r1}. Iterating, we have for all $n>k$ \begin{align*} \rho_n>\varepsilon^{1+\alpha+\cdots+\alpha^{n-k-1}}\rho_k^{\alpha^{n-k}}=\varepsilon^{\frac{\alpha^{n-k}-1}{\alpha-1}}\rho_k^{\alpha^{n-k}}. \end{align*} Using $\tau_{n+1}\ge\sigma_{n+1}=\tau_n\rho_n\ge \rho_n$ we obtain \begin{align*} 0=\alpha\theta=\lim_{n\to\infty}\alpha^{-n}\log\tau_{n+1} \ge \lim_{n\to\infty}\alpha^{-n}\log\rho_{n} \ge \alpha^{-k}\log\big(\varepsilon^{\frac{1}{\alpha-1}}\rho_k\big)>0 \end{align*} leading to a contradiction. \end{proof} Recall from Lemma~\ref{diverges} that the series~\eqref{div7} plays an important r\^ ole for non-occurrence of monopoly. In the following lemma we explore the behaviour of that series in various cases of the subcritical regime. \begin{lemma} \label{omegazero} If $(\rho_n)$ is bounded then $\theta=0$ and there is $c>0$ such that for all $m\in\mathbb N_0$ \begin{align*} \sum_{n=m}^{\infty}\frac{\sigma_{n+1}}{\tau_n^{\alpha}}\le \frac{c}{\tau_m^{\alpha-1}}. \end{align*} If $\rho_n\to\infty$, $\theta=0$, and $\lambda<1$ then \begin{align} \label{sercon} \sum_{n=0}^{\infty}\frac{\sigma_{n+1}}{\tau_n^{\alpha}}<\infty. \end{align} If $\rho_n\to\infty$, $\theta=0$, and $\lambda>1$ then \begin{align} \label{sercon1} \sum_{n=0}^{\infty}\frac{\sigma_{n+1}}{\tau_n^{\alpha}}=\infty. \end{align} \end{lemma} \begin{proof} First, suppose $(\rho_n)$ is bounded by some constant $\rho$. We have \begin{align*} \alpha^{-n}\log\tau_n=\alpha^{-n}\Big(\log\tau_0+\sum_{i=0}^{n-1}\log (1+\rho_i)\Big) \le \alpha^{-n}\big(\log\tau_0+n\log(1+\rho)\big)\to 0 \end{align*} implying $\theta=0$. Further, we have \begin{align*} \sum_{n=m}^{\infty}\frac{\sigma_{n+1}}{\tau_n^{\alpha}} =\sum_{n=m}^{\infty}\frac{(1+\rho_n)^{\alpha}\sigma_{n+1}}{\tau_{n+1}^{\alpha}} \le (1+\rho)^{\alpha}\sum_{n=m}^{\infty}\frac{\sigma_{n+1}}{\tau_{n+1}^{\alpha}} \le (1+\rho)^{\alpha}\int_{\tau_m}^{\infty}\frac{dx}{x^{\alpha}} =\frac{(1+\rho)^{\alpha}}{\alpha-1}\cdot\frac{1}{\tau_m^{\alpha-1}}. \end{align*} \smallskip Now suppose $\rho_n\to\infty$ and $\theta=0$. Observe that \begin{align} \label{ooo} \frac{\sigma_{n+1}}{\tau_n^{\alpha}} =\frac{\sigma_n}{\tau_{n-1}^{\alpha}}\cdot\frac{\rho_n}{\rho_{n-1}^{\alpha}} \Big(\frac{\rho_{n-1}}{1+\rho_{n-1}}\Big)^{\alpha-1}. \end{align} Using $\rho_n\to\infty$ and~\eqref{st} we obtain similarly to~\eqref{r1} \begin{align*} \limsup_{n\to\infty}\frac{\rho_n}{\rho_{n-1}^{\alpha}} \Big(\frac{\rho_{n-1}}{1+\rho_{n-1}}\Big)^{\alpha-1}=\lambda, \end{align*} which implies convergence~\eqref{sercon} and divergence~\eqref{sercon1} by the ratio test together with~\eqref{ooo}. \end{proof} \smallskip The next lemma gives an elementary upper bound for the noise terms $(\varepsilon_n)$. \begin{lemma} \label{le1} Almost surely, $\varepsilon_{n+1}\le n$ eventually for all $n$. \end{lemma} \begin{proof} By L\' evy's extension of the Borel-Cantelli Lemmas, see~\cite[\S 12.15]{DW}, \begin{align*} \{\varepsilon_{n+1}> n\text{ i.o.}\}=\Big\{\sum_{n=1}^{\infty}\P_{\mathcal{F}_n}(\varepsilon_{n+1}> n)=\infty\Big\}. \end{align*} By Chebychev's inequality and using~\eqref{norm} we have \begin{align*} \P_{\mathcal{F}_n}(\varepsilon_{n+1}> n)\le\P_{\mathcal{F}_n}(\varepsilon_{n+1}^2> n^2)\le \frac{1}{n^2} \end{align*} implying that the series converges almost surely. \end{proof} The final result of this section is a CLT statement about the random series of $(\varepsilon_n)$. \begin{lemma} \label{l:liminf0} Let $(\xi_n)$ be a sequence of random variables adapted to the filtration $(\mathcal{F}_n)$. Suppose \begin{align} \label{bd} |\xi_i|\le \zeta_n a_i\qquad\text{for all }i\ge n, \end{align} for all $n$ almost surely, where $(\zeta_n)$ is an almost surely positive and square-integrable $(\mathcal{F}_n)$-adapted sequence, and $(a_n)$ is a deterministic square-summable sequence. Then, almost surely, \begin{align} \label{se10} \sum_{i=0}^{\infty}\xi_i\varepsilon_{i+1}<\infty \end{align} and \begin{align} \label{se20} \liminf_{n\to\infty}\ \frac{1}{\zeta_n\sqrt{A_n}}\sum_{i=n}^{\infty}\xi_i\varepsilon_{i+1}<\infty, \end{align} where \begin{align*} A_n=\sum_{i=n}^{\infty}a_i^2. \end{align*} \end{lemma} \begin{proof} Let $n\in\mathbb N_0$. For each $m\ge n$, denote \begin{align*} S^{\ssup n}_m=\sum_{i=n}^m\xi_i\varepsilon_{i+1}. \end{align*} Observe that $S^{\ssup n}$ is a martingale bounded in $L^2$ since by~\eqref{se10} for all $m$ \begin{align} \label{mo20} \mathbb E \big(S_m^{\ssup n}\big)^2 \le \mathbb E\zeta_n^2\, \sum_{i=n}^{\infty}a_i^2, \end{align} which is finite as $(a_i)$ is summable and $\zeta_n$ is square-integrable. Hence, as $m\to\infty$, $S^{\ssup n}_m$ converges almost surely. This in particular implies~\eqref{se10}. \smallskip Observe that~\eqref{se20} is equivalent to showing that \begin{align} \label{se205} \P\Big(\bigcap_{k=1}^{\infty}\bigcup_{N=1}^{\infty}\bigcap_{n\ge N} \Big\{ \frac{1}{\zeta_n\sqrt{A_n}}\sum_{i=n}^{\infty}\xi_i\varepsilon_{i+1}>k \Big\}\Big)=0, \end{align} or, equivalently, \begin{align} \label{mo30} \lim_{k\to\infty}\lim_{N\to\infty}\P\Big(\bigcap_{n\ge N} \Big\{ \frac{1}{\zeta_n\sqrt{A_n}}\sum_{i=n}^{\infty}\xi_i\varepsilon_{i+1}>k \Big\}\Big)=0. \end{align} By Chebychev's inequality and using~\eqref{norm}, \eqref{bd} and $L^2$-boundedness of the martingales $S^{\ssup N}$ conditionally on $\mathcal{F}_{N}$ in the same way it was done in~\eqref{mo20}, we have \begin{align*} \P\Big(\bigcap_{n\ge N} &\Big\{\frac{1}{\zeta_n\sqrt{A_n}}\sum_{i=n}^{\infty}\xi_i\varepsilon_{i+1}>k\Big\}\Big) \le \P\Big(\frac{1}{\zeta_N\sqrt{A_N}}\sum_{i=N}^{\infty}\xi_i\varepsilon_{i+1}>k\Big)\\ &\le \mathbb E\,\P_{\mathcal{F}_{N}}\Big(\Big(\sum_{i=N}^{\infty}\xi_i\varepsilon_{i+1}\Big)^2> k^2\zeta_N^2A_N\Big) \le \frac{1}{k^2}\mathbb E\Big[\frac{1}{\zeta_N^2A_N}\mathbb E_{\mathcal{F}_{N}}\Big(\sum_{i=N}^{\infty}\xi_i\varepsilon_{i+1}\Big)^2\Big] \le \frac{1}{k^2}. \end{align*} Combining this with~\eqref{mo30} we obtain~\eqref{se205}. \end{proof} \bigskip \section{Subcritical regime} \label{s:sub} The aim of this section is to prove Theorem~\ref{main1}. We will split it in two propositions corresponding to the cases when $(\rho_n)$ is bounded and when it tends to infinity with $\lambda<1$. The remaining case of $\rho_n\to \infty$ with $\lambda>1$ will follow easily from the same method as in Section~\ref{s:nomo}. \smallskip It follows from~\eqref{psibounds} and~\eqref{iteration1} that for all $n\in\mathbb N_0$ \begin{align} T_{n+1} &\le T_n+2^{\alpha-1}\sigma_{n+1}\Theta_n^{\alpha} +\varepsilon_{n+1}\sqrt{\sigma_{n+1}P_n(1-P_n)}\notag\\ &= T_n+2^{\alpha-1}T_n^{\alpha}\frac{\sigma_{n+1}}{\tau_n^{\alpha}} +\varepsilon_{n+1}\sqrt{\sigma_{n+1}P_n(1-P_n)}. \label{itit} \end{align} This iterative upper bound will play an important role in showing monopoly. \smallskip The first proposition deals with the case when $(\rho_n)$ is bounded. This is an easier case as $(\sigma_n)$ is growing not too fast so that the sums \begin{align} \label{aprox} \sum_{i=n}^{\infty}\frac{\sigma_{i+1}}{\tau_i^{\alpha}}\approx \int_{\tau_n}^{\infty}\frac{dx}{x^{\alpha}}=\frac{1}{\alpha-1}\cdot \frac{1}{\tau_n^{\alpha-1}} \end{align} can be approximated accurately enough by the corresponding integrals. \begin{prop} \label{p:mod1} Suppose $\alpha>1$, {\rm (S)} is satisfied, and $(\rho_n)$ is bounded. Then $\P(\mathcal{M})=1$. \end{prop} \begin{proof} By symmetry and Theorem~\ref{main0}, without loss of generality it suffices to consider the event $\mathcal{E}=\{\Theta_n\to 0\}$ and prove that $T_n$ is bounded on $\mathcal{E}$. \smallskip Consider the event $\mathcal{E}$ and assume that $T_n\to\infty$. Then for all $n\in\mathbb N_0$ \begin{align} \label{ko1} \sum_{i=n}^{\infty}\frac{T_{i+1}-T_i}{T_i^{\alpha}}\ge \int_{T_n}^{\infty}\frac{dx}{x^{\alpha}}=\frac{1}{\alpha-1}\cdot\frac{1}{T_n^{\alpha-1}}. \end{align} On the other hand, it follows from~\eqref{itit} that \begin{align} \label{fdif1} \frac{T_{i+1}-T_i}{T_i^{\alpha}}\le 2^{\alpha-1}\frac{\sigma_{i+1}}{\tau_i^{\alpha}}+\xi_{i}\varepsilon_{i+1}, \end{align} where \begin{align*} \xi_{i}=T_i^{-\alpha}\sqrt{\sigma_{i+1}P_i(1-P_i)}. \end{align*} By~\eqref{fdif1} and Lemma~\ref{omegazero} we have with some $c>0$ \begin{align} \label{ko2} \sum_{i=n}^{\infty}\frac{T_{i+1}-T_i}{T_i^{\alpha}} \le 2^{\alpha-1}\sum_{i=n}^{\infty}\frac{\sigma_{i+1}}{\tau_i^{\alpha}} +\sum_{i=n}^{\infty}\xi_{i}\varepsilon_{i+1} \le \frac{c}{\tau_n^{\alpha-1}}+\sum_{i=n}^{\infty}\xi_{i}\varepsilon_{i+1}. \end{align} Combining this with~\eqref{ko1} we obtain \begin{align} \label{ko5} \frac{1}{\alpha-1}\cdot\frac{1}{T_n^{\alpha-1}}\le\frac{c}{\tau_n^{\alpha-1}}+ \sum_{i=n}^{\infty}\xi_{i}\varepsilon_{i+1}. \end{align} Observe that by~\eqref{psibounds} and using $T_i\ge T_n$ on the whole probability space we have \begin{align*} |\xi_{i}|\le T_i^{-\alpha}\sqrt{\sigma_{i+1}2^{\alpha-1}\Theta_i^{\alpha}} =T_i^{-\frac{\alpha}{2}}\sqrt{2^{\alpha-1}\frac{\sigma_{i+1}}{\tau_i^{\alpha}}} \le \zeta_n a_{i} \end{align*} for all $n$ and all $i\ge n$, where \begin{align*} \zeta_n=T_n^{-\frac{\alpha}{2}} \qquad\text{and}\qquad a_i=\sqrt{2^{\alpha-1}\frac{\sigma_{i+1}}{\tau_i^{\alpha}}}. \end{align*} The sequence $(\zeta_n)$ is clearly $(\mathcal{F}_n)$-adapted and square-integrable since it is bounded by one. By Lemma~\ref{omegazero} the sequence $(a_i)$ is square-summable, and \begin{align*} A_{n}=\sum_{i=n}^{\infty}a_{i}^2\le \frac{c2^{\alpha-1}}{\tau_{n}^{\alpha-1}}. \end{align*} Using this, we can rewrite and further estimate~\eqref{ko5} on $\mathcal{E}$ as \begin{align*} \frac{1}{\Theta_n^{\alpha-1}}-c(\alpha-1) &\le (\alpha-1)\tau_n^{\alpha-1}\sum_{i=n}^{\infty}\xi_{i}\varepsilon_{i+1}\\ &\le (\alpha-1)\tau_n^{\alpha-1}T_n^{-\frac{\alpha}{2}}\sqrt{\frac{c2^{\alpha-1}}{\tau_{n}^{\alpha-1}}} \Big[\frac{1}{\zeta_{n}\sqrt{A_{n}}}\sum_{i=n}^{\infty}\xi_{i}\varepsilon_{i+1}\Big]\\ &\le (\alpha-1)\sqrt{c 2^{\alpha-1}}\Theta_n^{-\frac{\alpha}{2}}\tau_n^{-\frac 1 2 } \Big[\frac{1}{\zeta_{n}\sqrt{A_{n}}}\sum_{i=n}^{\infty}\xi_{i}\varepsilon_{i+1}\Big]. \end{align*} By Lemma~\ref{l:liminf0} we obtain \begin{align*} \liminf_{n\to\infty}\Big[\frac{1}{\Theta_n^{\alpha-1}}- c(\alpha-1)\Big]\Theta_n^{\frac{\alpha}{2}}\sqrt{\tau_n} <\infty. \end{align*} Since $\Theta_n\to 0$ we obtain \begin{align} \label{c1} \liminf\limits_{n\to\infty}\Theta_n^{1-\frac{\alpha}{2}}\sqrt{\tau_n}<\infty. \end{align} If $\alpha\ge 2$ we get a contradiction as $\Theta_n\le 1$. If $1<\alpha<2$ then~\eqref{c1} implies \begin{align*} \liminf\limits_{n\to\infty}T_n\tau_n^{\frac{\alpha-1}{2-\alpha}}=0, \end{align*} and we get a contradiction as $T_n\ge 1$ and hence the value should be infinite. \end{proof} Now we turn to the case when $\rho_n\to\infty$, and the approximation~\eqref{aprox} is no longer valid. Instead, we will look at the iterations~\eqref{itit} in more detail. \begin{prop} \label{p:fast} Suppose $\alpha>1$, $\theta=0$, {\rm (S)} is satisfied, $\rho_n\to\infty$, and $\lambda<1$. Then $\P(\mathcal{M})=1.$ \end{prop} \begin{proof} By symmetry and Theorem~\ref{main0}, without loss of generality it suffices to consider the event \begin{align*} \mathcal{E}=\{\Theta_n\to 0\}\cap \{T_n\to\infty\} \end{align*} and prove that it has probability zero. \smallskip We begin by defining a small parameter $\delta$ in the following way. Let $q\in (\lambda,1)$. Let $\varepsilon>0$ be small enough so that \begin{align} \label{eeqq} (\lambda+\varepsilon)(1+2\varepsilon)<q. \end{align} Let $\gamma\in(\max\{0,\alpha-2\},\alpha-1)$ and let $\delta>0$ be small enough so that \begin{align} \label{ddd} \delta 2^{\alpha-1}<\varepsilon \end{align} and \begin{align} \label{ddd2} 2^{\frac{\alpha-1}{2}} \delta^{\frac{\gamma}{2(\alpha-1)}} \max_{m\in\mathbb N_0}\big[(m+1)q^{\frac{\gamma m}{2(\alpha-1)}}\big]<\varepsilon. \end{align} Further, we denote $\kappa_0=0$, and for each $n\in \mathbb N$, define the stopping times \begin{align*} \kappa_n=\inf\Big\{i>\kappa_{n-1}: \frac{\sigma_{i+1}}{\tau_i^{\alpha}}T_i^{\alpha}\le \delta T_i\Big\} \end{align*} to be the subsequent times when the second term in the iteration~\eqref{itit} is $\delta$-smaller than the first term. \smallskip The proof of the theorem now consist of the following four steps. First, we show that on $\mathcal{E}$ the $\delta$-negligibility of the second term occurs infinitely often, that is, all $\kappa_n$ are finite. Then we provide an upper bound for $\varepsilon_i, i\ge \kappa_n$, as a function of both $n$ and $i$. Using that bound, we observe that we can find a stopping time $\kappa_{\nu}$ such that for all subsequent times $n\ge \kappa_{\nu}$ the second term in~\eqref{itit} will be $\delta q^{n-\kappa_{\nu}}$-smaller than the first term. This much stronger domination of the first term finally allows us to show that $(T_n)$ is bounded on $\mathcal{E}$, which leads to a contradiction. \bigskip \emph{Step 1.} Let us prove that $\kappa_n<\infty$ almost surely for all $n$ on the event $\mathcal{E}$. Suppose this is not the case, that is, \begin{align*} \bar n=\sup\Big\{i\in \mathbb N:\frac{\sigma_{i+1}}{\tau_i^{\alpha}}T_i^{\alpha}\le \delta T_i\Big\} \end{align*} is finite on $\mathcal{E}$ with positive probability. Roughly speaking, this means that in the iteration~\eqref{itit} the second term plays the main r\^ ole eventually, and (ignoring the other terms and constants for the moment) for all $n>k>\hat n$ we have approximately \begin{align*} T_n &\preceq T_{n-1}^{\alpha}\frac{\sigma_n}{\tau_{n-1}^{\alpha}} \preceq T_k^{\alpha^{n-k}} \frac{\sigma_{n}}{\tau_{n-1}^{\alpha}} \Big(\frac{\sigma_{n-1}}{\tau_{n-2}^{\alpha}}\Big)^{\alpha} \cdots \Big(\frac{\sigma_{k+1}}{\tau_k^{\alpha}}\Big)^{\alpha^{n-k-1}} \le \tau_n\Theta_k^{\alpha^{n-k}}, \end{align*} where $\preceq$ stands for ``approximately less''. Since $\theta=0$ this would lead to \begin{align*} \limsup_{n\to\infty}\alpha^{-n}\log T_n \preceq \alpha^{-k}\log \Theta_k<0 \end{align*} for sufficiently large $k$ since $\Theta_k\to 0$ on $\mathcal{E}$. This, however, would be a contradiction to $T_n\to 0$. \smallskip Now we will make this argument precise. On the event $\{\bar n<\infty\}\cap \mathcal{E}$ by~\eqref{itit} we have for all $n\ge \bar n$ \begin{align*} T_{n+1} &\le \big(\delta^{-1}+2^{\alpha-1}\big)T_{n}^{\alpha}\frac{\sigma_{n+1}}{\tau_{n}^{\alpha}} +\varepsilon_{n+1}\sqrt{T_n^{\alpha}\frac{\sigma_{n+1}}{\tau_n^{\alpha}}(1-P_{n})} \le cT_{n}^{\alpha}\frac{\sigma_{n+1}}{\tau_{n}^{\alpha}}(1 +\hat\xi_n\varepsilon_{n+1}). \end{align*} where $c=\delta^{-1}+2^{\alpha-1}$ and \begin{align*} \hat \xi_n c^{-1}\Big[T_n^{\alpha}\frac{\sigma_{n+1}}{\tau_{n}^{\alpha}}\Big]^{-\frac{1}{2}}\sqrt{1-P_{n}} =c^{-1}\Big[\max\Big\{T_n^{\alpha}\frac{\sigma_{n+1}}{\tau_{n}^{\alpha}},\delta T_n\Big\}\Big]^{-\frac{1}{2}}\sqrt{1-P_{n}}. \end{align*} Iterating and using $\sigma_i\le\tau_i$ and $\log(1+x)\le x$ we obtain for all $k>\bar n$ and all $n>k$ \begin{align*} T_n &\le c^{1+\alpha+\cdots+\alpha^{n-k-1}}T_k^{\alpha^{n-k}} \frac{\sigma_{n}}{\tau_{n-1}^{\alpha}} \Big(\frac{\sigma_{n-1}}{\tau_{n-2}^{\alpha}}\Big)^{\alpha} \cdots \Big(\frac{\sigma_{k+1}}{\tau_k^{\alpha}}\Big)^{\alpha^{n-k-1}} \prod_{i=k}^{n-1} (1+\hat\xi_i\varepsilon_{i+1})^{\alpha^{n-i-1}}\\ &\le \big(c^{\frac{1}{\alpha-1}}\big)^{\alpha^{n-k}}\tau_n\Theta_k^{\alpha^{n-k}} \prod_{i=k}^{n-1} (1+\hat\xi_i\varepsilon_{i+1})^{\alpha^{n-i-1}}\\ &\le \tau_n\Big[c^{\frac{1}{\alpha-1}}\Theta_k\exp\Big\{ \sum_{i=k}^{n-1} \alpha^{k-i-1}\log(1 +\hat\xi_i\varepsilon_{i+1}]\Big\} \Big]^{\alpha^{n-k}}\\ &\le \tau_n\Big[c^{\frac{1}{\alpha-1}}\Theta_k\exp\Big\{\alpha^{k-1} \sum_{i=k}^{n-1} \xi_i\varepsilon_{i+1}\Big\} \Big]^{\alpha^{n-k}}, \end{align*} where \begin{align*} \xi_i=\alpha^{-i}\hat\xi_i. \end{align*} As $\theta=0$ this implies, for all $k\ge \bar n$, \begin{align} \label{sub} \limsup_{n\to\infty}\alpha^{-n}\log T_n \le \alpha^{-k}\log \Big[c^{\frac{1}{\alpha-1}}\Theta_k\exp\Big\{\alpha^{k-1} \sum_{i=k}^{\infty} \xi_i\varepsilon_{i+1}\Big\} \Big]. \end{align} It suffices to show that on the event $\mathcal{E}$ the expression on the right hand side is negative for some $k$ (depending on $\omega\in \mathcal{E}$), as it would imply that $T_n\to 0$ contradicting $T_n\ge 1$. Since $\Theta_k\to 0$ on $\mathcal{E}$, it amounts to proving that \begin{align} \label{sub9} \liminf_{k\to\infty}\Big[\alpha^{k-1} \sum_{i=k}^{\infty} \xi_i\varepsilon_{i+1}\Big]<\infty. \end{align} To show that this holds almost surely we use Lemma~\ref{l:liminf0}. Observe that since $T_i\ge 1$, on the whole probability space we have \begin{align*} |\xi_i|\le \alpha^{-i}|\hat \xi_i| \le \frac{\alpha^{-i}}{c\sqrt {\delta T_i}} \le \frac{\alpha^{-i}}{c\sqrt {\delta}}, \end{align*} for all $i\ge n$ and all $n$. With \begin{align*} A_n=\frac{1}{\delta c^2}\sum_{i=n}^{\infty}\alpha^{-2i}= \frac{\alpha^{-2n+2}}{\delta c^2(\alpha^2-1)} \end{align*} we have by Lemma~\ref{l:liminf0} \begin{align*} \liminf_{k\to\infty}\Big[\alpha^{k-1} \sum_{i=k}^{\infty} \xi_i\varepsilon_{i+1}\Big] &=\frac{1}{\sqrt{\delta c^2(\alpha^2-1)}}\liminf_{k\to\infty}\Big[\frac{1}{\sqrt{A_{k}}}\sum_{i=k}^{\infty} \xi_i\varepsilon_{i+1}\Big]<\infty \end{align*} as required in~\eqref{sub9}. \bigskip \emph{Step 2.} Let $(c_n)$ be a real-valued sequence tending to infinity. For each $n$, consider the event \begin{align*} \mathcal{E}_n=\big\{\kappa_n<\infty\text{ and }\varepsilon_{i}\le c_{i} (i-\kappa_n)\text{ for all }i>\kappa_n\big\}\cup\{\kappa_n=\infty\}. \end{align*} The aim of this step is to show that \begin{align} \label{io} \P(\mathcal{E}_n\text{ i.o.})=1, \end{align} for which it suffices to prove that \begin{align} \label{eqone} \lim_{n\to\infty}\P(\mathcal{E}_n)=1. \end{align} Using Chebychev's inequality, \eqref{norm}, and monotone convergence theorem we have \begin{align*} \P(\mathcal{E}_n) &=\lim_{j\to\infty}\P\big(\kappa_n<\infty\text{ and }\varepsilon_{i}\le c_{i} (i-\kappa_n)\text{ for all } \kappa_n< i\le \kappa_n+j\big)+\P(\kappa_n=\infty)\\ &\ge \lim_{j\to\infty}\mathbb E\Big[{\mathbf 1}\{\kappa_n<\infty\}\prod_{i=\kappa_n+1}^{\kappa_n+j-1}{\mathbf 1}\big\{\varepsilon_{i}\le c_{i} (i-\kappa_n)\big\}\cdot \Big(1-\P_{\mathcal{F}_{\kappa_n+j-1}}\big(\varepsilon_{\kappa_n+j}^2> c_{\kappa_n+j}^2 j^2\big)\Big)\Big]\\ &\phantom{aaall}+\P(\kappa_n=\infty)\\ &\ge \lim_{j\to\infty}\mathbb E\Big[{\mathbf 1}\{\kappa_n<\infty\}\prod_{i=\kappa_n+1}^{\kappa_n+j-1}{\mathbf 1}\big\{\varepsilon_{i}\le c_{i} (i-\kappa_n)\big\}\cdot\Big(1-\frac{1}{c_{\kappa_n+j}^2j^2}\Big)\Big]+\P(\kappa_n=\infty)\ge \cdots\\ &\ge\lim_{j\to\infty}\mathbb E\Big[{\mathbf 1}\{\kappa_n<\infty\}\prod_{i=\kappa_n+1}^{\kappa_n+j}\Big(1-\frac{1}{c_{i}^2(i-\kappa_n)^2}\Big)\Big]+\P(\kappa_n=\infty)\\ &=\mathbb E\Big[{\mathbf 1}\{\kappa_n<\infty\}\exp\Big\{\sum_{i=1}^{\infty}\log\Big(1-\frac{1}{c_{\kappa_n+i}^2i^2}\Big)\Big\}\Big]+\P(\kappa_n=\infty). \end{align*} Let $\hat c_n= \min \{c_i:i\ge n\}$ and observe that $\hat c_n\to\infty$. Hence for all sufficiently large $n$ we can estimate \begin{align*} \log(1-x)\ge-x-x^2 \end{align*} for all $x\in [0,\hat c_n^{-2}]$. Since $\kappa_n+i\ge n$ for all $n$ and $i$ we have $c_{\kappa_n+i}\ge \hat c_n$. Hence for all sufficiently large $n$ we obtain \begin{align*} \P(\mathcal{E}_n) &\ge \mathbb E\Big[{\mathbf 1}\{\kappa_n<\infty\}\exp\Big\{-\frac{1}{\hat c_{n}^2}\sum_{i=1}^{\infty}\frac{1}{i^2}-\frac{1}{\hat c_{n}^4}\sum_{i=1}^{\infty}\frac{1}{i^4}\Big\}\Big]+\P(\kappa_n=\infty)\to 1 \end{align*} as $n\to\infty$, implying~\eqref{eqone}. \bigskip \emph{Step 3.} Let us show that on the event $\mathcal{E}$ there is $\nu$ (depending on $\omega\in \mathcal{E}$) such that for all $n\ge \kappa_{\nu}$ \begin{align} \label{easy} \varepsilon_n\le n \end{align} and \begin{align} \label{ind1} \frac{\sigma_{n+1}}{\tau_{n}^{\alpha}}T_{n}^{\alpha-1}\le \delta q^{n-\kappa_{\nu}}. \end{align} Roughly speaking, we know that the first term of~\eqref{itit} dominates at time $\kappa_{n}$. Ignoring the remaining terms and taking into account~\eqref{st} we can write for $n>\kappa_{\nu}$ approximately \begin{align*} \frac{\sigma_{n+1}}{\tau_n^{\alpha}}T_n^{\alpha-1}\preceq \frac{\sigma_{n+1}}{\tau_n^{\alpha}}T_{n-1}^{\alpha-1} =\frac{\sigma_{n+1}\tau_{n-1}^{\alpha}}{\sigma_n\tau_n^{\alpha}}\cdot \frac{\sigma_{n}}{\tau_{n-1}^{\alpha}}T_n^{\alpha-1} \approx \lambda\cdot \frac{\sigma_{n}}{\tau_{n-1}^{\alpha}}T_n^{\alpha-1}, \end{align*} which demonstrates eventual exponential decay of these terms as pinpointed in~\eqref{ind1}. \smallskip To make this rigorous and prove the existence of $\nu$ satisfying~\eqref{easy} and~\eqref{ind1}, we define a sequence \begin{align*} c_n=\Big[\frac{\sigma_n}{\tau_{n-1}^{\alpha}}\Big]^{\frac{\gamma}{2(\alpha-1)}-\frac 1 2}, \quad n\in\mathbb N, \end{align*} and observe that $c_n\to \infty$ by Lemma~\ref{omegazero} and since $\gamma<\alpha-1$. Further, it follows from~\eqref{d:lambdan} and~\eqref{st} that \begin{align} \label{ls} \limsup_{n\to\infty}\frac{\sigma_{n+1}\tau_{n-1}^{\alpha}}{\sigma_{n}\tau_{n}^{\alpha}} =\limsup_{n\to\infty}\lambda_n=\lambda<1. \end{align} According to Steps 1 and 2, by Lemma~\ref{le1}, and by~\eqref{ls} we can pick $\nu$ so that $\kappa_{\nu}<\infty$, $\mathcal{E}_{\nu}$ occurs, \eqref{easy} holds, and for all $n>\kappa_{\nu}$ \begin{align} \label{leps} \frac{\sigma_{n+1}\tau_{n-1}^{\alpha}}{\sigma_{n}\tau_{n}^{\alpha}}<\lambda+\varepsilon. \end{align} Let us now prove~\eqref{ind1} by induction in $n$. It is true for $n=\kappa_{\nu}$ by definition of $\kappa_{\nu}$. For $n>\kappa_{\nu}$ using~\eqref{itit}, \eqref{psibounds}, and the fact that $\mathcal{E}_{\nu}$ occurs we obtain \begin{align} \frac{\sigma_{n+1}}{\tau_n^{\alpha}}T_n^{\alpha-1} &\le \frac{\sigma_{n+1}}{\tau_n^{\alpha}}\Big( T_{n-1}+2^{\alpha-1}T_{n-1}^{\alpha}\frac{\sigma_{n}}{\tau_{n-1}^{\alpha}} +\varepsilon_{n}\sqrt{\sigma_{n}P_{n-1}(1-P_{n-1})} \Big)^{\alpha-1}\notag\\ &\le \frac{\sigma_{n+1}}{\tau_n^{\alpha}} \Big(T_{n-1}+2^{\alpha-1}T_{n-1}^{\alpha}\frac{\sigma_n}{\tau_{n-1}^{\alpha}} +c_n (n-\kappa_{\nu})\sqrt{2^{\alpha-1}T_{n-1}^{\alpha}\frac{\sigma_n}{\tau_{n-1}^{\alpha}}}\Big)^{\alpha-1}\notag\\ &\le \frac{\sigma_{n+1}}{\tau_n^{\alpha}}T_{n-1}^{\alpha-1} \Big(1+2^{\alpha-1}T_{n-1}^{\alpha-1}\frac{\sigma_n}{\tau_{n-1}^{\alpha}} +c_n (n-\kappa_{\nu})\sqrt{2^{\alpha-1}T_{n-1}^{\alpha-2}\frac{\sigma_n}{\tau_{n-1}^{\alpha}}}\Big)^{\alpha-1}. \label{eps3} \end{align} Using~\eqref{ind1} for $n-1$ we have \begin{align} \label{eps12} \frac{\sigma_{n+1}}{\tau_{n}^{\alpha}}T_{n-1}^{\alpha-1} =\delta q^{n-\kappa_{\nu}-1}\frac{\sigma_{n+1}\tau_{n-1}^{\alpha}}{\tau_{n}^{\alpha}\sigma_n}. \end{align} Using~\eqref{ind1} for $n-1$, $q<1$, and~\eqref{ddd} we have \begin{align} \label{eps1} 2^{\alpha-1}T_{n-1}^{\alpha-1}\frac{\sigma_n}{\tau_{n-1}^{\alpha}} \le 2^{\alpha-1}\delta q^{n-\kappa_{\nu}-1}\le \delta 2^{\alpha-1}<\varepsilon. \end{align} Using~\eqref{ind1} for $n-1$, $T_{n-1}\ge 1$, and $\gamma>\alpha-2$ we estimate \begin{align*} T_{n-1}^{\alpha-2}\frac{\sigma_n}{\tau_{n-1}^{\alpha}} <T_{n-1}^{\gamma}\frac{\sigma_n}{\tau_{n-1}^{\alpha}} \le (\delta q^{n-\kappa_{\nu}-1})^{\frac{\gamma}{\alpha-1}} \Big[\frac{\sigma_n}{\tau_{n-1}^{\alpha}}\Big]^{1-\frac{\gamma}{\alpha-1}} \end{align*} and obtain by~\eqref{ddd2} and by the choice of $(c_n)$ \begin{align} \label{eps2} c_n(n-\kappa_{\nu})\sqrt{2^{\alpha-1}T_{n-1}^{\alpha-2}\frac{\sigma_n}{\tau_{n-1}^{\alpha}}} &\le 2^{\frac{\alpha-1}{2}}(n-\kappa_{\nu}) (\delta q^{n-\kappa_{\nu}-1})^{\frac{\gamma}{2(\alpha-1)}} <\varepsilon. \end{align} Substituting~\eqref{eps12}, \eqref{eps1}, and~\eqref{eps2} into~\eqref{eps3} as well as using~\eqref{leps} and~\eqref{eeqq} we obtain \begin{align*} \frac{\sigma_{n+1}}{\tau_n^{\alpha}}T_n^{\alpha-1} &\le \delta q^{n-\kappa_{\nu}-1}\frac{\sigma_{n+1}\tau_{n-1}^{\alpha}}{\tau_n^{\alpha}\sigma_n}(1+2\varepsilon) < \delta q^{n-\kappa_{\nu}-1}(\lambda+\varepsilon)(1+2\varepsilon) <\delta q^{n-\kappa_{\nu}}, \end{align*} completing the induction. \bigskip \emph{Step 4.} Finally, let us prove that $\P(\mathcal{E})=0$. For all $n\ge \kappa_{\nu}$ we have on the event $\mathcal{E}$ using~\eqref{itit}, \eqref{psibounds}, \eqref{easy},~\eqref{ind1}, and $T_{n-1}\ge 1$ \begin{align*} T_n &\le T_{n-1}\Big(1+2^{\alpha-1}T_{n-1}^{\alpha-1}\frac{\sigma_{n}}{\tau_{n-1}^{\alpha}}+n T_{n-1}^{-\frac 1 2}\sqrt{2^{\alpha-1}T_{n-1}^{\alpha-1}\frac{\sigma_{n}}{\tau_{n-1}^{\alpha}}}\Big)\\ &\le T_{n-1}\big(1+2^{\alpha-1}\delta q^{n-\kappa_{\nu}-1}+n\sqrt{2^{\alpha-1}\delta q^{n-\kappa_{\nu}-1}}\big)\le \cdots\\ &\le T_{\kappa_{\nu}}\prod_{i=1}^{n-\kappa_{\nu}}\Big(1+2^{\alpha-1}\delta q^{i-1}+(\kappa_{\nu}+i)\sqrt{2^{\alpha-1}\delta q^{i-1}}\Big)\\ &\le T_{\kappa_{\nu}}\prod_{i=1}^{\infty}\Big(1+2^{\alpha-1}\delta q^{i-1}+(\kappa_{\nu}+i)\sqrt{2^{\alpha-1}\delta q^{i-1}}\Big). \end{align*} As the infinite product converges, this implies that $T_n$ is bounded thus contradicting to the definition of the event $\mathcal{E}$. \end{proof} \bigskip \begin{proof}[Proof of Theorem~\ref{main1}] If $(\rho_n)$ is bounded the statement follows from Proposition~\ref{p:mod1}. If $\rho_n\to\infty$ and $\lambda<1$ the result follows from Proposition~\ref{p:fast}. If $\rho_n\to\infty$ and $\lambda>1$ it suffices to apply Lemmas~\ref{omegazero} and~\ref{diverges}. Finally, the last statement follows from Lemma~\ref{l:regular}. \end{proof} \smallskip \section{Critical regime} \label{s:cri} The aim of this section is to prove Theorem~\ref{main3}. For all $n$, denote \begin{align*} \phi_n=\tau_n e^{-\theta \alpha^n}. \end{align*} It is easy to see that \begin{align} \label{sec} \lim_{n\to\infty}\alpha^{-n}\log \phi_n=0 \end{align} and \begin{align} \label{imply} \frac{\tau_{n+1}}{\tau_n^{\alpha}}=\frac{\phi_{n+1}}{\phi_n^{\alpha}}, \quad n\in \mathbb N_0 \qquad\text{implying}\qquad \sum_{n=0}^{\infty}\frac{\tau_{n+1}}{\tau_n^{\alpha}}<\infty \quad\Leftrightarrow\quad \sum_{n=0}^{\infty}\frac{\tau_{n+1}}{\tau_n^{\alpha}}<\infty. \end{align} In the case when the above series diverge we will show that no monopoly occurs using the same method as in Section~\ref{s:nomo}. Otherwise, if the series converge, we will explicitly construct two non-trivial events such that on one of them we have monopoly and on the other one we don't. \begin{lemma} \label{l:lambda} If \begin{align*} \sum_{n=0}^{\infty}\frac{\phi_{n+1}}{\phi_n^{\alpha}}<\infty \end{align*} then $(\phi_n)$ is unbounded. \end{lemma} \begin{proof} Suppose $(\phi_n)$ is bounded by some $c<\infty$. Let $\varepsilon>0$ be such that $c\varepsilon^{\frac{1}{\alpha-1}}<1$ and let $k\in\mathbb N_0$ be such that for all $n\ge k$ \begin{align*} \phi_{n+1}\le \varepsilon\phi_n^{\alpha}, \end{align*} which is possible since the series converges. Iterating this inequality we obtain for all $n\ge k$ \begin{align*} \phi_n\le \varepsilon^{1+\alpha+\cdots+\alpha^{n-k-1}}\phi_k^{\alpha^{n-k}}=\varepsilon^{\frac{\alpha^{n-k}-1}{\alpha-1}}\phi_k^{\alpha^{n-k}}. \end{align*} This implies \begin{align*} \lim_{n\to\infty}\alpha^{-n}\log\phi_n =\alpha^{-k}\log\big(\phi_k\varepsilon^{\frac{1}{\alpha-1}}\big) \le\alpha^{-k}\log\big(c\varepsilon^{\frac{1}{\alpha-1}}\big)<0 \end{align*} contradicting~\eqref{sec}. \end{proof} \begin{prop} \label{p:cri1} Suppose $\alpha>1$ and $\theta\in (0,\infty)$. Then $\P(\mathcal{M})<1$. Further, if \begin{align} \label{s8} \sum_{n=0}^{\infty}\frac{\phi_{n+1}}{\phi_n^{\alpha}}=\infty \end{align} then $\P(\mathcal{M})=0$. \end{prop} \begin{proof} First, suppose that the series~\eqref{s8} diverges. By~\eqref{sec} we have \begin{align*} \frac{\sigma_{n+1}}{\tau_n^{\alpha}} =\frac{\tau_{n+1}-\tau_n}{\tau_n} =\frac{\phi_{n+1}-\phi_n e^{-\theta(\alpha-1)\alpha^n}}{\phi_n^{\alpha}}\sim\frac{\phi_{n+1}}{\phi_n^{\alpha}} \end{align*} as $n\to\infty$. Hence by~\eqref{s8} we have \begin{align} \label{div} \sum_{n=0}^{\infty}\frac{\sigma_{n+1}}{\tau_n^{\alpha}}=\infty \end{align} and $\P(\mathcal{M})=0$ by Lemma~\ref{diverges}. \medskip It remains to consider the case when the series on the left-hand side of~\eqref{s8} converges and show that $\P(\mathcal{M})<1$. To do so we will construct an event $\mathcal{E}$ such that $\P(\mathcal{E})>0$ and both $T_n$ and $\hat T_n=\tau_n-T_n$ tend to infinity on $\mathcal{E}$. \smallskip Let $\gamma>\frac{2}{\alpha-1}$ and, for all $n\in\mathbb N$, denote \begin{align} \label{defchi} \chi_n=\max_{1\le k\le n} k^{\gamma}\phi_k. \end{align} By Lemma~\ref{l:lambda} $\chi_n\to\infty$. Let $\beta\in (\frac{2+\gamma}{\alpha},\gamma)$. Choose $m\in\mathbb N$ large enough so that \begin{align} \label{he10} \tau_m-2\chi_m>1, \end{align} \begin{align} \label{he9} T_0\le \tau_m-\chi_m \qquad\text{and}\qquad T_0+\tau_m-\tau_0\ge \chi_m, \end{align} as well as \begin{align} \label{he1} \big(\phi_{n+1}-\phi_ne^{-\theta(\alpha-1)\alpha^{n}}\big)n^{\alpha\gamma}&> \phi_{n+1}n^{\alpha\beta},\\ \label{he1b} \chi_12^{-\gamma} n^{\alpha\beta-\gamma}&>n^2,\\ \label{he6} n^{\alpha\beta}\Big(1-\Big[\frac{2^{\gamma}n^{2-\alpha\beta+\gamma}}{\chi_1}\Big]^{1/2}\Big) &> (n+1)^{\gamma}, \end{align} for all $n\ge m$. Indeed, \eqref{he10} and \eqref{he9} are possible by~\eqref{sec} since $(\tau_n)$ grows faster than $(\chi_n)$, \eqref{he1} is possible since $\gamma>\beta$ and again by~\eqref{sec}, \eqref{he1b} is possible since $\alpha\beta-\gamma>2$, and~\eqref{he6} is possible as $2-\alpha\beta+\gamma<0$ and $\alpha\beta>\gamma$. Observe that the event \begin{align*} \mathcal{E}_0=\big\{T_m\in [\chi_m, \tau_m-\chi_m]\big\} \end{align*} occurs with positive probability by~\eqref{he10} and~\eqref{he9} since $T_m$ takes each value between $T_0$ and $T_0+\tau_m-\tau_0$ with positive probability. Let \begin{align*} \mathcal{E}=\mathcal{E}_0\cap\big\{|\varepsilon_{n+1}|\le n\text{ for all }n\ge m\big\}. \end{align*} By Chebychev's inequality and using~\eqref{norm} we obtain \begin{align*} \P(\mathcal{E}) &=\lim_{n\to\infty}\mathbb E\Big[{\mathbf 1}\{\mathcal{E}_0\}{\mathbf 1}\big\{|\varepsilon_{i+1}|\le i\text{ for all }m\le i<n\big\}\big(1-\P_{\mathcal{F}_{n}}(\varepsilon_{n+1}^2>n^2)\big)\Big]\notag\\ &\ge \lim_{n\to\infty}\mathbb E\Big[{\mathbf 1}\{\mathcal{E}_0\}{\mathbf 1}\big\{|\varepsilon_{i+1}|\le i\text{ for all }1\le i< n\big\}\Big(1-\frac{1}{n^2}\Big)\Big] \ge\cdots\notag\\ &\ge \P(\mathcal{E}_0)\prod_{n=m}^{\infty}\Big(1-\frac{1}{n^2}\Big)>0 \end{align*} since the infinite product converges. \smallskip Due to the symmetry of the relations \begin{align} \label{he4} T_{n+1}=T_n+\sigma_{n+1}\psi(\Theta_n)+\varepsilon_{n+1}\sqrt{\sigma_{n+1}\psi(\Theta_n)(1-\psi(\Theta_n))} \end{align} and \begin{align*} \hat T_{n+1}=\hat T_n+\sigma_{n+1}\psi(\hat \Theta_n)+\hat \varepsilon_{n+1}\sqrt{\sigma_{n+1}\psi(\hat \Theta_n)(1-\psi(\hat\Theta_n))}, \end{align*} where $\hat\Theta_n=\hat T_n/\tau_n$ and $\hat \varepsilon_n=-\varepsilon_n$, and since $\hat T_m$ satisfies the condition $\hat T_m\in [\chi_m, \tau_m-\chi_m]$ in the same way as $T_m$ on $\mathcal{E}_0$, we only need to prove that $T_n\to\infty$ on $\mathcal{E}$. \smallskip Since $\chi_n\to\infty$, it suffices to show that $T_n\ge \chi_n$ on $\mathcal{E}$ for all $n\ge m$, which we prove by induction. For $n=m$ this follows from the definition of $\mathcal{E}_0$. Let $n\ge m$. If $\chi_{n+1}=\chi_n$ then we have $T_{n+1}\ge T_n\ge \chi_n=\chi_{n+1}$ as required. If $\chi_{n+1}>\chi_n$ then \begin{align} \label{222} \phi_{n+1}>\frac{\chi_n}{(n+1)^\gamma}\ge \frac{\chi_1}{2^{\gamma}n^{\gamma}}. \end{align} It follows from~\eqref{he4} that on $\mathcal{E}$ \begin{align} T_{n+1} &\ge \sigma_{n+1}\psi(\Theta_n) -n\sqrt{\sigma_{n+1}\psi(\Theta_n)}. \label{he2} \end{align} Using~\eqref{psibounds}, the induction hypothesis $T_n\ge \chi_n$, $\chi_n\ge n^{\gamma}\phi_n$ following from the definition~\eqref{defchi}, and~\eqref{he1} we obtain \begin{align} \sigma_{n+1}\psi(\Theta_n) &\ge \sigma_{n+1}\Theta_n^{\alpha} = \frac{\tau_{n+1}-\tau_n}{\tau_n^{\alpha}}\,T_n^{\alpha} \ge\frac{\phi_{n+1}e^{\theta\alpha^{n+1}}-\phi_ne^{\theta\alpha^{n}}}{\phi_n^{\alpha}e^{\theta\alpha^{n+1}}}\chi_n^{\alpha}\notag\\ &\ge \big(\phi_{n+1}-\phi_ne^{-\theta(\alpha-1)\alpha^{n}}\big)n^{\alpha\gamma}>\phi_{n+1}n^{\alpha\beta}. \label{he3} \end{align} Since the function $x\mapsto x-n\sqrt x$ is increasing on $[n^2,\infty)$ \begin{align*} \phi_{n+1}n^{\alpha\beta}\ge \chi_12^{-\gamma} n^{\alpha\beta-\gamma}>n^2 \end{align*} by~\eqref{222} and \eqref{he1b}, we obtain by~\eqref{he2}, \eqref{he3}, \eqref{222}, and~\eqref{he6} that \begin{align*} T_{n+1} &\ge n^{\alpha\beta}\phi_{n+1}-n^{\frac{\alpha\beta}{2}+1}\sqrt{\phi_{n+1}} \ge n^{\alpha\beta}\phi_{n+1}\Big(1-\Big[\frac{n^{2-\alpha\beta}}{\phi_{n+1}}\Big]^{1/2}\Big)\\ &\ge n^{\alpha\beta}\phi_{n+1}\Big(1-\Big[\frac{2^{\gamma}n^{2-\alpha\beta+\gamma}}{\chi_1 }\Big]^{\frac 1 2}\Big) \ge (n+1)^{\gamma}\phi_{n+1}=\chi_{n+1} \end{align*} as required. \end{proof} \begin{prop} \label{p:cri2} Suppose $\alpha>1$ and $\theta\in (0,\infty)$. If \begin{align} \label{coco} \sum_{n=0}^{\infty}\frac{\phi_{n+1}}{\phi_n^{\alpha}}<\infty \end{align} then $\P(\mathcal{M})>0$. \end{prop} \begin{proof} For all $n\in\mathbb N_0$ it follows from~\eqref{psibounds}, \eqref{iteration1}, and $\sigma_{n+1}\le \tau_{n+1}$ that \begin{align} T_{n+1} &\le T_n+\tau_{n+1}2^{\alpha-1}\Theta_n^{\alpha}+\varepsilon_{n+1}\sqrt{\sigma_{n+1}P_n(1-P_n)}\notag\\ &\le T_n+2^{\alpha-1}T_n^{\alpha}\frac{\phi_{n+1}}{\phi_n^{\alpha}} +\varepsilon_{n+1}\sqrt{\sigma_{n+1}P_n(1-P_n)}, \label{itit7} \end{align} implying \begin{align} \label{fdif17} \frac{T_{n+1}-T_n}{T_n^{\alpha}}\le 2^{\alpha-1}\frac{\phi_{n+1}}{\phi_n^{\alpha}}+\xi_{n}\varepsilon_{n+1}, \end{align} where \begin{align*} \xi_{n}=T_n^{-\alpha}\sqrt{\sigma_{n+1}P_n(1-P_n)}. \end{align*} As the series~\eqref{coco} converges we can pick $m$ large enough so that \begin{align} \label{m8} 2^{\alpha-1}\sum_{n=m}^{\infty}\frac{\phi_{n+1}}{\phi_n^{\alpha}}<\frac{1}{\alpha-1}\cdot \frac{1}{T_0^{\alpha-1}}. \end{align} Consider the event \begin{align*} \mathcal{E}=\Big\{2^{\alpha-1}\sum_{n=m}^{\infty}\frac{\phi_{n+1}}{\phi_n^{\alpha}}<\frac{1}{\alpha-1}\cdot\frac{1}{T_m^{\alpha-1}}\text{ \rm and } \sum_{n=m}^{\infty}\xi_{n}\varepsilon_{n+1} \le 0\Big\}. \end{align*} We have $\P(\mathcal{E})>0$ by~\eqref{m8}, as the first condition can be achieved by $B_1=\cdots=B_m=0$ occurring with positive probability, and since the conditional expectation of the second series given $\mathcal{F}_m$ is zero almost surely. Let us show that $\mathcal{E}\subset \mathcal{M}$. Suppose $T_n\to\infty$ for some $\omega\in \mathcal{E}$. We have by~\eqref{fdif17} and the second condition defining the event $\mathcal{E}$ \begin{align*} 2^{\alpha-1}\sum_{n=m}^{\infty}\frac{\phi_{n+1}}{\phi_n^{\alpha}}\ge \sum_{n=m}^{\infty}\frac{T_{n+1}-T_n}{T_n^{\alpha}} \ge \int_{T_m}^{\infty}\frac{dx}{x^{\alpha}}=\frac{1}{\alpha-1}\cdot\frac{1}{T_m^{\alpha-1}} \end{align*} leading to a contradiction with the first condition defining the event $\mathcal{E}$. \end{proof} \bigskip \begin{proof}[Proof of Theorem~\ref{main3}] The statement follows from~\eqref{imply} and Propositions~\ref{p:cri1} and~\ref{p:cri2}. \end{proof} \bigskip

1f9c47ce4e5be20be1230473579592918c104807

\section{Introduction} The ability to systematically design and additively manufacture multi-scale structures that exhibit globally desired mechanical properties is one the frontiers of modern material science. The advancement of 3D printing has enabled researchers to create high resolution structures with complex geometries at several scales. This in turn, enabled the creation of metamaterials that carry properties that are otherwise not possible to find in nature. Broadly speaking, metamaterials are assemblies of microstructures that derive their effective properties not just from the bulk composition but also from geometric arrangements of the structure. Hence, by producing different geometries at the microscopic scale and tiling them together, often in a periodic way, we can create macrostructures with desired properties. These different geometric arrangements make metamaterials possess unusual properties such as enhanced stiffness and energy absorption capabilities \cite{Ha18}, \cite{Imb16}, \cite{Li18}, indentation resistance, greater yield strength \cite{LE93}, crashworthiness \cite{Hou15}, phononic performance \cite{BD17} as well as many other interesting properties. Since the seminal work of Lakes \cite{Lakes87} to design auxetic structures, significant efforts have been devoted to the development of auxetic materials. Shape and topology optimization methods \cite{AllHom}--\cite{AJT}, \cite{BS04}, \cite{WMW04}--\cite{WW04} arise as a natural way to design complex geometries. For periodic auxetic metamaterials the overall properties can be studied using homogenization theory where the effective coefficients computed take into account the bulk material composition as well as the geometry layout. Topology optimization using inverse homogenization exploits this fact in order to systematically identify optimal geometries for the unit cell. This in turn would produce a macrostructure with desired properties. The works of Wang, Mei, Wang \cite{WMW04}, Michailidis \cite{GM14}, Nika \& Constantinescu \cite{NC18}, Wang et.al. \cite{WW14}, among others, use inverse homogenization, the level set method and the Hadamard shape derivative to identify material regions and track boundary changes to systematically design auxetic shapes. Wang, Mei \& Wang \cite{WMW04}, Michailidis \cite{GM14}, design linear elastic and thermoelastic materials with negative Poisson's ratio while Nika \& Constantinescu \cite{NC18} design linear elastic multimaterials with negative Poisson's ratio. The effective material tensor that characterizes the auxetic macrostructure a priori has orthotropic symmetry. Hence, the macrostructure, in 2D, is characterized by two Young modulii, $E_1$, $E_2$ and two Poisson ratios $\nu_{12}$, $\nu_{21}$. The Poisson ratio $\nu_{12}$ characterizes contraction of the structure in the direction of $Oy$ axis when stretched in the direction of $Ox$ axis and in general $\nu_{12} \ne \nu_{21}$. In literature, most auxetic metamaterials obtain their auxetic behavior through centers of rotations (see \cite{KLB15}). This type of materials tend to carry cubic symmetry, namely, $\nu_{12}=\nu_{21}$. Another interesting class of auxetic materials are re-entrant materials. For most re-entrant structures $\nu_{12} \ne \nu_{21}$. The most well known re-entrant auxetic structure is the inverted honeycomb structure first introduced by Almgren \cite{Alm85}. The inverted honey comb structure was also introduced as a ``bi-mode'' extremal material which supports a stress with a negative determinant in Milton \& Cherkaev \cite{MC95}. For isotropic structures it is well known that the Poisson ratio remains bounded in $[-1,0.5]$. Since in general these are anisotropic materials, the bounds on Poisson's ratio are wider $(-\infty,+\infty)$ than in the isotropic case; indeed Poisson's ratios smaller than negative one have been reported in Lakes \cite{Lakes93}. In the topology optimization literature the auxetic shapes obtained tend to be exhibit cubic symmetry with the apparent Poisson's ratio, $\nu^* \in [-1,0.5]$. However, as was already mentioned, this need not be the case. The aim of this paper is to complete a design cycle for several auxetic materials. The cycle consists of three steps (i) the design of the microarchitecture, (ii) the manufacturing of the material and (iii) the testing of the material. The goal is to obtain domain microarchitectured materials with a prescribed elasticity tensor and apparent Poisson's ratio. In order to reach this goal topology optimization will be used for the material design process. Materials will be manufactured using a desktop stereolithography 3D printer and then tested on standard tensile machine. Insight into the local mechanical fields is obtained using digital image correlation. The paper is organized as follows. Section 2 presents the optimal design of the microarchitectured material. It provides some basic results needed from the theory of homogenization, relate the effective coefficients to the apparent Poisson ratio, and sets up the optimization problem to systematically identify optimal auxetic shapes. Section 3 present the optimal auxetic microstructures obtained and describes some of their properties as well as the additive manufacturing process. Section 4 is dedicated to the experimental testing of the structures and the interpretation of results using digital image correlation. Additionally, an appendix reviews the approach used to measure the apparent Poisson's ratio by DIC and the finite element method on periodic structures. \section{Optimal design of the microarchitectured material} The link between the properties of the material at the scale of the microarchitecture and the macroscopic scale of the material is described using classical homogeneization theory. This defines a precise mathematical framework which enables the application of the topology optimization algorithm to compute a microstructure such that the material has a prescibed elasticity tensor. Next we shall only refer to the main results used in the sequel and recommend the classical literature on homogenization for details and proofs (see for example Sanchez-Palencia \cite{SP80}, Bakhavalov \& Panasenko \cite{BP89}, Sanchez-Palencia, A. Zaoui \cite{SP87}, Allaire \cite{AllHom}, Cioranescu \& Donato \cite{CD00}, Mei \& Vernescu \cite{MV10} among others) Let us consider a linear, elastic body occupying a bounded domain $\Omega$ of $ {\mathbb R} ^N, N = 2, 3$. Its periodic microstructure is of size $\epsilon$ assumed to be small in comparison to the size of the domain and completely described by a rescaled unit cell $Y=\left( -1/2,1/2 \right)^N \subset {\mathbb R} ^N$. The material properties in $\Omega$ are represented by a periodic fourth order tensor $\mathbb{A}(\vc{y})$ with $\vc{y}=\vc{x}/\epsilon \in Y$ and $\vc{x} \in \Omega$ the current points in the respective domains. \begin{figure}[h] \begin{center} \begin{tikzpicture}[scale=1.0] \draw [step=0.5,thin,blue!40] (-2.6,-1.7) grid (2.6,1.7); \draw [semithick,blue!40] (0,0) ellipse (2.1 and 1.2); \draw [semithick,blue!80] (2.0,1.0) node [left] {$\Omega$}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(0.25,0.3) (0.3,0.2) (0.2,0.1) (0.2,0.3)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(0.75,0.3) (0.8,0.2) (0.7,0.1) (0.7,0.3)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(1.25,0.3) (1.3,0.2) (1.2,0.1) (1.2,0.3)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(0.25,0.8) (0.3,0.7) (0.2,0.6) (0.2,0.8)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(0.75,0.8) (0.8,0.7) (0.7,0.6) (0.7,0.8)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(-0.25,0.3) (-0.2,0.2) (-0.3,0.1) (-0.3,0.3)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(-0.75,0.3) (-0.7,0.2) (-0.8,0.1) (-0.8,0.3)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(-1.25,0.3) (-1.2,0.2) (-1.3,0.1) (-1.3,0.3)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(-0.25,0.8) (-0.2,0.7) (-0.3,0.6) (-0.3,0.8)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(-0.75,0.8) (-0.7,0.7) (-0.8,0.6) (-0.8,0.8)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(-0.25,-0.3) (-0.2,-0.2) (-0.3,-0.1) (-0.3,-0.3)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(-0.75,-0.3) (-0.7,-0.2) (-0.8,-0.1) (-0.8,-0.3)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(-1.25,-0.3) (-1.2,-0.2) (-1.3,-0.1) (-1.3,-0.3)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(-0.25,-0.8) (-0.2,-0.7) (-0.3,-0.6) (-0.3,-0.8)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(-0.75,-0.8) (-0.7,-0.7) (-0.8,-0.6) (-0.8,-0.8)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(0.25,-0.3) (0.3,-0.2) (0.2,-0.1) (0.2,-0.3)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(0.75,-0.3) (0.8,-0.2) (0.7,-0.1) (0.7,-0.3)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(1.25,-0.3) (1.3,-0.2) (1.2,-0.1) (1.2,-0.3)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(0.25,-0.8) (0.3,-0.7) (0.2,-0.6) (0.2,-0.8)}; \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(0.75,-0.8) (0.8,-0.7) (0.7,-0.6) (0.7,-0.8)}; \draw [->,blue!40] (1.5,0) -- (5,-1); \draw [->,blue!40] (1.5,0.5) -- (5,2); \draw [semithick,blue!40] (5,-1) -- (8,-1) -- (8,2) -- (5,2) -- (5,-1); \draw[semithick,red,fill=red] plot [smooth cycle] coordinates {(6.2,-0.3) (6.8,-0.1) (7,0.8) (6.3,1.2)}; \draw [semithick,blue!80] (8.2,2.3) node [left] {$Y$}; \draw [<->,semithick,blue!40] (8.2,-1) -- (8.2,2); \draw [semithick,blue!80] (8.2,0.5) node [right] {$1$}; \draw [<->,semithick,blue!40] (5,-1.2) -- (8,-1.2); \draw [semithick,blue!80] (6.5,-1.2) node [below] {$1$}; \end{tikzpicture} \end{center} \caption{Schematic of the elastic periodic composite material.} \label{fig:hom_schem} \end{figure} The homogenized coefficients $\mathbb{A}^H$ of the effective medium at the scale of $\Omega$ are obtained from the following explicit computation on the unit cell $Y$: \begin{equation}\label{hom:coef} A^H_{ijm\ell} = \int_{Y} \mathbb{A}(E^{ij} + \varepsilon(\vc{\chi}^{ij})):(E^{m\ell} + \varepsilon(\vc{\chi}^{m\ell})) \, d\vc{y} \end{equation} where $(\vc{e}_k)_{1 \le k \le N}$ is the canonical basis of $ {\mathbb R} ^N$. $\vc{E}^{m\ell}=\frac{1}{2}(\vc{e}_m \otimes \vc{e}_\ell + \vc{e}_{\ell} \otimes \vc{e}_m) \text{ with } m, \ell \in \{ 1, ... , N\}$ denote a vector basis of the mean deformations applied on the unit cell. $\vc{\chi}^{m\ell}$ are the corresponding displacements fields created by these deformations on the unit cell $Y$. As such the displacement fields $\vc{\chi}^{m\ell}$ are solution of the following linear elastic problems with periodic boundary conditions: \begin{equation} \label{local:sol} -\dv \left( { \mathbb{A} (\vc{E}^{m\ell} + \vc{\varepsilon}(\vc{\chi}^{m\ell})) } \right) = \vc{0} \text{ in } Y \end{equation} \noindent where the displacement $\vc{\chi}^{m\ell}(\vc{y})$ is periodic in $Y$ and has a zero average value on the unit cell $$ \mean{\vc{\chi}^{m\ell}}_Y = \vc{0}. $$ The loading is given by the distributed body force $- \dv \left( \mathbb{A} (\vc{E}^{m\ell}) \right) $, which stems directly from the mean deformation field $\vc{E}^{m\ell}$ and the distribution of elastic moduli $\mathbb{A}$ in the unit-cell. \begin{figure}[h] \centering \includegraphics[width=130mm]{MeanDeform2} \caption{A schematic view of boundary displacements corresponding to the three mean deformation tensors $\vc{E}^{m\ell}$ applied on a homogeneous unit cell $Y$: tensile deformations along the coordinate axes (a) $\vc{E}^{11}$, (b) $\vc{E}^{22}$ and (c) the simple shear deformation $\vc{E}^{12}$. } \label{MeanDeform} \end{figure} \subsection*{Orthotropic elastic materials} The homogenized tensor $\mathbb{A}^H$ in \eqref{hom:coef} carries a natural orthotropic material symmetry. The linear elastic constitutive equation on the unit cell relating the mean stress and strain tensors, denoted as $\vc{\sigma}^H$ and $\vc{\varepsilon}^H$ respectively, has therefore the following expression for the two dimensional problems under consideration: \[ \begin{pmatrix} \sigma^H_{11} \\[3pt] \sigma^H_{22} \\[3pt] \sigma^H_{12} \end{pmatrix} = \begin{pmatrix} A^H_{1111} & A^H_{1122} & 0 \\[3pt] A^H_{1122} & A^H_{2222} & 0 \\[3pt] 0 & 0 & 2\,A^H_{1212} \\[3pt] \end{pmatrix} \begin{pmatrix} \varepsilon^H_{11} \\[3pt] \varepsilon^H_{22} \\[3pt] \varepsilon^H_{12} \end{pmatrix} \] where $\vc{\sigma}^H$ and $\vc{\varepsilon}^H$ are the mean stress and strain respectively: $$ \vc{\sigma}^H = \mean{\vc{\sigma}}_Y \qquad \text \qquad \vc{\varepsilon}^H = \mean{\vc{\varepsilon}}_Y. $$ Alternatively, one could express the effective strain as a function of the effective stress with the following effective material tensor: \[ \begin{pmatrix} \varepsilon^H_{11} \\[3pt] \varepsilon^H_{22} \\[3pt] \varepsilon^H_{12} \end{pmatrix} = \begin{pmatrix} 1/E_1 & -\nu_{12}/E_2 & 0 \\[3pt] -\nu_{21}/E_1 & 1/E_2 & 0 \\[3pt] 0 & 0 & 1/{2\,G} \\[3pt] \end{pmatrix} \begin{pmatrix} \sigma^H_{11} \\[3pt] \sigma^H_{22} \\[3pt] \sigma^H_{12} \end{pmatrix} \] where $E_i$ and $\nu_{ij}$ denote the homogenized Young moduli and Poisson's ratios respectively. In order to simplify the notation, we will drop the $H$ superscript. Let us further remark, that by symmetry of the elastic compliance matrix, the following ratios have to be equal: $$ \frac{\nu_{12}}{E_2} = \frac{\nu_{21}}{E_1} $$ The elastic moduli, $A^H_{ijkl}$, can equally be expressed in terms of the compliance moduli, i.e. Young moduli and Poisson's ratios: $A^H_{1111}= (1-{\nu_{12}\nu_{21}})^{-1}E_1$, $A^H_{2222}= (1-{\nu_{12}\nu_{21}})^{-1}E_2$, $A^H_{1122}=\nu_{21}(1-{\nu_{12}\nu_{21}})^{-1}E_1$, $A^H_{2211}=\nu_{12}(1-{\nu_{12}\nu_{21}})^{-1}E_2$ with $A^H_{1122}=A^H_{2211}$ as can be easily obtained from the inversion of the corresponding matrices. A simple calculation immediately yields: \[ \nu_{12}=\frac{A^H_{1122}}{A^H_{2222}} \text{ and } \nu_{21}=\frac{A^H_{1122}}{A^H_{1111}}. \] Moreover, the homogenized Poisson's ratio $\nu_{ij}$ are equally denoted \textit{apparent Poisson's ratio } to highlight their reference to the homogenized unit cell. For example $\nu_{12}$ characterizes the contraction of the structure in the direction of $Oy$ axis when the cell stretched in the direction of $Ox$ axis and in general $\nu_{12} \ne \nu_{21}$. However, if the microarchitecture of the unit cell obeys cubic symmetry we have $A^H_{1111} = A^H_{2222}$ and we trivially obtain that $E_1=E_2=E^*$ and $\nu_{12} = \nu_{21}=\nu^*$. \subsection*{The optimization problem} Next we discuss the framework of the optimization problem without presenting the details of the algorithm which is beyond the scope of the paper. Let us assume that the unit cell $Y$ is a working domain and consider $\omega$ an open and bounded subset of $Y$ representing the shape of the architectured microstructure. The distance between the homogenized elastic moduli over the unit cell and target elastic moduli can be measured by the following cost functional, \begin{equation*} \label{objective} J(\omega) = \frac{1}{2} \norm{\mathbb{A}^H(\omega) - \mathbb{A}^t}^2_{\eta} \end{equation*} where $\norm{\cdot}_{\eta}$ is the weighted Euclidean norm, $\mathbb{A}^t$ is the target elastic tensor, and $\eta$ are the weight coefficients. We define a set of admissible shapes contained in the working domain $Y$ and have a fixed volume by $\mathcal{U}_{ad} = \left \{ \omega \subset Y \text{ such that } |\omega| = V^t \right \}.$ Hence, the optimal shape design of the microarchitecture can be formulated as the following optimization problem: \begin{gather} \begin{aligned} \label{opti:prob} & \inf_{\omega \subset \mathcal{U}_{ad}} J(\omega), \\ \vc{\chi}^{m\ell} & \text{ satisfies } \eqref{local:sol}. \end{aligned} \end{gather} \subsection*{Level set representation of the shape} The shape $\omega$ is represented in a fixed mesh inside the unit cell $Y$ using the the level set method. If we denote by $\phi$ a level set, then the corresponding shape $\omega$ is defined in the following way: $$ \displaystyle \begin{cases} \phi(\vc{y}) = 0 & \text{ if } \vc{y} \in \partial \omega \cap Y, \nonumber \\ \phi(\vc{y}) > 0 & \text{ if } \vc{y} \in Y \backslash \omega, \nonumber \\ \phi(\vc{y}) < 0 & \text{ if } \vc{y} \in \omega, \nonumber \end{cases} $$ See also {\sc Figure \ref{fig:levelset}} for a graphical example. \begin{figure}[h] \centering $\vcenter{\hbox{\includegraphics[width=3.0in]{levelset.jpg}}}$ \caption{Schematic representation of the level set $\phi$ of the domain $\omega$} \label{fig:levelset} \end{figure} The optimization of $J(\omega)$ is carried out by advecting an initial shape $\omega_0$ with velocity $v$ obtained from the shape derivative $\dpair{J'(\omega)}{\vc{\theta}}$ in the direction $\vc{\theta}$ (see Allaire, Jouve, Toader \cite{AJT}, Wang, Mei, Wang \cite{WMW04}). The advection is realized by solving the Hamilton-Jacobi equation, \[ \partial_t \phi + v|\grad \phi| = 0, \] where $v$ is the velocity of the interface computed from the shape derivative $\dpair{J'(\omega)}{\vc{\theta}}$. The numerical algorithm can be summarized in the following steps: {\it \begin{itemize} \item[(i)] Initialize the level set $\phi^0$ corresponding to the initial form $\omega_0$. \item[(ii)] Update the level set $\phi^0$ using the signed distance function $d_{\omega_0}$. \item[(iii)] Iterate until convergence for $k \ge 0$: \begin{itemize} \item[a.] Calculate the local solutions $\vc{w}^{m\ell}_k$ for $m,\ell=1,2$ by solving the linear elasticity problem in $Y$. \item[b.] Deform the domain $\omega_k$ by solving the above Hamilton-Jacobi equation. The new shape $\omega_{k+1}$ is characterized by the level set $\psi^{k+1}$ after a time step $\Delta t_k$. The time step $\Delta t_k$ is chosen so that $J(\omega^{k+1}) \le J(\omega^k)$. \end{itemize} \item[(iv)] If needed for stability reasons, re-initialize the level set functions $\phi^k$. \end{itemize}} \noindent The complete algorithm as well as several examples are presented in Nika \& Constantinescu \cite{NC18}. The interested reader is encouraged to consult the works of Allaire \cite{AllCon}, Allaire, Jouve \& Toader \cite{AJT}, Allaire et. al. \cite{ADDM}, Michailidis \cite{GM14}, Wang, Mei \& Wang \cite{WMW04} for more details about the mathematical results and algorithmic issues for the solution method. \subsection*{Examples of optimal microstructures} In all the examples that follow we have set the Young modulus for the void (or weak) material equal to $E_0=0.0001$ and for the strong material to $E_1=0.91$. The Poisson's ratio for both phases was set to $\nu=0.3$. The quadratic unit cell $Y$ was meshed with a structured symmetric grid of $100 \times 100$ quadrangular each formed of four equal triangular $P1$ elements. All computations were carried out using an in house programming of the preceding algorithm \cite{NC18} operating on {\tt FreeFEM++} software \cite{FH12}. We would like to remark that we only controlled the coefficients and is reported in Table \ref{table:Hcoef} with varying values for the three different structures. In all examples the target objective was defined only in terms of $A^H_{1111}, A^H_{1122}, A^H_{2222}$. The shear coefficient $A^H_{1212}$ as well as the $A^H_{1211}$ and $A^H_{1222}$ coefficients were left free. Therefore only the elastic moduli of the unit cell corresponding to the direction $11$ and $22$ directions of strain and stresses were controlled. However all targets had as an underlying objective an apparent Poisson ratio which will be discussed next. \noindent \textit{\bf Example 1}. The first microstructure to be optimized is a structure whose target apparent Poisson's ratio is equal $\nu^* = -1.0$. The volume constraint was set to $V^t=50\%$ and was enforced using a Lagrange multiplier computed by assuming that the optimality criterion was satisfied, namely $\dpair{J'(\omega)}{\vc{\theta}} + \ell \dpair{V'(\omega)}{\vc{\theta}}=0$. The Lagrange multiplier was updated by imposing $\ell^{n+1} = \frac{(\ell^n + \ell)}{2} + \epsilon (V-V^t),$ where $\epsilon$ is a small positive real number and $\ell = -\dpair{J'(\omega)}{\vc{\theta}} / \dpair{V'(\omega)}{\vc{\theta}}$. This technique has been initially proposed and discussed by Allaire and Pantz \cite{AP06}. We further note that for this structure we enforced symmetry only along the $Oy$ axis in the algorithm, by symmetrizing the shape after each iteration. \begin{figure}[h] \centering \begin{tabular}{ccc} \subf{\includegraphics[width=35mm]{initial1}} {Initial design} & \subf{\includegraphics[width=35mm]{macro1}} {Final design} & \subf{\includegraphics[width=35mm]{img_macro1}} {Macro-structure} \end{tabular} \caption{The design process of the material from initial guess to final macro-structure represented in the unit cell and as a periodic material. \protect \newboxsymbol{magenta}{magenta} Young modulus of $0.91$, \protect \newboxsymbol{yellow}{yellow} void.} \label{example1} \end{figure} \begin{figure}[h] \centering \begin{tabular}{cc} \subf{\includegraphics[width=45mm]{obj1}} {Objective} & \subf{\includegraphics[width=45mm]{vol1}} {Volume constraint} \end{tabular} \caption{Convergence history of the objective function and the volume constraint.} \label{example1:convergence} \end{figure} \noindent The initial and final shape of the microstructure on the unit-cell and as a periodic material are represented in figure \ref{example1}. The final shape can be characterized as an inverted honeycomb structure and looks similar to the designs imagined by Almgren \cite{Alm85}. Its homogenized coefficients, displayed in Table \ref{table:Hcoef}, show that the structure exhibits orthotropic symmetry and a simple calculation yields $\nu_{12} = -1.25$ and $\nu_{21}=-0.42$. Hence, the expansion of the structure along the $Oy$ axis when stretched in the $Ox$ axis is larger the expansion along the $Ox$ axis when stretched in the $Oy$ axis. This non-symmetric effect has been enabled as the symmetry relation was only imposed along the $Oy$ axis in the algorithm. \noindent The convergency history of the cost-functional and of the volume constraint displayed in {\sc Figure \ref{example1:convergence}} shows that the target coefficient where reached in slightly more than 20 iteration and that the later iteration contributed only to small improvements without bringing the cost functional to less than $0.06$ which corresponds to $92\%$ decrease of the initial value. The gap with respect to the target moduli can be read from Table \ref{table:Hcoef}. It is interesting to remark, that the final optimized microstructure has a shear moduli close to $0$. However the final apparent Poisson's ratio is close to the imagined target as will be discussed in the comparison with the printed samples. The volume constraint has a different evolution than the cost-functional with an initial increase given by the initial evolution of the holes and then a fast and a slow evolution which lies within the proposed range of the constraint. \noindent \textit{\bf Example 2}. The second microstructure to be optimized is a structure whose target apparent Poisson's ratio was also $\nu^*=-1.0$. The Young modulus and Poisson's ratio are the same as in example 1. The volume constraint was of an inequality type, and was set to $16\% \le V^t \le 60\%$. For the enforcement of the constraint we used an augmented Lagrangian. The Lagrange multipliers for the weight were updated the following way, $\ell^{n+1} = \ell^n + \beta (V-V^t),$ where $\beta$ is a penalty term that was updated every $10$ iterations. We point out that the augmented Lagrangian approach would require a re-computation of the shape derivative. As result, initial guesses for $\ell$ and $\beta$ would both have an effect on final optimal shape (see also the discussion on the algorithmic issues in Nika \& Constantinescu \cite{NC18}). \begin{figure}[h] \centering \begin{tabular}{ccc} \subf{\includegraphics[width=35mm]{initial2}} {Initial design} & \subf{\includegraphics[width=35mm]{micro2}} {Final design} & \subf{\includegraphics[width=35mm]{img_macro2}} {Macro-structure} \end{tabular} \caption{The design process of the material from initial guess to final macro-structure represented in the unit cell and as a periodic material. \protect \newboxsymbol{magenta}{magenta} Young modulus of $0.91$, \protect \newboxsymbol{yellow}{yellow} void.} \label{example2} \end{figure} \begin{figure}[h] \centering \begin{tabular}{cc} \subf{\includegraphics[width=45mm]{obj2}} {Objective} & \subf{\includegraphics[width=45mm]{vol2}} {Volume constraint} \end{tabular} \caption{Convergence history of the objective function and the volume constraint.} \label{example2:convergence} \end{figure} \noindent Let us further specify, that we enforced in the symmetry of the shape along both the $Ox$ axis and $Oy$ axis, by symmetrizing the shape during the algorithmic iterations. The resulting structure exhibits cubic symmetry with apparent Poisson's ratio $\nu^*=\nu_{12}=\nu_{21}=-0.42$. Hence, the expansion along the $Ox$ and $Oy$ axis is the same. As in the first example, one can read the target and final elastic moduli from Table \ref{table:Hcoef}. As before, the final optimized microstructure has a shear moduli close to $0$ and the final apparent Poisson ratio is close to the imagined target. \noindent The evolution of the cost-functional displayed in Figure \ref{example2:convergence} shows that the decrease was not monotonous and that several interventions in the step size and algorithmic details where necessary to reestablish convergence. Moreover these moments are correlated with the important variations in the volume constraint, however the final volume ratio at approximatevely $35\%$ lies in the middle of the imposed interval. \noindent \textit{\bf Example 3}. The third microstructure to be optimized is a structure whose target apparent Poisson's ration is $\nu^*=-0.5$. The volume constraint is set to $V^t = 40\%$ and is updated the same exact way as in the first example. Again, as in example 2, we symmetrized the structure along the $Ox$ and $Oy$ axis. \begin{figure}[h] \centering \begin{tabular}{ccc} \subf{\includegraphics[width=35mm]{initial3}} {Initial design} & \subf{\includegraphics[width=35mm]{micro3}} {Final design} & \subf{\includegraphics[width=35mm]{img_macro3}} {Macro-structure} \end{tabular} \caption{The design process of the material from initial guess to final macro-structure. \protect \newboxsymbol{magenta}{magenta} Young modulus of $0.91$, \protect \newboxsymbol{yellow}{yellow} void.} \label{example3} \end{figure} \begin{figure}[h] \centering \begin{tabular}{cc} \subf{\includegraphics[width=45mm]{obj3}} {Objective} & \subf{\includegraphics[width=45mm]{vol3}} {Volume constraint} \end{tabular} \caption{Convergence history of objective function and the volume constraints.} \end{figure} The structure exhibits cubic symmetry with apparent Poisson ratio $\nu^*=-0.47$. Moreover one can remark that the final elastic moduli are very close to the target and that this structure has a shear moduli which is of the same order of magnitude as the other moduli. \begin{table} \begin{tabularx}{\textwidth}{*{3}{>{\centering \arraybackslash}X}} Target tensor $\mathbb{A}^t$ & Homogenised tensor $\mathbb{A}^H$ & Unit cell\\ \hline {$\displaystyle \begin{pmatrix} 0.1 & -0.1 & 0 \\ -0.1 & 0.1 & 0 \\ 0 & 0 & G \\ \end{pmatrix}$} & {$\displaystyle \begin{pmatrix} 0.12 & -0.05 & 0 \\ -0.05 & 0.04 & 0 \\ 0 & 0 & 6\times 10^{-3} \\ \end{pmatrix}$} & \vspace{0.2cm} \includegraphics[width=0.15\columnwidth]{macro1}\\ {$\displaystyle \begin{pmatrix} 0.1 & -0.1 & 0 \\ -0.1 & 0.1 & 0 \\ 0 & 0 & G \\ \end{pmatrix}$} & {$\displaystyle \begin{pmatrix} 0.12 & -0.05 & 0 \\ -0.05 & 0.12 & 0 \\ 0 & 0 & 0.003 \\ \end{pmatrix}$} & \vspace{0.2cm} \includegraphics[width=0.15\columnwidth]{micro2}\\ {$\displaystyle \begin{pmatrix} 0.2 & -0.1 & 0 \\ -0.1 & 0.2 & 0 \\ 0 & 0 & G \\ \end{pmatrix}$} & {$\displaystyle \begin{pmatrix} 0.19 & -0.09 & 0 \\ -0.09 & 0.19 & 0 \\ 0 & 0 & 0.6 \\ \end{pmatrix}$} & \vspace{0.2cm} \includegraphics[width=0.15\columnwidth]{micro3} \end{tabularx} \caption{Target tensor values and homogenized tensor values computed numerically for each corresponding shape.} \label{table:Hcoef} \end{table} \section{Analysis of fabricated polymer structures} \subsection*{Manufacturing process: equipment and materials} The optimal shapes have been additively manufactured with digital light processing stereo-lithography technology (DLP) using a EMBER DLP 3D printer. A digital projector screen flashes a single image of each layer across the entire surface of the vat filled of photo-sensitive liquid resin at once, causing chains of molecules to link and thus forming solid polymer. The process is repeated until the 3D model is complete. Then the vat is drained of liquid, revealing the solidified model and the solid model is washed with a solvant. The printer has a resolution of $50 \, \mu m$, corresponding to 1 pixel in the digital projector screen, and a range of the processing layer thickness of $10-100 \, \mu m$. The largest processing build volume is $64 \, mm \times 40 \, mm \times 134 \, mm$ (note that $64 \, mm \times 40 \, mm$ correspond to a $1280 \times 800$ pixels picture). For a thickness of $25 \, \mu m$ per layer, the speed range is of $18 mm/h$. The printable minimal feature size of the specimens is announced at $0.4 \, mm$ corresponding roughly to $8$ pixels. We selected a rubber-like material, commercially denoted as GM08b, as the base material because of its compliant nature. Figure \ref{GM08B} displays a representative tensile stress--strain curve of the GM08b material. As expected for a rubber-like material it does not display an ideal linear elastic behavior, it exhibits a gradually variation of the stiffness with increasing strain. However, we can linearly approximate the stiffness with a $7$ {\em MPa} Young's modulus in accordance with one of the manufacturers data sheet (see \texttt{https://dl.airtable.com}). Moreover, let us further remark that the Poisson coefficient of the material has not been measured, but we should expect it to be close to the incompressible limit of $\nu \approx 0.5$. \begin{figure} \centering \includegraphics[width=100mm]{StressStrainGM08b} \caption{Uniaxial tensile test. Homogeneous stress-strain curves} \label{GM08B} \end{figure} The optimal shapes obtained in examples 1--3 (see {\sc Figure \ref{example1}, Figure \ref{example2}, Figure \ref{example3} } are represented by the final level set function. The later presents a smooth variation between values corresponding to the two materials in a neighborhood of their interface and therefore the levelset representation has been binarized and then extruded in the $Oz$ direction in order to create a 3D object. More precisely, the 3D-printed samples have been produced by the following procedure: {\it \begin{itemize} \item[(i)] Binarize the level set function representing by shape optimisation. As the later presents a smooth variation between values corresponding to the two materials in a neighborhood of their interface. \item[(ii)] Create a periodic array for each sample: $8 \times 6$ unit cells for example 1, $5 \times 4$ unit cells for example 2 and example 3, using the software ImageJ \cite{rueden2017imagej2}. The final result was a binarised $1280 \times 800$ pixels image (see Figure \ref{fig:3D} for details) . \item[(iii)] Extrude the preceding image to obtain the 3D sheet of the desired height. The final dimensions of the printed samples are $64 \, \times 38 \, \times 6 \, mm$ for example 1 and $64 \times 40 \, \times 6 \,  mm$ for example 2 and 3. \item[(iv)] Print the files with the following processing parameters: laser power was $5 \, W$, the exposure time $1\,s$ per layer and the layer thickness was $50 \, \mu m$. \item[(v)] Wash the samples in an isopropanol bath for $5\,min$ \item[(vi)] Post-cure the samples for $30\,min$ in an UV oven at 2000W. \end{itemize}} \begin{figure} \centering \begin{tabular}{ccc} \subf{\includegraphics[width=.413\columnwidth]{aux1sample}} {Example 1: Input design} & \subf{\includegraphics[width=.45\columnwidth]{aux1printed}} {Example 1: Fabricated result} \\ \subf{\includegraphics[width=.405\columnwidth]{aux2sample}} {Example 2: Input design} & \subf{\includegraphics[width=.45\columnwidth]{aux2printed}} {Example 2: Fabricated result} \\ \subf{\includegraphics[width=.41\columnwidth]{aux3sample}} {Example 3: Input design} & \subf{\includegraphics[width=.45\columnwidth]{aux3printed}} {Example 3: Fabricated result} \end{tabular} \caption{Fabricated auxetic microstructures with different apparent Poisson ratios (right) and the initial input design (left).} \label{fig:3D} \end{figure} \section{Testing and full-field displacement measurement using digital image correlation} A series of uniaxial static tensile tests were undertaken to assess the tensile properties of the auxetic lattice structures by using a home-made testing machine with a symmetric displacement of the two crossheads and equipped with a $100 \, N$ load cell. The tensile tests were performed at a rate of $0.05 \, mm/s$ up to $3\, mm$, which corresponds to a strain rate of $\dot{\varepsilon} = 10^-3 \, s^-1$ up to a maximal strain of $\varepsilon = 5\%$. Several pictures were taken during the tensile experiment and the complete displacement field of the specimen was computed using digital image correlation (DIC). The pictures with a resolution of $4904 \times 3280$ were obtained using a high-resolution digital camera (Schneider Optics 8-bit camera with a Makro-UNIFOC 100/77 lens) mounted on the tensile testing machine and grey scale pictures were recorded every second during the loading. The camera is mounted on a perpendicular axes with respect to the plane of the specimen, which enables the direct use of a 2D DIC. To improve the precision of the measurement, a white speckle pattern was placed on the sample by airbrushing. The DIC was done using the CorrelManuV 2D (CMV) software, developed by M. Bornert \cite{ABBC94}. The processed displacement field corresponds to a single unit cell in the middle of the structure at 5 different loading time steps, using a $100 \times 100$ grid, i.e. having $10 000$ measurement points. For each node, the subset size was set to $20 \times 20$ pixels, while the searching area was set to $100 \times 100$ pixels. The measurement included a computation without transformation, i.e. rotation of the subset window and a re-optimisation allowing transformations with a reduced searching area of $30 \times 30$ pixels. \begin{figure} \centering \includegraphics[width=100mm]{stress_strain} \caption{Stress strain curves for all three examples obtained by performing a uniaxial tensile test. All three materials exhibit linear behaviour for strains up to $5\%$.} \label{stress_strain} \end{figure} \subsection*{Finite Element Computations} A series of finite element computation were undertaken under the assumption small strains, large displacements and plane stress using the finite element solver Cast3M2018 (\texttt{http://http://www-cast3m.cea.fr}). The mesh was obtained using image processing from the binarized images of the optimal level set function and completed to the sample geometry including the grips. The sample was loaded by the imposing the displacements of the clamps of the tensile machine. Other boundaries were stress free edges. The elastic behavior was considered to be linear and corresponded to the linear approximation of the base material displayed in {\sc Figure \ref{stress_strain}}. \subsection*{Results} The stress strain response under a uniaxial tensile test for the three materials are displayed in {\sc Figure \ref{stress_strain}}. One can easily observe a linear behavior of the samples that up to a maximal strain of $5\%$ strain despite the nonlinearity of the rubber-like base material in the same strain range. This indicates that the samples have as expected structural deformation where different parts of the "lattice" behave as rigid struts and deformable hinges. This effect will be highlighted by the DIC measurements discussed later. Furthermore, one can directly observe a lateral expansion during the tensile extension indicating a negative apparent Poisson's ratio for all the samples. The precise measurements of the apparent Poisson's ratio corresponding to a single central unit cell are presented in {\sc figure \ref{APR_evolution}}. The precise method for the computation of the apparent Poisson's ratio of a single unit cell from DIC measurements was based on homogenization assumptions and is presented in detail in the appendix \ref{appendix}. The results show that the initial apparent Poisson ratio was for all samples close to the target value used in the optimization process and was not degraded during the manufacturing process. During tensile loading, the apparent Poisson's ratio tends to increase, indicating a decrease of the ``auxeticity" of the samples of up to increases by $10\%$ for a $5\%$ strain. Let us first remark, that the evolution is close to predictions of the deformation of the samples obtained by the finite element method under the assumption of large displacements. Second, one can remark that the evolution of the apparent Poisson ratio with applied strain has already been observed and discussed in \cite{clausen2015topology} on polymeric filament structure. Moreover, they arrived to correct the phenomenon up to $20 \%$ strain using a nonlinear material behavior in the optimization process, see \cite{clausen2015topology,WW14} for more details on the subject. In the case of the optimization procedure presented here, the extension to nonlinear material behavior is currently under works. The displacements fields obtained using DIC permit a further comparison with predictions and give an insight of the deformation mechanism in the samples, i.e. how the structure moves and deforms. {\sc Figures \ref{aux1uy}--\ref{aux3uy}} display the measured and the computed vertical displacement, i.e. the $u_y$ displacement component, of the central unit-cell. A comparison of the values and the shapes of the colormaps exhibits a good match between the measurements and finite element prediction. Moreover the displacement fields permits to better understand the local movements of the microstructure which conducts to the global auxeticity effect by combining almost rigid regions submitted to translations and rotations with local concentrated deformation exhibiting local hinges. \begin{figure}[H] \centering \includegraphics[width=120mm]{aux1uy} \caption{Dimensionless values of the vertical displacement field for the unit cell of the structure in example 1 as measured using Digital Image Correlation (right) and Finite Element Analysis (left). Scale bar is 1 mm.} \label{aux1uy} \end{figure} \begin{figure}[H] \centering \includegraphics[width=120mm]{aux2uy} \caption{Dimensionless values of the vertical displacement field for the unit cell of the structure in example 1 as measured using Digital Image Correlation (left) and Finite Element Analysis (right).} \label{aux2uy} \end{figure} \begin{figure}[H] \centering \includegraphics[width=120mm]{aux3uy} \caption{Dimensionless values of the vertical displacement field for the unit cell of the structure in example 3 as measured using Digital Image Correlation (left) and Finite Element Analysis (right).} \label{aux3uy} \end{figure} \begin{figure}[h] \centering \includegraphics[width=100mm]{APRv3} \caption{Evolution of the apparent Poisson ratio plotted as a function of the strain. We observe the all three materials lose their auxeticity as the uniaxial strain increases.} \label{APR_evolution} \end{figure} \section{Conclusion} This paper presented a complete design cycle for auxetic materials based on a topology optimization of the microstructure, the manufacturing of the material with a periodic microstructure using a commercial stereolithography printer and its testing. The topology optimization technique was based on the level set method and the Hadamard shape derivative to track the topological changes imposed by the variation of the cost functional. The results showed that optimal shapes could be directly printed without additional enhancement of the surface, which is a direct consequence of the in the smoothed interface technique used in the optimization. Moreover the manufactured materials had the designed mechanical behavior. The targeted apparent Poisson's ratios have been experimentally attained and the local material behavior was close to predictions obtained by the finite element method. However, the results showed that the nonlinear material behavior both in terms of the constitutive law and of large strains and displacements should be included in the design process in order to control the target materials in the large strains regime. The topology optimization process proved to be efficient tool to reach the desired apparent Poisson's ratios. However, the process was not as efficient in reaching a precise tensor of elastic moduli. The theoretical problem of reachable elasticity tensors has theoretically been solved using laminates in the seminal paper of Milton \& Cherkaev \cite{MC95} starting from two isotropic materials with arbitrary Young moduli. Given geometric and materials constraints that 3D printing introduces we are not certain that all targets of elastic moduli can be attained. In other words, the definition of the set of elastic moduli determined by all printable designs is a priori not know. Moreover, we are in need of a simple method to assess the set of realizable tensors when starting from a combination of real materials and given geometry constraints imposed by the 3D printer constraints. The designed examples had two or three non-zero eigenelastic moduli and indicate that the design of unimode, bimode, pentamode, etc. materials as theoretically proposed in \cite{MC95} could be pursued by topology optimization. The article presented a structural analysis of microarchitectured materials. Its key novelties are the microstructural optimization and the excellent match between numerical predictions and experimental observations. The results demonstrate that the employed design method can be extended over different scales, creating materials with hierachically architected microstructures. \section*{Acknowledgments} This work is financed by the french-swiss ANR-SNF project MechNanoTruss (ANR-15-CE29-0024-01). AC and GN would like to express their gratitude to Chiara Daraio for fruitfull discussion on the design of lattice structures and to Gregoire Allaire and Georgios Michailidis for lending their expertise on the numerical and algorithmic issues of the optimization. \bibliographystyle{amsplain}

2079596a3537c4e2e58763b3a1840e76fdd5d9ed

\section{Introduction}\label{sec:intro} A natural requirement for statistical methods for density estimation and clustering is for them to be robust under affine transformations of the data. Such a desideratum is exacerbated in multivariate problems where data components are incommensurable, that is not measured in the same physical unit, and for which, thus, the definition of a metric on the sample space requires the specification of constants relating units along different axes. As an illustrative example, consider astronomical data consisting of position and velocity of stars, thus living in the so-called phase-space: a metric on such a space can be defined by setting a dimensional constant to relate positions and velocities. In this setting, any sensible statistical procedure should be robust with respect to the specification of such a constant \citep{ascasibar2005numerical,maciejewski2009phase}. This is specially important considering that often scarce to no a priori guidance about dimensional constants might be available, thus making the model calibration a daunting task. The motivating example of this work comes indeed from astronomy, the dataset we consider consisting of measurements on a set of 139 stars, possibly belonging to a globular cluster called NGC~2419 \citep{Iba11}. Globular clusters are sets of stars orbiting some galactic center. The NGC~2419, showed in Figure~\ref{fig:glob}, is one of the furthest known globular clusters in the Milky Way. \begin{figure}[h] \centering \includegraphics[width=.9\textwidth]{Fig1.jpg} \caption{An image of the remote Milky Way globular cluster NGC~2419 (about 300\,000 light years away from the solar system). Picture by Bob Franke, with permission (www.bf-astro.com).} \label{fig:glob} \end{figure} For each star we observe a four-dimensional vector $(Y_1,Y_2,V,[{\rm Fe/H}])$, where $(Y_1,Y_2)$ is a two-dimensional projection on the plane of the sky of the position of the star, $V$ is its line of sight velocity and $[{\rm Fe/H}]$ its metallicity, a measure of the abundance of iron relative to hydrogen. Out of these four components, only $Y_1$ and $Y_2$ are measured in the same physical unit, while dimensional constants need to be specified in order to relate position, velocity and metallicity. A key question arising with these data consists in identifying the stars that, among the 139 observed, can be rightfully considered as belonging to NGC~2419: a correct classification would be pivotal in the study of the globular cluster dynamics. Astronomers expect the large majority of the observed stars to belong to the cluster: the remaining ones, called field stars or contaminants, are Milky Way stars, unrelated to the cluster, that happen to appear projected in the same region of the plane of the sky. In general the contaminants have different kinematic and chemical properties with respect to the cluster members. Considering the nature of the problem, this research question can be formalised as an unsupervised classification problem, the goal being the identification of the stars which belong to the largest cluster, which can be interpreted as the NGC~2419 globular cluster. Admittedly, the terms of such a classification problem are not limited to the considered dataset but, on the contrary, are ubiquitous in astronomy and, more in general, might arise in any field where data components are incommensurable. Bayesian nonparametric methods for density estimation and clustering have been successfully applied in a wide range of fields, including genetics \citep{huelsenbeck2007inference}, bioinformatics \citep{medvedovic2002bayesian}, clinical trials \citep{xu2017decision}, econometrics \citep{otranto2002nonparametric}, to cite but a few. In this work we focus on the Dirichlet process mixture (DPM) model introduced by \citet{Lo84}, arguably the most popular Bayesian nonparametric model. Although its properties have been thoroughly studied \citep[see, e.g.,][]{hjort2010bayesian}, little attention has been dedicated to its robustness under data transformations \citep[see][]{arbel2013bayesian}. To the best of our knowledge, only \citet{bean2016transformations} and \citet{shi2018} study the effect of data transformation under a DPM model. The goal of \citet{bean2016transformations} is to transform the sample so to facilitate the estimation of univariate densities on a new scale and thus to improve the performance of the methodology; \citet{shi2018}, instead, study the consistency of DPM models under affine data transformation, when investigating the properties of the so-called low information omnibus prior for DPM models they introduce. In this paper we investigate the effect of affine transformations of the data on location-scale DPM of multivariate Gaussians (DPM-G) \citep{Mue96}, which will be introduced in Section~\ref{sec:model}. This is a very commonly used class of DPM models whose asymptotic properties have been studied by \citet{Wu10} and \citet{Can17}, among others. While rescaling the data, often for numerical convenience, is a common practice, the robustness of multivariate DPM-G models under such transformations remains essentially unaddressed to date. We fill this gap by formally studying robustness properties for a flexible specification of DPM-G models, under affine transformations of the data. Specifically, our contribution is two-fold: first, we formalise the intuitive idea that a location-scale DPM-G model on a given dataset induces a location-scale DPM-G model on rescaled data and we provide the parameters mapping for the transformed DPM-G model; second, we introduce the notion of asymptotic robustness under affine transformations of the data and show that, under mild assumptions on the true data generating process, DPM-G models feature such robustness property. As a by-product, we show that the original assumptions of \citet{Wu10} and \citet{Can17} for ensuring posterior consistency of Dirichlet process mixtures can be simplified by removing a redundant assumption regarding the finite entropy of the model. This result, proven in Lemma~\ref{lem:improve}, can be of independent interest. Our theoretical results are supported by an extensive simulation study, focusing on both density and clustering estimation. These findings make the DPM-G model a suitable candidate to deal with problems where an informed choice of the relative scale of different dimensions seems prohibitive. We thus fit a DPM-G model to the NGC~2419 dataset and show that it provides interesting insight on the classification problem motivating this work. The rest of the paper is organised as follows. In Section~\ref{sec:model} we describe the modelling framework and introduce the notation used throughout the paper. Section~\ref{sec:theory} presents the main results of the work, with two-fold focus on finite sample properties on the one hand, and large sample asymptotics on the other. A thorough simulation study is presented in Section~\ref{sec:simulation} while Section~\ref{sec:astro} is dedicated to the analysis of the NGC~2419 dataset. Conclusions are discussed in Section~\ref{sec:conclusions}. Proofs of all results are postponed to Appendix~\ref{sec:proofs}. \section{Modelling framework}\label{sec:model} Let $\mathbf{X}^{(n)}:=(\mathbf{X}_1,\dots,\mathbf{X}_n)$ be a sample of $d$-dimensional observations $\mathbf{X}_i:= (X_{i,1},\ldots,X_{i,d})^\intercal$ defined on some probability space $(\Omega,\mathscr{A},\mathds{P})$ and taking values in $\mathds{R}^d$. Consider an invertible affine transformation $g:\mathds{R}^d \longrightarrow \mathds{R}^d$, that is $g(\mathbf{x})=\mathbf{C}\mathbf{x}+\mathbf{b}$ where $\mathbf{C}$ is an invertible matrix of dimension $d\times d$ and $\mathbf{b}$ a $d$-dimensional column vector. The nature of the transformation $g$ is such that, if applied to a random vector $\mathbf{X}$ with probability density function $f$, it gives rise to a new random vector $g(\mathbf{X})$ with probability density function $f_g=|\det(\mathbf{C})|^{-1} f\circ g^{-1}$. Henceforth we denote by $\mathscr{F}$ the space of all density functions with support on $\mathds{R}^d$. The DPM model \citep{Lo84} defines a random density taking values in $\mathscr{F}$ as \begin{equation* \tilde{f}(\mathbf{x}) = \int_{\Theta} k(\mathbf{x};\boldsymbol{\theta})\mathrm{d} \tilde P(\boldsymbol{\theta}), \end{equation*} where $k(\mathbf{x};\boldsymbol{\theta})$ is a kernel on $\mathds{R}^d$ parameterized by $\boldsymbol{\theta}\in\Theta$, $\tilde P$ is a Dirichlet process (DP) with parameters $\alpha$ (precision parameter) and $P_0:=\mathds{E}[\tilde P]$ (base measure), a distribution defined on $\Theta$ \citep{Fer73}. The almost sure discreteness of $\tilde P$ allows the random density $\tilde f$ to be rewritten as \begin{equation* \tilde{f}(\mathbf{x}) = \sum_{i=1}^{\infty} w_i k(\mathbf{x};\boldsymbol{\theta}_i), \end{equation*} where the random atoms $\boldsymbol{\theta}_i$ are i.i.d.~from $P_0$, and the random jumps $w_i$, independent of the atoms, admit the following stick-breaking representation \citep{Set94}: given a set of random weights $v_i\stackrel{\mathrm{iid}}{\sim}\text{Beta}(1,\alpha)$ (independent of the atoms $\boldsymbol{\theta}_i$), then $w_1=v_1$ and, for $j\geq 2$, $w_j=v_j \prod_{i=1}^{j-1}(1-v_i)$. While several kernels $k(\mathbf{x};\boldsymbol{\theta})$ have been considered in the literature, including e.g. skew-normal \citep{canale2016bayesian}, Weibull \citep{kottas2006nonparametric}, Poisson \citep{krnjajic2008parametric}, here we focus on the convenient and commonly adopted Gaussian specification of \citet{Esc95} and \citet{Mue96}. In the latter case, $k(\mathbf{x};\boldsymbol{\theta})$ represents a $d$-dimensional Gaussian kernel $\phi_d(\mathbf{x}; \boldsymbol{\mu},\boldsymbol{\Sigma})$, provided that $\boldsymbol{\theta}=(\boldsymbol{\mu},\boldsymbol{\Sigma})$, where the column vector $\boldsymbol{\mu}$ and the matrix $\boldsymbol{\Sigma}$ represent, respectively, mean vector and covariance matrix of the Gaussian kernel. This specification defines the model referred to as $d$-dimensional location-scale Dirichlet process mixture of Gaussians (DPM-G), which can be represented in hierarchical form as \begin{gather \mathbf{X}_i\mid \boldsymbol{\theta}_i=(\boldsymbol{\mu}_i,\boldsymbol{\Sigma}_i) \stackrel{\mathrm{ind}}{\sim} \phi_d(\mathbf{x}_i; \boldsymbol{\mu}_i,\boldsymbol{\Sigma}_i),\nonumber\\ \boldsymbol{\theta}_i\mid \tilde P \stackrel{\mathrm{iid}}{\sim} \tilde P,\label{eq:DPmmls}\\ \tilde P \sim DP(\alpha,P_0).\nonumber \end{gather} The almost sure discreteness of $\tilde P$ implies that the vector $\boldsymbol{\theta}^{(n)}:=(\boldsymbol{\theta}_1,\ldots,\boldsymbol{\theta}_n)$ might show ties with positive probability, thus leading to a partition of $\boldsymbol{\theta}^{(n)}$ into $K_n\leq n$ distinct values. This, in turn, leads to a partition of the set of observations $\mathbf{X}^{(n)}$, obtained by grouping two observations $\mathbf{X}_{i_1}$ and $\mathbf{X}_{i_2}$ together if and only if $\boldsymbol{\theta}_{i_1}=\boldsymbol{\theta}_{i_2}$. This observation implies that the posterior distribution of the random density $\tilde f$ carries useful information on the clustering structure of the data, thus making DPM-G models convenient tools for density and clustering estimation problems. Although other specifications for the base measure can be considered \citep[see, e.g.,][]{Gor10}, we choose to work within the framework set forth by \cite{Mue96} where $P_0$ is defined as the product of two independent distributions for the location parameter $\boldsymbol{\mu}$ and the scale parameter $\boldsymbol{\Sigma}$, namely a multivariate normal and an inverse-Wishart distribution, that is \begin{equation}\label{eq:base} P_0(\mathrm{d} \boldsymbol{\mu}, \mathrm{d} \boldsymbol{\Sigma};\boldsymbol{\pi})= N_d(\mathrm{d} \boldsymbol{\mu}; \mathbf{m}_0, \mathbf{B}_0) \times IW(\mathrm{d} \boldsymbol{\Sigma}; \nu_0, \mathbf{S}_0). \end{equation} For the sake of compactness, we use the notation $\boldsymbol{\pi}:= (\mathbf{m}_0, \mathbf{B}_0, \nu_0, \mathbf{S}_0)$ to denote the vector of hyperparameters characterising the base measure $P_0$. We denote by $\Pi$ the prior distribution induced on $\mathscr{F}$ by the DPM-G model \eqref{eq:DPmmls} with base measure \eqref{eq:base}. \section{Theoretical results\label{sec:theory}} \subsection{DPM-G model and affine transformations of the data}\label{sec:affine} Let $\tilde{f}_{\boldsymbol{\pi}}$ be a DPM-G model defined in \eqref{eq:DPmmls}, with base measure \eqref{eq:base} and hyperparameters $\boldsymbol{\pi}$. The next result shows that, for any invertible affine transformation $g(\mathbf{x})= \mathbf{C} \mathbf{x}+\mathbf{b}$, there exists a specification $\boldsymbol{\pi}_{g}:=(\mathbf{m}_0^{(g)},\mathbf{B}_0^{(g)},\nu_0^{(g)},\mathbf{S}_0^{(g)})$ of the hyperparameters characterising the base measure in \eqref{eq:base}, such that the deterministic relation $\tilde f_{\boldsymbol{\pi}_g}=|\det(\mathbf{C})|^{-1}\tilde f_{\boldsymbol{\pi}}\circ g^{-1}$ holds. That is, for every $\omega\in\Omega$ and given a random vector $\mathbf{X}$ distributed according to $\tilde{f}_{\boldsymbol{\pi}}(\omega)$, we have that $\tilde{f}_{\boldsymbol{\pi}_g}(\omega)$ is the density of the transformed random vector $g(\mathbf{X})$. \begin{prop}\label{prop:exa} Let $\tilde{f}_{\boldsymbol{\pi}}$ be a location-scale DPM-G model defined as in \eqref{eq:DPmmls}, with base measure \eqref{eq:base} and hyperparameters $\boldsymbol{\pi}= (\mathbf{m}_0, \mathbf{B}_0, \nu_0, \mathbf{S}_0)$. For any invertible affine transformation $g(\mathbf{x})=\mathbf{C}\mathbf{x}+\mathbf{b}$, we have the deterministic relation \begin{equation*} \tilde f_{\boldsymbol{\pi}_g}=|\det(\mathbf{C})|^{-1}\tilde f_{\boldsymbol{\pi}}\circ g^{-1}, \end{equation*} where $\boldsymbol{\pi}_{g} :=(\mathbf{C} \mathbf{m}_0+\mathbf{b},\mathbf{C} \mathbf{B}_0 \mathbf{C}^\intercal,\nu_0,\mathbf{C} \mathbf{S}_0 \mathbf{C}^\intercal)$. \end{prop} While Proposition~\ref{prop:exa} can be derived from general properties of the Dirichlet process \citep[see][]{lijoi2009distributional}, a direct proof is provided in Appendix~\ref{subsec:pro1}. This result implies that, for any invertible affine transformation $g$, modelling the set of observations $\mathbf{X}^{(n)}$ with a DPM-G model \eqref{eq:DPmmls}, with base measure \eqref{eq:base} and hyperparameters $\boldsymbol{\pi}$, is equivalent with assuming the same model with transformed hyperparameters $\boldsymbol{\pi}_g$, for the transformed observations $g(\mathbf{X})^{(n)}:=(g(\mathbf{X}_1),\ldots,g(\mathbf{X}_n))$. As a by-product, the same posterior inference can be drawn conditionally on both the original and the transformed set of observations, as the conditional distribution of the random density $\tilde f_{\boldsymbol{\pi}_g}$, given $g(\mathbf{X})^{(n)}$, coincides with the conditional distribution of $|\det(\mathbf{C})|^{-1}\tilde{f}_{\boldsymbol{\pi}} \circ g^{-1} $, given $\mathbf{X}^{(n)}$. Proposition~\ref{prop:exa} thus provides a formal justification for the procedure of transforming data, e.g. via standardisation or normalisation, often adopted to achieve numerical efficiency: as long as the prior specification of the hyperparameters of a DPM-G model respects the condition of Proposition~\ref{prop:exa}, transforming the data does not affect posterior inference. \subsubsection{Empirical Bayes approach} The elicitation of an honest prior, thus independent of the data, for the hyperparameters $\boldsymbol{\pi}$ of the base measure \eqref{eq:base} of a DPM model is in general a difficult task. A popular practice, therefore, consists in setting the hyperparameters equal to some empirical estimates $\hat{\boldsymbol{\pi}}(\mathbf{X}^{(n)})$, by applying the so-called empirical Bayes approach \citep[see, e.g.,][]{Leh06}. Recent investigations \citep{Pet14,Don18} provide a theoretical justification of this hybrid procedure by shedding light on its asymptotic properties. We show here that this procedure satisfies the assumptions of Proposition~\ref{prop:exa} and, thus, guarantees that posterior Bayesian inference, under an empirical Bayes approach, is not affected by affine transformations to the data.\\ A commonly used empirical Bayes approach for specifying the hyperparameters $\boldsymbol{\pi}$ of a DPM-G model, defined as in \eqref{eq:DPmmls} and \eqref{eq:base}, consists in setting \begin{equation}\label{eq:EB} \mathbf{m}_0=\overline{\mathbf{X}},\qquad\quad \mathbf{B}_0=\frac{1}{\gamma_1}\mathbf{S}_\mathbf{X}^2,\qquad\quad \mathbf{S}_0= \frac{\nu_0 - d - 1}{\gamma_2} \mathbf{S}_\mathbf{X}^2, \end{equation} where $\overline{\mathbf{X}}=\sum_{i=1}^n \mathbf{X}_i/n$ and $\mathbf{S}_\mathbf{X}^2=\sum_{i=1}^n(\mathbf{X}_i-\overline{\mathbf{X}})(\mathbf{X}_i-\overline{\mathbf{X}})^\intercal/(n-1)$ are the sample mean vector and the sample covariance matrix, respectively, and $\gamma_1,\gamma_2>0$, $\nu_0>d+1$. This specification for the hyperparameters $\boldsymbol{\pi}$ has a straightforward interpretation. Namely, the parameter $\mathbf{m}_0$, mean of the prior guess distribution of $\boldsymbol{\mu}$, can be interpreted as the overall mean value and, in absence of available prior information, set equal to the observed sample mean. Similarly, the parameter $\mathbf{B}_0$, covariance matrix of the prior guess distribution of $\boldsymbol{\mu}$, is set equal to a penalised version of the sample covariance matrix $\mathbf{S}^2_\mathbf{X}$, where $\gamma_1$ takes on the interpretation of the size of the ideal prior sample upon which the prior guess on the distribution of $\boldsymbol{\mu}$ is based. Similarly, the hyperparameter $\mathbf{S}_0$ is set equal to a penalised version of the sample covariance matrix $\mathbf{S}_\mathbf{X}^2$, choice that corresponds to the prior guess that the covariance matrix of each component of the mixture coincides with a rescaled version of the sample covariance matrix. Specifically, $\mathbf{S}_0= \mathbf{S}^2_\mathbf{X} (\nu_0 - d - 1)/\gamma_2$ follows by setting $\mathds{E}[\boldsymbol{\Sigma}]=\mathbf{S}_\mathbf{X}^2/\gamma_2$ and observing that, by standard properties of the inverse-Wishart distribution, $\mathds{E}[\boldsymbol{\Sigma}]=\mathbf{S}_0/(\nu_0 - d - 1)$. Finally the parameter $\nu_0$ takes on the interpretation of the size of an ideal prior sample upon which the prior guess $\mathbf{S}_0$ is based. Next we focus on the setting of the hyperparameters $\boldsymbol{\pi}_g$, given the transformed observations $g(\mathbf{X})^{(n)}$. The same empirical Bayes procedure adopted in \eqref{eq:EB} leads to \begin{equation*}\label{eq:EB2} \mathbf{m}_0^{(g)}=\overline{g(\mathbf{X})}=\mathbf{C} \mathbf{m}_0 + \mathbf{b},\qquad \mathbf{B}^{(g)}_0=\frac{1}{\gamma_1}\mathbf{S}_{g(\mathbf{X})}^2,\qquad \mathbf{S}^{(g)}_0= \frac{\nu_0 - d - 1}{\gamma_2} \mathbf{S}_{g(\mathbf{X})}^2. \end{equation*} Observing that $\mathbf{S}_{g(\mathbf{X})}^2=\mathbf{C} \mathbf{S}_\mathbf{X}^2 \mathbf{C}^\intercal$ and setting $\nu_0^{(g)}=\nu_0$ shows that the described empirical Bayes procedure corresponds to $\boldsymbol{\pi}_g=(\mathbf{C} \mathbf{m}_0+\mathbf{b},\mathbf{C} \mathbf{B}_0 \mathbf{C}^\intercal,\nu_0,\mathbf{C} \mathbf{S}_0 \mathbf{C}^\intercal)$ and, thus, by Proposition~\ref{prop:exa}, $\tilde f_{\boldsymbol{\pi}_g}=|\det(\mathbf{C})|^{-1}\tilde f_{\boldsymbol{\pi}}\circ g^{-1}$. \subsection{Large $n$ asymptotic robustness}\label{sec:asymptotics} We investigate the effect of affine transformations of the data on DPM-G models by studying the asymptotic behaviour of the resulting posterior distribution in the large sample size regime. To this end, we fit the same DPM-G model $\tilde{f}_{\boldsymbol{\pi}}$, defined in \eqref{eq:DPmmls} and \eqref{eq:base}, to two versions of the data, that is $\mathbf{X}^{(n)}$ and $g(\mathbf{X})^{(n)}$, by using the exact same specification for the hyperparameters $\boldsymbol{\pi}$. Under this setting, the assumptions of Proposition~\ref{prop:exa} are not met and the posterior distributions obtained by conditioning on the two sets of observations are different random distributions which, thus, might lead to different statistical conclusions. The main result of this section shows that, under mild conditions on the true generating distribution of the observations, the posterior distributions obtained by conditioning $\tilde{f}_{\boldsymbol{\pi}}$ on the two sets of observations $\mathbf{X}^{(n)}$ and $g(\mathbf{X})^{(n)}$, become more and more similar, up to an affine reparametrisation, as the sample size $n$ grows. More specifically we show that the probability mass of the joint distribution of these two conditional random densities concentrates in a neighbourhood of $\{(f_1,f_2)\in\mathcal{F}\times\mathcal{F}\text{ s.t. }f_1=|\det(\mathbf{C})| f_2 \circ g\}$ as $n$ goes to infinity. Henceforth we will say that the DPM-G model \eqref{eq:DPmmls} with base measure \eqref{eq:base} is asymptotically robust to affine transformation of the data. The rest of the section formalises and discusses this result. We consider a metric $\rho$ on $\mathscr{F}$ which can be equivalently defined as the Hellinger distance $\rho(f_1,f_2)=\{\int (\sqrt{f_1(\mathbf{x})}-\sqrt{f_2(\mathbf{x})})^2\mathrm{d} \mathbf{x} \}^{1/2}$ or the $L^1$ distance $\rho(f_1,f_2)=\int |f_1(\mathbf{x})-f_2(\mathbf{x}))|\mathrm{d} \mathbf{x}$ between densities $f_1$ and $f_2$ in $\mathscr{F}$, and we denote by $\|\cdot\|$ the Euclidean norm on $\mathds{R}^d$. Moreover, we adopt here the usual frequentist validation approach in the large $n$ regime, working `as if' the observations $\mathbf{X}^{(n)}$ were generated from a true and fixed data generating process $F^*$ \citep[see for instance][]{rousseau2016frequentist}. We introduce the notation $F_{n}^*$ to denote the $n$-fold product measure $F^*\times \cdots \times F^*$, and we assume that $F^*$ admits a density function with respect to the Lebesgue measure, denoted by $f^*$. In the setting we consider, the same model $\tilde{f}_{\boldsymbol{\pi}}$ defined in \eqref{eq:DPmmls} and \eqref{eq:base} is fitted to $\mathbf{X}^{(n)}$ and $g(\mathbf{X})^{(n)}$, thus leading to two distinct posterior random densities, with distributions on $\mathscr{F}$ denoted by $\Pi(\,\cdot\, \mid \mathbf{X}^{(n)})$ and $\Pi(\,\cdot\, \mid g(\mathbf{X})^{(n)})$, respectively. We use the notation $\Pi_2(\cdot\mid \mathbf{X}^{(n)})$ to refer to their joint posterior distribution on $\mathscr{F}\times\mathscr{F}$. \begin{theo}\label{thm:asym} Let $f^*\in\mathscr{F}$, true generating density of $\mathbf{X}^{(n)}$, satisfy the conditions \begin{enumerate} \item[{\normalfont A1}.] $0 < f^*(\mathbf{x}) < M$, for some constant $M$ and for all $\mathbf{x} \in \mathds{R}^d$, \item[{\normalfont A2}.] for some $\eta > 0$, $\int \|\mathbf{x}\|^{2(1+\eta)}f^*(\mathbf{x})\mathrm{d}\mathbf{x} < \infty$, \item[{\normalfont A3}.] $\mathbf{x}\mapsto f^*(\mathbf{x})\log^2(\varphi_\delta (\mathbf{x}))$ is bounded on $\mathds{R}^d$, where $\varphi_\delta(\mathbf{x}) = \inf_{\{\mathbf{t}\,:\,\|\mathbf{t}-\mathbf{x}\|<\delta\}}f^*(\mathbf{t})$. \end{enumerate} Let $g:\mathds{R}^d \longrightarrow \mathds{R}^d$ be an invertible affine transformation and $\tilde{f}_{\boldsymbol{\pi}}$ be the random density induced by a DPM-G as \eqref{eq:DPmmls} with base measure \eqref{eq:base} where $\nu_0>(d + 1)(2d - 3)$. Then, for any $\varepsilon>0$, \begin{equation*} \Pi_2((f_1,f_2): \rho(f_1,|\det(\mathbf{C})| f_2 \circ g)<\varepsilon \mid \mathbf{X}^{(n)})\longrightarrow 1 \end{equation*} in $F_{n}^*$-probability, as $n\to\infty$. \end{theo} It is worth stressing that, while in line with the usual posterior consistency approach the existence of a true data generating process $F^*$ is postulated, the focus of Theorem~\ref{thm:asym} is not on the asymptotic behaviour of the posterior distribution with respect to the true data generating process but rather on the relative behaviour of two posterior distributions, obtained by conditioning the same model on two sets of observations which coincide up to an affine transformation. More specifically, according to Theorem~\ref{thm:asym}, when the sample size grows, the joint distribution $\Pi_2(\cdot\mid \mathbf{X}^{(n)})$ concentrates its mass on a subset of the space $\mathscr{F}\times\mathscr{F}$ where the distance $\rho$ between $f_1$ and $|\det(\mathbf{C})| f_2 \circ g$ is smaller than $\varepsilon$. In other terms, the two posterior distributions get similar, up to the affine transformation, as $n$ becomes large. The assumptions of Theorem~\ref{thm:asym} refer to the true generating distribution $f^*$ of $\mathbf{X}^{(n)}$. Assumption A1 requires $f^*$ to be bounded and fully supported on $\mathds{R}^d$. Assumption A2 requires the tails of $f^*$ to be thin enough for some moment of order strictly larger than two to exist. Such an assumption is not met, for example, by a Student's $t$-distribution with two degrees of freedom, case which will be considered in the simulation study of Section~\ref{sec:simulation}. Finally, assumption A3 is a weak condition ensuring local regularity of the entropy of $f^*$. The proof of Theorem~\ref{thm:asym} is based on previous results proved by \cite{Wu08} and \cite{Can17} in order to derive the so-called Kullback--Leibler property at $f^*$ for some mixtures of Gaussians models. Importantly, in Lemma~\ref{lem:improve} (see Appendix~\ref{thm:asym}), we improve upon their results by showing that the set of assumptions required by \cite{Wu08} and \cite{Can17} can be reduced to the simpler set of assumptions A1, A2 and A3 of Theorem~\ref{thm:asym} by removing a redundant assumption. More specifically, we prove that A1, A2 and A3 imply that $f^*$ has finite entropy and regular local entropy, conditions required in the aforementioned works. \section{Simulation study}\label{sec:simulation} We ran an extensive simulation study with a two-fold goal: 1) providing empirical support to our result on the large $n$ asymptotic robustness of a DPM-G model, under affine transformations of the data; 2) investigating whether an analogous robustness property holds when DPM-G models are adopted to make inference on the clustering structure of the data. To this end, we considered two distinct data-generating distributions, which allowed us to highlight different facets of DPM-G models. In the first case, data are generated from a mixture of bivariate Gaussians, distribution which satisfies the conditions of Theorem~\ref{thm:asym}. This study complements our asymptotic result with a numerical investigation of the finite $n$ behaviour of DPM-G models, when data undergo an affine transformation. Moreover, the same data are used to perform a numerical study on the effect of data transformation and sample size on the number of clusters on the estimated partition. While not directly related to theoretical results of Section~\ref{sec:theory}, this part of the study is relevant in view of the astronomical application of Section~\ref{sec:astro} where a DPM-G model will be used for unsupervised clustering. The second scenario we considered does not satisfy the set of assumptions of Theorem~\ref{thm:asym}, as data are generated from univariate Student's $t$-distribution with two degrees of freedom, thus breaking assumption A2. Our study, in this case, aims at assessing the robustness of DPM-G models when the sufficient conditions of Theorem~\ref{thm:asym} are not met. \subsection{Data from mixture of Gaussians}\label{subsec:mix} The first part of the simulation study focuses on the analysis of data generated from a mixture of Gaussians. Specifically, we considered three sample sizes, namely $n = 100$, $n = 300$ and $n = 1\,000$, and we generated 100 samples $\mathbf{X}^{(n)}$, for each $n$, from a mixture of two Gaussian components with density function \begin{equation* f(\mathbf{x}) = \frac{1}{2} \phi_2\left(\mathbf{x}; \mathbf{m}_1, \mathbf{S}_1\right) + \frac{1}{2} \phi_2\left(\mathbf{x};\mathbf{m}_2,\mathbf{S}_2\right), \end{equation*} where $\phi_2(\cdot;\mathbf{m},\mathbf{S})$ denotes the density function of a two-dimensional Gaussian distribution with mean vector $\mathbf{m}$ and covariance matrix $\mathbf{S}$, and the two components of the mixture are characterized by the parameters \begin{equation*} \mathbf{m}_1=(-2,-2),\quad\mathbf{S}_1=\begin{bmatrix} 1 & 0.85 \\ 0.85 & 1 \end{bmatrix},\quad\mathbf{m}_2=(2,2),\quad \mathbf{S}_2=\begin{bmatrix} 1 & 0 \\ 0 & 1 \end{bmatrix}.\end{equation*} \noindent In order to test the robustness of the model under affine transformations of the data, we compressed or stretched the generated datasets by using five different constants, namely $c = 1/5$, $c = 1/2$, $c = 1$, $c = 2$ and $c = 5$. For each constant, we multiplied the simulated data by $c$, thus % obtaining a transformed dataset $\mathbf{X}_c^{(n)} := c \mathbf{X}^{(n)}$. We then fitted a DPM-G model, specified as in \eqref{eq:DPmmls} and \eqref{eq:base}, to each one of the $5\times 3\times 100 = 1\,500$ resulting datasets. In order to enhance the flexibility of the model, we completed its specification by setting a normal/inverse-Wishart prior distribution for the hyperparameters $(\mathbf{m}_0,\mathbf{B}_0)$ of the base measure \eqref{eq:base}. Namely, we set $\mathbf{B}_0 \sim IW(4, \text{diag}(\mathbf{15}))$ and $\mathbf{m}_0\mid\mathbf{B}_0 \sim N(0, \mathbf{B}_0)$, specification chosen so that $\mathds{E}[\boldsymbol{\mu}] = \mathbf{0}$ and to guarantee a prior guess on the location component $\boldsymbol{\mu}$ flat enough to cover the support of the non-transformed data. As for the scale component of the base measure \eqref{eq:base}, we set $(\nu_0,\mathbf{S}_0)=(4, \text{diag}(\mathbf 1))$. Finally, the precision parameter $\alpha$ of the Dirichlet process was set equal to 1. Realisations of the mean of the posterior distribution were obtained by means of a Gibbs sampler relying on a Blackwell--McQueen P\'olya urn scheme \citep[see][]{Mue96}, implemented in the \href{https://github.com/rcorradin/AFFINEpack}{\textsf{AFFINEpack}} \textsf{R} package\footnote{\label{foot:package}The package is available at \href{https://github.com/rcorradin/AFFINEpack}{https://github.com/rcorradin/AFFINEpack} and can be installed via \textsf{devtools}. For reproducibility, the code is available at \href{https://github.com/rcorradin/Affine}{https://github.com/rcorradin/Affine}.}. For each replicate, posterior inference was drawn based on $5\,000$ iterations, obtained after discarding the first $2\,500$. Convergence of the chains was assessed by visually investigating the traceplots of some randomly selected replicates, which did not provide indication against it. \begin{figure}[h] \centering \includegraphics[width=0.99\textwidth]{Fig2} \caption{Simulation study, data generated from a mixture of Gaussians. Based on a single replicate of the samples $\mathbf{X}^{(100)}$, $\mathbf{X}^{(300)}$ and $\mathbf{X}^{(1\,000)}$, scatter plots of the data (grey dots), contour plots of the estimated densities based on a DPM-G model (red curves) and contour plots for the expected prior density (blue filled curves). Left to right: rescaling constant $c=1/5$, $c=1/2$, $c=1$, $c=2$, $c=5$. Top to bottom: sample size $n = 100$, $n = 300$, $n = 1\,000$.} \label{fig:splotsim} \end{figure} Figure~\ref{fig:splotsim} shows, for every $n\in\{100,300,1\,000\}$ and $c\in\{1/5,1/2,1,2,5\}$, a contour plot of the estimated posterior densities. The difference between estimated densities, across different values of $c$, is apparent when $n=100$, with the two extreme cases, namely $c=1/5$ and $c=5$, displaying very different contour lines and possibly suggesting a different number of modes in the estimated density. For larger sample sizes, this difference is less evident and, when $n=1\,000$, the contour plots are hardly distinguishable. These qualitative observations are in agreement with the large $n$ asymptotic results of Theorem~\ref{thm:asym}. The plots of Figure~\ref{fig:splotsim} refer to a single realisation of the samples $\mathbf{X}^{(100)}$, $\mathbf{X}^{(300)}$ and $\mathbf{X}^{(1\,000)}$ considered in the simulation study, although qualitatively similar results can be found in almost any replicate. The findings drawn from a visual inspection of Figure~\ref{fig:splotsim} were confirmed by assessing the distance between estimated posterior densities. Specifically, for any considered sample size $n$ and for any pair of values $c_1$ and $c_2$ taken by the constant $c$, we approximately evaluated the $L^1$ distance between the suitably rescaled estimated posterior densities obtained conditionally on $\mathbf{X}_{c_1}^{(n)}$ and on $\mathbf{X}_{c_2}^{(n)}$. The results of such analysis are shown in Figure~\ref{fig:L1plot} and indicate that, as the sample size grows, the difference in terms of $L^1$ distance strictly decreases. \begin{figure}[h] \centering \includegraphics[width=0.99\textwidth]{Fig3} \caption{Simulation study, data generated from a mixture of Gaussians. Normalised $L^1$ distances (all distances are divided by the largest observed distance) between suitably rescaled estimated densities, conditionally on data rescaled by means of different constants $c_1$ ($X$ axis) and $c_2$ ($Y$ axis), where $c_1$ and $c_2$ denote the scaling factors used to transform the data, averaged over $100$ replications. Left to right: sample size $n=100$, sample size $n=300$, sample size $n=1\,000$.} \label{fig:L1plot} \end{figure} The posterior distribution of the random density induced by a DPM-G model provides interesting insight also on the clustering structure of the data. The second goal of the simulation study, thus, consisted in investigating the impact of the scaling factor $c$ on the estimated number of groups in the partition induced on the data. To this end, for each considered $n$ and $c$, we estimated $\hat K_n^{(\text{VI})}$, the number of groups in the optimal partition estimated using a procedure introduced by \cite{Wad17} and based on the variation of information loss function. In light of known inconsistency results for the posterior distribution of the number of components under a DPM-G model \citep[see, for instance,][]{miller2013simple}, the numerical findings of this part of the simulation study contribute to shed some light on the large $n$ behaviour of $\hat{K}_n^{\text{(VI)}}$. The average values for this quantity, over 100 replicates, are reported in Table~\ref{tab:simresult}. \begin{table}[h] \caption{Simulation study, data generated from a mixture of Gaussians. Averages over 100 replicates for $\hat K_n^{(\text{VI})}$, the number of clusters of the optimal partition estimated by means of \citet{Wad17}'s variation of information method. Left to right: rescaling constant $c=1/5$, $c=1/2$, $c=1$, $c=2$, $c=5$. Top to bottom: sample size $n = 100$, $n = 300$, $n = 1\,000$.} \label{tab:simresult} \centering \begin{tabular}{clllll} \hline \multicolumn{1}{l}{} & $c = 1/5$ & $c = 1/2$ & $c = 1$ & $c = 2$ & $c = 5$ \\ \hline \multicolumn{1}{l}{$n=100$} &1.81 &2.04 &2.84 &5.96 &10.52 \\ \multicolumn{1}{l}{$n=300$} &2.00 &2.03 &2.20 &2.82 &5.18 \\ \multicolumn{1}{l}{$n=1\,000$} &2.00 &2.00 &2.04 &2.05 &2.12 \\ \hline \end{tabular} \end{table} There appears to be a clear trend suggesting that a larger scaling constant $c$ leads to a larger $\hat K_n^{(\text{VI})}$: this finding is consistent with the fact that, if the data are stretched while the prior specification is kept unchanged, then we expect the estimated posterior density to need a larger number of Gaussian components to cover the support of the sample. For the purpose of this simulation study the main quantity of interest is the ratio between the estimated number of groups under any two distinct values $c_1$ and $c_2$ for the scaling constant $c$, that is $\hat K_{n,c_1}^{(\text{VI})}/\hat K_{n,c_2}^{(\text{VI})}$. The results presented in Table~\ref{tab:simresult} clearly indicate that, as the sample size $n$ becomes large, such ratios tend to approach $1$. This suggests that the large $n$ robustness property of the DPM-G model nicely translates to an equivalent notion of robustness in terms of the estimated number of groups $\hat K_{n}^{(\text{VI})}$ in the data. \subsection{Data from Student's $t$-distribution}\label{sec:misst} The second part of the simulation study deals with the same simulation scenarios ($n\in\{100,300,1\,000\}$ and $c=\{1/5,1/2,1,2,5\}$) considered in Section~\ref{subsec:mix}, with the difference that data are generated from a Student's $t$-distribution with two degrees of freedom. It is important to stress that such a distribution does not have finite variance and therefore does not meet assumption A2 of Theorem~\ref{thm:asym}. Also in this case we considered 100 replicates for each considered simulation scenario. We analysed each dataset with a univariate version of the DPM-G model specified in \eqref{eq:DPmmls} and \eqref{eq:base}. That is, we considered a univariate Gaussian kernel and a base measure defined as the product of two independent distributions, a univariate normal distribution for the location parameter $\mu\sim N(m_0, s_0^2)$ and an inverse-gamma distribution for the scale parameter $\sigma^2\sim IG(a_0,b_0)$. The model specification is completed by setting $a_0 = 2$ and $b_0 = 1$, so that $\mathds{E}[\sigma^2] = 1$, and by considering a normal/inverse-gamma distribution for the hyperparameters $(m_0, s_0^2)$, specifically $s_0^2 \sim IG(2,1)$ and $m_0 \mid s_0^2 \sim N(0, s_0^2)$. Finally, the precision parameter $\alpha$ of the Dirichlet process was set equal to 1. Realisations of the mean of the posterior distribution were obtained by means of a Gibbs sampler relying on a Blackwell--McQueen P\'olya urn scheme\footnote{See footnote~\ref{foot:package}.}. Posterior inference was drawn based on $5\,000$ iterations, after a burn-in period of $2\,500$ iterations. We assessed the convergence of the chains by visually investigating traceplots, which did not provide indication against it. Also for these data, for any considered sample size $n$ and for any pair of values $c_1$ and $c_2$ taken by the constant $c$, we approximately evaluated the $L^1$ distance between the suitably rescaled estimated posterior densities obtained conditionally on $\mathbf{X}_{c_1}^{(n)}$ and on $\mathbf{X}_{c_2}^{(n)}$. The results of such analysis are displayed in Figure~\ref{fig:L1plot_tstudent} and indicate that, as the sample size grows, the $L^1$ distance decreases. This qualitative findings suggest that asymptotic robustness might hold also for data generated from a distribution not meeting the assumptions of Theorem~\ref{thm:asym}. \begin{figure}[h] \centering \includegraphics[width=0.99\textwidth]{Fig4} \caption{Simulation study, data generated from a Student's $t$-distribution. Normalised $L^1$ distances (all distances are divided by the largest observed distance) between suitably rescaled estimated densities, conditionally on data rescaled by means of different constants $c_1$ ($X$ axis) and $c_2$ ($Y$ axis), where $c_1$ and $c_2$ denote the scaling factors used to transform the data, averaged over $100$ replications. Left to right: sample size $n=100$, sample size $n=300$, sample size $n=1\,000$.} \label{fig:L1plot_tstudent} \end{figure} \section{Astronomical data}\label{sec:astro} The large $n$ asymptotic robustness to affine transformation of the DPM-G model makes it a suitable candidate also for analysing data whose components are not commensurable and for which an informed choice of the relative scale of different dimensions seems prohibitive. We fitted the DPM-G model, specified as in \eqref{eq:DPmmls} and with base measure \eqref{eq:base}, to the NGC~2419 dataset described in Section~\ref{sec:intro}. The ultimate goal of our analysis consists in classifying stars as belonging to the NGC~2419 globular cluster or as being contaminants: an accurate classification is crucial for the astronomers to study the dynamics of the globular cluster. Since the large majority of the stars in the dataset is expected to belong to the globular cluster, with only a few of them being contaminants, we will identify the globular cluster as the largest group in the estimated partition of the dataset. Prior to any analysis, data were standardised component by component, the legitimacy of such procedure following from the robustness results of Theorem~\ref{thm:asym}. Hyperprior distributions were specified for the location parameter of the base measure \eqref{eq:base} and on the DP precision parameter $\alpha$. Specifically, $\mathbf{B}_0 \sim IW(6, \text{diag}(\mathbf{15}))$ and $\mathbf{m}_0\mid\mathbf{B}_0 \sim N(0, \mathbf{B}_0)$, specification chosen to guarantee a prior guess on the location component $\boldsymbol{\mu}$ flat enough to cover the support of the data and centered at $\mathbf 0$. In addition, $\alpha$ was given a gamma prior distribution with unit shape parameter and rate parameter equal to $5.26$, so that, a priori, $\alpha_0:=\mathds{E}[\alpha] \simeq 0.19$. This leads to an expected number of components $K_n$ in a sample of size $n=139$ from a DP equal to $\sum_{i=1}^{n}\alpha_0/(\alpha_0 + i - 1) \simeq 2$, thus reflecting the prior opinion of astronomers who would expect two distinct groups of stars in the dataset. Finally, as far as the scale component of the base measure \eqref{eq:base} is concerned, we set $(\nu_0,\mathbf{S}_0)=(26,\text{diag}(\mathbf{21}))$, where the number of degrees of freedom $\nu_0=26$ of the inverse-Wishart distribution was chosen so that to satisfy the conditions of Theorem~\ref{thm:asym} and, in turn, the scale matrix $\mathbf{S}_0=\text{diag}(\mathbf{21})$ so that $\mathds{E}[\boldsymbol{\Sigma}] = \text{diag} (\mathbf 1)$. Realisations of the mean of the posterior distribution were obtained by means of a Gibbs sampler relying on a Blackwell--McQueen P\'olya urn scheme\footnote{See footnote~\ref{foot:package}.}. In turn, posterior inference was drawn based on $20\,000$ iterations, after a burn-in period of $5\,000$ iterations. Convergence of the chains was assessed by visually investigating traceplots, which did not provide indication against it. \begin{figure}[h] \centering \includegraphics[width=0.99\textwidth]{BP} \caption{NGC~2419 data. Contour plots of the bivariate log marginal densities estimated via DPM-G model (log densities are used for better visualization). Partition estimated via DPM-G model combined with \citet{Wad17}'s variation of information method. Five groups are detected: the largest group (grey dots), group A (blue triangles), group B (red triangles), group C (one orange triangle), group D (one green triangle).} \label{fig:dens} \end{figure} Figure~\ref{fig:dens} displays contour plots for the six two-dimensional projections of the estimated posterior density, with the scatter plots of the dataset with individual observations coloured according to their membership in the optimal partition estimated via the variation of information method of \citet{Wad17}, and labeled as main group (grey circles) and other groups (coloured triangles). The estimated partition is composed of five groups. The largest one, identified as the globular cluster, consists of 124 stars. The remaining 15 stars are thus considered contaminants and are further divided into four groups, one composed by eight stars (group A), one containing five stars (group B) and two singletons (groups C and D). A visual investigation of Figure~\ref{fig:dens} suggests that stars in group A differ from those in the globular cluster in terms of metallicity and position, with the contaminants characterised by larger values for $[{\rm Fe/H}]$ and smaller values for $Y_1$ and $Y_2$. Stars in group B differ from the globular cluster in terms of velocity and metallicity, with the contaminants showing larger values for $V$ and $[{\rm Fe/H}]$. Finally, groups C and D are singletons, the first one being characterised by a high metallicity and an extremely small value for the velocity, the second one showing large values for both metallicity and location $Y_1$. Our unsupervised statistical clustering can be compared to the clustering of \citet{Iba11} (described in their Figure 4) obtained by means of ad hoc physical considerations. Specifically, once the best fitting physical model, in the class of either Newtonian or Modified Newtonian Dynamics models, is detected, they use it in order to compute the average values of the physical variables describing the stars. Stars are then assigned to the globular cluster based on a comparison between their velocity and the average model velocity: those lying close enough are deemed to belong to the cluster, while the others are considered as potential contaminants. For the latter, the evidence of being contaminants is measured by evaluating how distant their metallicity is from the average model one. Two classifications are then proposed: the first one assigns to the globular cluster only the 118 stars for which the evidence seems strong, the second and less conservative strategy classifies as belonging to the globular cluster a total of 130 stars. Following this distinction and for the sake of simplicity, we summarise the results of \citet{Iba11}'s analysis, by devising three groups of stars: \begin{itemize} \item[-] \emph{globular cluster}: 118 stars deemed to belong to the globular cluster, \item[-] \textit{likely globular cluster}: 12 stars assigned to the globular cluster only when the less conservative procedure is adopted, \item[-] \textit{contaminants}: 9 stars with strong evidence of being contaminants. \end{itemize} \renewcommand{\arraystretch}{1.5} \begin{table}[h] \caption{NGC~2419 data. Comparison between the groups identified by \citet{Iba11} and the groups estimated via DPM-G model.} \label{tab:VSibata} \centering \begin{tabular}{llc*{5}{c}} \toprule & && \multicolumn{5}{c}{DPM-G groups} \\ & && \textit{largest} & \textit{A} & \textit{B} & \textit{C} & \textit{D} \\ \midrule &&\multicolumn{1}{c|}{\textit{total}}&\textit{124}&\textit{8}&\textit{5}&\textit{1}&\textit{1}\\ \cline{3-8} \multirow{3}{*}{\rotatebox[origin=c]{90}{Ibata et al.}\hspace{.3cm}\rotatebox[origin=c]{90}{groups}} & \textit{globular cluster} &\multicolumn{1}{c|}{\textit{118}}& 114 & 4 & 0 &0 &0 \\ & \textit{likely globular cluster} &\multicolumn{1}{c|}{\textit{12}}& 10 & 1 &0 &0 & 1 \\ & \textit{contaminants} & \multicolumn{1}{c|}{\textit{9}}& 0 & 3 &5 & 1 &0 \\ \bottomrule \end{tabular} \end{table} \def\Ibata{Ibata et al.\xspace} For the purpose of comparison, we report in Table~\ref{tab:VSibata} the confusion matrix of the groups obtained via the DPM-G model against the groups detected by \Ibata All of the 124 stars belonging to the largest group of the partition estimated based on the DPM-G model belong to the groups identified as \textit{globular cluster} or \textit{likely globular cluster} by \Ibata At the same time, out of the nine stars classified as contaminants by \Ibata, the approach based on the DPM-G model assigns none to the globular cluster, three to group A, five stars to group B, which is composed only by stars considered contaminants in \Ibata, and the star of group C, which shows an extremely small value for the velocity variable. Finally, group D contains only one star, which is not considered a contaminant by \Ibata \begin{figure}[h] \centering \begin{subfigure}{.49\textwidth} \centering \includegraphics[width=1\textwidth]{LP} \end{subfigure} \begin{subfigure}{.49\textwidth} \centering \includegraphics[width=1\textwidth]{UP} \end{subfigure}\caption{\label{fig:uncert}NGC~2419 data. Lower bound (left) and upper bound (right) of the credible ball on the partitions' space, estimated via DPM-G model combined with \citet{Wad17}'s variation of information method.}% \end{figure} In order to characterize the uncertainty associated to the estimated optimal partition displayed in Figure~\ref{fig:dens}, we considered a 95\% posterior credible ball in the space of partitions, based on the variation of information metric \citep[see][for details]{Wad17}. Figure~\ref{fig:uncert} shows the vertical lower bound and the vertical upper bound of such credible ball: the first one is the partition in the credible ball which, among those with the largest number of clusters, is the most distant from the optimal partition; the latter one is the partition in the credible ball which, among those with the smallest number of clusters, is the most distant from the optimal one. The lower bound displays a total of $13$ groups, with the largest one, object of interest in our analysis, composed by $119$ observations. The upper bound instead is composed of only $3$ groups, with the largest one counting $121$ observations. The largest groups in the two vertical bounds share $116$ observations, thus showing a limited variability, as far as the size of the largest cluster, main object of our analysis, is concerned. This nicely suggests that the adopted procedure for differentiating stars belonging to the globular cluster and contaminants can be considered robust. Finally, further insight on the clustering structure of the data is provided by Figure~\ref{fig:heat}, which shows the heatmap representation of the posterior similarity matrix obtained from the MCMC output. In agreement with the partition obtained by applying the approach of \citet{Wad17}, one main group, identified with the globular cluster, can be clearly detected in Figure~\ref{fig:heat}. As for the remaining stars, arguably the contaminants, there seem to be two well defined groups, A and B, and a few stars whose group membership is less certain. \begin{figure}[h \centering \includegraphics[width=0.8\textwidth]{Fig7} \caption{NGC~2419 data. Heatmap representation of the posterior similarity matrix obtained based on DPM-G model.}\label{fig:heat} \end{figure} \section{Conclusions}\label{sec:conclusions} The purpose of this paper was to investigate the behaviour of the multivariate DPM-G model when affine transformations are applied to the data. To this end we focused on the DPM-G model with independent normal and inverse-Wishart specification for the base measure. Our investigation covered both the finite sample size and the asymptotic framework. Specifically, in Proposition~\ref{prop:exa}, given any affine transformation $g$, an explicit model specification, depending on $g$, was derived so to ensure coherence between posterior inferences carried out based on a dataset or its transformation via $g$. We then considered a different setting where the specification of the model is assumed independent of the specific transformation $g$. In this case, we formalised the notion of asymptotic robustness of a model under transformations of the data and identified mild conditions on the true data generating distributions which are sufficient to ensure that the DPM-G model features such a property. Specifically, Theorem~\ref{thm:asym} shows that the posterior distributions obtained conditionally on a dataset or any affine transformation of it, become more and more similar as the sample size grows to infinity. Inference on densities and, as suggested by the simulation study, on the clustering structure underlying the data, thus becomes increasingly less dependent on the affine transformation applied to the data, as the sample size grows. As a special case, Theorem~\ref{thm:asym} implies that posterior inference based DPM-G models is asymptotically robust to data transformations commonly adopted for the sake of numerical efficiency, such as standardisation or normalisation. This observation is particularly relevant when dealing with the astronomical unsupervised clustering problem motivating this work. Due to the lack of prior information on the dimensional constants relating different physical units, we resorted to a standardisation of each component of the data and chose an arbitrary model specification. Prior information was available in the form of the experts' prior opinion on the expected number of groups in the dataset and was used to elicit the hyperprior distribution for $\alpha$, the precision parameter of the DP. \section*{Acknowledgements} This work was developed in the framework of the Ulysses Program for French-Irish collaborations (43135ZK) and the Grenoble Alpes Data Institute. The authors wish to thank \href{https://www.unibo.it/sitoweb/carlo.nipoti/en}{Carlo Nipoti} for suggesting the motivating astronomical problem, and \href{http://mistis.inrialpes.fr/people/girard/}{St\'ephane Girard} for helpful discussions on the set of assumptions of Theorem~\ref{thm:asym}. The authors are also grateful to \href{http://bf-astro.com/index.htm}{Bob Franke} for the picture in Figure~\ref{fig:glob}.

c202ef49c4ff62e02c6ae6567fca4928ab8b4d98

\section{Introduction} \label{intro} In this paper we propose an interactive method for probabilistic classification of pixels, which can be used for segmentation of 2D and 3D images. Our approach is especially advantageous for detecting patterns, a situation regularly occurring in images of materials and medical samples. Such images often show a collection of objects which are to be separated from the background. One example we extensively use in this paper is concerned with detection of fibres in volumetric data of composite materials, see Fig~\ref{fig:fibres}. Another example is detection and segmentation of cells in histological images. When segmenting images showing a collection of similar objects, an established strategy involves extensive modelling of the objects' appearance, usually leading to a highly specialised method. Another common strategy is to learn the appearance of the objects from a large amount of prelabelled data, often with high computational requirements during the training phase. Here we aim for a general method that requires limited computation, as well as modest user-labelling. Our method fits into the framework of semi-supervised learning, combining two ingredients: a model for image content created in an unsupervised manner from the image features, and a modest input from the user. When a user marks a structure in the image as belonging to a class, our method propagates the marks to similar structures in the rest of the image. The output is a layered image which at every pixel position contains the probabilities of belonging to each of the defined classes. We call this output \emph{pixelwise probabilities} of belonging to segmentation classes. From pixelwise probabilities, the segmentation is readily obtained by selecting the most probable class for each pixel. The method is highly flexible and captures the features which are of interest to the user; an example with various image features is shown in Fig.~\ref{fig:bag}. Our approach allows easy segmentation of complex structures, that would otherwise require the development of algorithms targeted at specific problems. An important property of our model is real-time feedback, allowing the user to place new markings strategically, depending on the current result. For this to work without delay, the segmentation must be updated very fast. Our method relies on an efficient update of the parameters used for pixel classification, and equally efficient update of the classification results. With results shown promptly, the user can continue adding marks until the desired outcome is learned by the algorithm. Having learned the desired outcome, the classification model can be applied to other images of the same type in an unsupervised manner, that is, without additional user input. Such pipeline has many applications, for example, in microscopy or when segmenting slices from a volumetric image. Our prototype implementation, including a graphical user interface, is in Matlab, and we made the code available through MathWorks File Exchange (search for InSegt) - \emph{to come}. \subsection{Related work} Benefits of user input with real-time feedback have been recognised in image segmentation. A comprehensive summary of interactive approaches can be found in Boykov \citep{boykov2014computer}. Here we review some important advances to place our method in the existing framework, and to explain how our method differs from the current trends in interactive segmentation. Early interactive techniques for segmentation of highly complex images include intelligent scissors \citep{mortensen1995intelligent} or live wire \citep{falcao1998user}, where the user cuts out an object by placing markers along its boundary. Based on edge information, the algorithm traces the boundary by finding the shortest path in an edge-weighted graph. These algorithms are computationally cheap, but require a lot of user effort to obtain a segmentation. Less user input is required when using interactive graph cuts \citep{boykov2001interactive}\cite{boykov2006graph}, which often give very impressive results with only a few seeds provided by the user. The algorithm separates the foreground from the background based on the boundary and region properties of segments. In the GrabCut method \citep{rother2004grabcut} the user provides a bounding rectangle, often leading to very precise foreground-background separation. Optional editing using brush strokes can be carried out to correct finer details. Extensions of GrabCut include shape priors \citep{price2010geodesic} and and improvement to graph cut energy representation \citep{tang2013grabcut}. An alternative to combinatorial graph-based solutions is the use of continuous representation of segmentation boundaries. Such interactive active contours often minimise an energy functional in a variational framework \cite{unger2008tvseg}\cite{santner2011interactive}. Common to the described methods is the focus on segmenting relatively large foreground objects, which justifies using regularisation on the length or the curvature of the segmentation boundary. In some applications it is, however, not possible to use a strong regulariser. For example, when segmenting the circular fibres shown in \myfigref{fig:fibres}, regularisation could remove or merge small regions. The need for segmenting a number of small objects is often seen in areas like microscopy for life science or materials science. Appearance of such images can vary significantly, with texture as well as intensity carrying information that is useful for obtaining the desired segmentation. A specialist would use such clues to distinguish amongst structures, but automating the segmentation task typically requires highly sophisticated and problem-adapted methods. While there are situations which justify the development of a specialised method, in many cases a reasonable result with modest interactive effort would be preferred. When segmenting small image structures, e.g.\ cells, a well-suited approach is classification of pixels. This is the basis for the ilastik segmentation tool \citep{sommer2011ilastik}, which employs a random forest classifier \cite{breiman2001random} trained on image features including colour, edges, orientation and texture. The features are computed from the image before starting the interactive labelling of image structures, while parameters of the random forest classifier are learned from the manual labelling. When a user updates the labels to improve the segmentation, the parameters of the classifier need to be re-learned, which is computationally costly and causes a noticeable delay in the feedback. Another specialised tool for segmentation of microscopy images is the trainable Weka segmentation \cite{arganda2017trainable} (a part of the Fiji \cite{schindelin2012fiji} distribution of ImageJ) which utilises a data mining and machine learning toolkit for solving pixel classification problems. A user can choose from a variety of image features and interactively re-train the classifier. Frameworks using neural networks are increasingly popular in pixel classification, and often yield impressive results \cite{lecun2015deep}. A neural network operates on features extracted locally from the image. This input is fed through a series of multidimensional linear functions, with a non-linear activation between them, ending up in a probabilistic output. The weights of the linear functions need to be trained by optimising the performance on the usually large set of prelabelled data. This provides an extreme flexibility to the method and, provided an adequate training, neural networks may solve pixel classification problems as accurately as specialists. However, neural networks are dependent on large training sets and require computationally costly training, which makes them less convenient for the task of segmenting a small set of images. Our approach shares some similarities with neural networks. We also feed the input through linear functions with non-linear steps in between. However, we use the features extracted from the image to construct the linear functions in a preprocessing step. The functions are then kept fixed, while they operate on the interactively provided user input, resulting in a probabilistic output. Due to the fixed linear functions, our method is not as adaptable as neural networks. For example, our approach is less fit for semantic segmentation of photographs. Nevertheless, we achieve excellent results when segmenting patterned images, without requiring a large set of labelled data and without performing a costly optimisation during interactive update. The foundation of our method is a linear operator encoding image content. The linear operator is described in terms of a dictionary, as it assigns image pixels to dictionary pixels. The relation between the image and the dictionary can be formulated as a bipartite graph and represented using a biadjacency matrix. The approach has been used for evolving deformable models in \cite{dahl2014dictionary}\cite{dahl2015dictionary}\cite{dahl2017probabilistic}. An early version of the method, without the interactive update, proved valuable for quantifying composite materials \cite{emerson2018statistical}. In this work we use the image--dictionary relationship to propagate the brush strokes provided by the users. \setlength{\sub}{0.495\linewidth} \begin{figure} \centering \includegraphics[width=\sub]{MFigures/glass_fibers/glass_labeling_image_image.png}\hspace{\vs}% \includegraphics[width=\sub]{MFigures/glass_fibers/glass_segmentation_image.png}% \\\vspace{\vs}% \includegraphics[width=\sub]{MFigures/carbon_fibers/carbon_labeling_image_image.png}\hspace{\vs}% \includegraphics[width=\sub]{MFigures/carbon_fibers/carbon_segmentation_image.png}% \caption{Detecting glass and carbon fibers using our interactive pattern-based method. On the \emph{left} input images and a very small subset of pixels manually labelled as either being close to a fibre centre (cyan), or not being close to a fibre centre (magenta). On the \emph{right}, the manual labelling has been propagated to the whole image and the result is obtained by selecting the most probable class for each pixel.} \label{fig:fibres} \end{figure} \setlength{\sub}{0.1377\textwidth} \begin{figure*}% \centering% \includegraphics[width=\sub]{BFigures/Texture1_A_labeling.png}\hspace{\vs}% \includegraphics[width=\sub]{BFigures/Texture1_B_labeling.png}\hspace{\vs}% \includegraphics[width=\sub]{BFigures/Texture1_C2_labeling.png}\hspace{\vs}% \includegraphics[width=\sub]{BFigures/Texture1_D_labeling.png}\hspace{\vs}% \includegraphics[width=\sub]{BFigures/Texture1_D5_labeling.png}\hspace{\vs}% \includegraphics[width=\sub]{BFigures/Texture1_D4_labeling.png}\hspace{\vs}% \includegraphics[width=\sub]{BFigures/Texture1_D6_labeling.png}% \\\vspace{\vs}% \includegraphics[width=\sub]{BFigures/Texture1_A_segmentation.png}\hspace{\vs}% \includegraphics[width=\sub]{BFigures/Texture1_B_segmentation.png}\hspace{\vs}% \includegraphics[width=\sub]{BFigures/Texture1_C2_segmentation.png}\hspace{\vs}% \includegraphics[width=\sub]{BFigures/Texture1_D_segmentation.png}\hspace{\vs}% \includegraphics[width=\sub]{BFigures/Texture1_D5_segmentation.png}\hspace{\vs}% \includegraphics[width=\sub]{BFigures/Texture1_D4_segmentation.png}\hspace{\vs}% \includegraphics[width=\sub]{BFigures/Texture1_D6_segmentation.png}% \caption{An example demonstrating the flexibility of our method. In the \emph{top} row, two manually labelled classes (cyan and magenta), corresponding to different image features. In the \emph{bottom} row, a resulting image segmentation obtained by binarizing a probability image.} \label{fig:bag} \end{figure*} \section{Method} \label{sec:method} Our method combines two sources of information, the structure in the image and the user-provided partial labelling. The structure in the image is captured in the preprocessing step, namely clustering, which we describe in \ref{subsec:preprocessing}. After that, in \ref{subsec:transformation}, we explain how clustering is used for transforming the user-provided partial labelling into pixelwise probabilities of belonging to each of the classes. The central part of our segmentation, the interactive update, covered in \ref{subsec:update}, is obtained by immediately displaying the result of the transformation and allowing the user to repeatedly improve the partial labelling. Postprocessing choices, covered in \ref{subsec:postprocessing}, are concerned with the outputs of the interactive update. The first and most obvious output is a probability image. The probability image can give the image segmentation, but other postprocessing methods may be utilised as well. For the second output, which we call \emph{dictionary probabilities}, the user-provided partial labellings are propagated to the dictionary patches which were obtained in the preprocessing step. This encodes the learned information used for transforming the intensity image to the probability image, and can be used for subsequent automatic processing of similar images. Our method comes in a range of flavours, governed by the features used for clustering and the strategy used for calculating pixelwise probabilities. In this section we only explain the simplest variant, the other possibilities are covered in Sec.~\ref{sec:implementation}. \emph{Notation.} Throughout the paper we consider an image $\image$ defined on an $\imsizeX $-by-$\imsizeY$ image grid with pixel values in either grayscale or RGB colour space. During the interactive part, the user will be placing marks in the image grid, to indicate the pixels which belong to one of the $C$ segmentation classes. We chose to represent this user-provided information with a layered \emph{label image} $L$, where $L(x,y,c) = 1$ if the user indicated that pixel $(x,y)$ belongs to class $c$, and 0 otherwise. \subsection{Clustering image patches} \label{subsec:preprocessing} The aim of preprocessing is to find the structures in the image without considering the user-provided labels. In the framework of semi-supervised learning, a cluster assumption states that, if points are in the same cluster, they are likely to be of the same class -- which does not imply that each class forms a single cluster \cite{chapelle2009semi}. For our purpose, we assume that image features tend to form discrete clusters and that image features in the same cluster are more likely to share a class. However, we do not assume that each class is represented by only one cluster, so we will need many more clusters than classes. Therefore, we create a multitude of clusters to capture the variety of features present in the image. In Sec.~\ref{sec:implementation} we will explain the implementation details and some more advanced ways of accomplishing clustering. In this section we outline the basic approach, which operates on intensity patches. For this case, only two parameters are required: the number of clusters $\dictsize$ and the size of the patches $\patchsize$. The number of clusters should be large, measured in hundreds or thousands, and is roughly reflecting the variability in the image. The size of the patches should reflect the scale of the distinctive image features and could, for example, be 9 pixels. For simplicity, we always assume that the size of the image patches $\patchsize$ is odd and patches are centred around the central pixel. For clustering, we extract patches of size $\patchsize$-by-$\patchsize$ from the image $\image$, treat each patch as a vector containing the pixel intensities and group those vectors into $\dictsize$ clusters, e.g.\ using $k$-means clustering based on Euclidean distance. The resulting collection of cluster centres represents the content of the image. As these basic elements are inferred by grouping features from image, we call the collection of $\dictsize$ cluster centres an \emph{intensity dictionary}, and each of its elements (each cluster centre) is denoted \emph{dictionary patch}. Every image pixel $(x,y)$ in the centre of an $\patchsize $-by-$\patchsize$ image patch is, by means of clustering, uniquely assigned to one cluster. We represent this using an \emph{assignment image} $\assignim$. For boundary pixels we define $A(x,y) = 0$. \subsection{Relation between image and dictionary} \label{subsec:transformation} According to the cluster assumption, image patches assigned to the same dictionary patch are more likely to belong to the same class. Unique for our method is that we use this assumption on a pixel level, and not on a patch level. That is, if two image patches are assigned to the same dictionary patch, their \emph{corresponding} pixels (i.e.\ the pixels at the same position in the patch) are more likely to belong to the same class. In other words, for every dictionary patch there is a certain (unknown) classification of its pixels, which all assigned patches are likely to share. To exploit this assumption, we define a binary relation between corresponding pixels assigned to the same dictionary pixel. For example, a central pixel of an image patch assigned to a certain dictionary patch relates to central pixels of all other patches assigned to the same dictionary patch. Likewise, the pixel directly above the central pixel relates to corresponding pixels in other patches, and the similar relation extends to all positions in a patch. This results in $M^2K$ cliques of pixels, one for every pixel in the intensity dictionary. Due to the overlap between image patches, every non-boundary pixel belongs to $M^2$ different cliques. The central part of our method is concerned with transforming a user-provided partial labelling to pixelwise probabilities. The transformation matrix we use has a very simple decomposition, which makes our method efficient and allows for immediate feedback to the user. The construction of the transformation matrix is therefore fundamental for our method. However, describing how this matrix is constructed provides little intuition about our method, so we start by motivating our approach. As covered previously, the assignment image $A$, obtained in an unsupervised manner, contains information on clusters of structures in the image $I$. At the same time, image $I$ is accompanied by the user-provided partial labelling $L$. To combine the two sources of information, we create a dictionary of labels to accompany our intensity dictionary. For each dictionary patch $k\in\{1,\dots,K\}$ we use $A$ to identify the locations of all image patches assigned to it. At those locations in the image grid we extract all corresponding patches but from the labelling image $L$. For the set of related labelling patches we compute a pixelwise average for every layer. As a result, every $M$-by-$M$ dictionary patch now has a corresponding $M$-by-$M$ labelling representation consisting of $C$ layers. When the image is fully labelled, the label image $L$ sums to one in every pixel, as only one out of $C$ classes has a label of 1. Consequently, the labelling representation of every dictionary patch also sums to one in every pixel. However, due to the pixelwise averaging, the values of this representation are not binary, they instead encode the normalised frequency of a dictionary pixel being labelled as belonging to class $c$ in the current labelling image. For this reason, we think of this labelling representation as of pixelwise probabilities of belonging to class $c$, and we call them \emph{dictionary probabilities}. Dictionary probabilities can now be pasted back into an $\imsizeX $-by-$\imsizeY$ image grid, again using the location information from $A$, and again averaged in every pixel. This results in an $\imsizeX $-by-$\imsizeY$ probability image $P$ consisting of $C$ layers, where $P$ is a diffused version of $L$. In other words, we use the self-similarity information encoded by $A$ to propagate the user-provided markings from $L$ onto the rest of the image. In light of this motivation, now we turn to explaining the construction of the transformation matrices used for efficient computation of dictionary probabilities and image probabilities. Fundamental for this transformation is the relation between the $\imsizeX $-by-$\imsizeY$ image grid and the $\patchsize$-by-$ \patchsize$-by-$\dictsize$ dictionary grid. This relation will be encoded using an $n$-by-$m$ biadjacency matrix $\mathbf{B}$, where $n = XY$ and $m = M^2K$. For this purpose, we need a linear (single) index for the pixels in the image and the pixels in the dictionary grid. The linear index of an image pixel $(x,y)$ is \begin{equation} i = x + (y-1)\imsizeX\,.\label{eq:image_linear_index} \end{equation} As for the dictionary grid, we use $(0,0,k)$ for the central pixel of the $k$-th dictionary element, and coordinates of other pixels in the patch are defined in terms of within-patch displacements $\Delta x$ and $\Delta y$, both from $\{-s,\dots,0,\dots,s\}$ with $s = (M-1)/2$. A dictionary pixel at coordinates $(\Delta x,\Delta y,k)$ has a linear index \begin{equation} j = (\Delta x+s) + (\Delta y+s) \patchsize + (k-1)\patchsize^2\,.\label{eq:dict_linear_index} \end{equation} Each assignment of an image patch centered around $(x,y)$ to a $k$-th dictionary patch centered around $(0,0,k)$ induces a relation between the $\patchsize^2$ image pixels and the $\patchsize^2$ dictionary pixels, see \myfigref{fig:ID_relation_smaller}. Using $\sim$ for denoting a relation between image pixels and dictionary pixels gives \begin{equation} \assignim(x,y)=k \quad \Rightarrow \quad \begin{array}{l} (x+\Delta x,y+\Delta y)\sim(\Delta x,\Delta y,k),\\ \text{for all }\Delta x \text{ and }\Delta y\end{array} \,.\label{eq:relation} \end{equation} Since image patches are overlapping, every non-boundary image pixel relates to $\patchsize^2$ dictionary pixels. Image pixels in a boundary relate to less than $\patchsize^2$ dictionary pixels, and the four corner pixels relate to only one dictionary pixel. In total there are $(\imsizeX-2s)(\imsizeY-2s)\patchsize^2$ relations between the image pixels and the dictionary pixels. We represent the relations between $n$ image pixels and $m$ dictionary pixels using an $n$-by-$m$ biadjacency matrix $\biadjacency$, with elements \begin{equation} b_{ij} = \left\{\begin{array}{l l} 1 & \quad i\sim j \\ 0 & \quad \mathrm{otherwise} \end{array}\right. \quad , \end{equation} where $i$ and $j$ are linear indices of an image pixel and a dictionary pixel. The algorithm for constructing $\biadjacency$ is summarised in Alg.~\ref{alg:biadjacency}. \begin{algorithm} \caption{Construction of $\biadjacency$} \begin{algorithmic}[1] \State Initiate $\biadjacency$ as an $n$-by-$m$ matrix with $b_{ij}=0$ \For{an non-boundary pixel $(x,y)$} \State Retrieve pixel assignment $k = A(x,y)$ \For{within-patch displacement $(\Delta x,\Delta y)$} \State compute $i$ for $(x+\Delta x,y + \Delta y)$ using \myeqref{eq:image_linear_index \State compute $j$ for $(\Delta x,\Delta y,k)$ using \myeqref{eq:dict_linear_index} \State assign $b_{ij} = 1$ \EndFor \EndFor \end{algorithmic} \label{alg:biadjacency} \end{algorithm} The biadjacency matrix $\biadjacency$ defines the linear mapping used to propagate the information from the image to the dictionary and vice versa. Consider a quantity defined on the image grid (e.g. user-provided markings indicating pixels which belong to class 1) arranged into a length $n$ vector $\mathbf{v}$ such that the $i$-th element contains the value of the $i$-th image pixel. Propagating these values to the dictionary is carried out by calculating a length $m$ vector \begin{equation} \mathbf{d} = \mathrm{diag}(\biadjacency^\mathsf{T}\mathbf{1}_{n\times1})^{-1}\biadjacency^\mathsf{T} \mathbf{v}\,, \label{eq:L_to_D} \end{equation} where $\mathbf{1}$ denotes a column vector of ones, while $\mathrm{diag}(\cdot)$ denotes a diagonal matrix with the diagonal defined by the argument. The $j$-th element of $\mathbf{d}$ contains the value of the $j$-th dictionary pixel computed by averaging the values of the related image pixels. The summation is accomplished by multiplying with $\biadjacency^\mathsf{T}$ while the diagonal matrix accomplishes the division with the total number of related pixels. For this reason we define the $m$-by-$n$ transformation matrix for mapping from the image to the dictionary as \begin{equation} \imtodict = \mathrm{diag}(\biadjacency^\mathsf{T}\mathbf{1}_{n\times1})^{-1}\biadjacency^\mathsf{T} \,. \label{eq:T1} \end{equation} Similarly, mapping from the dictionary to the image is given by the $n$-by-$m$ matrix \begin{equation} \dicttoim = \mathrm{diag}(\biadjacency\mathbf{1}_{m\times1})^{-1}\biadjacency\,. \label{eq:T2} \end{equation} \begin{figure} \begin{center} \input{tikz/ID_relation_smaller_new} \end{tikzpicture} \end{center} \caption{\label{fig:relations} A subset of relations between a $9\times 6$ image and a $3\times3\times4$ dictionary caused by the framed patch centered around the pixel shaded darker being assigned to the first dictionary patch. \label{fig:ID_relation_smaller}} \end{figure} Those two transformation matrices are fundamental for our method. The propagation of user-provided markings (as described in the motivational paragraphs) is computed as \begin{equation} \mathbf{P} = \dicttoim \imtodict \mathbf{L} \,, \label{eq:L2P} \end{equation} where $\mathbf{L}$ is the user-provided labelling $L$ arranged in a $n$-by-$C$ matrix, while the resulting $n$-by-$C$ matrix $\mathbf{P}$ needs to be arranged back into a layered image $P$ \subsection{Interactive update} \label{subsec:update} When equipping our method with the user-provided interactive update, we run into choices with regards to: i) the way in which we treat unlabelled pixels, ii) the number of applied diffusion steps, and the way of treating intermediate results between the steps, and iii) the possibility of changing the number of segmentation classes. After testing many types of interactive updates, we kept three main versions. In all versions the number of classes $C$ is chosen during initialisation and kept fixed during the update. The way in which we handle pixels that have not been labelled by the user is also common to all versions. Such pixels are initially assigned equal probability of belonging to each class. As a result, before the user places the first label, all probabilities are equal and no segmentation is possible. The user starts the interaction by choosing a pencil corresponding to one of the $C$ classes applies markings to some pixels. The partial labelling information is immediately transformed to the probability image and shown to the user as an image segmentation, such that every pixel is placed in the class with the highest probability. After the first pencil stroke, only one class will have values larger than $\frac{1}{C}$ in the label image $L$, and the same applies for the probability image $P$ computed using \eqref{eq:L2P}. Thus, at first, many pixels will belong to the first marked class and no pixels will be assigned to the classes that have not used yet. As user adds markings for the other classes, those will appear in probability image $P$. Thanks to the real-time feedback, the user can quickly improve the result by placing markings in misclassified regions (the regions that have been incorrectly classified). With many unlabelled pixels in $L$, the image $P$ will typically have many values that only differ slightly from $\frac{1}{C}$. Those small deviations carry the information needed for inferring the class of the unlabelled pixels. As for the number of applied diffusion steps, we use either one or two. When using two diffusion steps, instead of continuing to diffuse the (already diffused) probability image, we can apply additional non-linear operations in between the two diffusions. Very good results are obtained if we apply \emph{binarisation} of the labels between the two diffusion steps. For binarisation we identify the class of the highest probability for each pixel, and apply $\{0,1\}$ labelling. If there are pixels with no clear probability maximum, we let them retain their unresolved labels. Consequently, for the second iteration of the diffusion, many pixels act as labelled, and this improves the quality of the result. Another additional operation for the two-step diffusion involves the subset of pixels which has been labelled by the user. After the first diffusion step, the $\{0,1\}$ labelling of those pixels has probably dispersed, and some might even have a maximal probability in a class different from the markings indicated by the user. The operation of \emph{overwriting} imposes the original user-provided labelling to all labelled pixels in between the two diffusion steps. The options for the two-step diffusion, binarisation and overwriting are implemented in our segmentation tool, such that the user can quickly switch between the variants of the method and decide which one yields the best results for the data at hand. Likewise, the user can quickly determine whether the quality of the results is sufficient or additional markings should be placed. The user can choose to see the output of the classification displayed as a final segmentation based on the resulting probability image. Alternatively, there is an option for inspecting the $C$ probability images, which often gives a better insight into the quality of the result. In some cases, the final classification may seem incorrect, but the probability images do contain useful information which can be postprocessed for obtaining the desired result. \subsection{Postprocessing} \label{subsec:postprocessing} Our approach allows for various postprocessing options, which may be grouped into two postprocessing strategies. One strategy involves processing the probability image to obtain the segmentation, or detection of interesting features from the probability image. These operations are application-driven and examples are illustrated in Sec.~\ref{sec:results}. The second strategy involves reusing the information stored in the dictionary and the associated dictionary probabilities. The linear transformation \eqref{eq:L2P}, which is core to our method, first transforms the user-provided markings from $L$ to the dictionary space (using matrix $\imtodict$) and then back to the image space (using matrix $\dicttoim$). Consider only the first product \[\mathbf{D} = \imtodict\mathbf{L}\,.\] This is an $m$-by-$C$ matrix containing the pixelwise probabilities of the dictionary pixels (i.e. the dictionary probabilities) which can be useful for processing a previously unseen image similar to $I$. Processing a new image $\hat{I}$ requires extracting all $M$-by-$M$ patches for every pixel of $\hat{I}$ and assigning those patches to the \emph{existing} dictionary, i.e.\ the dictionary created using patches from $I$. Just like before, this assignment defines an image-to-dictionary and we can compute the two associated transformation matrices. Here we are interested in the dictionary-to-image transformation $\hat{\mathbf{T}}_2$. To compute the probability image corresponding to the unlabelled image $\hat{I}$ we therefore need to compute \[\hat{\mathbf{P}} = \hat{\mathbf{T}}_2\mathbf{D}\,.\] and rearrange the result into $\hat{P}$. This way of using our method fits into the framework of supervised learning. The original image $I$ and the computed labelling $L$ can in this context be seen as a (labelled) training set (ignoring the fact that the labelling is computed in a semi-supervised way). Our method is then capable of producing the probability image $\hat{P}$ for the new, unlabelled image $\hat{I}$. The approach will work as long as the initial clustering captures the features present in $\hat{I}$, which holds for similar images. \section{Implementation details} \label{sec:implementation} When developing a framework for interactive propagation of markings we made a number of implementational choices governed by the performance of our method. The part of the method concerning the update includes a few of options mentioned earlier (running the diffusion step twice and discretising between diffusion steps). As for the clustering step for preprocessing the data, our experience leads to two conclusions. First, our method is rather robust to the \emph{quality} of the clustering, so using an approximate clustering will generally not deteriorate the output. Second, the features used for clustering need to reflect the distinction in the appearance of the classes we want to separate. For many types of images, an intensity-based approach as sketched in Sec.~\ref{sec:method} will perform reasonably well. However, in challenging cases, more elaborate image features might provide better results. In this section, we briefly touch upon different possibilities. With the method being robust with respect to the quality of the clustering, we focus on efficiency when building the dictionary. Therefore, we choose to use a $k$-means tree \citep{nister2006scalable}, which is built from consecutive $k$-means clusterings. In this implementation, the size of the dictionary is defined in terms of the branching factor $\treebranch$ and the number of layers $\treelayer$. Since each node in the tree makes up a dictionary element, the total number of dictionary elements is given by $K = \frac{b^{t+1}-1}{b-1}$. Our experience is that good performance is obtained also without running the $k$-means until convergence for each three layer, and therefore a fixed number of iterations is chosen, e.g.~10 iterations. Furthermore, in order to limit the computational burden and memory usage, we extract only a subset of $\patchsize$-by-$\patchsize$ patches from the image when building the dictionary. As for producing $A$ given the clustering represented by a $k$-means tree, the patch vector is compared with the nodes in the first layer to find the match. The patch vector is then compared to the children of this node, and the most similar node is again chosen. This process is repeated until a leaf node or an empty node is reached. The patch vector is assigned to the most similar node along this path. Apart from clustering image patches, we also experimented with different image features. Some of the results we show in Sec.~\ref{sec:results} are based on SIFT \cite{lowe2004distinctive}, but other features can also be incorporated in our method. The approach is as follows. First, image features represented by vectors are extracted from all pixel positions in the image and clustered in $K$ clusters. For speed, it often suffices to consider only a subset of pixels for clustering, as long as we capture the variability in the image. Every position $(x,y)$ from the image grid can now be uniquely assigned to one of the $k$ clusters -- the cluster that is closest to the feature vector extracted at $(x,y)$. This results in an assignment image $A$. The only additional information we need for building the transformation matrices is a value $M$, which earlier represented the size of the extracted image patches. The value $M$ now determines the size of the overlap when linking the image to the dictionary. While we now freely chose $M$, it is reasonable to use a value which corresponds to the size of the extracted features. \section{Results} \label{sec:results} In Fig.~\ref{fig:nerves} we show a three-class classification of a volumetric X-ray image of peripheral nerves appearing as tubular structures. Using a purely intensity-based approach to pixel classification, it would be difficult to differentiate between the bright background and the bright regions inside the dark tubes. Furthermore, a significant bias field makes it difficult to choose a global threshold. Our approach utilises a very limited user input in just one slice of the volume to differentiate between three classes: background, tubes and inside. Moreover, the dictionary probabilities learned from processing one slice can be used for automatic classification of all other slices in the volume, yielding a volumetric segmentation. Fig.~\ref{fig:fibres2} shows an example of segmenting a volumetric image of a fibre composite into two classes: background class and fibre centre class. Using our method, a huge number of individual fibres can be segmented with a modest user input. The probability image of a fibre centre class precisely indicates a region for each fibre centre, and can readily be used in postprocessing for obtaining quantitative information about the spatial distribution of fibres. In this example we also use the result of single-slice segmentation for batch processing of a whole volume stack. In principle, this yields the centre line of each individual fibre. For comparison, we also show a result obtained by thresholding the intensity image. This nicely illustrates a challenge in segmenting densely packed fibres, when the image resolution does not suffice to clearly delineate the boundary of every individual fibre. In such a case, a successful segmentation requires utilisation of the repetitive patterns in the image. Our method accomplishes this via clustering. In Fig.~\ref{fig:onion} we show a three-class segmentation of onion cells. Since cell walls and nuclei both appear dark, a purely intensity-based method would not distinguish these two classes--a task which our method successfully solves with only a modest user input. In Fig.~\ref{fig:histo} we show the use of our method for counting cells in a stained microscopy image. Unlike other examples, this is a colour (RGB) image. To utilise colour information, the features extracted from every image patch contain three colour channels concatenated inyo a single feature vector. Since the final goal is to count and measure the size of cells, we postprocess the probability images to obtain individual cell segmentation. We detect the centre of each cell by computing the local maxima of the centre-class probability image and we estimate the extent of each cell by considering both the centre-class and boundary-class probability images, coupled with the distance from the previously detected cell centres. \setlength{\sub}{0.495\linewidth} \begin{figure}[t!] \centering \includegraphics[width=\sub]{NFigures/settings_labels_displayed.png}\hspace{\vs}% \includegraphics[width=\sub]{NFigures/settings_results_displayed.png}\\\vspace{\vs} \includegraphics[width=\sub]{NFigures/nerve_boundary_p.png}\hspace{\vs}% \includegraphics[width=\sub]{NFigures/nerve_centers_p.png}\\\vspace{\vs}% \includegraphics[width=\sub,trim={5.9cm 2.3cm 5.9cm 2.3cm},clip]{NFigures/nerve_boundary_g.png}\hspace{\vs}% \includegraphics[width=\sub,trim={5.9cm 2.3cm 5.9cm 2.3cm},clip]{NFigures/nerve_centers_g.png}% \caption{\label{fig:nerves} Volumetric segmentation of peripheral nerves. In the \emph{top} row, a slice from the volumetric data with overlayed limited user input and the three-class segmentation dividing the pixels into background (cyan), tubes (purple) and inside (magenta). The \emph{middle} row shows two layers of probability images, corresponding to the tube class and the inside class. High intensity indicates a hight probability of belonging to the class. The \emph{bottom} row shows the 3D visualization of the volumetric data obtained by processing a full image stack and assigning each voxel to the class of the highest probability. This experiment was performed using $M =9$, $K=4000$ and a clustering based on SIFT features. Small connected components of less than $10^4$ voxels were removed during postprocessing.} \end{figure} \begin{figure} \centering \includegraphics[width=\sub]{FFigures/settings_labels_displayed.png}\hspace{\vs}% \includegraphics[width=\sub]{FFigures/settings_results_displayed.png}\\\vspace{\vs}% \includegraphics[width=\sub]{FFigures/P_1.png}\hspace{\vs}% \includegraphics[width=\sub]{FFigures/fibres_segmentation_FR_G.png}\\\vspace{\vs}% \includegraphics[width=0.3266\linewidth,trim={7cm 7cm 7cm 7cm},clip]{FFigures/settings_labels_displayed.png}\hspace{\vs}% \includegraphics[width=0.3266\linewidth,trim={7cm 7cm 7cm 7cm},clip]{FFigures/P_1.png}\hspace{\vs}% \includegraphics[width=0.3266\linewidth,trim={7cm 7cm 7cm 7cm},clip]{FFigures/fibre_thres.png}% \caption{Volumetric segmentation of glass fibres. In the \emph{top} row a slice with manual input indicating fibre centres (magenta) on a background (cyan) class, together with a resulting two-class segmentation. The \emph{middle} row shows a layer of a probability image corresponding to the fibre centres class. On the \emph{middle} to the \emph{right} is the output of processing a full volumetric stack. From the 3D visualisation it is evident that fibres form clusters of different orientations. The \emph{bottom} row shows a zoom-in on the central part of the image slice, together with the corresponding probability image and (for comparison) a corresponding segmentation obtained by directly thresholding the image intensities. Settings used in this experiment are $M=9$ and $K=4000$.\label{fig:fibres2}} \end{figure} \begin{figure} \centering \includegraphics[width=\sub]{OFigures/settings_1_labels_displayed.png}\hspace{\vs}% \includegraphics[width=\sub]{OFigures/settings_1_results_displayed.png}\\\vspace{\vs} \includegraphics[width=\sub]{OFigures/settings_4_results_displayed.png}\hspace{\vs}% \includegraphics[width=\sub]{OFigures/settings_3_results_displayed.png} \caption{A three-class segmentation of an image showing onion cells. In the \emph{top} row an image with manual input and a segmentation into three classes: background (cyan), nucleus (purple) and wall (magenta). In the \emph{bottom} row the probability images for the wall and the nucleus class. Settings used are $M=9$ and $K=4000$.\label{fig:onion}} \end{figure} \begin{figure} \centering \includegraphics[width=\sub]{HFigures/org.png}\hspace{\vs}% \includegraphics[width=\sub]{HFigures/m1.png}\\\vspace{\vs} \includegraphics[width=\sub]{HFigures/m2.png}\hspace{\vs}% \includegraphics[width=\sub]{HFigures/segmentation_colored.png}\\\vspace{\vs} \caption{A three-class segmentation of a histopathology image. In the \emph{top} row an original colour image. The extent of the manual input and a corresponding segmentation into three classes are shown on the \emph{right}, with a frame cropped to show the central column of the image. Likewise, in the \emph{bottom} row we show the probability images for the two classes for the central part of the image. In the \emph{bottom right} the final result obtained through additional postprocessing to distinguish individual cells. Settings used are $M=5$ and $K=4000$.\label{fig:histo}} \end{figure} \section{Conclusion} We propose a method for interactive labeling of image pixels. Instrumental for our method is a pair of closely related transformations which propagate the information from the image grid to a dictionary, and back to the image. The transformations are constructed such that the propagation is strong between image pixels with similar appearance, captured by extracted image features. In this paper we present an algorithm for building an efficient matrix representation of those transformations, allowing a real-time processing. We demonstrate how propagation of user-provided labelling can be used for an interactive image segmentation. Furthermore, a segmentation of one image allows for subsequent automatic processing of similar images. With only a modest user input, our method can yield good results when segmenting patterned images. We find this extremely useful for many tasks in materials and life sciences. \label{sec:conclusion} \begin{small} \bibliographystyle{spphys}

a27efa38229845717d11a14c367c9d2ce1e3360a

\section{Introduction} Spoken language understanding (SLU) in current literature refers to the study of models that parse spoken queries into semantic frames. Semantic frames contain pieces of semantic units that best represent the speaker's intentions and are essential for the development of human-machine interfaces, such as virtual assistants. Scarcity of linguistic resources has been a recurring issue in many NLP tasks such as representation learning \cite{al2013polyglot}, neural machine translation (NMT) \cite{zoph2016transfer}, and SLU \cite{kurata2016labeled}. The issue is especially true for SLU, because creating manually annotated SLU datasets is costly but the domain space that might require new labeled datasets is near infinite. Even for domains with existing datasets, they might suffer from the data sparsity issue, which have long been plaguing many NLP tasks that require annotated linguistic datasets \cite{lai2015recurrent}. For example, most SLU datasets are not large enough cover all possible data label pairs. Furthermore, biased data collection methods could exacerbate the issue \cite{torralba2011unbiased}. Recent years, there have been significant advances in variational autoencoders (VAE) \cite{kingma2013auto} and other latent variable models for textual generation \cite{serban2017hierarchical,yu2017seqgan,hu2017toward,li2017adversarial}, prompting investigations into the possibility of improving model performances through generative data augmentation \cite{kafle2017data,kurata2016labeled,hou2018sequence}. In order to alleviate the data issues, data augmentation (DA) techniques that simply perform class-preserving transformation on data samples have been extensively used extensively \cite{simard2003neural,Krizhevsky2012,fadaee2017data}. However, such DA methods require full supervision and generated datasets lack variety and robustness. To reduce reliance on handcrafted transformation functions, there has been growing interest in leveraging the generative power of latent variable models to facilitate DA. These line of works deserve a category of its own, to which we refer as \textit{generative data augmentation} (GDA). Recent works have explored the idea for the SLU task \cite{kurata2016labeled,hou2018sequence}. In this paper, we formalize the notion of GDA by developing a general framework for the class of DA techniques in the SLU domain. Upon the framework, we propose a generative model specialized in the generation of SLU datasets. Finally, we wish to demonstrate the effectiveness of our approach through various experiments. In essence, our main contributions are three folds: \begin{enumerate} \item \textbf{The Generative DA Framework}: We develop a general framework of generative data augmentation specifically for the SLU task. During formulation, we posit the importance of prior approximation in generation sampling and propose a Monte Carlo-based method. Experiments show that the Monte Carlo-based estimation is superior compared to other approximation methods. \item \textbf{A Novel Model for Labeled Language Generation}: We propose a novel generative model for jointly synthesizing spoken utterances and their semantic annotations (slot labels and intents). We show that the synthetic samples generated from the model are not only natural and accurately annotated, but they improve SLU performances by a significant margin when used in the generative data augmentation framework. We also show that our model is better than the previous work \cite{kurata2016labeled}. \item \textbf{Substantiation with Extensive Experimentation}: We substantiate the general benefits of generative data augmentation with experiments and statistical testing on various SLU models and datasets. Results show that our approach produces extremely competitive performances for existing SLU models in the ATIS dataset. Our ablation studies also bring some important insights such as the optimal synthetic dataset size to light. \end{enumerate} \section{Related Work} \subsubsection{Deep Generative Models and Text Generation} Variational autoencoders (VAE) \cite{kingma2013auto,rezende2014stochastic} are deep latent Gaussian models applied with stochastic variational inference, a method which makes the models scalable to large datasets. Since its inception, many variations of the VAE model have been explored in the language domain. Notably, variational recurrent auto-encoders (VRAE) were first proposed by \cite{fabius2014variational}. Generative adversarial networks (GAN) are another class of latent variable models with implicit latent distribution \cite{goodfellow2014generative}. Advances have been made in applying the GAN model to text generation \cite{yu2017seqgan,fedus2018maskgan}. Recently, much attention has been drawn to the tasks of controllable generation and style transfer, which have been successfully explored in \cite{hu2017toward,shen2017style} using variational models. \subsubsection{Data Augmentation and Regularization} For data-hungry models, appropriate regularization is necessary to achieve high performance for many tasks. Model regularizers such as dropout \cite{srivastava2014dropout} and batch normalization \cite{ioffe2015batch} are widely accepted techniques to prevent model overfitting and promote noise robustness. Transfer learning is another regularization technique to enhance the generalization power of models that has achieved success across numerous domains and tasks \cite{pan2010survey}. Data augmentation (DA) is a separate class of regularization methods that create artificial training data to obtain better resulting models. Most DA techniques proposed in the literature can be categorized into either \textit{transformative} or \textit{generative} methods. Transformative data augmentation relies on unparameterized data-transforming functions embued with external knowledge to synthesize new class-preserving data points \cite{dao2018kernel}. Transformative DA is widely used in the vision domain. For example, images are randomly perturbed with linear transformations (rotation, shifting etc.) to boost performances in many vision-related tasks \cite{simard2003neural,Krizhevsky2012}. On the other hand, Generative DA (GDA) exploits the generative power of latent variable models to artificially create convincing data samples. With advances in powerful generative models such as VAEs and GANs, the potential to leverage them for data augmentation has gained much attention recently. Particularly, performance gains from generated datasets have been studied and documented in the VQA task \cite{kafle2017data}, general image classification \cite{ratner2017learning}, and selected SLU tasks \cite{kurata2016labeled,hou2018sequence}. However, relevant researches are hurdled by the architectural and experimental complexities. Nevertheless, our work is the first to explore idea of using variational generative models for DA. \subsubsection{Spoken Language Understanding} The SLU task is one of more mature research areas in NLP. Many works have focused on exploring neural architectures for the SLU task. Plain RNNs and LSTMs were first explored in \cite{Mesnil2015,yao2014spoken}. \cite{kurata2016leveraging} proposed sequence-to-sequence (Seq2Seq) models. Hybrid models between RNNs and CRFs were explored in \cite{huang2015bidirectional}. Joint language understanding models that jointly predict slot labels and intents gained significant traction since they had been first proposed in \cite{guo2014joint,goo2018slot}. Some works focused on translating advances in other NLP areas to the SLU task \cite{liu2016attention}. \tikzset{ prob/.style = {draw, thick, rounded corners, rectangle, minimum height=0.8cm, minimum width=1.2cm, node distance=1.2cm, font=\small}, dataset/.style = {draw, circle, node distance=1.2cm, minimum size=0.8cm, font=\small}, input/.style = {coordinate}, output/.style = {coordinate} } \begin{figure}[t] \centering \begin{tikzpicture}[auto, thick, node distance=2cm, >=triangle 45] \draw node [dataset] (original-words) {$\mathbf{w}$}; \draw node [dataset, below of=original-words] (original-labels) {$\mathbf{s}$}; \draw node [dataset, below of=original-labels] (original-intents) {$y$}; \draw [rounded corners, solid, name=original-dataset] ($(original-words.north west) + (-0.25cm, 0.25cm)$) rectangle ($(original-intents.south east) + (0.25cm, -0.25cm)$); \path ($(original-words.north west) + (0, 0.2cm)$) -- ($(original-words.north east) + (0, 0.2cm)$) node [midway, above] {$\mathcal{D}$}; \draw node [prob, left=0.7cm of original-words, minimum width=1.2cm] (true-prob) {$p \left( \mathbf{x} \right)$}; \draw node [prob, left=0.7cm of original-intents, minimum width=1.2cm] (false-prob) {$p^\star \left( \mathbf{x} \right)$}; \draw [-latex, dashed, above, font=\footnotesize, shorten >= 0.11cm] (false-prob) -- node {} (original-intents); \draw [-latex, dashdotted, left, font=\footnotesize] (true-prob) -- node {$\omega_b$} (false-prob); \draw node [prob, right=0.7cm of original-intents] (data-model) {$\hat{p} \left( \mathbf{x} \right)$}; \draw node [prob, right=0.7cm of original-words] (better-model) {$\hat{p}^\star \left( \mathbf{x} \right)$}; \draw [-latex, shorten <= 0.11cm] (original-intents) -- node {} (data-model); \draw [-latex, dashdotted, left, font=\footnotesize] (data-model) -- node {$\omega_d$} (better-model); \draw node [dataset, right=0.7cm of better-model] (new-words) {$\hat{\mathbf{w}}$}; \draw node [dataset, below of=new-words] (new-labels) {$\hat{\mathbf{s}}$}; \draw node [dataset, right=0.7cm of data-model] (new-intents) {$\hat{y}$}; \draw [rounded corners, name=new-dataset] ($(new-words.north west) + (-0.25cm, 0.25cm)$) rectangle ($(new-intents.south east) + (0.25cm, -0.25cm)$); \path ($(new-words.north west) + (0, 0.2cm)$) -- ($(new-words.north east) + (0, 0.2cm)$) node [midway, above] {$\hat{\mathcal{D}}$}; \draw [rounded corners, loosely dotted, name=jluva] ($(better-model.north west) + (-0.25cm, 0.6cm)$) rectangle ($(new-intents.south east) + (0.4cm, -0.4cm)$); \draw [-latex, dashed, left, font=\footnotesize, shorten >= 0.11cm] (better-model) -- node {} (new-words); \draw node [dataset, right=0.7cm of new-words] (aug-words) {$\mathbf{w}'$}; \draw node [dataset, right=0.7cm of new-labels] (aug-labels) {$\mathbf{s}'$}; \draw node [dataset, right=0.7cm of new-intents] (aug-intents) {$y'$}; \draw [rounded corners, name=aug-dataset] ($(aug-words.north west) + (-0.25cm, 0.25cm)$) rectangle ($(aug-intents.south east) + (0.25cm, -0.25cm)$); \path ($(aug-words.north west) + (0, 0.2cm)$) -- ($(aug-words.north east) + (0, 0.2cm)$) node [midway, above] {$\mathcal{D}'$}; \draw [-latex, dotted, shorten >= 0.3cm, shorten <= 0.3cm] (original-intents.south east) to [out=335, in=205] (aug-intents.south west); \draw [-latex, dotted, shorten >= 0.14cm, shorten <= 0.14cm] (new-labels.east) to [] (aug-labels.west); \end{tikzpicture} \par \caption{The general framework for generative language understanding data augmentation. Solid arrows (\sampleline{}) denote training, dashed arrows (\sampleline{dashed}) denote generation, dot-dashed arrows (\sampleline{dashdotted}) denote distortion, and dotted arrows (\sampleline{dotted}) denote data duplication. $\mathcal{D}'$ is the final augmented dataset for training SLU models. The goal of GDA (enclosed in loosely dotted lines) is to recover the true data distribution $p$ through sampling, as if the samples are drawn from the corrected model distribution.} \label{fig:da} \end{figure} \section{Model Description} In this section, we describe our generative data augmentation model and the underlying framework in detail. \subsection{Framework Formulation} We begin with some notations, then we formulate the overall generative data augmentation framework for the spoken language understanding task. \subsubsection{Notations} An utterance $\mathbf{w}$ is a sequence of words $\left( w_1, \ldots, w_{T_i} \right)$, where $T$ is the length of the utterance. For each utterance in a labeled dataset, an equally-long semantic slot sequence $\mathbf{s} = \left( s_1, \ldots, s_T \right)$ exists such that $s_i$ annotates the corresponding word $w_i$. The intent class of the utterance is denoted by $y$. A fully labeled language understanding dataset $\mathcal{D}$ is a collection of utterances and their respective annotations $\left\{ \left( \mathbf{w}_1, \mathbf{s}_1, y_1 \right), \ldots, \left( \mathbf{w}_n, \mathbf{s}_n, y_n \right)\right\}$, where $n$ is the size of the dataset. A data sample in $\mathcal{D}$ is denoted by $\mathbf{x} = \left( \mathbf{w}, \mathbf{s}, y \right)$. The set of all utterances present in $\mathcal{D}$ is denoted by $\mathcal{D}_w = \left\{\mathbf{w}_1, \ldots, \mathbf{w}_n\right\}$. Similarly, the set of slot label sequences and intent classes are denoted by $\mathcal{D}_s$ and $\mathcal{D}_y$. \subsubsection{Spoken Language Understanding} A spoken language understanding model is a discriminative model $\mathbf{S}$ fitted on labeled SLU datasets. Specifically, let $\psi$ to denote parameters of the prediction model. Given a training sample $\left(\mathbf{w}, \mathbf{s}, y\right)$, the training objective is as follows: \begin{equation} \mathcal{L}_{LU} \left( \psi ; \mathbf{w}, \mathbf{s}, y \right) = -\log p_{\psi} \left( \mathbf{s}, y \middle| \mathbf{w} \right). \label{eq:lu-nll} \end{equation} Given an utterance $\mathbf{w}$, predictions are made by finding the slot label sequence $\hat{\mathbf{s}}$ and the intent class $\hat{y}$ that maximize the loglikelihood: $\left( \hat{\mathbf{s}}, \hat{y} \right) = \argmax_{\mathbf{s}, y} {\log p_{\psi} \left( \mathbf{s}, y \middle| \mathbf{w} \right)}$. For non-joint SLU models, $p_\phi$ is factorizable: $p_\phi \left(\mathbf{s}, y \middle| \mathbf{w} \right) = p_\phi \left(\mathbf{s} \middle| \mathbf{w} \right) p_\phi \left( y \middle| \mathbf{w} \right)$. In recent years, joint language understanding has become a popular approach, as studies show a synergetic effect of jointly training slot filling and intent identification \cite{guo2014joint,chen2016syntax}. \subsubsection{Generative Data Augmentation} A general framework of generative data augmentation (GDA) is depicted in Figure \ref{fig:da}. Suppose that IID samples $\mathbf{x} \in \mathcal{D}$ were intended to be sampled from a true but unknown language distribution $p \left( \mathbf{x} \right) \in \mathcal{P}$, where $\mathcal{P}$ is the probability function space for $\mathbf{x}$. However, in real world cases, the actual distribution represented by the $\mathcal{D}_w$ could be distorted due to biases introduced during erroneous data collection process or due to under-sampling variance \cite{torralba2011unbiased}. Let such distortion be a function $\omega_b \in \Omega: \mathcal{P} \rightarrow \mathcal{P}$. The distorted data distribution $p^\star = \omega_b \left( p \right)$ diverges from the true distribution $p$, i.e. $d \left( p^\star, p \right) > 0$ where $d$ is some statistical distance measure such as KL-divergence. An ideal GDA counteracts the bias-introducing function $\omega_b$ and unearths the true distribution $p$ through unsupervised explorative sampling. Suppose that a joint language understanding model $\hat{p} \left( \mathbf{x} \right)$ is trained on $\mathbf{x} \sim p^\star \left( \mathbf{x} \right)$. Without the loss of generality, suppose that the model is expressive enough to perfectly capture the underlying distribution, i.e. $\hat{p} = p^\star$. We collect $m$ samples $\hat{\mathcal{D}} = \left\{\hat{\mathbf{x}}_1,\ldots,\hat{\mathbf{x}}_m\right\}$ drawn from $\hat{p} \left( \mathbf{x} \right)$ and combine them with the original dataset $\mathcal{D}$ to form an augmented dataset $\mathcal{D}'$ of size $n + m$. Na\"ive DA will not yield better SLU results as synthetic data samples $\hat{\mathbf{x}}$ follow the distorted data distribution $p^\star$ in the best case. However, an ideal explorative sampling method could distort the sampling distribution, as if $\hat{\mathbf{x}}$ were sampled from another distribution $\hat{p}^\star$, such that the new distribution is closer to the true distribution (i.e. $d \left( \hat{p}^\star, p \right) < d \left(\hat{p}, p\right)$). There exists a distortion function $\omega_d$ such that $\hat{p}^\star = \omega_d \left( \hat{p} \right)$. The ideal sampling method can be seen as a corrective function $\omega_d$ that undos the effect of $\omega_b$. In this paper, we propose and investigate different sampling methods $\omega_d$ for the maximal DA effect. These methods are described in model description sections. The implementation details are covered in the experiments sections \begin{algorithm}[!ht] \SetKwData{Left}{left}\SetKwData{This}{this}\SetKwData{Up}{up} \SetKwFunction{Union}{Union}\SetKwFunction{FindCompress}{FindCompress} \SetKwInOut{Input}{input}\SetKwInOut{Output}{output} \SetKwInOut{Given}{given} \Input{a sufficiently large number $m$} \Given{$\mathcal{D}_w$, $\theta$, $\phi$} \Output{synthetic utterance list $\mathbf{U}$} initialize $\mathbf{U}$ as an empty list\; \While{$\mathbf{U}$ has less than $m$ samples}{ sample a real utterance $\mathbf{w}$ from $\mathcal{D}_w$\; estimate the mean $\bar{\mathbf{z}}$ of the posterior $q_\phi \left(\mathbf{z} \middle| \mathbf{w}\right)$\; sample $\hat{\mathbf{w}}$ from the likelihood $p_\theta \left( \mathbf{w} \middle| \mathbf{\bar{z}} \right)$\; append $\hat{\mathbf{w}}$ to $\mathbf{U}$\; } \Return{$\mathbf{U}$} \caption{Monte Carlo posterior sampling.} \label{alg:posterior-sampling} \end{algorithm} \subsection{Joint Generative Model} In this subsection, we describe our generative model in detail. We begin with a standard VAE \cite{kingma2013auto} applied to utterances, then we extend the model by allowing it to generate other labels in a joint fashion. \subsubsection{Standard VAE} VAEs are latent variable models applied with amortized variational inference. Let $\theta$ be the parameters of the generator network (i.e. the decoder network), and let $\phi$ be the parameters of the recognition network (i.e. the encoder network). Specifically in the case of utterance learning, the goal is to maximize the log likelihood of sample utterances $\mathbf{w}$ in the dataset $\log p \left( \mathbf{w} \right) = \log \int p \left(\mathbf{w},\mathbf{z}\right) d\mathbf{z}$. However, since the marginalization is computationally intractable, we introduce a proxy network $q_\phi \left( \mathbf{z} \middle| \mathbf{w} \right)$ and subsequently minimize a training objective based on evidence lower bound (ELBO): \begin{align} \begin{split} \mathcal{L}_{VAE} \left( \theta, \phi; \mathbf{w} \right) = &\infdiv{q_\phi \left( \mathbf{z} \middle| \mathbf{w} \right)}{p \left( \mathbf{z} \right)}\\ &-\mathbb{E}_{\mathbf{z} \sim q_\phi \left( \mathbf{z} \middle| \mathbf{w} \right)} \left[\log p_\theta \left(\mathbf{w} \middle| \mathbf{z}\right)\right] \end{split} \label{eq:elbo} \end{align} In Equation \ref{eq:elbo}, the proxy distribution $q_\phi$ is kept close to the prior $p\left(\mathbf{z}\right)$, which we assume to be the standard multivariate Gaussian. Since the KL-divergence term is always positive, $\mathcal{L}_{VAE}$ is the upper bound for the reconstruction error under the particular choice of a proxy distribution $q_\phi$. The proposed generative model is based on VRAEs, in which the posterior of a sequence factorizes over sequence elements (i.e. words) based on the Markov Chain assumption: $p_\theta \left( \mathbf{w} \middle| \mathbf{z} \right) = \prod_{i=1}^T p_\theta \left( w_i | w_1,\ldots,w_{i-1},\mathbf{z} \right)$. VAEs can be optimized using gradient-descent methods with the reparameterization trick \cite{kingma2013auto}. \tikzset{ vector/.style = {draw, rectangle, minimum height=0.9cm, minimum width=0.15cm, node distance=0.6cm, line width=0.5pt}, point/.style = {draw, circle, inner sep=0pt, minimum size=0.1cm, line width=0.3pt}, label/.style = {font=\footnotesize} } \begin{figure}[t] \centering \begin{tikzpicture}[auto] \draw node [vector] (h1enc) {}; \draw node [vector, right of=h1enc] (h2enc) {}; \draw node [vector, right of=h2enc] (h3enc) {}; \draw node [vector, right of=h3enc] (h4enc) {}; \draw node [vector, right of=h4enc] (h5enc) {}; \draw node [point, below=0.35cm of h1enc] (w1in) {}; \draw node [point, below=0.35cm of h2enc] (w2in) {}; \draw node [point, below=0.35cm of h3enc] (w3in) {}; \draw node [point, below=0.35cm of h4enc] (w4in) {}; \draw node [point, below=0.35cm of h5enc] (w5in) {}; \draw node [label, below=0.05cm of w1in] {$w_1$}; \draw node [label, below=0.05cm of w2in] {$w_2$}; \draw node [label, below=0.05cm of w3in] {$w_3$}; \draw node [label, below=0.05cm of w4in] {$w_4$}; \draw node [label, below=0.05cm of w5in] {$w_5$}; \draw [-stealth] (w1in) -- node {} (h1enc); \draw [-stealth] (w2in) -- node {} (h2enc); \draw [-stealth] (w3in) -- node {} (h3enc); \draw [-stealth] (w4in) -- node {} (h4enc); \draw [-stealth] (w5in) -- node {} (h5enc); \draw [-stealth] (h1enc.55) -- node {} (h2enc.125); \draw [-stealth] (h2enc.55) -- node {} (h3enc.125); \draw [-stealth] (h3enc.55) -- node {} (h4enc.125); \draw [-stealth] (h4enc.55) -- node {} (h5enc.125); \draw [-stealth] (h5enc.235) -- node {} (h4enc.305); \draw [-stealth] (h4enc.235) -- node {} (h3enc.305); \draw [-stealth] (h3enc.235) -- node {} (h2enc.305); \draw [-stealth] (h2enc.235) -- node {} (h1enc.305); \draw [rounded corners, dashed, name=encoder] ($(h1enc.north west) + (-0.20cm, 0.20cm)$) rectangle ($(h5enc.south east) + (0.20cm, -0.20cm)$); \path ($(h1enc.north west) + (-0.15cm, 0.2cm)$) -- ($(h5enc.north east) + (0.1cm, 0.2cm)$) node [near start, above, font=\footnotesize] {Encoder}; \draw node [vector, above right=-0.4cm and 0.5cm of h5enc] (mu) {}; \draw node [vector, below right=-0.4cm and 0.5cm of h5enc] (std) {}; \draw node [label, above left=-0.5cm and -0.05cm of mu] () {$\mu$}; \draw node [label, below left=-0.5cm and -0.05cm of std] () {$\sigma$}; \draw [-stealth, shorten <= 0.26cm] (h5enc.70) -- node {} (mu); \draw [-stealth, shorten <= 0.26cm] (h5enc.290) -- node {} (std); \draw node [vector, right=1.2cm of h5enc] (z) {}; \draw node [label, left=-0.06cm of z] {$z$}; \draw [-stealth, dotted] (mu) -- node {} (z.110); \draw [-stealth, dotted] (std) -- node {} (z.250); \draw node [vector, right of=z] (h1wdec) {}; \draw node [vector, right of=h1wdec] (h2wdec) {}; \draw node [vector, right of=h2wdec] (h3wdec) {}; \draw node [vector, right of=h3wdec] (h4wdec) {}; \draw node [vector, right of=h4wdec] (h5wdec) {}; \draw [-stealth] (z) -- node {} (h1wdec); \draw [-stealth] (h1wdec) -- node {} (h2wdec); \draw [-stealth] (h2wdec) -- node {} (h3wdec); \draw [-stealth] (h3wdec) -- node {} (h4wdec); \draw [-stealth] (h4wdec) -- node {} (h5wdec); \draw node [point, above=0.35cm of h1wdec] (w1out) {}; \draw node [point, above=0.35cm of h2wdec] (w2out) {}; \draw node [point, above=0.35cm of h3wdec] (w3out) {}; \draw node [point, above=0.35cm of h4wdec] (w4out) {}; \draw node [point, above=0.35cm of h5wdec] (w5out) {}; \draw node [label, right=-0.1cm of w1out] {$\hat{w}_1$}; \draw node [label, right=-0.1cm of w2out] {$\hat{w}_2$}; \draw node [label, right=-0.1cm of w3out] {$\hat{w}_3$}; \draw node [label, right=-0.1cm of w4out] {$\hat{w}_4$}; \draw node [label, right=-0.1cm of w5out] {$\hat{w}_5$}; \draw [-stealth] (h1wdec) -- node {} (w1out); \draw [-stealth] (h2wdec) -- node {} (w2out); \draw [-stealth] (h3wdec) -- node {} (w3out); \draw [-stealth] (h4wdec) -- node {} (w4out); \draw [-stealth] (h5wdec) -- node {} (w5out); \draw [-stealth] (h1wdec.north) |- ++(0, 0.1cm) -| ++(0.3cm, 0) |- ++(0, -1.25cm) -| node {} (h2wdec.south); \draw [-stealth] (h2wdec.north) |- ++(0, 0.1cm) -| ++(0.3cm, 0) |- ++(0, -1.25cm) -| node {} (h3wdec.south); \draw [-stealth] (h3wdec.north) |- ++(0, 0.1cm) -| ++(0.3cm, 0) |- ++(0, -1.25cm) -| node {} (h4wdec.south); \draw [-stealth] (h4wdec.north) |- ++(0, 0.1cm) -| ++(0.3cm, 0) |- ++(0, -1.25cm) -| node {} (h5wdec.south); \draw node [vector, above=0.8cm of h1wdec] (h1sdec) {}; \draw node [vector, above=0.8cm of h2wdec] (h2sdec) {}; \draw node [vector, above=0.8cm of h3wdec] (h3sdec) {}; \draw node [vector, above=0.8cm of h4wdec] (h4sdec) {}; \draw node [vector, above=0.8cm of h5wdec] (h5sdec) {}; \draw [-stealth] (w1out) -- node {} (h1sdec.south); \draw [-stealth] (w2out) -- node {} (h2sdec.south); \draw [-stealth] (w3out) -- node {} (h3sdec.south); \draw [-stealth] (w4out) -- node {} (h4sdec.south); \draw [-stealth] (w5out) -- node {} (h5sdec.south); \draw [-stealth] (h1sdec) -- node {} (h2sdec); \draw [-stealth] (h2sdec) -- node {} (h3sdec); \draw [-stealth] (h3sdec) -- node {} (h4sdec); \draw [-stealth] (h4sdec) -- node {} (h5sdec); \draw [-stealth] (z) |- node {} (h1sdec); \draw node [point, above=0.35cm of h1sdec] (s1out) {}; \draw node [point, above=0.35cm of h2sdec] (s2out) {}; \draw node [point, above=0.35cm of h3sdec] (s3out) {}; \draw node [point, above=0.35cm of h4sdec] (s4out) {}; \draw node [point, above=0.35cm of h5sdec] (s5out) {}; \draw node [label, above=0cm of s1out] {$\hat{s}_1$}; \draw node [label, above=0cm of s2out] {$\hat{s}_2$}; \draw node [label, above=0cm of s3out] {$\hat{s}_3$}; \draw node [label, above=0cm of s4out] {$\hat{s}_4$}; \draw node [label, above=0cm of s5out] {$\hat{s}_5$}; \draw [-stealth] (h1sdec) -- node {} (s1out); \draw [-stealth] (h2sdec) -- node {} (s2out); \draw [-stealth] (h3sdec) -- node {} (s3out); \draw [-stealth] (h4sdec) -- node {} (s4out); \draw [-stealth] (h5sdec) -- node {} (s5out); \draw node [vector, below=0.6cm of h1wdec] (h1idec) {}; \draw node [vector, below=0.6cm of h2wdec] (h2idec) {}; \draw node [vector, below=0.6cm of h3wdec] (h3idec) {}; \draw node [vector, below=0.6cm of h4wdec] (h4idec) {}; \draw node [vector, below=0.6cm of h5wdec] (h5idec) {}; \draw [-stealth] (h1idec) -- node {} (h2idec); \draw [-stealth] (h2idec) -- node {} (h3idec); \draw [-stealth] (h3idec) -- node {} (h4idec); \draw [-stealth] (h4idec) -- node {} (h5idec); \draw [-stealth] (z.south) |- node {} (h1idec); \draw node [point, right=0.3cm of h5idec] (iout) {}; \draw node [label, above=0.0cm of iout] {$\hat{y}$}; \draw [-stealth] (h5idec) -- node {} (iout); \draw [-stealth] (h1wdec.north) |- ++(0, 0.1cm) -| ++(0.3cm, 0) |- ++(0, -1.35cm) -| node {} (h1idec.north); \draw [-stealth] (h2wdec.north) |- ++(0, 0.1cm) -| ++(0.3cm, 0) |- ++(0, -1.35cm) -| node {} (h2idec.north); \draw [-stealth] (h3wdec.north) |- ++(0, 0.1cm) -| ++(0.3cm, 0) |- ++(0, -1.35cm) -| node {} (h3idec.north); \draw [-stealth] (h4wdec.north) |- ++(0, 0.1cm) -| ++(0.3cm, 0) |- ++(0, -1.35cm) -| node {} (h4idec.north); \draw [-stealth] (h5wdec.north) |- ++(0, 0.1cm) -| ++(0.3cm, 0) |- ++(0, -1.35cm) -| node {} (h5idec.north); \draw [rounded corners, dashed, name=encoder] ($(h1sdec.north west) + (-0.22cm, 0.15cm)$) rectangle ($(h5idec.south east) + (0.55cm, -0.2cm)$); \path ($(h5sdec.north east) + (0.1cm, 0.3cm)$) -- ($(h5idec.north east) + (0.2cm, 0cm)$) node [near start, above, font=\footnotesize, rotate=270] {Decoders}; \end{tikzpicture} \caption{Joint language understanding variational autoencoder (JLUVA). The VAE model consists of a BiLSTM-Max encoder and three uni-directional decoders. Note that the fully connected layers and embedding layers are omitted for clarity.} \label{fig:jluva} \end{figure} 6 \subsubsection{The Sampling Problem} Given the parameters $\theta_\mathcal{D}$ and $\phi_\mathcal{D}$ that are optimized for all $\mathbf{w} \in \mathcal{D}_w$, our goal is to sample plausible utterances $\hat{\mathbf{w}}$ from the distribution of $\mathbf{w}$ believed by the model: \begin{equation} \hat{\mathbf{w}} \sim p_{\theta_\mathcal{D},\phi_\mathcal{D}}\left(\mathbf{w}\right) =\int p_{\theta_\mathcal{D}}\left(\mathbf{w} \middle| \mathbf{z}\right) p_{\theta_\mathcal{D},\phi_\mathcal{D}}\left(\mathbf{z}\right)d\mathbf{z} \label{eq:general-sampling} \end{equation} As evident in Equation \ref{eq:general-sampling}, the marginal likelihood estimation requires us to infer the marginal probability of the latent variable $p_{\theta_\mathcal{D},\phi_\mathcal{D}}\left(\mathbf{z}\right)$, which can be estimated by marginalizing the joint probability from the recognition network. \begin{equation} p_{\theta_\mathcal{D},\phi_\mathcal{D}}\left(\mathbf{z}\right) = \mathbb{E}_{\mathbf{w} \sim p \left( \mathbf{w} \right)} \left[q_{\phi_\mathcal{D}}\left(\mathbf{z}\middle|\mathbf{w}\right)\right] \label{eq:marginalize-encoder} \end{equation} However, Equation \ref{eq:marginalize-encoder} cannot be solved analytically, as the true distribution of $w$ is unknown. Hence, some form of approximation is required to sample utterances from the latent variable model. The approximation approach will likely have an impact on the quality of generated utterances, thereby determine the effect of data augmentation. Here, we describe two options. The first is to approximate the marginal probability of the latent variable with the prior $p \left( \mathbf{z} \right)$, the standard multivariate Gaussian. However, this na\"ive approximation will likely yield homogeneous and uninteresting utterances due to over-simplication of the latent variable space. In real world scenarios, the KLD loss term in Equation \ref{eq:elbo} is still large after convergence. Alternatively, the other option is to approximate using the Monte Carlo method. Under the Monte Carlo approach (Algorithm \ref{alg:posterior-sampling}), the marginal likelihood is calculated deterministically for each utterance $w$ sampled from the dataset $\mathcal{D}$. According to the law of large numbers, the marginal likelihood $p_{\theta_\mathcal{D},\phi_\mathcal{D}} \left( \mathbf{w} \right)$ converges to the empirical mean, thereby providing an unbiased distribution for sampling $\mathbf{w}$. \subsubsection{Exploratory Sampling} In our general framework for GDA, remind that the sampling method is required to be exploratory, such that the biases in datasets are counteracted. translating to better performances in resulting models. Hence, an ideal exploratory sampling approach is unbiased but has increased sampling variance. Intuitively, we can sample the latent variable $\mathbf{z}$ from the Gaussian encoded by the recognizer in place of analytically estimating the mean in Algorithm \ref{alg:posterior-sampling}. Suppose that $\boldsymbol{\mu}$ and $\boldsymbol{\sigma}$ are mean and standard deviation vectors encoded by the recognizer. Then we sample $\mathbf{z}$ from $\mathcal{N} \left(\boldsymbol{\mu}\left(\mathbf{w}\right), \lambda_s \cdot \boldsymbol{\sigma}\left(\mathbf{w}\right)\right)$, where the scaling hyperparameter $\lambda_s$ controls the level of exploration exhibited by the generator. This unbiased empirical estimation of the posterior helps generate realistic but more varied utterances. \subsubsection{Joint Language Understanding VAE} Starting from a VAE for encoding and decoding utterances, Joint Language Understanding VAE (JLUVA) extends the model by predicting slot labels and intent classes. The generation of slot labels and intents are conditioned on the latent variable $\mathbf{z}$ and the generated utterance $\hat{\mathbf{w}}$ (Figure \ref{fig:jluva}). The benefits of having conditional dependence on $\mathbf{z}$ during labeling is documented in \cite{kurata2016leveraging}. The modified training objective for the language understanding task is as follows. \begin{table} \centering \begin{tabular}{lcccccc} \hline Dataset & \#Splits & Train & Val & Test \\ \hline\hline ATIS-small & 35 & 127 - 128 & 500 & 893 \\ ATIS-medium & 9 & 497 - 498 & 500 & 893 \\ ATIS & 1 & 4,478 & 500 & 893 \\ Snips & 1 & 13,084 & 700 & 700 \\ MIT Movie Eng & 1 & 8,798 & 977 & 2,443 \\ MIT Movie Trivia & 1 & 7,035 & 781 & 1,953 \\ MIT Restaurant & 1 & 6,894 & 766 & 1,521 \\ \hline \end{tabular} \caption{Dataset statistics. Training sets of ATIS (Small) and ATIS (Medium) have been chunked from the training set of ATIS (Full). } \label{tab:datastats} \end{table} \begin{equation} \mathcal{L}_{LU} \left( \phi, \psi ; \mathbf{w}, \mathbf{s}, y \right) = -\mathbb{E}_{\mathbf{z} \sim q_\phi} \left[ \log p_{\psi} \left( \mathbf{s}, y \middle| \hat{\mathbf{w}}, \mathbf{z} \right) \right] \label{eq:lu-new-nll} \end{equation} The joint training objective of the entire model is specified in terms of the training objective of the VAE component (Equation \ref{eq:elbo}) and the negative log-likelihood of the discriminatory component (Equation \ref{eq:lu-new-nll}): \begin{table*} \centering \begin{threeparttable} \begin{tabular}{l|ccc|ccc|ccc} \hline & \multicolumn{3}{c}{Slot Filling (F1)} & \multicolumn{3}{c}{Intent (F1)} & \multicolumn{3}{c}{Semantic (Acc.)} \\ \cline{2-10} Model + Sampling Approach & Small & Med. & Full & Small & Med. & Full & Small & Med. & Full \\ \hline \hline \begin{tabular}{@{}@{} l}Baseline (No Augmentation) \end{tabular} & 72.57\tnote{\ddag} & 88.28\tnote{\ddag} & 95.34 & 82.65 & 90.59\tnote{\dag} & 97.21 & 35.09\tnote{\ddag} & 65.18\tnote{\ddag} & 85.95 \\ \begin{tabular}{@{}@{} l}Encoder-Decoder + Additive\tnote{*} \end{tabular} & 74.79\tnote{\dag} & 89.13\tnote{\ddag} & 95.20& - & - & - & - & - & - \\ \begin{tabular}{@{}@{} l}JLUVA + Additive (Ours) \end{tabular} & 74.14\tnote{\ddag} & 89.13\tnote{\ddag} & 95.40 & 83.46 & \textbf{90.97} & 97.04 & 38.58 & 66.75 & 85.81 \\ \begin{tabular}{@{}@{} l}JLUVA + Standard Gaussian (Ours) \end{tabular} & 70.72\tnote{\ddag} & 86.90\tnote{\ddag} & 94.91\tnote{\ddag} & 78.67\tnote{\ddag} & 86.90\tnote{\ddag} & 96.90 & 32.46\tnote{\ddag} & 61.12\tnote{\ddag} & 84.62\tnote{\ddag} \\ \begin{tabular}{@{}@{} l}JLUVA + Posterior (Ours) \end{tabular} & \textbf{74.92} & \textbf{89.27} & \textbf{95.51} & \textbf{83.65} & 90.95 & \textbf{97.24} & \textbf{39.43} & \textbf{67.05} & \textbf{86.26} \\ \hline \end{tabular} \begin{tablenotes} \item[*]{\footnotesize \cite{kurata2016labeled}} \item[\dag]{\footnotesize $p<0.1$} \item[\ddag]{\footnotesize $p<0.01$} \end{tablenotes} \end{threeparttable} \caption{ Data scarcity results for the ATIS dataset. We use the baseline BiLSTM model as the control SLU model. Results are averaged over multiple runs and compared to the best of our approaches (JLUVA + Posterior). The differences are tested for statistical significance. } \label{tab:data-scarce} \end{table*} \begin{align} \begin{split} \mathcal{L} \left( \theta, \phi, \psi; \mathbf{w}, \mathbf{s}, y \right) = &\infdiv{q_\phi \left( \mathbf{z} \middle| \mathbf{w} \right)}{p_\theta \left( \mathbf{z} \middle| \mathbf{w} \right)}\\ &-\mathbb{E}_{\mathbf{z} \sim q_\phi} \left[\log p_\theta \left(\mathbf{w} \middle| \mathbf{z}\right)\right] \\ &-\mathbb{E}_{\mathbf{z} \sim q_\phi} \left[ \log p_{\psi} \left( \mathbf{s}, y \middle| \hat{\mathbf{w}}, \mathbf{z} \right) \right] \end{split} \label{eq:jluva-loss} \end{align} We obtain the optimal parameters $\theta^*, \phi^*, \psi^*$ by minimizing Equation \ref{eq:jluva-loss} (i.e. $\argmin_{\theta, \phi, \psi}\mathcal{L}$) with respect to a real dataset $\mathcal{D}$. During the data generation process, we sample $z^\star$ from an approximated prior $p^\star \left( \mathbf{z} \right)$ which depends on the approximation strategy (e.g. posterior sampling). Then we perform inference on the posterior network $p_\theta \left( \mathbf{w} \middle| \mathbf{z^\star} \right)$ to estimate the language distribution. A synthetic utterance $\hat{\mathbf{w}}$ is sampled from said distribution and is used to infer the slot label and intent distribution from the relevant networks, i.e. $p \left( \mathbf{s}, y \middle| \mathbf{z}, \hat{\mathbf{w}} \right)$. The most probable $\hat{\mathbf{s}}$ and $\hat{y}$ are combined with $\hat{\mathbf{w}}$ to form a generated sample set $\left( \hat{\mathbf{w}}, \hat{\mathbf{s}}, \hat{y} \right)$. This generation process is repeated until sufficient synthetic data samples are collected. \section{Experiments} \label{sec:exp} In this section, we outline the design, execution, results and analysis of all experiments pertaining to testing the effectiveness of our GDA approach. \subsection{Datasets} In this paper, we carry out experiments on the following language understanding datasets. \begin{itemize} \item \textbf{ATIS}: Airline Travel Information System (ATIS) \cite{hemphill1990atis} is a representative dataset in the SLU task, providing well-founded comparative environment for our experiments. \item \textbf{Snips}: The snips dataset is an open source virtual-assistant corpus. The dataset contains user queries from various domains such as manipulating playlists or booking restaurants. \item \textbf{MIT Restaurant (MR)}: This single-domain dataset specializes in spoken queries related to booking restaurants. \item \textbf{MIT Movie}: The MIT movie corpus consists of two single-domain datasets: the movie eng (ME) and movie trivia (MT) datasets. While both datasets contain queries about film information, the trivia queries are more complex and specific. \end{itemize} All of the datasets are annotated with slot labels and intent classes except the MIT datasets. The detailed statistics of each dataset are shown in Table \ref{tab:datastats}. In order to simulate a data scarce environment (similar to the experimental design proposed in \cite{chen2016syntax}), we randomly chunk the ATIS training set into equal-sized smaller splits. For the small dataset the training set is chunked into 35 pieces, and for the medium dataset it is chunked into 9 pieces. The sizes of the small and medium training splits approximately correspond those mentioned in the previous work \cite{chen2016syntax}. \subsection{Experimental Settings} Here, we describe the methodological and implementation details for testing the GDA approach under the framework. \subsubsection{General Experimental Flow} Since we observe a high variance in performance gains among different runs of the same generative model (e.g. Figure \ref{fig:ablation}), we need to approach the experimental designs with a more conservative stance. The general experimental methodology is as follows. \begin{itemize} \item $N_G$ identical generative models are trained with different initial seeds on the same training split. \item $m$ utterance samples are drawn from each model to create $N_G$ augmented datasets $\mathcal{D}'_1,\ldots,\mathcal{D}_{N_G}'$. \item $N_L$ identical SLU models are train for \textit{each} augmented dataset $\mathcal{D}'$. All models are validated against the evaluation results on the same validation split. Best model from each SLU model is evaluated on the test set. \item We collect the statistics of all $N_G \times N_L$ results and perform comparative analyses. \end{itemize} \subsubsection{Implementation Details} For all of our models, the word embeddings $W_w$, slot label embeddings $W_s$, and intent embeddings $W_y$ dimensions were 300, 200, and 100 respectively and they were trained with the rest of the network. The word embeddings had been initialized with the GloVe vectors \cite{pennington2014glove}. For the generative model, the encoder is a single-layer BiLSTM-Max model \cite{conneau2017supervised}, which encodes the word embeddings of word tokens $w_i \in \mathbf{w}$ in both directions and produces the final hidden state by applying max-pooling-over-time on combined encoder hidden outputs $\mathbf{h}^{\left(e\right)}_1,\ldots,\mathbf{h}^{\left(e\right)}_T$ (1024 hidden dimensions). The decoders are uni-directional single-layer LSTMs with the same hidden dimensions (1024). Let $\mathbf{h}^{\left(w\right)}_t$, $\mathbf{h}^{\left(s\right)}_t$, and $\mathbf{h}^{\left(y\right)}_t$ be the hidden outputs of word, slot label, and intent decoders at time step $t$ respectively. We perform dot products between respective embeddings and the hidden outputs to obtain logits (e.g. $\mathbf{o}^{\left(w\right)}_t = W_w \mathbf{h}^{\left(w\right)}_t$ etc.). The likelihood of each token at each time step $t$ is obtained by applying the softmax on the logits: $$p \left( w_t \middle| \mathbf{w}_{<t}, \mathbf{z} \right) = \frac{e^{o^{\left(w\right)}_{t,w_t}}}{\sum_{w' \in V_w}{ e^{o^{\left(w\right)}_{t,w'}}}}.$$ Where $V_w$ is the vocabulary set of utterance words. During generation, the beam search algorithm is used to search for the most likely sequence candidates using the conditional token distributions. The beam search size was set to 15 and the utterances were sampled from top-1 ($k_b$) candidate(s) to reduce variance. Exploratory hyperparameter $\lambda_s$ was 0.18. To feasibly train the model, we employ the teacher-forcing strategy, in which the LU network is trained on the ground truth utterance $\mathbf{w}$ instead of the predicted sequence $\hat{\mathbf{w}}$. We applied KLD annealing and the decoder word dropout \cite{bowman2016generating}. KLD annealing rate ($k_d$) was 0.03 and word dropout rate $p_w$ was 0.5. We used Adam \cite{kingma2014adam} optimizer with 0.001 initial learning rate. \subsubsection{SLU Models} \tikzset{ vector/.style = {draw, rectangle, minimum height=1.2cm, minimum width=0.15cm, node distance=0.6cm, line width=0.5pt}, point/.style = {draw, circle, inner sep=0pt, minimum size=0.1cm, line width=0.3pt}, label/.style = {font=\footnotesize} } For the baseline SLU model, we implemented a simple BiLSTM model. A bidirectional LSTM cell encodes an utterance into a fixed size representation $h$. A fully connected layer translates the hidden outputs $h_t$ of the BiLSTM to slot scores for all time step $t$. The softmax function is applied to the logits to produce $p \left( s_t \middle| \mathbf{w}_{\leq t} \right)$. The final hidden representation $h$ of the input utterance is obtained by applying max-pooling-over-time on all hidden outputs. Another fully connected layer and a softmax function maps $h_t$ to the intent distribution $p \left( y \middle| \mathbf{w} \right)$. This simple baseline was able to achieve 95.32 in the slot filling f1-score. \begin{table} \centering \begin{threeparttable} \begin{tabular}{l|ccc|ccc} \hline & \multicolumn{3}{c}{Slot-Gated (Full)} & \multicolumn{3}{c}{Slot-Gated (Intent)} \\ \cline{2-7} Dataset & Slot & Intent & SF & Slot & Intent & SF \\ \hline \hline ATIS & 95.3\tnote{\ddag} & 94.9\tnote{\ddag} & 84.3\tnote{\ddag} & 95.4\tnote{\ddag} & 94.7\tnote{\ddag}& 83.5\tnote{\ddag} \\ ATIS+ & \textbf{95.7} & \textbf{95.6} & \textbf{85.4} & \textbf{95.6} & $\textbf{95.6}$ & \textbf{84.8} \\ \hline Snips & 88.2\tnote{\ddag} & 97.0 & 74.9\tnote{\ddag} & 88.2 & \textbf{96.9} & 74.6 \\ Snips+ & \textbf{89.3} & \textbf{97.3} & \textbf{76.4} & \textbf{88.3} & 96.7 & \textbf{74.6} \\ \hline ME & 82.2\tnote{\ddag} & - & 63.6\tnote{\ddag} & 81.8\tnote{\ddag} & - & 62.1\tnote{\ddag} \\ ME+ & \textbf{82.9} & - & \textbf{64.5} & \textbf{82.8} & - & \textbf{63.3} \\ \hline MT & 63.5\tnote{\ddag} & - & 24.0\tnote{\ddag} & 62.8\tnote{\ddag} & - & 24.4\tnote{\ddag} \\ MT+ & \textbf{65.7} & - & \textbf{27.4} & \textbf{65.0} & - & \textbf{27.5} \\ \hline MR & 72.6\tnote{\dag} & - & 52.8\tnote{\dag} & 72.1\tnote{\ddag} & - & 51.8\tnote{\ddag} \\ MR+ & \textbf{73.0} & - & \textbf{53.4} & \textbf{73.0} & - & \textbf{52.9} \\ \hline \end{tabular} \begin{tablenotes} \item[\dag]{\footnotesize $p<0.1$} \item[\ddag]{\footnotesize $p<0.01$} \end{tablenotes} \end{threeparttable} \caption{Mean data augmentation results on various SLU tasks tested using the slot-gated \cite{goo2018slot} SLU models. Datasets are augmented (prefixed by +) using our proposed generative model. The results have been aggregated and are tested for statistical significance. } \label{tab:general} \end{table} For other SLU models, we consider the slot-gated SLU model \cite{goo2018slot}, which incorporates the attention and the gating mechanism into the LU network. We found the model suitable for our task, as the model is reasonably complex and distinctive from our simple baseline. Furthermore, the code for running the model is publicly available and the results are readily reproducible. We were able to obtain similar or even better results on our environment (Table \ref{tab:atis-slot}). This difference might be due to differing data preprocessing methods. SLU performance is measure by (1) slot filling f1-score (evaluated using the conlleval perl script), (2) intent identification f1-sore, and (3) semantic frame formulation. f1-score measures the correctness of predicted slot labels. \subsection{Generative Data Augmentation Results} In this section, we describe and present two experiments that test the GDA approach under variety of experimental settings: data scarce scenarios, varied SLU models, and varied datasets. \subsubsection{Data Scarce Scenario} \begin{table} \centering \begin{threeparttable} \begin{tabular}{llcc} \hline Dataset & Model & Slot (F1) & Intent (F1)\\ \hline\hline ATIS & JLUVA & 94.44 & 97.09 \\ \hline ATIS & BiLSTM (Baseline)& 95.34 & 97.21 \\ ATIS & Deep LSTM\tnote{a} & 95.66 & - \\ ATIS & Slot-Gated (Full)\tnote{b,d} & 95.66 & 96.08 \\ ATIS & Att. Encoder-Decoder\tnote{c} & 95.87 & \textbf{98.43} \\ ATIS & Att. BiRNN\tnote{c} & 95.98 & 98.21 \\ \hline ATIS+ & BiLSTM (Baseline) & 95.75 & 97.54 \\ ATIS+ & Slot-Gated (Full)\tnote{b,d} & \textbf{96.04} & 96.75 \\ \hline \end{tabular} \begin{tablenotes} \item[a]{\footnotesize \cite{kurata2016leveraging}} \item[b]{\footnotesize \cite{goo2018slot}} \item[c]{\footnotesize \cite{liu2016attention}} \item[d]{\footnotesize run on our environment} \end{tablenotes} \end{threeparttable} \par \caption{Comparisons of the best slot filling and intent detection results for the ATIS dataset.} \label{tab:atis-slot} \end{table} For the first experiment, we test whether our GDA approach performs better than the previous work 1) under the regular condition (full datasets) 2) and data scarce scenarios. We compare our model to a deterministic encoder-decoder model (Seq2Seq) proposed in \cite{kurata2016labeled}. The two decoders of the model learn to decode utterances and slot labels from an encoded representation of the utterance. For the full dataset, we conduct the standard experiments with $N_G=3, N_L=3$ and $m=10000$, synthetic dataset size. For small and medium datasets, each experiment is repeated $N_L=3$ times for \textit{all} $N_T$ training splits. The final result is aggregated from $N_T \times N_L$ runs (i.e. 105 runs for ATIS-small and 27 runs for ATIS-medium). Results are presented in \ref{tab:data-scarce}. According to the results, our approach performed better than all other baselines at the statistically significant level for small and medium datasets. The performance gain of our approach diminishes for the full dataset. This is likely due to the homogeneous nature of the ATIS dataset, leaving little room for the GDA to explore. Although we could not achieve statistically significant improvement on the full dataset, we note that our approach never experiences performance degradation for any dataset size and evaluation measure. \subsubsection{GDA on Other SLU Models and Datasets} We test GDA with various combinations of SLU models and datasets (Table \ref{tab:general}). There were statistically significant improvements in language understanding performances across most datasets and SLU models. Comparing these results with the data scarcity results in Table \ref{tab:data-scarce}, we observe two trends: (1) the more difficult the dataset is to model (e.g. MIT Movie Trivia) and (2) the more expressive the SLU model, the more drastic the improvements are. For example, the improvement rate between ATIS and ATIS+ for full attention-based Slot-Gated model was only 0.39\%, whereas the improvement rate increased nearly ten-fold (3.54\%) between MIT Movie Eng and MIT Movie Eng+ for the same model. We also observe a positive correlation between model complexity and performance gains. For example, the performance improvement was more significant for the slot-gated model than the simple baseline model for the ATIS dataset. This suggests that the performance-boosting benefits from synthetic datasets can be more easily captured by more expressive models. This is also supported by generally better performances achieved by the slot-gated full attention model, as the full attention variant is the more complex one. \subsection{Comparison to Other State-of-the-art Results} In this study, we compare the best LU performance achieved by our generative approach on the ATIS task to other state-of-the-art results in literature (Table \ref{tab:atis-slot}). We chose the best performing run out of all runs carried out from the previous experiments ($N_G=3,N_L=3, m=10000$) and report its results in Table \ref{tab:atis-slot}. In the best case, our approach was able to boost the slot filling performance for the slot-gated (full) model by 0.38. Remarkably, our best results outpeformed more complex models, further supporting the idea of data-centric regularization. We also evaluate the SLU performance of JLUVA by performing deterministic inference (i.e. $\mathbf{z} = \boldsymbol{\mu}$). We find that the LU performance by itself is not competitive. This eliminates the possibility that the performance gains in our approach are attributed to JLUVA being a more expressive model and therefore acting as a teacher network. \subsection{Ablation Studies} In the ablation studies, we carry out two separate comparative experiments on variations of our generative model. \subsubsection{Sampling Methods} The following sampling approaches are considered. \begin{itemize} \item \textbf{Monte-Carlo Posterior Sampling (Ours)}: $\mathbf{z}$ is sampled from the empirical expectation of the model, which is estimated by inferring posteriors from random utterance samples. (Algorithm \ref{alg:posterior-sampling}) \item \textbf{Standard Gaussian}: $\mathbf{z}$ is sampled from the assumed prior, the standard multivariate Gaussian. \item \textbf{Additive Sampling}: First, the latent representation $\mathbf{z}_\mathbf{w}$ of a random utterance $\mathbf{w}$ is sampled. Then $\mathbf{z}_\mathbf{w}$ is disturbed by a perturbation vector $\boldsymbol{\alpha} \sim \mathcal{U}\left(-0.2,0.2\right)$. It was proposed for the deterministic model in \cite{kurata2016labeled}. \end{itemize} \begin{figure} \centering \begin{subfigure}[b]{0.23\textwidth} \resizebox{\textwidth}{!}{ \begin{tikzpicture} \begin{axis} [xmode=log, log ticks with fixed point, xtick pos=left, ytick pos=left, xlabel={Synthetic to Real Data Ratio ($r$)}, ylabel={}, ylabel shift=5pt, label style={font=\large}] \addplot [box plot median] table {aaai2019-kmyoo-data-augmentation-ablation-slot.dat}; \addplot [box plot box] table {aaai2019-kmyoo-data-augmentation-ablation-slot.dat}; \addplot [box plot top whisker] table {aaai2019-kmyoo-data-augmentation-ablation-slot.dat}; \addplot [box plot bottom whisker] table {aaai2019-kmyoo-data-augmentation-ablation-slot.dat}; \draw[dashed] ({rel axis cs:0,0}|-{axis cs:0,0}) -- ({rel axis cs:1,0}|-{axis cs:0,0}); \end{axis} \end{tikzpicture} } \caption{Slot Filling} \label{fig:ablation-slot} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \resizebox{\textwidth}{!}{ \begin{tikzpicture} \begin{axis} [xmode=log, log ticks with fixed point, xtick pos=left, ytick pos=left, xlabel={Synthetic to Real Data Ratio ($r$)}, ylabel={}, ylabel shift=5pt, label style={font=\large}] \addplot [box plot median] table {aaai2019-kmyoo-data-augmentation-ablation-intent.dat}; \addplot [box plot box] table {aaai2019-kmyoo-data-augmentation-ablation-intent.dat}; \addplot [box plot top whisker] table {aaai2019-kmyoo-data-augmentation-ablation-intent.dat}; \addplot [box plot bottom whisker] table {aaai2019-kmyoo-data-augmentation-ablation-intent.dat}; \draw[dashed] ({rel axis cs:0,0}|-{axis cs:0,0}) -- ({rel axis cs:1,0}|-{axis cs:0,0}); \end{axis} \end{tikzpicture} } \caption{Intent Classification} \label{fig:ablation-intent} \end{subfigure} \par \caption{The impact of synthetic data to real data ratio on the relative improvements in SLU performance. The vertical axis shows the relative performance gains, compared to the non-augmented baseline (dashed horizontal lines). For each box plot, the height of the box depicts the variance and outer whiskers mark the minimum and the maximum.} \label{fig:ablation} \end{figure} The results in Table \ref{tab:data-scarce} confirm that exploratory Monte-Carlo sampling based on scaled posterior distribution ($\lambda_s=0.18$) provides the greatest benefit to the language understanding models for the ATIS and the data-scarce datasets. We note that the additive perturbation, despite its simplicity in nature, performs reasonably well compared to our approach. This suggests the exploratory sampling approaches are not only limited to Gaussian distributions. On the other hand, over-simplified and biased approximation of the prior such as standard multivariate Gaussian, could rather cause performance degradation. This also highlights the fact that the choice of sampling approach has a significant impact on the generative quality and thereby the resulting performances. \subsubsection{Synthetic Data Ratio} To gain further insights into generative DA, we conduct regressional experiments to expose the underlying relationship between the relative synthetic data size and the performance improvements. Let $m$ be the size of the synthetic dataset used to augment the original dataset of size $n$. The \textit{synthetic to real data ratio} $r$ is $m/n$. For each run, we conduct the standard experiment procedure ($N_G=10, N_L=5$) on a ATIS-small dataset with JLUVA as the generative model and the simple BiLSTM as the SLU model. We repeat the experiments for all $r \in \left\{0.08, 0.78, 1.56, 3.90, 7.81, 15.6, 39.06, 78.13 \right\}$. From the box plots of our results (Figure \ref{fig:ablation}), we make two observations. First, the maximum marginal improvement is achieved around $10 \leq r \leq 20$ for all evaluation measures. Also, the improvements appear to plateau around $r=50$. Second, The variance starts off relatively small when $r < 1$, but it quickly grows as $r$ increases and peaks around $5 \leq r \leq 20$. The variance appears to shrink again after $r > 20$. A plausible explanation for the apparent trend of the variance is that increasing $r$ enhances the chance to generate performance-boosting key utterances, until no novel instances of such utterances are samplable from the generator, at which point further increasing $r$ only increases the chance to generate already known utterances, thereby reducing the variance. This also explains the plateauing phenomenon. \section{Conclusion} In this paper, we formulated the generic framework for generative data augmentation (GDA) and derived analytically the most effective sampling approach for generating performance-boosting instances from our proposed generative model, Joint Language Understanding Variational Autoencoder (JLUVA). Based on the positive experimental results, we believe that our approach could bring immediate benefits to SLU researchers and the industry by reducing the cost of building new SLU datasets and improve performances of existing SLU models. Although our work has primarily been motivated by the data issues in SLU datasets, we would like to invite researchers to explore the potential of applying GDA in other NLP tasks, such as neural machine translation and natural language inference. Similar to the work done by \citeauthor{dao2018kernel} on the analysis of class-preserving transformative DAs using the kernel theory \cite{dao2018kernel}, our work also calls for deeper theoretical analysis on the mechanism of data-centric regularization techniques. We wish to address these issues in our future work. \newpage \fontsize{9pt}{10pt} \selectfont

ee18544ad12f52395fe089fb01dc41c327731e4d

\section{Introduction} \label{sec:Introduction} The shaping of output pulses from detectors operating in count mode is widely used to resolve overlapped detector pulses and improve their energy resolution. The shaping of detector signals is primarily realised using digital signal processing algorithms \cite{smith2003digital, proakis2001digital}. There are a number of standard shapers \cite{jordanov1994digital,jordanov1994digital1} which convert detector pulses with a short rise time~$\tau_r$ and a long decaying tail $\tau_t$ to shorter pulses whose full width at half maximum (FWHM) is $T_s$. At $T_s\gg\tau_r$, shapers provide well-formed symmetrical output pulses at a low noise level, which ensure unbiased and accurate measurements of pulse amplitudes at count rates lower than $\approx1/T_s$. At shorter pulse widths ($T_s$), the form of the shaped pulses is altered, output noises are increased, and the inferred pulse amplitudes are biased from their real values. These factors restrict the output count rate of standard shapers to a value significantly lower than $1/\tau_r$. The proposed true Gaussian shaper can shape detector pulses into a Gaussian form with $T_s<\tau_r$ to achieve an output count rate for pulse processing several times higher compared with standard shapers. A shaper is defined by its transmission function, calculated as the ratio of the Fourier transforms of the output true Gaussian pulse and the input detector pulse. The capabilities of true Gaussian shaping were tested in an AXAS-D system \cite{KETEKAXAS} with a silicon drift detector (SDD) H7 VITUS \cite{KETEKVITUS} developed by KETEK GmbH in 2012 for high throughput measurements of soft X-ray spectra. The system will be referred hereafter in this paper as the KETEK spectrometer. We will show that the proposed true Gaussian shapers can achieve the maximal output count rate of the spectrometer system, which is several times higher than that obtained with standard shapers, and maintains the energy resolution of the measurements. The layout of this paper is as follows. First, the measured impulse response of the KETEK spectrometer is presented in Section 2 prior the analysis of true Gaussian shapers, because the impulse response determines the transmission function of the shaper. Two analytical approximations of the impulse response are introduced and are used further in the analysis. In addition, the experimental noise characteristics of the detector are presented in the same section, and are used further for estimation of the dead time, energy resolution and output count rate of spectrometers with true Gaussian shapers. The measured impulse response allows for calculation of the transmission function of true Gaussian shapers, presented in Section 3, for output Gaussian pulses of different widths ($T_s$). The shaper exhibited a very strong suppression of signals at frequencies higher than $1/T_s$ for pulse widths $T_s<<\tau_r$. This is the strongly suppressed transmission characteristic at high frequencies that provides the short true Gaussian pulses in the shaper output. Gaussian pulses are compared with the output pulses of trapezoidal shapers in Section 4. Trapezoidal pulses keep a symmetrical and undistorted form when their width $T_s>>\tau_r$. Distortions in these short trapezoidal pulses restrict the output count rate of spectrometer systems. The amplification of detector noise in true Gaussian and trapezoidal shapers is analysed in Section 5 while taking into account the measured noise spectrum of the KETEK system. The dead times of true Gaussian and trapezoidal shapers are calculated in Section 6. For large pulse widths, the calculated dead time coincided with the peaking plus flat top times of trapezoidal shapers. The concept of resolving time is introduced in Section 7 to specify the shortest time interval between two overlapped pulses when their inferred and actual amplitudes differ less than the output shaper noise. The dead and resolving times, along with the amplitude measurement errors, are given in the aforementioned section as a function of the width of the shaped pulses. The output count rate and energy resolution of the KETEK spectrometer with the true Gaussian and trapezoidal shapers are calculated in Section 8 for a wide range of input count rates. The effects of amplitude variations in the shaped pulses with charge collection times in the SDD sensor are analysed in Section 9 for the KETEK spectrometer and a fast SDD with a larger sensor area. The design of a true trapezoidal shaper is discussed in Section 10. The results of this work are summarised in Section 11. \section{Detector impulse response and noise} \label{sec:Responses} A true Gaussian shaper is specified by its transmission function, which is the ratio of the Fourier transforms of the output and input pulses. The proposed shaper should always be designed for the actual impulse response of the employed detector for a single quantum, whereas standard shaper are designed for ideal input pulses with zero rise time \cite{jordanov1994digital,jordanov1994digital1}. Therefore, the proposed shaper is analysed with both actual and ideal input pulses to compare it with standard shapers. The characteristics of the actual detector impulse response and its approximations are considered below. An AXAS-D spectrometer \cite{KETEKAXAS} with a silicon drift detector of standard class H7 VITUS \cite{KETEKVITUS} made by KETEK GmbH was selected for testing and verifying the proposed shaper. The sensitive area of the detector was 7~mm$^2$. The system was developed for spectral measurements of soft X-ray radiation at output count rates up to $3\cdot10^5~s^{-1}$ at an energy resolution of ~139~eV at FWHM for a photon energy of 5895~eV. The analogue output of the AXAS-D system was used as the input of the tested shapers. The analogue impulse response on a single quantum was measured while using a $^{55}Fe$ radiation source emitting $\approx10^4$ photons with an energy of 5895~eV per second on the detector area. Approximately 4000~pulses from these photons were digitized by a custom built ADC with a resolution of 12~bits and a sampling rate of 50~MHz. The passband of the digitizer ranged from 0 to 25~MHz with a signal amplitude damping of $\omega^{-3}$ outside the band. The normalized response of the KETEK system, hereafter referred to as $S_R$ pulse, is shown in \figref{fig:stp}a by a black solid curve. The pulse was interpolated at a sampling period of 2~ns and with an amplitude accuracy of 0.08 \% by averaging the recorded pulses after fine synchronisation of their leading edges. The FWHM of the pulse is 4.6~$\mu$s. The rise time between 0.1 and 0.9 of the pulse amplitude was measured to be 203~ns. The tail of the pulse after scaling by a factor of 100 (plotted in the figure with a dark green dashed curve at $t>10^4$ ns) shows an undershoot of 0.2 \%. The collection times of the electron clouds created by photon impacts in the H7 absorption layer \cite{Gatti1987, Metzger2004} were estimated to be under 6~ns (see Section 9). They are considerably less than the rise time of $\approx$200~ns of the detector response determined using filters of the AXAS-D preamplifier. Therefore, the rise time is slightly affected by variations in the charge collection time. Estimated experimental variations of the rise time of the impulse response of 5895~eV photons were $\approx$10~ns and were fully accounted for the noise of the KETEK spectrometer. \begin{figure*}[h] \centering \includegraphics[width=\linewidth]{Fig1} \caption{\\a. Impulse response $S_R$ of KETEK spectrometer and its approximations\\b. Fourier spectra of the impulse response, its approximations and detector noises} \label{fig:stp} \end{figure*} Small rise time variations of $\approx$2~ns were measured by the KETEK spectrometer by collecting high-energy 100-keV photons from the high temperature plasma of the FT-2 tokamak ~\cite{lashkul2001effect}. These variations were also accounted for when determining the preamplifier noise. The amplitude variations of the shaped pulses with the charge collection time of the H7 detector and for SDDs with a larger sensor area are analysed in Section 9. Two analytical approximations of the impulse response ($S_R$) are also shown in the aforementioned figure. The simplest approximation of the pulse $S_1=A_1\cdot exp(-t/\tau_t)$ with zero rise time and an exponentially decaying tail, $\tau_t$=5.95~$\mu$s, is shown in \figref{fig:stp}a by a blue dashed curve. The exponential pulse can be converted with standard digital shapers \cite{jordanov1994digital,jordanov1994digital1} to a symmetrical output pulse of trapezoidal, cusp, or pseudo-Gaussian forms. Trapezoidal shapers are known for having the lowest delta noise \cite{Goulding1983} and for their unbiased amplitude measurements of overlapped pulses. Therefore, they are widely used in detection systems, and are thus compared with true Gaussian shapers in the paper in terms of count rate, resolution, and biasing of pulse amplitude measurements. A more accurate approximation is one that takes into account a finite rise time of the impulse response, but assumes instant charge collection from the electron clouds in the SDD sensor layer: $S_2=A_2\cdot exp(-t/\tau_t) (1-exp(-t/\tau_f))$. The leading edge front time $\tau_f$=0.10~$\mu$s was determined by the integration time of the RC-CR filter of the pre-amplifier in the SDD output \cite{Knoll2010}, which forms the impulse response of the KETEK spectrometer. The $S_2$ form is very close to the measured pulse, as shown in the aforementioned figure by a red dashed-dotted curve which shows the difference between the approximated and real pulses, $S_2-S_R$. The most important difference between these pulses happens in first 50~ns after their start. The nonlinear growth of the $S_R$ pulse is related to the finite charge collection time of electron clouds by the SDD anode and integration of the anode pulse by the reset pre-amplifier shown in \figref{fig:RC-CR filter}. Fourier spectra of the $S_R$ pulse and its approximations are plotted in \figref{fig:stp}b. The spectrum of the first approximation, $F_1 \propto \tau_f/(1-\omega\tau_f)$, is shown by a blue dashed curve, and it can be seen that it decays slowly with frequency as $\omega^{-1}$. The second approximation, $F_2 \propto \tau_t/(1-\omega\tau_t)/(1+\tau_f/\tau_t-\omega\tau_f)$, is shown by a red dashed-dotted curve, and it provides a higher suppression of the output signals at frequencies $\omega>1/\tau_f$. The response spectrum, $F_R$, is presented by a black solid curve. It is significantly damped above $f_c\approx$8~MHz compared with the spectra of the approximations. This difference underlies in the restriction of the output pulse widths of standard shapers, as will be shown further on. The noise spectrum in the output of the detector system is shown by a green thick dashed curve. The noise was measured as the voltage variations of the ADC output for a bandwidth of 25~MHz during a time interval of 15 ms when the detector was in darkness. The standard deviation of the electronic noise was estimated to be approximately 55~eV. The amplitude of the noise spectrum decays with $\omega^{-3}$ for frequencies above the cut-off frequency, $f_c$=8~MHz. The signal spectrum $F_R$ under the noise level in the high-frequency band was measured by averaging many response pulses. The measured parameters of the KETEK spectrometer impulse response were used for the design of a true Gaussian shaper of a given pulse width. \section{True Gaussian shaper of detector pulses } \label{sec:Shaper} Trapezoidal, cusp, and pseudo-Gaussian standard shapers \cite{jordanov1994digital,jordanov1994digital1} are designed for shaping pulses with an exponentially decaying tail and a very short rise time $\tau_r$, i.e. pulses of type $S_1$. The shaped pulses get an asymmetrical form when the shaping width approaches the rise time $\tau_r$ of the input pulse. The actual width of the output shaped pulses is restricted by the rise time of the input pulses, as will be shown in Sections 4 and 5. The amplitude spectra of input pulses with longer rise times ($\tau_r\approx T_s$) decay with $\omega^{-2}$ for frequencies above $1/\tau_f$, as shown in \figref{fig:stp}b. The spectra of pulses with an ideal trapezoidal or cusp forms also decay above their characteristic frequency with $\omega^{-2}$. Therefore, the damped spectra of the input pulses significantly affect the shape and width of the output pulses when the inverse rise time is close to the characteristic frequency of the shaper. In other words, the output pulses get wider and more distorted when the rise time approaches the characteristic width to be provided by the shaper. Indeed, the symmetrical output pulses of standard shapers are much wider than the rise time of the input pulse, as illustrated in Section 4. The cause of this restriction is the slow decaying spectra of ideally shaped pulses in the high-frequency band $\omega>1/\tau_f$. Short-shaped pulses are possible if their spectra are less affected by the damped spectra of the input pulses in this band. This requirement is best satisfied in true Gaussian pulses, $S_G=exp(-(t/\tau_G)^2/2)$, having a spectral amplitude with a Gaussian form: $F_G = \tau_G\cdot~exp(-(\omega\tau_G)^2/2)$. The transmission function of the true Gaussian shaper is defined as the ratio of the Fourier transforms of the output Gaussian and input $S_R$ pulses: \begin{align} & K_G(\omega, T_s)=\frac{F_G(\omega, T_s)}{F_R(\omega)} \end{align} Here, $T_s$ is the width of the Gaussian pulse at FWHM. The transmission function is calculated for the Gaussian pulse with an amplitude equal to that of the input pulse and centred on the $S_R$ time interval. Gaussian values outside this interval are replaced by zeros. This positioning of the Gaussian pulse minimizes its jumps at the start and end of the $S_R$ time interval and ensures that the transmission function satisfies the causal principle. The impulse response $H_G(t, T_s)$ of the true Gaussian shaper for its implementation as a FIR filter is found as the inverse Fourier transform of the transmission function $K_G(\omega, T_s)$. \begin{figure*}[h] \centering \includegraphics[width=\linewidth]{Fig2} \caption{Transmission (a) and impulse response (b) of true Gaussian shapers at different pulse widths} \label{fig:Responses} \end{figure*} The transmission and impulse response functions of the true Gaussian shaper calculated for the $S_R$ pulse are plotted in \figref{fig:Responses} versus frequency and time, normalized to the FWHM of the shaped pulse $T_s$. The impulse responses were multiplied by the pulse widths $T_s$ to present the plots in a closer scale. A high gain of $K_G(\omega, T_s)$ around $f\approx1/T_s$ can be accounted for by the difference between the rise and peaking times of the input and output pulses. The short peaking times of Gaussian pulses require a high amplification in the shaper around the characteristic frequencies $f\approx1/T_s$ of these pulses. Output noise is also added in this band, and when the noise level approaches the pulse's amplitude, the true Gaussian shaping fails. The experimental detector signal $S_D(t)$ of length $N_D$ is a sum of the responses on photons randomly arriving in time. This signal is converted to a sequence of Gaussian pulses $S_G(t)$ using the inverse Fourier transform of the product of the signal spectrum and the transmission function or by convolving `the impulse response $H_G(t, T_s)$ and the signal $S_D(t)$ digitized at a sampling period of $\Delta t$: \begin{align} & S_G(n)=\sum^{P}_{k=0}H_G(k,T_s)S_D(n-k), ~~~~ t=n\Delta t \end{align} The shaper transmission function $K_G(\omega,T_s)$ must have the same length as the spectrum of the signal. For that matter, the impulse response $S_R$ is expanded to a length of $N_D$ using an exponential approximation of its tail. The shaping procedure uses three fast Fourier transforms, whose total computational complexity is proportional to $N_F=3\cdot N_D\cdot log_2(N_D)$. This complexity is reduced by one third for repeating measurement intervals of the same length. The impulse response function $H_G(t, T_s)$ of the shaper or its FIR filter kernel can be truncated to a time region of several times the width $T_s$ around its maximum (see \figref{fig:Responses}b). The number of non-zero terms P of the function is determined by the sampling frequency of the input signal, the Gaussian width $T_s$, and the required conversion accuracy. About 30 non-zero terms are needed to convert input signals digitized at a sampling rate of 50~MHz to Gaussian pulses of 90~ns FWHM with an accuracy of 0.1 \% of their maximum values. The computational complexity of this convolution is $\propto N_D\cdot P$. Thus, the computational complexity of the FIR filter and the inverse transform are equal when the signal length is approximately $2^{P/3}$ samples. This estimation is in accordance with \cite{smith2003digital}. True Gaussian shaping in real time requires digital signal processors (DSPs) of moderate computing power, e.g. a DSP operating at a frequency of hundreds of MHz is able to convolve an $S_R$ pulse digitized at a sampling rate of 50~MHz to a Gaussian form with the use of a 30-term kernel. The optimal convolution to a true Gaussian shape is a special issue, which should be addressed in another paper. In this work, the inverse Fourier transform was employed for accurate analysis of the true Gaussian shaper with actual recorded and simulated signals of the detection system. \section{Forms of the shaped pulses} \label{sec:Forms} Output pulses of the true Gaussian and trapezoidal shapers are presented in \figref{fig:Forms}a and b for the response pulses of the AXAS-D pre-amplifier, plotted in the figure by black dashed curves. The trapezoidal output was calculated with a recursive algorithm which converts an exponential pulse into a symmetrical trapezoidal form \cite{jordanov1994digital1}. The output signal of the shaper was formed as subtractions of the properly delayed and scaled digitized input signal. The shaper converts a signal with an infinite exponential tail into a symmetrical trapezoidal pulse with gain M, determined by the time constant of the input pulse tail and the sampling period; see (5) in \cite{jordanov1994digital1} for more information. The trapezoidal pulse was normalized so that its amplitude equalled that of the input $S_R$ pulse. The conversion of $S_R$ pulses with an undershoot and not exactly an exponential tail results in an asymmetrical output trapezoidal pulse which is followed by an undershot decaying tail. One can improve the form of the output tail by changing the gain coefficient M in the algorithm (see, for example, \cite{Dey2014}) but the tail cannot be completely eliminated. Further in the paper, the trapezoidal algorithm is applied with the gain value calculated for the exponential tail of the $S_1$ approximation of the detector impulse response. \begin{figure*}[h] \centering \includegraphics[width=\linewidth]{Fig3} \caption{Normalized output pulses of true Gaussian and trapezoidal shapers\\ a. Shaping width 1.6 $\mu$s,\\ b. Shaping width 0.08 $\mu$s} \label{fig:Forms} \end{figure*} Both shapers can provide output pulses of the same width if the width of their outputs is larger than the rise time of the input pulses. Output pulses of the shapers at $\tau_{FWHM}$=1.6~$\mu$s are plotted in \figref{fig:Forms}a. The Gaussian pulse is plotted in a red solid curve. Two trapezoidal pulses of different forms and the same FWHM are plotted in blue dashed-dotted and green dashed curves. The first form is a Gaussian-like one provided by a trapezoidal shaper which time parameters are set so that to equate the width and slopes of the trapezoidal pulse to those of the Gaussian pulse measured at half maximum. The peaking and flat top times of this shaper are as follow: \begin{align} & \tau_{Tp} = \frac{\tau_{FWHM}}{2ln(2)}, ~~~~~\tau_{Tt}= \tau_{FWHM}-\tau_{Tp}\approx 0.28\tau_{FWHM} \end{align} The second trapezoidal pulse is set to a rectangular-like form with short 40-ns fronts and a long 1560-ns flat top and the same FWHM. The actual output pulse has curved leading and falling edges which are considerably distorted from a pure trapezoidal form. The output noise of this shaper was three times higher than the noise levels of the two previous ones. A Gaussian output pulse of 0.08~$\mu$s at FWHM is shown in \figref{fig:Forms}b by a red solid curve. A Gaussian-like trapezoidal pulse at the same noise level is shown in a blue dashed-dotted curve. The parameters of this shaper set in accordance with (3) are $\tau_{Tp}$ =20 ns and $\tau_{Tt}$ =40 ns. This pulse is strongly distorted from the trapezoidal form and its FWHM is larger than that calculated from the settings. The output noise level of the shapers, to be discussed in Section~\ref{sec:Noise}, is $\approx$8 times higher than that of the shapers with wider output pulses. The asymmetrical form of the short Gaussian-like trapezoidal pulse is certainly less advantageous for the detection of overlapped pulses. A trapezoidal pulse with a flat-top setting elongated to 240~ns, shown by a green dashed curve in \figref{fig:Forms}b, also has a strongly distorted form determined by its short 40-ns peaking time. The output noise of the Gaussian-like trapezoidal shaper is much less than the noise of rectangular-like trapezoidal shaper because of its longer peaking time. The noise is close to the minimal level provided by trapezoidal shapers at the given pulse width. Therefore, all further comparisons will be between true Gaussian and Gaussian-like trapezoidal shapers. Both shapers are solely characterized by the width of the output pulse at FWHM which is the sum of peaking and flat top times of an ideal trapezoidal pulse. Trapezoidal pulses converted from the $S_R$ pulse have a long tail with an undershoot of $\approx$0.5 \% of the pulse amplitude caused by a ~0.2 \% undershoot of the $S_R$ input. This undershoot tail does not affect pulse amplitude measurements at low count rates, but should be taken into account for high count rate operation of the spectrometer, as discussed in Section~\ref{sec:Detection}. \section{Noise factor of shapers} \label{sec:Noise} Amplification of noise in a shaper is determined by its transmission function. The noise factor of shapers is defined as the ratio of the standard deviations of the output and input electronic noise measured in volts when the spectrometer is in darkness. The voltage standard deviation is expressed in photon energy by using a scale factor determined during calibration. The standard deviation of the output noise of the KETEK spectrometer, as noted earlier, is 55~eV. The spectrum of this noise compared with the spectrum of the impulse response to a photon with 5895~eV is shown in \figref{fig:stp}b. The noise spectrum of output shaped pulses is the transmission function of the shaper (1) times the input noise spectrum. \begin{align} & Noise_{Out}(\omega, T_s)=K_G(\omega, T_s)\cdot Noise_{In}(\omega) \end{align} Thus, the noise factor of the the shaper explicitly depends on the width of the shaped pulse $T_SR$. The output noise of the Gaussian shaper was calculated as the inverse Fourier transform of the output noise spectrum (4). The output noise of the trapezoidal shapers is calculated from the same input noise with the use of the recursive algorithm \cite{jordanov1994digital1} applied to generate trapezoidal response of a given pulse width from the input signal. \begin{figure}[h] \centering \includegraphics[width=\linewidth]{Fig4} \caption{Noise factors of true Gaussian and Gaussian-like trapezoidal shapers. The shaped pulse width is given in terms of FWHM.} \label{fig:Noises} \end{figure} The noise factors of the shapers for the $S_R$ pulse and its two approximations are plotted in \figref{fig:Noises} versus the FWHM of the shaped pulses. The FWHM of distorted trapezoidal pulses were measured but not calculated from the set peaking and flat top times. True Gaussian and Gaussian-like trapezoidal shapers exhibit the same noise factors for the exponential input pulse $S_1$. These noise factors are shown in the figure by green dashed-dotted curves without and with circle marks, respectively. At large pulse widths the factors decrease inversely with the square root of the peaking time of the output pulses. The roll-off of the noise factor at small widths can be accounted for less amplification of the shapers at low frequencies and noise damping at high frequencies, i.e. noise decreases when the noise dumping frequency $f_c=\omega_c/2\pi$ becomes less than the frequency $f\approx1/T_s$ corresponding to maximum of the transmission function (see \figref{fig:stp}b and \figref{fig:Responses}a). Both shapers have the same noise factors if the FWHM of the shaped pulses is longer than the front time $\tau_f\approx$100~ns of the input pulses of $S_2$. These factors are plotted for the true Gaussian and trapezoidal shapers by dashed blue curves without and with open triangle marks, respectively. The shortest strongly asymmetrical output pulse of the trapezoidal shaper had a width of $\tau_f\approx$100 ns at FWHM. True Gaussian shaping allows for much shorter symmetrical output pulses, as can be seen from the red solid and blue dashed curves in \figref{fig:Noises}. The width of the trapezoidal output pulses converted from the $S_R$ pulses is also limited to 100~ns at FWHM as for the $S_2$ input pulses, but their form is even more distorted. True Gaussian shapers provide shorter symmetrical pulses, although at a higher noise factor, as shown by the red solid curve in \figref{fig:Noises}. The noise factor of the trapezoidal shaper is plotted by red bottom-up triangles. Thus, the true Gaussian shaper provides symmetrical output pulses which are much shorter than the rise time of the input pulse and, therefore, more advantageous for the detection of strongly overlapped input pulses. The noise factor of the shaper depends strongly on the form of the input pulse and the noise spectrum of the detector, which can restrict possible settings of the shaper. The detection capabilities of true Gaussian shapers are analysed in the following sections of this paper. \section{Dead time of true Gaussian shapers} \label{sec:DeadTime} The detection of closely overlapped photons and the maximal output count rate of the spectrometer is determined by its dead time \cite{Knoll2010, AMPTEKGloss, AMPTEKSDD, AMPTEKHigh}, i.e. the time interval after registration of a photon during which the system is not able to register any subsequent photons. The dead time of a trapezoidal shaper with an output pulse of undisturbed form and without pile up rejection is the sum of the peaking and flat top times of the pulse, see (2) in \cite{AMPTEKSDD}. However, this standard dead time is not certain for highly distorted short trapezoidal pulses with non-linear fronts. The peaking time is not directly applicable to the calculation of the dead time of true Gaussian shapers either. Therefore, the dead times of true Gaussian and Gaussian-like trapezoidal shapers were found in this work by means of a numerical model while taking into account the output noise of the shapers. The numerical model considers the impulse response of the KETEK system on two subsequent photons of similar energy and a time lag T. The detector signals are digitized at a sampling period $\tau_s$ and converted by true Gaussian or trapezoidal shapers to shorter pulses. The dead time of the shapers is found, in the proposed model, as the smallest time interval between two subsequent output pulses when: \begin{enumerate}[(1)] \item the sum of two single pulses has two maxima corresponding to the pulse peaks, separated by a valley; \item the heights of both maxima compared with the valley are larger than twice the standard deviation of electronic noise at the shaper output. \end{enumerate} The statistical errors of the pulse amplitude measurements are determined by the charge statistics of pair creation in the detector \cite{lechner1996pair,Schlosser2010} and electronic noises. The input electronic noises of the shapers are modelled as normal white noise smoothed by a third-order low-pass filter to fit the experimental noise spectrum shown in \figref{fig:stp}b. The standard deviation of the smoothed electronic noises was 55~eV. Instrumental or biasing errors are defined as the differences between the mean inferred peaks of two overlapped shaped pulses and their actual mean amplitudes. These parameters were calculated for a pair of pulses with mean amplitudes of 5895~eV each versus the output pulse widths of the true Gaussian and trapezoidal shapers. Input signals were digitized at a sampling frequency of 50~MHz. The dead times of the shapers, shown in \figref{fig:DeadTime}a, were nearly equal to the durations of the shaped pulses at FWHM. Both shapers had their lowest dead time determined by the sampling frequency and pulse amplitude for given detector parameters. The minimal dead time of the true Gaussian shaper was approximately half of that of the trapezoidal shapers: 60~ns versus 100~ns, respectively. The set peaking time of the trapezoidal shapers is shown in the plot by a black solid curve. Note, that short asymmetrical trapezoidal pulses are characterized by a single dead time value in accordance with the definition given in the beginning of this Section. \begin{figure}[H] \centering \includegraphics[width=\linewidth]{Fig5} \caption{Dead times (a), measurement errors of single peaks (b) and biasing errors (c) of two peaks lagged by the dead time interval from each other for true Gaussian and Gaussian-like trapezoidal shapers. The widths of shaped pulses are given in terms of FWHM.} \label{fig:DeadTime} \end{figure} The measurement errors of the pulse amplitudes were estimated as the quadratic sum of the output electronic noise of the shapers and the Fano statistics of electron hole pairs created by photon impacts in the detector area. The peak spread caused by Fano noises, estimated in \cite{lechner1996pair, Schlosser2010}, were 118~eV at FWHM. Both shapers had the same measurement errors for the peak amplitudes, and an energy resolution of $\approx$150~eV at FWHM, measured at the large shaping widths, as shown in \figref{fig:DeadTime}b. This noise level corresponds to the experimental value inferred from the experimental data for large shaping widths. The electronic noises at the shaper outputs start to dominate over the Fano noises for small shaping widths. The true Gaussian shaper performed better in terms of measurement error when the shaped pulses were shorter than the rising time $\tau_r \approx$ 200~ns of the impulse response~$S_R$. The inferred amplitudes of two close pulses may be biased from their actual values. The ideal trapezoidal shaper is known as a biasing-free filter for large pulse widths, as confirmed in \figref{fig:DeadTime}c. Distortions of short trapezoidal pulses result in considerable biasing of the mean amplitude of the pair. A blue dashed curve with open circles in \figref{fig:DeadTime}c represents the biasing amplitudes of the second pulse in a pair, separated from the first peak by the dead time interval. On the other hand, the amplitude of the first pulse remains unbiased, as depicted in the plot by a blue solid curve with solid circles. This biasing limits the application of trapezoidal shapers with KETEK-like response pulses for dead times larger than $\approx$1~$\mu$s. On the contrary, true Gaussian shaping results in lower and equal biasing amplitudes for two short pulses separated by the dead time, as shown in the plot by the red curve with cross marks. For larger widths, the inferred amplitudes of two closely overlapped Gaussian pulses become strongly biased. The advantageous capabilities of the true Gaussian shapers of detection of very close pulses allows a considerable increase of the maximal output count rate of pulse processing, as will be shown in the next sections. Very shot true Gaussian pulses can be employed for pulse detection at poor amplitude resolution in the fast channel of the spectrometer \cite{AMPTEKSDD, AMPTEKHigh}. The found dead time interval is used for rejection too close pulses. The remained pulses are used for the pile-up of true Gaussian pulses measured in slow channels at a higher amplitude resolution. \section{Resolving time of true Gaussian shapers} \label{sec:ResolveTime} Accurate measurements of the pulse amplitudes of both trapezoidal and Gaussian pulses in terms of amplitude error and bias require a time lag between the pair somewhat larger than the dead time. We hereby define as resolving time the smallest time interval between two detected subsequent pulses when \begin{enumerate}[(3)] \item the inferred amplitudes of the overlapped detected pulses differ from the actual amplitude less than twice the standard deviation of the output electronic noise. \end{enumerate} A true Gaussian pulse is resolved correctly when its time separation to both the preceding and successive pulses is larger than the resolving time. The resolving time of the true Gaussian shapers is presented in \figref{fig:ResolveTime} by a red solid curve versus their dead time. The dead times coincide with resolving times for very short Gaussian pulses and become larger as the pulse width is increased. On the contrary, wide trapezoidal pulses have equal resolving and dead times. The resolving time of shorter trapezoidal pulses is split in two values. The first one relates to the first pulse of the pair and is labeled in the figure as 'Trapezoidal 1'. Its amplitude is less affected by the successive pulse and therefore the successive resolving time remains close to the dead time of the pair. The second time relates to resolution of the second pulse of the pair. Its amplitude can be strongly affected by a long tail of the preceding pulse. Therefore, the resolving time of the second pulse increases with shortening the dead time as shown in the figure by a blue dash curve. \begin{figure}[h] \centering \includegraphics[width=\linewidth]{Fig6} \caption{Resolving time of true Gaussian and Gaussian-like trapezoidal shapers.} \label{fig:ResolveTime} \end{figure} Thus, trapezoidal shapers are better for shaped pulses of long or moderate widths, whereas Gaussian shapers are advantageous when operating with very short shaped pulses. The minimal resolving time of trapezoidal shapers is restricted by about the rise time of the $S_R$ pulse ($\tau_p \approx$480~ns), defined as the time interval between its maximum level and 1 \% of it (see \figref{fig:ResolveTime}). True Gaussian shaping reduces this limit below 100~ns and significantly increases the output count rate of the KETEK spectrometer. \section{Detection of pulses at a high count rate} \label{sec:Detection} The true Gaussian and trapezoidal shapers should be compared at high count rate operation, but the available radiation source $^{55}Fe$ could not provide a photon flux of the required intensity of $\approx 10^7$ photons per second on the detector area. Therefore, detector signals at high photon fluxes were simulated using the characteristics of the KETEK system measured at lower count rates of $\approx 10^4$ 1/s. A radiation source was simulated to emit photons of 5895~eV energy, distributed in time according to the Poisson law. The inverse mean time interval between subsequent photons which impact the SDD sensor is the input count rate of the SDD. The output count rate of a detection system is defined as the inverse mean time interval between subsequent resolved impulse responses. The energy resolution of the shaped pulses is determined by the pair creation statistics in the detector layer and the electronic noise \cite{lechner1996pair} (see Sections 6 and 7). The measured standard deviation of the electronic noise of the KETEK system amounted to 55~eV. The electronic noises of the detection system were modelled as white noise smoothed by a third-order low-pass filter with a cut-off frequency of 8~MHz. The spectrum of the modelled noises corresponded to that shown in \figref{fig:stp}b. The simulated impulse responses at the output of the KETEK system were synchronised with impact photons. The response amplitudes are normally distributed around the mean photon energy with the standard deviation of the Fano noises \cite{Schlosser2010}. The sum of the simulated impulse responses and noises were converted by the true Gaussian and trapezoidal shapers to a sequence of shorter pulses in the way described above. The shaped signals were analysed using a numerical code which returns the times and amplitudes of the peaks detected above the threshold level, defined as five standard deviations of the output shaper noise. \begin{figure*} \centering \includegraphics[width=\linewidth]{Fig7} \caption{Output count rate with true Gaussian and Gaussian-like trapezoidal shapers for large (a) and small (b) shaping widths.} \label{fig:CountRates} \end{figure*} A true Gaussian shaper with output pulses of 60 ns was served as a fast channel for pile-up selection of peaks detected in slow channels of the Gaussian and trapezoidal shapers, as described in \cite{Goulding1983,AMPTEKHigh}. The peaks detected in the slow channels were tested one by one for the time intervals to their closest preceding and successive neighbors found in the fast channel. The peaks separated less than the resolving time were removed from the counted set. The preceding and successive resolving times were applied for the pile-up selection of trapezoidal pulses. The peaks selected in the counted set will be hereafter referred to as the resolved peaks. The output count rate, energy resolution and biasing were calculated for 1000 input pulses. Therefore, the simulated data have some spread about their mean values. The count rates of the resolved peaks for small and large shaping widths are shown in two plots in \figref{fig:CountRates} for the true Gaussian and trapezoidal shapers. The output rates fit well to the theoretical expression $C_{out}=C_{in}\cdot exp(-2C_{in}\cdot \tau_R)$ \cite{Knoll2010, AMPTEKSDD}, where $C_{out}$ and $C_{in}$ are the output and input count rates, respectively, and $\tau_R$ is the resolving time of the system. The output rates that are equal to the input rates are plotted in the figure by the upper dashed-dotted straight line. The maximal output count rates $C_{in}/e$ achieved for $C_{in}^{max}=1/2\tau_R$ are plotted by the lower dashed-dotted straight lines. The calculated output count rates are presented only within these limits. \begin{figure}[!h] \centering \includegraphics[width=\linewidth]{Fig8} \caption{Amplitude resolution of pulses resolved with true Gaussian and Gaussian-like trapezoidal shapers} \label{fig:Resolution} \end{figure} The first plot in the figure presents data for shaping widths nearly equal or larger than the rise time of the $S_R$ pulse. The output count rates of the true Gaussian shapers are lower than those of the trapezoidal shaper, as expected from~\figref{fig:ResolveTime}. \begin{figure}[h] \centering \includegraphics[width=\linewidth]{Fig9} \caption{Biasing errors of pulse amplitude measurements with true Gaussian and Gaussian-like trapezoidal shapers} \label{fig:Bias} \end{figure} The second plot in the figure presents the count rates calculated for small shaping widths, where true Gaussian shapers are superior and provide a lower resolving time. At small widths, the maximal output count rate of the Gaussian shapers exceeds $3\cdot10^6$~counts per second, which is more than three times higher than that calculated for the trapezoidal shapers and ten times higher the maximal rate specified by the manufacturer \cite{KETEKVITUS}. The highest output count rate, $>4\cdot10^6$~counts per second, was achieved with Gaussian pulses of 70~ns at FWHM. Obtaining higher rates is impossible because the shaping noises increase as the pulse widths decrease. The spectral resolution of the resolved pulse amplitudes is shown in \figref{fig:Resolution}. The data are also restricted by the maximal output count rate and the curves are marked in the same way as in \figref{fig:CountRates}. The resolution is expressed in terms of the FWHM of the spectral peak around the mean photon energy of 5895~eV. The resolutions of the both types of shapers providing long shaped pulses are in the range of the technical specifications of the tested system. At higher rates the resolution of the true Gaussian and trapezoidal shapers of the same pulse widths are similar for count rates lower than the maximal output rate of the trapezoidal shapers. The calculated resolution and output count rate of all trapezoidal shapers presented in the Section correspond well to those of a fast SDD spectrometer developed by AMPTEK \cite{AMPTEKHigh}. The true Gaussian shapers surpass significantly the maximal output count rate of the trapezoidal shapers of the same pulse width without losses of energy resolution. \begin{figure}[h] \centering \includegraphics[width=\linewidth]{Fig10} \caption{Tails of the KETEK and shaped pulses with and without digital pole-zero cancellation} \label{fig:UndershootCancellation} \end{figure} The inferred amplitudes can be biased from the mean amplitude of the input pulses. The biased amplitudes of the shaped signals for the true Gaussian and trapezoidal shapers are shown in \figref{fig:Bias}. True Gaussian shaping provides a much lower bias of the resolved pules than trapezoidal shaping. The large negative biasing amplitudes of trapezoidal pulses are caused by the accumulation of long undershoot tails of many pulses. These undershoots come from a 0.2 \% undershoot of the $S_R$ pulse, as shown in \figref{fig:stp}a. The undershoot of this pulse is not perfectly cancelled with the pole-zero circuit of the preamplifier \cite{Knoll2010}. The further cancellation was done with a digital reversed pole-zero circuit which was found to significantly reduce the tail of the KETEK spectrometer impulse response, as shown in \figref{fig:UndershootCancellation}. The measured undershoot tail is shown by a solid black curve. The output tail after the digital cancellation filter is plotted by a dashed-dotted red curve. The tails of the trapezoidal pulses converted from the measured and filtered $S_R$ pulses are plotted in dashed blue and dotted magenta curves, respectively. The cancellation filter allows for a reduction of biasing in the trapezoidal amplitudes down to the biasing of Gaussian pulses, as shown in \figref{fig:Bias}, and does not change other characteristics of the trapezoidal shapers. Note that the pole-zero cancellation does not affect the biasing in closely overlapped trapezoidal pulses considered in Section \ref{sec:DeadTime} and shown in \figref{fig:DeadTime}c. \section{Variations in the amplitudes of shaped pulses} \label{sec:Variations} An impact photon creates an electron cloud in the sensor layer of the SDD. The cloud drifts in the electric field to the SDD anode, expending in size. Therefore, the charge collection time of the cloud depends on the distance between the anode and the impact point \cite{Gatti1987, Prigozhin2012}. Thus, distribution of collected photons across a large detector area can result in the ballistic deficit of shaped pulses, i.e. variations of their amplitudes with the charge collection time. \begin{figure}[h] \centering \includegraphics[width=\linewidth]{Fig11} \caption{Reset pre-amplifier and RC-CR filter of SDD anode signals} \label{fig:RC-CR filter} \end{figure} The amplitude variations of the shaped pulses were analysed by means of the following model. First, the charge collection time is estimated with the use of \cite{Gatti1987} for the KETEK H7 SDD in terms of the rms widths of the anode current pulses $\tau_C$. These pulses are integrated by a reset preamplifier, as shown in \figref{fig:RC-CR filter}, which converts them to step-wise voltage signals. The rise time of these signals is determined by the charge collection time, and their amplitude is proportional to the photon energy used. The step-wise signals are shaped by means of an RC-CR filter connected to the output of the reset preamplifier, as shown in \figref{fig:RC-CR filter}. \begin{figure}[h] \centering \includegraphics[width=\linewidth]{Fig12} \caption{Amplitude variations of shaped pulses of the KETEK spectrometer} \label{fig:AmplVariationsH7} \end{figure} The maximal rms width of the current pulses of the KETEK H7 SDD found from \cite{Gatti1987} is 6~ns. The integration time of the filter $\tau_i =R_iC_i$ determines the rise time of the impulse response, whereas its decay is specified by the differentiation time $\tau_d =R_dC_d$. The measured impulse response $S_R$ of a small area of the H7 SDD was fitted using the model for $\tau_i$ =112~ns and $\tau_d$ =5945~ns. The output amplitudes of a true Gaussian shaper and two trapezoidal shapers were calculated using the model and the suggested parameters of the KETEK spectrometer. A Gaussian shaper of width $\tau_{G0}$ was designed and calibrated to the photon energy at the lowest charge collection time of the detector, and was applied for collection times in the full range. The first trapezoidal shaper converted the SDD pulses into a Gaussian-like trapezoidal form. The second shaper was designed for trapezoidal pulses with the lowest peaking time (one sampling period set for the leading edge of the pulse) and longest flat top. The FWHM of both trapezoidal shapers were set to that of the Gaussian pulse $\tau_{G0}$. The amplitude variations of the shaped pulses versus the charge collection time $\tau_C$ of the KETEK H7 SDD were calculated for a sampling frequency of 50~MHz and are shown in \figref{fig:AmplVariationsH7}. The amplitudes of the SDD pulses are plotted in a black curve and are independent from $\tau_C$. The amplitudes of Gaussian pulses of 70~ns at FWHM decreased by 200~eV with $\tau_C$, as shown by the red solid curve. This is one third of the spectrometer resolution calculated in the previous section for the same pulse width. The amplitude variations of both types of trapezoidal pulses are plotted in the figure by green solid triangles and blue solid squares. They ranged from 300 to 600~eV and were larger than the calculated energy resolution of both types of trapezoidal shapers. \begin{figure}[h] \centering \includegraphics[width=\linewidth]{Fig13} \caption{Normalized range of amplitude variations with shaped pulse width at fixed integration times for KETEK spectrometer. The widths of shaped pulses are given in terms of FWHM.} \label{fig:AmplVariationsKETEK} \end{figure} Gaussian amplitude variations decrease with shaped pulse width, and for $\tau_{G0}$=200~ns they amount to $\approx$30~eV, i.e. they are negligible with the SDD resolution at low count rates. They are plotted in the figure by a red dashed curve. The first type of trapezoidal shaper mentioned exhibited the same variations, shown in the figure by green open triangles over the red dashed curve. The second type of trapezoidal shaper with rectangular output pulses had variations several times larger, which are plotted by a blue dashed curve with open squares in the figure. \begin{figure}[h] \centering \includegraphics[width=\linewidth]{Fig14} \caption{Normalized range of amplitude variations with shaped pulse width at fixed integration times of AMPTEK XR100 SDD. The widths of shaped pulses are given in terms of FWHM.} \label{fig:AmplVariationsAMPTEK} \end{figure} The model used shows that amplitude variations of Gaussian pulses are independent of the integration time $\tau_i$. The amplitude variations of trapezoidal pulses depend both on integration time and on pulse width, as shown in \figref{fig:AmplVariationsKETEK}. Here, the range of amplitude variations at a fixed integration time was normalized to the mean amplitude of the shaped pulses of the KETEK spectrometer. The minimal width of the trapezoidal pulses is limited to $\approx80 \%$ of the integration time $\tau_i$. The amplitude variations of short trapezoidal pulses are considerably larger than that of Gaussian pulses. The variations of the first type of trapezoidal pulses decreased with width and were close to those of Gaussian pulses, as shown by the green curves with triangles in the figure. The amplitude variations of the second type of trapezoidal pulses decreased slowly with width and remained significantly larger than that of Gaussian pulses for small and moderate pulse widths. The form of rectangular-like trapezoidal pulses is more affected by variations of the charge collection time because their peaking time is closer to the range of the charge collection time. Thus, the amplitude variations of short true Gaussian pulses can be neglected when compared with the energy resolution of the KETEK spectrometer. The variations of the short trapezoidal pulses are much larger and therefore the energy resolution of the spectrometers operating with short pulses at high count rates can be reduced. The amplitude variations might be higher in SDDs with a larger area. The variations in charge collection time in an AMPTEK XR100-SDD \cite{AMPTEKXR100}, estimated in a similar way \cite{Gatti1987}, amounted to $\approx20$~ns. This value corresponds to direct measurements carried out in \cite{Prigozhin2012}. Normalized amplitude variations calculated for this detector at integration times of 100~ns and 300~ns are shown in \figref{fig:AmplVariationsAMPTEK}. Larger variations of the charge collection time in the XR100 SDD result in higher variations of the shaped pulse amplitudes. At short pulse widths, the variations exceed the energy resolution of the detection system. True Gaussian shapers provide the smallest variations at high count rates. \section{Discussion} \label{sec:Discussion} The symmetrical form of true Gaussian pulses is one of the factors which enable a high count rate when using them. One can try to design a true trapezoidal shaper in a similar way, i.e. with the use of a transmission function found as the ratio of Fourier transforms of the output trapezoidal and input pulses. Such a trapezoidal shaper was developed for scintillator pulses with a short rising time \cite{Redus2014, RedusPrivate}. The spectrum of a symmetrical trapezoidal pulse is much wider than that of a Gaussian pulse of the same width and decays slowly in the high frequency band: $F_{trap}=4/(\omega^2T)~sin(\omega(\tau_p+T)/2)sin(\omega~T/2) \approx 4/(\omega^2T) $. Here, $\tau_p$ is the peaking time of the pulse and T is its flat top duration. The amplification of the shaper at high frequencies $f\ge~1/\tau_p$ must be high enough to form a true trapezoidal pulse of a width less than the rise time of the input pulse. In this case, the shaper transmission has little slope or is not even damped at high frequencies, resulting high output noise. The same conclusion holds true for cusp-like and pseudo-Gaussian shapers. They use parabolic approximations of the output pulse form, in which the Fourier spectra decays with $\omega^{-3}$. The large noise factors of standard shapers with output pulses of true forms limit their application in spectrometers at high count rates. \section{Conclusion} A new digital true Gaussian shaper for the detection of pulses at high count rates has been proposed in this paper. The shaper was analysed for the impulse response of an AXAS-D system with a H7 VITUS SDD, made by KETEK GmbH, in terms of noise factor, dead and resolving times, energy resolution and output count rate of the measurements. The analysis employed numerical models based on calibrations and measured characteristics of the KETEK spectrometer. The main advantages of the proposed true Gaussian shaper compared with standard shapers are as follows: 1. the width of symmetrical output pulses of the Gaussian shapers can be reduced well below the rise time of the input pulse to overcome the minimal width of output of standards shapers, 2. short Gaussian pulses allow detection and accurate measurements of amplitudes of strongly overlapped input pulses, 3. true Gaussian shapers provide a significant increase of the output count rate of SDD spectrometers while maintaining their energy resolution. 4. Amplitude variations of Gaussian pulses with charge collection time of electron clouds by the SDD anode are less or similar to those provided by standard shapers. \section*{Acknowledgments} This work was supported by a grant of the Russian Science Foundation (17-12-01110) and by the Ioffe Institute.

f0ea237d46e187851f79f46068c1c02acce2cd07

\section{Introduction} \label{sec:intro} It is a challenging problem to reconstruct the magnetic field and plasma together in the solar atmosphere. Usually in the corona, magnetic field is expected to dominate over plasma because of the low plasma $\beta$ \citep{g01}. The magnetic field is then modeled by the so-called force-free assumption \citep{ws12}. However, in the photosphere and lower chromosphere, there always exists high $\beta$ regions where the pressure gradient and gravity are also important. Still under the assumption of stationary state, the more general extrapolation which takes into account the non-magnetic-force is called the magneto-hydro-static (MHS) extrapolation. While sophisticated approaches of force-free extrapolation have been developped in the past few decades: \cite{s64}, \cite{s67} for potential field; \cite{ch77}, \cite{s78} for linear force-free field (LFFF); and \cite{s81}, \cite{w90}, \cite{wsr00}, \cite{ys00}, \cite{rak02}, \cite{wn03}, \cite{w04a}, \cite{w04b}, \cite{vkk05}, \cite{aba06}, \cite{wis06}, \cite{hw08}, \cite{jf12}, \cite{imp14}, \cite{gxk16} for nonlinear force-free field (NLFFF), much less papers addressed the MHS extrapolation. In the generic case, the MHS equations are not soluble analytically. However, a special class of MHS equilibria can be obtained by the following ansatz: \begin{equation} \nabla\times \mathbf{B}=\alpha_{0}\mathbf{B}+f(z)\nabla B_{z}\times \mathbf{e_{z}}\label{eq:linmhs_current}, \end{equation} where the first term is a field line parallel current and the second term defines the current perpendicular to the gravity \citep{l85}. For this special form of the current, the MHS equations can be solved by the separation of variables \citep{l85,l91,l92,nr99} or a Fast-Fourier Transform \citep{a81}. This is the so-called linear MHS model, which reduces to a LFFF for $f(z)=0$. \cite{ads98} modeled the magnetic field using MHS equations derived by \cite{l92}, taking into account the pressure and gravity. The parameters in the linear MHS model, $\alpha$ and $a$, are constant in the entire computational region and a scale-height of 2 Mm was used. The authors pointed out main properties of magneto-static configurations computed with this model, namely that the field aligned part of the current density contains two parts, the $\alpha {\bf B}$ term and the horizontal currents. Different from linear force-free fields, where the current density is strictly parallel to the magnetic field, this property adds some nonlinearity regarding the field aligned currents. Another interesting property pointed out by \cite{ads98} is that (using their Eq. (4)) the changes in plasma pressure (compared to the background atmosphere model) is as stronger as more vertical the field is. This property is consistent with the observation of a reduced plasma pressure in strong field regions like sunspots. We would like to point out that the linear MHS model requires global constants $a$ and $\alpha$ and this excludes strong localized concentration of electric current and Lorentz forces. While \cite{ads98} used the linear MHS configuration to model solar structures, the main emphasis of our paper is to develop and test a nonlinear MHS code, which does not have such limitations. As we are not aware of exact nonlinear MHS solutions in 3D, we, however, test the code by comparison with a linear MHS model. For the general MHS equations, the computationally expensive numerical codes are required. Different numerical methods have been developed for this aim, e.g., \cite{gr58} solved a system of linear equations iteratively to approach the solution of nonlinear MHS equations. An advantage of the Grad-Rubin approach is that underlying mathematical problem is well posed. A disadvantage of providing certain boundary conditions (currents or $\alpha$ in NLFFF, additional pressure in MHS) is that in reality the boundary data on both footpoints are not consistent due to measurement errors. This can lead to large differences between the solutions computed from positive and negative footpoints as shown in \cite{sdm08} for the force-free approach. The Grad-Rubin method has been extended to solve the MHS equations with gravity by \cite{gw13} and \cite{gbb16}. \cite{wn06} developed an optimization principle for computing the magnetic field and plasma pressure consistently without considering gravity. The method was tested by application to a semi-analytic MHS solution which is axisymmetric. \cite{zwd13} modeled the MHS equilibria through magnetohydrodynamic (MHD) relaxation method. The method was tested by a Sun-like numerical model and H$\alpha$ fibril observation in the chromosphere \citep{zwd16} On the other hand, high spatial resolution data make MHS extrapolations necessary to extrapolate photospheric vector magnetograms upwards and resolve thereby the physics of the upper photosphere and chromosphere. Compared with the height, say about 2 Mm, of the non-force-free layer, the common spatial resolution of vector magnetograms (e.g., 700km for SDO/HMI) was too low to resolve it. The modeling of this thin layer, however, becomes possible with the unprecedented small pixel size of 40 km from the \textsc{Sunrise}/IMaX observation. \cite{wnn15}/\cite{wnn17} applied a linear MHS model to a quiet/an active region using the line-of-sight (LOS)/vector magnetogram from IMaX observation during its first/second flight in 2009/2013. In this paper we present a more general optimization model, where the magnetic field, plasma pressure and density are computed self-consistently. The photospheric boundary (vector magnetogram, typical pressure and density in the quiet region) is the only input of this model, which makes it applicable to real data. The basic equations are described in Section \ref{sec:BasicEquation}, a method used to update the pressure at the bottom boundary during optimization is presented in Section \ref{sec:BotPressure}, then the algorithm is presented in Section \ref{sec:algorithm}, an analytic linear MHS solution for testing the code is described in Section \ref{sec:reference_solution}, the results are presented in Section \ref{sec:results}. In Section \ref{sec:discussion} we present our conclusions and discuss some questions in the future application. \section{Basic Equations} \label{sec:BasicEquation} The MHS equations are given by: \begin{eqnarray} \frac{1}{\mu_{0}}(\nabla \times \mathbf{B})\times \mathbf{B}-\nabla p - \rho g\mathbf{\hat{z}} & = & 0, \label{eq:force_balance}\\ \nabla \cdot \mathbf{B} & = & 0, \label{divergence_free} \end{eqnarray} where $\mathbf{B}$, $p$, $\rho$, $g$ and $\mu_{0}$ are magnetic field, plasma pressure, plasma density, gravitational acceleration and vacuum permeability, respectively. As the gravitational acceleration changes only $0.57\%$ (from 272.407 to 273.975 $m/s^{2}$) in the 2 Mm non-force-free layer, $g$ is treated as a constant. We define the functional \begin{equation} L(\mathbf{B},p,\rho)=\int_{V}\omega_{a}B^{2}\Omega_{a}^{2}+\omega_{b}B^{2}\Omega_{b}^{2}dV,\label{eq:L} \end{equation} with \begin{eqnarray} \mathbf{\Omega_{a}} &=& B^{-2}\left[\frac{1}{\mu_{0}}(\nabla \times \mathbf{B})\times \mathbf{B}-\nabla p - \rho g \mathbf{\hat{z}}\right], \\ \mathbf{\Omega_{b}} &=& B^{-2}[(\nabla \cdot \mathbf{B})\mathbf{B}], \end{eqnarray} where $\omega_{a}$ and $\omega_{b}$ are the weighting functions with cos-profile. The problem of solving Eq. (\ref{eq:force_balance}-\ref{divergence_free}) is replaced with following minimization problem: \begin{eqnarray} minimize \quad &L&(\mathbf{B},p,\rho) \\ subject \ to: p &>& 0 \label{eq:p_constraint} \\ \rho &>& 0 \label{eq:d_constraint} \end{eqnarray} The constraints can be eliminated by using the variable transformation to $p$ and $\rho$ \begin{eqnarray} p=Q^{2}, \label{eq:Q}\\ \rho=\frac{R^{2}}{gH_{s}}, \label{eq:R} \end{eqnarray} where pressure scale-height $H_{s}$ is a constant. $H_{s}$ and $g$ in Eq. (\ref{eq:R}) are used to make $R$ has the dimension of $B$ and $Q$. Then the above constrained optimization problem is changed to an unconstrained one: \begin{equation} minimize \quad L(\mathbf{B},Q,R). \end{equation} According to \cite{wsr00}, \cite{w04a} and \cite{wn06}, the optimization can be simply extended to solve the MHS equations with gravity. Taking the functional derivative of the functional (\ref{eq:L}) with respect to an iteration parameter $t$ leads to: \begin{equation} \frac{1}{2}\frac{dL}{dt}=-\int_{V}\left(\frac{\partial \mathbf{B}}{\partial t}\cdot \mathbf{\tilde{F}}+\frac{\partial Q}{\partial t}F_{1}+\frac{\partial R}{\partial t}F_{2}\right)dV-\oint_{S}\left(\frac{\partial \mathbf{B}}{\partial t}\cdot \mathbf{\tilde{G}}+\frac{\partial Q}{\partial t}G_{1}\right)dS,\label{eq:dL} \end{equation} where $\mathbf{\tilde{F}},\ F_{1},\ F_{2},\ \mathbf{\tilde{G}}$ and $\mathbf{G_{1}}$ are defined in Appendix \ref{sec:variables}. If $\mathbf{B},\ p,\ \rho$ are fixed on the boundary of the computation box, $L$ can be minimized by solving the equations \begin{equation} \frac{\partial \mathbf{B}}{\partial t}=\mu_{1}\mathbf{\tilde{F}},\quad \frac{\partial Q}{\partial t}=\mu_{2}F_{1},\quad \frac{\partial R}{\partial t}=\mu_{3}F_{2}\label{eq:ite_bpd} \end{equation} iteratively. In the paper, $\mu_{1}=\mu_{2}=\mu_{3}=1$. \section{Consistent evolution of Pressure on the boundary} \label{sec:BotPressure} Because of the observational limitation, only the vector magnetogram on the photosphere can be used as boundary input. The weighting functions diminish the effect of the unknown top and lateral boundaries \citep{w04a}, but different from NLFFF extrapolation we need additional information regarding the plasma pressure and density on the bottom boundary. Because the gravitational force is only in vertical direction, we derive the following simplified MHS equations on the 2D photospheric layer: \begin{equation} \nabla_{ph} p=\mathbf{f}_{ph} \label{eq:FB_ph}, \end{equation} where $\mathbf{f}_{ph}$ is the 2D Lorentz force on the photosphere and $\nabla_{ph}=\mathbf{\hat{x}}\partial_{x}+\mathbf{\hat{y}}\partial_{y}$. Taken another divergence operation on both sides of Eq. (\ref{eq:FB_ph}) results in the following Poisson's equation: \begin{equation} \Delta_{ph} p=\nabla\cdot\mathbf{f}_{ph}\label{eq:poisson}, \end{equation} where $\Delta_{ph}=\partial_{x}^{2}+\partial_{y}^{2}$ is the 2D Laplacian. If we knowing the Lorentz force, the pressure is determined when the pressure on the 4 edges of the bottom plane (xy-plane) is assigned; the typical pressure of quiet region can be used as the pressure on the edges if the computation box is much larger than the active region. Although we do not know the Lorentz force of the MHS equilibria to be extrapolated, we can compute it at any step during the optimization. Then an iterative approach can be designed to update the pressure on the photosphere consistently with magnetic field (detailed description of the algorithm in Section \ref{sec:algorithm}). From another perspective, any vector field can be decomposed into curl-free and divergence-free components (Helmholtz decomposition). For the Lorentz force on the photosphere, however, it is curl-free if the stationary state is maintained. But the Lorentz force has divergence-free component during the optimization. Taking additional divergence operation to Eq. (\ref{eq:FB_ph}) extracts the curl-free component of the Lorentz force. The curl-free component of the Lorentz force determines the pressure. So far, we used information regarding the Lorentz force during optimization to update the bottom pressure. It looks like that the density $\rho$ can be easily computed from force balance in z-direction: $\rho=\left(\frac{1}{\mu_{0}}(\nabla\times \mathbf{B})\times \mathbf{B}-\nabla p\right)_{z}/g$. However, the test shows no improvement of the results. We will further study this issue in the future. In this paper, the bottom density is uniform and fixed during optimization. \section{Numerical Implementation} \label{sec:algorithm} We have developed a code to compute 3D-MHS equilibria, based on the previous optimization code \citep{w04a,wn06}. \begin{enumerate} \item Calculate a NLFFF by using vector magnetogram. \item Insert an isothermal gravity stratified atmosphere. The $p$ and $\rho$ on the photosphere are uniformly distributed. \item Iterate for $\mathbf{B},\ p$ and $\rho$ by Eq. (\ref{eq:ite_bpd}). This step repeated until L reaches its minimum\label{item:ite_bpd}. \item Update $p$ on the photosphere by solving Poisson's Eq. (\ref{eq:poisson}) with Lorentz force computed from $\mathbf{J}\times\mathbf{B}$, and repeat from step \ref{item:ite_bpd}. If $p$ is not changed for the giving tolerance, iteration stops and output $\mathbf{B},\ p$ and $\rho$. \end{enumerate} \section{Reference MHS solution} \label{sec:reference_solution} \cite{l85,l91} presented a class of analytic solutions of the 3D static, magnetized atmospheres. The solutions are characterized by two parts of electric currents as described in Eq. (\ref{eq:linmhs_current}), namely the component parallel to the magnetic field and the component perpendicular to the gravitational field. Assume that $f(z)$ has the form \begin{equation} f(z)=a\exp^{-\kappa z}, \end{equation} where $a$ and $\kappa$ control the magnitude and effective height of Lorentz force. Using Fourier transforming $\mathbf{B}$ with respect to $x$ and $y$, Eq. (\ref{eq:linmhs_current}) can be solved by the separation of variables with LOS magnetogram as bottom boundary. With this magnetic structure, the pressure and density have the following distribution: \begin{eqnarray} p&=&p_{0}(z)-\frac{1}{2\mu_{0}}f(z)B_{z}^{2},\\ \rho&=&-\frac{1}{g}\frac{dp_{0}}{dz}+\frac{1}{\mu_{0}g}\left[\frac{df}{dz}\frac{B_{z}^{2}}{2}+f(\mathbf{B}\cdot\nabla)B_{z}\right]. \end{eqnarray} It is apparent from the above two equations that the plane-parallel hydro-static atmosphere ($\rho_{0}=-\frac{1}{g}\frac{dp_{0}}{dz}$, $p_{0}$) is disturbed by the magnetic field. The pressure is weak in strong $B_{z}$ region with $f>0$. To determine all the variables in the computation box, we use the following parameter set: \cite{ll90} LOS magnetogram labeled $n=m=1$, $\Phi=\frac{\pi}{4}$ and $l=0.3$ in their notation; field line parallel linear current with $\alpha=-3.0$ and non-magnetic force with $a=0.5$; $\kappa=0.02$ means the effective height of Lorentz force is 50 grids; the background atmosphere with $\rho_{0(z=0)}=9.0\times 10^{-4} kg/m^{3}$ and temperature $T_{0(z)}=$ 6000/5500/7840K at the height 0/0.5/1.28Mm (use linear interpolation to derive inter point temperature). For more sophisticated modelling of the vertical temperature profile see \cite{val81}. Then a linear MHS solution is generated in the Cartesian box (unit: Mm) \begin{equation} V=\left\lbrace(x,y,z)|-1.6\le x\le 1.6,\ -1.6\le y\le 1.6,\ 0\le z\le 1.28\right\rbrace. \end{equation} All above parameters are chosen to mimic a small magnetic pole on the sun. The grid points $80\times80\times32$ are used to resolve this reference model. The grid size is 40km which is the same with \textsc{Sunrise}/IMaX data. \section{Results} \label{sec:results} We use the figures of merit introduced by \cite{sdm06} to quantify the difference between the reconstructed magnetic field $\mathbf{B}$ and the reference one $\mathbf{b}$, and supplement these with C-value between field lines, linear Pearson correlation coefficients both for the 3D and LOS integration (along the $z$ axis) of plasma pressure ($corr3D.p,\ corr2D.p$) and density ($corr3D.\rho,\ corr2D.\rho$). They are defined as: \begin{enumerate} \item[$\cdot$] vector correlation \begin{equation} C_{vec}=\displaystyle\sum_{i}\mathbf{B}_{i}\cdot \mathbf{b}_{i}/\left(\displaystyle\sum_{i}|\mathbf{B}_{i}|^2\displaystyle\sum_{i}|\mathbf{b}_{i}|^2\right)^{\frac{1}{2}}, \end{equation} \item[$\cdot$] Cauchy-Schwarz inequality \begin{equation} C_{CS}=\frac{1}{N}\displaystyle\sum_{i}\frac{\mathbf{B}_{i}\cdot \mathbf{b}_{i}}{|\mathbf{B}_{i}||\mathbf{b}_{i}|}, \end{equation} \item[$\cdot$] normalized vector error \begin{equation} E_{N}=\displaystyle\sum_{i}|\mathbf{B}_{i}-\mathbf{b}_{i}|/\displaystyle\sum_{i}|\mathbf{B}_{i}|, \end{equation} \item[$\cdot$] mean vector error \begin{equation} E_{M}=\frac{1}{N}\displaystyle\sum_{i}\frac{|\mathbf{B}_{i}||\mathbf{b}_{i}|}{\mathbf{B}_{i}}, \end{equation} where N is the number of grid points in the computation box. \item[$\cdot$] C-value \begin{equation} C=\frac{1}{l^{2}}\int_{0}^{l}\sqrt{\left(\bf{r}_{ref}(\tau)-\bf{r}_{extrapol}(\tau)\right)^{2}}d\tau, \end{equation} where C is a measure of how well the reference and extrapolated field lines agree. It is the integration along the field line (total length $l$) from the geometrical length $\tau=0$ to $\tau=l$. The C-value has been used by \cite{wn02} and \cite{wls05} to compare the magnetic field lines with observed loops. \item[$\cdot$] Pearson's correlation coefficient \begin{equation} corr.=\frac{cov(Q,q)}{\sigma_{Q}\sigma_{q}}, \end{equation} where $cov$ is the covariance, $\sigma$ is the standard deviation, Q and q are the extrapolated solution and reference model respectively. \end{enumerate} \subsection{Test I: all boundary conditions provided} \begin{figure \plotone{allbnd1.pdf} \caption{Magnetic field for test I with different models. The field lines start from the same seeds which are uniformly distributed in the bottom plane.\label{fig:allbnd1}} \end{figure} \begin{figure \plotone{linecompareallbndR2.pdf} \caption{Field lines of reference model (white), NLFFF extrapolation (blue) and MHS extrapolation (yellow) with the same start points on the bottom boundary. Notice that the white and yellow lines alsmost coincide with each other. \label{fig:allcompare}} \end{figure} Figure \ref{fig:allbnd1} shows the overall magnetic field line patterns from different models for test I using all boundary conditions. From Fig. \ref{fig:allcompare}, we clearly see that the MHS extrapolation produce better field lines than the NLFFF extrapolation. See also Table \ref{tab:linecompareallbnd} of the C-values of the individual field line. The mean C-values of NLFFF and MHS extrapolated lines are 0.162 and 0.016, with corresponding standard deviations of 0.128 and 0.026, respectively. The above comparisons show how the Lorentz force affects the field line patterns. The ordering of the figures of merit (see Table \ref{tab:allbnd}) agrees with the conclusion from the above visual quality. Fig. \ref{fig:allbnd2} shows LOS integration of plasma pressure and density along the z-axis. We also notice that the MHS extrapolation need 5 times more steps and 6 times more CPU time than NLFFF extrapolation. \begin{deluxetable}{cccccccccc} \tablecaption{Model resuls for test I in which all boundary conditions are specified.}\label{tab:allbnd} \tablecolumns{10} \tablewidth{0pt} \tablehead{\colhead{Model} & \colhead{$C_{vec}$} & \colhead{$C_{cs}$} & \colhead{$1-E_{m}$} & \colhead{$1-E_{n}$} & \colhead{$corr2D.p$} & \colhead{$corr2D.\rho$} & \colhead{$corr3D.p$} & \colhead{$corr3D.\rho$} & \colhead{step ($\times10^{3}$)}} \startdata Potential & 0.8911 & 0.7841 & 0.4952 & 0.4080 & / & / & / & / & / \\ NLFFF & 0.9875 & 0.9747 & 0.8531 & 0.8405 & / & / & / & / & 110\\ MHS & 0.9979 & 0.9911 & 0.9492 & 0.9237 & 0.9988 & 0.9993 & 1.0000 & 0.9999 & 590\\ \enddata \end{deluxetable} \begin{figure \plotone{allbnd2R2.pdf} \caption{LOS integration of the plasma pressure (top) and density (bottom) in the central field of view $x,y\in [-1.2,\ 1.2]$ (unit: Mm). Left/right panels correspond to the reference/reconstructed solution. \label{fig:allbnd2}} \end{figure} \begin{deluxetable}{|c|cc|c|cc|c|cccc} \tablecaption{C-values of individual field line of test I. The footpoints of 30 lines randomly distributed in the negative region ($B_{z}<0$).\label{tab:linecompareallbnd}} \tablecolumns{9} \tablewidth{0pt} \tablehead{\colhead{No.} & \colhead{$C_{n\!l\!f\!\!f\!\!f}$} & \colhead{$C_{m\!h\!s}$} & \colhead{No.} & \colhead{$C_{n\!l\!f\!\!f\!\!f}$} & \colhead{$C_{m\!h\!s}$} & \colhead{No.} & \colhead{$C_{n\!l\!f\!\!f\!\!f}$} & \colhead{$C_{m\!h\!s}$} & \colhead{$<\!C_{n\!l\!f\!\!f\!\!f}\!>\!\pm\sigma $} & \colhead{$<\!C_{m\!h\!s}\!>\!\pm\sigma$}} \startdata 1 & 0.059 & 0.007 & 11 & 0.277 & 0.009 & 21 & 0.003 & 0.003 & 0.162$\pm$0.128 & 0.016$\pm$0.026\\ 2 & 0.105 & 0.011 & 12 & 0.103 & 0.012 & 22 & 0.025 & 0.003 \\ 3 & 0.131 & 0.016 & 13 & 0.191 & 0.009 & 23 & 0.189 & 0.024 \\ 4 & 0.104 & 0.017 & 14 & 0.164 & 0.006 & 24 & 0.393 & 0.076 \\ 5 & 0.055 & 0.019 & 15 & 0.248 & 0.010 & 25 & 0.084 & 0.005 \\ 6 & 0.013 & 0.001 & 16 & 0.234 & 0.013 & 26 & 0.035 & 0.005 \\ 7 & 0.214 & 0.006 & 17 & 0.133 & 0.012 & 27 & 0.026 & 0.006 \\ 8 & 0.255 & 0.012 & 18 & 0.520 & 0.014 & 28 & 0.031 & 0.007 \\ 9 & 0.250 & 0.012 & 19 & 0.221 & 0.009 & 29 & 0.392 & 0.138 \\ 10 & 0.262 & 0.012 & 20 & 0.025 & 0.003 & 30 & 0.118 & 0.015 \\ \enddata \end{deluxetable} \subsection{Test II: bottom vector magnetogram with weighted boundary layer} In test II, we only use the bottom vector magnetogram as the boundary input, which mimics the real situation. In this test, the totally $80\times80\times32$ grids of the box consist of the inner region ($64\times64\times24$) and layer ($nd=8$ grids) at the lateral and top boundaries with cos-profile weighting functions \citep{w04a}. To see if the pressure update on the photosphere improves the result, we perform two test runs for MHS extrapolation. The difference between them is: in one of the runs, the pressure is uniform and fixed on the photosphere during optimization; while in the other run, we update the pressure using the method mentioned in Section \ref{sec:BotPressure}. Figure \ref{fig:botbnd1} shows the overall magnetic field line patterns from different models for test II using bottom magnetogram. From Fig. \ref{fig:botcompare}, we can see that the MHS extrapolation produce better field lines than the NLFFF extrapolation. See also Table \ref{tab:linecomparebotbnd} of the C-values of the individual field line. The mean C-values of NLFFF and MHS extrapolated lines are 0.103 and 0.059, with corresponding standard deviations of 0.061 and 0.049, respectively. Although the field line geometry difference between the two MHS extrapolations is not large, the integration of plasma pressure and density along z-axis (see Fig. \ref{fig:botbnd2}) shows rather large differences. Updating the bottom pressure significantly improves the pressure and density results. Table \ref{tab:botbnd} shows that the ordering of the figures of merit agrees with the previous visual judgment. We notice that, for pressure and density, the correlation of 2D integration is a better index than correlation of 3D distribution. Because in 3D, the almost gravity stratified atmosphere ensures the high correlation between the reference model and reconstructed solution. It is also good to see the improvement in density result even we do not use the density information on the bottom boundary. The process of MHS extrapolation is optimizing the magnetic field and plasma. Fig. \ref{fig:PreDen} shows how far the final plasma deviate from the initially gravity stratified atmosphere. We can see the final solution is close to the gravity stratified atmosphere at the low height. When $z$ increases, the difference becomes larger. To check if the MHS equations are fulfilled in the extrapolated solution of test II, the field line components of $-\nabla p$ and $\rho \bf g$ are calculated. Defining \begin{equation} Ratio=\frac{{\bf\hat{B}}\cdot(-\nabla p+\rho\bf{g})}{|\nabla p|+|\rho {\bf g}|}, \end{equation} where ${\bf\hat{B}}={\bf B}/B$ is the unit vector along the magnetic field line. For an MHS equilibrium, $Ratio=0$ at anywhere. Here we compute $Ratio$ along four field lines (the same lines in Fig. \ref{fig:botcompare}). For totally 272 points, the mean $Ratio$ is 0.86\textperthousand\ with standard deviation of 0.96\textperthousand\ (see Fig. \ref{fig:Ratio}). The extremely small $Ratio$ means the recovered plasma satisfy the field line component of the MHS equation at high accuracy. \begin{figure \plotone{botbnd12nd.pdf} \caption{Magnetic field in the inner region (smaller box) for test II with different models. \label{fig:botbnd1}} \end{figure} \begin{figure \plotone{linecomparebotbndR2.pdf} \caption{The same with Fig. \ref{fig:allcompare} except red/yellow lines represent the magnetic field from MHS extrapolation with uniform/update bottom pressure. \label{fig:botcompare}} \end{figure} \begin{figure \plotone{botbnd2R2.pdf} \caption{LOS integration of the plasma pressure (top) and density (bottom) in the central field of view $x,y\in [-1.2,\ 1.2]$ (unit: Mm). Left panels are the reference results. Middle/Right panels correspond to the reconstructed solutions with updating/uniform bottom pressure. \label{fig:botbnd2}} \end{figure} \begin{figure \plotone{PreDenR2.pdf} \caption{Plasma pressure (top) and density (bottom) change along vertical axis of test II. Top/bottom panels show (from left to right): plane average pressure/density along z-axis, pressure/density along $(x,y,z)=(-0.8,\ 0,\ *)$ and $(x,y,z)=(0.8,\ 0,\ *)$. Black and blue lines correspond to the gravity stratified and final solution of the atmosphere, respectively. \label{fig:PreDen}} \end{figure} \begin{figure \plotone{Ratio.pdf} \caption{$Ratio$ along 4 field lines (the same lines in Fig. \ref{fig:botcompare}) with totally 272 points.} \label{fig:Ratio} \end{figure} \begin{deluxetable*}{cccccccccc} \tablecaption{Results of inner region for test II with only bottom vector magnetogram specified. \label{tab:botbnd}} \tablecolumns{10} \tablewidth{0pt} \tablehead{\colhead{Model} & \colhead{$C_{vec}$} & \colhead{$C_{cs}$} & \colhead{$1-E_{m}$} & \colhead{$1-E_{n}$} & \colhead{$corr2D.p$} & \colhead{$corr2D.\rho$} & \colhead{$corr3D.p$} & \colhead{$corr3D.\rho$} & \colhead{step ($\times10^{3}$)}} \startdata Initial state\tablenotemark{a} & 0.9880 & 0.9679 & 0.8052 & 0.7538 & 0.0000 & 0.0000 & 0.9979 & 0.9978 & 6\\ Uniform-$p$\tablenotemark{b} & 0.9917 & 0.9724 & 0.8505 & 0.7916 & 0.9694 & 0.6695 & 0.9994 & 0.9957 & 153\\ Update-$p$\tablenotemark{c} & 0.9921 & 0.9728 & 0.8596 & 0.7977 & 0.9831 & 0.9683 & 0.9998 & 0.9982 & 191\\ \enddata \tablenotetext{a}{Initial state consists of a NLFFF and an isothermal gravity stratified atmosphere.} \tablenotetext{b}{Uniform bottom pressure during optimization.} \tablenotetext{c}{Update bottom pressure during optimization.} \end{deluxetable*} \begin{deluxetable}{|c|cc|c|cc|c|cccc} \tablecaption{The same with Table \ref{tab:linecompareallbnd} of test II. \label{tab:linecomparebotbnd}} \tablecolumns{9} \tablewidth{0pt} \tablehead{\colhead{No.} & \colhead{$C_{n\!l\!f\!\!f\!\!f}$} & \colhead{$C_{m\!h\!s}$\tablenotemark{a}} & \colhead{No.} & \colhead{$C_{n\!l\!f\!\!f\!\!f}$} & \colhead{$C_{m\!h\!s}$} & \colhead{No.} & \colhead{$C_{n\!l\!f\!\!f\!\!f}$} & \colhead{$C_{m\!h\!s}$} & \colhead{$<\!C_{n\!l\!f\!\!f\!\!f}\!>\!\pm\sigma $} & \colhead{$<\!C_{m\!h\!s}\!>\!\pm\sigma$}} \startdata 1 & 0.058 & 0.011 & 11 & 0.176 & 0.054 & 21 & 0.039 & 0.036 & 0.103$\pm$0.061 & 0.059$\pm$0.049\\ 2 & 0.106 & 0.025 & 12 & 0.100 & 0.018 & 22 & 0.048 & 0.047 \\ 3 & 0.133 & 0.029 & 13 & 0.092 & 0.043 & 23 & 0.195 & 0.193 \\ 4 & 0.116 & 0.025 & 14 & 0.039 & 0.051 & 24 & 0.092 & 0.039 \\ 5 & 0.057 & 0.015 & 15 & 0.132 & 0.119 & 25 & 0.031 & 0.050 \\ 6 & 0.014 & 0.010 & 16 & 0.088 & 0.065 & 26 & 0.024 & 0.017 \\ 7 & 0.173 & 0.058 & 17 & 0.105 & 0.044 & 27 & 0.035 & 0.019 \\ 8 & 0.181 & 0.080 & 18 & 0.193 & 0.044 & 28 & 0.058 & 0.051 \\ 9 & 0.148 & 0.072 & 19 & 0.033 & 0.040 & 29 & 0.155 & 0.151 \\ 10 & 0.170 & 0.076 & 20 & 0.067 & 0.073 & 30 & 0.222 & 0.207 \\ \enddata \tablenotetext{a}{MHS extrapolation with updating bottom pressure.} \end{deluxetable} \subsubsection{Influence of initial conditions} Here we investigate in the dependence of the result on the choice of initial condition. The two initial magnetic field we use are (1) potential field \citep{s78} and (2) NLFFF produced by optimization code \citep{w04a}, while the two initial atmospheres are (1) isothermal atmosphere and (2) more realistic 1D model described in section \ref{sec:reference_solution}. This results in 4 combinations. The choice of the initial magnetic field configuration has a significant influence on the resulting magnetic field and plasma equilibrium. Similar conclusions were found in previous studies for NLFFF \citep{w04a,sdm06}: a starting state which is near to the true solution leads to a better result. That means we better use a multigrid approach to give a better starting state, similar as used as a standard in NLFFF extrapolation. However, notice that the initial potential field results in a somewhat more accurate density solution. Unlike the magnetic field, the choice of the initial atmosphere has negligible influence on the results. Either isothermal atmosphere or sun-like atmosphere gives almost the same solution in this test. \begin{deluxetable*}{cccccccc} \tablecaption{Results of inner region with different initial conditions. \label{tab:inicondition}} \tablecolumns{7} \tablewidth{0pt} \tablehead{ \colhead{case} & \colhead{$C_{vec}$} & \colhead{$C_{cs}$} & \colhead{$1-E_{m}$} & \colhead{$1-E_{n}$} & \colhead{$corr2D.p$} & \colhead{$corr2D.\rho$} & } \startdata case I\tablenotemark{a} & 0.9921 & 0.9728 & 0.8596 & 0.7977 & 0.9831 & 0.9683\\ case II\tablenotemark{b} & 0.9920 & 0.9727 & 0.8594 & 0.7975 & 0.9832 & 0.9695\\ case III\tablenotemark{c} & 0.9166 & 0.8648 & 0.5648 & 0.5038 & 0.9655 & 0.9749\\ case IV\tablenotemark{d} & 0.9165 & 0.8648 & 0.5647 & 0.5038 & 0.9653 & 0.9765\\ \enddata \tablenotetext{a}{Initial state consists of NLFFF and isothermal atmosphere.} \tablenotetext{b}{Initial state consists of NLFFF and sun-like atmosphere.} \tablenotetext{c}{Initial state consists of potential field and sun-like atmosphere.} \tablenotetext{d}{Initial state consists of potential field and isothermal atmosphere.} \end{deluxetable*} \subsubsection{Influence of noise} Until now, we input the magnetic field on the bottom boundary as it is known exactly. However, this is not the case when the real vector magnetogram is used. In this subsection, we study the influence of the noise of the bottom magnetic field by adding some random noise ($2\%$ in $B_{z}$, $nl$ in $B_{x}$ and $B_{y}$) to the magnetogram. $nl=5\%, 10\%, 15\%, 20\%$ are noise levels of transverse field for different test runs. The same cos-profile weighting functions and boundary layer $nd=8$ are used in these test runs. \begin{deluxetable*}{cccccccc} \tablecaption{Results with different noise level $nl$. \label{tab:noise}} \tablecolumns{7} \tablewidth{0pt} \tablehead{ \colhead{noise level} & \colhead{$C_{vec}$} & \colhead{$C_{cs}$} & \colhead{$1-E_{m}$} & \colhead{$1-E_{n}$} & \colhead{$corr2D.p$} & \colhead{$corr2D.\rho$} & } \startdata No noise & 0.9921 & 0.9728 & 0.8596 & 0.7977 & 0.9831 & 0.9683\\ 5\% & 0.9920 & 0.9729 & 0.8576 & 0.7972 & 0.9804 & 0.9535\\ 10\% & 0.9913 & 0.9722 & 0.8514 & 0.7918 & 0.9767 & 0.9236\\ 15\% & 0.9905 & 0.9715 & 0.8453 & 0.7879 & 0.9729 & 0.8851\\ 20\% & 0.9881 & 0.9688 & 0.8316 & 0.7754 & 0.9615 & 0.8223\\ \enddata \end{deluxetable*} Table \ref{tab:noise} shows the results. The random noise of magnetic field is independent of neighboring grids. This leads to high frequent noise of current and Lorentz force on the photosphere, which makes the extrapolation inaccurate. As a result, all metrics are getting worse with increasing noise. \section{Discussion and conclusions} \label{sec:discussion} In this work, we have generalized the optimization method to apply to MHS equilibria. Compare with NLFFF approach, MHS optimization confronts two new challenges: (1) how to ensure positive pressure and density; (2) how to deal with boundary pressure and density. The first problem is actually how to deal with positivity constraint in optimization. This constraint can be eliminated by the variable transformation of Eq. (\ref{eq:Q},\ref{eq:R}). The second problem is more complex because no measurement of plasma pressure and density is available. Some information, however, are included in the data of the vector magnetogram. Based on the assumption of the force balance in the bottom plane, we obtain the Poisson's Eq. (\ref{eq:poisson}) for computing pressure on the photosphere. Then we design an algorithm to update the bottom pressure consistently within the optimization procedure. In test II, we need 18 times update of bottom pressure, and most steps (153K in totally 191K) are in the first round of $L$ minimization. We conclude from above tests: (1) The MHS equilibria are reconstructed at relatively high accuracy by the generalized optimization principle for iterating magnetic field, plasma pressure and mass density simultaneously; (2) update the bottom pressure by using Lorentz force significantly improves the results of MHS extrapolation; (3) the initial choice of magnetic field influences the final results significantly, whereas, MHS extrapolation using a NLFFF model as the initial condition produces much better results than using a potential field. We also test our code with vanishing $f_{z}$ by setting $a=0.0$. The model with $a=0.0$ is a LFFF. As supposed, our code can recover the LFFF at almost the same accuracy with the results obtained by NLFFF approach. Notice that the bottom density is still uniformly distributed in the current extrapolation. We would like to address this issue in future article. In test II, the MHS extrapolation takes about 6.5 CPU hours on a 2.1GHz processor. An application to IMaX vector magnetogram embedded by HMI data (about $2000\times2000$ grids, see \cite{wnn17}) needs large amount of computational resources. For practical application, we can limit calculation to the \textsc{Sunrise}-FOV ($936\times 936$) to reduce the computation. A multigrid approach is likely enable faster convergence with high resolution magnetograms. Furthermore, the MHS model should be restricted to the non-force-free layer (about 2 Mm above the photosphere) to reduce the computation time. In the force-free corona above, computational less expensive NLFFF extrapolations can be used. \acknowledgments We appreciate the very constructive comments from the anonymous referee and the inspiring discussions with Bernd Inhester. This work was supported by DFG-grant WI 3211/4-1.

75ed9091cd698d0d8f7913d18e143c21e02c19bf

\section{Introduction} \label{intro} Control of spin population and its simultaneous non-destructive detection play a crucial role in diverse scientific fields such as atom interferometry \cite{Cronin2009}, precision magnetometry \cite{Brask2015}, atomic clocks \cite{Wynands2005}, quantum simulation \cite{Georgescu2014} and quantum information processing \cite{Monroe2002}. While external magnetic fields and optical pumping can be used to manipulate the spin polarization and population in an atomic system, spin noise spectroscopy (SNS) \cite{Muller2010a, Zapasskii2013, Sinitsyn2016} provides a means of the detection of such spin coherences by probing spontaneous spin fluctuations of the system via an off-resonant laser beam, and is relatively non-perturbative as compared to the traditional absorption spectroscopy. In this work, we show that SNS can measure different spin properties of rubidium (Rb) vapor not only in equilibrium systems, but also in systems driven out of equilibrium. The random fluctuations over space and time are prevalent in a wide variety of physical systems, and they can be a valuable resource for probing the characteristic nature and internal structure of the systems. Examples of such fluctuations or noise include the Brownian motion of pollen grains in water, the Johnson noise due to thermal agitation of the electrons in an electrical conductor \cite{Johnson1927}, the intensity fluctuations in the emission of random lasers \cite{Sharma06,Gomes2016} and the stochastic fluctuations in clonal cellular constituents \cite{Munsky2009, Swaminathan1982}. The optical SNS technique has been developed by Aleksandrov and Zapasskii \cite{Aleksandrov1981} to passively probe intrinsic spin fluctuations or magnetization noise in a thermal ensemble of spins. These spin ensembles can be made of electron spins in atomic systems and spins of electrons or holes in semiconductors and other solid-state materials. A study of SNS in alkali atomic vapor of Rb and potassium (K) in thermodynamic equilibrium was carried out by \cite{Crooker2004}, which indicated that the electronic and nuclear g-factors, isotope abundance ratios, nuclear moments and hyperfine splittings could be measured. The first successful application of SNS to a solid-state system was performed by \cite{Oestreich2005}, for measuring the electron's Lande g-factor and spin relaxation time in a $n$-doped GaAs semiconductor. There has been significant progress in the recent years to extend the applicability of SNS \cite{Mihaila2006, Katsoprinakis2007, Muller2008, Crooker2009, Crooker2010, Muller2010b, Yan2012, Zapasskii2013b, Pershin2013, Berski2013, Dahbashi2014, Poltavtsev2014, Yang2014, Roy2015, Lucivero2016, Sterin2018}. First, we perform the SNS of Rb atoms in thermal equilibrium. We demonstrate accurate measurements of several physical quantities such as electron's g-factor, nuclear g-factor, isotope abundance ratios, and develop precision magnetometry with our thermal Rb vapor in the presence of a static magnetic field perpendicular to probe laser. While prior measurements have been reported \cite{Crooker2004, Mihaila2006} of several of these quantities using SNS technique, we are able to refine some of these estimates especially for isotope abundance ratios and nuclear g-factor. We then apply an optical pumping beam to control relative spin population in the ground state hyperfine levels of Rb atoms. This is realized by an on-resonance pump beam nearly co-propagating with the far-detuned probe beam. The optical pumping drives the system out of equilibrium. We then show that SNS can be used to determine spin imbalance in different ground state hyperfine levels without disturbing the non-equilibrium steady-state of the system. We also show that the spin noise (SN) spectra from the optically pumped atoms have better resolution than typical absorption spectra from the same system. Therefore, the SNS can be used in resolving spectral lines of a non-equilibrium system in the presence of various spectral broadening mechanisms. This paper is organized as follows: We provide a brief theoretical description of SNS in Sec.~\ref{theory}. In Sec.~\ref{setup}, the experimental set-up and measurement methods are presented in detail. In Sec.~\ref{equi}, we describe measurements and results of SNS in Rb vapor in thermal equilibrium. The results of SNS for optically pumped Rb vapor are given in Sec.~\ref{outEq}. The final Sec.~\ref{conc} comprises a conclusion and an outlook. \section{Theory} \label{theory} Let us consider an ensemble of non-interacting spins in thermal equilibrium at some temperature. The equilibrium is achieved through interactions between these spins and a thermal bath surrounding them. The presence of the thermal bath induces fluctuations in spin polarization over time. Nevertheless, the time-averaged value of the spin polarization or magnetization, $\langle M(t)\rangle_{T\rightarrow \infty}$ ($T$ is the total averaging time) along any arbitrary quantization axis is zero for a paramagnetic system. However, the variance of the magnetization is non-zero. Within the optical SNS technique, a linearly polarized laser light on passing through such a paramagnetic sample can passively detect these magnetization fluctuations along the light propagation in its time-resolved Faraday rotation noise \cite{Crooker2004, Zapasskii2013}. Such detection is feasible as the magnetization fluctuations in a paramagnetic sample alter its optical properties which lead to Faraday rotation noise. The probe beam is kept far-detuned (with a detuning $\delta_p$) from any allowed optical transition of the medium to ensure negligible scattering by the medium making SNS a relatively non-invasive technique. The intrinsic fluctuations of the spin polarization reveal the characteristic relaxation times of the system. In the presence of a constant magnetic field, the spontaneous spin polarization precesses at the Larmor frequency about the magnetic field. Assuming a single exponential spin relaxation time $T_2$ and a magnetic field $B_{\bot}$ being orthogonal to the probe laser propagation ($\widehat{z}$), we can obtain the temporal correlation of magnetization along the probe beam from the Bloch equation: \begin{equation} \langle M_z(t)M_z(0)\rangle \propto \cos(\nu_{L} t)e^{-t/T_2},\label{Eqn:Lorentz_SNS} \end{equation} where the Larmor frequency $\nu_{L}=g_F\mu_BB_{\bot}/h$, $g_F$ is the $g$-factor of the hyperfine $F$-levels, $\mu_B$ is the Bohr magneton and $h$ is the Planck's constant. The measured Faraday rotation fluctuation $\langle \theta_F(t)\theta_F(0)\rangle$ ($\theta_F(t)$ is the Faraday rotation angle at time $t$) is a direct probe of the magnetization fluctuation $\langle M_z(t)M_z(0)\rangle$ of the system in thermal equilibrium and its Fourier transform to spectral frequency $\nu$ is the power spectral density $P(\nu)$ of the spin noise. Therefore, \begin{eqnarray} P(\nu>0)&=&\int_{0}^{\infty}dt\: \cos(\nu t)\langle \theta_F(t)\theta_F(0)\rangle \nonumber \\ &\propto& \frac{1/T_2}{(\nu-\nu_L)^2+1/T_2^2}, \label{SNpower} \end{eqnarray} where we have used Eq.~\ref{Eqn:Lorentz_SNS} in the last line. So, the SN power spectrum has a Lorentzian lineshape centred at $\nu_L$ in frequency domain (refer to the peaks in the SN amplitude spectrum $\sqrt{P(\nu>0)}$ in Fig.~\ref{fig:Experimental_set-up}(b) for $^{87}$Rb or $^{85}$Rb) and its full width at half maxima (FWHM) is proportional to $1/T_2$. The energy $E_{F,m_F}$ of hyperfine $F$-levels for alkali atoms in an arbitrary magnetic field $B_{\bot}$ has an exact expression following Breit and Rabi \cite{Mockler1961}, \begin{eqnarray} \label{Eqn:Breit-Rabi} E_{F=I\pm\frac{1}{2},m_F}&=&-\frac{h\Delta_{\text{hf}}}{2(2I+1)} + g_I\mu_B B_{\bot} m_F \nonumber \\ &\pm& \frac{h\Delta_{\text{hf}}}{2} \sqrt{1+\frac{4m_F}{2I+1}x+x^2}, \end{eqnarray} where $h\Delta_{\text{hf}}$, $g_I$ and $m_F$ are the zero-field hyperfine separation between the levels $F=I+\frac{1}{2}$ and $F=I-\frac{1}{2}$, the nuclear $g$-factor and the magnetic quantum number, respectively. Here, $x=(g_J-g_I)\mu_B B_{\bot}/(h\Delta_{\text{hf}})$ where $g_J$ is the Lande $g$-factor and the nuclear spin $I = 3/2$ for $^{87}$Rb. Since, the SNS detects the spin coherences between different Zeeman sub-levels ($\bigtriangleup m_F=\pm 1$), the frequencies of different magnetic resonance peaks have a nonlinear dependence on $B_{\bot}$ \cite{steck87}. The integrated SN power over frequency, $\chi \equiv \int d\nu P(\nu >0 )$, depends on the probe detuning as $\chi \propto \delta_p^{-2}$ \cite{Mihaila2006} and is symmetric about the atomic resonance frequency for a far-detuned probe beam (where $\delta_p \gg \Gamma$, $\Gamma$ being the width of the absorption spectra). However, this integrated SN power $\chi$ becomes asymmetric over $\delta_p$ for $\delta_p \approx 0$ due to the non-vanishing coherences between the ground and excited state hyperfine levels of the atoms \cite{Horn2011}. This asymmetry in $\chi$ is shown later in this paper in Fig.~\ref{fig:SNA_detuning}(b,d). The asymmetry in $\chi$ also depends on the homogeneous and inhomogeneous broadening present in the medium \cite{Horn2011, Zapasskii2013b}. \section{Experimental set-up and methods} \label{setup} \begin{figure} \centering \includegraphics[scale=0.4]{1_a_b} \caption{(a) Schematic of the experimental set-up for measuring spin noise (SN) spectrum. A probe beam along $\hat{z}$ is focused by a plano-convex lens before entering the vapor cell. The transmitted probe beam is sent through a polarimetric set-up comprising of a half-wave plate (HWP) and a polarizing beam splitter (PBS), and then it is collected by a balanced photo-detector which is connected to a spectrum analyzer. A constant magnetic field $B_\bot$ along $\hat{x}$ is applied on the atomic vapor using two magnetic coils in Helmholtz configuration. (b) A typical SN amplitude spectrum at $B_\bot = 6.95$ G and vapor cell temperature of $105^{\circ}$C is shown for a probe detuning $\delta_p=-10.2$ GHz. The stronger and weaker peaks are identified with $^{87}$Rb $(|g_F| = 1/2)$ and $^{85}$Rb $(|g_F| = 1/3)$ respectively.} \label{fig:Experimental_set-up} \end{figure} The schematic of the experimental set-up is shown in Fig.~\ref{fig:Experimental_set-up}(a). A linearly polarized probe laser beam with tunable frequency $\nu_p$ is sent through a 20 mm long glass cell containing enriched $^{87}$Rb vapor. This probe beam is derived from a grating stabilized external cavity diode laser in Littrow configuration having an instantaneous linewidth below 1 MHz. The probe beam with a Gaussian profile is focused inside the atomic vapor to a $1/e^2$ waist size of 45 $\mu$m at the focal plane and a Rayleigh range of 4 mm. In order to study the dependence of the SN spectra on the probe beam detuning $\delta_p$, the probe frequency $\nu_p$ is varied over a large range of frequencies ($\sim$ 25 GHz). The relevant energy-levels of $^{87}$Rb are depicted in Fig.~\ref{fig:Energy_Diagram}. Special care is taken so that the laser operates in a mode-hop free regime. The frequency $\nu_p$ of the probe beam is measured using a commercial wavelength meter (HighFinesse, model-WSU2) with a relative accuracy of $\pm$ 1 MHz. \begin{figure} \centering \includegraphics[scale=0.38]{2_EnergyDiagram} \caption{Energy level diagram for $D_2$ transition of $^{87}$Rb atoms. The two ground state hyperfine levels ($F = 1$ and $F = 2$) are separated by $\sim 6.8$ GHz. The probe laser frequency $\nu_p$ is detuned by $\delta_p$ from the $F = 2 \rightarrow F' = 3$ transition, i.e., $\delta_p = \nu_{p} - \nu_{F = 2 \rightarrow F' = 3}$, where $\nu_{F = 2 \rightarrow F' = 3}$ is the frequency between $F = 2$ and $F' = 3$ hyperfine levels. The Zeeman sub-levels of the ground state hyperfine levels in the presence of external magnetic field are depicted. The selected transitions for optical pumping of atoms by a control laser of frequency $\nu_c$ are also shown.} \label{fig:Energy_Diagram} \end{figure} The glass cell is filled with neon buffer gas at a pressure of $200$ mbar in order to make the medium diffusive for Rb atoms. This increases the transverse transit time of the atoms across the probe beam from $\sim$ 200 ns to $\sim$ 100 $\mu$s providing sufficient time for acquiring time resolved Faraday rotation signal for the accurate detection of the atomic properties. We collect each real-time Faraday rotation signal for relatively longer time duration than the transit time. However, the spin life-time of Rb atoms at room temperature is of the order of milliseconds, \cite{Arditi1964} which is much longer than the above transverse transit time ($\sim$ 100 $\mu$s). The inert gas neon is chosen because the collisions between Rb and neon do not change the spin state of the Rb atoms. The vapor cell is connected to a controllable heater in order to vary the number density of the atomic spin ensemble. The atoms are subjected to a uniform, constant magnetic field ($B_\bot \hat{x}$) perpendicular to the direction of propagation of the probe beam ($\hat{z}$). This field, generated using two circular coils in Helmholtz configuration, is uniform along the length of the cell within $\pm 0.4\%$. The entire experimental set-up is shielded with a mild-steel box ($\mu/\mu_0 = 2000$) to avoid any unwanted stray magnetic field. The probe beam after passing through the glass cell is separated into $s$- and $p$-polarized components using a polarization sensitive set-up as shown in Fig.~\ref{fig:Experimental_set-up}(a). The two components are then fed into the two ports of a balanced photo detector (Newport model no. 1807-FS) that has a 3 dB bandwidth of 80 MHz and a good common mode rejection ratio of $25$ dB. The output of the balanced detector is directly connected to a spectrum analyzer (Agilent CSA Spectrum Analyzer Model no. N1996A, frequency range $100$ kHz - $3$ GHz) whose resolution bandwidth is adjusted between $100$ Hz to $1$ kHz. The spectrum analyzer was set on continuous averaging mode for two to five minutes for recording various SN spectra presented in this paper. \section{Measurements and Results in equilibrium} \label{equi} A typical spin noise spectrum of Rb atoms in thermal equilibrium at low $B_\bot$ ($=6.95$ G) is presented in Fig.~\ref{fig:Experimental_set-up}(b). This signal was obtained with the vapor cell at 105$^{\circ}$C and a $p$-polarized ($\hat{\varepsilon}\parallel\hat{x}$), 300 $\mu W$ probe beam. The probe beam is red-detuned by 10.2 GHz with respect to the $D_2$ transition (at $\sim$ 780 nm) of $^{87}$Rb, $F=2\rightarrow F'=3$. Two distinct noise peaks, one at 3.24 MHz and another at 4.87 MHz, are observed and identified as arising due to spin fluctuations among the intra-hyperfine Zeeman sublevels ($\bigtriangleup F=0$, $\bigtriangleup m_F=\pm1$) of $^{85}$Rb and $^{87}$Rb, respectively. The photon shot noise background of $\sim$ 350 nV.Hz$^{-1/2}$ is subtracted from the noise spectrum. The observed SN peaks are very narrow ($<$ 100 kHz) and the peak positions (which occur at the $\nu_L$) can be detected with a precision of $\sim 1$ part in $10^5$. This makes it possible to employ SNS for a variety of precision measurements as we demonstrate in the following sections. \subsection{Measurements of g-factors and isotope abundance} \begin{figure} \centering \includegraphics[scale=0.42]{3_a_b} \caption{ (a) Spin noise (SN) spectra at various $B_\bot$ are presented. The linear shift of the noise peaks with $B_\bot$ suggests a linear Zeeman effect of the ground state hyperfine levels in that $B_\bot$ range. (b) A 2-D false color mapping shows the SN peak positions for $^{85}$Rb and $^{87}$Rb as a function of $B_\bot$. The bright (faint) trace is for $^{87}$Rb ($^{85}$Rb) SN signal. The noise signal strength of $^{87}$Rb and $^{85}$Rb for each spectrum is plotted after normalizing the signal by the SN peak strength of $^{87}$Rb. The slopes of these traces reveal $|g_F|$ of Rb isotopes.} \label{fig:SNS Signal} \end{figure} Fig.~\ref{fig:SNS Signal}(a) shows SN spectra at three representative values of $B_\bot$ illustrating that the two noise peaks, corresponding to $^{87}$Rb and $^{85}$Rb, shift in positions with $B_\bot$. Fig.~\ref{fig:SNS Signal}(b) gives the variation in the position of these noise peaks as a function of $B_\bot$. The bright (faint) trace corresponds to the spin noise peak positions of $^{87}$Rb ($^{85}$Rb) atoms. The linear dependence of the peak positions on $B_\bot$ indicates that the system is in the linear Zeeman regime. The slope of these lines give a measure of the g-factor $|g_F|$ for the ground state hyperfine levels. The g-factors obtained from our measurements are $|g_F| = 0.500(\pm 0.001)$ for $^{87}$Rb and $|g_F| = 0.333(\pm 0.001)$ for $^{85}$Rb which are in excellent agreement with the theoretical values. Our measurements were made with an enriched $^{87}$Rb vapor cell. Traditional absorption spectroscopy did not show the presence of the isotope $^{85}$Rb. However, SN spectra clearly indicate the presence of both isotopes in the cell. From the ratio of the total integrated SN power ($\chi$) of the two peaks, we estimate an abundance ratio of $^{87}$Rb $:$ $^{85}$Rb $= 11:1$ from Fig.~\ref{fig:Experimental_set-up}(b). This shows that SNS is a very sensitive technique for detecting abundance ratios of various isotopes with high precision even when present in minute quantities. \subsection{Precision magnetometry} \begin{figure} \centering \includegraphics[scale=0.42]{4_a_b} \caption{ Broadening and splitting of the spin noise (SN) spectrum with increasing $B_\bot$. (a) SN spectra for $^{87}$Rb at $B_\bot = 2.7$ G, $31.27$ G, $125$ G. The origin of the frequency in these spectra is shifted to the central Larmor frequency $\nu_L=g_F\mu_BB_{\bot}/h$. The other parameters are: probe power = $400 \mu$W, $\delta_p = -10.6$ GHz, cell temperature = $105^{\circ}$C. (b) Visual realization of the nonlinear Zeeman effect of ground state hyperfine levels with increasing $B_\bot$. Each spectrum is normalized by the strongest peak in the SN signal.} \label{fig:Quadratic_Zeeman} \end{figure} On increasing the magnitude of the applied magnetic field ($B_\bot$), the spin noise spectra is observed (Fig.~\ref{fig:Quadratic_Zeeman}(a)) to broaden ($B_\bot\sim 25-40$ G) and to split into well-resolved peaks at even higher $B_\bot$ ($> 60$ G). At such high fields, the system is clearly in non-linear Zeeman regime (Eq.~\ref{Eqn:Breit-Rabi}). A false-color mapping of the measured nonlinear Zeeman splitting of the ground state hyperfine levels of $^{87}$Rb atoms as a function of $B_\bot$ is shown in Fig.~\ref{fig:Quadratic_Zeeman}(b). The individual noise peaks in the SN spectrum for a higher $B_\bot$ $(> 60$ G$)$ can be identified as the transitions between different Zeeman sub-levels of the ground state hyperfine levels. These are shown in the inset of Fig.~\ref{fig:magnetometry} where $P1$ denotes the magnetic resonance frequency between ($F = 2, m_F = 2) \leftrightarrow (F = 2, m_F = 1$) and $P2$ for the magnetic resonance frequency between ($F = 2, m_F = 1) \leftrightarrow (F = 2, m_F = 0$) and so on. Fig.~\ref{fig:Quadratic_Zeeman}(b) shows nonlinear dependence of each noise peak frequency on $B_\bot$. However, the sum of all four noise peak frequencies, \begin{eqnarray} \label{Eqn:S} S=P1 + P2 + P3 + P4 = \frac{\mu_B}{h}(3 g_I + g_J) B_{\bot}, \end{eqnarray} depends linearly on $B_{\bot}$. Using the values of $P1, P2, P3, P4$ determined from the measured SN spectrum one can estimate $B_{\bot}$ using Eq.~\ref{Eqn:S}, substituting the values of $\mu_B,h,g_I,g_J$, which are already known to high precision. As the observed SN peaks are extremely narrow (FWHM $< 100$ kHz) and the peak positions can be determined with an accuracy of one part in $10^5$, we can therefore measure an external magnetic field within that same order of relative error, of one part in $10^5$, for the range of $B_{\bot}$ where noise peaks are separable. Thus SNS provides a simple means of precision magnetometry. \begin{figure} \centering \includegraphics[scale=0.4]{5_PrecisionMagnetometry} \caption{Measurement of magnetic field $B_{\bot}$ and $\Delta_{\text{hf}}$ using spin noise spectroscopy. $P1,P2,P3$ and $P4$ indicate the position of spin noise peaks (corresponding to Zeeman sub-levels of $F = 2$) as can be seen from the raw data in the inset. The measured frequency separation between different noise peaks are plotted against measured $\alpha S$ (refer to the text). The bold lines are obtained from the Breit-Rabi formula in Eq.~\ref{Eqn:Breit-Rabi1} with known parameters for error analysis in magnetic field measurements. The experimental parameters are the same as in Fig.~\ref{fig:Quadratic_Zeeman}. } \label{fig:magnetometry} \end{figure} Now we rewrite Eq.~\ref{Eqn:Breit-Rabi} for $^{87}$Rb as \begin{eqnarray} \label{Eqn:Breit-Rabi1} E_{F=2,m_F}&=&-\frac{h\Delta_{\text{hf}}}{8} + \frac{h g_I}{(g_J-g_I)} (\alpha S) m_F\nonumber \\ &+&\frac{h}{2} \sqrt{\Delta_{\text{hf}}^2+\Delta_{\text{hf}}(\alpha S)m_F+(\alpha S)^2}, \end{eqnarray} where $\alpha=(g_{J}-g_{I})/(g_{J}+3g_{I})$. In Fig.~\ref{fig:magnetometry}, we plot, as a function of $\alpha S$, the noise peak separations (($P4-P1$), ($P4-P2$) ... ($P2-P1$)) calculated from Eq.~\ref{Eqn:Breit-Rabi1} using known values of $h,g_I,g_J,\Delta_{\text{hf}}$. Superposed on the plot are the experimentally obtained noise peak separations shown as solid symbols. Then, we note down the x-errors between the experimentally obtained peak separations and those from the calculated curves. The root-mean-square value of the x-errors gives an estimate of the error in measuring the external magnetic field. We find that the error is within $500~\mu$G in our measurement range of $B_{\bot}$ between $60$ G to $150$ G. This accuracy surpasses the standard Hall probe based magnetometers by nearly two orders of magnitude. Moreover, this high precision measurement of magnetic field is an in-situ detection, without requiring the physical placement of a separate probe. In Fig.~\ref{fig:magnetometry}, we have also fitted the experimentally obtained values of ($P4-P1$), ($P4-P2$) ... ($P2-P1$) using Eq.~\ref{Eqn:Breit-Rabi1} keeping $\Delta_{\text{hf}}$ as a free parameter. From these fittings, we extract the value of the zero-field hyperfine constant $\Delta_{\text{hf}} \sim 6805.5 (\pm 7.2)$ MHz. \subsection{Nuclear g-factor from SNS} In the presence of $B_{\bot}$, the energy separations between similar magnetic (Zeeman) transitions from different hyperfine ground states $(F, m_F)$ such as, $(2,1) \leftrightarrow (2,0)$ and $(1,1) \leftrightarrow (1,0)$ or $(2,0) \leftrightarrow (2,-1)$ and $(1,0) \leftrightarrow (1,-1)$ in Fig.~\ref{fig:Energy_Diagram}, are determined by the second term in Eq.~\ref{Eqn:Breit-Rabi} arising out of the nuclear spin. However, since the value of nuclear g-factor ($g_I$) is small, the contribution of this term is negligible for low magnetic fields. Therefore, the SN peaks from $(2,1) \leftrightarrow (2,0)$ and $(1,1) \leftrightarrow (1,0)$ are almost unresolved for $B_{\bot}< 150$ G in our case. At high magnetic fields ($>150$ G), we can resolve the SN peaks from all available Zeeman transitions when their separations are more than the width of the individual peaks. Six distinct SN peaks from the allowed $\bigtriangleup F = 0$, $\bigtriangleup m_F = \pm1$ transitions of $^{87}$Rb are observed in Fig.~\ref{fig:NuclearZeeman} at $B_{\bot} = 160$ G. The value of $g_I$ can be precisely obtained by measuring the separation between $(2,1) \leftrightarrow (2,0)$ and $(1,1) \leftrightarrow (1,0)$ (also $(2,0) \leftrightarrow (2,-1)$ and $(1,0) \leftrightarrow (1,-1)$). From a series of such measurements, the experimentally estimated $g_I$ for $^{87}$Rb in our experiment is $-0.00100627(\pm 0.00002558)$, where the quantity in the bracket refers to the $1\sigma$ error. \begin{figure} \centering \includegraphics[scale=0.4]{6_NuclearZeeman} \caption{Spin noise spectrum with resolved all-allowed Zeeman transitions ($\bigtriangleup F = 0$, $\bigtriangleup m_F = \pm1$) of ground state hyperfine levels in $^{87}$Rb. The parameters are $B_\bot = 160$ G, probe power = $750~\mu$W, $\delta_p = -10.6$ GHz, cell temperature = $105^{\circ}$C. The value of nuclear g-factor $(g_I)$ is precisely obtained and reported in the text by measuring the separation between $(2,1) \leftrightarrow (2,0)$ and $(1,1) \leftrightarrow (1,0)$ (also $(2,0) \leftrightarrow (2,-1)$ and $(1,0) \leftrightarrow (1,-1)$) in a series of measurements.} \label{fig:NuclearZeeman} \end{figure} \section{Measurements and Results in optically pumped vapor} \label{outEq} Thus far, we have explored SNS in an equilibrium thermal vapor. We now report the measurements on out-of-equilibrium systems where we use optical fields to manipulate spin populations in the different ground state hyperfine levels. Recently, the SNS was employed to detect couplings and correlations between different spin coherences in a non-equilibrium atomic vapor \cite{Glasenapp2014}. In the experiment in \cite{Glasenapp2014}, the Zeeman sub-levels of the ground state hyperfine levels of $^{41}$K are driven by a weak radio frequency magnetic field which brings the vapor out of equilibrium. In contrast, we apply an optical control beam between the ground state hyperfine levels and the excited state hyperfine levels to drive as well as control the spin populations. The control beam is linearly polarized and almost co-propagating with the probe beam. In our experiment, the control beam is derived from an independent tunable external cavity diode laser. The Rb atoms in the vapor cell are optically pumped to the desired ground state hyperfine levels by tuning the frequency $\nu_c$ and the intensity $I_c$ of the pump laser. Substantial modifications of the SNS signals are observed depending upon the relative ground state hyperfine level populations of the vapor. \subsection{Detection of spin imbalance} \begin{figure} \centering \includegraphics[scale=0.43]{7_InAndOutEquilibriumSNS} \caption{ Spin noise (SN) spectra in and out of equilibrium $^{87}$Rb atoms. (a) SN spectra with a pump beam on resonant to $F = 1 \rightarrow F' = 2$ optical transition for various pump beam intensities $I_c$. (b) SN spectrum in thermal equilibrium (without optical pump beam). (c) SN spectra with a pump beam on resonant to $F = 2 \rightarrow F' = 2$ optical transition for different pump beam intensities $I_c$. For all panels, $B_\bot = 160$ G, $\delta_p=-10.6$ GHz and cell temperature = 105$^{\circ}$C.} \label{fig:OpticalPumping} \end{figure} Here, we implement the off-resonant SNS to probe the spin imbalance in an optically driven system without applying further perturbation \cite{Happer1967}. In the absence of the pump beam, six spin coherences are seen in the SN spectrum in Fig.~\ref{fig:OpticalPumping}(b) as is expected from an ensemble of atoms with population in both the ground state hyperfine levels ($F = 1,2$). On setting the frequency $\nu_c$ of the control beam on resonance to the $F = 1 \rightarrow F' = 2$ transition of $^{87}$Rb (see Fig.~\ref{fig:Energy_Diagram}), a fraction of atoms is pumped out of the ground $F = 1$ level depending on the intensity $I_c$ of the pump. This is evident in Fig.~\ref{fig:OpticalPumping}(a) where only four SN peaks related to $F = 2, \bigtriangleup m_F = \pm 1$ are observed at the highest $I_c$. On the other hand, when the atoms are pumped out of the ground $F = 2$ level using a pump beam resonant with the $F = 2 \rightarrow F' = 2$ transition, the SN spectrum reduces to two peaks related to $F = 1, \bigtriangleup m_F = \pm 1$ spin coherences. This is shown in Fig.~\ref{fig:OpticalPumping}(c) for different pump beam intensities. This clearly shows that the spin populations in different ground state hyperfine levels is reflected in the SN spectra. Such relatively non-invasive detection of spin states in a non-equilibrium atomic system may find applications in atom interferometry \cite{Cronin2009}, atomic clocks \cite{Wynands2005} and gravimetry \cite{Bertoldi2006}. \subsection{Resolving spectral lines} The enriched $^{87}$Rb vapor cell used in our experiments contained a buffer gas (neon), and thus the conventional absorption spectra that we measure suffers broadening mainly due to homogeneous pressure broadening \cite{Romalis1997, Couture2008, Li2016} and modestly due to inhomogeneous Doppler broadening. The transmission of the probe through the atomic vapor at 90$^{\circ}$C is studied in the absence and presence of an optical control beam as shown in Fig.~\ref{fig:SNA_detuning}(a,c). The detuning $\delta_p$ of the probe beam was varied over a wide range $(-10$ GHz to $12$ GHz$)$ which covers both the $F=2 \rightarrow F'$ and $F = 1 \rightarrow F'$ transition lines. In the absence of a pump beam in Fig.~\ref{fig:SNA_detuning}(a), the probe transmission as a function of $\delta_p$ shows a single dip situated between the transition lines. The integrated SN power $\chi$ from the Rb vapor also shows a single dip in Fig.~\ref{fig:SNA_detuning}(b) in the absence of optical pumping. Thus, both the absorption spectroscopy and SNS fail to detect $F=2 \rightarrow F'$ and $F = 1 \rightarrow F'$ transition lines separately. Nevertheless, the dip in $\chi$ seems to indicate a red-detuned $F=2 \rightarrow F'$ transition as expected in the presence of neon buffer gas \cite{Li2016}. In the case of an optically pumped atomic ensemble, the probe transmission (Fig.~\ref{fig:SNA_detuning}(c)) can detect the above two optical transitions independently. In Fig.~\ref{fig:SNA_detuning}(d), we show the integrated SN power $\chi$ with the probe detuning $\delta_p$ when the atoms are optically pumped to either $F=2$ or $F=1$ level by a pump intensity $I_c > 50 I_{\text sat}$. We observe a dip in each integrated SN power near $F=2 \rightarrow F'$ or $F = 1 \rightarrow F'$ transition lines. Therefore, the SNS can also be used to resolve the spectral lines in a driven atomic system \cite{Zapasskii2013b}. Moreover, SNS has a better resolution (around three times in Fig.~\ref{fig:SNA_detuning}(d) than Fig.~\ref{fig:SNA_detuning}(c)) over the absorption spectroscopy \cite{Zapasskii2013b}. \begin{figure} \centering \includegraphics[scale=0.35]{8_ResolvingSpectralLines} \caption{Comparison between absorption spectroscopy and spin noise spectroscopy in resolving spectral lines in a buffer gas filled Rb vapor cell in the absence (a,b) and presence (c,d) of optical pumping. (a,c) Probe transmission vs. probe beam detuning, $\delta_p$ defined in Fig.~\ref{fig:Energy_Diagram}. (b,d) Integrated spin noise (SN) power $\chi$ from $^{87}$Rb atoms vs. $\delta_p$. Red triangles (blue squares) depict probe transmission and integrated SN power $\chi$ from optically pumped $F=2~(F=1)$ atoms. The lines joining the data points are a guide to the eye. For all panels, $B_\bot = 7.12$ G and cell temperature = $90^{\circ}$C.} \label{fig:SNA_detuning} \end{figure} \begin{figure} \centering \includegraphics[scale=0.38]{9_PumpDetuningVariation} \caption{Resolving spectral lines of $^{87}$Rb atoms in the presence of neon buffer gas by tuning the pump beam frequency $\nu_c$. Here, the probe beam detuning $\delta_p = -10$ GHz and the pump beam intensity $I_c \sim 50 I_{\text sat}$. The black dotted horizontal line represents the integrated spin noise power $\chi$ without optical pumping. The line joining the data points are a guide to the eye. } \label{fig:PumpBeamDiffDetuning} \end{figure} We can also detect these transitions by tuning the frequency $\nu_c$ of the pump beam instead of the probe beam. Here we keep the probe detuning $\delta_p$ fixed at $-10$ GHz from $F = 2 \rightarrow F'$ transition (and around $-16.8$ GHz detuned from $F = 1 \rightarrow F'$ transition). Therefore, most of the contribution in the SN signal comes from the $F = 2$ level. We tune the frequency $\nu_c$ of the pump beam from $-10$ GHz to 10 GHz around $F = 2 \rightarrow F'=3$ transition. We plot the integrated SN power $\chi$ as a function of the pump beam detuning in Fig.~\ref{fig:PumpBeamDiffDetuning}, and observe a clear dip near $F = 2 \rightarrow F'$ transition and a prominent peak near $F = 1 \rightarrow F'$ transition. Therefore, unlike conventional spectroscopy, in the case of SNS, we have the freedom to scan the pump beam for detecting the spectral lines instead of applying the pump beam at a particular known frequency as in Fig.~\ref{fig:SNA_detuning}(c,d). This, we believe, will be of particular advantage, when we wish to probe local environment-induced energy level shifts, or in resolving ground state levels in complex molecular and condensed matter systems where one has incomplete knowledge of energy levels. \section{Conclusion and Outlook} \label{conc} We explore the SNS technique in atomic vapor of Rb in thermal equilibrium as well as in a system driven out of equilibrium by optical pumping. Optical pumping is a commonly used technique in atomic and optical science and technology. To the best of our knowledge, the present study is the first implementation of SNS in an optically pumped atomic system. Our effort to combine the SNS and the optical pumping has the potential for use in magnetometry with alkali atoms \cite{Allred2002}. There are important applications of such {\it{precision magnetometry}}, e.g., in cold atom experiments where narrow Feshbach resonances \cite{Chin2010} are used extensively with the resonances occurring at magnetic fields ranging from a few Gauss to a few hundred Gauss. In those experiments, measuring the external fields with ultra-high precision will be hugely beneficial in fixing the interaction energy scale. We also extend the applicability of SNS in precision measurements of various atomic, nuclear and magnetic properties in equilibrium systems. The relatively non-perturbative nature of SNS makes it a versatile non-invasive detection technique which can be utilized in a wide range of physical systems in atomic, molecular and condensed matter systems. Recently, some measurements of spin polarization in ultracold atoms via Faraday rotation of a far-detuned probe beam were carried out as a non-destructive imaging technique \cite{Gajdacz2013, Gajdacz2016, Yamamoto2017, Palacios2018}. We are interested to further implement SNS using Faraday rotation noise in ultracold atoms and Bose-Einstein condensates where it may have significant application in quantum non-demolition measurements. However, acquiring sufficient time-resolved Faraday-rotation noise from such ultracold atomic systems poses a challenge. \begin{acknowledgements} The authors acknowledge funding support from Department of Science and Technology, India. The authors also acknowledge the contribution of Meena M. S. for the help with electronics and mechanical workshop of Raman Research Institute for the hardware development. D.R. acknowledges funding from the Department of Science and Technology, India, via the Ramanujan Fellowship. \end{acknowledgements}

b541043558632a161bf7214c10c99e55db9e6485

\section{Introduction} \label{intro-sec} The information-theoretic notion of entropy captures the idea of the typical ``uncertainty'' in a sample of a given probability measure. An important class of probability measures in model theory, combinatorics, and probability theory are the $\Sym{{\EM{{\mbb{N}}}}}$-invariant measures on the space of structures, in a countable language, with underlying set ${\EM{{\mbb{N}}}}$. We call such measures \emph{invariant measures}, and consider the entropy of such measures. When the support of a probability measure is uncountable, as is usually the case for invariant measures, the entropy of the measure is infinite. But any such measure $\mu$ can be approximated by its projections to the spaces of structures on initial segments of ${\EM{{\mbb{N}}}}$. We define the \emph{entropy function} of $\mu$ to be the function from ${\EM{{\mbb{N}}}}$ to ${\EM{{\mbb{R}}}} \cup \{\infty\}$ taking $n$ to the entropy of the measure induced by $\mu$ on structures with underlying set $\{0, \dots, n-1\}$. In this paper, we study the growth of entropy functions. The growth of the entropy function of an invariant measure has been studied by Aldous \cite[Chapter~15]{MR883646} in the setting of exchangeable arrays and by Janson \cite[\S10 and \S D.2]{MR3043217} and Hatami--Norine \cite[Theorem~1.1]{MR3073488} in the case where the measure is concentrated on the space of graphs. The leading coefficient of the entropy function has been used to study large deviations for Erd\H{o}s--R\'enyi\ random graphs and exponential random graphs in \cite{MR2825532} and \cite{MR3127871}, and phase transitions in exponential random graphs in \cite{MR3083277}. For additional results on the entropy functions of invariant measures concentrated on the space of graphs, see \cite{MR3742179}. Further, ergodic invariant measures are the focus of a project that treats them model-theoretically as a notion of ``symmetric probabilistic structures''; see \cite{MR3515800}, \cite{MR3425980}, \cite{MR3564374}, \cite{AFPcompleteclassification}, and \cite{2017arXiv171009336A}. The entropy function of such a measure is one gauge of its complexity. In this paper we primarily consider invariant measures, for a countable relational language, that are \emph{non-redundant}, namely those that concentrate on structures in which a relation can hold of a tuple only when the tuple consists of distinct elements. In \S\ref{qf-interdef-subsec} we develop machinery for quantifier-free interdefinitions that allows us to show, in Proposition~\ref{nonredundant-entropy-function}, that every entropy function of an invariant measure for a countable language is the entropy function of some non-redundant invariant measure for a countable relational language. In Section~\ref{section-finite}, we study the case of invariant measures for countable relational languages ${\EM{L}}$ with all relations having the same arity $k\ge 1$. The Aldous--Hoover--Kallenberg theorem provides a representation, which we call an \emph{extended ${\EM{L}}$-hypergraphon}, of a non-redundant invariant measure for ${\EM{L}}$. Our main technical result in this section, Theorem~\ref{Entropy-from-hypergraphon}, shows that under a certain condition (which is satisfied, e.g., when $L$ is finite), the entropy function of the invariant measure corresponding to an extended ${\EM{L}}$-hypergraphon $W$ grows like $C n^k + o(n^k)$, where the constant $C$ can be calculated from $W$. When the invariant measure is concentrated on the space of graphs, Theorem~\ref{Entropy-from-hypergraphon} reduces to a result first observed by Aldous in \cite[Chapter~15]{MR883646}, and also by Janson in \cite[Theorem~D.5]{MR3043217}. The proof of Theorem~\ref{Entropy-from-hypergraphon} follows closely that of Janson. In Section~\ref{randomfree-sec} we consider invariant measures that arise via sampling from a Borel $k$-uniform hypergraph for $k\ge2$. The entropy functions of such measures grow like $o(n^k)$, as we show in Lemma~\ref{random-free-has-leading-entropy-zero}. We prove, in Theorem~\ref{Random-free has arbitrary growth rates}, that for every function $\gamma(n) \in o(n^k)$ there is an invariant measure sampled from a Borel $k$-uniform hypergraph whose entropy function grows faster than $\gamma(n)$. Our theorem generalizes a result of Hatami--Norine \cite[Theorem~1.1]{MR3073488}, who prove it in the case where $k = 2$ (hence where the Borel hypergraph is simply a Borel graph). Hatami--Norine's proof analyzes the random graph obtained by subsampling from an appropriate ``blow up'' of the Rado graph, i.e., the unique countable homogeneous-universal graph. Our proof takes an analogous path, using the unique countable homogeneous-universal $k$-uniform hypergraph. Finally, in Section~\ref{nonredundant-sec} we consider invariant measures for countable languages that may be of unbounded arity. We provide an upper bound on the entropy functions of non-redundant invariant measures for a given countable relational language, in terms of the number of relations of each arity. We do so by calculating the entropy function of a particular non-redundant invariant measure whose entropy function is maximal among those for that language. This calculation also demonstrates that whereas the growth of the entropy function of an invariant measure for a finite language is polynomial, there are ${\EM{{\mbb{R}}}}$-valued entropy functions that grow arbitrarily fast. \subsection{Preliminaries} \emph{In this subsection, let ${\EM{L}}$ be a countable language and let ${\mathfrak{n}}\in{\EM{{\mbb{N}}}} \cup \{{\EM{{\mbb{N}}}}\}$.} For $n\in{\EM{{\mbb{N}}}}$, write $[n]{\EM{\ :=\ }} \{0, \ldots, n-1\}$, and $[{\EM{{\mbb{N}}}}] {\EM{\ :=\ }} {\EM{{\mbb{N}}}}$. Let $\Sym{{\mathfrak{n}}}$ denote the collection of permutations of $[{\mathfrak{n}}]$. For $k \in {\EM{{\mbb{N}}}}$, define $\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}({\mathfrak{n}}) {\EM{\ :=\ }} \{Y \subseteq [{\mathfrak{n}}] \,:\, |Y| < k\}$ and $\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{k}({\mathfrak{n}}) {\EM{\ :=\ }}$ $\{Y \subseteq [{\mathfrak{n}}] \,:\, |Y| = k\}$, and let $\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{\leq k}({\mathfrak{n}}) {\EM{\ :=\ }} \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{< k}({\mathfrak{n}}) \cup \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{k}({\mathfrak{n}})$. We order each of these sets using \emph{shortlex} order, i.e., ordered by size, with sets of the same size ordered lexicographically. We write $\Lomega{\EM{\omega}}({\EM{L}})$ to denote the collection of first-order ${\EM{L}}$-formulas. An \emph{${\EM{L}}$-theory} is a collection of first-order ${\EM{L}}$-sentences. A theory is $\Pi_1$ when every sentence is quantifier-free or of the form $(\forall {\EM{\textbf{x}}})\varphi({\EM{\textbf{x}}})$ where ${\EM{\textbf{x}}}$ is a tuple of variables and $\varphi$ is quantifier-free. The \emph{maximum arity} of $L$ is the maximum arity, when it exists, of a relation symbol or function symbol in $L$, where we consider constant symbols to be function symbols of arity $0$. Fix a probability space $(\Omega, {\EM{\mc{F}}}, {\EM{\mathbb{P}}})$. Suppose $(D, {\EM{\mathcal{D}}})$ is a measurable space. A \emph{$D$-valued random variable} $Z$, also called a \emph{random element in} $D$, is an $({\EM{\mc{F}}},{\EM{\mathcal{D}}})$-measurable function $Z \colon \Omega \to D$. The \emph{distribution} of $Z$ is the probability measure ${\EM{\mathbb{P}}}\circ Z^{-1}$. Given an event $B \in {\EM{\mathcal{D}}}$, we say that $B$ holds \emph{almost surely} when ${\EM{\mathbb{P}}}(B) = 1$, and abbreviate this \emph{a.s.} Typically $B$ will be specified indirectly via some property of random variables; for example, we say that two random variables are equal a.s.\ when the subset of $\Omega$ on which they are equal has full measure. For a topological space $S$, let ${\EM{\mathcal{P}}}(S)$ denote the space of Borel probability measures on $S$, with $\sigma$-algebra given by the weak topology. We use the symbol $\bigwedge$ for conjunctions of probabilistic events, as well as for conjunctions of logical formulas. See \cite{MR1876169} for further background and notation from probability theory. We write $\lambda$ to denote the uniform (Lebesgue) measure on $[0, 1]$ and on finite powers of $[0,1]$. Suppose a variable $x$ takes values in some finite power of $[0,1]$. We say that an expression involving $x$ holds \emph{almost everywhere}, abbreviated \emph{a.e.}, when it holds of $x$ on all but a $\lambda$-null subset. Let $\Str_{\EM{L}}({\mathfrak{n}})$ denote the collection of ${\EM{L}}$-structures that have underlying set $[{\mathfrak{n}}]$. We will consider $\Str_{\EM{L}}({\mathfrak{n}})$ as a measure space with the $\sigma$-algebra generated by the topology given by basic clopen sets of the form \[ \{{\EM{\mathcal{M}}} \in \Str_{\EM{L}}({\mathfrak{n}}) \,:\, {\EM{\mathcal{M}}} \vDash \varphi(x_0, \dots, x_{\ell-1})\} \] when $\varphi\in \Lomega{\EM{\omega}}(L)$ is a quantifier-free formula with $\ell$ free variables and $x_0, \dots, x_{\ell-1} \in [{\mathfrak{n}}]$. \begin{definition} Suppose $\varphi \in \Lomega{\EM{\omega}}({\EM{L}})$ has $\ell$ free variables and $r_0, \dots, r_{\ell-1} \in [{\mathfrak{n}}]$. Define the \defn{extent} (on $[{\mathfrak{n}}]$) of $\varphi(r_0, \dots, r_{\ell-1})$ to be \[ \Extent{\varphi(r_0, \dots, r_{\ell-1})}_{[{\mathfrak{n}}]} {\EM{\ :=\ }} \{ {\EM{\mathcal{M}}} \in \Str_{\EM{L}}({\mathfrak{n}}) \,:\, {\EM{\mathcal{M}}} \vDash \varphi(r_0, \dots, r_{\ell-1})\}. \] When ${\mathfrak{n}}={\EM{{\mbb{N}}}}$ we will sometimes omit the subscript $[{\mathfrak{n}}]$. For an ${\EM{L}}$-theory $T$, define $\Extent{T}_{[{\mathfrak{n}}]} {\EM{\ :=\ }} \bigcap_{\rho\in T}\Extent{\rho}_{[{\mathfrak{n}}]}$. \end{definition} There is a natural action of $\Sym{{\mathfrak{n}}}$ on $\Str_{\EM{L}}({\mathfrak{n}})$ called \defn{the logic action}, defined as follows: for $\sigma\in\Sym{{\mathfrak{n}}}$ and ${\EM{\mathcal{M}}}\in\Str_{\EM{L}}({\mathfrak{n}})$, let $\sigma \cdot {\EM{\mathcal{M}}}$ be the structure ${\EM{\mathcal{N}}}\in\Str_{\EM{L}}({\mathfrak{n}})$ for which \[ R^{\EM{\mathcal{N}}}(r_0, \ldots, r_{k-1}) \quad \text{if and only if} \quad R^{\EM{\mathcal{M}}}\bigl(\sigma^{-1}(r_0), \ldots, \sigma^{-1}(r_{k-1}) \bigr) \] for all relation symbols $R\in L$ and $r_0, \ldots, r_{k-1} \in{\mathfrak{n}}$, where $k$ is the arity of $R$, and similarly with constant and function symbols. Note that the orbit under the logic action of any structure in $\Str_{\EM{L}}({\mathfrak{n}})$ is its isomorphism class. By Scott's isomorphism theorem, every such orbit is Borel. For more details on the logic action, see \cite[\S16.C]{MR1321597}. We say a probability measure $\mu$ on $\Str_{{\EM{L}}}({\mathfrak{n}})$ is \defn{invariant} if it is invariant under the logic action of $\Sym{{\mathfrak{n}}}$, i.e., if $\mu(B) = \mu(\sigma \cdot B)$ for every Borel $B \subseteq \Str_{\EM{L}}({\mathfrak{n}})$ and every $\sigma \in\Sym{{\mathfrak{n}}}$. We call such a probability measure an \emph{invariant measure for ${\EM{L}}$}. An invariant measure $\mu$ is \defn{ergodic} if $\mu(X) = 0$ or $\mu(X) = 1$ whenever $\mu(X\triangle\sigma(X)) = 0$ for all $\sigma \in\Sym{{\EM{{\mbb{N}}}}}$. Every ergodic invariant measure on $\Str_{\EM{L}}({\mathfrak{n}})$ is an \emph{extreme point} in the simplex of invariant measures on $\Str_{\EM{L}}({\mathfrak{n}})$, and any invariant measure on $\Str_{\EM{L}}({\mathfrak{n}})$ can be decomposed as a mixture of ergodic ones (see \cite[Lemma~A1.2 and Theorem~A1.3]{MR2161313}). For $n\in{\EM{{\mbb{N}}}}$, any probability measure $\mu$ on $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$ induces a probability measure $\mu_n$ on $\Str_{\EM{L}}(n)$ such that for any Borel set $B \subseteq \Str_{\EM{L}}(n)$, we have \[ \mu_n(B) {\EM{\ :=\ }} \mu(\{{\EM{\mathcal{M}}} \in \Str_{\EM{L}}({\EM{{\mbb{N}}}}) \,:\, {\EM{\mathcal{M}}}|_{[n]} \in B\}), \] where ${\EM{\mathcal{M}}}|_{[n]}\in\Str_{\EM{L}}(n)$ denotes the restriction of ${\EM{\mathcal{M}}}$ to $[n]$. Further, by the Kolmogorov consistency theorem, $\mu$ is uniquely determined by the collection $\<\mu_n\>_{n \in {\EM{{\mbb{N}}}}}$. Towards defining the entropy function of an invariant measure, we give the standard definition of the entropy of a probability measure. For this definition, we use the convention that $-0 \log_2(0) = 0$. \begin{definition} \label{entropy-definition-finite-set} Let $\nu$ be a probability measure on a standard Borel space $S$, and let $A{\EM{\ :=\ }} \{s\in S \,:\, \nu(\{s\}) > 0\}$ be its (countable) set of atoms. If $\nu$ is purely atomic, i.e., $\nu(A) = 1$, then the \defn{entropy} of $\nu$ is given by \[ \EntropyFin{\nu}{\EM{\ :=\ }} -\sum_{x \in A} \nu(\{x\}) \log_2(\nu(\{x\})). \] Otherwise, let \[ \EntropyFin{\nu}{\EM{\ :=\ }} \infty. \] For any random variable $X$ with distribution $\nu$, define $\EntropyFin{X} {\EM{\ :=\ }} \EntropyFin{\nu}$. \end{definition} \begin{definition} The \defn{joint entropy} of a pair of random variables $X$ and $Y$, written $h(X,Y)$, is defined to be the entropy of the joint distribution of $(X,Y)$. Similarly, $h(\<X_i\>_{i\in I})$ is defined to be the entropy of the joint distribution of the sequence $\<X_i\>_{i\in I}$. \end{definition} \begin{definition} Given random variables $X$ and $Y$, the \defn{conditional entropy} of $X$ given $Y$, written $h(X\in \cdot\ |\,Y)$, is defined to be the function \[ y \mapsto h({\EM{\mathbb{P}}}(X\in\cdot\ |\,Y = y)).\] \end{definition} We have defined the entropy of a random variable to be the entropy of its distribution. When the random variable takes values in a space of measures, instead of considering the entropy of the random variable directly, we sometimes need the random variable, defined below, that is obtained by taking the entropies of these measures themselves. \begin{definition} Suppose $\chi$ is a measure-valued random variable. The \defn{random entropy} of $\chi$, written $H(\chi)$, is a random variable defined by \[ H(\chi)(\varpi) = h(\chi(\varpi)). \] for $\varpi\in\Omega$. \end{definition} This notion allows us to define random conditional entropy. \begin{definition} Given random variables $X$ and $Y$, the \defn{random conditional entropy} of $X$ given $Y$, written $H(X\,|\,Y)$, is defined by \[ H(X\,|\,Y) = H({\EM{\mathbb{E}}}(X\,|\,Y)), \] i.e., the random entropy of the random measure ${\EM{\mathbb{E}}}(X\,|\,Y)$, the conditional expectation of $X$ given $Y$. \end{definition} The following three lemmas are standard facts about entropy, which we will need later. \begin{lemma}[\textnormal{see {\cite[Theorem~2.2.1]{MR2239987}}}] \label{chainrule} For random variables $X$ and $Y$, we have \[ \EntropyFin{X, Y} = \EntropyFin{X} + {\EM{\mathbb{E}}}(\RandomEntropyFin{Y \ |\ X}). \] \end{lemma} \begin{lemma}[\textnormal{see {\cite[Theorem~2.6.5]{MR2239987}}}] \label{cond-reduces-entropy} For random variables $X$ and $Y$, we have \[ \EntropyFin{X} \ge {\EM{\mathbb{E}}}(\RandomEntropyFin{X \ |\ Y}),\] with equality if and only if $X$ and $Y$ are independent. \end{lemma} \begin{lemma}[\textnormal{see {\cite[Theorem~2.6.6]{MR2239987}}}] \label{joint-entropy-less-than-sum-entropy} Let $\<X_i\>_{i \in I}$ be a sequence of random variables. Then \[ \EntropyFin{\<X_i\>_{i \in I}} \leq \sum_{i \in I} \EntropyFin{X_i}, \] with equality if and only if the $X_i$ are independent. \end{lemma} We now define the entropy function of an invariant measure on $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$. \begin{definition} Let $\mu$ be an invariant measure on $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$. The \defn{entropy function} of $\mu$ is defined to be the function $\Entropy{\mu}$ from ${\EM{{\mbb{N}}}}$ to ${\EM{{\mbb{R}}}} \cup \{\infty\}$ given by \[ \Entropy{\mu}(n) {\EM{\ :=\ }} \EntropyFin{\mu_n}. \] \end{definition} The following basic property of the entropy function is immediate. \begin{lemma} Let $\mu$ be an invariant measure on $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$. Then $\Entropy{\mu}$ is a non-decreasing function. \end{lemma} In this paper we will mainly be interested in invariant measures whose entropy functions take values in ${\EM{{\mbb{R}}}}$. The following lemma is immediate because in a finite language, there are only finitely many structures of each finite size. \begin{lemma} \label{realvalued-lemma} Suppose ${\EM{L}}$ is finite, and let $\mu$ be an invariant measure on $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$. Then $\Entropy{\mu}$ is ${\EM{{\mbb{R}}}}$-valued. \end{lemma} In Section~\ref{nonredundant-sec} we will strengthen this lemma by showing that there is a polynomial upper bound of $O(n^k)$ on $\Entropy{\mu}(n)$, where $k$ is the maximum arity of ${\EM{L}}$. The notion of \emph{non-redundance} for an invariant measure, which we introduce next, will be key throughout this paper. \begin{definition} \label{Thnr-def} Suppose ${\EM{L}}$ is relational. For each $R \in {\EM{L}}$ define $\vartheta_R$ to be the formula \[ (\forall x_0, \dots, x_{k-1}) \Big(R(x_0, \dots, x_{k-1}) \rightarrow \bigwedge_{i < j < k} (x_i \neq x_j)\Big), \] where $k$ is the arity of $R$. Define the $\Pi_1$ ${\EM{L}}$-theory \[ \mathrm{Th}_\mathrm{nr}({\EM{L}}) {\EM{\ :=\ }} \{\vartheta_R \,:\, R \in{\EM{L}}\}.\] An ${\EM{L}}$-structure ${\EM{\mathcal{M}}}$ is \defn{non-redundant} when ${\EM{\mathcal{M}}} \vDash \mathrm{Th}_\mathrm{nr}({\EM{L}})$. An invariant measure $\mu$ on $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$ is \defn{non-redundant} when \[ \mu(\Extent{\mathrm{Th}_\mathrm{nr}({\EM{L}})}) = 1. \] \end{definition} For example, any $k$-uniform hypergraph is non-redundant, as is any directed graph without self-loops. The following straightforward lemma provides conditions under which the entropy function takes values in ${\EM{{\mbb{R}}}}$. \begin{lemma} \label{non-redundant-makes-it-real-valued} Suppose ${\EM{L}}$ has finitely many relations of any given arity, and let $\mu$ be a non-redundant invariant measure on $\Str_{{\EM{L}}}({\EM{{\mbb{N}}}})$. Then for every $n \in {\EM{{\mbb{N}}}}$ there is a finite set $A_n \subseteq \Str_{{\EM{L}}}(n)$ such that $\mu_n$ concentrates on $A_n$. In particular, $\Entropy{\mu}(n)\in{\EM{{\mbb{R}}}}$, and so $\Entropy{\mu}$ is ${\EM{{\mbb{R}}}}$-valued. \end{lemma} In Section~\ref{nonredundant-sec} we will consider the case of non-redundant invariant measures for a relational language, and will strengthen this lemma by providing an explicit upper bound on $\Entropy{\mu}(n)$. \subsection{Quantifier-free interdefinitions} \label{qf-interdef-subsec} The notion of \emph{quantifier-free interdefinability}, which we define below, is a variant of the standard notion of interdefinability from the setting of $\aleph_0$-categorical theories (see, e.g., {\cite[\S1]{MR831437}}). It provides a method for translating invariant measures concentrated on the extent of a given $\Pi_1$ theory to invariant measures concentrated on the extent of a target $\Pi_1$ theory, in a way that preserves the entropy function. We use this machinery to show, in Proposition~\ref{nonredundant-entropy-function}, that every entropy function of an invariant measure already occurs as the entropy function of a non-redundant invariant measure for a relational language. \emph{Throughout this subsection, $L_0$ and $L_1$ will be countable languages, sometimes with further restrictions.} \begin{definition} \label{interdefinability} Suppose $T_0$ is an $L_0$-theory and $T_1$ is an $L_1$-theory. A \defn{quantifier-free interdefinition} between $T_0$ and $T_1$ is a pair $\Psi=(\Psi_0, \Psi_1)$ of maps \begin{eqnarray*} \Psi_0 &\colon & \Lomega{\EM{\omega}}(L_0) \to \Lomega{\EM{\omega}}(L_1) \qquad \text{and}\\ \Psi_1 & \colon & \Lomega{\EM{\omega}}(L_1) \to \Lomega{\EM{\omega}}(L_0) \end{eqnarray*} such that for $j \in \{0, 1\}$, the formula $\Psi_j(\eta)$ is quantifier-free whenever $\eta\in\Lomega{\EM{\omega}}(L_{1-j})$ is quantifier-free, and further, \begin{eqnarray*} T_{1-j} & \vdash & \ \Psi_{j} \circ \Psi_{1-j} (\rho) \, \leftrightarrow \, \rho,\\ T_{1-j} & \vdash & \ \Psi_{j} (x=y) \, \leftrightarrow \, (x=y),\\ T_{1-j} & \vdash & \ \neg \Psi_j(\chi) \, \leftrightarrow \, \Psi_j(\neg \chi), \\ T_{1-j} & \vdash & \ \Psi_j(\chi \EM{\wedge} \varphi) \, \leftrightarrow \, \bigl(\Psi_j(\chi) \EM{\wedge} \Psi_j(\varphi)\bigr), \qquad \text{and} \\ T_{1-j} & \vdash& \ (\exists x) \Psi_j(\psi(x)) \, \leftrightarrow \, \Psi_j\bigl((\exists x) \psi(x)\bigr) \end{eqnarray*} for all $\rho\in\Lomega{\EM{\omega}}(L_{1-j})$ and $\chi,\,\varphi,\, \psi(x) \in \Lomega{\EM{\omega}}(L_j)$, and such that the free variables of $\Psi_j(\upsilon)$ are the same as those of $\upsilon$ for every $\upsilon\in \Lomega{\EM{\omega}}(L_j)$. For ${\mathfrak{n}}\in{\EM{{\mbb{N}}}}\cup\{{\EM{{\mbb{N}}}}\}$, the interdefinition $\Psi$ induces maps $\Psi^*_{j,{\mathfrak{n}}}\colon \Extent{T_j}_{[{\mathfrak{n}}]} \to \Extent{T_{1-j}}_{[{\mathfrak{n}}]}$ for $j\in\{0,1\}$ satisfying, for any structure ${\EM{\mathcal{N}}}\vDash T_j$, tuple ${\EM{\textbf{m}}}\in {\mathfrak{n}}$, and formula $\varphi\in\Lomega{\EM{\omega}}(L_{1-j})$, \[ \Psi^*_{j,{\mathfrak{n}}}({\EM{\mathcal{N}}}) \vDash \varphi({\EM{\textbf{m}}}) \qquad \text{if and only if}\qquad {\EM{\mathcal{N}}} \vDash \Psi_{1-j}(\varphi)({\EM{\textbf{m}}}). \] It is immediate that each $\Psi^*_{j,{\mathfrak{n}}}$ is a bijection. \end{definition} Quantifier-free interdefinitions between $\Pi_1$ theories preserve entropy functions, as we now show. \begin{lemma} \label{Interdefinitions-preserve-entropy} Let $\Psi = (\Psi_0, \Psi_1)$ be a quantifier-free interdefinition between a $\Pi_1$ $L_0$-theory $T_0$ and a $\Pi_1$ $L_1$-theory $T_1$. Suppose $\mu$ is an invariant measure on $\Str_{{\EM{L}}_0}({\EM{{\mbb{N}}}})$ concentrated on $\Extent{T_0}$ and let $\nu$ be the pushforward of $\mu$ along $\Psi^*_{0,{\EM{{\mbb{N}}}}}$. Then $\nu$ is an invariant measure on $\Str_{{\EM{L}}_1}({\EM{{\mbb{N}}}})$ concentrated on $\Extent{T_1}$, and $\Entropy{\nu} = \Entropy{\mu}$. \end{lemma} \begin{proof} It is immediate that $\nu$ is an invariant measure on $\Str_{{\EM{L}}_1}({\EM{{\mbb{N}}}})$ concentrated on $\Extent{T_1}$, by the definition of pushforward and the fact that $\Psi^*_{0,{\EM{{\mbb{N}}}}}$ is a bijection between $\Extent{T_0}$ and $\Extent{T_1}$. We will next show, for each $n\in{\EM{{\mbb{N}}}}$, that there is a measure-preserving bijection between the atoms of $\mu_n$ and the atoms of $\nu_n$. By Definition~\ref{entropy-definition-finite-set}, this will establish that $h(\mu_n) = h(\nu_n)$ for each $n\in{\EM{{\mbb{N}}}}$, and so $\Entropy{\mu} = \Entropy{\nu}$. Let $n\in{\EM{{\mbb{N}}}}$. If ${\EM{\mathcal{M}}}\in\Extent{T_0}_{\EM{{\mbb{N}}}}$ then ${\EM{\mathcal{M}}}|_{[n]} \in \Extent{T_0}_{[n]}$, because $\mu$ concentrates on $\Extent{T_0}_{\EM{{\mbb{N}}}}$ and $T_0$ is $\Pi_1$. Hence the measure $\mu_n$ concentrates on $\Extent{T_0}_{[n]}$. Similarly, $\nu_n$ concentrates on $\Extent{T_1}_{[n]}$. Therefore every atom of $\mu_n$ is in $\Extent{T_0}_{[n]}$ and every atom of $\nu_n$ is in $\Extent{T_1}_{[n]}$. Write $\nu^*_n$ for the pushforward of $\mu_n$ along $\Psi^*_{0,[n]}$. It is clear that $\Psi^*_{0, [n]}$ is a measure-preserving bijection between the atoms of $\mu_n$ and the atoms of $\nu^*_n$. Let ${\EM{\mathcal{B}}}$ be an atom of $\nu_n$; we have ${\EM{\mathcal{B}}} \in \Extent{T_1}_{[n]}$ because $\nu_n({\EM{\mathcal{B}}}) > 0$. We will show that $\nu_n({\EM{\mathcal{B}}}) = \nu^*_n({\EM{\mathcal{B}}})$. Let ${\EM{\mathcal{A}}} = \Psi_{1, [n]}^*({\EM{\mathcal{B}}})$. For ${\EM{\mathcal{N}}}\in\Extent{T_1}_{\EM{{\mbb{N}}}}$, by Definition~\ref{interdefinability} we have ${\EM{\mathcal{N}}}|_{[n]} = {\EM{\mathcal{B}}}$ if and only if $\Psi_{1, {\EM{{\mbb{N}}}}}^*({\EM{\mathcal{N}}})|_{[n]} = {\EM{\mathcal{A}}}$. By this fact, the definitions of $\mu_n$, $\nu_n$, and pushforward, the surjectivity of $\Psi^*_{1, {\EM{{\mbb{N}}}}}$, and the fact that $\mu$ and $\nu$ are concentrated on $\Extent{T_0}_{\EM{{\mbb{N}}}}$ and $\Extent{T_1}_{\EM{{\mbb{N}}}}$ respectively, we have \begin{eqnarray*} \nu_n({\EM{\mathcal{B}}}) & = & \nu(\{{\EM{\mathcal{N}}} \in \Extent{T_1}_{\EM{{\mbb{N}}}} \,:\, {\EM{\mathcal{N}}}|_{[n]} = {\EM{\mathcal{B}}}\})\\ & = & \nu(\{{\EM{\mathcal{N}}} \in \Extent{T_1}_{\EM{{\mbb{N}}}} \,:\, \Psi_{1, {\EM{{\mbb{N}}}}}^*({\EM{\mathcal{N}}})|_{[n]} = {\EM{\mathcal{A}}}\})\\ & = & \mu(\{{\EM{\mathcal{M}}} \in \Extent{T_0}_{\EM{{\mbb{N}}}} \,:\, {\EM{\mathcal{M}}}|_{[n]} = {\EM{\mathcal{A}}}\})\\ & = & \mu_n({\EM{\mathcal{A}}}). \end{eqnarray*} But $\mu_n({\EM{\mathcal{A}}}) = \nu_n^*({\EM{\mathcal{B}}})$ by the definition of pushforward. Hence $\nu_n({\EM{\mathcal{B}}}) = \nu_n^*({\EM{\mathcal{B}}})$, as desired. \end{proof} Non-redundant invariant measures are defined only for relational languages. Hence towards showing, in Proposition~\ref{nonredundant-entropy-function}, that every entropy function is the entropy function of a non-redundant invariant measure, we need a quantifier-free interdefinition between the empty theory in an arbitrary language and a particular $\Pi_1$ theory in a relational language. There is a standard such interdefinition that maps every function symbol to a relation symbol representing the graph of the function, but it entails an increase in arity, as each function symbol of arity $\ell$ is mapped to a relation symbol of arity $\ell+1$. While this standard interdefinition would suffice for our purposes in Proposition~\ref{nonredundant-entropy-function}, here we provide a more parsimonious interdefinition, to highlight a connection with what is known about invariant measures for languages containing function symbols. Namely, using results from \cite{AFPcompleteclassification}, in Lemmas \ref{AFP-consequence} and \ref{atomic definition with functions} we show how to avoid increasing the arity of symbols, by providing a quantifier-free interdefinition that replaces each function symbol of arity $\ell$ with finitely many relation symbols, each of arity~$\ell$. As we will see in Lemma~\ref{AFP-consequence}, every invariant measure is concentrated on structures in which every function is a ``selector'', sometimes called a ``choice function'', i.e., a function for which the output is always one of the inputs. For example, the only unary selector is the identity function. We consider constant symbols to be $0$-ary function symbols; observe that no constant is a selector. \begin{definition} \label{Thfun-def} Let $f \in {\EM{L}}_0$ be a function symbol, and let $\ell$ be the arity of $f$. Define $\theta_f$ to be the sentence \[ (\forall x_0, \dots, x_{\ell-1}) \Big( \bigvee_{i \in [\ell]} f(x_0, \dots, x_{\ell-1}) = x_i \Bigl ), \] asserting that $f$ is a \emph{selector}, and define the $\Pi_1$ ${\EM{L}}_0$-theory \[ \mathrm{Th}_\mathrm{sel}({\EM{L}}_0) {\EM{\ :=\ }} \{\theta_f \,:\, f \in{\EM{L}}_0\text{ is a function symbol}\} . \] \end{definition} \begin{lemma} \label{AFP-consequence} If $\mu$ is an invariant measure on $\Str_{{\EM{L}}_0}({\EM{{\mbb{N}}}})$ then $\mu(\Extent{\mathrm{Th}_\mathrm{sel}({\EM{L}}_0)}) = 1$. \end{lemma} \begin{proof} Let $\nu$ be an ergodic invariant measure on $\Str_{{\EM{L}}_0}({\EM{{\mbb{N}}}})$. By \cite[Lemma~2.4]{AFPcompleteclassification}, \[ \Th(\nu){\EM{\ :=\ }} \{ \varphi \in \Lomega{\EM{\omega}}(L) \,:\, \nu(\Extent{\varphi}) = 1\}, \] is a complete deductively-closed first-order theory. For a function symbol $f\in {\EM{L}}_0$, if $\neg \theta_f \in \Th(\nu)$ then $\Th(\nu)$ has non-trivial definable closure, contradicting \cite[Proposition~6.1]{AFPcompleteclassification}, and so we must have $\theta_f \in \Th(\nu)$. Therefore $\nu(\Extent{\mathrm{Th}_\mathrm{sel}({\EM{L}}_0)}) = 1$. Because $\mu$ is a mixture of ergodic invariant measures $\nu$, we also have $\mu(\Extent{\mathrm{Th}_\mathrm{sel}({\EM{L}}_0)}) = 1$. \end{proof} We now provide the desired quantifier-free interdefinition. \begin{lemma} \label{atomic definition with functions} Suppose ${\EM{L}}_1$ is the relational language consisting of \begin{itemize} \item each relation symbol in ${\EM{L}}_0$, along with \item a relation symbol $E_{f, i}$ for each function symbol $f \in {\EM{L}}_0$ and $i\in [\ell]$, where $\ell$ is the arity of $f$. \end{itemize} Let $T$ be the $\Pi_1$ ${\EM{L}}_1$-theory consisting of the sentences \[ (\forall x_0, \dots, x_{\ell-1}) \bigvee_{i \in [\ell]}E_{f, i}(x_0, \dots, x_{\ell-1}), \] and \[ (\forall x_0, \dots, x_{\ell-1}) \bigwedge_{i < j \in [\ell]}\neg \bigl(E_{f, i}(x_0, \dots, x_{\ell-1}) \EM{\wedge} E_{f, j}(x_0, \dots, x_{\ell-1})\bigr), \] for each function symbol $f \in {\EM{L}}_0$, where $\ell$ is the arity of $f$. Then there is a quantifier-free interdefinition $(\Psi_0,\Psi_1)$ between $\mathrm{Th}_\mathrm{sel}({\EM{L}}_0)$ and $T$. \end{lemma} \begin{proof} For each relation symbol $R\in{\EM{L}}_0$, let $\Psi_1(R(x_0, \dots, x_{k-1}))$ be the formula $R(x_0, \dots, x_{k-1})$, where $k$ is the arity of $R$. For each function symbol $f\in {\EM{L}}_0$ and $i \in [\ell]$, let $\Psi_1(E_{f, i}(x_0, \dots, x_{\ell-1}))$ be the formula $f(x_0, \dots, x_{\ell-1}) = x_i$, where $\ell$ is the arity of $f$. It is easy to define the analogous map $\Psi_0$ and check that $(\Psi_0, \Psi_1)$ is a quantifier-free interdefinition between $\mathrm{Th}_\mathrm{sel}({\EM{L}}_0)$ and $T$. \end{proof} Next we provide a quantifier-free interdefinition which, combined with the previous results, will allow us to prove Proposition~\ref{nonredundant-entropy-function}. \begin{lemma} \label{Reduction to non-redundant} Suppose ${\EM{L}}_0$ is relational. Then there is a countable relational language ${\EM{L}}_1$ and a quantifier-free interdefinition $(\Psi_0, \Psi_1)$ between the empty ${\EM{L}}_0$-theory and the $\Pi_1$ ${\EM{L}}_1$-theory $\mathrm{Th}_\mathrm{nr}({\EM{L}}_1)$. \end{lemma} \begin{proof} For each relation symbol $R\in {\EM{L}}_0$ and equivalence relation $E$ on $[k]$, where $k\in{\EM{{\mbb{N}}}}$ is the arity of $R$, we define the following. Let $\ell$ be the number of $E$-equivalence classes. Let $f_E\colon [k] \to [k]$ send each $i\in[k]$ to the least element of its $E$-equivalence class, and let $\<y^E_0, \dots, y^E_{\ell-1}\>$ be the increasing enumeration of the image of $f_E$. Let $R_{E}$ be an $\ell$-ary relation symbol and let \[ \phi_{E}(x_0, \dots, x_{k-1}) {\EM{\ :=\ }} R_{E}(x_{y_0^E}, \dots, x_{y_{\ell-1}^E}) \EM{\wedge} \bigwedge_{\substack{i, j \in [k]\\ i E j}} x_i = x_j \EM{\wedge} \bigwedge_{\substack{i, j \in [k]\\ \neg(i E j)}} x_{i} \neq x_{j} . \] Define ${\EM{L}}_1$ to be the collection of all such symbols $R_{E}$. Recall the $\Pi_1$ theory $\mathrm{Th}_\mathrm{nr}({\EM{L}}_1)$, as defined in Definition~\ref{Thnr-def}. Define the map $\Psi_0$ on atomic ${\EM{L}}_0$-formulas by \[ \Psi_0\bigl(R(x_0, \dots, x_{k-1})\bigr) {\EM{\ :=\ }} \bigvee_{\substack{E\text{~is an equivalence}\\\text{relation on }[k]} } \phi_{E}(x_0, \dots, x_{k-1}), \] for each $R\in{\EM{L}}_0$, where $k$ is the arity of $R$. Define the map $\Psi_1$ on atomic ${\EM{L}}_1$-formulas by \[ \Psi_1\bigl(R_E(x_0, \dots, x_{\ell-1})\bigr) {\EM{\ :=\ }} R(x_{f_E(0)}, \dots, x_{f_E(k-1)}) \] for each $R_E\in{\EM{L}}_1$, where $\ell$ is the arity of $R_E$. Extend $\Psi_0$ and $\Psi_1$ to all formulas in $\Lomega{\EM{\omega}}({\EM{L}}_0)$ and $\Lomega{\EM{\omega}}({\EM{L}}_1)$, respectively, in the natural way. Observe that for any ${\EM{\mathcal{M}}}\in \Str_{L_0}({\EM{{\mbb{N}}}})$, the $L_1$-structure $\Psi^*_{0,{\EM{{\mbb{N}}}}}({\EM{\mathcal{M}}})$ is non-redundant, and conversely, every non-redundant structure in $\Str_{L_1}({\EM{{\mbb{N}}}})$ is in the image of $\Psi^*_{0,{\EM{{\mbb{N}}}}}$. One can check that $(\Psi_0, \Psi_1)$ is a quantifier-free interdefinition between the empty ${\EM{L}}_0$-theory and $\mathrm{Th}_\mathrm{nr}({\EM{L}}_1)$. \end{proof} We can now show that every entropy function is the entropy function of some non-redundant invariant measure. \begin{proposition} \label{nonredundant-entropy-function} Let $\mu$ be an invariant measure on $\Str_{{\EM{L}}_0}({\EM{{\mbb{N}}}})$. Then there is a countable relational language ${\EM{L}}_1$ and a non-redundant invariant measure $\nu$ on $\Str_{{\EM{L}}_1}({\EM{{\mbb{N}}}})$ such that $\Entropy{\mu} = \Entropy{\nu}$. Further, if $L_0$ is finite and of maximum arity $k$, then so is $L_1$. \end{proposition} \begin{proof} Recall the $\Pi_1$ $L_0$-theory $\mathrm{Th}_\mathrm{sel}({\EM{L}}_0)$ from Definition~\ref{Thfun-def}. First observe by Lemma~\ref{AFP-consequence} that $\mu(\Extent{\mathrm{Th}_\mathrm{sel}({\EM{L}}_0)}) = 1$. By Lemma~\ref{atomic definition with functions} there is a quantifier-free interdefinition $(\Psi_0, \Psi_1)$ between $\mathrm{Th}_\mathrm{sel}({\EM{L}}_0)$ and some $\Pi_1$ theory $T$ in a countable relational language $L'$. When $L_0$ is finite and of maximum arity $k$, so is $L'$. By Lemma~\ref{Reduction to non-redundant} there is a countable relational language ${\EM{L}}_1$, and a quantifier-free interdefinition $\Theta = (\Theta_0, \Theta_1)$ between the empty $L'$-theory and the $\Pi_1$ ${\EM{L}}_1$-theory $\mathrm{Th}_\mathrm{nr}$. It is easy to see from Definition~\ref{interdefinability} that $\Theta$ is also a quantifier-free interdefinition between $T$ and its image $\Theta_0(T)$, which is a $\Pi_1$ ${\EM{L}}_1$-theory. Again, when $L'$ is finite and of maximum arity $k$, so is $L_1$. By two applications of Lemma~\ref{Interdefinitions-preserve-entropy}, there is an invariant measure $\nu$ on $\Str_{L_1}({\EM{{\mbb{N}}}})$ that is concentrated on $\Extent{{\Theta_0}(T)}$ and that satisfies $\Entropy{\mu} = \Entropy{\nu}$. Finally, $\nu$ is non-redundant because ${\Theta_0}(T)$ contains $\mathrm{Th}_\mathrm{nr}$. \end{proof} \section{Invariant measures arising from extended $L$-hypergraphons} \label{section-finite} In this section we study the growth of entropy functions of non-redundant invariant measures for a relational language $L$ whose relation symbols all have the same arity $k \ge 1$. By a variant of the Aldous--Hoover--Kallenberg theorem, these invariant measures are precisely the ones that arise as the distribution of a certain random $L$-structure $G({\EM{{\mbb{N}}}}, W)$, where $W$ is a type of measurable function called an \emph{extended $L$-hypergraphon}. In Theorem~\ref{Entropy-from-hypergraphon} we express the growth of the entropy function of $G({\EM{{\mbb{N}}}}, W)$ in terms of $W$. This result generalizes a theorem of Aldous \cite[Chapter~15]{MR883646} and Janson \cite[Theorem~D.5]{MR3043217}, and our argument mirrors that of Janson. As an immediate consequence of Theorem~\ref{Entropy-from-hypergraphon} and the Aldous--Hoover--Kallenberg theorem, we see in Corollary~\ref{finite-relational-cor} that when $L$ is finite, the entropy function of any non-redundant invariant measure for ${\EM{L}}$ is $O(n^k)$. \emph{For the rest of this section, fix $k\ge1$ and let ${\EM{L}}$ be a countable relational language (possibly infinite) all of whose relations have arity $k$.} We will define an extended $L$-hypergraphon to be a probability kernel from $[0,1]^{\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)}$ to a space of quantifier-free $k$-types, satisfying a specific coherence condition. \begin{definition} A \defn{complete non-redundant quantifier-free ${\EM{L}}$-type} with free variables $x_0, \dots, x_{k-1}$ is a maximal consistent collection of atomic formulas or negations of atomic formulas containing $\{ x_i \neq x_j \,:\, i < j < k\}$, and whose free variables are all contained in $\{x_0, \dots, x_{k-1}\}$. Let $\EM{\mathfrak{qf}^{\mathrm{nr}}}_{\EM{L}}$ be the space of complete non-redundant quantifier-free ${\EM{L}}$-types with free variables $x_0, \dots, x_{k-1}$, where a subbasic clopen set consists of complete non-redundant quantifier-free ${\EM{L}}$-types containing a given atomic formula or negated atomic formula all of whose free variables are contained in $\{x_0, \dots, x_{k-1}\}$. \end{definition} Note that every complete non-redundant quantifier-free ${\EM{L}}$-type in the above sense implies a complete quantifier-free $\Lomega{\EM{\omega}_1}({\EM{L}})$-type. In the next definition, we introduce the notion of an extended ${\EM{L}}$-hypergraphon. As we describe below, this generalizes the standard notion of a hypergraphon \cite[\S23.3]{MR3012035}, which is a higher-arity version of a graphon. The definition of extended $L$-hypergraphon involves a coherence condition, specified in terms of an action of $\Sym{k}$. In Definition~\ref{zeta-def}, we describe how an extended ${\EM{L}}$-hypergraphon gives rise to a distribution on ${\EM{L}}$-structures by determining, for each $k$-tuple of elements, the distribution on its quantifier-free type. The coherence condition ensures that the order in which the $k$-tuple is specified does not affect the resulting distribution on its quantifier-free type. Consider the action of $\Sym{k}$ on $\EM{\mathfrak{qf}^{\mathrm{nr}}}_{\EM{L}}$ given by \[ \sigma \cdot u = \{\varphi(x_{\sigma(i_0)}, \ldots, x_{\sigma(i_{\ell-1})}) \,:\, \varphi(x_{i_0}, \ldots, x_{i_{\ell-1}})\in u \textrm{~and~} \ell \le k\} \] for $\sigma\in\Sym{k}$ and $u \in \EM{\mathfrak{qf}^{\mathrm{nr}}}_{\EM{L}}$. Note that this action of $\Sym{k}$ extends to ${\EM{\mathcal{P}}}(\EM{\mathfrak{qf}^{\mathrm{nr}}}_{\EM{L}})$ in the natural way, namely, \[ (\sigma \cdot \nu) (B) = \nu(\{\sigma \cdot u \,:\, u\in B\}) \] for $\nu \in {\EM{\mathcal{P}}}(\EM{\mathfrak{qf}^{\mathrm{nr}}}_{\EM{L}})$ and Borel $B \subseteq \EM{\mathfrak{qf}^{\mathrm{nr}}}_{\EM{L}}$. \begin{definition} \label{def-extended-hypergraphon} An \defn{extended ${\EM{L}}$-hypergraphon} is a measurable map \[ W\colon[0,1]^{\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)} \to {\EM{\mathcal{P}}}(\EM{\mathfrak{qf}^{\mathrm{nr}}}_{\EM{L}}) \] such that for any $\<x_F\>_{F\in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)}\in [0,1]^{\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)}$ and $\sigma \in \Sym{k}$, \[ W(\<x_{\sigma(F)}\>_{F\in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)}) = \sigma \cdot W(\<x_F\>_{F\in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)}). \] \end{definition} The next technical lemma and definitions show how an extended $L$-hyper\-graphon $W$ gives rise to a random non-redundant $L$-structure $G({\EM{{\mbb{N}}}}, W)$ whose distribution is an invariant measure on $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$. The following is a special case of a standard result from probability theory about the randomization of a kernel. \begin{lemma}[\textnormal{see {\cite[Lemma~3.22]{MR1876169}}}] \label{Hypergraphon to AH representation} Let $W\colon[0,1]^{\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)} \to {\EM{\mathcal{P}}}(\EM{\mathfrak{qf}^{\mathrm{nr}}}_{\EM{L}})$ be an extended ${\EM{L}}$-hypergraphon. There is a measurable function $W^*\colon[0,1]^{\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)}\times [0,1] \to \EM{\mathfrak{qf}^{\mathrm{nr}}}_{\EM{L}}$ such that whenever $\zeta$ is a uniform random variable in $[0,1]$, then for all $t \in [0,1]^{\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)}$, the random variable $W^*(t, \zeta)$ has distribution $W(t)$. \end{lemma} Definition~\ref{hat-def} introduces notation that we will use throughout this section and the next. \begin{definition} \label{hat-def} Let ${\mathfrak{n}} \in {\EM{{\mbb{N}}}} \cup \{{\EM{{\mbb{N}}}}\}$, and suppose ${\mathfrak{n}}\geq k$ (when ${\mathfrak{n}} \in {\EM{{\mbb{N}}}}$). Let $J\in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_k({\mathfrak{n}})$, and let $\tau_J\colon [k] \to J$ be the unique increasing bijection from $[k]$ to $J$. Define $\hat{X}_{J}$ to be the sequence $\<X_{\tau_J(F)}\>_{F \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)}$ consisting of terms of the form $X_I$, where $I$ is from $\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}({\mathfrak{n}})$. \end{definition} Recall that $L$ consists of relation symbols all of arity $k$, and so in order to define a non-redundant $L$-structure, it suffices to describe the complete quantifier-free type of every strictly increasing $k$-tuple in the structure. \begin{definition} \label{zeta-def} Let $W$ be an extended ${\EM{L}}$-hypergraphon and ${\mathfrak{n}} \in {\EM{{\mbb{N}}}} \cup \{{\EM{{\mbb{N}}}}\}$, and suppose ${\mathfrak{n}}\geq k$ (when ${\mathfrak{n}} \in {\EM{{\mbb{N}}}}$). \begin{itemize} \item Define $M({\mathfrak{n}}, W)\colon [0, 1]^{\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{\leq k}({\mathfrak{n}})} \to \Str_{\EM{L}}({\mathfrak{n}})$ to be the map such that for all $\<x_D\>_{D\in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{\leq k}({\mathfrak{n}})} \in [0, 1]^{\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{\leq k}({\mathfrak{n}})}$ and every $J\in\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_k({\mathfrak{n}})$, the quan\-tifier-\linebreak{}free type of the tuple $\<\tau_J(0), \ldots, \tau_J(k-1)\>$ in the $L$-structure \linebreak $M({\mathfrak{n}}, W)(\<x_D\>_{D \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{\le k}({\mathfrak{n}})})$ is $W^*(\hat{x}_{J}, x_J)$. \item Let $\< \zeta_D \>_{D\in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{\leq k}({\mathfrak{n}})}$ be an i.i.d.\ sequence of uniform random \linebreak variables in $[0, 1]$. Define $G({\mathfrak{n}}, W)$ to be the random ${\EM{L}}$-structure $M({\mathfrak{n}}, W)(\< \zeta_D \>_{D\in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{\leq k}({\mathfrak{n}})})$. \item For $J \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_k({\mathfrak{n}})$, define the random variable $E^W_J {\EM{\ :=\ }} W^*(\hat{\zeta}_{J}, \zeta_J)$. \end{itemize} \end{definition} In summary, $G({\mathfrak{n}}, W)$ is the random $L$-structure with underlying set $[{\mathfrak{n}}]$ whose quantifier-free $k$-types are given by the random variables $E^W_J$ for $J \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_k({\mathfrak{n}})$. It is easy to check that the distribution of $G({\mathfrak{n}},W)$ does not depend on the specific choice of i.i.d.\ uniform $\< \zeta_D \>_{D\in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{\leq k}({\mathfrak{n}})}$ or on the specific function $W^*$ satisfying Lemma~\ref{Hypergraphon to AH representation}. Observe that for an extended $L$-hypergraphon $W$, the distribution of $G({\EM{{\mbb{N}}}}, W)$ is a probability measure on $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$ that is invariant because \linebreak $\< \zeta_D \>_{D\in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{\leq k}({\EM{{\mbb{N}}}})}$ is i.i.d., and is non-redundant because $W^*$ takes values in $\EM{\mathfrak{qf}^{\mathrm{nr}}}_{\EM{L}}$. In fact, as Theorem~\ref{main-hypergraphon-theorem} asserts, every non-redundant invariant measure on $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$ arises from some $W$ in this way; this result is a variant of the usual Aldous--Hoover--Kallenberg theorem, and is a specialization of \cite[Theorem~2.37]{AckermanAutM}. \begin{theorem}[{Aldous--Hoover--Kallenberg theorem}] \label{main-hypergraphon-theorem} A non-redun\-dant probability measure $\mu$ on $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$ is invariant under the action of $\Sym{{\EM{{\mbb{N}}}}}$ if and only if $\mu$ is the distribution of $G({\EM{{\mbb{N}}}}, W)$ for some extended ${\EM{L}}$-hypergraphon $W$. \end{theorem} For more details on the usual Aldous--Hoover--Kallenberg theorem, which is stated in terms of exchangeable arrays, see \cite[Chapter~7]{MR2161313} and the historical notes to that chapter. There is an essentially equivalent statement, in terms of hypergraphons, that provides a correspondence between hypergraphons (of arity $k$) and ergodic invariant measures on the space of $k$-uniform graphs (see \cite{MR2463439} and \cite{MR2426176}). The version that we have stated above in Theorem~\ref{main-hypergraphon-theorem} provides an analogous correspondence between extended $L$-hypergraphons and non-redundant invariant measures on $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$. Note that, unlike those invariant measures that arise from hypergraphons, the ones arising in Theorem~\ref{main-hypergraphon-theorem} can in general be mixtures of ergodic invariant measures, and can be concentrated on structures that include multiple relations (though finitely many, of the same arity) which need not be symmetric. In fact, we could have defined extended ${\EM{L}}$-hypergraphons, and then described the associated sampling procedure $G({\EM{{\mbb{N}}}}, W)$ and stated the Aldous--Hoover--Kallenberg theorem, for the even more general case that does not stipulate non-redundance. However, the invariant measure resulting from the sampling procedure would then be sensitive to measure $0$ changes in the extended ${\EM{L}}$-hypergraphon, unlike the situation for standard hypergraphons or the usual Aldous--Hoover--Kallenberg representation, and so we have restricted our definitions and results to the non-redundant case. Using the machinery we have developed above, we can now prove the main theorems of this section. By Theorem~\ref{main-hypergraphon-theorem}, in order to study the entropy function of a non-redundant invariant measure on $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$, we may ask for a suitable extended $L$-hypergraphon $W$ and then analyze the entropy function of $G({\EM{{\mbb{N}}}}, W)$. In Theorem~\ref{Entropy-from-hypergraphon}, we consider such a $G({\EM{{\mbb{N}}}}, W)$, and express the leading term of its entropy function as a function of $W$. As a consequence, in Corollary~\ref{finite-relational-cor}, when $L$ is finite we obtain bounds on the entropy function of a non-redundant invariant measure for $L$. The proof of Theorem~\ref{Entropy-from-hypergraphon} closely follows that of \cite[Theorem~D.5]{MR3043217}. \begin{theorem} \label{Entropy-from-hypergraphon} Fix an extended ${\EM{L}}$-hypergraphon $W$. Suppose that \linebreak $C {\EM{\ :=\ }} \int \EntropyFin{W(\hat{z}_{[k]})}\ d\lambda(\hat{z}_{[k]})$ is finite. Then \[ \lim_{n \to \infty} \frac{\Entropy{G({\EM{{\mbb{N}}}}, W)}(n)}{|\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{k}(n)|} = \int \EntropyFin{W(\hat{z}_{[k]})}\ d\lambda(\hat{z}_{[k]}). \] In particular, $\Entropy{G({\EM{{\mbb{N}}}}, W)}(n) = C n^k + o(n^k)$ for some constant $C$. \end{theorem} \begin{proof} Observe that for all $n \in {\EM{{\mbb{N}}}}$, we have $\Entropy{G({\EM{{\mbb{N}}}}, W)}(n) = h(G(n, W))$. We first show a lower bound on $\EntropyFin{G(n, W)}$ for all $n\ge k$. For $J \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_k(n)$, the random variables $E^W_J$ are independent conditioned on $\<\zeta_I\>_{I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}$ and so, as the conditional entropy of conditionally independent random variables is additive, we have \begin{align*} \RandomEntropyFinbig{G(n, W) \ \big|\ \< \zeta_I\>_{I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}} \ &=\ \sum_{J \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_k(n)} \RandomEntropyFinbig{E^W_J \ \big|\ \<\zeta_I\>_{I \in\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}} \\ \ &=\ \sum_{J \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_k(n)} \RandomEntropyFinbig{{\EM{\mathbb{E}}}\bigl(W(\hat{\zeta}_{J}) \ \big|\ \<\zeta_I\>_{I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}\bigr)}\\ \ &= \ \sum_{J \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_k(n)} \RandomEntropyFin{W(\hat{\zeta}_{J})} \end{align*} a.s., where the last equality follows from the fact that every random variable in the sequence $\hat{\zeta}_{J}$ occurs within $\< \zeta_I\>_{I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}$. By Lemma~\ref{cond-reduces-entropy}, and then taking expectations of the first and last terms in the previous chain of equalities, we have \begin{align*} \label{Entropy sum of G(n W) given Bs} \EntropyFinbig{G(n, W)} \ &\geq \ {\EM{\mathbb{E}}}\bigl(\RandomEntropyFinbig{G(n, W) \ \big|\ \<\zeta_I\>_{I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}}\bigr) \\ \ &= \ \sum_{J \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_k(n)} {\EM{\mathbb{E}}}\bigl(\RandomEntropyFin{W(\hat{\zeta}_{J})}\bigr). \end{align*} Because the distribution of the random variable $\hat{\zeta}_{J}$ is the same for all $J \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{k}(n)$, we have \begin{align*} \sum_{J \in\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_k(n)} {\EM{\mathbb{E}}}\bigl(\RandomEntropyFin{W(\hat{\zeta}_{J})}\bigr) \ &= \ |\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{k}(n)| \cdot {\EM{\mathbb{E}}}\bigl(\RandomEntropyFin{W(\hat{\zeta}_{[k]})}\bigr) \\ \ &= \ |\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{k}(n)| \cdot \int \EntropyFin{W(\hat{z}_{[k]})} \ d\lambda(\hat{z}_{[k]}). \end{align*} This equation, together with the previous inequality, yields \[ \frac{\EntropyFin{G(n, W)}}{|\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{k}(n)|} \ \geq \ \int \EntropyFin{W(\hat{z}_{[k]})}\ d\lambda(\hat{z}_{[k]}). \] Now we show an upper bound on $\EntropyFin{G(n, W)}$ for all $n\ge k$. Let $r \in {\EM{{\mbb{N}}}}$ be positive. For $I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)$, define $Y_I {\EM{\ :=\ }} \lfloor r \cdot \zeta_I \rfloor$, so that $Y_I = \ell$ precisely when $\frac{\ell}{r} \le \zeta_I < \frac{\ell + 1}{r}$. Then by Lemma~\ref{chainrule}, we have \begin{align*} \label{Entropy sum of G(n W) given Bs} \EntropyFinbig{G(n, W)} \ &\leq\ \EntropyFinbig{G(n, W), \<Y_I\>_{I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}} \\ \ &=\ \EntropyFinbig{\<Y_I\>_{I \in\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}} + {\EM{\mathbb{E}}}\Bigl(\RandomEntropyFinbig{G(n, W) \ \big|\ \<Y_I\>_{I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}}\Bigr). \end{align*} The first term in this last expression is straightforward to calculate; we have \[ \EntropyFinbig{\<Y_I\>_{I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}} = \bigl |\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)\bigr | \cdot \log_2(r), \] as the $Y_I$ are uniformly distributed on $[r]$ and independent (because the $\zeta_I$ are i.i.d.\ uniform on $[0,1]$). We next calculate the second term; we will show that \begin{align*} {\EM{\mathbb{E}}}\Bigl(\RandomEntropyFinbig{G(n, W) \ |\ \<Y_I\>_{I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}}\Bigr) \ &=\ |\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{k}(n)|\cdot \int \EntropyFin{W_r(\hat{z}_{[k]})}\ d\lambda(\hat{z}_{[k]}). \end{align*} By (the standard conditional extension of) Lemma~\ref{joint-entropy-less-than-sum-entropy}, we have \begin{align*} \RandomEntropyFinbig{G(n, W) \ \big|\ \<Y_I\>_{I \in\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}} \ &\leq\ \sum_{J \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_k(n)}\RandomEntropyFinbig{E^W_J \ \big|\ \<Y_I\>_{I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}} \\ \ &=\ \sum_{J \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_k(n)}\RandomEntropyFinbig{E^W_J \ \big|\ \hat{Y}_{J}} \end{align*} a.s., where the last equality follows from the fact that the only random variables in $\<\zeta_I\>_{I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}$ on which a given $E^W_J$ depends are those among $\hat{\zeta}_J$. Given a function $\alpha\colon \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k) \to [r]$, define \begin{align*} w_r(\alpha) \ {\EM{\ :=\ }}\ r^{|\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)|} \cdot \int W(\hat{z}_{[k]}) \cdot \prod_{F \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)} {\mathbbm{1}}_{\bigl[\frac{\alpha(F)}{r},\frac{\alpha(F) + 1}{r}\bigr)}(z_{F})\ d \lambda(\hat{z}_{[k]}), \end{align*} where ${\mathbbm{1}}_S$ denotes the characteristic function of a set $S$ (in this case, a half-open interval). Observe that \begin{align*} w_r(\alpha) \ = \ {\EM{\mathbb{E}}}\Big(W(\hat{\zeta}_{[k]}) \ \Big|\ \bigwedge_{F \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)}Y_{F} = \alpha(F)\Big) \end{align*} a.s. (We may form this conditional expectation because the set of signed measures on $\EM{\mathfrak{qf}^{\mathrm{nr}}}_{\EM{L}}$ is a Banach space, and hence integration on this space is well-defined.) For $\hat{z}_{[k]} \in [0, 1]^{\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)}$, define $\beta_{\hat z_{[k]}}\colon\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k) \to [r]$ by $\beta_{\hat z_{[k]}}(F) = \lfloor r \cdot z_F \rfloor$ for $F \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)$, and let $W_r(\hat{z}_{[k]}) {\EM{\ :=\ }} w_r(\beta_{\hat z_{[k]}})$. Observe that while we have defined continuum-many instances of $\beta_{\hat z_{[k]}}$, they range over the merely finitely many functions from $\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)$ to $[r]$. Further note that $W_r$ is a step function, and that as $r\to \infty$, the function $W_r$ converges to $W$ pointwise a.e. We can think of $W_r$ as the result of discretizing $W$ along blocks of width $1/r$. Recall the function $\tau_J$ defined in Definition~\ref{hat-def}. For any $J \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_k(n)$, we have \begin{eqnarray*} {\EM{\mathbb{P}}}\Big(E^W_J \hspace*{-2pt} \in \cdot \ \Big|\ \hspace*{-5pt} \bigwedge_{F \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)} \hspace*{-10pt} Y_{\tau_J(F)} = \beta_{\hat z_{[k]}}(F)\Big) \hspace*{-2pt} \ &=&\ {\EM{\mathbb{E}}}\Big(W(\hat{\zeta}_{J}) \ \Big|\ \hspace*{-5pt} \bigwedge_{F \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)} \hspace*{-10pt} Y_{\tau_J(F)} = \beta_{\hat z_{[k]}}(F)\Big ) \\ \ &=&\ {\EM{\mathbb{E}}}\Big(W(\hat{\zeta}_{[k]}) \ \Big|\ \hspace*{-5pt} \bigwedge_{F \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)} \hspace*{-10pt} Y_{F} = \beta_{\hat z_{[k]}}(F)\Big ) \\ \ &=&\ w_r(\beta_{\hat z_{[k]}}) \end{eqnarray*} a.s., because $\tau_J$ is a bijection from $[k]$ to $J$ and the distribution of the random variable $\hat{\zeta}_{J}$ is the same as that of $\hat{\zeta}_{[k]}$. Therefore \[ \EntropyFinBig{E^W_J \in \cdot \ \Big|\ \bigwedge_{F \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)} Y_{\tau_J(F)} = \beta_{\hat z_{[k]}}(F)} = \EntropyFin{w_r(\beta_{\hat z_{[k]}})}. \] Summing both sides over all possible choices of $\beta_{\hat z_{[k]}}$, we have \[ \sum_{\alpha\colon \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k) \to [r]}\EntropyFinBig{E^W_J \in \cdot \ \Big|\ \bigwedge_{F \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)} Y_{\tau_J(F)} = \alpha(F)} = \hspace*{-7pt} \sum_{\alpha\colon\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k) \to [r]} \hspace*{-4pt} \EntropyFin{w_r(\alpha)}). \] As before, \[ \RandomEntropyFinbig{E^W_J \ \big|\ \<Y_I\>_{I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}} = \RandomEntropyFinbig{E^W_J \ \big|\ \hat{Y}_J}. \] One can directly show that \begin{align*} {\EM{\mathbb{E}}}\Bigl(\RandomEntropyFinbig{E^W_J \ \big|\ \hat{Y}_J}\Bigr) = r^{-|\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)|}\hspace*{-10pt} \sum_{\alpha\colon \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k) \to [r]} \hspace*{-3pt} \EntropyFinBig{E^W_J \in \cdot \ \Big|\ \hspace*{-5pt}\bigwedge_{F \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)} \hspace*{-8pt} Y_{\tau_J(F)} = \alpha(F)}. \end{align*} Hence \begin{align*} {\EM{\mathbb{E}}}\Bigl(\RandomEntropyFinbig{E^W_J \ \big|\ \<Y_I\>_{I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}}\Bigr) \ &=\ r^{-|\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)|} \cdot \sum_{\alpha\colon\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k) \to [r]} \EntropyFin{w_r(\alpha)} \\ \ &=\ \int \EntropyFinbig{W_r(\hat{z}_{[k]})}\ d\lambda(\hat{z}_{[k]}). \end{align*} But once again by (an extension of) Lemma~\ref{joint-entropy-less-than-sum-entropy}, we have \begin{align*} {\EM{\mathbb{E}}}\Bigl(\RandomEntropyFinbig{G(n, W) \ |\ \<Y_I\>_{I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}}\Bigr) \ &\leq\ \sum_{J \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_k(n)}{\EM{\mathbb{E}}}\Bigl(\RandomEntropyFinbig{E^W_J \ \big|\ \<Y_I\>_{I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)}}\Bigr) \\ \ &=\ |\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{k}(n)|\cdot \int \EntropyFin{W_r(\hat{z}_{[k]})}\ d\lambda(\hat{z}_{[k]}). \end{align*} Putting together our two calculations, we obtain \[ \EntropyFin{G(n, W)} \ \leq\ |\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)| \cdot \log_2(r) + |\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{k}(n)| \cdot \int \EntropyFin{W_r(\hat{z}_{[k]})}\ d\lambda(\hat{z}_{[k]}). \] Hence for each positive $r$, we have \[ \frac{\EntropyFin{G(n, W)}}{|\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{k}(n)|} \ \leq\ \frac{|\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(n)|}{|\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_k(n)|} \cdot \log_2(r) + \int \EntropyFin{W_r(\hat{z}_{[k]})}\ d\lambda(\hat{z}_{[k]}), \] and so \[ \limsup_{n \to \infty}\frac{\EntropyFin{G(n, W)}}{|\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{k}(n)|} \ \leq\ \int \EntropyFin{W_r(\hat{z}_{[k]})}\ d\lambda(\hat{z}_{[k]}). \] Letting $r \to \infty$, recall that $W_r \to W$ a.e., and so by the dominated convergence theorem, we have \[ \limsup_{n \to \infty}\frac{\EntropyFin{G(n, W)}}{|\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{k}(n)|} \ \leq\ \int \EntropyFin{W(\hat{z}_{[k]})}\ d\lambda(\hat{z}_{[k]}). \] Combining our lower and upper bounds, we have \[ \lim_{n \to \infty} \frac{\EntropyFin{G({\EM{{\mbb{N}}}}, W)}}{|\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{k}(n)|} = \int \EntropyFin{W(\hat{z}_{[k]})}\ d\lambda(\hat{z}_{[k]}), \] as desired. Finally, because $\lim_{n \to \infty} \frac{|\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_k(n)|}{n^k} = 1$, we have \[\Entropy{G({\EM{{\mbb{N}}}}, W)}(n) = Cn^k + o(n^k),\] where $C = \int \EntropyFin{W(\hat{z}_{[k] })}\ d\lambda(\hat{z}_{[k]})$. \end{proof} As a corollary, when $L$ is finite, we obtain a bound on the growth of entropy functions of non-redundant invariant measures for $L$. \begin{corollary} \label{finite-relational-cor} Suppose $L$ is finite, and let $\mu$ be a non-redundant invariant measure on $\Str_{{\EM{L}}}({\EM{{\mbb{N}}}})$. Then $\Entropy{\mu}(n) = Cn^k + o(n^k)$ for some constant $C$. In particular, $\Entropy{\mu}(n) = O(n^k)$. \end{corollary} \begin{proof} By Theorem~\ref{main-hypergraphon-theorem} there is some extended $L$-hypergraphon $W$ such that $\mu$ is the distribution of the random $L$-structure $G({\EM{{\mbb{N}}}}, W)$. Because $L$ is finite, $W$ is bounded, and so $\int \EntropyFin{W(\hat{z}_{[k]})}\ d\lambda(\hat{z}_{[k]})$ is finite. Hence the entropy function $\Entropy{\mu}$ is of the desired form by Theorem~\ref{Entropy-from-hypergraphon}. \end{proof} \section{Invariant measures sampled from a Borel hypergraph} \label{randomfree-sec} We have seen in Corollary~\ref{finite-relational-cor} that for a finite relational language ${\EM{L}}$ all of whose relation symbols have the same arity $k \ge 1$, the entropy function of an invariant measure on $\Str_{{\EM{L}}}({\EM{{\mbb{N}}}})$ is of the form $Cn^k + o(n^k)$, where $C$ is a constant depending on the invariant measure. In this section we consider the situation where $C=0$. For $k=1$, consider an invariant measure $\mu$ on $\Str_{{\EM{L}}}$, and let $W$ be an extended $L$-hypergraphon such that $\mu$ is the distribution of $G({\EM{{\mbb{N}}}}, W)$. By Theorem~\ref{Entropy-from-hypergraphon}, we have $\Entropy{\mu}(n) = C n + o(n)$, where $C = \int \EntropyFin{W(z_{\emptyset})}\ d\lambda(z_{\emptyset})$. Suppose $C=0$. Then $h(W(z_\emptyset)) = 0$ for a.e.\ $z_\emptyset$, and so $W(z_\emptyset)$ is a point mass a.e. Hence $G({\EM{{\mbb{N}}}}, W)$ is a random $L$-structure where every element of ${\EM{{\mbb{N}}}}$ has the same quantifier-free $1$-type, a.s. Therefore $\mu$ is a mixture of finitely many point masses, and $\Entropy{\mu}(n)$ is a constant that does not depend on $n$. In summary, for $k=1$, the only possible entropy functions of sublinear growth are the constant functions. Theorem~\ref{Random-free has arbitrary growth rates}, the main result of this section, states that in contrast, for $k>1$, Corollary~\ref{finite-relational-cor} is tight in the sense that for any given function $\gamma$ that is $o(n^k)$, there is some non-redundant invariant measure whose entropy function is $o(n^k)$ but grows faster than $\gamma$. As discussed in the introduction, this result is a generalization of \cite[Theorem~1.1]{MR3073488}, and the arguments in this section closely follow their proof. \emph{For the rest of this section, fix $k \ge 2$ and let ${\EM{L}}$ be the language of $k$-uniform hypergraphs, i.e., ${\EM{L}} = \{E\}$ where $E$ is a $k$-ary relation symbol.} A \emph{$k$-uniform hypergraph} is a non-redundant ${\EM{L}}$-structure satisfying \[ \bigwedge_{\sigma\in \Sym{k}} (\forall x_0, \dots, x_{k-1})\ \big(E(x_0, \dots, x_{k-1}) \leftrightarrow E(x_{\sigma(0)}, \dots, x_{\sigma(k-1)})\big), \] and we call the instantiation of $E$ its \emph{edge set}. By a \emph{Borel hypergraph}, we mean a $k$-uniform hypergraph ${\EM{\mathcal{M}}}$ whose underlying set is $[0,1]$ and such that for any atomic formula $\varphi$ in the language of hypergraphs, the set $\{{\EM{\ol{a}}} \in {\EM{\mathcal{M}}}\,:\, {\EM{\mathcal{M}}} \vDash \varphi({\EM{\ol{a}}})\}$ of realizations of $\varphi$ in ${\EM{\mathcal{M}}}$ is Borel. Any extended $L$-hypergraphon $W$ \emph{yields} a non-redundant invariant measure, namely the distribution of the random $L$-structure $G({\EM{{\mbb{N}}}}, W)$. We say that $W$ \emph{induces a Borel hypergraph} when this invariant measure can be obtained by sampling a random subhypergraph of some Borel hypergraph, as we make precise in Definition~\ref{def-induces-a-borel-hypergraph}. We will see, in Lemma~\ref{random-free-has-leading-entropy-zero}, that when $W$ induces a Borel hypergraph, it yields an invariant measure whose entropy function is $o(n^k)$. The main construction of this section, in Theorem~\ref{Random-free has arbitrary growth rates}, builds non-redundant invariant measures whose entropy functions have arbitrarily high growth within $o(n^k)$ by sampling from certain extended $L$-hypergraphons that induce Borel hypergraphs. Observe that there are only two non-redundant quantifier-free $k$-types in $L$ that are consistent with the theory of $k$-uniform hypergraphs. Let $u_\top, u_\bot \in \EM{\mathfrak{qf}^{\mathrm{nr}}}_{\EM{L}}$ be the unique non-redundant quantifier-free types containing \[ \bigwedge_{\sigma \in \Sym{k}} E(x_{\sigma(0)}, \dots, x_{\sigma(k-1)}) \] and \[ \bigwedge_{\sigma \in \Sym{k}} \neg E(x_{\sigma(0)}, \dots, x_{\sigma(k-1)}), \] respectively, and let $\delta_\top, \delta_\bot \in {\EM{\mathcal{P}}}(\EM{\mathfrak{qf}^{\mathrm{nr}}}_{\EM{L}})$ be the respective point masses concentrated on them. Recall from Definition~\ref{hat-def} that $\hat{x}_{[k]}$ denotes the tuple of variables $\<x_I\>_{I \in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)}$. \begin{definition} \label{def-induces-a-borel-hypergraph} \quad We say that an extended ${\EM{L}}$-hypergraphon \linebreak $W\colon[0,1]^{\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)} \to {\EM{\mathcal{P}}}(\EM{\mathfrak{qf}^{\mathrm{nr}}}_{\EM{L}})$ \defn{induces a Borel hypergraph} if \begin{itemize} \item for a.e.\ pair of sequences $\hat{x}_{[k]}$, $\hat{y}_{[k]}$ of elements of $[0,1]$ with $x_{\{i\}} = y_{\{i\}}$ for all $i\in[k]$, we have $W(\hat{x}_{[k]}) = W(\hat{y}_{[k]})$, and \item for a.e.\ sequence $\hat{x}_{[k]}$ of elements of $[0,1]$, the distribution $W(\hat{x}_{[k]})$ is either $\delta_\top$ or $\delta_\bot$. \end{itemize} \end{definition} It follows that an extended ${\EM{L}}$-hypergraphon $W$ induces a Borel hypergraph precisely when there is a Borel hypergraph ${\EM{\mathcal{B}}}$ such that $W(\hat{x}_{[k]})$ is a point mass concentrated on the quantifier-free type of $\<x_{\{0\}}, \dots, x_{\{k-1\}}\>$ in ${\EM{\mathcal{B}}}$. In this case we say that $W$ \emph{induces the Borel hypergraph} ${\EM{\mathcal{B}}}$. The notion of an extended $L$-hypergraphon inducing a Borel hypergraph is closely related, in the case $k=2$, to that of a graphon being random-free. A \emph{graphon} is a symmetric Borel function from $[0,1]^2$ to $[0,1]$; as described in \cite{MR2815610} and \cite{MR3043217}, it is called \emph{random-free} when it is $\{0,1\}$-valued a.e. Every graphon gives rise to a random undirected graph on ${\EM{{\mbb{N}}}}$ whose distribution is an invariant measure. For $k=2$, an extended $L$-hypergraphon $W$ that yields an invariant measure concentrated on undirected graphs can be expressed as a mixture of invariant measures, each obtained via a graphon. In the case where such a $W$ corresponds to a single graphon, $W$ induces a Borel hypergraph precisely when the corresponding graphon is random-free. The notion of a random-free graphon also essentially appeared, in the context of separate exchangeability, in work of Aldous in \cite[Proposition~3.6]{MR637937} and \cite[(14.15) and p.~133]{MR883646}, and Diaconis--Freedman \cite[(4.10)]{MR640207}. Further, Kallenberg \cite{MR1702867} describes, for all $k\ge2$, the similar notion of a \emph{simple array}; this corresponds to our notion of inducing a Borel hypergraph, in the case where the distribution of the simple array is ergodic. Random-free graphons arise as well in \cite{MR2724668} and \cite[\S6.1]{MR3515800}, which consider invariant measures that are concentrated on a given orbit of the logic action. Theorem~\ref{Entropy-from-hypergraphon} implies that any extended $L$-hypergraphon $W$ yields an invariant measure whose entropy function is $O(n^k)$. Observe that there are extended $L$-hypergraphons achieving this upper bound, i.e., that yield an invariant measure whose entropy function is $\Omega(n^k)$. For example, one can directly calculate that the Erd\H{o}s--R\'enyi\ extended $L$-hypergraphon given by the constant function \[ W_\mathrm{ER}(\hat{x}_{[k]}) = \mathrm{Uniform}(\{u_{\top}, u_\bot\}) \] satisfies $\Entropy{G({\EM{{\mbb{N}}}}, W_\mathrm{ER})}(n) = {n\choose k}$. However, we now show that this growth rate cannot be achieved for an extended $L$-hypergraphon that induces a Borel hypergraph. \begin{lemma} \label{random-free-has-leading-entropy-zero} Let $W$ be an extended $L$-hypergraphon, and suppose $W$ induces a Borel hypergraph. Then $\Entropy{G({\EM{{\mbb{N}}}}, W)}(n) = o(n^k)$. \end{lemma} \begin{proof} Because $W$ induces a Borel hypergraph, it takes the value $\delta_\top$ or $\delta_\bot$ a.e. But $\EntropyFin{\delta_{\top}} = \EntropyFin{\delta_{\bot}} = 0$, and so $\int \EntropyFin{W(\hat{z}_{[k]})}\ d\lambda(\hat{z}_{[k]}) =0$. Hence by Theorem~\ref{Entropy-from-hypergraphon}, we have $\Entropy{G({\EM{{\mbb{N}}}}, W)}(n) = o(n^k)$. \end{proof} As noted previously, Aldous \cite{MR883646} and Janson \cite{MR3043217} have versions of Theorem~\ref{Entropy-from-hypergraphon} for $k=2$. They also observe that their respective results immediately imply that a random-free graphon yields an invariant measure with entropy function that is $o(n^2)$; their proofs are similar to that of Lemma~\ref{random-free-has-leading-entropy-zero}. Their setting involves working with graphons, which yield ergodic invariant measures. Under the restriction of ergodicity, it is easily seen that the converse of Lemma~\ref{random-free-has-leading-entropy-zero} holds for $k=2$, as noted by Janson \cite[Theorem~10.16]{MR3043217}. In contrast, the converse of Lemma~\ref{random-free-has-leading-entropy-zero} itself does not hold, as our notion of extended $L$-hypergraphon allows for ones that yield non-ergodic invariant measures. For example, consider the extended $L$-hypergraphon \[ W_{\text{\larger[-2]$\blacksquare \square$}} (\hat{x}_{[k]}) = \begin{cases} \delta_{\top}& \text{ if } x_\emptyset < \frac12,\\ \delta_{\bot}& \text{ otherwise}. \end{cases} \] Define the extended $L$-hypergraphons $W_{\text{\larger[-2]$\blacksquare$}} (\hat{x}_{[k]}) = \delta_{\top}$ and $W_{\text{\larger[-2]$\square$}} (\hat{x}_{[k]}) = \delta_{\bot}$, which each induce a Borel hypergraph. The random hypergraph $G({\EM{{\mbb{N}}}}, W_{\text{\larger[-2]$\blacksquare \square$}})$ is the complete hypergraph or the empty hypergraph, each with probability $\frac12$, and so its distribution is a non-trivial mixture of the distributions of $G({\EM{{\mbb{N}}}}, W_{\text{\larger[-2]$\blacksquare$}})$ and $G({\EM{{\mbb{N}}}}, W_{\text{\larger[-2]$\square$}})$, hence a non-ergodic invariant measure. The extended $L$-hypergraphon $W_{\text{\larger[-2]$\blacksquare \square$}}$ does not induce a Borel hypergraph as it depends on the variable $x_\emptyset$, yet the entropy function of $G({\EM{{\mbb{N}}}}, W_{\text{\larger[-2]$\blacksquare \square$}})$ is $o(n^k)$. But in fact, for $k\ge3$, there is a more interesting obstruction to a converse of Lemma~\ref{random-free-has-leading-entropy-zero}, even among extended $L$-hypergraphons that yield ergodic invariant measures. The following example for $k=3$ (which is easily generalized to larger values of $k$) makes fundamental use of the variables indexed by pairs from $[3]$, and yet also yields an invariant measure whose entropy function is $o(n^3)$: \[ W_{\!\triangle}(\hat{x}_{[3]}) = \begin{cases} \delta_{\top}& \text{ if } x_{\{0,1\}}< \frac12 \text{ and } x_{\{0,2\}}< \frac12 \text{ and } x_{\{1,2\}}< \frac12, \\ \delta_{\bot}& \text{ otherwise}. \end{cases} \] The random hypergraph $G({\EM{{\mbb{N}}}},W_{\!\triangle})$ can be thought of as first building a ``virtual'' Erd\H{o}s--R\'enyi\ graph with independent $2$-edge probabilities $\frac12$, then adding a $3$-edge for each triangle existing in the graph, and then throwing away the virtual $2$-edges. (For more about this example, see \cite[p.~92]{MR2426176} and \cite[Example~23.11]{MR3012035}.) In Lemma~\ref{random-free-has-leading-entropy-zero}, we established that any extended $L$-hypergraphon $W$ that induces a Borel hypergraph is such that the entropy function of $G({\EM{{\mbb{N}}}},W)$ is $o(n^k)$. We now proceed to show that there are such $W$ for which the growth of $\Entropy{G({\EM{{\mbb{N}}}},W)}(n)$ is arbitrarily close to $n^k$. We first define a kind of ``blow up'' that creates an extended $L$-hypergraphon from a countably infinite $L$-structure. We will use this notion in Lemma~\ref{Bound on conditional entropy}. Blow ups are a standard technique for expanding a countable structure into a continuum-sized structure that has a positive-measure worth of ``copies'' of each point from the original. (See, e.g., the use of step function graphons in \cite{MR3012035}.) \begin{definition} \label{piblowup-def} Let ${\EM{\mathcal{M}}} \in \Str_{{\EM{L}}}({\EM{{\mbb{N}}}})$ and let $\pi\colon [0,1]\to {\EM{{\mbb{N}}}}$ be a Borel map such that $\lambda(\pi^{-1}(i)) > 0$ for all $i\in{\EM{{\mbb{N}}}}$. The \defn{$\pi$-blow up of ${\EM{\mathcal{M}}}$} is defined to be the extended $L$-hypergraphon $W\colon[0,1]^{\ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{<k}(k)} \to \{\delta_{\top}, \delta_{\bot}\}$ given by \[ W(\hat{x}_{[k]}) = \begin{cases} \delta_{\top}& \text{ if } {\EM{\mathcal{M}}} \vDash E\bigl(\pi(x_{\{0\}}), \pi(x_{\{1\}}), \dots, \pi(x_{\{k-1\}})\bigr),\\ \delta_{\bot}& \text{ otherwise}. \end{cases} \] \end{definition} Observe that such a $W$ induces a Borel hypergraph, as on every input it outputs either the value $\delta_\top$ or $\delta_\bot$, and it depends only on the variables $x_{\{0\}}, x_{\{1\}}, \dots, x_{\{k-1\}}$. In particular, $W$ induces the Borel hypergraph ${\EM{\mathcal{B}}}$ with underlying set $[0,1]$ given by: \[ {\EM{\mathcal{B}}} \vDash E(x_0, x_1,\dots, x_{k-1}) \] if and only if \[ {\EM{\mathcal{M}}} \vDash E\bigl(\pi(x_0), \pi(x_1), \dots, \pi(x_{k-1})\bigr). \] Hence we can think of $\pi^{-1}$ as a Borel partition of the unit interval into positive measure pieces, such that each element of ${\EM{\mathcal{M}}}$ is ``blown up'' into a piece of the partition in ${\EM{\mathcal{B}}}$, and each piece of the partition arises in this way. Next, we proceed to construct an extended $L$-hypergraphon $W$ that yields an invariant measure whose entropy function has the desired growth. Analogously to Hatami--Norine \cite{MR3073488}, we let $W$ be the blow up of a particular countably infinite structure, the \emph{Rado $k$-hypergraph}, a well-known generalization of the \emph{transversal-uniform graph} used in \cite{MR3073488}. The Rado $k$-hypergraph is the countable homogeneous-universal $k$-uniform hypergraph, namely, the unique (up to isomorphism) countable $k$-uniform hypergraph satisfying the so-called ``Alice's restaurant'' axioms. These axioms state that for any possible way of extending a finite induced subhypergraph of the Rado $k$-hypergraph by one vertex to obtain a $k$-uniform hypergraph, there is some element of the Rado $k$-hypergraph that realizes this extension. We now describe an inductive construction of an instantiation $\mathcal{R}_k$ of the Rado $k$-hypergraph, with underlying set ${\EM{{\mbb{N}}}}$. At each stage $\ell\in{\EM{{\mbb{N}}}}$ we define a finite set $A_\ell$ of new vertices, which we call \emph{generation} $\ell$, and build a $k$-uniform hypergraph $G_\ell$ with underlying set $V_\ell {\EM{\ :=\ }} \bigcup_{j\le \ell} A_j$. Stage $0$: Let $A_0 {\EM{\ :=\ }} \{0\}$ consist of a single vertex, and let $G_0$ be the empty hypergraph with vertex set $V_0 = A_0$. Stage $\ell>0$: Let $A_\ell$ consist of one new vertex $a_X$ for each subset $X$ of unordered $(k-1)$-tuples from $V_{\ell-1}$, with the elements of $A_\ell$ chosen to be the consecutive least elements of ${\EM{{\mbb{N}}}}$ not yet used. Let $G_\ell$ be the $k$-uniform hypergraph on vertex set $V_\ell = V_{\ell-1} \cup A_\ell$ whose edges are those of $G_{\ell-1}$ along with, for each such $X$ and every unordered tuple $d\in X$, an edge consisting of $a_X$ and the $k-1$ vertices in $d$. Define $\mathcal{R}_k$ to be the union of the hypergraphs $G_\ell$, i.e., the hypergraph with vertex set $\bigcup_{\ell\in{\EM{{\mbb{N}}}}} V_\ell$ and edge set $\bigcup_{\ell\in{\EM{{\mbb{N}}}}} E^{G_\ell}$. Observe that $\mathcal{R}_k$ is a $k$-uniform hypergraph with underlying set ${\EM{{\mbb{N}}}}$ that satisfies the Alice's restaurant axioms: Given a finite induced subhypergraph $D$ of $\mathcal{R}_k$, all one-vertex extensions of $D$ are realized in stage $\ell+1$, for any $\ell$ such that $V_\ell$ contains the vertices of $D$. The following lemma, which we will use in the proof of Theorem~\ref{Random-free has arbitrary growth rates}, provides a lower bound on the entropy function of a random hypergraph sampled from a blow up of $\mathcal{R}_k$ in the case where elements of $\mathcal{R}_k$ belonging to the same generation get blown up to sets of equal measure. Both the statement and proof of the lemma are directly analogous to those of \cite[Lemma~2.1]{MR3073488}. Recall from Definition~\ref{zeta-def} that $\<\zeta_D\>_{D\in \ensuremath{\mathrel{\raisebox{2pt}{\scaleobj{1.3}{\wp}}}}_{\le k}({\EM{{\mbb{N}}}})}$ is the collection of i.i.d.\ uniform random variables in $[0,1]$ in terms of which the random $L$-structure $G({\EM{{\mbb{N}}}},W)$ is defined. \begin{lemma} \label{Bound on conditional entropy} Let $\pi\colon [0,1]\to {\EM{{\mbb{N}}}}$ be a Borel map such that $\lambda(\pi^{-1}(i)) > 0$ for all $i\in{\EM{{\mbb{N}}}}$, and let $W$ be a $\pi$-blow up of $\mathcal{R}_k$. Suppose that $\lambda(\pi^{-1}(a)) = \lambda(\pi^{-1}(b))$ for all $\ell\in{\EM{{\mbb{N}}}}$ and $a,b\in A_\ell$. Then for all $n \in {\EM{{\mbb{N}}}}$ and $\rho\colon[n] \to {\EM{{\mbb{N}}}}$, we have \[ \RandomEntropyFinBig{G(n, W) \ \Big|\ \bigwedge_{j \in [n]} \pi(\zeta_{\{j\}}) \in A_{\rho(j)}} \ \geq \ {\bigl|\rho([n])\bigr| \choose {k}}, \] a.s. \end{lemma} \begin{proof} Let $S\subseteq [n]$ be maximal such that $\rho$ is injective on $S$, and write $G(S,W)$ to denote the random induced substructure of $G(n, W)$ with underlying set $S$. Then \begin{align*} \RandomEntropyFinBig{ G(n, W) \ \Big|\ \bigwedge_{j \in [n]} \pi(\zeta_{\{j\}}) \in A_{\rho(j)}} \ \geq\ \RandomEntropyFinBig{ G(S, W) \ \Big|\ \bigwedge_{j \in S} \pi(\zeta_{\{j\}}) \in A_{\rho(j)}}. \end{align*} Consider the random measure \begin{align*} \label{condprob-uniform} \tag{$\dagger$} {\EM{\mathbb{E}}}\Big( G(S, W) \ \Big|\ \bigwedge_{j \in S} \pi(\zeta_{\{j\}}) \in A_{\rho(j)}\Big). \end{align*} Since $W$ is a blow up of $\mathcal{R}_k$, the random hypergraph $G(S,W)$ is a $|S|$-element sample with replacement from $\mathcal{R}_k$, with the vertices relabeled by $S$. By the injectivity of $\rho$, the condition in \eqref{condprob-uniform} constrains the elements of $G(S,W)$ to be obtained from distinct generations of $\mathcal{R}_k$, which implies that the random distribution on hypergraphs given by \eqref{condprob-uniform} is actually sampled from $\mathcal{R}_k$ without replacement a.s. Further, each $\zeta_{\{j\}}$ is uniform, and $\pi^{-1}$ assigns sets of equal measure to vertices in $\mathcal{R}_k$ of the same generation. So \eqref{condprob-uniform} is a.s.\ the distribution of the random hypergraph $Q$ with underlying set $S$ obtained by, for each $j\in S$, uniformly selecting a vertex of $\mathcal{R}_k$ from among those in generation $\rho(j)$, and taking the edges induced from $\mathcal{R}_k$. Let $\ell \in {\EM{{\mbb{N}}}}$, and consider a subset $U\subseteq V_\ell$. Write $T$ for the set of unordered $(k-1)$-tuples from $U$. Suppose $\ell' > \ell$. For every subset $X \subseteq T$, exactly a $2^{-|T|}$-fraction of the vertices in $A_{\ell'}$ form an edge with every $(k-1)$-tuple in $X$ and with no $(k-1)$-tuple in $T \setminus X$. For a vertex $v$ selected uniformly at random from $A_{\ell'}$, let $G_{U, v}$ be the (not necessarily induced) random subhypergraph of $G_{\ell'}$ that has vertex set $U \cup \{v\}$ and whose edges are those in $G_{\ell'}$ that consist of $v$ along with $k-1$ vertices from $U$. Then $G_{U,v}$ is equally likely to be any of the hypergraphs with vertex set $U \cup \{v\}$ whose edges all include $v$. For each $j\in S$, define $S_j {\EM{\ :=\ }} \{i \in S \,:\, \rho(i) < \rho(j)\}$, and let $Q_j$ be the (not necessarily induced) random subhypergraph of $Q$ that has vertex set $S_j \cup \{j\}$ and whose edges are those in $Q$ that consist of $j$ along with $k-1$ vertices from $S_j$. Then for each $j\in S$, the random hypergraph $Q_j$ is equally likely to be any given hypergraph $F_j$ with vertex set $S_j \cup \{j\}$ all of whose edges include $j$. Further, for any choice of hypergraphs $\{F_j\,:\, j\in S\}$ as above, the events $Q_j = F_j$ for $j\in S$ are independent because the random variables $\zeta_{\{j\}}$ are independent. Now, for any hypergraph $F$ with underlying set $S$, we can write $F$ as the union, over $j\in S$, of the subhypergraph of $F$ that has vertex set $S_j \cup \{j\}$ and whose edges are those in $F$ consisting of $j$ along with $k-1$ vertices from $S_j$. We therefore see that $Q$ is equally likely to be any hypergraph on vertex set $S$. In summary, the random distribution \eqref{condprob-uniform} is a.s.\ the uniform measure on $k$-uniform hypergraphs with underlying set $S$. Hence \[ \RandomEntropyFinBig{ G(S, W) \ \Big|\ \bigwedge_{j \in S} \pi(\zeta_{\{j\}}) \in A_{\rho(j)}} \ =\ {{|S|}\choose {k}} \] a.s., establishing the lemma. \end{proof} We now prove the main result of this section, which asserts that the entropy function of an invariant measure can have arbitrarily large growth rate within $o(n^k)$. This result is a higher-arity version of \cite[Theorem~1.1]{MR3073488}, and its proof proceeds via the same steps. We include the proof here (with appropriately modified parameters and notation) for completeness. \begin{theorem} \label{Random-free has arbitrary growth rates} Suppose $\gamma\colon {\EM{{\mbb{N}}}} \to [0,1]$ is a function such that $\lim_{n\to \infty} \gamma(n) = 0$. Then there is an extended $L$-hypergraphon $W$ that induces a Borel hypergraph and is such that $\Entropy{G({\EM{{\mbb{N}}}}, W)} (n) = o(n^k)$ and $\Entropy{G({\EM{{\mbb{N}}}}, W)} (n) = \Omega(\gamma(n) \cdot n^k)$. \end{theorem} \begin{proof} We will define a Borel map $\pi\colon [0,1]\to {\EM{{\mbb{N}}}}$ satisfying $\lambda(\pi^{-1}(i)) > 0$ for all $i\in{\EM{{\mbb{N}}}}$ in such a way that the $\pi$-blow up of $\mathcal{R}_k$, which we denote by $W$, has the desired properties. By the observation that follows Definition~\ref{piblowup-def}, if $\lambda(\pi^{-1}(i)) > 0$ for all $i\in{\EM{{\mbb{N}}}}$, then the $\pi$-blow up $W$ induces a Borel hypergraph, and so by Lemma~\ref{random-free-has-leading-entropy-zero}, we have $\Entropy{G({\EM{{\mbb{N}}}}, W)} (n) = o(n^k)$. Hence it suffices to construct $\pi$ satisfying $\lambda(\pi^{-1}(i)) > 0$ for all $i\in{\EM{{\mbb{N}}}}$ in such a way that $\Entropy{G({\EM{{\mbb{N}}}}, W)}(n) = $ \linebreak $\Omega(\gamma(n) \cdot n^k)$. For positive $r\in{\EM{{\mbb{N}}}}$, define \[ g_r {\EM{\ :=\ }} \max\bigl\{\{2^{r+3} k \} \cup \{n \in{\EM{{\mbb{N}}}} \,:\, \gamma(n) > 2^{-(r+1)k-3k-1} k^{-k} \}\bigr\}. \] Note that $\lim_{n\to \infty} \gamma(n) = 0$ and so for each $r$ there are only finitely many $n$ such that $\gamma(n) > 2^{-(r+1)k-3k-1} k^{-k}$; hence $g_r$ is well-defined. Observe that for all $n\ge g_1 + 1$, the inequalities \begin{align*} \label{ellstar} \tag{$\star$} n > 2^{r+2} k \qquad \text{and}\qquad \gamma(n) \le 2^{-r k-3k-1} k^{-k} \end{align*} hold when $r = 1$. The remainder of the proof establishes that for all such $n$, we have $\Entropy{G({\EM{{\mbb{N}}}}, W)}(n) \ \geq \ n^k \cdot \gamma(n)$. Fix $n\ge g_1+1$. We have seen that there is at least one $r$ satisfying \eqref{ellstar}; on the other hand, there are only finitely many choices of $r$ for which \eqref{ellstar} holds. Let $q$ be the largest such $r$. Then either $n < 2^{q+3} k$ or \linebreak $\gamma(n) > 2^{-(q+1)k-3k-1} k^{-k}$, and so $n \le g_q$ by the definition of $g_q$. For each $r\ge1$, define \[\textstyle \Gamma_r {\EM{\ :=\ }} \bigl\{\ell\in{\EM{{\mbb{N}}}} \,:\, \sum_{i = 1}^{r-1} g_i \le \ell < \sum_{i=1}^r g_i\bigr\}, \] so that $\{\Gamma_r\}_{r\ge 1}$ is a partition of ${\EM{{\mbb{N}}}}$. For every $\ell\in \Gamma_r$, let $\alpha_\ell {\EM{\ :=\ }} \frac{1}{g_r2^r}$. Observe that for each $r\ge1$, we have $|\Gamma_r| = g_r$, and so $\sum_{\ell\in{\EM{{\mbb{N}}}}}\alpha_{\ell} = 1$. As a consequence, there is a Borel map $\pi\colon [0,1]\to {\EM{{\mbb{N}}}}$ such that for all $a\in\mathcal{R}_k$, we have \[ \lambda(\pi^{-1}(a)) = \frac{\alpha_\ell}{|A_\ell|} \] where $\ell$ is such that $a\in A_\ell$. In other words, vertices in $\mathcal{R}_k$ of the same generation $A_\ell$ are blown up to sets of the same positive measure, and the entire generation $A_\ell$ is blown up to a set of measure $\alpha_\ell$. We may therefore apply Lemma~\ref{Bound on conditional entropy}. Hence for any $\rho\colon[n] \to {\EM{{\mbb{N}}}}$, we have \[ \RandomEntropyFinBig{ G(n, W) \ \Big|\ \bigwedge_{j \in [n]} \pi(\zeta_{\{j\}}) \in A_{\rho(j)}} \geq {{|\rho([n])|}\choose {k}} \] a.s. By Lemma~\ref{cond-reduces-entropy}, we have \begin{align*} \label{Random-free has arbitrary growth rates: Equation 1} \Entropy{G({\EM{{\mbb{N}}}}, W)}(n) \ &\geq \ {\EM{\mathbb{E}}}\Bigl(\RandomEntropyFinBig{G(n, W) \ \Big|\ \bigvee_{\rho\colon[n] \to {\EM{{\mbb{N}}}}} \Bigl(\bigwedge_{j \in [n]} \pi(\zeta_{\{j\}}) \in A_{\rho(j)}\Bigr) \Bigr) } \\ &\geq\sum_{\rho\colon[n] \to {\EM{{\mbb{N}}}}} {\EM{\mathbb{P}}} \Bigl(\bigwedge_{j \in [n]} \pi(\zeta_{\{j\}}) \in A_{\rho(j)}\Bigr) \cdot {{|\rho([n])|}\choose {k}}\\ \ &\geq \ {\EM{\mathbb{P}}}\Bigl(\bigl|Z| \geq n\cdot 2^{-q-2}\Bigr) \cdot {{\lceil n\cdot 2^{-q-2} \rceil}\choose {k}}, \tag{$\ddag$} \end{align*} where we define the random set $Z {\EM{\ :=\ }} \bigcup_{j\in [n]}\{\ell \in {\EM{{\mbb{N}}}} \,:\, \pi(\zeta_{\{j\}}) \in A_\ell \}$. Now define the random quantity $X {\EM{\ :=\ }} |\{Z \cap \Gamma_q ]\}|$, and note that we always have $X \leq |Z|$. Because $\<\zeta_{\{j \}}\>_{j \in [n]}$ is an i.i.d.\ uniform sequence, we have \begin{align*} {\EM{\mathbb{E}}}(X) \ &= \sum_{\ell \in \Gamma_q} {\EM{\mathbb{P}}}\Bigl(\bigvee_{j \in [n]} \pi(\zeta_{\{j \}}) \in A_\ell\Bigr) \\ \ &= \sum_{\ell \in \Gamma_q} (1 - (1-\alpha_\ell)^n)\\ \ &= g_q \cdot \left(1 - \left(1-\frac{1}{g_q 2^q}\right)^n\right). \end{align*} Observe that $(1- x)^n \ \leq\ 1 - nx + n^2x^2 \ \leq\ 1-\frac{nx}{2}$ holds for all $x \in [0, \frac{1}{2n}]$. Since $n \le g_q$ and $q\ge 1$, we have $\frac{1}{g_q 2^q} \in [0, \frac{1}{2n}]$, and so \[ 1 - \left(1-\frac{1}{g_q 2^q}\right)^n \ \geq\ \frac{n}{g_q 2^{q+1}}. \] Putting these together, we get \[ {\EM{\mathbb{E}}}(X) \ \geq\ n\cdot 2^{-q-1}. \] By Chebyshev's inequality, we have \[ {\EM{\mathbb{P}}}\left[|X - {\EM{\mathbb{E}}}(X)| \geq \frac{{\EM{\mathbb{E}}}(X)}{2}\right] \le \frac{4 \Var(X)}{({\EM{\mathbb{E}}}(X))^2}. \] Hence, \begin{align*} 1 - \frac{4 \Var(X)}{({\EM{\mathbb{E}}}(X))^2} \ &\le\ {\EM{\mathbb{P}}}\left[|X - {\EM{\mathbb{E}}}(X)| < \frac{{\EM{\mathbb{E}}}(X)}{2}\right] \\ \ &\le\ {\EM{\mathbb{P}}}\left[X > \frac{{\EM{\mathbb{E}}}(X)}{2} \right]\\ \ &\le\ {\EM{\mathbb{P}}}(X > n\cdot 2^{-q-2}) \\ \ &\le\ {\EM{\mathbb{P}}}(|Z| > n\cdot 2^{-q-2}). \end{align*} For distinct $\ell, \ell' \in \Gamma_q$, the events $\ell \in Z$ and $\ell' \in Z$ have negative correlation, which implies that $\Var(X) \leq {\EM{\mathbb{E}}}(X)$. Hence \[ {\EM{\mathbb{P}}}(|Z| \ > \ n\cdot 2^{-q-2}) \ \geq\ 1 - \frac{4}{{\EM{\mathbb{E}}}(X)} \ \geq\ 1 - \frac{4}{n\cdot 2^{-q-1}} \ \geq\ \frac{1}{2}. \] Finally, substituting in \eqref{Random-free has arbitrary growth rates: Equation 1} and recalling that, by \eqref{ellstar} for $r = q$, we have $n \cdot 2^{-q-2} > k$ and $\gamma(n) \le 2^{-qk-3k-1} k^{-k}$, we obtain \begin{align*} \Entropy{G({\EM{{\mbb{N}}}}, W)}(n) \ &\geq\ \frac{1}{2} \cdot {{\lceil n\cdot 2^{-q-2} \rceil}\choose {k}} \\ \ &\geq\ \frac{1}{2} \cdot \frac{(n\cdot 2^{-q-2} - k)^k}{k^k} \\ \ &\geq\ \frac{1}{2} \cdot (n\cdot 2^{-q-3})^k k^{-k} \\ \ &=\ n^k \cdot 2^{-qk-3k-1} k^{-k} \\ \ &\geq\ n^k \cdot \gamma(n), \end{align*} as desired. \end{proof} \section{Non-redundant invariant measures} \label{nonredundant-sec} In this section we consider entropy functions of invariant measures for countable languages that may be of unbounded arity. If an invariant measure fails to be non-redundant, then its entropy function may take the value $\infty$ even when there are only finitely many relation symbols of each arity in the language. Hence we restrict to the case of non-redundant invariant measures for relational languages, and provide an upper bound on the entropy function in terms of the number of relation symbols of each arity. \emph{For the rest of this section, let ${\EM{L}}$ be a countable relational language (possibly infinite).} We show, in Proposition~\ref{nonred-bound}, that the entropy function of a non-redundant invariant measure for ${\EM{L}}$ is dominated by that of a particular random ${\EM{L}}$-structure that generalizes the Erd\H{o}s--R\'enyi\ random graph having edge probability $\frac12$. We also calculate, in Lemma~\ref{nonred-binomial}, the entropy function of such a maximal entropy structure explicitly in terms of the number of relation symbols of each arity in ${\EM{L}}$. In the case where ${\EM{L}}$ has finitely many relation symbols of each arity, this provides a more precise version of Lemma~\ref{non-redundant-makes-it-real-valued}, which states that such an entropy function takes values in ${\EM{{\mbb{R}}}}$. Moreover, this calculation shows that there are ${\EM{{\mbb{R}}}}$-valued entropy functions that grow arbitrarily fast, in contrast to the situation for finite languages, where the growth is at most polynomial. We now define the \emph{uniform non-redundant measure} for ${\EM{L}}$, which is the distribution of a random structure obtained by independently flipping a fair coin to decide every relation on a tuple of distinct elements, and setting all relations on tuples with repeated elements to false. For a relation symbol $R\in{\EM{L}}$, write $\EM{\mathrm{ar}}(R)$ to denote its arity. \begin{definition} Given a set $X$, define \[\mathrm{NR}_{{\EM{L}}, X} {\EM{\ :=\ }} \{ \<R, {\EM{\ol{x}}}\> \,:\, R \in {\EM{L}},\ {\EM{\ol{x}}}\in X\text{ has distinct entries, and}\ |{\EM{\ol{x}}}| = \EM{\mathrm{ar}}(R) \}. \] Let $\{\xi_{R({\EM{\ol{x}}})}\,:\, \<R, {\EM{\ol{x}}}\> \in \mathrm{NR}_{{\EM{L}}, {\EM{{\mbb{N}}}}} \}$ be a collection of i.i.d.\ uniform $\{\top, \bot\}$-valued random variables and let $\Xi$ be the $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$-valued random variable given by \[ \Xi \vDash R({\EM{\ol{x}}})\qquad \text{if and only if} \qquad \<R, {\EM{\ol{x}}}\> \in \mathrm{NR}_{{\EM{L}}, {\EM{{\mbb{N}}}}} \quad \text{and} \quad \xi_{R({\EM{\ol{x}}})} = \top \] for every $R\in{\EM{L}}$ and $\EM{\mathrm{ar}}(R)$-tuple ${\EM{\ol{x}}}$ of elements from ${\EM{{\mbb{N}}}}$. Define $\UniformMeasure{{\EM{L}}}$ to be the distribution of $\Xi$, and call it the \defn{uniform non-redundant measure} for ${\EM{L}}$. \end{definition} It is easy to see that $\UniformMeasure{{\EM{L}}}$ is both invariant and non-redundant. Note that for $n\in{\EM{{\mbb{N}}}}$, the measure $(\UniformMeasure{{\EM{L}}})_n$ is the uniform distribution on non-redundant structures in $\Str_{\EM{L}}(n)$. Let $\mathfrak{a}_{{\EM{L}}}\colon {\EM{{\mbb{N}}}}\to{\EM{{\mbb{N}}}}\cup \{\infty\}$ be the function sending each $n\in{\EM{{\mbb{N}}}}$ to the number of relation symbols in ${\EM{L}}$ having arity $n$. In the following lemma we calculate the entropy function of $\UniformMeasure{{\EM{L}}}$ in terms of the function $\mathfrak{a}_{{\EM{L}}}$. \begin{lemma} \label{nonred-binomial} For any $n \in {\EM{{\mbb{N}}}}$, we have \[ \Entropy{\UniformMeasure{{\EM{L}}}}(n) \ = \ \sum_{r \leq n} {n \choose r} \cdot r! \cdot \mathfrak{a}_{{\EM{L}}}(r). \] \end{lemma} \begin{proof} The invariant measure $\UniformMeasure{{\EM{L}}}$ is non-redundant, and so for each $n\in{\EM{{\mbb{N}}}}$, the invariant measure $(\UniformMeasure{{\EM{L}}})_n$ is concentrated on elements of $\Str_{\EM{L}}(n)$ in which no relations of arity greater than $n$ hold. Hence $(\UniformMeasure{{\EM{L}}})_n$ is determined by the set $\{\xi_{R({\EM{\ol{x}}})}\,:\, \<R, {\EM{\ol{x}}}\> \in \mathrm{NR}_{{\EM{L}}, [n]}\}$ of random variables. Because these random variables are independent, we have \begin{align*} \Entropy{\UniformMeasure{{\EM{L}}}}(n) \ &= \sum_{(R, {\EM{\ol{x}}}) \in \mathrm{NR}_{{\EM{L}}, [n]}} \EntropyFin{\xi_{R({\EM{\ol{x}}})}} \\ &= \ \bigl|\mathrm{NR}_{{\EM{L}}, [n]}\bigr| \\ &= \ \sum_{r \leq n} {n \choose r} \cdot r! \cdot \mathfrak{a}_{{\EM{L}}}(r), \end{align*} as the number of $r$-tuples from $[n]$ consisting of distinct elements is ${n \choose r} \cdot r!$, and there are $\mathfrak{a}_{{\EM{L}}}(r)$-many relation symbols of arity $r$ in $L$. \end{proof} Observe that when $L$ has only finitely many relation symbols of each arity, Lemma~\ref{nonred-binomial} shows that $\Entropy{\UniformMeasure{{\EM{L}}}}$ is ${\EM{{\mbb{R}}}}$-valued. Hence, by varying the choice of such an $L$ (and hence the function $\mathfrak{a}_{{\EM{L}}}$), we can obtain ${\EM{{\mbb{R}}}}$-valued entropy functions that grow arbitrarily fast. We now show that $\UniformMeasure{{\EM{L}}}$ has the fastest growing entropy function among non-redundant invariant measures on $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$. \begin{proposition} \label{nonred-bound} Let $\nu$ be a non-redundant invariant measure on $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$. Then \[ \Entropy{\nu}(n) \ \leq \ \Entropy{\UniformMeasure{{\EM{L}}}}(n) \] for all $n \in {\EM{{\mbb{N}}}}$. \end{proposition} \begin{proof} Suppose ${\EM{\mathcal{N}}}$ is a random $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$-structure with distribution $\nu$. For $R \in {\EM{L}}$ and ${\EM{\ol{x}}} \in {\EM{{\mbb{N}}}}$ with $|{\EM{\ol{x}}}| = \EM{\mathrm{ar}}(R)$, recall that the random instantiation $R^{\EM{\mathcal{N}}}$ satisfies \[ {R^{\EM{\mathcal{N}}}({\EM{\ol{x}}})} = \top \qquad \text{if and only if} \qquad {\EM{\mathcal{N}}}\vDash R({\EM{\ol{x}}}). \] By Lemma~\ref{joint-entropy-less-than-sum-entropy} we then have, for $n \in {\EM{{\mbb{N}}}}$, \[ \Entropy{\nu}(n) \ = \ \Entropy{{\EM{\mathcal{N}}}}(n) \ \leq \sum_{\substack{R \in {\EM{L}}, \ {\EM{\ol{x}}} \in [n], \text{ and }\\|{\EM{\ol{x}}}| = \EM{\mathrm{ar}}(R)}} \EntropyFinbig{{R^{\EM{\mathcal{N}}}({\EM{\ol{x}}})}}. \] If ${\EM{\ol{x}}}$ has duplicate entries then we know that $\EntropyFinbig{R^{\EM{\mathcal{N}}}({\EM{\ol{x}}})} = 0$ as ${\EM{\mathcal{N}}}$ is non-redundant a.s. Further, if ${\EM{\ol{x}}}$ has no duplicate entries then $\EntropyFinbig{{R^{\EM{\mathcal{N}}}({\EM{\ol{x}}})}} \leq \EntropyFin{\xi_{R({\EM{\ol{x}}})}}$ as the distribution of $\xi_{R({\EM{\ol{x}}})}$ is uniform on $\{\top, \bot\}$, and this is the distribution with maximal entropy on $\{\top, \bot\}$. Therefore, we have \begin{align*} \sum_{\substack{R \in {\EM{L}}, \ {\EM{\ol{x}}} \in [n], \text{ and }\\|{\EM{\ol{x}}}| = \EM{\mathrm{ar}}(R)}} \EntropyFinbig{{R^{\EM{\mathcal{N}}}({\EM{\ol{x}}})}} \ &\leq \sum_{(R, {\EM{\ol{x}}}) \in \mathrm{NR}_{{\EM{L}}, [n]}} \hspace*{-10pt} \EntropyFin{\xi_{R({\EM{\ol{x}}})}}\\ \ &=\ \Entropy{\Xi}(n) \\ \ &=\ \Entropy{\UniformMeasure{{\EM{L}}}}(n), \end{align*} and so $\Entropy{\nu}(n) \ \le \ \Entropy{\UniformMeasure{{\EM{L}}}}(n),$ as desired. \end{proof} Putting together the previous lemma and proposition we immediately obtain the following bound. \begin{corollary} \label{strongLemma} Let $\nu$ be a non-redundant invariant measure on $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$. Then \[ \Entropy{\nu}(n) \leq \sum_{r \leq n} {n \choose r} \cdot r! \cdot \mathfrak{a}_{{\EM{L}}}(r)\] for all $n \in {\EM{{\mbb{N}}}}$. In particular, when $L$ is finite with maximum arity $k$, we have $\Entropy{\nu}(n) = O(n^k)$. \end{corollary} Recall from Lemma~\ref{non-redundant-makes-it-real-valued} that when $L$ has only finitely many relation symbols of each arity, any non-redundant invariant measure on $\Str_{\EM{L}}({\EM{{\mbb{N}}}})$ has an ${\EM{{\mbb{R}}}}$-valued entropy function. Corollary~\ref{strongLemma} improves this by providing an explicit upper bound. In fact, we do not need the invariant measure to be non-redundant, nor the language to be relational, to obtain the final line of Corollary~\ref{strongLemma}; by Proposition~\ref{nonredundant-entropy-function} we see that the polynomial bound $O(n^k)$ holds for the entropy function of an arbitrary invariant measure for a finite language (possibly with constant or function symbols) of maximum arity $k$, thereby strengthening Lemma~\ref{realvalued-lemma}. An interesting question for future work is to characterize precisely those functions from ${\EM{{\mbb{N}}}}$ to ${\EM{{\mbb{R}}}} \cup \{\infty\}$ that can be the entropy function of an invariant measure. By Proposition~\ref{nonredundant-entropy-function}, it suffices to consider entropy functions of non-redundant invariant measures for relational languages. \section*{Acknowledgements} The authors thank Jan Reimann and Daniel M.\ Roy for helpful conversations, and the anonymous referee for their comments. \providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace} \providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR } \providecommand{\MRhref}[2]{% \href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2} } \providecommand{\href}[2]{#2}

1ba2bac0d085b1d23bb085d91f93f0ffc91e731a

\section{Introduction} Massless Dirac fermions are observed in condensed matter physics, in graphene \cite{Novoselov2005,Zhang2005}, organic conductors\cite{Tajima2000,Katayama2006,Kajita2014}, and the surface of the 3D topological insulators\cite{Fu2007,Hsieh2009}. When a two-dimensional system has an inversion symmetry and a time reversal symmetry, massless Dirac points ($\pm \mathbf{k}_{\textrm{D}}$) appear as a pair. The minimal model for the massless Dirac fermions is written as\cite{Kobayashi2007,Goerbig2008,Kobayashi2009} \begin{equation} \mathcal{H}_{\mathrm{D}} = \begin{pmatrix} w_{0x} q_x +w_{0y} q_y & w_x q_x \mp i w_y q_y\\ w_x q_x \pm i w_y q_y & w_{0x} q_x +w_{0y} q_y \end{pmatrix}, \label{eqeq1} \end{equation} where \begin{equation} \mathbf{q} = \mathbf{k} \mp \mathbf{k}_{\textrm{D}}. \end{equation} \begin{figure}[b] \hspace{-0.5cm} \centering \includegraphics[width=0.4\textwidth]{fig1new.eps} \\ \caption{(Color online) Energy as a function of $q_x$ and $q_y$ at $B=0$ at the three-quarter Dirac point. Parameters are $w_x = 0.4$, $w_y=1$, $w_{0x}=-w_x$, $w_{0y}=0$, $\alpha_2'=\alpha_2''=0.01$, } \label{figbandtqD} \end{figure} Two bands touch at the Dirac points. When $w_{0x} = 0$ and $w_{0y} = 0$, the linear energy dispersion near the Dirac point (Dirac cone) is not tilted. By the finite $w_{0x}$ or $w_{0y}$, the Dirac cone is tilted, and if the condition \begin{equation} \left(\frac{w_{0x}} {w_x}\right)^2 + \left(\frac{w_{0y}} {w_y}\right)^2 = 1, \end{equation} is fulfilled, the Dirac cone is critically tilted, i.e., the conical edge of the Dirac cone is horizontal in one direction. In that case we have to take into account the quadratic terms in the tilted direction, except for the special case that the quadratic terms vanish by symmetry or by accident. Generally the quadratic terms exist as we have found previously\cite{Kishigi2017} in the tight-binding model with pressure-dependent hoppings for the organic conductor, $\alpha$-(BEDT-TTF)$_2$I$_3$. The energy near the critically tilted Dirac point is shown in Fig.~\ref{figbandtqD}. Since the energy of the upper band depends linearly in three directions (for example, $-q_x$ and $\pm q_y$) and quadratically in one direction (for example, $+ q_x$) in that case, we call the critically tilted Dirac point as the ``three-quarter Dirac point''.\cite{Kishigi2017} It has been known that when two-Dirac points merge at the time-reversal-invariant momentum, the energy depends linearly in two directions and quadratically in two directions, and it is called the semi-Dirac point\cite{Hasegawa2006,Dietl2008,Montambaux2009,Banerjee2009,Delplace2010}. Previously we have shown that the energy in a magnetic field (the Landau level) at the three-quarter Dirac point depends on the quantum number $n$ and the magnetic field $B$ as \begin{equation} \epsilon_n \propto (n B)^{\frac{4}{5}} , \end{equation} by calculating the energy of the tight-binding model for $\alpha$-(BEDT-TTF)$_2$I$_3$ in a magnetic field numerically\cite{Kishigi2017}. In that paper we have explained these $n$ and $B$ dependences of Landau levels by using the semi-classical quantization rule. In this paper we study the Landau quantization at the three-quarter Dirac point in a numerical study and an analytical treatment with a crude approximation. The Dirac cone is taken to be critically tilted in the $k_x$ direction, i.e., $w_{0x} = -w_x$, and $w_{0y}=0$ in Eq.~(\ref{eqeq1}). For simplicity we take $w_{x}>0$, $w_{y}>0$, and we introduce the quadratic terms in the $q_x$ direction ($\alpha_2' q_x^2$ in diagonal elements and $\alpha_2'' q_x^2$ in off-diagonal elements). Then the three-quarter Dirac Hamiltonian we study in this paper is \begin{equation} \mathcal{H}_{\mathrm{tqD}} = \begin{pmatrix} - w_x q_x + \alpha_2' q_x^2 & w_x q_x +\alpha_2'' q_x^2 - i w_y q_y\\ w_x q_x +\alpha_2'' q_x^2 + i w_y q_y & - w_x q_x + \alpha_2' q_x^2 \end{pmatrix}. \label{eqtqd0} \end{equation} \section{three-quarter Dirac point} \subsection{energy at $B=0$} In the absence of the magnetic field the energy is obtained by \begin{equation} \mathcal{H}_{\mathrm{tqD}} \Psi = E (\mathbf{q}) \Psi , \label{eqeq0} \end{equation} where $\Psi$ is a wave function which has two components, $\psi_1$ and $\Psi_2$. The eigenvalues of $\mathcal{H}_{\mathrm{tqD}}$ is obtained as $E(\mathbf{q})=\varepsilon^0_{\mathrm{tqD}_{\pm}} (\mathbf{q})$; \begin{align} \varepsilon^0_{\textrm{tqD}_{\pm}} (\mathbf{q}) =& -w_x q_x + \alpha_{2}' q_x^2 \nonumber \\ & \pm \sqrt{(w_x q_x+\alpha_{2}'' q_x^2)^2 + (w_y q_y)^2}, \end{align} which are plotted in Fig.~\ref{figbandtqD}. There exist the upper band ($\varepsilon^0_{\textrm{tqD}_+}(\mathbf{q})$) and the lower band ($\varepsilon^0_{\textrm{tqD}_-}(\mathbf{q})$). These two bands touch at $\mathbf{q}=(0,0)$. Along the $q_x$ axis, the linear term disappears in $\varepsilon^0_{\textrm{tqD}_{+}} (\mathbf{q})$ and $\varepsilon^0_{\textrm{tqD}_{-}} (\mathbf{q})$ for $q_x >0$ and $q_x <0$, respectively, whereas in other three directions the linear term exists; \begin{align} \varepsilon_{\textrm{tqD}_+} (q_x, q_y=0) &= \left\{ \begin{array}{ll} \alpha_2 q_x^2 & \mbox{ if $q_x>0$} \\ 2 w_x |q_x| + \tilde{\alpha}_2 q_x^2 & \mbox{ if $q_x < 0$} \end{array} \right. \label{eq07} \\ \varepsilon_{\textrm{tqD}_+} (q_x=0, q_y) &= w_y |q_y| \\ \varepsilon_{\textrm{tqD}_-} (q_x, q_y=0) &= \left\{ \begin{array}{ll} -2w_x q_x + \tilde{\alpha}_2 q_x^2 & \mbox{ if $q_x > 0$} \\ \alpha_2 q_x^2 & \mbox{ if $q_x < 0$} \end{array} \right. \label{eq09} \\ \varepsilon_{\textrm{tqD}_-} (q_x=0, q_y) &= - w_y |q_y| , \end{align} where \begin{equation} \alpha_2 = \alpha_2' + |\alpha_2''| , \label{eqalpha2} \end{equation} and \begin{equation} \tilde{\alpha}_2 = \alpha_2' - |\alpha_2''| . \label{eqtildealpha2} \end{equation} If $\alpha_2>0$, $\mathbf{q}=0$ is a local minimum of $\varepsilon_{\textrm{tqD}_+}$ with the linear dispersion in three directions ($q_x<0$, $q_y>0$ and $q_y <0$) and quadratic dispersion in one direction ($q_x>0$). Note that the three-quarter Dirac point is neither the local maximum nor the local minimum of $\varepsilon_{\textrm{tqD}_-}$ if $\alpha_2>0$. If $\alpha_2<0$, the three-quarter Dirac point is the local maximum of $\varepsilon_{\textrm{tqD}_-}$, but it is neither the local maximum nor the local minimum of $\varepsilon_{\textrm{tqD}_+}$. \subsection{numerical results of the energy at $B > 0$, using boundary condition at $y>0$} \label{sectionbc1} Hereafter we study the case $\alpha_2 >0$, i.e., the three-quarter Dirac point is the minimum of $\varepsilon^0_{\mathrm{tqD}_+}$, as shown in Fig.~\ref{figbandtqD}. In this case it is expected that when the magnetic field is applied, there are the almost-localized bound states (the Landau levels) at $E>0$, since there exists a closed Fermi surface at $E>0$ in the $\varepsilon^0_{\mathrm{tqD}_+}$ band, and the semiclassical Landau quantization is expected for the closed orbit. On the other hand, the Fermi surface in the $\varepsilon^0_{\mathrm{tqD}_-}$ band is open and a continuous energy is expected in the $\varepsilon^0_{\mathrm{tqD}_-}$ band even in the presence of the magnetic field. Quantum mechanically the Landau levels in the $\varepsilon^0_{\mathrm{tqD}_+}$ band couple to the continuous energy in the $\varepsilon^0_{\mathrm{tqD}_-}$ band by quantum tunneling. In this subsection we show that the coupling between the almost-localized Landau levels and the continuous energy cannot be neglected for the quantized energy with the small quantum number, $n$, but it becomes small for the larger values of $n$. In the presence of the magnetic field $\mathbf{B}$ ($\mathbf{B} = \nabla \times \mathbf{A}$, where $\mathbf{A}$ is the vector potential), we replace $q_x$ and $q_y$ as \begin{align} q_x &\rightarrow -i \hbar \frac{\partial}{\partial x} + e A_x, \\ q_y &\rightarrow -i \hbar \frac{\partial}{\partial y} + e A_y . \end{align} We study the case that the uniform magnetic field $B>0$ is applied along the $z$ direction. We take the vector potential as \begin{equation} \mathbf{A} = (-B y,0,0). \end{equation} Since there is no explicit $x$ in Eq.~(\ref{eqeq0}), we can write \begin{equation} \Psi(x,y) = e^{i k_x x} \begin{pmatrix} \Psi_1(y) \\ \Psi_2(y) \end{pmatrix} . \end{equation} In this case we take \begin{equation} q_x \rightarrow \hbar k_x - eB y \equiv - eB \ell \bar{y}, \label{eqqxBy} \end{equation} where the magnetic length $\ell$ is defined as usual, \begin{equation} \ell=\sqrt{\frac{\hbar}{eB}}, \end{equation} and $\bar{y}$ is the dimensionless length. Hereafter we write $\bar{y}$ as $y$ for simplicity. Then the equation we study is \begin{equation} \begin{pmatrix} (\tilde{\cal{H}}_{\textrm{tqD}} )_{11} & (\tilde{\cal{H}}_{\textrm{tqD}} )_{12} \\ (\tilde{\cal{H}}_{\textrm{tqD}} )_{21} & (\tilde{\cal{H}}_{\textrm{tqD}} )_{22} \end{pmatrix} \begin{pmatrix} \Psi_1({y}) \\ \Psi_2({y}) \end{pmatrix} = E \begin{pmatrix} \Psi_1({y}) \\ \Psi_2({y}) \end{pmatrix} , \label{eqeq2} \end{equation} where \begin{align} (\tilde{\cal{H}}_{\textrm{tqD}} )_{11} &= w_y\sqrt{\hbar eB} \left( \frac{w_x}{w_y} {y} + \frac{\alpha_2' \sqrt{\hbar eB}}{w_y} {y}^2 \right), \label{eqh11}\\ (\tilde{\cal{H}}_{\textrm{tqD}} )_{12} &= w_y\sqrt{\hbar eB} \left( - \frac{d}{d {y}} - \frac{w_x}{w_y} {y} + \frac{\alpha_2'' \sqrt{\hbar eB}}{w_y} {y}^2 \right) , \label{eqh12} \\ (\tilde{\cal{H}}_{\textrm{tqD}} )_{21} &= w_y\sqrt{\hbar eB} \left( \frac{d}{d {y}} - \frac{w_x}{w_y} {y} + \frac{\alpha_2'' \sqrt{\hbar eB}}{w_y} {y}^2 \right) , \label{eqh21}\\ (\tilde{\cal{H}}_{\textrm{tqD}} )_{22} &= (\tilde{\cal{H}}_{\textrm{tqD}} )_{11} . \label{eqh22} \end{align where $w_y\sqrt{\hbar eB} = \hbar w_y/\ell$ is the energy scale for the massless Dirac fermions. There are other dimensionless parameters, $w_x/w_y$, $\alpha_2'\sqrt{\hbar e B}/w_y$, and $\alpha_2''\sqrt{\hbar e B}/w_y$. We assume that $w_x/w_y$ is order of $1$ and we mainly study the case $w_x=w_y$ in this paper. Other two dimensionless parameters are taken to be small, i.e., \begin{align} \alpha_2' \sqrt{\hbar e B}/w_y &\ll 1 ,\\ \alpha_2'' \sqrt{\hbar e B}/w_y &\ll 1 . \end{align} We will show that the sum of these small dimensionless parameters ($\alpha_2 \sqrt{\hbar eB}/w_y$) plays an important role in the quantization of energies for almost localized states in the magnetic field, but the difference ($\tilde{\alpha}_2 \sqrt{\hbar eB}/w_y$) is irrelevant when these parameters are small. In other words there is another length scale $\ell \alpha_2\sqrt{\hbar eB}/w_y = \hbar \alpha_2/w_y$. \begin{figure}[tb] \centering {\raggedright \ (a) \hspace{4.0cm} (b) \\ } \vspace{0.5cm} \includegraphics[width=0.2\textwidth]{fig2anew.eps} \hfil \includegraphics[width=0.2\textwidth]{fig2bnew.eps}\\ \vspace{0.5cm} {\raggedright \ (c) \hspace{4.0cm} (d) \\ } \vspace{0.5cm} \includegraphics[width=0.2\textwidth]{fig2cnew.eps} \hfil \includegraphics[width=0.2\textwidth]{fig2dnew.eps} \caption{(Color online) Wave functions in the three-quarter Dirac point obtained numerically with $E/(w_y \sqrt{\hbar eB})=0.3$ (a), $0.31$ (b), $0.32$ (c), and $0.33$ (d). Parameters are $w_x = w_y=1$, $w_{0x}=-1$, $w_{0y}=0$, $\alpha_2' \sqrt{\hbar eB}/w_y =\alpha_2' \sqrt{\hbar eB}/w_y=0.01$, and $B=1$. The boundary condition at $y=20$ is taken to be Eq.~(\ref{eqbcy20}). } \label{fige0bx} \end{figure} \begin{figure}[tb] \centering {\raggedright \ (a) \hspace{4.0cm} (b) \\ } \vspace{0.5cm} \includegraphics[width=0.2\textwidth]{fig3anew.eps}\hfil \includegraphics[width=0.2\textwidth]{fig3bnew.eps \caption{(Color online) Wave functions in the three-quarter Dirac point obtained numerically with $E/(w_y \sqrt{\hbar eB})=0.3105$ (a), and $0.3108$ (b) Parameters are $w_x = w_y=1$, $w_{0x}=-1$, $w_{0y}=0$, $\alpha_2'\sqrt{\hbar eB}/w_y=\alpha_2''\sqrt{\hbar eB}/w_y=0.01$, and $B=1$. The boundary condition at $y=20$ is taken to be Eq.~(\ref{eqbcy20}). } \label{fige0bx2} \end{figure} \begin{figure}[bt] \centering {\raggedright \ (a) \hspace{3.5cm} (b) \\ } \vspace{0.4cm} \includegraphics[width=0.2\textwidth]{fig4anew.eps} \hfil \includegraphics[width=0.2\textwidth]{fig4bnew.eps}\\ \vspace{0.5cm} {\raggedright \ (c) \\ } \vspace{0.4cm} \hspace{0.4cm} \includegraphics[width=0.2\textwidth]{fig4cnew.eps}\hfill \mbox{ } \caption{(Color online) Wave functions in the three-quarter Dirac point obtained numerically with $E/(w_y \sqrt{\hbar eB})=0.52906$ (a), $0.529065$ (b), and $0.52907$ (c). Parameters are $w_x = w_y=1$, $w_{0x}=-1$, $w_{0y}=0$, $\alpha_2'\sqrt{\hbar eB}/w_y=\alpha_2''\sqrt{\hbar eB}/w_y=0.01$, and $B=1$. The boundary condition at $y=20$ is taken to be Eq.~(\ref{eqbcy20}). } \label{fige1bxx} \end{figure} We seek the solution of Eq.(\ref{eqeq2}) with $E \geq 0$ which satisfies the conditions that at $y \to -\infty$ \begin{align} \Psi_1(y) \to 0, \label{eqymininf1}\\ \Psi_2(y) \to 0. \label{eqymininf2} \end{align} Note that $y \to -\infty$ corresponds to $q_x \to + \infty$, as seen in Eq.~(\ref{eqqxBy}). When $y \to +\infty$, $\Psi_1(y)$ and $\Psi_2(y)$ do not have to vanish because the lower band becomes positive when $q_x \to -\infty$ at $B=0$ as seen in Fig.~\ref{figbandtqD}. Therefore, the conditions, Eqs.~(\ref{eqymininf1}) and (\ref{eqymininf2}) at $y \to - \infty$, do not make the energy quantized. There is the solution for any value of $E$, but the conditions, Eqs.~(\ref{eqymininf1}) and (\ref{eqymininf2}) at $y \to - \infty$, make the restriction for the solutions. We solve the differential equations, Eq.~(\ref{eqeq2}), numerically by the fourth-order Runge-Kutta method in this and the next subsections. We take the step size in the Runge-Kutta method to be $0.01$. Since Eq.~(\ref{eqeq2}) is the real linear differential equations, the solutions can be taken as real functions, and the solutions multiplied by any constant values give the same solutions. Therefore, for each value of $E$ the only adjustable parameter to obtain the solution numerically by the Runge-Kutta method starting from a fixed $y=y_{+}>0$ and decreasing $y$ is the ratio $\Psi_2(y_{+})/\Psi_1(y_{+})$. In this subsection we take $y_{+}=20$. It is convenient to parametrize the ratio in terms of the angle $\theta$ defined by \begin{equation} \frac{\Psi_2(y=20)}{\Psi_1(y=20)} = \tan \theta . \label{eqbcy20} \end{equation} The numerically obtained solution diverges as $y$ becomes a negative large value, if the chosen $\theta$ is not a suitable value for the given $E$. Only when $\theta$ is the correct value for $E$, the numerically obtained solution becomes zero as $y \to -\infty$. In this way we determine $\theta$ for any given $E>0$. The boundary condition $\theta$ depends on the choice of $y_{+}$ and it does not have an important meaning. The $E$-dependence of $\theta$, however, is important to obtain the almost-localized state. When $E$ is changed continuously, $\theta$ changes continuously. Note that the energy is semi-classically quantized by the magnetic field, since the closed Fermi surface $\varepsilon_{\textrm{tqD}_{+}}^0(\mathbf{q})$ exists at $B=0$. Quantum mechanically, these quantized states in $y \lesssim 0$ couple to the continuous-energy states, which exist mainly in $y\gtrsim 0$, by tunneling. With this mixing of the states $\theta$ changes by $\pi$ in the small region of energy variation. Note that $\theta$ and $\theta + n \pi$ with integer $n$ give the same condition. We show some examples of the solutions for $0.3 \leq E /(w_y \sqrt{\hbar eB}) \leq 0.33$ in Fig.~\ref{fige0bx} and Fig.~\ref{fige0bx2} and for $0.52906 \leq E /(w_y \sqrt{\hbar eB}) \leq 0.52907$ in Fig.~\ref{fige1bxx}, where we have normalized the wave functions numerically as \begin{equation} \int_{-10}^{20} \left( |\Psi_1(y)|^2 + |\Psi_2(y)|^2 \right) dy = 1. \end{equation} \begin{figure}[bt] \centering {\raggedright \ (a) \\ } \includegraphics[width=0.33\textwidth]{fig5anew.eps}\\ \vspace{0.4cm} {\raggedright \ (b) \hspace{3.5cm} (c) \\ } \vspace{0.4cm} \includegraphics[width=0.2\textwidth]{fig5bnew.eps} \hfil \includegraphics[width=0.2\textwidth]{fig5cnew.eps} \caption{(Color online) Boundary condition $\theta$ at $y=20$ (Eq.~(\ref{eqbcy20})) as a function of energy for $w_x = w_y=1$, $w_{0x}=-1$, $w_{0y}=0$, $\alpha_2'\sqrt{\hbar eB}/w_y=\alpha_2''\sqrt{\hbar eB}/w_y=0.01$, and $B=1$. (b) and (c) are the close-up of (a) near the energy of the almost-localized states at $y \lesssim 0$. } \label{figevsthetax} \end{figure} Nearly-localized states in $y \lesssim 0$ exist at $E/(w_y \sqrt{\hbar eB}) \approx 0.3108$ and $0.529065$. The wave functions $(\Psi_1(y), \Psi_2(y))$ at $ E= 0.3105$ and $0.3108$ with the suitable boundary conditions have one node of $\Psi_1(y)$ and $\Psi_2(y)$ in $y \lesssim 0$, as seen in Fig.~\ref{fige0bx2} (a) and (b), and the wave functions at $E/(w_y \sqrt{\hbar eB}) = 0.52906$, $0.529065$, and $0.52807$ have two nodes in $y \lesssim 0$, as seen in Fig.~\ref{fige1bxx} (a) - (c). Therefore, $E/(w_y \sqrt{\hbar eB}) \approx 0.3108$ and $E/(w_y \sqrt{\hbar eB}) \approx 0.529065$ are the nearly-localized state energies with $n=1$ and $n=2$, respectively. Due to the tunneling these nearly-localized states are not completely localized in the region $y \lesssim 0$, which corresponds to the region $q_x \gtrsim 0$ in the case of $B=0$ (see Eq.~(\ref{eqqxBy}) and Fig.~\ref{figbandtqD}). This interpretation of the nearly-localized states in three-quarter Dirac point is justified by plotting $\theta$ as a function of energy (Fig.~\ref{figevsthetax}). As seen in Fig.~\ref{figevsthetax}, $\theta$ changes continuously as $E$ increases. When the energy is close to one of the energies of the nearly-localized states, $\theta$ changes by $\pi$ in a narrow range of $E$. At $n=2$ ($E/(w_y \sqrt{\hbar eB}) \approx 0.529065$) $\theta$ changes in a narrower range of the energy $E$ than at $n=1$ ($E/(w_y \sqrt{\hbar eB}) \approx 0.3108$). The narrowing of the range in $\theta$ is reasonable because the tunneling of the almost-localized state at $y \lesssim 0$ into the region of $y \gtrless 0$ is weaker at $n=2$ than at $n=1$. \begin{figure}[tb] \centering {\raggedright \hspace{0.8cm} (a) \\ } \vspace{1.0cm} \includegraphics[width=0.45\textwidth]{fig6anew.eps} {\raggedright \hspace{0.8cm} (b) \\ } \vspace{1.0cm} \includegraphics[width=0.45\textwidth]{fig6bnew.eps} \caption{(Color online) Green lines are $\varepsilon_{\mathrm{tqD}_\pm}(q_x,q_y=0)/(w_y \sqrt{\hbar eB})$ as a function of $y = -q_x/(\ell eB)=-q_x/\sqrt{\hbar eB}$, i.e. $E/(w_y \sqrt{\hbar eB}) = (\alpha_2 \sqrt{\hbar eB}/w_y) y^2$ and $E/(w_y \sqrt{\hbar eB}) = (2 w_x/w_y) y + (\tilde{\alpha}_2 \sqrt{\hbar eB}/w_y) y^2$. Classical particles can exist in the cyan-shaded regions. Wave functions of almost-localized state with the quantum number $n=1$, $2$, $3$, and $4$ near the three-quarter Dirac point are plotted as functions of $y$. Zero of the wave functions are shifted to their energies. Wave functions are calculated with the boundary conditions at $y=-15$ as in section II C. Parameters are $B=1$, $w_x = w_y=1$, $w_{0x}=-1$, $w_{0y}=0$ in (a) and (b), and $\alpha_2'\sqrt{\hbar eB}/w_y=\alpha_2''\sqrt{\hbar eB}/w_y=0.005$ ($\alpha_2\sqrt{\hbar eB}/w_y=0.01$) in (a) and $\alpha_2'\sqrt{\hbar eB}/w_y=\alpha_2''\sqrt{\hbar eB}/w_y=0.0025$ ($\alpha_2\sqrt{\hbar eB}/w_y=0.005$) in (b). } \label{figphiByE} \end{figure} In Fig.~\ref{figphiByE}, we plot $\varepsilon_{\textrm{tqD}_\pm}(q_x,q_y=0)$ at $B=0$ (Eqs.~(\ref{eq07}) and (\ref{eq09})) with replacing $q_x \to -eB \ell y$ (Eq.~(\ref{eqqxBy})) divided by the energy scale of the massless Dirac fermions ($w_y\sqrt{\hbar eB}$) as a function of dimensionless length $y$ for the dimensionless parameter $\alpha_2\sqrt{\hbar eB} /wy=0.01$ (a) and $0.005$ (b). In these figures we also plot the wave functions of the almost-localized states at $y \lesssim 0$ with the quantum number $n=1 - 4$, which are calculated using the boundary condition at $y=y_{-}<0$ discussed in the next subsection. Classically, electrons can exist in the cyan-shaded regions in Fig.~\ref{figphiByE}, and they can exist only by the quantum tunneling effect in the white regions. For the larger energy (larger quantum number $n$) the width and the hight of the classically-forbidden region (white region in Fig.~\ref{figphiByE}) is larger. As a result the tunneling of the almost localized state with the larger quantum number at $y \lesssim 0$ into the $y \gtrsim 0$ region becomes smaller. Therefore, the numerical solutions of the bound states $n \geq 3$ are difficult to obtain by using the boundary condition at $y=y_{+} >0$, Eq.~(\ref{eqbcy20}), since $\theta$ changes by $\pi$ in a very narrow region in energy. On the other hand, the almost-localized state with the quantum number $n=1$ couples strongly to the continuous states at $y \gtrsim 0$ as seen in Fig.~\ref{figevsthetax}(b), and the energy of the almost-localized state is ``broadened''. In the next subsection we use the boundary condition at $y<0$ to obtain the energy of the bound states. \subsection{numerical results of energy at $B > 0$, using boundary condition at $y<0$} \label{sectionbc2} \begin{figure}[tb] \begin{center} \includegraphics[width=0.45\textwidth]{fig7new.eps} \end{center} \caption{(Color online) Wave function $\Psi_1(y)$ in the three-quarter Dirac point obtained numerically by using the boundary condition at $y=-10$. Parameters are $w_x = w_y=1$, $w_{0x}=-1$, $w_{0y}=0$, $\alpha_2'\sqrt{\hbar eB}/w_y=\alpha_2''\sqrt{\hbar eB}/w_y=0.01$, and $B=1$. We take several values of $E$, and we find that an eigenvalue for the nearly localized state at $y<0$ exists in the region $0.310 < E/(w_y \sqrt{\hbar eB}) < 0.311$. } \label{figen1x} \end{figure} As shown in the previous subsection, it is difficult to obtain the energy of the almost-localized states at $y \lesssim 0$ with a large quantum number $n$ in Eq.(\ref{eqeq2}) by using the boundary condition at $y = y_{+} >0$, since the boundary condition changes in a very narrow region and the energy of the almost-localized states at $y \lesssim 0$ may be overlooked. Therefore, we try to obtain the energy by using the boundary conditions at $y<0$. We study the solutions of Eq.(\ref{eqeq2}) at $y \to -\infty$, assuming \begin{equation} \Psi_j(y) = c_j(y) e^{-g(y)}, \end{equation} ($j=1,2$) and \begin{align} \frac{d \Psi_j}{dy} &= \left(- \frac{dg(y)}{dy} c_j(y) +\frac{d c_j(y)}{d y} \right) e^{-g(y)} \nonumber \\ &\sim - \frac{dg(y)}{dy} c_j e^{-g(y)} , \end{align} as $y \to -\infty$. Then we obtain the equation \begin{equation} \begin{pmatrix} (\tilde{\cal{H}}_{\textrm{tqD}} )_{11} & (\tilde{\cal{F}}_{\textrm{tqD}} )_{12} \\ (\tilde{\cal{F}}_{\textrm{tqD}} )_{21} & (\tilde{\cal{H}}_{\textrm{tqD}} )_{22} \end{pmatrix} \begin{pmatrix} \Psi_1(y) \\ \Psi_2(y) \end{pmatrix} \approx E \begin{pmatrix} \Psi_1(y) \\ \Psi_2(y) \end{pmatrix} , \label{eqapprox} \end{equation} where $(\tilde{\cal{H}}_{\textrm{tqD}} )_{11}$ and $(\tilde{\cal{H}}_{\textrm{tqD}} )_{22}$ are given in Eqs.~(\ref{eqh11}) and (\ref{eqh22}) and \begin{align} (\tilde{\cal{F}}_{\textrm{tqD}} )_{12} &= w_y \sqrt{\hbar eB} \left( \frac{d g}{dy} -\frac{w_x}{w_y} y + \frac{\alpha'' \sqrt{\hbar eB}}{w_y} y^2 \right) , \label{eqh12b} \\ (\tilde{\cal{F}}_{\textrm{tqD}} )_{21} &= w_y \sqrt{\hbar eB} \left( - \frac{d g}{dy} -\frac{w_x}{w_y} y + \frac{\alpha'' \sqrt{\hbar eB}}{w_y} y^2 \right) . \label{eqh21b} \end{align} The nontrivial solution exists when the condition \begin{equation} \det \begin{pmatrix} (\tilde{\cal{H}}_{\textrm{tqD}} )_{11} -E & (\tilde{\cal{F}}_{\textrm{tqD}} )_{12} \\ (\tilde{\cal{F}}_{\textrm{tqD}} )_{21} & (\tilde{\cal{H}}_{\textrm{tqD}} )_{22} -E \end{pmatrix} =0, \end{equation} is fulfilled, i.e., \begin{align} \left( \frac{d g}{dy} \right)^2 &= \left( \frac{\alpha_2 \sqrt{\hbar eB}}{w_y} y^2 - \frac{E}{w_y \sqrt{\hbar eB}} \right) \nonumber \\ & \times \left(- \frac{2 w_x}{w_y} y - \frac{\tilde{\alpha}_2 \sqrt{\hbar eB}}{w_y} y^2 + \frac{E}{w_y\sqrt{\hbar eB}} \right), \end{align} In the simple case that $\alpha_2 >0$, $\tilde{\alpha}_2=0$ (i.e., $\alpha'=\alpha''=\alpha_2/2$), and large $|y|$, we can neglect terms proportional to $E$. Then the approximate solution is \begin{equation} g(y) \sim \pm \frac{2 \sqrt{2 \alpha_2 w_x \sqrt{\hbar eB}}}{5 w_y} (-y)^{\frac{5}{2}} + \textrm{const.} \end{equation} The solution which does not diverge at $y \to -\infty$ is obtained as \begin{equation} \begin{pmatrix} \Psi_1 \\ \Psi_2 \end{pmatrix} \sim \exp \left( - \frac{2 \sqrt{2 \alpha_2 w_x \sqrt{\hbar eB}}}{5 w_y} (-y)^{\frac{5}{2}} \right) \begin{pmatrix} c_1 \\ c_2 \end{pmatrix} . \label{eqf12} \end{equation} Inserting Eq.~(\ref{eqf12}) into Eq.~(\ref{eqapprox}) we obtain the approximate boundary condition at $y \to -\infty$ as \begin{align} \frac{\Psi_1(y)}{\Psi_2(y)} &\sim \frac{c_1(y)}{c_2(y)} \nonumber \\ &\sim \frac{w_x+{\alpha_2'} B\sqrt{\hbar eB} y}{w_x-{\alpha_2''} \sqrt{\hbar eB} y -\sqrt{2\alpha_2 w_x \sqrt{\hbar eB (-y)}}} . \end{align} With this boundary conditions at $y=y_{-}=-10.0$ we solve the differential equation Eq.~(\ref{eqeq2}) numerically in the Runge-Kutta method with increasing $y$. When we take $E$ to be one of the correct values of the Landau levels, the wave function is nearly localized at $y \lesssim 0$ and tunnels to $y>0$ very little. On the other hand, if we take the different value of $E$, the wave function becomes large as $y$ is increased at $y>0$, although it does not diverge. As shown in Fig.~\ref{figen1x}, the wave function in the region $y>0$ calculated numerically with the boundary condition at $y_{-}=-10$ becomes small only when we take the correct eigenvalue $0.310 < E < 0.311$. This value is consistent with the $n=1$ eigenvalue obtained numerically with the boundary condition at $y_{+}=20$ (Fig.~\ref{fige0bx2}). We also check numerically that the solution is not sensitive to the boundary condition; numerically the same result is obtained even when we take $\Psi_1=0$ and $\Psi_2 \neq 0$ at $y=y_{-}=-10$. The independence on the boundary condition can be understood as follows. As seen in section II B, the coupling between the nearly-localized state at $y \lesssim 0$ and the continuous state as $y \gtrsim 0$ is small for $n \geq 2$. In section II~B we first fixed the energy and obtain the wave functions not divergent at $y \to -\infty$ by changing the boundary condition at $y_{+}>0$ ($\theta$ at $y_{+}=20$). In this section we first take the approximate boundary condition at $y_{-}=-10$, and obtain the energy which gives the smallest amplitude of oscillations of the wave function at $y>0$. Even though the boundary condition is not exact, suitable linear combination of the nearly-localized state at $y \lesssim 0$ and continuous state as $y \gtrsim 0$ may give the non-divergent solution with the given boundary condition at $y=y_{-}$, if the energy is the correct energy of the nearly-localized state at $y \lesssim 0$. In Fig.~\ref{fige0} we show the wave functions for nearly-localized states with quantum numbers $n=0$ -- $6$. \begin{figure}[tb] \centering {\raggedright \ (a) \hspace{3.5cm} (b) \\ } \vspace{0.4cm} \includegraphics[width=0.2\textwidth]{fig8anew.eps} \hfil \includegraphics[width=0.2\textwidth]{fig8bnew.eps}\\ \vspace{0.5cm} {\raggedright \ (c) \hspace{3.5cm} (d) \\ } \vspace{0.4cm} \includegraphics[width=0.2\textwidth]{fig8cnew.eps} \hfil \includegraphics[width=0.2\textwidth]{fig8dnew.eps}\\ \vspace{0.5cm} {\raggedright \ (e) \hspace{3.5cm} (f) \\ } \vspace{0.4cm} \includegraphics[width=0.2\textwidth]{fig8enew.eps} \hfil \includegraphics[width=0.2\textwidth]{fig8fnew.eps}\\ \vspace{0.5cm} {\raggedright \ (g) \\ } \vspace{0.4cm} \hspace{0.4cm} \includegraphics[width=0.2\textwidth]{fig8gnew.eps} \hfill \mbox{ } \caption{(Color online) Wave functions of nearly localized eigenstates at $y<0$ with quantum number (a) $n=0$, (b) $n=1$, (c) $n=2$, (d) $n=3$, (e) $n=4$, (f) $n=5$, and (g) $n=6$ in the three-quarter Dirac point obtained numerically with the boundary condition at $y=-10$. Parameters are $w_x = w_y=1$, $w_{0x}=-1$, $w_{0y}=0$, $\alpha_2'\sqrt{\hbar eB}/w_y=\alpha_2''\sqrt{\hbar eB}/w_y=0.01$, and $B=1$. } \label{fige0} \end{figure} For $n=0$, i.e. $E=0$, both components of the wave function have a broad peak at $y=0$, although each component of the wave functions is not small at $y>0$, as shown in Fig.~\ref{fige0}(a). The oscillation of the wave function at $y>0$ can be understood as the continuous energy states at $y>0$. Since the upper band touches the lower band at the three-quarter Dirac point without the boundary barrier, the nearly-localized state at $y<0$ goes through to the region $y>0$. We will discuss the $n=0$ state in the next section. The eigenstate for $n \geq 1$ is obtained by taking the suitable value of $E$, which minimize the amplitude of oscillation of the wave function in the region $y>0$. We find the tunneling through the barrier is smaller as $n$ becomes larger, as we have discussed in the previous subsection. We also calculate the energy as a function of quantum number $n$ with different choice of parameters $\alpha_2'\sqrt{\hbar eB}/w_y=0.02$ and $\alpha_2''=0$ from these used in Fig.~\ref{fige0} ($\alpha_2'\sqrt{\hbar eB}/w_y=\alpha_2''\sqrt{\hbar eB}/w_y=0.01$). We plot the energy as a function of $n$ in Fig.~\ref{figEvsn}. We obtain \begin{equation} E_n \propto n^{\frac{4}{5}}. \label{eqpropto} \end{equation} In Figs.~\ref{figEvsalpha} and \ref{figEvsB} we plot the energy as a function of $\alpha_2$ and $B$, respectively. We obtain \begin{equation} E_n \propto \alpha_2^{\frac{3}{5}} (n B)^{\frac{4}{5}}. \label{eqquantizedE} \end{equation} We have previously obtained $n$ and $B$ dependence at the three-quarter Dirac point (Eq.~(\ref{eqquantizedE})) in the tight-binding model of $\alpha$-(BEDT-TTF)$_2$I$_3$ at the critical pressure\cite{Kishigi2017}. \begin{figure}[tb] \centering \includegraphics[width=0.4\textwidth]{fig9new.eps} \caption{(Color online) Energy as a function of the quantum number $n$ for the three-quarter Dirac point. Parameters are $w_x = w_y=1$, $w_{0x}=-1$, $w_{0y}=0$, and $B=1$. We take two choices of parameters giving the same $\alpha_2$, $\alpha_2'\sqrt{\hbar eB}/w_y=\alpha_2''\sqrt{\hbar eB}/w_y=0.01$ and $\alpha_2'\sqrt{\hbar eB}/w_y=0.02$, $\alpha_2''=0$. The obtained values of the energy is well fitted by the red broken line ($E \propto 0.3 n^{\frac{4}{5}}$). } \label{figEvsn} \end{figure} \begin{figure}[tb] \hspace{1cm} \begin{center} \includegraphics[width=0.4\textwidth]{fig10new.eps} \end{center} \caption{(Color online) Energy with the quantum number $n=3$ as a function of $\alpha_2$ for the three-quarter Dirac point. Parameters are $w_x = w_y=1$, $w_{0x}=-1$, $w_{0y}=0$, $\alpha_2''=0$, and $B=1$. } \label{figEvsalpha} \end{figure} \begin{figure}[tb] \begin{center} \vspace{5mm} \includegraphics[width=0.4\textwidth]{fig11new.eps} \end{center} \caption{(Color online) Dimensionless energy ($E/(w_y\sqrt{\hbar e B_0})$ with $B_0=1$) as a function of magnetic field $B$ at the three quarter Dirac point. It is well fitted as a function of $B^{\frac{4}{5}}$. } \label{figEvsB} \end{figure} \subsection{analytical study with approximation in the magnetic-field- and $\alpha_2$-dependence of the Landau levels at the three-quarter Dirac point} In this subsection we give the analytical derivation of Eq.~(\ref{eqquantizedE}). Taking a sum and a difference, we obtain the equations \begin{align} & \left[ \frac{\alpha_2\sqrt{\hbar eB}}{w_y} {y}^2 -\frac{E}{w_y\sqrt{\hbar eB}} \right] \left(\Psi_1+\Psi_2\right) \nonumber \\ &+ \frac{d}{d{y}} \left( \Psi_1 - \Psi_2 \right) = 0 , \label{eqeqanalA} \\ &\left[ \frac{ 2 w_x }{w_y} y + \frac{\tilde{\alpha}_2\sqrt{\hbar eB}}{w_y} {y}^2 - \frac{E}{w_y \sqrt{\hbar eB}} \right] \left( \Psi_1-\Psi_2 \right) \nonumber \\ &- \frac{d}{d{y}} \left( \Psi_1 + \Psi_2 \right) = 0 . \label{eqeqanalB} \end{align} In the three-quarter Dirac case studied in this paper the term proportional to ${y}$ in Eq.~(\ref{eqeqanalA}) does not exist and the term proportional to ${y}^2$ in Eq.~(\ref{eqeqanalA}) cannot be neglected, while the term proportional to ${y}^2$ in Eq.~(\ref{eqeqanalB}) can be neglected. Then there appear dimensionless parameters $\alpha_2 \sqrt{\hbar eB}/w_y$ and $w_x/w_y$. The energy depends not only the energy scale $w_y\sqrt {\hbar eB}$ but also these dimensionless parameters. Therefore, we may expect \begin{equation} E \propto \left( w_y \sqrt{\hbar e B} \right) \left(\frac{\alpha_2 \sqrt{\hbar eB}}{w_y}\right)^{\beta} \left( \frac{w_x}{w_y} \right) ^{\eta} n^\delta, \label{eqdimE} \end{equation} where $n$ is the quantum number of the almost localized state at $y \lesssim 0$. We determine the exponents, $\beta$, $\eta$ and $\delta$. We take \begin{equation} \beta > 0 , \end{equation} in order to obtain $E \to 0$ as $\alpha_2 \to 0$. The almost-localized state has the finite absolute value of $|\Psi_1+\Psi_2|$ in the region \begin{equation} -{y}_0 \lesssim {y} \lesssim 0, \end{equation} and it is exponentially small in the region \begin{equation} {y} \lesssim -{y}_0, \end{equation} where the dimensionless length ${y}_0$ is determined by the equation \begin{equation} \frac{\alpha_2 \sqrt{\hbar eB}}{w_y} y_0^2 = \frac{E}{w_y \sqrt{\hbar e B}} \sim \left(\frac{\alpha_2 \sqrt{\hbar eB}}{w_y}\right)^{\beta} \left( \frac{w_x}{w_y} \right) ^{\eta} n^{\delta}. \end{equation} Then $y_0$ depends on the dimensionless parameters as \begin{equation} y_0 \sim \left(\frac{\alpha_2 \sqrt{\hbar eB}}{w_y}\right)^{\frac{\beta-1}{2}} \left(\frac{w_x}{w_y} \right)^{\frac{\eta}{2}} n^{\frac{1}{2} \delta}. \end{equation} We expect \begin{equation} \left\langle \left| \frac{d}{dy} (\Psi_1-\Psi_2) \right| \right\rangle \sim \frac{2 c n}{ {y}_0} \Bigl\langle \sqrt{|\Psi_1-\Psi_2|^2} \Bigr\rangle , \label{eqapprox0} \end{equation} where $\langle \cdots \rangle$ is the spacial average in ${y}_0 \lesssim {y} \lesssim 0$ and $c$ is a dimensionless constant of order $1$. This approximation is not justified for small $n$. However, we may consider that $(\Psi_1-\Psi_2)$ changes sign $n$ times in the length of $y_0$, i.e., $(\Psi_1-\Psi_2)$ changes from $\pm c \left\langle \sqrt{|\Psi_1-\Psi_2|^2} \right\rangle$ to $\mp c \left\langle \sqrt{|\Psi_1-\Psi_2|^2} \right\rangle$ periodically in the half period ($y_0/n$). Approximating the oscillation of $(\Psi_1-\Psi_2)$ by a triangle wave, we obtain Eq.~(\ref{eqapprox0}). This crude approximation will give an approximate dependence on $n$ and $y_0$ in Eq.~(\ref{eqapprox0}) in the limit of $n \gg 1$. With this approximation we obtain \begin{equation} \frac{\Bigl\langle \sqrt{|\Psi_1-\Psi_2|^2} \Bigr\rangle}{\Bigl\langle \sqrt{|\Psi_1+\Psi_2|^2} \Bigr\rangle} \sim \left(\frac{\alpha_2\sqrt{\hbar eB}}{w_y} \right)^{\frac{3\beta-1}{2}} \left( \frac{w_x}{w_y} \right)^{\frac{3}{2}\eta} (2 c n)^{\frac{3\delta-2}{2}}, \label{eqbeta1} \end{equation} by taking the spacial average in Eq.~(\ref{eqeqanalA}). Next, we examine Eq.~(\ref{eqeqanalB}) in the same way. The second term and the third term in the coefficient of $\Psi_1-\Psi_2$ in Eq.~(\ref{eqeqanalB}), which depend on the dimensionless parameter as $\left(\alpha_2 \sqrt{\hbar eB}/w_y\right)^{\beta}$, can be neglected with respect to the first term in the coefficient of $\Psi_1-\Psi_2$, since we study the case \begin{equation} \frac{\alpha_2 \sqrt{\hbar eB}}{w_y} \ll \frac{w_x}{w_y}. \end{equation} Then we obtain \begin{equation} \frac{\Bigl\langle \sqrt{|\Psi_1-\Psi_2|^2} \Bigr\rangle}{\Bigl\langle \sqrt{|\Psi_1+\Psi_2|^2} \Bigr\rangle} \sim \left(\frac{\alpha_2\sqrt{\hbar eB}}{w_y} \right)^{1-\beta} \left( \frac{w_x}{w_y} \right)^{1-\eta} (2 c n)^{-\delta-1}. \label{eqbeta2} \end{equation} Comparing Eq.~(\ref{eqbeta1}) and Eq.~(\ref{eqbeta2}), we obtain \begin{equation} \beta=\frac{3}{5}, \end{equation} \begin{equation} \eta=-\frac{2}{5}, \end{equation} and \begin{equation} \delta=\frac{4}{5}. \end{equation} Inserting these exponents in Eq.~(\ref{eqdimE}), we obtain \begin{equation} E \sim w_x^{-\frac{2}{5}}w_y^{\frac{4}{5}} \alpha_2^{\frac{3}{5}} (n \hbar eB)^{\frac{4}{5}}. \label{eqeq22x} \end{equation} In Appendix we give a simpler derivation of Eq.~(\ref{eqeq22x}). This result is consistent with the result obtained by the semiclassical quantization rule in the previous paper\cite{Kishigi2017}, in which the energy is quantized as \begin{equation} A(E_n) \propto (n+ \gamma) B, \end{equation} where $\gamma$ is a phase factor ($\gamma=1/2$ for 2D free electrons and semi-Dirac fermions and $\gamma=0$ for Dirac fermions and three-quarter Dirac fermions) and $A(\varepsilon)$ is the area of the Fermi surface in the 2D $\mathbf{k}$-space at $B=0$ with the Fermi energy $\varepsilon$. The area, $A(\epsilon)$, and the density of states, $D(\epsilon)$, are related by \begin{equation} \frac{1}{4\pi^2} \frac{d A(\varepsilon)}{d \varepsilon} = D(\varepsilon). \end{equation} We plot $A(\varepsilon)$ and $D(\varepsilon)$ in Fig.~\ref{figtilted2}. In the three-quarter Dirac case, we have obtained\cite{Kishigi2017} \begin{equation} A(\varepsilon) \propto \alpha_2^{-\frac{3}{4}} \varepsilon^{\frac{5}{4}}, \end{equation} in the limit $\epsilon \to 0$, and \begin{equation} E_n \propto \alpha_2^{\frac{3}{5}} (nB)^{\frac{4}{5}}. \end{equation} \begin{figure}[bt] \centering {\raggedright \ (a) \hspace{4.0cm} (b) \\ } \vspace{0.5cm} \includegraphics[width=0.2\textwidth]{fig12a.eps} \hfil \includegraphics[width=0.2\textwidth]{fig12b.eps} \caption{(Color online) Schematic plot of the areas of the Fermi surface and the density of states as a function of energy for the three-quarter Dirac fermion, massless 2D Dirac fermion, 2D free fermion, and 2D semi Dirac fermion\cite{Hasegawa2006,Dietl2008}. The density of states are scaled to be 1 at $\epsilon = 1$. } \label{figtilted2} \end{figure} \section{finite energy gap and $n=0$ state} In this section we study the $n=0$ state by introducing the energy gap in the three-quarter Dirac point, which may be caused by a difference of the site energy in two sublattices, \begin{equation} \mathcal{H}_{\textrm{tqD}}^{\prime} = \mathcal{H}_{\textrm{tqD}} + \begin{pmatrix} \Delta & 0 \\ 0 & -\Delta \end{pmatrix}, \label{eqtqDgap} \end{equation} where $2 |\Delta|$ is the energy gap at the three-quarter Dirac point. Note that the minimum of the upper band is not at the three-quarter Dirac point ($\mathbf{q}=0$) and the minimum energy of the upper band is not $|\Delta|$. Then the equation we study at $B \neq 0$ is \begin{align} \begin{pmatrix} (\tilde{\cal{H}}_{\textrm{tqD}} )_{11} + \Delta & (\tilde{\cal{H}}_{\textrm{tqD}} )_{12} \\ (\tilde{\cal{H}}_{\textrm{tqD}} )_{21} & (\tilde{\cal{H}}_{\textrm{tqD}} )_{22} - \Delta \end{pmatrix} \begin{pmatrix} \Psi_1(y) \\ \Psi_2(y) \end{pmatrix} =& E \begin{pmatrix} \Psi_1(y) \\ \Psi_2(y) \end{pmatrix} . \label{eqeq2Delta} \end{align} \begin{figure}[bt] \centering \includegraphics[width=0.4\textwidth]{fig13new.eps} \caption{(Color online) Energy at $B=0$ as a function of $q_x$ and $q_y$ with parameters $w_x =0.4$, $w_y=1$, $\alpha_{2}'=\alpha_{2}''=0.01$, $w_{0x}=-w_x$, $w_{0y}=0$, and $\Delta=\pm0.3$. \\ } \label{fig2qddelx} \end{figure} \begin{figure}[bt] \vspace{0.5cm} \centering \includegraphics[width=0.42\textwidth]{fig14new.eps} \caption{(Color online) Boundary condition $\theta$ at $y=y_{+}=20$ (Eq.~(\ref{eqbcy20})), which makes $|\Psi_{1,2} (y)| \to 0$ at $y \to -\infty$, as a function of the energy. We take parameters $w_x = w_y=1$, $\alpha_{2}'\sqrt{\hbar eB}/w_y=\alpha_{2}''\sqrt{\hbar eB}/w_y=0.01$, $\Delta=0.3$, $0.2$, $\cdots$, $-0.3$, $-0.4$, and $B=1$. } \label{figfig14} \end{figure} \begin{figure}[tb] \centering {\raggedright \ (a) \hspace{3.5cm} (b) \\ } \vspace{0.4cm} \includegraphics[width=0.2\textwidth]{fig15anew.eps} \hfil \includegraphics[width=0.2\textwidth]{fig15bnew.eps} \\ \vspace{0.5cm {\raggedright \ (c) \hspace{3.5cm} (d) \\ } \vspace{0.4cm} \includegraphics[width=0.2\textwidth]{fig15cnew.eps} \hfil \includegraphics[width=0.2\textwidth]{fig15dnew.eps}\\ \caption{(Color online) Wave functions for (a) $\Delta=-0.1$, $E/(w_y\sqrt{\hbar eB})=0.06$, (b) $\Delta=-0.3$, $E/(w_y\sqrt{\hbar eB})= 0.10$, (c) $\Delta=0.1$, $E/(w_y\sqrt{\hbar eB})=0.06$, and (d) $\Delta=0.3$, $E/(w_y\sqrt{\hbar eB})= 0.10$, Other parameters are $w_x = w_y=1$, $\alpha_{2}'\sqrt{\hbar eB}/w_y=\alpha_{2}''\sqrt{\hbar eB}/w_y=0.01$, and $B=1$. Boundary conditions at $y=20$ ($\theta$) are taken as in the case of $\Delta=0$ in section II B. The wave functions have large amplitudes in $y \lesssim 0$ region, when $\Delta<0$ [(a) and (b)], while no peaks are seen in $y \lesssim 0$ region, when $\Delta > 0$ [(c) and (d)]. } \label{figfig15} \end{figure} Although the energy dispersion at $B=0$ does not depend on the sign of $\Delta$, the quantized energies at $B \neq 0$ are not the same for $\pm \Delta \neq 0$. We take $\alpha_2'=\alpha_2''=0.01$ and $B=1$ and we calculate the wave functions numerically with the boundary condition at $y=20$, as in Section \ref{sectionbc1}. We plot the boundary condition $\theta$ to exist a non-divergent solution as a function of energy in Fig.~\ref{figfig14}. For $\Delta \lesssim -0.1$, $\theta$ changes in a narrow region of $E$, which indicate that an almost-localized state exists at $y \lesssim 0$ as shown in Fig.~\ref{figfig15}~(a) and (b), while the variation of $\theta$ as a function of $E$ becomes broad for $\Delta \geq -0.1$, which indicate that an almost-localized state at $y \lesssim 0$ couples strongly to the continuous energy state at $y>0$ and an almost-localized state ceases to exist at $y \lesssim 0$ as shown in Fig.~\ref{figfig15}~(c) and (d). We think that the eigenstate with $n=0$ does not exist when $\Delta>0$, but the almost-localized state exists at $y \lesssim 0$ when $\Delta \leq 0$. The effect of the tunneling would become important as $\Delta$ approaches to zero and the almost-localized $n=0$ state at $y \lesssim 0$ couples strongly to the continuous energy levels in $y>0$. This situation that the $n=0$ mode exists only when $\Delta \leq 0$ is similar to the model studied by Haldane\cite{Haldane1988}, where the zero mode exists either upper band or lower band depending on the sign of the mass, which is $\Delta$ in the present model, and the direction of the magnetic field. In our model the nearly bound state with $n=0$ exists when $\Delta\leq 0$. The $n=0$ ($E=0$) state at $\Delta=0$ in Fig.~\ref{fige0}(a) is understood as the zero-mode of the almost-localized state at three-quarter Dirac point, which couples strongly to the continuous states at $y \gtrsim 0$. Note that the simultaneous changes of $B \leftrightarrow -B$, $y \leftrightarrow -y$, $\Psi_1 \leftrightarrow \Psi_2$, and $\Delta \leftrightarrow -\Delta$ do not change Eq.~(\ref{eqeq2Delta}). \section{Summary} We study the quantized energy at the three-quarter Dirac point in the presence of external magnetic field $B$. We obtain that the quantized energy is proportional to $\alpha_2^{\frac{3}{5}}(nB)^{\frac{4}{5}}$ (Eq.~(\ref{eqquantizedE})) by calculating the solution of the differential equation (Eq.~(\ref{eqeq2})) numerically. We also obtain the approximate result in the limit of $|\alpha_2 \sqrt{\hbar eB} /w_y| \ll 1$ as $E \propto w_x^{-\frac{2}{5}} w_y^{\frac{4}{5}} \alpha_2^{\frac{3}{5}} (n \hbar eB)^{\frac{4}{5}}$ (Eq.~(\ref{eqeq22x})), which is consistent with the result obtained in the previous paper\cite{Kishigi2017} by using the semiclassical quantization rule. We show that the zero mode exists by studying the finite-gap system. Since the three-quarter Dirac points with the finite gap appear as a pair when the time-reversal symmetry is not broken at $B=0$, sign of $\Delta$ is positive at one finite-gap three-quarter Dirac point and negative at another point. As a result, there is one zero mode in the system when $B \neq 0$ and $\Delta \neq 0$. The quantization of the energy in the three-quarter Dirac point in a magnetic field can be observed experimentally in quasi-two-dimensional organic superconductor $\alpha$-(BEDT-TTF)$_2$I$_3$\cite{Kishigi2017} and ultra cold Fermi gas on a tunable optical lattice\cite{Tarruell2012}.

1db6fc28e70e5f700ece7668eb2341099c83fc22

\section{Introduction} For electric conductors placed in an external magnetic field, the Lorentz field felt by electrons can lead to a voltage transverse to the electric current, which is known as the Hall effect \cite{hall1879new}. In magnetic systems, the exchange interaction between electron spin and magnetic moments can give rise to additional topological contributions to the Hall effect. In a ferromagnetic (FM) system with a strong spin-orbit coupling (SOC), the additional Hall contribution is induced by Berry phases accumulated by the adiabatic motion of quasiparticle on the Fermi surface in the momentum space and normally known as (intrinsic) anomalous Hall effect (AHE) \cite{nagaosa2010anomalous,karplus1954hall}. On the other hand, when electrons propagate through chiral magnetic structures, e.g. skyrmions, in the real space, they can also feel Berry phase due to the magnetization texture, leading to the so-called ``topological Hall effect (THE)'' (also known as ``geometric Hall effect'') \cite{machida2007unconventional,kanazawa2015discretized,neubauer2009topological,oveshnikov2015berry,ishizuka2018spin}. Both Hall phenomena originate from Berry phase contribution and thus are topological. An intriguing question is how to understand the topological contribution to the Hall effect in a magnetic skyrmion system with strong SOC, where the Berry phase exists in both the real and momentum spaces. Topological insulator (TI) films with magnetic doping, dubbed ``magnetic topological insulator (MTI)'' below, provide an ideal platform to explore the interplay between AHE and THE. The coexistence of strong SOC and ferromagnetism in MTI films can result in a strong AHE \cite{jungwirth2002anomalous}, and the Hall resistance can even achieve the quantized value when the chemical potential is tuned into the magnetization gap of surface states. Such phenomenon, known as the quantum anomalous Hall (QAH) effect \cite{yu2010quantized, haldane1988model, liu2008quantum, liu2016quantum}, has been experimentally observed in Cr or V doped (Bi,Sb)$_2$Te$_3$ films \cite{chang2013experimental}. Furthermore, the surface states in TI film can also mediate Dzyaloshinsky-Moriya (DM) interaction between magnetic moments due to spin-momentum locking \cite{zhu2011electrically, ye2010spin}. As a result, chiral magnetic structures, such as skyrmion, are also possible. Indeed, recent experiments on Cr-doped-(Bi,Sb)$_2$Te$_3$/(Bi,Sb)$_2$Te$_3$ structure \cite{yasuda2016geometric} and Mn-doped Bi$_2$Te$_3$ \cite{liu2017dimensional} have observed a hump in the Hall resistance hysteresis loop at a small magnetic field. The hump structure is attributed to THE while the Hall hysteresis loop implies AHE. Therefore, the interplay between the AHE from ferromagnetism and the THE from magnetic skyrmion will be substantial to understand the electron transport phenomena in MTI films. In addition, MTI is normally highly disordered due to magnetic doping and it is not well understood how the disorder influences the THE in such strong spin-orbit coupled materials. In this work, we numerically study the magneto-transport of MTI films with a magnetic skyrmion based on a four-band model by combining the iterative Green's function method and the Landauer-Buttiker formalism. Our numerical results suggest that (1) both AHE and THE can coexist in our model system and the total Hall effect can be decomposed into the summation of these two effects; (2) in the clean limit, the topological Hall conductance (THC) almost remains constant but the topological Hall resistance (THR) can increase due to the reduction of longitudinal conductance when the SOC is increasing; (3) in the disorder limit, both the THC and THR are increasing with increasing SOC, while longitudinal conductance is not influenced much by SOC. In addition to numerical simulations, we also studied the scattering cross-section of a skyrmion texture analytically with the second-order Born approximation to provide a more theoretical understanding of this system. Our results are organized as the following. In Sec.\;II, we will describe our model Hamiltonian for the quantum well states in MTI films. In Sec.\;III, we will give our numerical results based on Landauer-Buttiker formalism for the model Hamiltonian and present the corresponding theoretical analysis. The calculation of scattering cross-section will be performed in Sec.\;IV to provide the additional theoretical understanding of the asymmetric scattering for our model Hamiltonian. The disorder effect is numerically calculated and discussed in Sec.\;V. The conclusion will be drawn in Sec.\;VI. \section{ Model Hamiltonian} The TI films can be modeled by a 3D four-band model in a quantum well (QW) with an infinite potential along the $z$ direction \cite{zhang2009topological,liu2010model}. The confinement effect along the $z$ direction can be approximated by choosing $\langle k_z\rangle=0$ and $\langle k_z^2\rangle=(n_b \pi /d)^2$, where $n_b$ is an integer to label the sub-band index and $d$ is the width of the QW \cite{liu2010oscillatory}. As shown in the Appendix\;A \cite{appendix}, we project the 3D four-band model into the subspace spanned by these QW sub-bands and obtain a 2D four-band BHZ-like model given by \eq{ H_0(\bm{k}) = M(\bm{k})\sigma_0\tau_z + \alpha(p_x \sigma_x + p_y \sigma_y)\tau_x, \label{eq:h0} } on the basis $(\ket{+ \uparrow},\ket{+ \downarrow}, \ket{- \uparrow}, \ket{- \downarrow})$, where $\sigma$ and $\tau$ are Pauli matrices for spin and orbital subspaces. $M(\bm{k})=M_0+Bk_\perp^2$ and $\alpha$ labels the SOC strength. The parameter $M_0=M+B_1 (n_b \pi /d)^2$ (See appendix for details) depends on the integer sub-band index $n_b$. Depending on the sub-band index $n_b$ (assuming $M<0$ and $B_1>0$), the four band model for the QW sub-bands can be in the inverted regime if $M_0 B<0$ or in the normal regime $M_0 B>0$. It should be mentioned that the Hamiltonian (\ref{eq:h0}) is block-diagonal with one block set by the basis $(\ket{+ \uparrow}, \ket{- \downarrow})$ and the other block by the basis $(\ket{+ \downarrow}, \ket{- \uparrow})$. These two blocks are related to each other by time-reversal symmetry and are degenerate. This degeneracy will be broken when introducing ferromagnetism or magnetic skyrmion into the system. Since we focus on the transport regime dominated by these QW states in this work, we expect that multiple QW sub-bands with different $n_b$ will be present at the Fermi energy. To simplify the problem, we treat QW sub-bands in the Hamiltonian (Eq.\;\ref{eq:h0}) with different $n_b$ independently. Thus, we may choose $M_0$ as an independent parameter and discuss below the transport behaviors for the parameter $M_0$ in different regimes. The SOC term ($\alpha$ term) couples the state $\ket{+ \uparrow}$ $(\ket{+ \downarrow})$ to the state $\ket{- \downarrow}$ $(\ket{- \uparrow})$ in different orbital basis, which is different from the conventional Rashba SOC where the SOC term couples different spin states in the same orbital basis. Here we adopted the Hamiltonian form (Eq.\;\ref{eq:h0}) in Ref.\;\onlinecite{zhang2009topological}, which is equivalent to the more standard Hamiltonian form given in Ref.\;\onlinecite{liu2010model} up to a unitary transformation $U=\textrm{Diag} (1,1,-i,i)$. For MTI, the exchange interaction between electron spin and magnetic moment is given by \eq{ H_{\textrm{ex}}(x,y) = \bm{m} (x, y) \cdot \bm{\sigma} \tau_0, \label{eq:Hex} } where $\bm{m}(x,y)$ represents the magnetization. The magnetic skyrmion texture can be taken into account by choosing the $\bm{m}$ configuration \eq{ \begin{split} \bm m &= m_0(\sin \theta \cos \phi ,\sin \theta \sin \phi ,\cos \theta )\\ \theta &= \pi \tanh \left(\frac{\rho }{R}\right) \\ \phi &= n \varphi + \eta, \label{eq:skyrmionmag} \end{split} } where $m_0$ represents the magnetization strength, $\theta$ and $\phi$ label the magnetization direction, $\rho$ and $\varphi$ define the spatial polar coordinates with $(x,y)=(\rho\cos\varphi,\rho\sin\varphi)$, and $R$ is the radius of the skyrmion. The chirality of the skyrmion is characterized by the integer number $n$, which is chosen to be $+1$ (a single skyrmion) or $-1$ (a single anti-skyrmion) below. The parameter $\eta$ denotes the helicity phase, which is an irrelevant parameter. It should be pointed out that the unitary transformation $U$ should be applied to the Hamiltonian (\ref{eq:Hex}) in order to be consistent with the Hamiltonian (\ref{eq:h0}). However, we find the THE only depends on the chirality of skyrmion texture, which is unchanged under the transformation $U$, and thus we can still use the Hamiltonian (\ref{eq:Hex}) to describe skyrmion texture. Physically, the magnetic skyrmion can be energetically stabilized by the interplay of the Zeeman coupling and DM interaction in MTI films \cite{koshibae2016theory}. In this study, we assume the skyrmion structure in our system (Fig.\;\ref{fig:config}(a)) and focus on the influence of skyrmion on magneto-transport. \begin{figure}[ht] \centering \includegraphics[width=0.35\textwidth]{figs/config.png} \includegraphics[width=0.23\textwidth]{figs/spectra.png} \includegraphics[width=0.23\textwidth]{figs/spectrainv.png} \caption{(a) Schematic configuration of the system. The square region in the center represents the sample, and the extended transparent edges represent semi-infinite FM leads. A skyrmion of $m=1, \eta=\pi/2, R = 0.4 L$ is shown at the origin. Cones point to the local magnetic moment direction. (b, c) Band dispersions for the Hamiltonian Eq.\;\ref{eq:hamiltonian} for two parameter sets (i) and (ii) (see main text). A periodic boundary condition is applied on the $y$ direction. The shaded regions show the energy range for the transport calculations in Fig.\;\ref{fig:the01m1} to \ref{fig:THEenhancement}. } \label{fig:config} \end{figure} Due to the absence of translation symmetry in a system with a single skyrmion, we numerically explore magneto-transport directly in the real space. To perform such calculation, we implement the tight-binding regularization on the Hamiltonian (\ref{eq:h0}) and (\ref{eq:Hex}), which is given by \eq{ \begin{split} \hat{H} &= \hat{H}_0 + \sum_i \Psi_i^\dagger H_{\textrm{ex}} \Psi_i ,\\ \hat{H}_0 &= \sum_i \Psi_{i}^\dagger \epsilon_i \Psi_{i} + \sum_{\langle i,j\rangle}(V_{ij} \Psi_{j}^\dagger \Psi_{i} + \text{h.c.}), \label{eq:hamiltonian} \end{split} } where $\Psi^\dagger_{i}$ presents $(c^\dagger_{i,+\uparrow}, c^\dagger_{i,+\downarrow}, c^\dagger_{i,-\uparrow}, c^\dagger_{i,-\downarrow})$ at the position $ i = (i_x,i_y)$. $\epsilon_i$ is the on-site energy and $V_{ij}$ is the hopping matrix between the nearest neighbors $i$ and $j$. Both $\epsilon_i$ and $V_{ij}$ are 4 by 4 matrices and their detailed forms can be related to those in the continuous model (Eq.\;\ref{eq:h0}), as listed in the Appendix\;B\cite{appendix}. The consistency between the tight-binding Hamiltonian (\ref{eq:hamiltonian}) and the continuous Hamiltonian (\ref{eq:h0}) and (\ref{eq:Hex}) is also discussed in the Appendix\;B\cite{appendix}. We consider a 2D square lattice with the side length $L$. The skyrmion texture is located at the center of the lattice, as shown in Fig.\;\ref{fig:config}(a). Four semi-infinite leads with the width $L$, labeled as 1 to 4 in Fig.\;\ref{fig:config}, are attached to each side of the square lattice. We adopt the recursive Green's function method \cite{datta1997electronic} to evaluate the transmission coefficient $T_{pq}$ between the leads $p$ and $q$ ($p,q=1,2,3,4$). The relationship between currents and voltages is calculated using the Landauer-B\"uttiker formalism \begin{eqnarray}\label{eq:LB} I_p=e^2/h \sum_{q\neq p}\left(T_{qp}V_p-T_{pq}V_q\right). \end{eqnarray} Due to the charge conservation of the whole system, the matrix $T$ is singular. Without loss of generality, we set $V_4=0$ and remove the corresponding column/row in $T$. To set up a Hall configuration, we consider a current flow from the lead 1 to 3, as setting $I_1 = -I_3 = I$ and $I_2=0$. Voltages of the leads are calculated through $(V_1,V_2,V_3)^\intercal = \frac{h}{e^2}T^{-1} (I_1,I_2,I_3)^\intercal$, where the matrix $T$ is the $3\times 3$ transmission matrix. The longitudinal resistance $R_{xx}$ and the Hall resistance $R_{xy}$ can be extracted by $R_{xx} = (V_1-V_3)/I$ and $R_{xy} = V_2/I$. \section{Numerical results and Analysis in the Clean limit} \begin{figure*}[tp] \centering \includegraphics[width=0.45\textwidth]{figs/the01m1.png} \includegraphics[width=0.45\textwidth]{figs/the01m1inv.png} \caption{(a1, b1) Red, blue and yellow lines represent Hall resistance in FM, skyrmion and anti-skyrmion cases as the function of Fermi energy, for the parameter sets (i) and (ii), respectively. SOC strength $\alpha/t=2$. (a2, b2) Blue and yellow lines are the extracted AHE and THE contribution $R_{\textrm{AHE}}$ and $R_{\textrm{THE}}$ from Eq.\;\ref{eq:Hall_skyrmion_decomposition}. Red lines are the same in (a1, b1) representing the FM contribution $R_{\textrm{FH}}$ as a comparison. Note that $R_{\textrm{AHE}}$ and $R_{\textrm{FH}}$ are close in value in a wide range of energy, suggesting the skyrmion texture has a minimum reduction on the AHE effect. } \label{fig:the01m1} \end{figure*} \begin{figure*}[tp] \centering \includegraphics[width=0.45\textwidth]{figs/theg01m1.png} \includegraphics[width=0.45\textwidth]{figs/theg01m1inv.png} \caption{The same system as Fig.\;\ref{fig:the01m1} expressed in conductances. (a1, b1) the Hall conductances for the parameter sets (i) and (ii), respectively. (a2, b2) the extracted decomposition into AHE and THE contribution. The ferromagnetic system is marked with red lines. Note that $G_{\textrm{FM}}$ is also close to $G_{\textrm{AHE}}$, showing that the decomposition approximation works for both resistance and conductance. } \label{fig:theg01m1} \end{figure*} Fig.\;\ref{fig:the01m1}(a1) and (b1) reveal the Hall resistance as a function of the Fermi energy in different parameter regimes. Here we choose a system of length $L =80 a_0$ and Skyrmion radius $R = 0.4 L$. Hopping parameters are expressed in term of the nearest-neighbor hopping strength $t \equiv B/a_0^2$ (See Appendix\;B\cite{appendix}). Here we consider two sets of parameters, one for the trivial regime, denoted as (i), and the other for the QAH regime, denoted as (ii) below. While all the other parameters are the same ($m_0=4/3,\alpha/t=2, a_0=1, B=1$) for the parameter sets (i) and (ii), the parameter $M_0$ is chosen to be different ($M_0 = 2$ for (i) and $M_0=1$ for (ii)). To see the topological property of the full Hamiltonian with these two parameter sets, we may consider the FM case with $\bm{m}(x,y)=m_0(0,0,1)$. For the parameter set (i), we notice that both blocks of bands in the Hamiltonian (\ref{eq:hamiltonian}) are in the normal regime since $(M_0\pm m_0)B>0$. In contrast, the system for the parameter set (ii) is in the QAH regime since one block is in the normal regime $(M_0+m_0)B>0$ while the other is in the inverted regime $(M_0-m_0)B<0$. The corresponding energy dispersions for a slab configuration with these two parameter sets are shown in Fig.\;1(b) and (c), from which one can see a full gap for the parameter set (i) and gapless chiral edge states appear in the bulk gap for the parameter set (ii). In this work, we focus on the transport behavior of the metallic regime when the Fermi energy $E_f$ crosses one valence band top ($E_f/t \in [-2, -0.7]$ for the parameter set (i) and $[-2.6,-1]$ for the parameter set (ii) in Fig.\;1(b) and (c)). For the purpose of the quantized conductance within the gap, the parameter set (i) and (ii) represent a comparison between a trivial and a non-trivial gap. The difference is briefly discussed in Appendix\;D\cite{appendix}. For the transport calculation, three magnetic configurations, namely ferromagnetism ($\hat m = -m_0 \hat z$), a skyrmion ($n=+1$) and an anti-skyrmion ($n=-1$) are considered and the corresponding Hall resistances $R_{\textrm{FM}}$ , $R_{\textrm{Sk},+1}$ and $R_{\textrm{Sk},-1}$ are shown by the yellow, red and blue lines in Fig.\;\ref{fig:the01m1}(a1) and (b1) for the parameter sets (i) and (ii), respectively. One can clearly see that $R_{\textrm{Sk},+1}$ is much larger than $R_{\textrm{FM}}$, while $R_{\textrm{Sk},-1}$ has the opposite sign. For the FM case, the Hall resistance only originates from the AHE, while in the skyrmion cases, both THE and AHE can contribute due to the coexistence of strong SOC and chiral magnetic structure. We expect the THE (AHE) is dependent (independent) on the chirality of the skyrmions. Therefore, we can decompose the Hall resistances $R_{\textrm{Sk},\pm 1}$ into chirality dependent part, $R_{\textrm{THE}}$, and independent part, $R_{\textrm{AHE}}$, \begin{equation} R_{\textrm{Sk},n}=R_{\textrm{AHE}}+ n R_{\textrm{THE}}, \label{eq:Hall_skyrmion_decomposition} \end{equation} where the index $n$ stands for the chirality of the skyrmion. Based on the decomposition of Eq.\;\ref{eq:Hall_skyrmion_decomposition}, Fig.\;\ref{fig:the01m1}\;(a2) and (b2) depict $R_{\textrm{AHE}}$ (blue line) and $R_{\textrm{THE}}$ (yellow line) as a function of Fermi energy for the parameter sets (i) and (ii), respectively. In addition, $R_{\textrm{FM}}$ is shown by the red line. Fig.\;\ref{fig:the01m1}\;(a2) and (b2) show the following features. (1) $R_{\textrm{FM}}$ generally shows a similar behavior as the blue line of $R_{\textrm{AHE}}$ (except that the Fermi energy is close to band gap), and thus the magnetic skyrmion does not have a strong influence on the AHE in the metallic regime and validates the decomposition of the Hall resistance. (2) $R_{\textrm{AHE}}$ is much larger for the parameter set (i) than that for (ii), due to the energy range (shaded area in Fig.\;\ref{fig:config}(b,c)) is closer to band center for parameter set (ii). (3) For both parameter sets, we notice that $R_{\textrm{THE}}$ increases rapidly when the Fermi energy is tuned towards the valence band top. We next turn to the Hall conductance with the same parameters, as shown in Fig.\;\ref{fig:theg01m1}. Here the red, blue and yellow lines are for the Hall conductance with the FM, skyrmion ($n=+1$) and anti-skyrmion ($n=-1$) configurations in Fig.\;\ref{fig:theg01m1}\;(a1) and (b1) for two parameter sets. A similar decomposition \begin{equation} G_{\textrm{Sk},n}=G_{\textrm{AHE}}+n G_{\textrm{THE}} \label{eq:Hall_skyrmion_decomposition_conductance} \end{equation} is considered and the corresponding $G_{\textrm{AHE}}$ and $G_{\textrm{THE}}$ are plotted in Fig.\;\ref{fig:theg01m1}\;(a2) and (b2), which show the following features. (1) The decomposition of Hall conductance also remains valid in most energy ranges, as indicated by the coincidence between $G_{\textrm{AHE}}$ and $G_{\textrm{FM}}$ in most energy ranges. (2) In contrast to Hall resistance, the Hall conductance is almost a constant in the whole metallic region for both parameter sets. (3) $G_{\textrm{AHE}}$ (or $G_{\textrm{FM}}$) is smaller for the parameter set (i) compared to that for (ii) while $G_{\textrm{THE}}$ is comparable for both parameter sets. Fig.\;\ref{fig:THEenhancement}(a) and (b) ((c) and (d)) reveal the THE contribution $R_{\textrm{THE}}$ ($G_{\textrm{THE}}$) from the decomposition of Eq.\;\ref{eq:Hall_skyrmion_decomposition} (Eq.\;\ref{eq:Hall_skyrmion_decomposition_conductance}) as a function of the Fermi energy for different SOC strength $\alpha$ for both parameter sets. An enhancement of $R_{\textrm{THE}}$ is found while $G_{\textrm{THE}}$ remains almost unchanged when increasing the SOC strength or the Fermi energy for both parameter sets. $G_{\textrm{THE}}$ is only found to drop when the Fermi energy is close to the band gap (insulating regime). \begin{figure}[htp] \centering \includegraphics[width=0.5\textwidth]{figs/THERG.png} \caption{(a) and (b) show THR $R_{\textrm{THE}}$ as the function of Fermi energy and SOC strength $\alpha$, for the parameter sets (i) and (ii), respectively. (c) and (d) show THC $G_{\textrm{THE}}$ decomposed in the similar way. } \label{fig:THEenhancement} \end{figure} To understand our numerical results, we will next present a theoretical analysis of the transport behavior of the model based on our numerical simulation of the Landauer-Buttiker formalism. The symmetry property of the transmission matrix $T_{pq}$ in the Eq.\;\ref{eq:LB} will be first analyzed. For the FM and $n=+1$ skyrmion cases, the system respects the $C_4$ rotation symmetry, while for the anti-skyrmion with $n=-1$, the system possesses the $S_4$ improper rotation symmetry. In all cases, we find the transmission matrix elements can be characterized by three independent parameters based on the following relations \eq{ \begin{split} T_{13} = T_{24} = T_{31} = T_{42} \equiv -a \\ T_{14} = T_{21} = T_{32} = T_{43} \equiv -b \\ T_{12} = T_{23} = T_{34} = T_{41} \equiv -c \\ \end{split} } where $a$, $b$ and $c$ can be understood as the probability of electronic modes going straight, turning left and turning right, respectively, after they entered the spin-textured structure from any of the leads. With this simplification, direct calculations from the Landauer-B\"uttiker formalism give rise to the Hall and longitudinal resistance as \eq{ \begin{split} R_{xy} &= \frac{(b-c)}{2 a^2+2 a (b+c)+b^2+c^2} \approx \frac{b-c}{2 (a+b)^2} \equiv \frac{\Delta}{2 \beta^2} \\ R_{xx} &= \frac{(2 a+b+c)}{2 a^2+2 a (b+c)+b^2+c^2} \approx \frac{1}{a+b} \equiv \frac{1}{\beta}, \label{eq:resistance_analytical} \end{split} } where we define the parameters $\beta \equiv a+b$ to be the transmission probability and $\Delta \equiv (b-c)$ to be the asymmetric scattering between the left and right directions. We further assume $\Delta\ll b\approx c$, which can be justified based on our numerical calculations for both parameter sets. The corresponding Hall and longitudinal conductance is \eq{ \begin{split} G_{xy} \approx -\Delta/2 \\ G_{xx} \approx \beta. \label{eq:conductance_analytical} \end{split} } \begin{figure}[htp] \centering \includegraphics[width=0.5\textwidth]{figs/ABD.png} \caption{(a), (b) and (c) show $\Delta$ as a function of $E$ for FM, skyrmion and antiskyrmion cases. (d), (e) and (f) reveal the energy dependence of $\beta$ for FM, skyrmion and antiskyrmion cases. Here we consider the parameter set (i). } \label{fig:ABD} \end{figure} \begin{figure}[htp] \centering \includegraphics[width=0.5\textwidth]{figs/DeltaDecompose.png} \caption{(a) and (b) show the decomposition of $\Delta$ into the THE ($\Delta_{\textrm{THE}}$) and AHE part ($\Delta_{\textrm{AHE}}$), respectively. Here we choose the parameter set (i). } \label{fig:DeltaDecompose} \end{figure} Eq.\;(\ref{eq:resistance_analytical}) and (\ref{eq:conductance_analytical}) are the basis for the analysis below. We can see that $\beta$ is related to the forward transmission and determines the longitudinal conductance $G_{xx}$ while $\Delta$ represents the asymmetry between the left and right scattering and determines the Hall conductance $G_{xy}$. In Fig.\;\ref{fig:ABD}, we demonstrate the behaviors of $\beta$ and $\Delta$ for the parameter set (i) as an example, from which we can understand the behaviors of AHE and THE. Below we will analyze the behavior of $\Delta$ first (Fig.\;\ref{fig:ABD} (d) - (f)). In Fig.\;\ref{fig:ABD}(d), one can see that $\Delta$ increases when increasing the SOC parameter $\alpha$ in the FM case. From Fig.\;\ref{fig:ABD}(e) and (f), we find the value of $\Delta$ is much larger when there is a skyrmion or anti-skyrmion as compared to the FM case. We may also consider a decomposition $ \Delta_{\textrm{Sk},n}=\Delta_{\textrm{AHE}}+ n \Delta_{\textrm{THE}}$ and the corresponding $\Delta_{\textrm{THE}}$ and $\Delta_{\textrm{AHE}}$ are plotted in Fig.\;\ref{fig:DeltaDecompose}(a) and (b), respectively. One can see that all the curves for $\Delta_{\textrm{THE}}$ fall into one line and thus are independent of the SOC parameter $\alpha$, while $\Delta_{\textrm{AHE}}$ increases rapidly with $\alpha$. From Eq.\;(\ref{eq:conductance_analytical}), we expect that $G_{xy}$ exhibits a similar behavior as $\Delta$, which was indeed revealed in Fig.\;\ref{fig:THEenhancement}(c). Therefore, we conclude that SOC mainly increases the AHE contribution, but has little influence on the THE contribution to our four-band model in the clean limit. Next, let us analyze the behavior of transmission $\beta$. For all three spin textures, the values of $\beta$ and the dependence of $\beta$ on SOC and the Fermi energy are quite similar, and thus we would not specify the magnetic texture for the discussion below. As expected, $\beta$ is decreasing when tuning Fermi energy to the valence band top (more insulating). In addition, we find a rapid decreasing of $\beta$ when increasing the SOC parameter $\alpha$ in Fig.\;\ref{fig:ABD}. This can be understood as the following. SOC tends to induce precession of electron spin and rotate it from the easy axis set by local magnetization. As a consequence, a strong scattering can be induced by the exchange coupling between electron spin and magnetic moments, and thus reduces transmission. It turns out that the reduction of transmission $\beta$ has a substantial influence on the behavior of Hall resistance $R_{xy}$. From Eq.\;\ref{eq:resistance_analytical}, we can see that $R_{xy}$ depends on the ratio between $\Delta$ and $\beta^2$. Therefore, although increasing SOC does not enhance $\Delta$, it reduces $\beta$, and thus increases $R_{xy}$ as shown in Fig.\;\ref{fig:THEenhancement}. (For a full figure of $R_{xy}$ for all SOC and spin textures, please see Appendix Fig.\;C1.) This analysis leads to the following conclusion for our model in the clean limit: (1) SOC does not have much influence on THE and (2) the behavior of $R_{\textrm{THE}}$ is mainly determined by the forward transmission $\beta$, rather than the asymmetric scattering $\Delta$. \section{Analytical results of cross section} In this section, we will provide more physical understanding on THE for our four-band model by analytically calculating the differential cross section of this system. We notice that topological surface states scattered by a magnetic skyrmion have been studied in Ref. \onlinecite{denisov2016electron, araki2017skyrmion}, while we focus on bulk QW states here. Due to the presence of spin-polarized background $S_z=\hat{z}$, we can treat $\bar{H}_0=H_0-m_0\sigma_z\otimes\tau_0$ as the unperturbed Hamiltonian, and take $\hat{V}(\bm{r}) =- (\bm{m}(\bm{r})\cdot\bm{\sigma}-m_0\sigma_z)\otimes\tau_0$ as the perturbation (scattering potential), where $\bm{m}(\bm{r})$ has been defined above in Eq. (\ref{eq:skyrmionmag}) and $m_0$ can be regarded as the exchange coupling strength between the conduction electron and local magnetic moment. The differential cross section of electron scattering is given by \begin{equation} \left(\frac{d\sigma(\phi)}{d\phi}\right)_{\alpha\beta} = |F_{\alpha\beta}(\bm{p},\bm{p'})|^2 \end{equation} where $\phi$ is the scattering angle. $\psi_{{\bm p}'\alpha}$ and $\psi_{{\bm p}\beta}$ are eigenstates of $\bar{H}_0$ that describe the incident and scattered states respectively with ${\bm p}$ and ${\bm p}'$ is the associated momenta. $F_{\alpha\beta}=F_{\alpha\beta}(\bm{p},\bm{p'}) = \langle\psi_{\bm{p'}\alpha}|\hat{V} + \hat{V}\hat{G}_0\hat{V}|\psi_{\bm{p}\beta}\rangle$ is the scattering amplitude up to the second order Born approximation, where $\hat{G}_0$ is the Green's function associated with unperturbed Hamiltonian $\bar{H}_0$. Asymmetric component of $\left(\frac{d\sigma(\phi)}{d\phi}\right)_{\alpha\beta}$ with respect to $\phi$, being responsible for the Hall response, arises from cross-terms between the first and second Born approximation. A major difficulty in this calculation is the computation of $\langle\psi_{\bm{p'}\alpha}|\hat{V}\hat{G}_0\hat{V}|\psi_{\bm{p}\beta}\rangle$ due to the Bessel function-like Green's function in 2D. Here we use the momentum representation so that $ \langle\psi_{\bm{p'}\alpha}|\hat{V}\hat{G}\hat{V}|\psi_{\bm{p}\beta}\rangle = \int d\bm{p}_1 \langle\alpha_{\bm{p'}}|V(\bm{p'}-\bm{p}_1)G(\bm{p}_1)V(\bm{p}_1-\bm{p})|\beta_{\bm{p}}\rangle$. For simplicity, we use a Bloch skyrmion configuration and let its polar angle in Eq. (\ref{eq:skyrmionmag}) be $\theta=\pi\exp(-r/a)$, where $a$ is the radius of the skyrmion. In the analytical calculations, Bessel functions of $qr$ will be used throughout the whole calculation, where ${\bm q}={\bm p}-{\bm p}'$. In the small angle scattering assumption, $qr\ll1$, so that we can expand the Bessel functions with Fourier series and keep the lowest order terms. In the current calculation, we are interested in the situation where the Fermi surface intersects only with the lowest electron band of $\bar{H}$. Direct calculation shows that the asymmetric part of the corresponding differential cross section is given by \begin{eqnarray} \left(\frac{d\sigma}{d\phi}\right)^A\propto a^4D(\varepsilon_F)m_0^3\left[\left(1-\cos\omega\right)\cos^4\frac{\omega}{2}\right.\nonumber\\ \left. + \left(1+\cos\omega\right)\sin^4\frac{\omega}{2}\right]\sin\phi \end{eqnarray} where $D(\varepsilon_F)$ is the density of states at the Fermi energy, and $\omega=\arccos [(Bp^2+M_0-m_0)/\varepsilon_F]$. The cross-section as a function of scattering angle $\phi$ for different $m_0$ are shown in Fig.\;\ref{fig:CrossSection}. With a small or intermediate exchange coupling strength $m_0$, we find that the asymmetric component of the differential cross section increases with the SOC parameter $\alpha$, and eventually saturate at large $\alpha$ limit, as shown in Fig.\;\ref{fig:CrossSection}(a). On the other hand, when $m_0$ is large, the influence of SOC parameter $\alpha$ becomes negligible due to the dominant role of exchange coupling in inducing THE in this regime, as shown in Fig.\;\ref{fig:CrossSection}(b), and our numerical results are consistent with the analytical result in this regime. \begin{figure}[htp] \centering \includegraphics[width=0.5\textwidth]{figs/JH.png} \caption{Asymmetric part of the differential cross section as a function of scattering angle for (a) $m_0=5t$ and (b) $m_0=80t$. } \label{fig:CrossSection} \end{figure} \section{Disorder effect} We next examine the disorder effect on THE in MTI films, as shown in Fig.\;\ref{fig:Disorder}(a-d), which reveal new features compared to Fig.\;\ref{fig:THEenhancement} in the clean limit. To consider the disorder effect, we introduce a spin-independent uniformly-distributed random on-site potential term $H_{\textrm{d}}=\sum_i \Psi^\dagger_i V_{\textrm{d},i}\Psi_i$ where $V_{\textrm{d},i} = \textrm{Diag} (\mathcal{V}_{+}, \mathcal{V}_{+}, \mathcal{V}_{-}, \mathcal{V}_{-})$ and $\mathcal{V}_{\pm} \in \left[-V_{\textrm{imp}}/2, V_{\textrm{imp}}/2\right]$ with $0<V_{\textrm{imp}}\leq 2t$ chosen in our calculations. All the calculations are performed with the disorder average over 160 samples. After such a sample average, the uncertainty (dictated by the error bars in Fig.\;\ref{fig:THEenhancement}) is much smaller than its mean value. We implement a similar decomposition of Hall conductance and resistance, as specified in Eq.\;(\ref{eq:Hall_skyrmion_decomposition}) and (\ref{eq:Hall_skyrmion_decomposition_conductance}) for each individual run, and the disorder-averaged Hall resistance ($R_{\textrm{THE}}$) and conductance ($G_{\textrm{THE}}$) from the THE contribution are revealed in Fig.\;\ref{fig:Disorder}(a, b) and (c, d) for two parameter sets (i) and (ii), respectively. Here the circle, square and diamond label different disorder strength $V_{\textrm{imp}}=1, 1.5$ and $2$ in the unit of $t$, while the green and black colors are for different SOC strength ($\alpha/t = 0$ and $\alpha/t = 2$). The solid lines show the results in the clean limit for comparison. With increasing the disorder strength, one can clearly see the decreasing of both Hall resistance $R_{\textrm{THE}}$ and conductance $G_{\textrm{THE}}$. A striking feature emerges in the Hall conductance $G_{\textrm{THE}}$ when increasing disorder strength. $G_{\textrm{THE}}$ is unchanged for different SOC strength in the clean limit, as seen by the coincidence between black and green solid lines in Fig.\;\ref{fig:Disorder} (c) and (d). In contrast, for intermediate or strong disorder strength, $G_{\textrm{THE}}$ at a large SOC $\alpha/t = 2$ can be much larger than that at zero SOC, as shown by the green and black markers in Fig.\;\ref{fig:Disorder}c and d. This suggests that SOC can stabilize the THE against disorder scattering. We also analyze the disorder-averaged forward transmission $\beta$ and asymmetric scattering $\Delta$, shown in Fig.\;\ref{fig:DisorderABD}(a) and (b). Interestingly, we find that with increasing disorder scattering, forward transmission $\beta$, although being reduced for both SOC strengths, becomes comparable for $\alpha/t=0$ and $\alpha/t=2$ when $V_{\textrm{imp}}$ is increased above $1.5t$. This suggests that the mean-free path of electrons is determined by disorder, rather than SOC, at this disorder strength $V_{\textrm{imp}}/t=2$. In contrast, although the asymmetric scattering parameter $\Delta$ is reduced for both SOC strengths, its reduction is much slower when the SOC $\alpha$ is strong, which can be clearly seen by the fact that the green markers ($\alpha/t=2$) are above the black markers ($\alpha/t=0$) in Fig.\ref{fig:DisorderABD} (b). In contrast to the clean limit, in which SOC only enhances $R_{xy}$ but not $G_{xy}$ due to the reduction of transmission $\beta$, SOC mainly influences the asymmetric scattering $\Delta$ and thus will enhance both $R_{xy}$ and $G_{xy}$ in the disordered limit. Therefore, we conclude that the THE is stabilized by SOC in the disordered limit. \begin{figure}[htp] \centering \includegraphics[width=0.5\textwidth]{figs/Disorder.png} \includegraphics[width=0.5\textwidth]{figs/DisorderConductance.png} \caption{(a) and (c) reveal $R_{\textrm{THE}}$ and $G_{\textrm{THE}}$ for the parameter set (i) under different disorder strength ($V_{\textrm{imp}}=0, 1, 1.5, 2$ in unit of $t$), respectively. Green and black colors represent $\alpha/t = 2$ and $0$ cases, respectively. (b) and (d) are the same as (a)(c), except that we choose the parameter set (ii). } \label{fig:Disorder} \end{figure} \begin{figure}[htp] \centering \includegraphics[width=0.5\textwidth]{figs/DisorderABD.png} \caption{(a) and (b) reveal $\bar{\beta}$ and $\Delta_{\textrm{THE}}$ as a function of $E$ in both disordered and clean limit. Here the colors represent different SOC strengths (green for $\alpha/t = 2$ and black for $\alpha/t=0$), and the line, circle, square and diamond represent $V_{\textrm{imp}}=0, 1, 1.5, 2$, respectively. } \label{fig:DisorderABD} \end{figure} \section{Discussion and Conclusion} In summary, we have studied the AHE and THE in an MTI model with skyrmion configuration and revealed how the magneto-transport behaviors in such systems are influenced by SOC, Fermi energy and disorder through numerical calculations and theoretical analysis. In particular, our calculations demonstrate the importance of disorder effect in determining the role of SOC in the THE. Given the recent experimental efforts in MTI systems \cite{yasuda2016geometric,liu2017dimensional}, our numerical and theoretical results will provide a physical understanding of these magneto-transport measurements and may stimulate further experimental studies. It should be pointed out that the SOC term we used here preserves inversion symmetry and thus is different from the Rashba SOC, which breaks inversion symmetry and responses for DM interaction. Including Rashba SOC in the calculation may bring new features and will require further studies, which is beyond the scope of the current work. In light of the importance of the interplay between disorder scattering and SOC, it will be important to develop a more analytical theory (such as the Boltzman equation and diagram expansion calculation \cite{nagaosa2010anomalous}) to take into account random scattering of multiple magnetic skyrmions or skyrmion lattice, SOC and disorder scattering, which will be another future direction. We would like to point out that although the current calculation is on a specific skyrmion texture, the presence of transverse scattering should exist for any spin textures, including random state, with net chirality\cite{hou2017thermally}. It will be interesting to generalize this work to the investigation of THE-AHE crossover in other chiral systems. \section{Acknowledgement} We acknowledge the discussion with C.Z. Chang and M.H.W. Chan. C.X.L and J.X.Z acknowledge support from the Office of Naval Research (Grant No. N00014-15-1-2675 and renewal No. N00014-18-1-2793) and the U.S. Department of Energy (DOE), Office of Science, Basic Energy Sciences (BES) under award No. DE-SC0019064. Work at UNH was supported by the U.S. Department of Energy (DOE), Office of Science, Basic Energy Sciences (BES) under Award No. DE-SC0016424.

ac780c164215bc45b02ef5d3acd6f094e7149fe7

\section{Introduction} Application of the random matrix theory (RMT) to the statistics of the energy spectra by Wigner and Dyson was originally motivated by the complexity of spectra of heavy nuclei consisting of many strongly interacting particles. However about 1980 it became clear that quantum systems with just two degrees of freedom also have universal statistical spectral properties described by RMT provided the corresponding classical motion is chaotic, for an overview see Refs. \cite{Stockmann,Haake,Gutzwiller,Guhr}; this statement was formulated as the famous Bohigas-Giannoni-Schmit conjecture \cite{BGS}. Its understanding took many investigations, to mention but a few relevant here: the diagonal approximation \cite{Berry}, a heuristic derivation of the oscillatory part of the correlation function for the unitary unversality class \cite{BogK3}, the orbit partnership role in spectral correlations \cite{Sieber}, the semiclassical treatment of spin in the level statistics \cite{Bolte, Heuslerspin, BolteHar}, the periodic orbit theory of the small time form factor \cite{MullerPhysRev}. The complete semiclassical derivation of the two-point correlation function for the unitary and orthogonal universality classes is given in \cite{Heusler, Muller} and for the symplectic class in \cite{Braun}. Three- and higher-level correlation functions are considered in \cite{NagaoMuller}. The state-of-the-art calculation of spectral correlators is based on the method of generating functions imported from the field theory of disorded media, which are ratios of spectral determinants $\det\left( E+e-H\right)$ with slightly different energy offsets $e$ averaged over the central energy $E$. The fundamental role in the calculation is played by the so called Riemann-Siegel lookalike \cite{Keatin,KeaMul}. This is the asymptotic representation of a spectral determinant by an explicitly real sum over sets of periodic orbits, or \textquotedblleft pseudo-orbits\textquotedblright, truncated at pseudo-orbits with total period half the Heisenberg time $T_{H}$. The result is far from obvious and has been proven only for some rare cases \cite{Wal,Har}, beyond the Riemann-Siegel formula proper in the theory of the Riemann zeta function \cite{Tit,Edw,Ber500}. Its usage in problems of spectral statistics looks also like an overkill since the non-trivial boundary $T_{H}/2$ doesn't play any role; after the averaging, summation over the pseudo-orbits is invariably extended to infinite periods. Here we want to reformulate the theory in the form which avoids the use of the Riemann-Siegel lookalike. The idea is to complexify the generating function by letting its energy arguments have an imaginary part large enough for convergence of the pertinent Gutzwiller expansions like the Gutzwiller trace formula. It can be done in two ways depending on the sign of the imaginary parts of the energies $e$ which, after averaging, provides two different semiclassical asymptotics which can be analytically continued inside the critical strip. We will refer to these two ways as I and II, for an illustration, see Fig.\ \ref{fig} a). \begin{figure} \raisebox{1cm}{\includegraphics[width=8cm]{applrangeI.eps}}\includegraphics[width=6cm]{applrangeII.eps}\label{fig} \caption{Analogy between the semiclassical periodic orbit theory and the theory of the Riemann zeta function. a) The generating function: two different Gutzwiller expansions are applicable outside the critical strip. b) The Riemann zeta function $\zeta(s)$ case.} \end{figure} The semiclassical approximation of the generating function in the vicinity of the real energy axis smoothly matching both of these asymptotics is their sum since the \textquotedblleft wrong\textquotedblright\ component is exponentially subdominant for energies with sufficiently large absolute value of the imaginary part. The paper is built in the following way. In Section 2 we describe the semiclassical periodic orbit theory of the generating function justifying its composition of two parts responsible for the monotonic and oscillatory parts of the level-level correlation function. In Section 3 we remind how the diagonal approximation neglecting the orbit action correlation is introduced for the generating functions; unlike Berry's diagonal approximation it provides both the oscillatory and monotonic parts of the spectral correlator. In Section 4 we show that our approach applied to the spectral correlator of the Riemann zeta function $\zeta(s)$ and Dirichlet $L-$functions zeros reproduces in a more transparent way the results of Bogomolny and Keating \cite{BogK3} and Bogomolny \cite{Bog}. We apply here arguments similar to the ones used in the case of the periodic orbit theory, for an illustration see Fig.\ \ref{fig}b). In Section 5 we return to dynamical systems without time-reversal symmetry and show how the oscillatory part of the correlator including nonuniversal effects stemming from repetitions can be obtained by our approach. In section 6 we explain why in the presence of time-reversal symmetry the oscillatory part of the correlator cannot be obtained without taking into account periodic orbit action correlations explicitly. Finally we conclude in section 7. \section{Correlation function} The level-level correlation function can be defined as% \[ R\left( e\right) =\frac{\left\langle \rho\left( E-\frac{e}{2}\right) \rho\left( E+\frac{e}{2}\right) \right\rangle _{E}}{\left\langle \rho\left( E-\frac{e}{2}\right) \right\rangle _{E}\left\langle \rho\left( E+\frac{e}% {2}\right) \right\rangle _{E}}-1 \] where $\rho\left( E\right) $ is the level density; averaging is done over an interval of the reference energy $E$ large compared with the mean level spacing $\delta=1/\bar{\rho},\quad\bar{\rho}=\left\langle \rho\right\rangle _{E},$ but small in classical terms; $e$ is the energy offset. The non-trivial part of $R(e)$ \begin{equation} r\left( e\right) \equiv\left\langle \rho\left( E-\frac{e}{2}\right) \rho\left( E+\frac{e}{2}\right) \right\rangle _{E}\label{paircorrelation}% \end{equation} to which we shall refer below as the correlation function for brevity, can be expressed in terms of the averaged ratio of four spectral determinants, or the generating function,% \[ Z\left( e_{A},e_{B},e_{C},e_{D}\right) =\left\langle \frac{\det\left( E+e_{C}-H\right) \det\left( E-e_{D}-H\right) }{\det\left( E+e_{A}% -H\right) \det\left( E-e_{B}-H\right) }\right\rangle _{E}. \] The connection between these quantities is given by the formula,% \begin{equation} r(e)=-\frac{1}{2\pi^{2}}\mathrm{Re}\left. \frac{\partial^{2}Z}{\partial e_{A}\partial e_{B}}\right\vert _{\parallel},\label{ZtoR}% \end{equation} assuming that the arguments of $Z$ have imaginary parts of the same sign for $e_{A},e_{B}$ as well as for $e_{C},e_{D}$: otherwise the result of averaging with respect to $E$ would be zero. After the averaging, all $e_{X},\quad X=A,B,C,D,$ must be set to $e/2$ which is indicated by $\left( \parallel\right) $ in (\ref{ZtoR}); this is the so called \textquotedblleft columnwise\textquotedblright\ limit of \cite{Heusler}. Without loss of generality we can fix for example ${\rm{Im}}\,e_{A,B}>0$ with the limit value $0_{+}$. As regards the energy offsets in the numerator we can then choose either ${\rm{Im}}\,e_{C,D}>0$ or ${\rm{Im}}\,e_{C,D}<0$ with the limit values $0_{+}$ and $0_{-}$ respectively. The two corresponding limits of $Z$ must coincide as the latter quantity is well defined for real $e_C$ and $e_D$. On the other hand, changing the sign of ${\rm{Im}}\,% e_{C,D}$ is equivalent to the replacement $e_{C}\rightleftarrows-e_{D}$ in $Z$. We come thus to the so called Weyl symmetry relation valid for real $e_{C,D}$,% \begin{equation*} Z\left( e_{A},e_{B},e_{C},e_{D}\right) =Z\left( e_{A},e_{B}% ,-e_{D},-e_{C}\right)\hspace*{1cm} {\rm{with}}\hspace*{1cm} {\rm{Im}}\,e_{C,D} =0\nonumber . \end{equation*} The semiclassical asymptotics of the generating function follows from the Gutz\-willer periodic orbit expansion \cite{Gutzwiller} of the trace of the resolvent. It converges when ${\rm{Im}}\,E$ is positive and large enough (the value guaranteeing absolute convergence is half the Kolmogorov-Sinai entropy, in view of the exponential orbit proliferation) and is given by \begin{equation} {\rm{Tr}}\left( E-H\right) ^{-1} \sim-i\pi\bar{\rho}\left( E\right) -\frac{i}{\hbar}\sum_{p}T_{p}F_{p}\mathrm{e}^{iS_{p}\left( E\right) /\hbar}\label{TrResolvIMgt0}\hspace*{6mm} {\rm{for}}\hspace*{6mm}{\rm{Im}}\,E >0, \end{equation} where we neglected the contributions from the orbit repetitions. The classical action of the orbit $p$ is denoted by $S_p$, the duration of the periodic orbits $T_p$ and the stability prefactor by $F_{p}$, for their precise form see Ref.\ \cite{Gutzwiller}. The expansion of the spectral determinant then follows from the identity $\det\left( E-H\right) =\exp{\rm{Tr}}\ln(E-H)$ which gives,% \begin{eqnarray} &\det\left( E-H\right)\propto\exp{\rm{Tr}}\int^{E}dE^{\prime }\left( E^{\prime}-H\right) ^{-1}\label{Positive}\propto\exp\left[ -i\pi\bar{N}\left( E\right) \right]\prod_pd_p(E) ,\nonumber\\ &{\rm{with}}\hspace*{6mm} {\rm{Im}}E >0\hspace*{6mm} {\rm{and}}\hspace*{6mm} d_p(E) =\exp% \left[-F_{p}\mathrm{e}^{iS_{p}\left( E\right) /\hbar}\right], \end{eqnarray} where% \begin{equation}\label{count} \bar{N}\left( E\right) =\int^{E}dE^{\prime}\bar{\rho}\left( E^{\prime }\right) \end{equation} denotes the averaged level counting function. The spectral determinant with the argument whose imaginary part is negative, is obtained by complex conjugation, \begin{equation} \det\left( E-H\right) \propto\exp\left[ i\pi\bar{N}\left( E\right) \right] \prod_pd_p^\ast(E) \label{neggamma}\hspace*{6mm} {\rm{with}}\hspace*{6mm}{\rm{Im}}E <0.\nonumber \end{equation} The two expansions can be used to get the semiclassical approximation of the generating function. Considering that ${\rm{Im}}\,e_{A,B}>0$ we substitute (\ref{Positive}) for $\det(E+e_{A}-H)$ and (\ref{neggamma}) for $\det\left( E-e_{B}-H\right) $. As regards the numerator there are two possibilities. If we choose ${\rm{Im}}\,e_{C,D}>0$ we must use (\ref{Positive}) for $\det(E+e_{C}-H)$ and (\ref{neggamma}) for $\det (E-e_{D}-H)$. Expanding $\bar{N}\left( E+e_{A}\right) =\bar{N}\left( E\right) +e_{A}\bar{\rho}$ etc. we obtain the semiclassical asymptotics of the determinants ratio as \begin{eqnarray} Z & \sim Z^{I}\left( e_{A,}e_{B},e\,_{C},e_{D}\right) =\mathrm{e}% ^{i\pi\left( e_{A}+e_{B}-e_{C}-e_{D}\right) \bar{\rho}}\left\langle\prod_pz_p(e_{A,}e_{B},e_{C},e_{D})\right\rangle ,\nonumber\\ &\hspace*{6mm}{\rm{with}}\hspace*{6mm} {\rm{Im}}\,e_{C,D} >0 \hspace*{6mm}{\rm{and}}\hspace*{6mm} {\rm{Im}}\,e_{A,B}>0\nonumber \end{eqnarray} with the definition \begin{equation} z_p(e_{A,}e_{B},e\,_{C},e_{D})=d_p(E+e_{C})d_p^\ast(E-e_{D})d_p(E+e_{A})d_p^\ast(E-e_{B}). \end{equation} If we choose ${\rm{Im}}\,e_{C,D}<0$ we get an alternative expression, \begin{eqnarray} Z & \sim Z^{II}\left( e_{A,}e_{B},e\,_{C},e_{D}\right) =Z^{I}\left( e_{A,}e_{B},-e\,_{D},-e_{C}\right) =\mathrm{e}^{i\pi\left( e_{A}+e_{B}% +e_{C}+e_{D}\right) \bar{\rho}}\nonumber\\ & \times\left\langle\prod_pz_p(e_{A,}e_{B},-e_{D},-e_{C}) \right\rangle \hspace*{4mm}{\rm{with}}\hspace*{4mm} {\rm{Im}}\,e_{C,D} <0\hspace*{4mm}{\rm{and}}\hspace*{4mm}{\rm{Im}}\,e_{A,B}>0\nonumber. \end{eqnarray} The functions $Z^{I},Z^{II}$ resulting after the energy averaging are well-behaved analytic functions of their arguments in which ${\rm{Im}}\,% e_{A,B}$ can be brought to the limit $0_{+}$ whereas $e_{C,D}$ can take any complex value. This is motivated by the fact that the presence of imaginary parts of the energy arguments in the denominator is obligatory in order to avoid the poles; on the other hand $e_C$ and $e_D$ can take any real values. It is important that due to the prefactors $\exp[\mp i\pi\left( e_{C}+e_{D}\right) \bar{\rho}]$ the function $Z^{II}\left( e_{A,}e_{B},e\,_{C},e_{D}\right) $ analytically continued to the \textquotedblleft alien\textquotedblright\ domain with ${\rm{Im}}\,% e_{C,D}$ large and positive, is exponentially smaller than $Z^{I}\left( e_{A,}e_{B},e\,_{C},e_{D}\right) $. Similarly the analytic continuation of $Z^{I}$ to ${\rm{Im}}\,e_{C},e_{D}<0$ is exponentially small compared to $Z^{II} $. With $e_{C,D}$ in the vicinity of the real axis $Z^{I},Z^{II}$ are of the same order of magnitude. What we are interested in is the semiclassical asymptotics of the generating function for $e_{C,D}$ real. We know that its analytical continuation to $e_{C,D}$ complex tends to $Z^{I}$ for large ${\rm{Im}}\,e_{C,D}>0$, and to $Z^{II}$ for ${\rm{Im}}\,e_{C,D}<0$. Their sum has this property and yields thus the semiclassical asymptotic for $Z$, \[ Z\sim Z_{\rm{sc}}=Z^{I}+Z^{II} \] which uniformly approximates $Z$ for both signs of the imaginary parts and reduces either to $Z^{I}$ or to $Z^{II}$ when $\left\vert {\rm{Im}}\,% e_{C,D}\right\vert $ is large. Like the exact generating function it obeys the Weyl symmetry relation when $e_{C,D}$ are real, \[ Z_{\rm{sc}}\left( e_{A},e_{B},e_{C},e_{D}\right) =Z_{\rm{sc}}\left( e_{A}% ,e_{B},-e_{D},-e_{C}\right) . \] \section{Diagonal approximation}\label{diagappro} Consider $Z^{I}$ with all its arguments containing large positive imaginary parts safeguarding convergence of the Gutzwiller expansions. Then $Z^{I}$ can be written as the energy average of a converging product over periodic orbits. Expanding the classical actions like $S_{p}\left( E+e_{A}\right) =S_{p}\left( E\right) +T_{p}e_{A}$ etc. we can write it as,% \[ Z^{I}\left( e_{A,}e_{B},e\,_{C},e_{D}\right) =\mathrm{e}^{i\pi\left( e_{A}+e_{B}-e_{C}-e_{D}\right) \bar{\rho}}\left\langle \prod_{p}% z_{p}\right\rangle _{E}, \] where% \[ \fl z_{p}=\exp\left[ F_{p}\mathrm{e}^{iS_{p}\left( E\right) /\hbar}\left( \mathrm{e}^{iT_{p}e_{A}/\hbar}-\mathrm{e}^{iT_{p}e_{C}/\hbar}\right) +F_{p}^{\ast}\mathrm{e}^{-iS_{p}\left( E\right) /\hbar}\left( e^{iT_{p}e_{B}/\hbar}-e^{iT_{p}e_{D}/\hbar}\right) \right]. \] Neglecting correlations between $z_{p}$ corresponding to orbits which do not have exactly the same action we can replace the average of the product by the product of averages. In the absence of time reversal and spatial symmetry the orbit actions are non-degenerate. Disregarding the periodic orbit repetitions we get,% \begin{equation*} Z^{I} \approx\mathrm{e}^{i\pi\left( e_{A}+e_{B}-e_{C}-e_{D}\right) \bar{\rho}}\prod_{p}\left\langle z_{p}\right\rangle. \end{equation*} By the analogous steps as in \cite{Heusler} we obtain,% \begin{equation*} Z^{I} \approx\mathrm{e}^{i\pi\left( e_{A}+e_{B}-e_{C}-e_{D}\right) \bar{\rho}}\,\tilde{Z}_{GUE}^{I}\hspace*{6mm}{\rm{with}}\hspace*{6mm} \tilde{Z}_{GUE}^{I} =\frac{\left( e_{A}+e_{D}\right) \left( e_{B}% +e_{C}\right) }{\left( e_{A}+e_{B}\right) \left( e_{C}+e_{D}\right) }. \end{equation*} The second component of $Z$ is obtained by the interchange of $e_{C,D}$ and change of their signs (the Weyl symmetry operation),% \begin{equation*} Z^{II} \approx\mathrm{e}^{i\pi\left( e_{A}+e_{B}+e_{C}+e_{D}\right) \bar{\rho}}\tilde{Z}_{GUE}^{II}\hspace*{6mm}{\rm{with}}\hspace*{6mm} \tilde{Z}_{GUE}^{II} =-\frac{\left( e_{A}-e_{C}\right) \left( e_{B}-e_{D}\right) }{\left( e_{A}+e_{B}\right) \left( e_{C}+e_{D}\right) }. \end{equation*} The correlation function obtained as the real part of $\left. \partial _{e_{A}e_{B}}^{2}\left( Z^{I}+Z^{II}\right) \right\vert _{\parallel}$ coincides with the correlation function of the GUE ensemble, its monotonic and oscillatory parts produced by $Z^{I}$ and $Z^{II},$ respectively. We observe that the product $\left( e_{A}-e_{C}\right) \left( e_{B}-e_{D}\right) $ in $\tilde{Z}_{GUE}^{II}$ turns into zero in the columnwise limit $e_{A,B,C,D}\to e/2$ such that both differentiations must be spent on it,\thinspace\ hence $\left( \partial _{e_{A}e_{B}}^{2}Z^{II}\right) _{\parallel}=-\mathrm{e}^{i2\pi e\bar{\rho}% }/e^{2}$. The usual diagonal approximation as introduced by Berry \cite{Berry} neglects terms corresponding to different orbits in the expansion of the correlation function itself. Only the non-oscillatory part of the correlator is then recovered; the oscillatory part is regarded as the result of the periodic orbits action correlations. It was demonstrated by Keating \cite{Keating} and Bogomolny \cite{Bog} that the oscillatory contribution to the correlation function of the Riemann zeta zeros results from the tendency of the prime numbers (``periodic orbits" of the problem) to avoid one another described by the Hardy-Littlewood conjecture. The respective orbit correlation mechanism \cite{Argaman} in physical systems without time reversal may thus be called ``the Hardy-Littlewood lookalike"; its nature is still not quite clear. It is remarkable that we can completely neglect that correlation using instead the diagonal approximation for the Weyl-symmetric generating function, and get the same result. That equivalence seems to hold also for corrections to the universal RMT result; compare the Riemann zeta treatment by Bogomolny \cite{Bog}, Conrey and Snaith \cite{Conrey} and the one below in this paper. \section{Zeros of Riemann zeta function} Following the idea of Hilbert and P\'{o}lya, the non-trivial zeros of the Riemann zeta function may be interpreted as the eigenvalues of a self-adjoint Hamilton operator of a virtual dynamical system. The analogy between the density of zeros of the Riemann zeta and the Gutzwiller trace formula for chaotic quantum systems provides further support for this spectral interpretation. Another evidence for this interpretation provides the behavior of the pair correlation of the zeros of the Riemann zeta function first studied by Montgomery \cite{Montgomery}; they are the same as for the eigenvalues of the Gaussian unitary ensemble of random matrices (GUE) as confirmed by Odlyzko's numerical calculations \cite{Odlyzko}. The spectral properties of the Riemann zeta function and closely related $L$-functions were a subject of an innumerable number of investigations; in connection with the present paper we would like to mention the works of Keating \cite{Keating}, Bogomolny et al. \cite{BogK3,Bog,BogK1,BogK2,BogK4,BogK5} and Berry et al. \cite{BerK}. See also Ref.\ \cite{Conrey} for a review and the references therein. Here we show that the pair correlation function including the system-specific corrections to the GUE prediction can easily be obtained by the method described above. \subsection{The generating function} The Riemann zeta function can be defined by the Euler product over primes,% \[ \zeta\left( s\right) =\prod_{p}\frac{1}{1-p^{-s}} \] which converges provided ${\rm{Re}}\,s>1$ but can be analytically continued to the whole complex plane. According to the Riemann hypothesis, all non-trivial zeros of zeta have the form $s_{k}=\frac{1}{2}+iE_{k}$ with $E_{k}$ real; the spectrum is symmetric with respect to $E=0$. The corresponding spectral determinant $\Delta(E)=\prod_{k}A\left( E,E_{k}% \right) (E-E_{k})$ where $A\left( E,E_{k}\right) $ stands for a suitable regularizing function, is related to the Riemann zeta function by \begin{equation} \Delta(E)=B(E)e^{-i\pi\bar{N}(E)}\zeta\left( \frac{1}{2}-iE\right) \label{SpecDetDef}% \end{equation} with $\bar{N}(E)$ being the mean counting function, i.e.\ the integral of the mean zeros density $\bar{\rho}\left( E\right) $ with respect to $E$, and $B(E)$ a smooth function positive for real $E$ whose explicit form depends on the regularization choice \cite{Keating}. We need only the asymptotic expressions for large $E$:% \[ \bar{N}\left( E\right) \sim\frac{E}{2\pi}\left(\ln\frac{E}{2\pi}-1\right),\quad\bar{\rho }\left( E\right) =\frac{d\bar{N}}{dE}\sim\frac{1}{2\pi}\ln\frac{E}{2\pi}. \] The Euler product for the spectral determinant converges for ${\rm{Im}}\,E>1/2, $ and then% \begin{equation} \Delta\left( E\right) =B\left( E\right) \mathrm{e}^{-i\pi\bar{N}\left( E\right) }\prod_{p}\frac{1}{1-p^{-1/2+iE}}.\label{impos}% \end{equation} When ${\rm{Im}}\,E<-\frac{1}{2}$ the complex conjugation gives \begin{equation} \Delta\left( E\right) =B\left( E\right) \mathrm{e}^{i\pi\bar {N}\left( E\right) }\prod_{p}\frac{1}{1-p^{-1/2-iE}}.\label{imneg}% \end{equation} Calculating $\frac{d}{dE}\log\Delta\left( E\right) $ we would get the Gutzwiller-like formula for the trace of the virtual resolvent with primes as periodic orbits, analogous to (\ref{TrResolvIMgt0}) and its complex conjugate but taking into account the orbit repetitions. In fact we don't need that to construct the generating function since the expansion for \begin{equation*} Z =\left\langle \frac{\Delta\left( E+e_{C}\right) \Delta\left( E-e_{D}\right) }{\Delta\left( E+e_{A}\right) \Delta\left( E-e_{B}\right) }\right\rangle _{E}\hspace*{6mm}{\rm{with}}\, E\,{\rm{ real}},\,{\rm{Im}}\,e_{A,B} >0, \end{equation*} follows directly from the Euler product of $\zeta$. The generating function can be written in two versions. If ${\rm{Im}}\,e_{C,D}>0$ we use (\ref{impos}) for $\Delta\left( E+e_{C}\right) ,\Delta\left( E+e_{A}\right) $ and (\ref{imneg}) for the two other determinants. Let us assume all deviations $e$ small compared with $E\gg1$ such that $\bar{N}\left( E+e_{A}\right) \approx\bar{N}\left( E\right) +\bar{\rho}\,e_{A}$ etc. and $\left|{\rm{Im}}\,e_{X}\right|>1/2,\quad X=A,B,C,D$. Denoting% \[ a_{p}\equiv\frac{\mathrm{e}^{ie_{A}\ln p}}{\sqrt{p}} \] and similar for $b_{p},c_{p},d_{p}$ we have for ${\rm{Im}}\,e_{A,B,C,D}>1/2$,% \begin{eqnarray} Z^{I}\left( e_{A},e_{B},e_{C},e_{D}\right) & =\exp\left[ -i\pi\left( e_{C}+e_{D}-e_{A}-e_{B}\right) \bar{\rho}\right] \left\langle \prod_{p}% z_{p}\right\rangle _{E},\nonumber\\ \nonumber \hspace*{6mm}{\rm{with}}\hspace*{16mm} z_{p} & =\frac{\left( 1-a_{p}\mathrm{e}^{iE\ln p}\right) \left( 1-b_{p}\mathrm{e}^{-iE\ln p}\right) }{\left( 1-c_{p}\mathrm{e}^{iE\ln p}\right) \left( 1-d_{p}\mathrm{e}^{-iE\ln p}\right) },\nonumber \end{eqnarray} we neglected the $e$-dependence of the smooth function $B$. The alternative representation $Z^{II}$ with ${\rm{Im}}\,e_{C,D}$ negative is obtained from $Z^{I}$ by the Weyl substitution,% \begin{equation*} \fl Z^{II}\left( e_{A},e_{B},e_{C},e_{D}\right) =Z^{I}\left( e_{A}% ,e_{B},-e_{D},-e_{C}\right) \hspace*{4mm}{\rm{with}}\hspace*{4mm} {\rm{Im}}\,e_{A,B} >1/2,\quad{\rm{Im}}\,e_{C,D}<-1/2; \end{equation*} with the substitution rule $\left\{ c_{p},d_{p}\right\} \rightarrow \left\{ \left( d_{p}p\right) ^{-1},\left( c_{p}p\right) ^{-1}\right\} $. Now we introduce the diagonal approximation by the assumption that $z_{p}$ associated with different primes are uncorrelated such that the energy average of the product of $z_{p}$ can be replaced by the product of their averages writing \[ Z^{I}\approx\exp\left[ -i\pi\left( e_{C}+e_{D}-e_{A}-e_{B}\right) \bar {\rho}\right] \prod_{p}\left\langle z_{p}\right\rangle _{E}. \] The factors in the product are elementarily calculated,% \begin{eqnarray} \left\langle z_{p}\right\rangle _{E} & =\frac{\ln p}{2\pi}\int_{0}% ^{\frac{2\pi}{\ln p}}dE\,z_{p}=\frac{1}{2\pi i}% {\displaystyle\oint_{\left\vert u\right\vert =1}} \frac{du}{u}\frac{\left( 1-a_{p}u\right) \left( 1-b_{p}/u\right) }{\left( 1-c_{p}u\right) \left( 1-d_{p}/u\right) }\label{zpav}\nonumber\\ & =1+\frac{\left( b_{p}-d_{p}\right) \left( a_{p}-c_{p}\right) }% {1-c_{p}d_{p}}. \end{eqnarray} Considering that $\zeta\left( 1-i\left( e_{A}+e_{B}\right) \right) =\prod_{p}\left( 1-a_{p}b_{p}\right) ^{-1}$ etc, we write $Z^{I}=Z_{\zeta }^{I}\,\Phi^{I}$, where \begin{equation} \fl Z_{\zeta}^{I}=\exp\left[ -i\pi\left( e_{C}+e_{D}-e_{A}-e_{B}\right) \bar{\rho}\right] \frac{\zeta\left( 1-i\left( e_{A}+e_{B}\right) \right) \zeta\left( 1-i\left( e_{C}+e_{D}\right) \right) }{\zeta\left( 1-i\left( e_{C}+e_{B}\right) \right) \zeta\left( 1-i\left( e_{A}+e_{D}\right) \right) }.\label{zetz}% \end{equation} The function $\Phi^{I}$ is a product over primes,% \begin{eqnarray} \fl\Phi^{I}\left( e_{A},e_{B},e_{C},e_{D}\right) & =\prod_{p}\phi_{p}% ^{I},\quad\label{PhiI}\\ \fl \hspace*{16mm}{\rm{with}}\hspace*{6mm} \phi_{p}^{I} & =\frac{(1-a_{p}b_{p})(1-c_{p}d_{p})}{(1-b_{p}c_{p}% )(1-a_{p}d_{p})}\left\langle z_{p}\right\rangle =1-\frac{a_{p}b_{p}% (a_{p}-c_{p})(b_{p}-d_{p})}{(1-b_{p}c_{p})(1-a_{p}d_{p})}.\label{phip}% \end{eqnarray} The second component $Z^{II}$ is obtained by the Weyl substitution $e_{C}\rightleftarrows-e_{D}$ and can be written as $Z_{\zeta}^{II}\Phi^{II}$ with \[ \fl Z_{\zeta}^{II}=\exp\left[ -i\pi\left( -e_{C}-e_{D}-e_{A}-e_{B}\right) \bar{\rho}\right] \frac{\zeta\left( 1-i\left( e_{A}+e_{B}\right) \right) \zeta\left( 1+i\left( e_{C}+e_{D}\right) \right) }{\zeta\left( 1-i\left( e_{B}-e_{D}\right) \right) \zeta\left( 1-i\left( e_{A}-e_{C}\right) \right) } \] and $\Phi^{II}\left( e_{A},e_{B},e_{C},e_{D}\right) =\Phi^{I}\left( e_{A},e_{B},-e_{D},-e_{C}\right) .$ Considering the exponential factors in $Z_{\zeta}$ the sum% \begin{equation} Z\sim Z^{I}+Z^{II}\label{z1plusz2}% \end{equation} will smoothly interpolate between $Z^{I}$ and $Z^{II}$ as we change ${\rm{Im}}\,e_{C,D}$ from positive to negative values, providing thus the high-energy asymptotics of the generating function for real $e_{C,D}$. It coincides with the result obtained differently in \cite{Conrey}. \subsection{The correlation function} The correlation function based on Eq.\ (\ref{z1plusz2}) consists of two summands. Differentiation of $Z^{I}$ with the subsequent columnwise identification $(||) $ $e_{A}=e_{B}=e_{C}=e_{D}=e/2$ gives the contribution which Bogomolny \cite{Bog} refers as the diagonal one. It is easily checked that \begin{equation} \left( Z_{\zeta}^{I}\right) _{||}=1,\hspace*{1cm}\left( \frac{\partial Z^{I}}{\partial e_{A}}\right) _{||}=\left( \frac{\partial Z^{I}}{\partial e_{B}}\right) _{||}=0\label{pa1}% \end{equation} and for the second derivative \begin{equation} \left( \frac{\partial^{2}Z_{\zeta}^{I}}{\partial e_{A}\partial e_{B}}\right) _{||}=\frac{\partial^{2}\ln\zeta(1-ie)}{\partial e^{2}}=\frac{1}{e^{2}% }+(\gamma_{0}^{2}+2\gamma_{1})+O(e).\label{pa2}% \end{equation} For $\Phi^{I}$ we obtain \begin{equation} \Phi_{||}^{I}=1,\hspace*{1cm}\left( \frac{\partial\Phi^{I}}{\partial e_{A}% }\right) _{||}=\Phi_{||}^{I}\sum_{p}\left( \frac{1}{\phi_{p}^{I}}% \frac{\partial\phi_{p}^{I}}{\partial e_{A}}\right) _{||}=0\label{pa3}% \end{equation} and the same result for $\left( {\partial\Phi}^{I}{/\partial e_{B}}\right) _{||}$. For the second derivatives we get \begin{equation} \left( \frac{\partial^{2}\Phi^{I}}{\partial e_{A}\partial e_{B}}\right) _{||}=\left( \sum_{p}\frac{\partial}{\partial e_{B}}\left( \frac{1}{\phi _{p}^{I}}\frac{\partial\phi_{p}^{I}}{\partial e_{A}}\right) \right) _{||}=\sum_{p}\frac{\ln^{2}p}{\left( 1-p^{1-ie}\right) ^{2}}.\label{pa4}% \end{equation} In total we get for the contribution to the spectral correlation function (\ref{paircorrelation}) \begin{equation} r^{I}(\varepsilon)=-\frac{1}{2\pi^{2}}\mathrm{Re}\left( \frac{\partial ^{2}Z_{\zeta}^{I}}{\partial e_{A}\partial e_{B}}\right) _{||}-\frac{1}% {2\pi^{2}}\mathrm{Re}\left( \frac{\partial^{2}\Phi^{I}}{\partial e_{A}\partial e_{B}}\right) _{||}.% \end{equation} Taking into account the relations (\ref{pa1},\ref{pa2},\ref{pa3},\ref{pa4}), we confirm that this result is identical to the \textquotedblleft diagonal\textquotedblright\ one obtained by Bogomolny from the Hardy-Littlewood conjecture in \cite{Bog}. The second contribution to the spectral correlation function, termed conventionally off-diagonal but which we obtain remaining in the framework of uncorrelated primes, is provided by the second derivative of $Z^{II}$ with respect to $e_{A}$ and $e_{B}$ and columnwise identification. This identification implies that $e_{A}-e_{C}$ and $e_{B}-e_{D}$ go to zero, hence the following expansion for the inverse zeta functions entering $Z_{\zeta }^{II}$ can be approximated by \begin{equation} \zeta^{-1}(1-i(e_{A}-e_{C}))\zeta^{-1}(1-i(e_{B}-e_{D}))\approx-(e_{A}% -e_{C})(e_{B}-e_{D}) \end{equation} such that the only term surviving within this identification results from $\partial^{2}Z_{\zeta}^{II}/\partial e_{A}\partial e_{B}$ yielding $-|\zeta(1-ie)|^{2}$. We get in total for the oscillatory (\textquotedblleft off-diagonal\textquotedblright) contribution to the correlator, \begin{equation} r^{II}(e)=\frac{1}{2\pi^{2}}|\zeta(1-ie)|^{2}\mathrm{Re}\left[ \mathrm{e}% ^{2\pi i\bar{\varrho}e}\prod_{p}\left( 1-\frac{(1-p^{ie})^{2}}{(p-1)^{2}% }\right) \right] , \end{equation} where the product in the last equation results from $\left( \Phi^{II}\right) _{\parallel}=\Phi^{I}\left( e/2,e/2,-e/2,-e/2\right) $ with $\Phi^{I}$ given by Eq.\ (\ref{PhiI}). Again we reproduced the result of Bogomolny and Keating obtained in Ref.\ \cite{BogK3} in a heuristic way. \subsection{$L$-functions} $L$-functions are generalizations of the Riemann zeta function \begin{equation} L(s)=\sum_{n=1}^{\infty}\frac{\chi(n)}{n^{s}}=\prod_{p}\frac{1}{1-\frac {\chi(p)}{p^{s}}}\label{Ls}% \end{equation} including the Dirichlet character $\chi(n)$ defined as a function on integers periodic with the integer period $k$ such that $\chi(n+k)=\chi(n)$, multiplicative $\chi(n)\chi(m)=\chi(nm)$, equal to zero iff the greatest common divisor of $n$ and $k$ is greater than one. We restrict our considerations to real characters $\chi(n)=\pm1,0$. The generalized Riemann hypothesis states that all non-trivial zeros of $L(s)$ lie on the critical line $s=1/2+iE$. The correlation function of these zeros can be determined using the generating function in a similar way as for the Riemann zeta function starting from the product representation given in Eq.\ (\ref{Ls}). The generating function obtained under the assumption ${\rm{Im}}\,% e_{A,B,C,D}>0$ differs from the Riemann zeta case by the presence of characters in the factors $z_{p}$; these are now% \[ z_{p}=\frac{\left( 1-\chi(p)a_{p}\mathrm{e}^{iE\ln p}\right) \left( 1-\chi(p)b_{p}\mathrm{e}^{-iE\ln p}\right) }{\left( 1-\chi(p)c_{p}% \mathrm{e}^{iE\ln p}\right) \left( 1-\chi(p)d_{p}\mathrm{e}^{-iE\ln p}\right) }. \] As before, we introduce the diagonal approximation; taking into account that $\chi^{2}(p)=1,0$, we obtain that $\left\langle z_{p}\right\rangle $ is given by Eq.\ (\ref{zpav}) if $p$ is not a divisor of $k$, otherwise it drops out from the product $Z^{I}=\prod_{p}\left\langle z_{p}\right\rangle $. This implies that the part of the generating function obtained under the assumption ${\rm{Im}}\,e_{C,D}>0$ can be expressed as \[ Z^{I}=Z_{\zeta}^{I}(e_{A},e_{B},e_{C},e_{D})\Phi^{I}(e_{A},e_{B},e_{C}% ,e_{D})\Xi^{I}(e_{A},e_{B},e_{C},e_{D}) \] with $Z_{\zeta}^{I}$ and $\Phi^{I}$ are the same as in Eqs.\ (\ref{zetz}% ,\ref{phip}). The function $\Xi^{I}$ cancels the factors now missing in $Z^{I}$, it can be expressed as a finite product over the prime divisors of $k $ \begin{equation} \Xi^{I}(e_{A},e_{B},e_{C},e_{D})=\prod_{p/k}\frac{1}{\left\langle z_{p}% \right\rangle }=\prod_{p/k}\left( 1+\frac{(b_{p}-d_{p})(a_{p}-c_{p}% )}{1-c_{p}d_{p}}\right) ^{-1}.% \end{equation} Differentiation of $Z^{I}$ with respect to $e_{A}$ and $e_{B}$ creates within the $(\parallel)-$ identification the non-oscillatory contribution to the correlation function as given in Ref.\ \cite{BogK4}. The oscillatory part is obtained as $\left( \partial_{e_{A}e_{B}}^{2}Z^{II}\right) _{\parallel}$ with $Z^{II}(e_{A},e_{B},e_{C},e_{D})=Z^{I}(e_{A},e_{B},-e_{D},-e_{C});$ it contains the factor additional to the Riemann zeta case \begin{equation} \Psi^{\mathrm{off}}(e)\equiv\left. \Xi^{I}(e_{A},e_{B},-e_{D},-e_{C})\right\vert _{\parallel}=\prod_{p/k}\left( 1+\frac{(p^{ie/2}-p^{-ie/2})^{2}}{p-p^{-ie}% }\right) ^{-1}% \end{equation} which was earlier found in Ref.\ \cite{BogK4} based on the Hardy-Littlewood conjecture. We observe that the correlation function depends only on the period $k$ but not on the precise form of the characters $\chi(n)$. \section{Dynamical systems without time reversal; non-universal corrections} In this section we return to dynamical systems with classically chaotic counterpart. We show how the results from Ref.\ \cite{BogK3} can be obtained by the method proposed here including the nonuniversal effect of repetitions. The expression for the spectral determinant $\det(E-H)$ taking into account repetitions of periodic orbits is given by \cite{Haake,Keating} \begin{equation} \det(E-H)=B(E){\rm e}^{-i\bar{N}(E)}\prod_p\prod_{k=0}^\infty\left(1-\frac{{\rm e}^{iS_p(E)/\hbar}}{\Lambda_p^{k+1/2}}\right) \end{equation} expressed as product over the primitive periodic orbits $p$ with the actions $S_p(E)$ and the stability coefficients $\Lambda_p={\rm e}^{\lambda_p T_p}$ depending on the Lyapunov exponent $\lambda_p$ and the duration of $T_p$ of the orbit $p$. The function $\bar{N}(E)$ is given in Eq.\ (\ref{count}) and $B(E)$ again is a real function resulting from the regularization of the infinite product. Expanding again $S_p(E+e)\approx S_p(E)+e T_p$, $z_p$ is obtained in diagonal approximation as \begin{equation}\label{zpavn} \left\langle z_p\right\rangle=\frac{1}{2\pi}\int_0^{2\pi}d\phi \frac{\left({\rm e}^{ie_C T_p+i\phi}/\sqrt{\Lambda_p};\Lambda_p^{-1}\right)_\infty}{\left({\rm e}^{ie_A T_p+i\phi}/\sqrt{\Lambda_p};\Lambda_p^{-1}\right)_\infty} \frac{\left({\rm e}^{ie_D T_p-i\phi}/\sqrt{\Lambda_p};\Lambda_p^{-1}\right)_\infty}{\left({\rm e}^{ie_B T_p-i\phi}/\sqrt{\Lambda_p};\Lambda_p^{-1}\right)_\infty} \end{equation} with the $q$-Pochhammer symbols defined as \begin{equation} (x;q)_n=\prod_{k=0}^{n-1}\left(1-xq^k\right). \end{equation} The averaging of (\ref{zpavn}) with respect to ${\rm e}^{i\phi}$ is performed by the $q$-binomial formula of Gauss \cite{qbin} \begin{equation} \frac{\left(tx;q\right)_\infty}{\left(x;q\right)_\infty}=\sum_{n=0}^\infty\frac{\left(t;q\right)_n}{\left(q;q\right)_n}x^n \end{equation} yielding \begin{eqnarray} \fl \left\langle z_p\right\rangle&=&\frac{1}{2\pi}\int_0^{2\pi}\!\!\!d\phi\!\sum_{n,m=0}^\infty\!\!\!\frac{\left({\rm e}^{i(e_C-e_A)T_p};\Lambda_p^{-1}\right)_n}{\left(\Lambda_p^{-1};\Lambda_p^{-1}\right)_n}\! \left(\frac{{\rm e}^{i\phi+ie_AT_p}}{\sqrt{\Lambda_p}}\right)^n\! \frac{\left({\rm e}^{i(e_D-e_B)T_p};\Lambda_p^{-1}\right)_m}{\left(\Lambda_p^{-1};\Lambda_p^{-1}\right)_m}\!\left(\frac{{\rm e}^{-i\phi+ie_BT_p}}{\sqrt{\Lambda_p}}\right)^m\nonumber\\ \fl &=&\sum_{n=0}^\infty\frac{\left({\rm e}^{i(e_C-e_A)T_p};\Lambda_p^{-1}\right)_n}{\left(\Lambda_p^{-1};\Lambda_p^{-1}\right)_n} \frac{\left({\rm e}^{i(e_D-e_B)T_p};\Lambda_p^{-1}\right)_n}{\left(\Lambda_p^{-1};\Lambda_p^{-1}\right)_n}\left(\frac{{\rm e}^{i(e_A+e_B)T_p}}{\Lambda_p}\right)^n, \end{eqnarray} where the last sum is the definition of the $q$-hypergeometric function of Heine $_2\varphi_1({\rm e}^{i(e_C-e_A)T_p},{\rm e}^{i(e_D-e_B)T_p};1/\Lambda_p;1/\Lambda_p; {\rm e}^{i(e_A+e_B)T_p}/\Lambda_p)$ \cite{heinehyp,heinehyp1}. In the same way as before in Eq.\ (\ref{zetz}) we can factorize from $Z^I=\prod_p\left\langle z_p\right\rangle$ its poles and zeros such that the rest is a convergent product. Therefore we consider the inverse classical zeta functions \begin{equation} \frac{1}{Z_{\rm cl}(s)}=\prod_p\frac{1}{Z_{{\rm cl},p}(s)}=\prod_p\prod_{k=0}^\infty\left(1-\frac{{\rm e}^{sT_p}}{\Lambda_p^{k+1}}\right)^{k+1}. \end{equation} The component $Z^I$ of the generating function can then be written as \begin{eqnarray} \fl Z^I&=&\frac{Z_{\rm cl}(i(e_A+e_B))Z_{\rm cl}(i(e_C+e_D))}{Z_{\rm cl}(i(e_A+e_D))Z_{\rm cl}(i(e_B+e_C))} {\prod_p}\, \left(\frac{Z_{{\rm cl},p}(i(e_A+e_B))Z_{{\rm cl},p}(i(e_C+e_D))}{Z_{{\rm cl},p}(i(e_A+e_D))Z_{{\rm cl},p} (i(e_B+e_C))}\right)^{-1}\nonumber\\ \fl &&\times _2\hspace*{-1mm}\varphi_1({\rm e}^{i(e_C-e_A)T_p},{\rm e}^{i(e_D-e_B)T_p};1/\Lambda_p;1/\Lambda_p; {\rm e}^{i(e_A+e_B)T_p}/\Lambda_p){\rm e}^{-i\pi\overline{\rho}\left(e_C-e_B-e_A+e_D\right)}. \end{eqnarray} For $e_{A,B,C,D}\rightarrow 0$ \begin{equation} \frac{1}{Z_{\rm cl}(i(e_A+e_D))Z_{\rm cl}(i(e_B+e_C))} \rightarrow-(e_A+e_D)(e_B+e_C)T_0^2 \end{equation} with a certain reference time $T_0$ \cite{Haake}. The overall generating function is the sum $Z^I+Z^{II}$ where $Z^{II}$ is obtained by the Weyl substitution $e_{C}\rightleftarrows-e_{D}$. Then the spectral correlation function of \cite{BogK3} is obtained by Eq.\ (\ref{ZtoR}). \section{Systems with time reversal} If time reversal is allowed almost every periodic orbit exists in two versions with different sense of traversal and exactly the same action. Repeating the calculations of Section \ref{diagappro} we get the generating function in the diagonal approximation as $Z=Z_{diag}^{I}+Z_{diag}^{II}$, where% \begin{eqnarray} Z_{\mathrm{diag}}^{I}\left( e_{A,}e_{B},e_{C},e_{D}\right) & =e^{-i\pi\left( e_{C}+e_{D}-e_{A}-e_{B}\right) \bar{\rho}}\frac{\left( e_{A}+e_{D}\right) ^{2}\left( e_{B}+e_{C}\right) ^{2}}{\left( e_{A}+e_{B}\right) ^{2}\left( e_{C}+e_{D}\right) ^{2}}% ,\nonumber\\ Z_{\mathrm{diag}}^{II}\left( e_{A,}e_{B},e_{C},e_{D}\right) & =Z_{\mathrm{diag}}^{I}\left( e_{A,}e_{B},-e_{D},-e_{C}\right).\nonumber \end{eqnarray} The term $Z_{\mathrm{diag}}^{II}\propto\left( e_{A}-e_{C}\right) ^{2}\left( e_{B}-e_{D}\right) ^{2}$ does not contribute to the correlation function since its second mixed derivative by $e_{A,B}$ turns into zero in the columnwise limit. Therefore in the presence of the time reversal the oscillatory contribution to the correlation function is a truly off-diagonal effect different from the Hardy-Littlewood-like mechanism; attempts to obtain it in the framework of the diagonal approximation were unsuccessful \cite{BogK3}. The off-diagonal corrections to the generating function of the orthogonal case were obtained in \cite{Heusler,Muller}; the symplectic case was considered in \cite{Braun}. Their existence is due to the so called orbit partnership existing in the chaotic motion whose significance was realized after the discovery of the Sieber-Richter pairs \cite{Sieber}. The result has the form of an asymptotic expansion valid for large scaled energy deviations $\varepsilon_{X}=e_{X}2\pi\bar{\rho},\quad X=A,B,C,D$. Here the $e_{X}$ are assumed to be small in classical terms while the dimensionless variables $\varepsilon_{X}$ are considered to be large. In the orthogonal case the generating function can be written as $Z_{\mathrm{ortho}}=Z^{I}+Z^{II}$ with% \begin{eqnarray} Z^{I} & =Z_{\mathrm{diag}}^{I}\left( 1+Z_{off}^{I}\right) ,\nonumber\\ Z_{\mathrm{diag}}^{I} & =e^{i\left( \varepsilon_{A}+\varepsilon _{B}-\varepsilon_{C}-\varepsilon_{D}\right) /2}\frac{\left( \varepsilon _{A}+\varepsilon_{D}\right) ^{2}\left( \varepsilon_{B}+\varepsilon _{C}\right) ^{2}}{\left( \varepsilon_{A}+\varepsilon_{B}\right) ^{2}\left( \varepsilon_{C}+\varepsilon_{D}\right) ^{2}},\nonumber\\ Z_{\mathrm{off}}^{I} & \sim-\frac{\left( \varepsilon_{A}-\varepsilon _{C}\right) \left( \varepsilon_{B}-\varepsilon_{D}\right) }{\left( \varepsilon_{A}+\varepsilon_{D}\right) \left( \varepsilon_{B}+\varepsilon _{C}\right) }\sum_{n=1}^{\infty}\frac{\left( -2i\right) ^{n}\left( n-1\right) !}{\left( \varepsilon_{A}+\varepsilon_{B}\right) ^{n-1}}\left( \frac{1}{\varepsilon_{C}+\varepsilon_{D}}+\frac{n}{\varepsilon_{A}% +\varepsilon_{B}}\right)\nonumber \end{eqnarray} and $Z^{II}\left( \varepsilon_{A},\varepsilon_{B},\varepsilon_{C}% ,\varepsilon_{D}\right) =Z^{II}\left( \varepsilon_{A},\varepsilon _{B},-\varepsilon_{D},-\varepsilon_{C}\right) .$ The corresponding expansion in the symplectic case is obtained by the substitution,% \[ Z_{\mathrm{sympl}}\left( \varepsilon_{A},\varepsilon_{B},\varepsilon _{C},\varepsilon_{D}\right) =Z_{\mathrm{ortho}}\left( -\varepsilon _{C},-\varepsilon_{D},-\varepsilon_{A},-\varepsilon_{B}\right) . \] Differentiating the generating functions by $\varepsilon_{A,B}$ and then setting all arguments equal to $\varepsilon$ we get the correlation functions as a diverging series in $\varepsilon^{-1}$ which can be brought to a closed form by the Borel summation. This is essentially a term-by-term Fourier transform to the time domain followed by summation of the resulting converging series. Finally the inverse Fourier transform brings the result back to the energy domain. An outstanding feature of the symplectic case is that the asymptotic expansion of the resulting correlation function contains two oscillatory components proportional respectively to $e^{i2\varepsilon}$ and $e^{i\varepsilon}$, the latter not present in the input of the Borel summation; for explanation see \cite{Braun} . \section{Conclusion} We present here a new way to compute spectral correlation functions starting from semiclassical expressions. It is carried out via the auxiliary generating function which is an averaged ratio of four spectral determinants. The semiclassical asymptotics of the generating function consists of two terms responsible for the monotonic and oscillatory components of the correlator. Our new way is to explain this twofold structure by complexifying the arguments of the generating functions and bringing them to the boundaries of convergence of the Gutzwiller periodic orbit expansions of the spectral determinants. That could be done in two ways differing by the direction in which we move away from the real axis; in the example of the Riemann zeta zeros it would mean bringing the arguments of zetas in the numerator of the generating function either to the left or to the right border of the critical strip. In our approach after the energy averaging the two semiclassical approximations of the generating function become well-behaved analytical functions dominant in the respective complex half-planes; at the real axis the asymptotics of the generating function is given by their sum. We avoid the use of the Riemann-Siegel-lookalike formula for the spectral determinants, so far not proven beyond the Riemann zeta and quantum maps in finite Hilbert space. We avoid thus the appearance of additional terms caused by the sharp cut off at $T_H/2$ and disappearing in the final result. Our approach has the additional advantage of simplicity, e.g., our derivation of the correlation function of the Riemann zeta and $L$-functions zeros is probably the simplest existing. In the absence of the time reversal the oscillatory contribution to the correlation function can be found via the diagonal approximation to the generating function totally neglecting the periodic orbit correlations. On the other hand, if we choose to directly calculate the correlator via the periodic orbit sum the oscillatory components would exist only if the action correlation between the periodic orbits is taken into account \cite{Argaman}; in the case of the Riemann zeta zeros this is the correlation of primes following from the Hardy-Littlewood conjecture. It is still not quite clear what is the nature of that correlation in physical systems and why it does not reveal its presence if time reversal is allowed. It is also not clear why the existence of that correlation automatically follows from the Weyl symmetry of the generating function. \section*{Acknowledgements} We are grateful to F.\ Haake for remarks which stimulated the work on the subject. We also thank T.\ Guhr who read the manuscript and made useful comments and M.\ Akila, M.\ Bruckhoff for discussions. \section*{References:}

d8a35d3993eab98e3c97cfe63df47649c3330517

\section{Introduction} An interesting phenomenon in many resource allocation decisions in marketing is that, at the decision unit level, the data are still very sparse, despite the size of the overall data. To leverage the rest of the information in the big data, hierarchical Bayes (HB) \cite{Gelman, Rossi} provides a natural solution by statistically borrowing information with a shrinkage-based estimation at the individual unit level. There are two challenges when applying a HB model. First, the hierarchy structure needs to be determined in advance, which could be a challenge especially when the data do not possess a clear hierarchical affiliation relationship. Second, in practice, researchers tend to use only two or three levels for HB models because, for a fully Bayesian analysis, simulation-based approaches are necessary to obtain the joint posterior distribution. If there are too many levels, the model could be computationally expensive and very sensitive to the distribution assumptions and priors when applied to real-world data in order to converge. In this paper, we develop a new model that dynamically determines the hierarchy based on the input data. Meanwhile, by adopting empirical Bayes \cite{Casella}, we present an empirical approach to get inferences through the hierarchical structure. We show a two-phase system where flexible multi-level hierarchical models with deep hierarchy can be applied efficiently. Inspired by the loss concept in tree models (e.g., CART \cite{Breiman}), we propose a Dynamic Hierarchical Empirical Bayesian (DHEB) method that is capable of dynamically constructing the hierarchy. Specifically, each sub region in a layer of the hierarchy is treated as a node. The challenge is to find a natural way to merge the idea of loss function into the HB framework so that the estimates derived by the HB model are consistent with the optimal solutions for the loss function. To do so, we propose a loss function with a regularization term that incorporates the Bayesian concept of prior \cite{Rasmussen}. More details can be found in section 4.2. Given the loss function, instead of a fully Bayesian analysis, we present a stepwise method that practices empirical Bayes and builds a hierarchy dynamically from top to bottom. This proposed methodology combines the advantages of both (1) the hierarchical Bayesian model, which allows information borrowing from similar branches, and (2) a tree model, which helps define the structure using data. The performance of the proposed method is evaluated using a set of simulated data and real-world data from Adobe Advertising Cloud. We compare the proposed method with baseline models: weighted average, regularized linear regression, and fully HB models with different levels. All of the comparisons favor the proposed method against all its competitors in terms of prediction accuracy and efficiency. The rest of the paper is organized as follows. We will first describe the process of the sponsored search and the challenges faced when evaluating ads performance in section 2, followed by some related work in section 3. We then introduce our proposed method in section 4. Section 5 and section 6 provide the simulation and experimental results. Finally, we conclude the paper in section 7. \section{Background} \subsection{Sponsored Search} Sponsored search advertising is a kind of auction-based keyword advertising in search engines \cite{Lahaie}. Search engines decide the winners of the auctions based on their expected revenue. Meanwhile, advertisers need to understand what keywords are more valuable using performance measurements, such as number of impressions, click-through rate (CTR), conversion rate (CVR), revenue per click (RPC), cost per click (CPC), etc., so that they can manage their bids efficiently and allocate their budgets accordingly. Here, revenue is defined by advertisers' goals, which can be dollar revenue, number of orders, number of subscriptions, and so on. The winning ads are charged by user clicks, meaning that advertisers only pay when their ads are clicked by users. Search engines provide platforms for advertisers to manage their bids and apply targeting and budgeting decisions. Figure 1 illustrates a typical hierarchical structure of bid management. Advertisers first create an account and construct several campaigns in the account. For each campaign, advertisers can group keywords and ads into ad groups for targeting and management purposes. Ads are often shared by keywords in a common ad group. For each keyword, advertisers can also determine the matching types used between keywords and search queries, such as ``broad match,'' ``exact match,'' and ``other'' match types. Advertisers can set targeting criteria using geographic and demographic information at the ad group or campaign level. \begin{figure} \includegraphics[height=1.5in, width=2.5in]{search_engine} \caption{Hierarchical structure of bid management.} \end{figure} \subsection{RPC Prediction} In this paper, we focus on RPC prediction from the advertisers' perspective. First, we define ``bid units'' as the atomic units at which advertisers set their bids. Bid units are different from keywords because the same keywords can be targeted in multiple ad groups or campaigns and set with different bids. For example, in Figure 1, we consider ``Keyword 1 + Match Type 1'' under ``Ad Group 1'' as a bid unit and ``Keyword 1 + Match Type 1'' under ``Ad Group 2'' as another bid unit. The performance data we collect on the advertisers' side contain daily impressions, clicks, conversions and attributed revenue at the bid-unit level, and we remove the records with zero clicks because our goal is to predict the RPC for each bid unit. The problem is that, given the historical clicks and revenue data $\{x_{i,m},y_{i,m}; m=1,2,...,n_i\}$, we want to predict the next day's RPC for bid unit $i$. The features we can utilize are the hierarchical structure information of the bid units, such as corresponding campaigns, ad groups, and keywords, as well as some upper level variables. Here, upper level variables refer to the information above the bid-unit level, such as geo targeting at the campaign level, which is shared by the bid units under each campaign. A well-known challenge in the RPC prediction problem is that, at the bid-unit level, the data are very sparse. From the perspective of users' behaviors, the sparsity challenge is twofold. First, for a large number of bid units, only a small number of days record non-zero clicks. We name the sparsity of clicks as $x$-sparsity. Second, among all the bid units that are clicked, the majority does not generate any revenue for the advertiser. This sparsity of revenue is denoted as $y$-sparsity. To further illustrate this phenomenon, we examine one month of data for a client of Adobe Advertising Cloud. The average $x$-sparsity and $y$-sparsity are about 90\% and 98\%, meaning only 10\% of the dates collect click data and among the dates with click data, about 98\% have zero revenue. Thus, if we build models at the bid-unit level by pushing down the upper level variables, we tend to generate zero RPC predictions for most bid units. These sparse predictions are undesirable for online advertising for several reasons. First, the bid units have potentials. Previous records of value zero do not necessarily mean the following day still bears a zero, and these potentials would be fully ignored by sparse predictions, leading to an overfitting model. Second, sparse predictions do not help distinguish the bid units if limited resources need to be allocated to them. \begin{figure} \includegraphics[height=2in, width=3in]{hierarchy_choosing} \caption{Dynamic hierarchy choosing process.} \end{figure} \section{Related Work} Although RPC is a vital metric in advertiser bidding decisions, the RPC-related literature is limited, partly because of the confidentiality of revenue data. Among the few existing studies, the work most related to our study is \cite{Sodomka}, which proposed a hierarchical model for predicting value per click, where the hierarchy is fixed a \textit{priori} and defined by ad group, campaign, and account. A linear model is used at each layer, and the aggregated loss is minimized. On the other hand, extensive literature has studied CTR and CVR predictions and offered some attempts to utilize the data hierarchies/clusters in addressing data sparsity. Among those few attempts, \cite{Agarwal} assumed a predefined advertiser-publisher pair hierarchy and built a Poisson log-linear model for each node. Using the same data hierarchy, \cite{Agarwal2} proposed a tree-structured Markov model. Other than linear regression, \cite{Menon} modeled CTR from a collaborative filtering perspective of view. In addition to the preexisting advertiser hierarchy and publisher hierarchy, \cite{Lee} also considered clustering user-level information by grouping data within a specified Euclidean distance. To the best of our knowledge, all existing methods require a predetermined hierarchy, a \textit{priori} using the data structure and feature set in the data, which becomes a challenge when more user-defined features are involved. Our study provides a methodology that determines the hierarchical structure using information in the data so that the structure can be determined layer by layer during the model estimation process. Another contribution of our study to the literature is that the existing methods allow the child nodes to borrow information from their parents, mostly by combining the mean values of the parents and the children while ignoring the uncertainty of the mean values. In this paper, we propose a new method that allows the uncertainty to be incorporated before combining these values from parent or child nodes. \section{Methodology} In this section, we present the proposed methodology in detail. For illustration, we first demonstrate how a two-level Bayesian regression model can be utilized in the RPC prediction problem in section 4.1. Then, we introduce hierarchical shrinkage loss (HSL) for determining the hierarchy empirically in section 4.2. We finish our discussion of the proposed DHEB method in section 4.3. \subsection{Two-level Hierarchical Bayes} For each bid unit $i$, we denote its RPC $\beta_i$ as a random variable. Then we construct a linear regression model: \begin{displaymath} Y_i=\beta_i X_i+\epsilon_i, \end{displaymath} where $X_i=(x_{i,1},x_{i,2}, ... ,x_{i,n_i} )^T$ and $Y_i=(y_{i,1},y_{i,2}, ... ,y_{i,n_i} )^T$ are historical number of clicks and revenue, respectively, and $\epsilon_i\sim N(0,\sigma_{\epsilon_i}^2)$. Our goal is to make an estimation for $\beta_i$ for each bid unit. Under the Bayesian framework, we assume a prior distribution of parameter $\beta_i$, then combine the prior with the likelihood function to yield a posterior. Assume $\beta_i$ has a normal prior distribution: \begin{displaymath} \beta_{i_{prior}} \sim N(\mu_0,\sigma_0^2 ), \end{displaymath} where $\mu_0$, $\sigma_0$ are pre-specified hyper-parameters. Given the likelihood $Y_i |\beta_i,X_i,\sigma_{\epsilon_i}^2 \sim N(\beta_i X_i,\sigma_{\epsilon_i}^2 I_{n_i })$, where $I_{n_i}$ is an $n_i \times n_i$ identity matrix, the posterior for $\beta_i$ is: \begin{displaymath} \beta_{i_{post}} | Y_i,X_i,\sigma_{\epsilon_i}^2\propto P(Y_i | \beta_i,X_i,\sigma_{\epsilon_i}^2 )P(\beta_{i_{prior}}) \sim N(\mu_i,\sigma_i^2 ), \end{displaymath} \begin{equation} \mu_i = \frac{(\sigma_0^2)^{-1} \mu_0+(\sigma_{\epsilon_i}^2 )^{-1} X_i^T X_i (X_i^T X_i )^{-1} X_i^T Y_i}{(\sigma_0^2)^{-1}+(\sigma_{\epsilon_i}^2)^{-1} X_i^T X_i}, \end{equation} \begin{equation} (\sigma_i^2)^{-1}=(\sigma_0^2)^{-1}+(\sigma_{\epsilon_i}^2)^{-1} X_i^T X_i. \end{equation} By applying the same prior distribution for all $\beta_i$s and using the posterior mean as the predicted RPC for each bid unit, we get non-sparse predictions that contain information borrowed by incorporating a prior distribution. This prior information can be obtained by empirical Bayes leveraging the overall data. For data containing more features, a multi-level hierarchical Bayesian method is required to enable the propagation of information across the hierarchical structure and allow for information sharing among subgroups related in the hierarchy. For example, bid units in the same ad groups may intuitively perform more similarly; thus, it makes more sense for them to share the same prior distribution. We fix the bottom level of the hierarchy to be the bid-unit level in order to differentiate the various bid units. The question now is determining the appropriate intermediate levels as shown in the top row of Figure 2. In a conventional hierarchical Bayesian model, the hierarchy is predetermined by domain knowledge. In our application, although there is a hierarchical structure for bid management as we introduced in section 2.1, issues still exist when trying to set the hierarchy involving features without a natural hierarchy. For example, under each ad group, advertisers set multiple keywords to target, which indicates that we can create a hierarchy with ``Keyword'' under ``Ad Group.'' Nevertheless, a common keyword can also appear in different ad groups targeting different types of customers. In this case, it is reasonable to put ``Ad Group'' under ``Keyword'' as well. This situation then calls for a data-driven approach to determine the hierarchy structure for the HB model. \subsection{Hierarchical Shrinkage Loss} Intuitively, it is similar to tree splitting using categorical variables in tree models, which grows a tree according to a certain predefined loss. In the interest of visualization and brevity, we use the terminology ``node'' as in tree models. The root node contains the whole population with all bid units. If we use ``Keyword'' as the first splitting feature and there are $n$ unique keywords in the data, then the root node will be split to $n$ child nodes, with each containing the bid units that share the same keyword. For bid units in each child node, we estimate the same RPC for them, and we assume that child nodes under a common parent node share the same prior distribution; thus, we use the term ``parent information'' to represent the ``prior information.'' Based on the observation of the posterior mean (1), which is a weighted average of parent information and information of itself, we develop the hierarchical shrinkage loss (HSL): \begin{equation} L_p (l,\beta)=\sum_{k=1}^{n_l}h(\alpha_{p_{lk}} f(\beta_{p_{lk}},X_{p_{lk}},Y_{p_{lk}}) + \gamma_{p_{lk}} g(\beta_{p_{lk}},\beta_p)), \end{equation} where $p$ denotes the parent node; $p_l=\{p_{l1},p_{l2}, ... ,p_{ln_l}\}$ denotes the child nodes of $p$ when splitting by feature $l$; $\beta_{p_{lk}}$ and $\beta_p$ represent the RPC predictions in child node $p_{lk}$ and parent node $p$, respectively; $X_{p_{lk}}$ and $Y_{p_{lk}}$ are the data in child node $p_{lk}$; $f$ and $g$ are functions measuring the within-node loss and loss to the parent node; $\alpha_{p_{lk}}$ and $\beta_{p_{lk}}$ represent the importance of the two losses; and $h(x)$ is a scalar function that transforms x to the order of interest. There are two terms in HSL: the first measures the weighted information loss within each child node, and the second considers the discrepancy between the estimators of the child node and the parent node. The estimator of each child node then considers not only the data within itself, but also the information of its parent, who also inherits from its parent according to the hierarchy. This additional hierarchy information leads to a more stable model as information from a larger subgroup is used. \subsection{Dynamic Hierarchical Empirical Bayes} In this section, we illustrate how DHEB builds a hierarchy using HSL. In the multi-level hierarchical Bayesian method, it is assumed that the parameters of the child nodes under the same parent node are from a common prior distribution and the prior information flows through the hierarchy. In a fully Bayesian analysis, a complete joint posterior distribution is generated according to the predetermined hierarchy, and simulations are usually applied to get inferences. This process can be computationally expensive. Instead of a fully Bayesian analysis, we employ empirical Bayes to grow the hierarchy from top to bottom. The proposed method not only provides a method for determining the hierarchy, but also presents an efficient way to get inferences. We illustrate how to construct a loss function to choose the splitting features for the intermediate levels using an example in the bottom row of Figure 2. Suppose we are in the node ``Keyword 1'' and want to decide which feature to use for the further subdivision. Assume we use ``Geo'' as the splitting feature and split the data for each ``Geo'' as a child node $j$. Here, we use $j$ to differentiate from bid unit $i$ in section 4.1. Similar to section 4.1, we assume all RPCs $\beta_j$s under ``Keyword 1'' across different ``Geo's'' are related and generated from a common prior distribution, which is $\beta_{j_{prior}} \sim N(\mu_0,\sigma_0^2)$. Then the posterior distribution of $\beta_j$ for each ``Geo'' node is $\beta_{j_{post}} | Y_j,X_j,\sigma_{\epsilon_j}^2 \sim N(\mu_j,\sigma_j^2 )$, where \begin{equation} \mu_j = \frac{(\sigma_0^2)^{-1} \mu_0+(\sigma_{\epsilon_j}^2 )^{-1} X_j^T X_j (X_j^T X_j )^{-1} X_j^T Y_j}{(\sigma_0^2)^{-1}+(\sigma_{\epsilon_j}^2)^{-1} X_j^T X_j}, \end{equation} \begin{equation} (\sigma_j^2)^{-1}=(\sigma_0^2)^{-1}+(\sigma_{\epsilon_j}^2)^{-1} X_j^T X_j. \end{equation} Using the posterior mean $\mu_j$ as an estimate for $\beta_j$ in each child node, we can construct a loss function by degenerating (3) to the current layer as follows: \begin{align} L_p (l ,\beta)=&\sum_{j=1}^n n_j ((\sigma_{\epsilon_j}^2 )^{-1} X_j^T X_j (\beta_j-(X_j^T X_j )^{-1} X_j^T Y_j )^2+\nonumber\\ &(\sigma_0^2)^{-1} (\beta_j - \mu_0 )^2 ) \nonumber\\ =& \sum_{j=1}^n n_j (\alpha_j f(\beta_j,X_j,Y_j )+\gamma_j g(\beta_j,\mu_0 )),\label{eqn:1} \end{align} where the generic functions in (3) are \\ $f(\beta_j,X_j,Y_j )=(\beta_j-(X_j^T X_j )^{-1} X_j^T Y_j )^2$, $g(\beta_j,\mu_0 )=(\beta_j - \mu_0 )^2$, $\alpha_j=(\sigma_{\epsilon_j}^2 )^{-1} X_j^T X_j$, $\gamma_j=(\sigma_0^2)^{-1}$, and $h(x)=nx$, with node $p_{lk}$ denoted as $j$ for short. The optimal solution of $\beta_j$ would be $\mu_j$. Function $f$ represents the difference between the parameters of the child nodes and the OLS estimations based on the sample data. Function $g$ measures the difference between the parameters of the child nodes and parent node, which is represented by the prior mean. The weights of the two losses $\alpha_j$ and $\gamma_j$ are inversely proportional to the variance of the OLS estimator and prior variance. The basic idea is intuitive: If the prior variance is larger, it provides noisier information regarding the $\beta_j$ estimates and, hence, its contribution is smaller than the case when the prior variance is smaller. Similarly, if the sample data are divergent and noisy, they will get less weight. $h(x)=nx$, where $n$ is the number of observations in the node. We multiply the loss for each child node by the number of observations in the node because we shrink the loss to one node level by $f$ and $g$. In order to make the losses for different splitting features comparable, we calculate the loss at the individual observation level and treat the loss at one node level as a representation for all the observations in this node. Once we have the loss function, we can decide which feature to use for partition as \begin{equation} l^*=\arg\!\min_l L_p (l,\hat{\beta}) \end{equation} Suppose we choose ``Geo'' for the second level and we need to decide the splitting variable for the third level. We assume the posterior distribution of $\beta_j$ as the prior distribution of $\beta_{jk}$ under ``Geo $j$'' and apply the same method recursively, which is $\beta_{{jk}_{prior}} \sim \beta_{j_{post}}$ (Figure 2, bottom row right). To get loss (6), both prior distribution of $\beta_j$ and regression variance $\sigma_{\epsilon_j}^2$ are assumed known; therefore, sample data should be used to get estimations. For prior distribution, only the parameters in the root node are necessary because the posterior of the parent node would be used as prior for its child nodes. Empirical Bayes can be applied when there is a lack of prior knowledge. Here, we give an example by using the sample mean as the prior mean and weighted sample variance as the prior variance: $\mu_0=\frac{\sum_m y_m}{\sum_m x_m}$, $\sigma_0^2=\frac{\sum_m x_m (\frac{y_m}{x_m} -\mu_0 )^2}{\sum_m x_m - 1}$, where $m$ denotes total historical data for all bid units. The variance $\sigma_{\epsilon_j}^2$ needs to be estimated in each node which can be given by: $\hat{\beta}_{j,OLS}=(X_j^T X_j )^{-1} X_j^T Y_j$, $\sigma_{\epsilon_j}^2=\frac{1}{n_j-1} (Y_j-\hat{\beta}_{j,OLS} X_j )^T (Y_j-\hat{\beta}_{j,OLS} X_j )$, where $\hat{\beta}_{j,OLS}$ is the OLS estimator and $n_j$ is the number of observations for node $j$. Another problem is when to stop splitting. Here, we propose a stopping criterion: \begin{equation} \frac{SSE(p_{l^*} )}{SSE(p)}>r, \end{equation} where $SSE(p)=||Y_p-\hat{\beta}_p X_p ||^2$ and $SSE(p_{l^*} )=\sum_{j\in p_{l^*}}||Y_j- \hat{\beta}_j X_j ||^2$, denoting the sum of squared errors for the parent node $p$ and child nodes $p_{l^*}$. This means a node will stop growing when the total sum of squared errors does not decrease by a certain ratio $1-r$. The final step would be attaching the bid-unit level to the bottom of the chosen hierarchy. The procedure loops the leaf nodes of the hierarchy and subdivides them into child nodes, with each node containing the data for a specific bid unit. The proposed DHEB also provides an approach to get inferences through a hierarchy. If we have a fixed hierarchy, we can apply equations (4) and (5) to get stepwise posterior distributions from the root to bottom levels and then obtain inferences. \begin{algorithm} \caption{Dynamic Hierarchical Empirical Bayes} \label{DHEB} Initialize a set of nodes to split $Q=\{root\}$\ \For{node $p$ in $Q$} { get splitting variable $l^*$ and create child nodes $C = p_{l^*}$ according to (7)\\ \If{stopping criterion (8) is satisfied}{ $C = \varnothing$ } \Else{attach $C$ to $p$} $Q = Q\backslash\{p\}\bigcup C$\ } \For{leaf node $ln$ in leaves} { \For{bid unit $bu$ in $ln$} { create child node $c$ for $bu$ and attach it to $ln$\ } } \end{algorithm} \section{Simulation Results} \begin{figure} \includegraphics[height=1.8in, width=3.4in]{simulation_results} \caption{Simulation results.} \end{figure} In this section, we evaluate the proposed method on several simulated datasets. On each dataset, we conduct an analysis using 6 models: \begin{enumerate} \item Weighted average (WA): The predicted RPC of each bid unit is the weighted average of historical RPC using the number of clicks as weights. \item Regularized linear regression (RLR): This fits a regularized linear regression by pushing down all the upper level features. \item Two-level HB (2HB): This model was discussed in section 4.2. The hierarchy is ``Root - Bid Unit.'' \item Three-level HB (3HB): This model first predefines a three-level fixed hierarchy. Then we use Rstan \cite{Stan} to do the posterior sampling and treat the posterior means at the bottom level as predictions. We limit the hierarchy to three levels because the more levels we have, the more computationally expensive the model is and, when using real-world data, many assumptions may not be satisfied, which makes the simulation difficult to converge. \item Multi-level fixed hierarchical empirical Bayes with true hierarchy (FHEB): We fix the hierarchy as the true hierarchy used during the simulation and apply the same inference approach as the proposed method. \item DHEB: This is our proposed method. \end{enumerate} We apply these six models on a set of simulated data. Data are generated by the procedure as follows: \begin{enumerate} \item Create 100 bid units and 4 upper level features (named A, B, C, and D); each feature has 10 to 20 categories. Set the date range to 6 months (i.e., from ``2017-01-01'' to ``2017-06-30''). \item Assume the implicit hierarchy is A - B - C - D - Bid Unit. \item Set the top prior mean $\mu_0$ and variance $\sigma_0^2$. \item For nodes in the intermediate levels, generate the mean $\mu_c$ for child nodes from $N(\mu_p,\sigma_p^2 )$, which is the parent node distribution. The variance of child nodes $\sigma_c^2$ is predetermined. Here, we just use the same variance as $\sigma_0^2$. \item For the bottom bid-unit level, we apply (4) to generate $\mu_c$ and set RPC for this bid unit as $\mu_c$. We generate a list of clicks $X$ with length $n$, then revenue $Y=RPC\times X+\epsilon$, where $\epsilon \sim N(0,\sigma_{\epsilon}^2)$ and $\sigma_{\epsilon}^2$ is predetermined. \item $x$-sparsity is determined by $n$, which is the number of observations we generated in (5). The $x$-sparsity is higher as $n$ is smaller because we fix the date range. $y$-sparsity is denoted by $s$, and we will randomly set $s$ of the revenue to be zero. \item We apply 9 pairs of $n$ and $s$ combinations and generate 10 datasets for each combination. \end{enumerate} We use two months of data to predict RPCs for next day and test in a rolling way for 30 days. The performance metric is: \begin{equation} \text{AVG-MSE} =\frac{1}{N}(\sum_{t=1}^N \frac{1}{N_t} \sum_{i=1}^{N_t}(\hat{\beta}_{it} \times x_{it} - y_{it})^2), \end{equation} where $N=30$ is the number of testing days; $N_t$ is the number of bid units on day $t$; $(x_{it}, y_{it})$ is the true data for bid unit $i$ on day $t$; and $\hat{\beta}_{it}$ is the predicted RPC. For 3HB, we apply a hierarchy as ``Root - A - Bid Unit''; for FHEB, we deploy the true hierarchy. Because WA is a baseline method, we calculate the improvement of the other 5 models relative to WA. The improvement is represented by the reduced percentage of AVG-MSE and would be negative if the AVG-MSE of WA is smaller. The top row of Figure 3 shows the comparison of the 5 models for different $n$ and $s$. As we can see, for different combinations of $n$ and $s$, FHEB outperforms all the other models, with DHEB ranking second. As both $x$-sparsity and $y$-sparsity increase, the benefits obtained from FHEB and DHEB become greater. For time complexity, we plot the ratio between the running time of the other 5 models and WA, as shown in the bottom row of Figure 3. For 3HB, it takes a much longer time to do the sampling, and the model does not perform best due to the limited number of levels. In practice, if we are confident about what the true hierarchy is, FHEB provides an approach to get predictions without worrying about time complexity. If we are not sure about how to build the hierarchy, DHEB can determine the hierarchy empirically and give desirable predictions. \section{Experimental Results} \subsection{Model Comparison} The data come from multiple online advertising campaigns owned by a common advertiser. In the bid-unit level, the performance data contain the number of clicks and collected revenue, ranging from ``2017-09-01'' to ``2017-12-23''. In addition, the data record the structural features of the keywords as shown in Table 1. Here, instead of regular hierarchical features, we introduce ``Day of Week,'' which provides an additional group for the daily data by indicating whether the day is Monday, Tuesday, etc. ``Geo'' represents geo targeting for the campaigns; it only has one unique category in this dataset. When a hierarchy is established in a hierarchical model, some features have a natural relationship with each other, such as ``Search Engine,'' ``Account,'' ``Campaign,'' and ``Ad Group.'' However, for ``Geo,'' ``Keyword,'' ``Match Type,'' and ``Day of Week,'' it is hard to determine their positions and order. In addition, it may not be necessary to include all structural features in the hierarchy. \begin{table} \caption{Structural features statistics} \label{tab:commands} \begin{tabular}{ccl} \toprule Feature & Symbol & \# Categories\\ \midrule Search Engine & SE & 3 \\ Account & AC & 3\\ Geo & GEO & 1\\ Campaign & CP & 156\\ Ad Group & AG & 2847\\ Keyword & KW & 3208\\ Match Type & MT & 3\\ Day of Week & DOW & 7\\ Bid Unit & BU & 5148\\ \bottomrule \end{tabular} \end{table} We compare the 6 models and use the same evaluation metric as in section 5. For 3HB, we apply a hierarchy as ``Root - Campaign - Bid Unit''; for FHEB, we chose a hierarchy by domain knowledge, which is ``Root - Search Engine - Account - Campaign - Ad group - Keyword - Match Type - Day of Week - Bid Unit.'' We dropped ``Geo'' because it is unique. In Figure 4, the left plot shows the AVG-MSE improvement compared with WA. As we can see, DHEB outperforms other methods. The middle plot demonstrates the time complexity compared with WA. 3HB takes a much longer time due to the sampling process. \begin{figure} \includegraphics[height=0.9in, width=3.2in]{experimental_results} \caption{Experimental results.} \end{figure} \begin{figure} \includegraphics[height=1.5in, width=3.2in]{experimental_hierarchy} \caption{Chosen hierarchy.} \end{figure} \subsection{Two-phase System} \begin{figure} \includegraphics[height=1.5in, width=2in]{system} \caption{Two-phase system.} \end{figure} Figure 5 shows the hierarchy trained on several testing days. As we can see, the hierarchy does not change frequently over time, which makes sense as there is only one day difference between the training data for two consecutive days. There are three modules: (1) data collection: obtaining bid units features, the daily number of clicks, and revenue in history; (2) model training: training the DHEB model and building a hierarchy; (3) prediction serving: giving RPC prediction based on the hierarchy determined in model training. Module (2) is the most time-comsuming part. We separate these three modules into offline and online phases as shown in Figure 6, where in offline phase, we do model training and in online phase, we do prediction serving based on the trained hierarchy. Given the fact that the hierarchy determined by DHEB does not change a lot in a short period, we schedule the offline phase in a low frequency and run the online phase in real time. We introduce parameter $T$ as the period of the offline phase. $T=1$ means we run the offline phase every day, $T=2$ means every other day, and so on. The right plot in Figure 4 consists of AVG-MSE improvement compared to $T=1$ for different values of $T$. An appropriate $T$ would be 4, which will reduce the time complexity without making many sacrifices in model accuracy. \section{Conclusions} In this paper, we propose a Dynamic Hierarchical Empirical Bayesian (DHEB) method to build a multi-level hierarchical model to overcome the sparsity challenge in online advertising data. The proposed method provides a way to choose hierarchical levels by incorporating a loss function, such as the function used in tree models. The method is also equipped with an empirical Bayesian approach to get inferences through a hierarchy. It is applicable in many practical problems where data are sparse and hierarchical structure can be leveraged to obtain shrinkage-based estimations. In addition, the proposed regularized loss function can be applied in traditional tree models as well as other tree-based methods, as an approach to borrow information from the parent node in order to deal with data sparseness. We also present a two-phase system which can serve prediction in real time. \bibliographystyle{ACM-Reference-Format}

456f3ebfb97252e3f7945ff262eb4bae78755080

\section{Introduction} We consider a three-players zero-sum game with two strategic variables. Three players are Players A, B and C. Two strategic variables are $t_i$ and $s_i$, $i=A, B, C$. They are related by invertible functions. The game is symmetric for Players A and B in the sense that they have the same payoff functions. On the other hand, Player C may have a different payoff function. Thus, the game is \emph{partially asymmetric}. In Section 3 we will show the following main results. \begin{enumerate} \item The equilibrium when all players choose $t_i$'s is equivalent to the equilibrium when Players A and B choose $t_i$'s and Player C chooses $s_C$ as their strategic variables. \item The equilibrium when all players choose $s_i$'s is equivalent to the equilibrium when Players A and B choose $s_i$'s and Player C chooses $t_C$ as their strategic variables. \end{enumerate} An example of three-players zero-sum game with two strategic variables is a relative profit maximization game in a three firms oligopoly with differentiated goods. See Section \ref{ex}. In that section we will show; \begin{enumerate} \item The equilibrium when all players choose $t_i$'s is not equivalent to the equilibrium when Players A and C choose $t_i$'s and Player B chooses $s_B$ as their strategic variables. \item The equilibrium when all players choose $t_i$'s is not equivalent to the equilibrium when Players A and B choose $s_i$'s and Player C chooses $t_C$ as their strategic variables. \item The equilibrium when all players choose $s_i$'s is not equivalent to the equilibrium when Players A and B choose $t_i$'s and Player C chooses $s_C$ as their strategic variables. \item The equilibrium when all players choose $s_i$'s is not equivalent to the equilibrium when Players A and C choose $s_i$'s and Player B chooses $t_B$ as their strategic variables. \item The equilibrium when all players choose $t_i$'s is not equivalent to the equilibrium when all players $s_i$'s. \end{enumerate} In a symmetric game, in which all players have the same payoff functions, they are all equivalent\footnote{\cite{hst}.}. In the next section we present a model of this paper and prove a preliminary result which is a variation of Sion's minimax theorem. \section{The model} We consider a three-players zero-sum game with two strategic variables. Three players are Players A, B and C. Two strategic variables are $t_i$ and $s_i$, $i=A, B, C$. The game is symmetric for Players A and B in the sense that they have the same payoff functions. On the other hand, Player C may have a different payoff function. $t_i$ is chosen from $T_i$ and $s_i$ is chosen from $S_i$. $T_i$ and $S_i$ are convex and compact sets in linear topological spaces, respectively, for each $i\in \{A, B, C\}$. The relations of the strategic variables are represented by \begin{equation*} s_i=f_i(t_A, t_B, t_C),\ i=A, B, C, \end{equation*} and \begin{equation*} t_i=g_i(s_A, s_B, s_C),\ i=A, B, C. \end{equation*} $(f_A, f_B, f_C)$ and $(g_A, g_B, g_C)$ are continuous, invertible, one-to-one and onto functions. When Players A and B choose $t_A$ and $t_B$ and Player C chooses $s_C$, then $t_C$ is determined according to \begin{equation*} t_C=g_C(f_A(t_A, t_B, t_C),f_B(t_A, t_B, t_C), s_C). \end{equation*} We denote this $t_C$ by $t_C(t_A, t_B, s_C)$. When Players A and B choose $s_A$ and $s_B$ and Player C chooses $t_C$, then $t_A$ and $t_B$ are determined according to \begin{equation*} \left\{ \begin{array}{l} t_A=g_A(s_A, s_B, f_C(t_A, t_B, t_C))\\ t_B=g_B(s_A, s_B, f_C(t_A, t_B, t_C)). \end{array}% \right. \end{equation*} We denote these $t_A$ and $t_B$ by $t_A(s_A, s_B, t_C)$ and $t_B(s_A, s_B, t_C)$. When all players choose $s_{A}$, $s_{B}$ and $s_{C}$, $t_{A}$, then $t_{B}$ and $t_{C}$ are determined according to \begin{equation*} t_{A}=g_{A}(s_{A},s_{B},s_{C}),\ t_{B}=g_{B}(s_{A},s_{B},s_{C}),\ t_{C}=g_{C}(s_{A},s_{B},s_{C}). \end{equation*}% Denote these $t_{A}$, $t_{B}$ and $t_{C}$ by $t_{A}(s_{A},s_{B},s_{C})$, $% t_{B}(s_{A},s_{B},s_{C})$ and $t_{C}(s_{A},s_{B},s_{C})$. The payoff function of Player $i$ is $u_i,\ i=A, B, C$. It is written as \begin{equation*} u_i(t_A, t_B, t_C),\ i\in \{A, B, C\}. \end{equation*} We assume \begin{quote} $u_i: T_1\times T_2\times T_3\Rightarrow \mathbb{R}$ for each $i\in \{A, B, C\}$ is continuous on $T_1\times T_2\times T_3$. Thus, it is continuous on $S_1\times S_2\times S_3$ through $f_i,\ i=A, B, C$. It is quasi-concave on $T_i$ and $S_i$ for a strategy of each other player, and quasi-convex on $T_j,\ j\neq i$ and $S_j,\ j\neq i$ for each $t_i$ and $s_i$. \end{quote} We do not assume differentiability of the payoff functions. Symmetry of the game for Players A and B means that in the payoff function of each player, Players A and B are interchangeable. Since the game is a zero-sum game, the sum of the values of the payoff functions of the players is zero. We assume that all $T_i$'s are identical, and all $S_i$'s are identical. Denote them by $T$ and $S$. Sion's minimax theorem (\cite{sion}, \cite{komiya}, \cite{kind}) for a continuous function is stated as follows. \begin{lemma} Let $X$ and $Y$ be non-void convex and compact subsets of two linear topological spaces, and let $f:X\times Y \rightarrow \mathbb{R}$ be a function that is continuous and quasi-concave in the first variable and continuous and quasi-convex in the second variable, then \begin{equation*} \max_{x\in X}\min_{y\in Y}f(x,y)=\min_{y\in Y}\max_{x\in X}f(x,y). \end{equation*} \label{l1} \end{lemma} We follow the description of Sion's theorem in \cite{kind}. Applying this lemma to the situation of this paper, we have the following relations. \begin{equation*} \max_{t_{A}\in T}\min_{t_{B}\in T}u_{A}(t_{A},t_{B},t_{C})=\min_{t_{B}\in T}\max_{t_{A}\in T}u_{A}(t_{A},t_{B},t_{C}),\ \max_{t_{B}\in T}\min_{t_{A}\in T}u_{B}(t_{A},t_{B},t_{C})=\min_{t_{A}\in T}\max_{t_{B}\in T}u_{B}(t_{A},t_{B},t_{C}). \end{equation*}% \begin{equation*} \max_{t_{A}\in T}\min_{t_{C}\in T}u_{A}(t_{A},t_{B},t_{C})=\min_{t_{C}\in T}\max_{t_{A}\in T}u_{A}(t_{A},t_{B},t_{C}),\ \max_{t_{B}\in T}\min_{t_{C}\in T}u_{B}(t_{A},t_{B},t_{C})=\min_{t_{C}\in T}\max_{t_{B}\in T}u_{B}(t_{A},t_{B},t_{C}). \end{equation*}% \begin{align*} & \max_{t_{A}\in T}\min_{t_{B}\in T}u_{A}(t_{A},t_{B},t_{C}(t_{A},t_{B},s_{C}))=\min_{t_{B}\in T}\max_{t_{A}\in T}u_{A}(t_{A},t_{B},t_{C}(t_{A},t_{B},s_{C})), \\ & \max_{t_{B}\in T}\min_{t_{A}\in T}u_{B}(t_{A},t_{B},t_{C}(t_{A},t_{B},s_{C}))=\min_{t_{A}\in T}\max_{t_{B}\in T}u_{B}(t_{A},t_{B},t_{C}(t_{A},t_{B},s_{C})). \end{align*}% \begin{align*} & \max_{t_{A}\in T}\min_{s_{C}\in S}u_{A}(t_{A},t_{B},t_{C}(t_{A},t_{B},s_{C}))=\min_{s_{C}\in S}\max_{t_{A}\in T}u_{A}(t_{A},t_{B},t_{C}(t_{A},t_{B},s_{C})),\\ & \max_{t_{B}\in T}\min_{s_{C}\in S}u_{B}(t_{A},t_{B},t_{C}(t_{A},t_{B},s_{C}))=\min_{s_{C}\in S}\max_{t_{B}\in T}u_{B}(t_{A},t_{B},t_{C}(t_{A},t_{B},s_{C})), \\ \end{align*}% Further we show the following result. \begin{lemma} \begin{align*} &\min_{t_C\in T}\max_{t_A\in T}u_A(t_A, t_B, t_C)=\min_{s_C\in S}\max_{t_A\in T}u_A(t_A, t_B, t_C(t_A,t_B,s_C)) \\ &=\max_{t_A\in T}\min_{s_C\in S}u_A(t_A, t_B, t_C(t_A,t_B,s_C))=\max_{t_A\in T}\min_{t_C\in T}u_A(t_A, t_B, t_C), \end{align*} and \begin{align*} &\min_{t_C\in T}\max_{t_B\in T}u_B(t_A, t_B, t_C)=\min_{s_C\in S}\max_{t_B\in T}u_B(t_A, t_B, t_C(t_A,t_B,s_C)) \\ &=\max_{t_B\in T}\min_{s_C\in S}u_B(t_A, t_B, t_C(t_A,t_B,s_C))=\max_{t_B\in T}\min_{t_C\in T}u_B(t_A, t_B, t_C). \end{align*} \label{l3} \end{lemma} \begin{proof} $\max_{t_A\in T}u_A(t_A, t_B, t_C(t_A,t_B,s_C))$ is the maximum of $u_A$ with respect to $t_A$ given $t_B$ and $s_C$. Let $\tilde{t}_A(s_C)=\arg\max_{t_A\in T}u_A(t_A, t_B, t_C(t_A,t_B,s_C))$, and fix the value of $t_C$ at \begin{equation} t_C^0=g_C(f_A(\tilde{t}_A(s_C), t_B, t_C^0), f_B(\tilde{t}_A(s_C), t_B, t_C^0), s_C).\label{tc1} \end{equation} We have \begin{align*} \max_{t_A\in T}u_A(t_A, t_B, t_C^0)\geq u_A(\tilde{t}_A(s_C), t_B, t_C^0)=\max_{t_A\in T}u_A(t_A, t_B, t_C(t_A,t_B,s_C)), \end{align*} where $\max_{t_A\in T}u_A(t_A, t_B, t_C^0)$ is the maximum of $u_A$ with respect to $t_A$ given the value of $t_C$ at $t_C^0$. We assume that $\tilde{t}_A(s_C)=\arg\max_{t_A\in T}u_A(t_A, t_B, t_C(t_A,t_B,s_C))$ is single-valued. By the maximum theorem and continuity of $u_A$, $\tilde{t}_A(s_C)$ is continuous, then any value of $t_C^0$ can be realized by appropriately choosing $s_C$ given $t_B$ according to (\ref{tc1}). Therefore, \begin{equation} \min_{t_C\in T}\max_{t_A\in T}u_A(t_A, t_B, t_C)\geq \min_{s_C\in S}\max_{t_A\in T}u_A(t_A, t_B, t_C(t_A,t_B,s_C)).\label{4-11} \end{equation} On the other hand, $\max_{t_A\in T}u_A(t_A, t_B, t_C)$ is the maximum of $u_A$ with respect to $t_A$ given $t_B$ and $t_C$. Let $\tilde{t}_A(t_C)=\arg\max_{t_A\in T}u_A(t_A, t_B, t_C)$, and fix the value of $s_C$ at \begin{equation} s_C^0=f_C(\tilde{t}_A(t_C), t_B, t_C).\label{sc1} \end{equation} Thus, we have \begin{align*} \max_{t_A\in T}u_A(t_A, t_B, t_C(t_A,t_B,s_C^0))\geq u_A(\tilde{t}_A(s_C), t_B, t_C(t_A,t_B,s_C^0))=\max_{t_A\in T}u_A(t_A, t_B, t_C), \end{align*} where $\max_{t_A\in T}u_A(t_A, t_B, t_C(t_A,t_B,s_C^0))$ is the maximum of $u_A$ with respect to $t_A$ given the value of $s_C$ at $s_C^0$. We assume that $\tilde{t}_A(t_C)=\arg\max_{t_A\in T}u_A(t_A, t_B, t_C)$ is single-valued. By the maximum theorem and continuity of $u_A$, $\tilde{t}_A(t_C)$ is continuous, then any value of $s_C^0$ can be realized by appropriately choosing $t_C$ given $t_B$ according to (\ref{sc1}). Therefore, \begin{equation} \min_{s_C\in S}\max_{t_A\in T}u_A(t_A, t_B, t_C(t_A,t_B,s_C))\geq \min_{t_C\in S}\max_{t_A\in T}u_A(t_A, t_B, t_C).\label{4-21} \end{equation} Combining (\ref{4-11}) and (\ref{4-21}), we get \[\min_{s_C\in S}\max_{t_A\in T}u_A(t_A, t_B, t_C(t_A,t_B,s_C))=\min_{t_C\in S}\max_{t_A\in T}u_A(t_A, t_B, t_C).\] Since any value of $s_C$ can be realized by appropriately choosing $t_C$ given $t_A$ and $t_B$, we have \[\min_{s_C\in S}u_A(t_A, t_B, t_C(t_A,t_B,s_C))=\min_{t_C\in S}u_A(t_A, t_B, t_C).\] Thus, \[\max_{t_A\in T}\min_{s_C\in S}u_A(t_A, t_B, t_C(t_A,t_B,s_C))=\max_{t_A\in T}\min_{t_C\in S}u_A(t_A, t_B, t_C).\] Therefore, \begin{align*} &\min_{t_C\in T}\max_{t_A\in T}u_A(t_A, t_B, t_C)=\min_{s_C\in S}\max_{t_A\in T}u_A(t_A, t_B, t_C(t_A,t_B,s_C)),\\ =&\max_{t_A\in T}\min_{s_C\in S}u_A(t_A, t_B, t_C(t_A,t_B,s_C))=\max_{t_A\in T}\min_{t_C\in T}u_A(t_A, t_B, t_C), \end{align*} given $t_B$. By similar procedures, we can show \begin{align*} &\min_{t_C\in T}\max_{t_B\in T}u_B(t_A, t_B, t_C)=\min_{s_C\in S}\max_{t_B\in T}u_B(t_A, t_B, t_C(t_A,t_B,s_C)),\\ =&\max_{t_B\in T}\min_{s_C\in S}u_B(t_A, t_B, t_C(t_A,t_B,s_C))=\max_{t_B\in T}\min_{t_C\in T}u_B(t_A, t_B, t_C), \end{align*} given $t_A$. \end{proof} \section{The main results} In this section we present the following main result of this paper. \begin{theorem} The equilibrium when all players choose $t_i$'s is equivalent to the equilibrium when Player C chooses $s_C$ and Players A and B choose $t_i$'s as their strategic variables.\label{t1} \end{theorem} \begin{proof} \begin{enumerate} \item Consider a situation $(t_A,t_B,t_C)=(t,t,t_C)$. By symmetry for Players A and B, \[\max_{t_A\in T}u_A(t_A, t, t_C)=\max_{t_B\in T}u_B(t,t_B,t_C),\] and \[\arg\max_{t_A\in T}u_A(t_A, t, t_C)=\arg\max_{t_B\in T}u_B(t,t_B,t_C)\in T,\] given $t_C$. Let \[t_C(t)=\arg\max_{t_C\in T}u_C(t,t,t_C).\] We assume that it is a single-valued continuous function. Consider the following function. \[t\rightarrow \arg\max_{t_A\in T}u_A(t_A, t, t_C),\ \mathrm{given}\ t_C.\] This function is continuous and $T$ is compact. Thus, there exists a fixed point given $t_C$. Denote it by $t^*(t_C)$, then \[t^*(t_C)=\arg\max_{t_A\in T}u_A(t_A, t^*(t_C), t_C)=\arg\max_{t_B\in T}u_B(t^*(t_C), t_B,t_C),\ \mathrm{given}\ t_C.\] Now we consider the following function. \[t\rightarrow t^*(t_C(t)).\] This also has a fixed point. Denote it by $t^*$ and $t_C(t^*)$ by $t_C^*$, then we have \[t^*=\arg\max_{t_A\in T}u_A(t_A, t^*, t_C^*)=\arg\max_{t_B\in T}u_B(t^*, t_B,t_C^*),\] \[t_C^*=\arg\max_{t_C\in T}u_C(t^*,t^*,t_C).\] \[\max_{t_A\in T}u_A(t_A, t^*, t_C^*)=u_A(t^*, t^*, t_C^*)=\max_{t_B\in T}u_B(t^*, t_B, t_C^*)=u_B(t^*, t^*, t_C^*),\] and \[\max_{t_C\in T}u_C(t^*, t^*, t_C)=u_C(t^*, t^*, t_C^*).\] $(t_A,t_B,t_C)=(t^*,t^*,t_C^*)$ is a Nash equilibrium when all players choose $t_i$'s \item Because the game is zero-sum, \[u_A(t^*, t^*, t_C)+u_B(t^*, t^*,t_C)+u_C(t^*, t^*,t_C)=0.\] By symmetry $u_A(t^*, t^*,t_C)=u_B(t^*, t^*,t_C)$. Thus, \[2u_A(t^*, t^*,t_C)+u_C(t^*, t^*,t_C)=0.\] This means \[2u_A(t^*, t^*,t_C)=-u_C(t^*, t^*,t_C),\] and \[2\min_{t_C\in T}u_A(t^*, t^*,t_C)=-\max_{t_C\in T}u_C(t^*, t^*,t_C).\] From this and symmetry for Players A and B, we get \[\arg\min_{t_C\in T}u_A(t^*, t^*,t_C)=\arg\min_{t_C\in T}u_B(t^*, t^*,t_C)=\arg\max_{t_C\in T}u_C(t^*, t^*,t_C)=t_C^*.\] We have \[\min_{t_C\in T}u_A(t^*, t^*, t_C^*)=u_A(t^*, t^*, t_C^*)=\max_{t_A\in T}u_A(t_A,t^*,t_C^*),\] \[\min_{t_C\in T}u_B(t^*, t^*, t_C^*)=u_B(t^*, t^*, t_C^*)=\max_{t_B\in T}u_B(t^*,t_B,t_C^*).\] Therefore, \[\min_{t_C\in T}\max_{t_A\in T}u_A(t_A, t^*, t_C)\leq \max_{t_A\in T}u_A(t_A, t^*, t_C^*)=\min_{t_C\in T}u_A(t^*,t^*, t_C)\leq \max_{t_A\in T}\min_{t_C\in T}u_A(t_A,t^*, t_C).\] From Lemma \ref{l3} we obtain \begin{align} &\min_{t_C\in T}\max_{t_A\in T}u_A(t_A, t^*, t_C)=\max_{t_A\in T}u_A(t_A, t^*, t_C^*)=\min_{t_C\in T}u_A(t^*,t^*, t_C)\label{l3-1}\\ &=\max_{t_A\in T}\min_{t_C\in T}u_A(t_A,t^*, t_C)=\min_{s_C\in S}\max_{t_A\in T}u_A(t_A, t^*, t_C(t_A,t^*,s_C))\notag\\ &=\max_{t_A\in T}\min_{s_C\in T}u_A(t_A,t^*, t_C(t_A,t^*,s_C)).\notag \end{align} \item Let \[s^0_C(t^*)=f_C(t^*,t^*,t_C^*).\] Since any value of $s_C$ can be realized by appropriately choosing $t_C$, \begin{equation} \min_{s_C\in S}u_A(t^*,t^*,t_C(t^*,t^*,s_C))=\min_{t_C\in T}u_A(t^*,t^*,t_C)=u_A(t^*,t^*,t_C^*).\label{z1} \end{equation} Thus, \[\arg\min_{s_C\in S}u_A(t^*,t^*,t_C(t^*,t^*,s_C))=s^0_C(t^*).\] (\ref{l3-1}) and (\ref{z1}) mean \begin{equation} \min_{s_C\in S}\max_{t_A\in T}u_A(t_A,t^*,t_C(t_A,t^*,s_C))=\min_{s_C\in S}u_A(t^*,t^*,t_C(t^*,t^*,s_C)).\label{z2} \end{equation} We have \[\max_{t_A\in T}u_A(t_A,t^*,t_C(t_A,t^*,s_C))\geq u_A(t^*,t^*,t_C(t^*,t^*,s_C)).\] Therefore, \[\arg\min_{s_C\in S}\max_{t_A\in T}u_A(t_A,t^*,t_C(t_A,t^*,s_C))=\arg\min_{s_C\in S}u_A(t^*,t^*,t_C(t^*,t^*,s_C))=s^0_C(t^*)\] Thus, by (\ref{z2}) \begin{align*} &\min_{s_C\in S}\max_{t_A\in T}u_A(t_A,t^*,t_C(t_A,t^*,s_C))=\max_{t_A\in T}u_A(t_A,t^*,t_C(t^*,t^*,s^0_C(t^*)))\\ =&\min_{s_C\in S}u_A(t^*,t^*,t_C(t^*,t^*,s_C))=u_A(t^*,t^*,t_C(t^*,t^*,s^0_C(t^*))). \end{align*} Therefore, \begin{equation} \arg\max_{t_A\in T}u_A(t_A,t^*,t_C(t_A,t^*,s^0_C(t^*)))=t^*.\label{t1-2} \end{equation} By symmetry for Players A and B, \begin{equation} \arg\max_{t_B\in T}u_B(t^*,t_B,t_C(t^*,t_B,s^0_C(t^*)))=t^*.\label{t1-3} \end{equation} On the other hand, because any value of $s_C$ is realized by appropriately choosing $t_C$, \[\max_{s_C\in S}u_C(t^*,t^*,t_C(t^*,t^*,s_C))=\max_{t_C\in T}u_C(t^*,t^*,t_C)=u_C(t^*,t^*,t_C^*).\] Therefore, \begin{equation} \arg\max_{s_C\in S}u_C(t^*,t^*,t_C(t^*,t^*,s_C))=s^0_C(t^*)=f_C(t^*,t^*,t_C^*).\label{t1-1} \end{equation} From (\ref{t1-2}), (\ref{t1-3}) and (\ref{t1-1}), $(t^*,t^*,t_C(t^*,t^*,s^0_C(t^*)))$ is a Nash equilibrium which is equivalent to $(t^*,t^*,t_C^*)$. \end{enumerate \end{proof} Interchanging $t_i$ and $s_i$ for each player, we can show \begin{theorem} The equilibrium when all players choose $s_i$'s is equivalent to the equilibrium when Player C chooses $t_C$ and Players A and B choose $s_i$'s as their strategic variables.\label{t2} \end{theorem} \section{Various examples}\label{ex} Consider a game of relative profit maximization under oligopoly including three firms with differentiated goods\footnote{% About relative profit maximization in an oligopoly see \cite{mm}, \cite{ebl2}, \cite{eb2}, \cite{st}, \cite{eb1}, \cite{ebl1} and \cite{redondo}}. It is a three-players zero-sum game with two strategic variables. The firms are A, B and C. The strategic variables are the outputs and the prices of their goods. We consider the following six patterns of competition. \begin{enumerate} \item Pattern 1: All firms determine their outputs. It is a Cournot case. The inverse demand functions are \begin{equation*} p_A=a-x_A-bx_B-bx_C, \end{equation*} \begin{equation*} p_B=a-x_B-bx_A-bx_C, \end{equation*} and \begin{equation*} p_C=a-x_C-bx_A-bx_B, \end{equation*} where $0<b<1$. $p_A$, $p_B$ and $p_C$ are the prices of the goods of Firms A, B and C, and $x_A$, $x_B$ and $x_C$ are the outputs of them. \item Pattern 2: Firms A and B determine their outputs, and Firm C determines the price of its good. From the inverse demand functions, \begin{equation*} p_A=(1-b)a+b^2x_B-bx_B+b^2x_A-x_A+bp_C, \end{equation*} \begin{equation*} p_B=(1-b)a+b^2x_B-x_B+b^2x_A-bx_A+bp_C, \end{equation*} and \begin{equation*} x_C=a-bx_B-bx_A-p_C \end{equation*} are derived. \item Pattern 3: Firms A and C determine their outputs, and Firm B determines the price of its good. From the inverse demand functions, \begin{equation*} p_A=(1-b)a+b^2x_C-bx_C+b^2x_A-x_A+bp_B, \end{equation*} \begin{equation*} p_C=(1-b)a+b^2x_C-x_C+b^2x_A-bx_A+bp_B, \end{equation*} and \begin{equation*} x_B=a-bx_C-bx_A-p_B \end{equation*} are derived. \item Pattern 4: Firms A and B determine the prices of their goods, and Firm C determines its output. From the above inverse demand functions, we obtain \begin{equation*} p_C=\frac{(1-b)a+2b^2x_C-bx_C-x_C+bp_A+bp_B}{1+b}, \end{equation*} \begin{equation*} x_B=\frac{(1-b)a+b^2x_C-bx_C+bp_A-p_B}{(1-b)(1+b)}, \end{equation*} and \begin{equation*} x_A=\frac{(1-b)a+b^2x_C-bx_C-p_A+bp_B}{(1-b)(1+b}. \end{equation*} \item Pattern 5: Firms A and C determine the prices of their goods, and Firm B determines its output. From the above inverse demand functions, we obtain \begin{equation*} p_B=\frac{(1-b)a+2b^2x_B-bx_B-x_B+bp_C+bp_A}{1+b}, \end{equation*} \begin{equation*} x_A=\frac{(1-b)a+b^2x_B-bx_B+bp_C-p_A}{(1-b)(1+b)}, \end{equation*} and \begin{equation*} x_C=\frac{(1-b)a+b^2x_B-bx_B-p_C+bp_A}{(1-b)(1+b}. \end{equation*} \item Pattern 6: All firms determine the prices of their goods. It is a Bertrand case. From the inverse demand functions, the direct demand functions are derived as follows; \begin{equation*} x_A=\frac{(1-b)a-(1+b)p_A+b(p_A+p_C)}{(1-b)(1+2b)}, \end{equation*} \begin{equation*} x_B=\frac{(1-b)a-(1+b)p_B+b(p_B+p_C)}{(1-b)(1+2b)}, \end{equation*} and \begin{equation*} x_C=\frac{(1-b)a-(1+b)p_C+b(p_A+p_B)}{(1-b)(1+2b)}. \end{equation*} \end{enumerate} The absolute profits of the firms are \begin{equation*} \pi_A=p_Ax_A-c_Ax_A, \end{equation*} \begin{equation*} \pi_B=p_Bx_B-c_Bx_B, \end{equation*} and \begin{equation*} \pi_C=p_Cx_C-c_Cx_C. \end{equation*} $c_A$, $c_B$ and $c_C$ are the constant marginal costs of Firms A, B and C. The relative profits of the firms are \begin{equation*} \psi_A=\pi_A-\frac{\pi_B+\pi_C}{2}, \end{equation*} \begin{equation*} \psi_B=\pi_B-\frac{\pi_A+\pi_C}{2}, \end{equation*} and \begin{equation*} \psi_C=\pi_C-\frac{\pi_A+\pi_B}{2}. \end{equation*} The firms determine the values of their strategic variables to maximize the relative profits. We see \begin{equation*} \psi_A+\psi_B+\psi_C=0, \end{equation*} so the game is zero-sum. We assume $c_A=c_B$, that is, the game is symmetric for Firms A and B. However, $c_C$ is not equal to $c_A$. Thus, the game is partially asymmetric. We calculate the equilibrium outputs of the firms in the above six patterns. \begin{enumerate} \item Pattern 1 \[x_A=\frac{bc_C-4c_A-ab+4a}{(4-b)(b+2)},\] \[x_B=\frac{bc_C-4c_A-ab+4a}{(4-b)(b+2)},\] \[x_C=\frac{bc_C+4c_C-2bc_A+ab-4a}{(4-b)(b+2)}.\] \item Pattern 2 \[x_A=\frac{bc_C-4c_A-ab+4a}{(4-b)(b+2)},\] \[x_B=\frac{bc_C-4c_A-ab+4a}{(4-b)(b+2)},\] \[x_C=\frac{bc_C+4c_C-2bc_A+ab-4a}{(b-4)(b+2)}.\] \item Pattern 3 \[x_A=\frac{5b^2c_C+4bc_C-3b^3c_A+6b^2c_A+4bc_A-16c_A+3ab^3-11ab^2-8ab+16a}{(4-b)(1-b)(b+2)(3b+4)},\] \[x_B=\frac{bc_C-4c_A-ab+4a}{(4-b)(b+2)},\] \[x_C=\frac{7b^2c_C-16c_C-3b^3c_A+4b^2c_A+8bc_A+3ab^3-11ab^2-8ab+16a}{(4-b)(1-b)(b+2)(3b+4)}.\] \item Pattern 4 \[x_A=\frac{2b^2c_C+bc_C+3b^2c_A-2bc_A-4c_A-5ab^2+ab+4a}{(1-b)(b+2)(5b+4))},\] \[x_B=\frac{2b^2c_C+bc_C+3b^2c_A-2bc_A-4c_A-5ab^2+ab+4a}{(1-b)(b+2)(5b+4))},\] \[x_C=\frac{b^2c_C-3bc_C-4c_C+4b^2c_A+2bc_A-5ab^2+ab+4a}{(1-b)(b+2)(5b+4))}.\] \item Pattern 5 \[x_A=\frac{3b^2c_C-b^3c_C+4bc_C+6b^3c_A+16b^2c_A-12bc_A-16c_A-5ab^3-19ab^2+8ab+16a}{(1-b)(b+2)(b+4)(5b+4)},\] \[x_B=\frac{2b^2c_C+bc_C+3b^2c_A-2bc_A-4c_A-5ab^2+ab+4a}{(1-b)(b+2)(5b+4)},\] \[x_C=\frac{4b^3c_C+7b^2c_C-16bc_C-16c_C+b^3c_A+12b^2c_A+8bc_A-5ab^3-19ab^2+8ab+16a}{(1-b)(b+2)(b+4)(5b+4)}.\] \item Pattern 6 \[x_A=\frac{2b^2c_C+bc_C+3b^2c_A-2bc_A-4c_A-5ab^2+ab+4a}{(1-b)(b+2)(5b+4)},\] \[x_B=\frac{2b^2c_C+bc_C+3b^2c_A-2bc_A-4c_A-5ab^2+ab+4a}{(1-b)(b+2)(5b+4)},\] \[x_C=\frac{b^2c_C-3bc_C-4c_C+4b^2c_A+2bc_A-5ab^2+ab+4a}{(1-b)(b+2)(5b+4)}.\] \end{enumerate} We find that Pattern 1 is equivalent to Pattern 2 (an example of Theorem \ref{t1}), but it is not equivalent to Pattern 3, and that Pattern 6 is equivalent to Pattern 4 (an example of Theorem \ref{t2}), but it is not equivalent to Pattern 5. Pattern 1 (Cournot Pattern) and Pattern 6 (Bertrand) are not equivalent unless we have $c_C=c_A$. \section{Concluding Remarks} In this paper we have examined equilibria in a partially asymmetric three-players zero-sum game under various situations. We want to extend the results of this paper to a general multi-players zero-sum game. \section*{Acknowledgment} This work was supported by Japan Society for the Promotion of Science KAKENHI Grant Number 15K03481 and 18K01594.

49acd4a898523cb109033a462e311b7f38ffb832

\section{Introduction} A smooth mapping $f$ between manifolds is said to be \emph{stable} if for any mapping $g$ sufficiently close to $f$ (with respect to the Whitney $C^\infty$--topology) one can take diffeomorphisms $\Phi$ and $\phi$ of the source and the target manifolds, respectively, so that they satisfy $\phi\circ f\circ \Phi = g$. Stable mappings play an important role not only in the theory of singularities of differentiable mappings, but also in recent studies on topology of smooth manifolds (see \cite{SaekiYamamoto} and \cite{GKtrisection}, for example). In spite of its simple and natural definition, it is in general difficult to check whether a given smooth mapping is stable or not (cf.~\cite[\S.29]{Whitney}). The first breakthrough in detecting stability is due to Mather \cite{MatherII,MatherV}. In his seminal work, Mather introduced the notion of \emph{infinitesimal stability} and show that infinitesimal stability implies stability for \emph{proper} smooth mappings (\cite{MatherII}, for the definition of infinitesimal stability, see Subsection~\ref{Se:stability maps}). Mather further introduced two other notions of stability which we call \emph{strong stability} and \emph{local stability}: $f$ is strongly stable if there exists a continuous mapping $g\mapsto (\Phi_g,\phi_g)\in \operatorname{Diff}(N)\times \operatorname{Diff}(P)$ defined on a neighborhood of $f$ such that $\phi_g\circ f\circ \Phi_g = g$ for any $g$ in the neighborhood, while local stability is a local version of infinitesimal stability (we will give the precise definition of local stability in Subsection~\ref{Se:stability maps}). It was then shown in \cite{MatherV} that all the four stabilities are equivalent for \emph{proper} smooth mappings. Unfortunately, as Mather already pointed out, the four stabilities above are not equivalent for \emph{non-proper} mappings. Whereas it is relatively easy to check local or infinitesimal stability of smooth mappings (see \cite[\S.4 and 5]{MatherV}), the problem of detecting (strong) stability is much harder. As far as the author knows, the only result concerning stability of non-proper mappings is due to Dimca \cite{Dimca}: he gave a necessary and sufficient condition for stability of Morse functions \emph{defined on $\mathbb{R}$} (see Theorem~\ref{T:Dimca stability R}). As for strong stability, du Plessis and Vosegaard \cite{duPlessisVosegaard} gave a necessary and sufficient condition for a smooth mapping to be strongly \emph{and} infinitesimally stable. The purpose of this paper is to give a sufficient condition for (strong) stability of non-proper functions. In order to explain the main result of the paper, we will give several definitions. An \emph{open neighborhood of the end of $N$} is an open subset $V$ of $N$ whose complement is compact. A smooth mapping $f:N\to P$ is said to be \emph{end-trivial at $y\in P$} if for any compact subset $K\subset N$, there exist an open neighborhood $W\subset P$ of $y$, and an open neighborhood $V$ of the end of $N$ whose complement contains $K$, such that they satisfy the following conditions: \begin{enumerate} \item the set $f^{-1}(y)\cap V$ does not contain a critical point of $f$, \item there exists a diffeomorphism $\Phi:(f^{-1}(y)\cap V)\times W\to f^{-1}(W)\cap V$ such that $f\circ \Phi$ is the projection onto the second component. \end{enumerate} We denote by $\tau (f)$ the set of points at which $f$ is end-trivial. For a smooth mapping $f:N\to P$, let $\Sigma(f)\subset N$ be the set of points $x\in N$ with $\operatorname{rank}(df_x) < p$~($=\dim P$) and $\Delta(f) =f(\Sigma(f))$. A mapping $f$ is said to be \textit{quasi-proper} if there exists a neighborhood $V\subset P$ of $\Delta(f)$ such that the restriction $f|_{f^{-1}(V)}:f^{-1}(V)\to V$ is proper. \begin{theorem}\label{T:suff condi stability} Let $N$ be a smooth manifold without boundaries and $f:N\to \mathbb{R}$ be a Morse function\footnote{A function $f:N\to \mathbb{R}$ is a \emph{Morse function} if $f|_{\Sigma(f)}$ is injective and $f$ has a non-degenerate Hessian at any point in $\Sigma(f)$. Note that $f$ is a Morse function if and only if $f$ is locally stable. }. \begin{enumerate} \item A function $f$ is stable if $\Delta(f)$ is contained in $\tau(f)$. \item A function $f$ is strongly stable if and only if $f$ is quasi-proper. \end{enumerate} \end{theorem} \noindent We will give two remarks on this theorem. First, it is easy to verify that Dimca's condition for $f:\mathbb{R}\to \mathbb{R}$ in Theorem~\ref{T:Dimca stability R} is equivalent to our sufficient condition $\Delta(f)\subset \tau(f)$ (we can indeed show that the complement $\mathbb{R}\setminus \tau(f)$ is equal to $Z(f|_{\Sigma(f)})\cup L(f)$ for a function $f$ on $\mathbb{R}$). In particular, our sufficient condition is also necessary for stability of functions on $\mathbb{R}$. Second, du Plessis and Vosegaard \cite{duPlessisVosegaard} already showed that quasi-properness is a necessary condition for strong stability of smooth mappings, and we will indeed prove that a quasi-proper Morse function is strongly stable. The paper is organized as follows. We will give several definitions in Section~\ref{Se:preliminaries}. Section~\ref{Se:proof main thm} is devoted to the proof of Theorem~\ref{T:suff condi stability}. In Section~\ref{Se:application} we will give two applications of Theorem~\ref{T:suff condi stability}. First, we will explicitly give an example of strongly stable but not infinitesimally stable function. (See Theorem~\ref{T:ex strongly but not inf stable func}. We indeed prove that the function $F(x)=\exp(-x^2)\sin x$ has desired properties.) As far as the author knows, we did not know any reasonable condition guaranteeing only strong stability (as we mentioned, du Plessis and Vosegaard \cite{duPlessisVosegaard} gave a necessary and sufficient condition for strong \emph{and} infinitesimal stability of smooth mappings), in particular we could not obtain such an example. The second application concerns stability of Nash (i.e.~semi-algebraic smooth) functions. We will prove that a locally stable Nash function on $\mathbb{R}^n$ satisfying some mild condition (on its gradient) is stable (see Theorem~\ref{T:stability Nash func}). As a corollary of this result, we will show that the following function is stable for any $k\in \{1,\ldots,n-1\}$ (Example~\ref{Ex:modelMorse stable}): \[ G_k(x_1,\ldots,x_n) = \sum_{i=1}^{k} x_i^2 -\sum_{j=k+1}^n x_j^2. \] Note that we can immediately deduce from Mather's result \cite{MatherV} that $G_0$ and $G_n$ are stable since these functions are proper (while $G_k$ for $k \in \{1,\ldots,n-1\}$ is not proper). Relying on the result on stability of Nash functions, together with the result in \cite{Ichiki}, we will also prove that any Nash function on $\mathbb{R}^n$ becomes stable after generic linear perturbation (Corollary~\ref{T:stable after linear perturbation}). \section{Preliminaries}\label{Se:preliminaries} Throughout the paper, we will assume that manifolds are smooth and do not have boundaries unless otherwise noted. For manifolds $N$ and $P$, we denote the set of smooth mappings from $N$ to $P$ by $C^\infty(N,P)$ and let $\operatorname{Diff}(N)\subset C^\infty(N,N)$ be the set of self-diffeomorphisms of $N$ and $C^\infty(N) = C^\infty(N,\mathbb{R})$. Let $X,Y$ be topological spaces. A mapping $f:X\to Y$ is \textit{proper} if the preimage $f^{-1}(K)$ of any compact subset $K\subset Y$ is compact. Note that properness of $f$ is equivalent to the condition $Z(f)=\emptyset$, where $Z(f)\subset Y$ is the set of \textit{improper points} of $f$, defined as follows: \[ Z(f) =\left\{y\in Y~\left|~y=\lim_{n\to \infty}f(x_n),\hspace{.3em}\{x_n\}:\begin{minipage}[c]{44mm} sequence of points in $X$ without accumulation points \end{minipage} \right.\right\}. \] A smooth mapping $g:N\to P$ between manifolds $N$ and $P$ is \textit{quasi-proper} if there exists a neighborhood $V\subset P$ of $\Delta(g)$ such that the restriction $g|_{g^{-1}(V)}:g^{-1}(V)\to V$ is proper. This condition is equivalent to the condition $Z(g) \cap \Delta(g) =\emptyset$ (see \cite[Corollary 3.2.15]{duPlessisWall}). \subsection{Whitney $C^k$--topology}\label{Se:Whitney topology} For a non-negative integer $k$, we denote the $k$--jet bundle with the source $N$ and the target $P$ by $J^k(N,P)$. For a smooth mapping $f\in C^\infty(N,P)$, let $j^kf:N\to J^k(N,P)$ be the $k$--jet extension of $f$. For an open subset $U\subset J^k(N,P)$, we define the subset $M(U)\subset C^\infty(N,P)$ as follows: \[ M(U) = \left\{f\in C^\infty(N,P)~|~j^kf(N)\subset U \right\}. \] It is easy to see that the family $\mathcal{M}_k=\{M(U)\subset C^\infty(N,P)~|~U\subset J^k(N,P):\mbox{open}\}$ forms a basis for a topology $\tau W^k$ of $C^\infty(N,P)$, which we call the \textit{Whitney $C^k$--topology}. We further define the \textit{Whitney $C^\infty$--topology} $\tau W^\infty$ as a topology with an open basis $\bigcup_{k\geq 0} \mathcal{M}_k$. In what follows we will explain a neighborhood basis of $\tau W^k$ due to Mather (\cite{MatherII}). For a smooth mapping $h:\mathbb{R}^n \to \mathbb{R}^p$, $x\in \mathbb{R}^n$ and a positive integer $k$, we define a linear mapping \[ D^kh(x):(\mathbb{R}^n)^{\otimes k} \to \mathbb{R}^p \] by giving a value of $e_{i_1}\otimes\cdots\otimes e_{i_k}\in (\mathbb{R}^n)^{\otimes k}$ (where $\{e_1,\ldots,e_n\}$ is the standard basis of $\mathbb{R}^n$) as follows: \[ D^kh(x)(e_{i_1}\otimes\cdots\otimes e_{i_k}) = \left(\frac{\partial^k h_1}{\partial x_{i_1}\cdots \partial x_{i_k}}(x),\ldots,\frac{\partial^k h_p}{\partial x_{i_1}\cdots \partial x_{i_k}}(x)\right). \] Using the operator norm $\norm{D^k h(x)}$ of $Dh^k(x)$, we define $\norm{h}_{k,x}$ and $ \norm{h}_{k,X}$ for $X\subset \mathbb{R}^n$ as follows: \[ \norm{h}_{k,x} = \norm{h(x)} + \sum_{j=1}^{k}\norm{D^jh(x)} \mbox{ and }\norm{h}_{k,X} = \sup_{x\in X} \norm{h}_{k,x}. \] Note that for a function $f:\mathbb{R}^n\to \mathbb{R}$, $\norm{D^1f(x)}$ is equal to $\displaystyle \norm{df_x} = \sqrt{\sum_{i=1}^{n}\left(\frac{\partial f}{\partial x_i}(x)\right)^2}$ and $\norm{D^2f(x)}$ is equal to $\displaystyle \sqrt{\sum_{i=1}^{k} m_i\lambda_i^2}$, where $\lambda_1,\ldots,\lambda_k$ are the eigenvalues of the Hessian of $f$ at $x$ and $m_i$ is the multiplicity of $\lambda_i$. For a smooth mapping $f\in C^\infty(N,P)$, we take systems of coordinate neighborhoods $\varphi=\{(U_\alpha,\varphi_\alpha)\}_{\alpha\in A}$ and $\psi=\{(V_\alpha \psi_\alpha)\}_{\alpha\in A}$ of $N$ and $P$, respectively, and a locally finite covering $\{K_\alpha\}_{\alpha \in A}$ of $N$ consisting of compact subsets so that $K_\alpha \subset U_\alpha$ and $f(U_\alpha)\subset V_\alpha$ for each $\alpha \in A$. Let $\{\varepsilon_\alpha\}_{\alpha\in A}$ be a system of positive numbers. For each $\alpha \in A$, we define a subset $N_k(f,K_\alpha,\varphi_\alpha,\psi_\alpha,\varepsilon_\alpha)\subset C^\infty(N,P)$ as follows: \[ N_k(f,K_\alpha,\varphi_\alpha,\psi_\alpha,\varepsilon_\alpha)=\left\{g\in C^\infty(N,P) ~\left|~\norm{\psi_\alpha\circ f\circ \varphi_\alpha^{-1}-\psi_\alpha\circ g\circ \varphi_\alpha^{-1}}_{k,\varphi_\alpha(K_\alpha)} < \varepsilon_\alpha\right. \right\}. \] We further define $N_k(f,K,\varphi,\psi,\varepsilon) = \bigcap_{\alpha\in A} N_k(f,K_\alpha,\varphi_\alpha,\psi_\alpha,\varepsilon_\alpha)$. \begin{theorem}[{\cite[\S.4, Lemma 1]{MatherII}}]\label{T:basis Whitney topology} For $k<\infty$ the system \[ \{N_k(f,K,\varphi,\psi,\varepsilon)\subset C^\infty(N,P)~|~\varepsilon = \{\varepsilon_\alpha\}_{\alpha\in A}:\mbox{system of positive numbers}\} \] is a neighborhood basis of $f\in C^\infty(N,P)$ with respect to the topology $\tau W^k$. \end{theorem} \begin{remark} For a system $\{V_\alpha\}_{\alpha\in A}$, where $V_\alpha\subset C^\infty(K_\alpha,P)$ is an open subset with respect to the topology $\tau W^k$, we define a subset $\cap_\alpha V_\alpha\subset C^\infty(N,P)$ as follows: \[ \cap_\alpha V_\alpha = \{f\in C^\infty(N,P)~|~ \forall \alpha \in A, \hspace{.3em} f|_{K_\alpha} \in V_\alpha\}. \] \noindent We can easily deduce from Theorem~\ref{T:basis Whitney topology} that the following system is a basis of $\tau W^k$ for $k<\infty$: \[ \mathcal{N}_k = \left\{\cap_\alpha V_\alpha\subset C^\infty(N,P) ~\left|~\{V_\alpha\}_{\alpha\in A}\mbox{~:~system of open subsets as above} \right.\right\}. \] We can also define a system $\mathcal{N}_\infty$ of subsets of $C^\infty(N,P)$ in a similar manner, yet it is \emph{not} a basis of the topology $\tau W^\infty$ but produces a stronger topology of $C^\infty(N,P)$, which is called the \emph{very strong topology} in \cite{duPlessisVosegaard}. \end{remark} \subsection{Stability of smooth mappings}\label{Se:stability maps} A smooth mapping $f\in C^\infty(N,P)$ is \textit{stable} if there exists an open neighborhood $\mathcal{U}\subset C^\infty(N,P)$ of $f$ (with respect to the topology $\tau W^\infty$) and (not necessarily continuous) mappings $\Theta:\mathcal{U}\to \operatorname{Diff}(N)$ and $\theta:\mathcal{U}\to \operatorname{Diff}(P)$ such that $\theta(g)\circ g\circ\Theta(g) = f$ for $g\in \mathcal{U}$. A smooth mapping $f\in C^\infty(N,P)$ is \textit{strongly stable} if we can further make $\Theta$ and $\theta$ above continuous (with respect to the topologies $\tau W^\infty$) and satisfy the conditions $\Theta(f)=\operatorname{id}_N$ and $\theta(f)=\operatorname{id}_P$. For a vector bundle $E$ on $N$, we denote the set of smooth sections of $E$ by $\Gamma(E)$, which is a $C^\infty(N)$--module. We define a $C^\infty(N)$--module homomorphism $tf: \Gamma(TN) \to \Gamma(f^\ast TP)$ and a $C^\infty(P)$--module homomorphism $\omega f:\Gamma(TP)\to \Gamma(f^\ast TP)$ as follows: \[ tf(\xi) = df(\xi)\mbox{ and }\omega f(\eta) = \eta \circ f. \] A smooth mapping $f\in C^\infty(N,P)$ is \textit{infinitesimally stable} if the following equality holds: \[ \Gamma(f^\ast TP) = tf(\Gamma(TN)) + \omega f (\Gamma(TP)). \] For a subset $S\subset N$ we denote the set of germs of sections of a vector bundle $E$ on $N$ at $S$ by $\Gamma(E)_S$. The homomorphism $tf$ (resp.~$\omega f$) induces a homomorphism from $\Gamma(TN)_S$ (resp.~$\Gamma(TP)_{f(S)}$) to $\Gamma(f^\ast TP)$ in the obvious way. A smooth mapping $f\in C^\infty(N,P)$ is \textit{locally stable} if the following equality holds for any $y\in \Delta(f)$ and $S\subset f^{-1}(y)$ with $\sharp(S)\leq p+1$: \[ \Gamma(f^\ast TP)_S = tf(\Gamma(TN)_S) + \omega f (\Gamma(TP)_y). \] As we noted in the introduction, all of the four stabilities above are equivalent for a proper mapping $f\in C^\infty(N,P)$. In the rest of the subsection we will briefly review known results on stabilities for general (non-proper) mappings. We can immediately deduce from the definitions that strong stability (resp.~infinitesimal stability) implies stability (resp.~local stability). Mather \cite{MatherV} showed that stability implies local stability and a smooth mapping $f\in C^\infty(N,P)$ is infinitesimally stable if and only if it is locally stable and $f|_{\Sigma(f)}$ is proper. It was shown in \cite{duPlessisVosegaard} that any strongly stable mapping is quasi-proper. Furthermore, Dimca \cite{Dimca} gave a necessary and sufficient condition for stability of a function $f\in C^\infty(\mathbb{R},\mathbb{R})$: \begin{theorem}[\cite{Dimca}]\label{T:Dimca stability R} A locally stable function $f\in C^\infty(\mathbb{R},\mathbb{R})$ is stable if and only if the intersection $\Delta(f) \cap (Z(f|_{\Sigma(f)})\cup L(f))$ is empty, where $L(f)$ is defined as follows: \[ L(f) = \left\{y\in \mathbb{R}~\left|~y = \lim_{x\to \pm \infty} f(x) \right.\right\}. \] \end{theorem} \section{A sufficient condition for stability of smooth functions}\label{Se:proof main thm} In this section we will first prove (1) of Theorem~\ref{T:suff condi stability}. In the proof, for a given function $f$ satisfying the assumption in (1), we will construct several diffeomorphisms so that the composition of them with a function close to $f$ coincides with $f$. We will then observe that the algorithm for constructing diffeomorphisms also guarantees strong stability of quasi-proper Morse functions. Suppose that a function $f:N\to \mathbb{R}$ is locally stable and $\Delta(f)\subset \tau(f)$. We put $\Sigma=\Sigma(f)$ and $\Delta=\Delta(f)$. Since $f$ is locally stable, $\Sigma\subset N$ is closed and discrete. The set $Z(f|_{\Sigma})$ is contained in $\mathbb{R}\setminus \tau(f)$. Thus we can deduce from the assumption that $\Delta\subset \mathbb{R}$ is discrete and countable. We put $\Delta = \{y_i\}_{i\in I}$, where $I\subset \mathbb{N}$, and take $x_i\in \Sigma$ so that $f(x_i)$ is equal to $y_i$. We also take a sequence $\{K_n\}_{n\in \mathbb{N}}$ of codimension $0$ compact submanifolds of $N$ satisfying the following conditions: \begin{itemize} \item $\displaystyle\bigcup_{n\in \mathbb{N}} K_n = N$, \item $K_k\subset \operatorname{Int}(K_{k+1})$ for any $k\in \mathbb{N}$, \item $x_i\in K_i$ for any $i\in I$. \end{itemize} Since $y_i$ is contained in $\tau(f)$ for each $i\in I$, we can take a neighborhood $V_i$ of the end of $N$ whose complement contains $K_i$, $\nu_i>0$ and a diffeomorphism \[ \Phi_i:(f^{-1}(y_i)\cap V_i)\times (y_i-\nu_i,y_i+\nu_i)\to f^{-1}(y_i-\nu_i,y_i+\nu_i)\cap V_i \] which satisfies $f\circ \Phi_i= p_2$. Without loss of generality we can assume that $f^{-1}(y_i-\nu_i,y_i+\nu_i)\cap V_i=\emptyset$ if $y_i$ is not contained in $Z(f)$. By taking sufficiently small $\nu_i$'s, we can assume that $\nu_i$ is less than $1/n$ and $(y_i-\nu_i,y_i+\nu_i)\cap (y_j-\nu_j,y_j+\nu_j)=\emptyset$ for $i\neq j$. Since the complement $N\setminus V_i$ is compact, $N\setminus V_i$ is contained in $K_{d(i)-1}$ for some $d(i)\in \mathbb{N}$, which is larger than $d(i)>i$. We take a coordinate neighborhood $(U_i,\varphi_i)$ of $x_i$ satisfying the following conditions: \begin{itemize} \item $f\circ\varphi_i^{-1}(w_1,\ldots,w_n) = \pm w_1^2\pm \cdots \pm w_n^2 + y_i$, \item $U_i\cap \overline{\left(f^{-1}((y_i-\nu_i,y_i+\nu_i))\cap V_i\right)} =\emptyset$. \end{itemize} We also take a system of coordinate neighborhoods $\{(U_\alpha,\varphi_\alpha)\}_{\alpha\in A}$ of $N$ so that it satisfies: \begin{enumerate}[(U1)] \item $U_\alpha$ is relatively compact, \item there exists $\alpha(i)\in A$ such that $U_{\alpha(i)}=U_i\cap f^{-1}((y_i-\nu_i,y_i+\nu_i))$ and $\varphi_{\alpha(i)}=\varphi_i$, \item either of the followings holds for any $\alpha\in A$: \begin{itemize} \item there exists $i \in I$ such that $U_\alpha \subset f^{-1}((y_i-\nu_i,y_i+\nu_i))$, or \item $U_\alpha \subset K_i - U_{\alpha(i)}$ for any $i\in I$ with $U_\alpha \cap f^{-1}((y_i-\nu_i/2,y_i+\nu_i/2))\neq \emptyset$, \end{itemize} \item if there exists $i \in I$ such that $U_\alpha \subset f^{-1}((y_i-\nu_i,y_i+\nu_i))$, either of the followings holds: \begin{itemize} \item $U_\alpha \subset K_{d(i)}$, or \item there exists an open set $W_\alpha\subset f^{-1}(y_i)\cap V_i$ and a chart $\phi_\alpha:W_\alpha\to \mathbb{R}^{n-1}$ on $W_\alpha$ such that $U_\alpha = \Phi_i^{-1}(W_\alpha \times (y_i-\nu_i,y_i+\nu_i))$ and $\varphi_\alpha = (\phi_\alpha\times \operatorname{id})\circ \Phi_i$, \end{itemize} \item for each $i\in I$, there exist only finitely many $\alpha$'s satisfying the condition $U_\alpha \cap N_i\neq \emptyset$, where $N_i=f^{-1}([y_i-\nu_i/2,y_i+\nu_i/2])\cap K_{d(i)+1}$ (note that $N_i$ is the empty set if $y_i$ is not contained in $Z(f)$). \end{enumerate} \noindent We denote the preimage $\varphi_i^{-1}(B(r))$ by $Q_i(r)$, where $B(r)\subset \mathbb{R}^n$ is the open $n$--ball with radius $r$ centered at the origin. By modifying $\{(U_\alpha,\varphi_\alpha)\}_{\alpha\in A}$ if necessary, we can take a locally finite covering $\{L_\alpha\}_{\alpha\in A}$ of $N$ which satisfying the following conditions: \begin{enumerate}[(L1)] \item $L_\alpha$ is compact, \item $L_\alpha\subset U_\alpha$, \item $L_{\alpha(i)} = \overline{Q_i(\nu_i)}$ ($\subset U_{\alpha(i)}$), \item $L_\alpha\cap Q_i(\nu_i/2)=\emptyset$ if $\alpha\neq \alpha(i)$. \end{enumerate} \noindent Let $\displaystyle U = \bigcup_{i\in I} U_{\alpha(i)}$ and $\mathcal{V}_0 = C^\infty(N)$. In what follows, we regard $L_\alpha$ as an open subset of $\mathbb{R}^n$ via $\varphi_\alpha$, and also regard a function on $L_\alpha$ as one on an open subset of $\mathbb{R}^n$ in the obvious way. We define a subspace $\mathcal{V}_k$~($k=1,2,3,4$) of $C^\infty(N)$ as follows: {\allowdisplaybreaks \begin{align*} \mathcal{V}_1 = &\{g\in C^\infty(N)~|~ g|_{N-U}=f|_{N-U}\}, \\ \mathcal{V}_2 = &\{g\in C^\infty(N)~|~\Delta(g) =\Delta\}, \\ \mathcal{V}_3 = &\{g\in C^\infty(N)~|~ \Delta(g) = \Delta,\hspace{.3em} \Sigma(g) = \Sigma\},\\ \mathcal{V}_4 = &\{f\}. \end{align*} } \noindent We will prove (1) of Theorem~\ref{T:suff condi stability} by showing the following Claim~$k$~($k=1,2,3,4$): \V{.8em} \noindent {\bf Claim~$\boldsymbol{k}$.} There exist an open neighborhood $\mathcal{U}_{k}\subset \mathcal{V}_{k-1}$ of $f$ and a continuous mapping $\theta_k:\mathcal{U}_k\to \mathcal{V}_k$ such that $\theta_k(f)=f$ and $g$ is $\mathcal{A}$--equivalent to $\theta_k(g)$ for any $g\in \mathcal{U}_k$. \V{.8em} \noindent Note that the statement (1) of Theorem~\ref{T:suff condi stability} immediately follows from Claims~$1,\ldots,4$ (any mapping $g\in \mathcal{U}=(\theta_3\circ \theta_2\circ \theta_1)^{-1}(\mathcal{U}_4)$ is $\mathcal{A}$--equivalent to $f$). \V{1em} \noindent \textit{Proof of Claim~1}. We will use the following lemma: \begin{lemma}[{\cite[Theorem 3.6.1]{duPlessisWall}}]\label{T:strongstabilitysubmersion} Let $f:N\to P$ be $C^r$--mapping ($1\leq r\leq \infty$) and $U\subset N$ an open neighborhood of $\Sigma(f)$. There exists an open neighborhood $\mathcal{U}\subset C^\infty(N,P)$ of $f$ with respect to the $C^1$--topology and a mapping $\beta:\mathcal{U}\to \operatorname{Diff}(N)$ which is continuous with respect to the $C^s$--topologies for any $s\in \{1,\ldots, r\}$ which satisfy the following conditions: \begin{itemize} \item $\beta(f)$ is equal to the identity $\operatorname{id}_N$, \item $g\circ \beta(g)|_{N-U} $ is equal to $f|_{N-U}$ for any $g\in \mathcal{U}$, \item $\beta(g)(x)=x$ for any $x\in N-U$ with $f(x)=g(x)$. \end{itemize} \end{lemma} \noindent By applying this lemma to $f$ and the open set $U$, we can take an open neighborhood $\mathcal{U}_1\subset C^\infty(N)$ of $f$ and a mapping $\beta:\mathcal{U}_1\to \operatorname{Diff}(N)$ so that: \begin{itemize} \item $\beta$ is continuous with respect to the $C^s$--topologies for any $s\in \{1,\ldots, \infty\}$, \item $\beta(f)$ is equal to $\operatorname{id}_N$, and \item $g\circ \beta(g)|_{N-U} = f|_{N-U}$ for any $g\in \mathcal{U}_1$. \end{itemize} \noindent We define a mapping $\theta_1:\mathcal{U}_1\to \mathcal{V}_1$ by $\theta(g)=g\circ \beta(g)$, Since any diffeomorphism is proper, $\theta_1$ is continuous (see \cite[\S.2, Proposition 1]{MatherII}). The mapping $\theta_1$ satisfies the desired conditions. \V{1em} \noindent \textit{Proof of Claim~2}. Let $\gamma_i = (\nu_i/4)^{4/\nu_i}$ ($\nu_i$ appears when we take a diffeomorphism $\Phi_i$). We can easily check that $\gamma_i$ is less than or equal to $\nu_i/4<1/n$. For any $\alpha \in A\setminus \{\alpha(i)\}_{i\in I}$, we define $m_\alpha$ as follows: \[ m_\alpha = \min_{x\in L_\alpha} \norm{D^1f(x)}. \] Here, as we mentioned, we regard the function $f|_{L_\alpha}$ as that on $\varphi_\alpha(L_\alpha)\subset \mathbb{R}^n$. Note that $m_\alpha$ is greater than $0$ as $f$ has no critical points in $L_\alpha$. Using $m_\alpha$, we next define $\mu_\alpha>0$ as follows: \[ \mu_\alpha = \begin{cases} \dfrac{\gamma_i}{4n} & (\exists i \mbox{ s.t. }\alpha = \alpha(i))\\[5pt] \dfrac{m_\alpha}{2} & (\mbox{otherwise}). \end{cases} \] Let $\mu = \{\mu_\alpha\}_{\alpha\in A}$ and $\displaystyle \mathcal{U}_2 = \mathcal{V}_1 \cap N_2(f,L,\varphi,\operatorname{id}_\mathbb{R},\mu)$, which is an open neighborhood of $f$. The following inequality holds for $g\in \mathcal{U}_2$ and $x\in L_\alpha$: \[ \norm{D^1g(x)} \geq \norm{D^1f(x)}- \norm{g- f}_{2,L_\alpha} > \begin{cases} 2 \norm{\varphi_{i}(x)} - \dfrac{\gamma_i}{4} & (\exists i \mbox{ s.t. }\alpha = \alpha(i)) \\ \dfrac{m_\alpha}{2} > 0 & (\mbox{otherwise}). \end{cases} \] We can deduce from this inequality that all the critical points of $g$ are contained in $\displaystyle \bigcup_{i\in I} \overline{Q_i (\gamma_i/8)}$. Since $\norm{g-f}_{2,L_{\alpha(i)}}$ is less than $\nu_i/n$, using Lemma~\ref{T:crit pt unique} we can verify that there exists exactly one critical point of $g$ in $Q_i(\nu_i)$($\supset\overline{Q_i(\gamma_i/8)}$) for each $i\in I$. We denote the critical point of $g$ in $\overline{Q_i(\gamma_i/8)}$ by $x_{i,g}$, and let $y_{i,g} = g(x_{i,g})$. The norm $|y_i-y_{i,g}|$ can be estimated as follows: {\allowdisplaybreaks \begin{align*} |y_i-y_{i,g}| & = |f(x_i)-g(x_{i,g})|\\ & \leq |f(x_i)-f(x_{i,g})| + |f(x_{i,g})-g(x_{i,g})| \\ & < \left(\max_{x\in \overline{Q_i(\gamma_i/8)}}\norm{df_x} \right) \cdot |x_i-x_{i,g}| + \frac{\gamma_i}{4} < \frac{\gamma_i}{2}. \end{align*} } We first construct a diffeomorphism $\psi_g:\mathbb{R}\to \mathbb{R}$ such that $\Delta(\psi_g^{-1}\circ g)$ is equal to $\Delta$. Although the construction below is same as that in \cite{Dimca}, we will briefly explain the construction for completeness of this manuscript. We take a smooth function $\rho:\mathbb{R}\to \mathbb{R}$ so that: \begin{itemize} \item $\rho(t)= \rho(-t)$, \item $\rho|_{[0,\infty)}$ is monotone decreasing, \item the value $\rho(y)$ is equal to $1$ if $|y|<1$, and equal to $0$ if $|y|>2$, and \item $\lvert\rho'(y)\rvert<3/2$ for any $y\in \mathbb{R}$. \end{itemize} \noindent Using the function $\rho$, we define a function $h_i:\mathbb{R} \to \mathbb{R}$ as follows: \[ h_i(y) = \begin{cases} y + (y_{i,g}-y_i)\rho\left(\dfrac{4(y- y_i)}{\nu_i}\right) & (y\in (y_i-\nu_i, y_i+\nu_i)) \\ y & (y\not\in (y_i-\nu_i, y_i+\nu_i)). \end{cases} \] We further define a mapping $\psi_g:\mathbb{R}\to \mathbb{R}$ as follows: \[ \psi_g(y) = \begin{cases} h_i(y) & (y\in (y_i-\nu_i,y_i+\nu_i)) \\ y & (y \not\in \bigcup_{i\in I} (y_i-\nu_i,y_i+\nu_i)). \end{cases} \] As was shown in \cite{Dimca}, $\psi_g$ is a diffeomorphism and $\Delta(\psi_g^{-1}\circ g)=\Delta$. For each $i\in I$, we take a function $\eta_i:N\to[0,1]$ so that $\eta_i|_{K_{d(i)}} \equiv 0$ and $\eta_i|_{N-K_{d(i)+1}} \equiv 1$. Using $\eta_i$ we define a mapping \[ \widetilde{\eta}_{i,g}:(f^{-1}(y_i)\cap V_i)\times (y_i-\nu_i,y_i+\nu_i) \to (f^{-1}(y_i)\cap V_i)\times (y_i-\nu_i,y_i+\nu_i) \] as follows: \[ \widetilde{\eta}_{i,g}(x,y) = \left(x, y + (y_{i,g}-y_i)\rho\left(\frac{4(y-y_i)}{\nu_i}\right)\eta_i(x)\right). \] It is easy to see that $\widetilde{\eta}_{i,g}$ is a diffeomorphism. Note that $\widetilde{\eta}_{i,g}$ is the identity on $(f^{-1}(y_i)\cap (V_i\cap K_{d(i)}))\times (y_i-\nu_i,y_i+\nu_i)$ and equal to $\operatorname{id}\times \psi_g$ on $(f^{-1}(y_i)\cap K_{d(i)+1}^c)\times (y_i-\nu_i,y_i+\nu_i)$. We then define $\Psi_g^1:N\to N$ as follows: \[ \Psi_g^1(x) = \begin{cases} \Phi_i \circ \widetilde{\eta}_{i,g} \circ \Phi_i^{-1} (x) & \left(x\in f^{-1}((y_i-\nu_i,y_i+\nu_i))\cap V_i\right)\\ x & \left(x\not\in \bigcup_{i\in I} (f^{-1}((y_i-\nu_i,y_i+\nu_i))\cap V_i)\right). \end{cases} \] Note that if $f$ is quasi-proper, $f^{-1}((y_i-\nu_i,y_i+\nu_i))\cap V_i$ is empty for each $i\in I$, in particular $\Psi_g^1$ is the identity. We can easily verify that $\Psi_g^1$ is a bijection. The boundary $\partial (f^{-1}((y_i-\nu_i,y_i+\nu_i))\cap V_i)$ is contained in $K_{d(i)}\cup f^{-1}(\{y_i-\nu_i,y_i+\nu_i\})$. Since the k $\Phi_i\circ \widetilde{\eta}_{i,g}\circ \Phi_i^{-1}$ is the identity on a neighborhood of $f^{-1}((y_i-\nu_i,y_i+\nu_i))\cap V_i\cap (K_{d(i)}\cup f^{-1}(\{y_i-\nu_i,y_i+\nu_i\}))$, $\Psi_g^1$ is $C^\infty$ and its differential $d\Psi_g^1$ is regular on $f^{-1}((y_i-\nu_i,y_i+\nu_i))$. It is easy to see that $\Psi_g^1$ is the identity on $\displaystyle N-\left(\bigcup_{i\in I}f^{-1}((y_i-\nu_i,y_i+\nu_i))\right)$. Thus it is $C^\infty$ with regular differential on $\displaystyle N-\overline{\left(\bigcup_{i\in I}f^{-1}((y_i-\nu_i,y_i+\nu_i))\right)}$. In what follows we will check that $\Psi_g^1$ is $C^\infty$ at $x\in f^{-1}(y)$, where $\displaystyle y\in \bigcup_{i\in I}\{y_i-\nu_i,y_i+\nu_i\}$. We take a sufficiently small neighborhood of $y$ and let $\{y_{l(j)}\}_{j\in \mathbb{N}}$~($l(j)\in I$) be the critical values of $f$ contained in this neighborhood. Suppose that the sequence $\{y_{l(j)}\}_{j\in \mathbb{N}}$ is monotone increasing. We take $l\in \mathbb{N}$ so that $x$ is contained in $K_l$. Since $K_l$ is compact, there exists $j_0\in \mathbb{N}$ such that $x_{l(j)}$ is not contained in $K_l$ for any $j>j_0$. Since $x_{l(j)}$ is contained in $K_{l(j)}\subset N-V_{l(j)}$, $K_l\cap V_{l(j)}$ is empty for $j>j_0$. Thus there exists a neighborhood of $x$ in $K_l$ on which $\Psi_g^1$ is the identity. This shows that $\Psi_g^1$ is $C^\infty$ at $x$, and moreover its differential at $x$ is regular. We can eventually conclude that $\Psi_g^1$ is a diffeomorphism. Note that $\psi_g^{-1}\circ g \circ \Psi_g^1$ is contained in $\mathcal{V}_2$ for $g\in \mathcal{U}_2$. \begin{lemma}\label{T:continuity theta2} The mapping $\theta_2:\mathcal{U}_2 \to \mathcal{V}_2$ defined as $\theta_2(g) = \psi_g^{-1}\circ g \circ \Psi_g^1$ is continuous (with respect to the $C^s$--topologies for $2\leq s \leq \infty$). \end{lemma} \begin{remark}\label{Re:case quasi-proper} Since $f$ is not necessarily proper, the mapping $g\mapsto \psi_g^{-1}\circ g$ is not continuous in general. We can also check that the mapping $g\mapsto\Psi_g^1$ is not continuous in general. However, as we will show, the mapping $g\mapsto \psi_g$ is continuous. Furthermore, if $f$ is quasi-proper, the mapping $g\mapsto \psi_g^{-1}\circ g$ is continuous since $\Psi_g^1$ is the identity. \end{remark} \begin{proof}[Proof of Lemma~\ref{T:continuity theta2}] Throughout the proof, we will assume that any mapping space is endowed with the $C^s$--topology. We can easily check that $\psi_g^{-1}\circ g \circ \Psi_g^1$ is equal to $g$ on $\displaystyle N-\bigcup_{i\in I} f^{-1}((y_i-\nu_i/2,y_i+\nu_i/2))$ and $\displaystyle \bigcup_{i\in I}\left( f^{-1}((y_i-\nu_i/2,y_i+\nu_i/2))\cap K_{d(i)+1}^c\right)$, in particular on the complement $\displaystyle N\setminus \bigcup_{i\in I} N_i$, where $N_i$ is defined in the condition (U5) on $\{U_\alpha\}_{\alpha\in A}$. We first show the following two claims: \begin{enumerate} \item the mapping $\theta_2':\mathcal{U}_2\to C^\infty(N_i,N)$ defined as $\theta_2'(g)=\Psi_g^1|_{N_i}$ is continuous, and \item the mapping $\theta_2'':\mathcal{U}_2 \to C^\infty(\mathbb{R})$ defined as $\theta_2''(g)=\psi_g$ is continuous. \end{enumerate} Let $g\in \mathcal{U}_2$ and $\{\varepsilon_\alpha\}_{\alpha \in A}$ be a system of positive numbers. We can deduce from the conditions (U5) and (L2) that there exist only finitely many $\alpha$'s satisfying the condition $L_\alpha \cap N_i\neq\emptyset$, in particular we can take $\kappa_i = \min \left\{\varepsilon_\alpha ~|~ L_\alpha \cap N_i\neq\emptyset\right\}$. We can further deduce from the conditions (U3), (U4) and (L2) that $\Psi_g^1(L_\alpha)$ is contained in $ U_\alpha$ for any $\alpha$, and either of the followings hold for $\alpha$ with $L_\alpha\cap N_i\neq \emptyset$: \begin{itemize} \item $L_\alpha\cap N_i$ is contained in $K_{d(i)}$, \item $L_\alpha$ is contained in $V_i$. \end{itemize} \noindent Suppose that $L_\alpha\cap N_i\neq \emptyset$ and $L_\alpha$ is contained in $V_i$. By the condition (U4), we can obtain the following estimate for $h\in \mathcal{U}_2$: {\allowdisplaybreaks \begin{align*} & \norm{\varphi_\alpha\circ \Psi_h^1|_{N_i}\circ \varphi_\alpha^{-1}-\varphi_\alpha\circ \Psi_g^1|_{N_i}\circ \varphi_\alpha^{-1}}_{s,\varphi_\alpha(L_\alpha)} \\ = & \norm{(\phi_\alpha\times \operatorname{id} )\circ \tilde{\eta}_{i,h}\circ (\phi_\alpha\times \operatorname{id} )^{-1}-(\phi_\alpha\times \operatorname{id} )\circ \tilde{\eta}_{i,g}\circ (\phi_\alpha\times \operatorname{id} )^{-1}}_{s,(\phi_\alpha\times \operatorname{id} )\circ \Phi_i^{-1}(L_\alpha)} \\ \leq &\left| y_{i,h}-y_{i,g} \right| \cdot M_{i,s}, \end{align*} } where $M_{i,s}$ is a constant depending only on $i$ and $s$. We take a system $\{\delta_\alpha\}_{\alpha \in A}$ of positive numbers as follows: \[ \delta_\alpha = \begin{cases} \min\left\{\dfrac{\nu_i}{n},\dfrac{\kappa_i}{M_{i,s} \left(\sqrt{n}\norm{g}_{1,L_{\alpha(i)}} + 1\right)}\right\} & (\alpha = \alpha(i)) \\ 1 & (\mbox{otherwise}). \end{cases} \] By Lemma~\ref{T:evaluation critical point/value}, $\left|y_{i,h}-y_{i,g}\right|$ is less than $\left(\sqrt{n}\norm{g}_{1,L_{\alpha(i)}} + 1\right)\delta_{\alpha(i)}$ for any $h\in N_s(g,L,\varphi,\operatorname{id}_\mathbb{R},\delta)$. Thus we obtain: \[ \norm{\varphi_\alpha\circ \Psi_h^1|_{N_i}\circ \varphi_\alpha^{-1}-\varphi_\alpha\circ \Psi_g^1|_{N_i}\circ \varphi_\alpha^{-1}}_{s,\varphi_\alpha(L_\alpha)} < M_{i,s} \left(\sqrt{n}\norm{g}_{1,L_{\alpha(i)}} + 1\right)\delta_{\alpha(i)} < \varepsilon_\alpha. \] In particular, the diffeomorphism $\Psi_h^1|_{N_i}$ is contained in $N_s(\Psi_g^1,L_\alpha,\varphi_\alpha,\varphi_\alpha,\varepsilon_\alpha)$. Suppose that $L_\alpha\cap N_i$ is contained in $K_{d(i)}$. Since $\Psi_h^1$ is the identity on $N_i\cap K_{d(i)}$ for any $h\in N_s(g,L,\varphi,\operatorname{id}_\mathbb{R},\delta)$, $\theta_2'(h)$ is contained in $N_s(\Psi_g^1|_{N_i},L_\alpha,\varphi_\alpha,\varphi_\alpha,\varepsilon_\alpha)=N_s(\operatorname{id},L_\alpha,\varphi_\alpha,\varphi_\alpha,\varepsilon_\alpha)$. We can eventually conclude that $\theta_2'(N_s(g,L,\varphi,\operatorname{id},\delta))$ is contained in $ N_s(\Psi_g^1|_{N_i},L,\varphi,\varphi,\varepsilon)$. We can thus deduce from Lemma~\ref{T:basis Whitney topology} that $\theta_2'$ is continuous. We next prove continuity of $\theta_2''$. For a function $g\in \mathcal{U}_2$, we take a locally finite covering $L'=\{[r_\beta,s_\beta]\}_{\beta\in B}$ of $\mathbb{R}$ consisting of compact subsets so that for each $i\in I$, there exist only finitely many $\beta$'s satisfying the condition $(y_i-\nu_i,y_i+\nu_i)\cap [r_\beta,s_\beta] \neq \emptyset$. For a system $\{\varepsilon_\beta\}_{\beta\in B}$ of positive integers, we take a positive number $\kappa_i$ as follows: \[ \kappa_i = \min\left\{\varepsilon_\beta>0 ~\left|\right. (y_i-\nu_i,y_i+\nu_i)\cap [r_\beta,s_\beta] \neq \emptyset\right\}. \] We can take a constant $M_{i,s}'>0$, depending only on $i$ and $s$, such that the following holds for any $h\in \mathcal{U}_2$: \[ \norm{\psi_h-\psi_g}_{s,[r_\beta,s_\beta]} \leq \sup_{(y_i-\nu_i,y_i+\nu_i)\cap [r_\beta,s_\beta] \neq \emptyset} \left|y_{i,h} - y_{i,g}\right| M_{i,s}'. \] We take a system $\{\delta_\alpha\}_{\alpha\in A}$ of positive numbers as follows: \[ \delta_\alpha = \begin{cases} \min \left\{\dfrac{\nu_i}{n}, \dfrac{\kappa_i}{M_{i,s}'\left(\sqrt{n}\norm{g}_{1,L_{\alpha(i)}} + 1\right)}\right\} & (\alpha = \alpha(i)) \\ 1 & (\mbox{otherwise}). \end{cases} \] Using Lemma~\ref{T:evaluation critical point/value}, we can obtain the following inequality for $h \in N_s(g,L,\varphi,\psi,\delta)$: \[ \norm{\psi_h-\psi_g}_{s,[r_\beta,s_\beta]} < \kappa_i \leq \varepsilon_\beta. \] We can deduce from this inequality that $\theta_2''(N_s(g,L,\varphi,\psi,\delta))$ is contained in $N_s(\psi_g,L',\operatorname{id},\operatorname{id},\varepsilon)$. Thus, the mapping $\theta_2''$ is continuous. Since the mappings $\theta_2'$ and $\theta_2''$ are continuous, the following mapping is also continuous: \[ \tau_i:\mathcal{U}_2\to C^\infty(N_i),\hspace{.5em} g \mapsto \psi_g^{-1}\circ g\circ \Psi_g^1|_{N_i}. \] Thus, for any $g\in \mathcal{U}_2$ and any system $\{\varepsilon_\alpha\}_{\alpha\in A}$ of positive numbers, we can take a system $\{\delta_\alpha^i\}_{\alpha\in A}$ so that $\tau_i(\mathcal{U}_2\cap N_s(g,L,\varphi,\operatorname{id},\delta^i))$ is contained in $N_s(\tau_i(g),L,\varphi,\operatorname{id},\varepsilon/2)$. By the definitions of $\psi_h$ and $\Psi_h^1$, the function $\tau_i(h)$ depends only on $y_{i,h}$. In particular, $\tau_i(\mathcal{U}_2\cap N_k(g,L,\varphi,\operatorname{id},\gamma)) $ is in $ N_s(\tau_i(g),L,\varphi,\operatorname{id},\varepsilon)$ for any system $\{\gamma_\alpha\}_{\alpha\in A}$ with $\gamma_{\alpha(i)}=\delta^i_{\alpha(i)}$ ($i\in I$). We take a system $\{\tilde{\delta}_\alpha\}_\alpha$ as follows: \[ \tilde{\delta}_\alpha = \begin{cases} \delta^i_{\alpha(i)} & (\exists i \mbox{ s.t. }\alpha = \alpha(i)) \\ 1 & (\mbox{otherwise}). \end{cases} \] For any $h \in \mathcal{U}_2\cap N_s(g,L,\varphi,\operatorname{id},\tilde{\delta})$, the support $\displaystyle L_\alpha \cap \operatorname{supp}(\theta_2(g)-\theta_2(h))$ is contained in $\bigcup_{i\in I}(N_i \cap L_\alpha)$, in particular the norm $\norm{\theta_2(h)- \theta_2(g)}_{s,L_\alpha}$ can be estimated as follows: {\allowdisplaybreaks \begin{align*} \norm{\theta_2(h)- \theta_2(g)}_{s,L_\alpha} =& \norm{\theta_2(h)- \theta_2(g)}_{s,\cup_{i\in I} (N_i\cap L_\alpha)}\\ \leq & \sup_{i\in I}\norm{\theta_2(h)- \theta_2(g)}_{s,N_i\cap L_\alpha}\\ =& \sup_{i\in I}\norm{\tau_i(h)- \tau_i(g)}_{s,N_i\cap L_\alpha} \leq \varepsilon_\alpha/2 <\varepsilon_\alpha. \end{align*} } We thus obtain: \[ \theta_2(N_s(g,L,\varphi,\operatorname{id},\tilde{\delta}))\subset N_s(\theta_2(g),L,\varphi,\operatorname{id},\varepsilon) \] This completes the proof of continuity of $\theta_2$. \end{proof} \V{1em} \noindent \textit{Proof of Claim~3}. We will use the following lemma in the proof: \begin{lemma}[{\cite[\S.7, Lemma 2]{MatherII}}]\label{T:continuity integral vectfield} Let $U$ be a manifold without boundaries and $\pi_U:U\times I\to U$ be the projection. There exists an open neighborhood $\mathcal{O}_U \subset \Gamma(\pi_U^\ast TU)$ of the zero-section (with respect to the topology $\tau W^0$) such that we can take a mapping \[ \theta:\mathcal{O}_U\to C^\infty(U\times I ,U) \] by taking a flow of a vector field in $\mathcal{O}_U$ and $\theta$ is continuous with respect to the topologies $\tau W^s$ (for $0\leq s \leq \infty$). \end{lemma} Let $\mu = \{\mu_\alpha\}_{\alpha\in A}$ be the system we took in the proof of Claim 2 and $\displaystyle \mathcal{U}_3' = \mathcal{V}_2 \cap N_2(f,L,\varphi,\operatorname{id}_\mathbb{R},\mu)$, which is an open neighborhood of $f$ in $\mathcal{V}_2$. By Lemma~\ref{T:evaluation critical point/value}, for any $g\in \mathcal{U}_3'$ and $i\in I$ there exists exactly one critical point $x_{i,g}$ of $g$ in $\overline{Q_i(\gamma_i/8)}$. We define a vector field $X_g$ on $N$ as follows: \[ X_g(w) = \begin{cases} \lambda(w)(\varphi_i(x_{i,g}) - \varphi_i(x_i)) & (w\in \overline{Q_i(\gamma_i/4)})\\ 0 & (\mbox{otherwise}), \end{cases} \] where $\lambda:N\to [0,1]$ is a non-negative valued function satisfying the conditions $\lambda|_{\overline{Q_i(\gamma_i/8)}}\equiv 1$ and $\lambda|_{N-\overline{Q_i(\gamma_i/4)}}\equiv 0$. (Note that we identify a vector field on $U_{\alpha(i)}$($\supset\overline{Q_i(\gamma_i/4)}$) with an element in $C^\infty(U_{\alpha(i)})^n$ via the chart $\varphi_i$.) We define a mapping $\chi:\mathcal{U}_3'\to \Gamma(\pi_N^\ast TN)$ by $\chi(g) =X_g\circ \pi_N$. By Lemma~\ref{T:continuity integral vectfield} we can take an open neighborhood $\mathcal{O}_N\subset \Gamma(\pi_N^\ast TN)$ of the zero-section so that the mapping $\theta: \mathcal{O}_N\to C^\infty(N\times I ,N)$ defined by taking a flow is continuous. Let $\mathcal{U}_3 = \chi^{-1}(\mathcal{O}_N)$ and $\Psi_g^2=\theta(\chi(g))$ for $g\in \mathcal{U}_3$. We define a mapping $\theta_3:\mathcal{U}_3 \to \mathcal{V}_3$ as $\theta_3(g) = g\circ \Psi_g^2$. \begin{lemma}\label{T:continuity theta3} The set $\mathcal{U}_3$ is an open neighborhood of $f$ in $\mathcal{V}_2$ and $\theta_3$ is continuous (with respect to the topologies $\tau W^s$ for $2\leq s \leq \infty$). \end{lemma} \begin{proof} It is enough to show that $\chi$ is continuous. The systems $\{(U_\alpha\times I, \varphi_\alpha\times \operatorname{id})\}_{\alpha\in A}$ and $\{(\pi_{TN}^{-1}(U_\alpha), d\varphi_\alpha)\}_{\alpha\in A}$ are the coordinate neighborhood systems of $N\times I$ and $TN$, respectively (here we identify $T\mathbb{R}^n$ with $\mathbb{R}^{2n}$ in the obvious way). Note that $\{L_\alpha\times I\}_{\alpha\in A}$ is a locally finite covering of $N\times I$ consisting of compact subsets, and $L_\alpha\times I$ is contained in $U_\alpha\times I$ for any $\alpha\in A$. The following holds for any $g, h \in \mathcal{U}_3'$: \[ \norm{d\varphi_{\alpha(i)} \circ (\chi(h)-\chi(g)) \circ (\varphi_{\alpha(i)}\times \operatorname{id})^{-1})}_{s,L_{\alpha(i)}\times I} =\left|x_{i,h} - x_{i,g}\right| \cdot \norm{\lambda\circ \varphi_i^{-1}}_{s,L_{\alpha(i)}}. \] The norm $\norm{\lambda\circ \varphi_i^{-1}}_{s,L_{\alpha(i)}}$ depends only on $s$ and $i$. Thus, using Lemma~\ref{T:evaluation critical point/value}, for any system $\varepsilon=\{\varepsilon_\alpha\}_{\alpha\in A}$ of positive integers, we can take a system $\delta$ such that $\chi (N_k(g,L,\varphi,\operatorname{id},\delta))$ is contained in $N_k(X_g,L\times I,\varphi\times \operatorname{id}, d\varphi, \varepsilon)$. Hence $\chi$ is continuous. \end{proof} \V{1em} \noindent \textit{Proof of Claim~4}. Here, we will show that there exists an open neighborhood $\mathcal{U}_4\subset \mathcal{V}_3$ of $f$ such that any $g\in \mathcal{U}_4$ is $\mathcal{A}$--equivalent to $f$. For a vector bundle $E$ over $N$, we define a $C^\infty(N)$--module $\Gamma_k(E)$ ($k\geq 0$) as follows (recall that $\Sigma$ is the critical point set of $f$): \[ \Gamma_k(E)=\{\xi\in \Gamma(E)~|~ j^k\xi|_{\Sigma}=0\}. \] Let $\pi_N=\pi:N\times I \to N$ be the projection and $\Gamma_k(\pi^\ast E)=\{\xi \in \Gamma(\pi^\ast E)~|~ j^k\xi|_{\Sigma\times I}=0\}$ for $k\geq 0$, which is a $C^\infty(N\times I)$--module. We regard $\Gamma_k(E)$ as a subset of $\Gamma_k(\pi^\ast E)$ via the injection $\pi^\ast:\Gamma_k(E)\to \Gamma_k(\pi^\ast E)$ Universality of tensor products yields a homomorphism \[ \iota_k:\Gamma_k(E)\otimes_{C^\infty(N)} C^\infty(N\times I) \to \Gamma_k(\pi^\ast E). \] \begin{lemma}\label{T:surj iota_i} The homomorphism $\iota_k$ is surjective. \end{lemma} \begin{proof} We take a partition of the unity $\{\varrho_i~|~i\in I\}\cup\{\varrho_V\}$ of $N$ so that they satisfy the following conditions: \begin{itemize} \item $\displaystyle \varrho_i \equiv 1$ in $\overline{Q_i(\nu_i/2)}$ and $\displaystyle \varrho_i \equiv 0$ outside $\overline{Q_i(\nu_i)}=L_{\alpha(i)}$, \item $\sqrt[n]{\varrho_i}$ is $C^\infty$ for any $n\in \mathbb{Z}_{\geq 0}$. \end{itemize} We denote the function $p_l\circ \varphi_i: L_{\alpha(i)}\to \mathbb{R}$ by $y_l^i$, where $p_l:\mathbb{R}^n\to \mathbb{R}$ be the projection onto the $l$--th component. We first prove Lemma~\ref{T:surj iota_i} under the assumption that $E$ is a trivial line bundle. In this case we can identify $\Gamma_k(\pi^\ast E)$ with the module $\{h\in C^\infty(N\times I)~|~ j^kh|_{\Sigma\times I}=0\}$. Let $h$ be a function in this module. Since $j^k\varrho_V|_\Sigma=0$ for any $k\geq 0$, the function $\varrho_V h$ is contained in the image of $\iota_k$. The support of the function $\varrho_i h$ is contained in $\overline{Q_i(\nu_i)}\times I$, so we can regard this function as that on $\mathbb{R}^n\times I$. Since the $k$--jet $j^k \varrho_i h|_{\Sigma\times I}$ vanishes, we can decompose this function into the sum of other functions as follows: {\allowdisplaybreaks \begin{align*} (\varrho_i h)(x,t) &= (\sqrt[k+1]{\varrho_i}^{k+1}h)(x,t)\\ &=\left(\sqrt[k+1]{\varrho_i}(x)\right)^{k+1}\int_{0}^1\frac{d}{ds}\left(h(sx,t)\right)ds \\ &= \left(\sqrt[k+1]{\varrho_i}(x)\right)^{k+1}\sum_{l_1=1}^n y_{l_1}^i(x)h_{l_1}^i(x,t) \hspace{.3em}\left(\mbox{where }h_{l_1}^i(x,t)=\int_{0}^1\frac{\partial}{\partial x_{l_1}}\left(h(sx,t)\right)ds\right)\\ &= \left(\sqrt[k+1]{\varrho_i}(x)\right)^{k+1}\sum_{l_1=1}^n y_{l_1}^i(x)\int_0^1 \frac{d}{ds}\left(h_{l_1}^i(sx,t)\right)ds \\ &=\cdots = \sum_{l_1,\ldots,l_k=1}^n (\sqrt[k+1]{\varrho_i}y_{l_1}^i)(x)\cdots (\sqrt[k+1]{\varrho_i}y_{l_k}^i)(x)(\sqrt[k+1]{\varrho_i}h_{l_1,\ldots,l_k}^i)(x,t). \end{align*} } Since the support of $\sqrt[k+1]{\varrho_i}$ are contained in $\overline{Q_i(\nu(i))}\times I$, we can extend the functions $\sqrt[k+1]{\varrho_i}y_{l_j}^i$ and $\sqrt[k+1]{\varrho_i}h_{l_1,\ldots,l_k}^i$ to those on $N\times I$, which we denote by the same symbols. We then obtain: {\allowdisplaybreaks \begin{align*} h &= \varrho_V h +\sum_{i\in I}\varrho_i h \\ &= \varrho_V h + \sum_{l_1,\ldots,l_k=1}^n \left(\sum_{i\in I}\sqrt[k+1]{\varrho_i}y_{l_1}^i\cdots \sqrt[k+1]{\varrho_i}y_{l_k}^i\right) \cdot\left(\sum_{i\in I}\sqrt[k+1]{\varrho_i}h_{l_1,\ldots,l_k}^i\right)\in \operatorname{Im} (\iota_k). \end{align*} } Thus, the homomorphism $\iota_k$ is surjective. We next show the lemma for a general vector bundle $E$. If $E$ admits a direct-sum decomposition $E= E_1\oplus E_2$, the homomorphism $\iota_k$ for $E$ is surjective if and only if $\iota_k$ for $E_1$ and $E_2$ are both surjective. Any vector bundle $E$ is a direct-sum summand of a trivial bundle, which is a direct-sum of trivial line bundles. Since we have already shown the lemma for a trivial line bundle, the statement for $E$ also holds. \end{proof} Although the homomorphism $tf : \Gamma(TN) \to \Gamma(f^\ast T\mathbb{R})$ is \textit{not} surjective, as its image is contained in $\Gamma_0(f^\ast T\mathbb{R})$, we can show the following lemma in the same way as that in the proof of \cite[Ch.~III, Proposition 2.2]{GG}: \begin{lemma}\label{T:surj tf for Morse} The mapping $tf : \Gamma_{k-1}(TN) \to \Gamma_k(f^\ast T\mathbb{R})$ is surjective for any $k\geq 1$. \end{lemma} \noindent We define $t'f:\Gamma_{k-1}(\pi^\ast TN) \to \Gamma_{k}(\pi^\ast f^\ast T\mathbb{R})$ as in \cite[\S.7]{MatherII}. \begin{lemma}\label{T:surj t'f} The mapping $t'f$ is surjective. \end{lemma} \begin{proof} We can easily verify that the following diagram commutes: \[ \begin{CD} \Gamma_{k-1}(TN)\otimes_{C^\infty(N)} C^\infty(N\times I) @>\iota_{k-1}>> \Gamma_{k-1}(\pi^\ast TN) \\ @Vtf\otimes \operatorname{id} VV @V t'f VV \\ \Gamma_{k}(f^\ast T\mathbb{R})\otimes_{C^\infty(N)} C^\infty(N\times I) @>\tilde{\iota}_k>> \Gamma_{k}(\pi^\ast f^\ast T\mathbb{R}) \end{CD} \] Since all the mappings except for $t'f$ are surjective by Lemmas \ref{T:surj iota_i} and \ref{T:surj tf for Morse}, so is $t'f$. \end{proof} \noindent We define the set $X$ as follows: \[ X=\{g\in C^\infty(N\times I)~|~ g_0 = f, \hspace{.3em} \Sigma(g_t)=\Sigma,\hspace{.3em}\Delta(g_t) = \Delta\}. \] Let $x_0= f\circ \pi\in X$ and $\Gamma_k^X(\pi^\ast TN)$ be the set of germs of continuous mappings (with respect to the topologies $\tau W^\infty$) from $X$ to $\Gamma_k(\pi^\ast TN)$ at $x_0$. We define $C^X(N\times I)$ and $C^X_k(N\times I)$ in a similar manner. We further define $\Gamma_k^X(j^\ast T\mathbb{R})$ as follows: \[ \Gamma_k^X(j^\ast T\mathbb{R}) = \left\{\xi:(X,x_0) \to C^\infty(N\times I, T\mathbb{R})~\left|~j^k\xi(x)|_{\Sigma\times I} = 0,\hspace{.3em}\pi \circ \xi(x) = x \mbox{ for any }x\in X\right.\right\}. \] It is easy to see that $\Gamma_k^X(j^\ast T\mathbb{R})$ is isomorphic to $C_k^X(N\times I)$ as $C^X(N\times I)$--modules. In what follows we will identify them via this isomorphism. We also regard $C_k^\infty(N\times I)$ as a subset of $C_k^X(N\times I)$ consisting of constant map-germs. \begin{lemma}\label{T:fg C_i^X} The $C^X(N\times I)$--module $C_k^X(N\times I)$ is finitely generated. Furthermore, we can take a finite generating set of $C_k^X(N\times I)$ consisting of elements in $C_k^\infty(N\times I)$. \end{lemma} \begin{proof} We will prove Lemma~\ref{T:fg C_i^X} by induction on $k$. Let $h\in C_0^X(N\times I)$ and $\{\varrho_i~|~i\in I\}\cup \{\varrho_V\}$ be the partition of the unity given in the proof of Lemma~\ref{T:surj iota_i}. By the definition, $\varrho_V$ is contained in $C_0^\infty(N\times I)$. As in the proof of Lemma~\ref{T:surj iota_i}, we can decompose $h$ as follows: \[ \sum_{i\in I} \varrho_i h = \sum_{l=1}^{n} \left( \sum_{i\in I}\sqrt{\varrho_i}h_{l}^i \right)\left( \sum_{i\in I} \sqrt{\varrho_i} y_l^i \right), \] where $\sqrt{\varrho_i}h_l^i$ is a map-germ defined as follows: \[ \sqrt{\varrho_i}h_l^i:(X,x_0) \to C^\infty(N\times I), \hspace{.3em} \left(\sqrt{\varrho_i}h_l^i(g)\right)(x,t) = \begin{cases} \sqrt{\varrho_i}(x)\bigints_{\hspace{.25em}0}^{1}\dfrac{\partial }{\partial x_l}\left( h(g) (sx,t)\right) ds & (x\in L_{\alpha(i)}) \\ 0 & (\mbox{otherwise}). \end{cases} \] Since $\displaystyle \sum_{i\in I} \sqrt{\varrho_i} y_l^i$ is an element in $C_0^\infty(N\times I)$, the decomposition of $\displaystyle\sum_{i\in I} \varrho_i h$ above (together with the decomposition $\displaystyle h = \varrho_V h + \sum_{i\in I} \varrho_i h$) yields the desired statement, provided that $\sqrt{\varrho_i}h_l^i$ is a continuous map-germ. The map-germ $\sqrt{\varrho_i}h_l^i$ can be obtained by composing $h$ to the following mappings: \begin{enumerate} \item the composing mapping $C_0^\infty(L_{\alpha(i)}\times I) \to C_0^\infty(\overline{B(\nu_i)}\times I)$ defined by $\xi_1 \to \xi_1\circ \varphi_i^{-1}$, \item the composing mapping $C_0^\infty(\overline{B(\nu_i)}\times I) \to C_0^\infty(I\times \overline{B(\nu_i)}\times I)$ defined by $\xi_2 \mapsto \xi_2 \circ \sigma$, where $\sigma:I\times \overline{B(\nu_i)}\times I\to \overline{B(\nu_i)}\times I$ is defined by $\sigma(s,x,t) = (sx,t)$ (note that $\sigma$ is proper), \item the differentiating mapping $C_0^\infty(I\times \overline{B(\nu_i)}\times I)\to C^\infty(I\times \overline{B(\nu_i)}\times I)$ defined by $\xi_3 \mapsto \dfrac{\partial \xi_3}{\partial x_l}$, \item the integrating mapping $C^\infty(I\times \overline{B(\nu_i)}\times I)\to C^\infty(\overline{B(\nu_i)}\times I)$ defined by $\displaystyle \xi_4\mapsto \int_0^1\xi_4ds$, \item the inverse mapping of (1), \item the multiplication mapping $C^\infty(L_{\alpha(i)}\times I) \to C_\partial^\infty(L_{\alpha(i)}\times I)$ defined by $\displaystyle \xi_6 \mapsto \sqrt{\varrho_i}\xi_6$, where $C_\partial^\infty(L_{\alpha(i)}\times I)$ is a subset of $C^\infty(L_{\alpha(i)}\times I)$ consisting of functions which are $0$ on a neighborhood of $\partial L_{\alpha(i)}\times I$. \item the extension mapping $C_\partial^\infty(L_{\alpha(i)}\times I) \to C^\infty(N\times I)$. \end{enumerate} \noindent We can easily check that these mappings are all continuous (with respect to the topologies $\tau W^\infty$). Thus, $h_l^i$ is also continuous. For $h\in C_k^\infty(N\times I)$ with general $k$, we can obtain the following decomposition in the same way as above: \[ h = \varrho_V h + \sum_{i\in I} \varrho_i h, \hspace{.5em} \sum_{i\in I} \varrho_i h = \sum_{l=1}^{n} \left( \sum_{i\in I}\sqrt{\varrho_i} h_l^i \right)\left( \sum_{i\in I} \sqrt{\varrho_i} y_l^i \right). \] The function $\varrho_V$ is contained in $C_k^\infty(N\times I)$. Moreover, we can deduce from construction of $\sqrt{\varrho_i}h_l^i$ that the map-germ $\displaystyle \sum_{i\in I}\sqrt{\varrho_i} h_l^i $ is contained in $C^X_{k-1}(N\times I)$, which has a finite generating set in $C^\infty_{k-1}(N\times I)$ by the induction hypothesis. Thus we can take a finite set in $C_k^\infty(N\times I)$ so that any $h\in C_k^\infty(N\times I)$ is a linear combination of the elements in this set (whose coefficients are in $C^X(N\times I)$). \end{proof} We define $\mathrm{Ev}:C^X(N\times I)\to C^\infty(N\times I)$ and $\mathrm{Ev}_k:C_k^X(N\times I)\to C_k^\infty(N\times I)$ by evaluating a map-germ at $x_0$. \begin{lemma}\label{T:relation kernel evaluation} For any $k\geq 0$, $\operatorname{Ker}(\mathrm{Ev}_k) = \operatorname{Ker}(\mathrm{Ev})\cdot C_k^X(N\times I)$. \end{lemma} \begin{proof} It is clear that the module $\operatorname{Ker}(\mathrm{Ev})\cdot C_k^X(N\times I)$ is contained in $\operatorname{Ker}(\mathrm{Ev}_k)$. On the other hand, by applying the procedure in the proof of Lemma~\ref{T:fg C_i^X} for decomposing a map-germ to $h \in \operatorname{Ker}(\mathrm{Ev}_k)$, we can decompose $h$ into the sum of finite elements in $\operatorname{Ker}(\mathrm{Ev})\cdot C_k^X(N\times I)$. \end{proof} We define a $C^X(N\times I)$--module homomorphism $t\overline{j}:\Gamma_{k-1}^X(\pi^\ast TN) \to \Gamma_k^X(j^\ast T\mathbb{R})$ as follows: \[ t\overline{j}([\xi])(x) = t'x (\xi(x)). \] \begin{lemma}\label{T:surj toverlinej} The mapping $t\overline{j}$ is surjective and $\operatorname{Ker}(\mathrm{Ev}_k:\Gamma_k^X(j^\ast T\mathbb{R})\to \Gamma_k((f\circ \pi)^\ast T\mathbb{R}))$ is equal to $t\overline{j}(\operatorname{Ker}(\mathrm{Ev})\cdot \Gamma_{k-1}^X(\pi^\ast TN))$. \end{lemma} \begin{proof} For surjectivity of $t\overline{j}$, it is sufficient to see that a generating set of $\Gamma_k^X(j^\ast T\mathbb{R})$ (as a $C^X(N\times I)$--module) is contained in the image $t\overline{j}(\Gamma_{k-1}^X(\pi^\ast TN))$ since $t\overline{j}$ is a $C^X(N\times I)$--module homomorphism. The following diagram commutes: \[ \xymatrix{ \Gamma_{k-1}^X(\pi^\ast TN) \ar[r]^{t\overline{j}} & \Gamma_k^X(j^\ast T\mathbb{R}) \\ \Gamma_{k-1}(\pi^\ast TN) \ar[u]\ar[r]^{t'f} & \Gamma_k((f\circ \pi)^\ast T\mathbb{R}), \ar[u] } \] where the vertical arrows are the inclusion mappings (note that we regard $\Gamma_l(E)$ as a subset of $\Gamma_l^X(E)$ consisting of constant map-germs). Since the mapping $t'f$ is surjective by Lemma~\ref{T:surj t'f}, the image of $t\overline{j}$ contains $\Gamma_i((f\circ \pi)^\ast T\mathbb{R}) \subset \Gamma_i^X(j^\ast T\mathbb{R})$, in which we can take a generating set of $\Gamma_i^X(j^\ast T\mathbb{R})$ by Lemma~\ref{T:fg C_i^X}. As for the statement on the kernel, we can first obtain $\operatorname{Ker}(\mathrm{Ev}_k) = \operatorname{Ker}(\mathrm{Ev})\cdot \Gamma_k^X(j^\ast T\mathbb{R})$ by Lemma~\ref{T:relation kernel evaluation}. Since $\Gamma_k^X(j^\ast T\mathbb{R})$ is equal to $t\overline{j}(\Gamma_{k-1}^X(\pi^\ast TN))$, we obtain: \[ \operatorname{Ker}(\mathrm{Ev}_k) = \operatorname{Ker}(\mathrm{Ev})\cdot t\overline{j}(\Gamma_{k-1}^X(\pi^\ast TN)) = t\overline{j}(\operatorname{Ker}(\mathrm{Ev})\Gamma_{k-1}^X(\pi^\ast TN)). \] \end{proof} Let $g$ be a function in $\mathcal{V}_3$. Since $\Sigma(g)=\Sigma$ and $\Delta(g)=\Delta$, the following function is an element in $X$: \[ \overline{g}:N\times I \to \mathbb{R}, \hspace{.5em}\overline{g}(x,t) = tg(x) + (1-t) f(x). \] Furthermore, the mapping $\theta_4':\mathcal{V}_3\to X$ defined by $\theta_4'(g)= \overline{g}$ is continuous (cf.~\cite[\S.2, Propositions 2 and 3]{MatherII}). For $G\in X$, we define $\xi(G)\in \Gamma_1(G^\ast T\mathbb{R})$ as follows: \[ \xi(G)(x,t) = \frac{d}{dt}\left(G(x,t) \right). \] It is easy to check that the mapping $\xi:X\to C^\infty(N\times I, T\mathbb{R})$ defined as above is continuous, in particular $\xi$ represent an element in $\Gamma^X_1(j^\ast T\mathbb{R})$. Since $\xi(x_0)$ is equal to $0$, $[\xi]$ is an element in $\operatorname{Ker}(\mathrm{Ev}_1)$. By Lemma~\ref{T:surj toverlinej}, there exists $\zeta\in \operatorname{Ker}(\mathrm{Ev})\cdot \Gamma_0^X(\pi^\ast TN)$ such that $[\xi]=t\overline{j}(\zeta)$. Let $\tilde{\zeta}$ be a representative of $\zeta$ (which is a mapping from a neighborhood of $x_0$ in $X$ to $\operatorname{Ker}(\mathrm{Ev})\cdot \Gamma_0(\pi^\ast TN)$), and $\mathcal{O}_N\subset \Gamma(\pi^\ast TN)$ a neighborhood of the zero-section we can take by Lemma~\ref{T:continuity integral vectfield}. Since $\tilde{\zeta}(x_0) = 0$, we can take an open neighborhood $\mathcal{U}_4'\subset X$ of $x_0$ so that the image $\tilde{\zeta}(\mathcal{U}_4')$ is contained in $\mathcal{O}_N$. Let $\mathcal{U}_4 = (\theta_4')^{-1}(\mathcal{U}_4')$, and $\Psi_g^3=\theta(\tilde{\zeta}(\theta_4'(g)))$ for $g\in \mathcal{U}_4$. We can verify that $g\circ \Psi_g^3$ is equal to $f$ for any $g\in \mathcal{U}_4$ (cf.~\cite[\S.7]{MatherII}), and we eventually complete the proof of (1) of Theorem~\ref{T:suff condi stability}. In summary, for $g\in \mathcal{U}=(\theta_3\circ \theta_2\circ \theta_1)^{-1}(\mathcal{U}_4)$, we have constructed self-diffeomorphisms \[ \beta(g), \Psi_{\theta_1(g)}^1, \Psi_{\theta_2\circ \theta_1(g)}^2, \Psi_{\theta_3\circ \theta_2,\theta_1(g)}^3:N\to N \] and a self-diffeomorphism $\psi_g:\mathbb{R}\to \mathbb{R}$ satisfying the following: \[ \psi_g^{-1}\circ g\circ \beta(g)\circ \Psi_{\theta_1(g)}^1 \circ \Psi_{\theta_2\circ \theta_1(g)}^2 \circ \Psi_{\theta_3\circ \theta_2,\theta_1(g)}^3= f. \] We denote the diffeomorphisms $\beta(g)\circ \Psi_{\theta_1(g)}^1 \circ \Psi_{\theta_2\circ \theta_1(g)}^2 \circ \Psi_{\theta_3\circ \theta_2,\theta_1(g)}^3$ and $\psi_g^{-1}$ by $\beta_s(g)$ and $\beta_t(g)$, respectively. We can directly deduce from the definitions that $\beta_s(f)$ and $\beta_t(f)$ are both equal to the identity mappings. As shown in the proof of Lemma~\ref{T:continuity theta2}, the mapping $\beta_t:\mathcal{U}\to \operatorname{Diff}(\mathbb{R})$ is continuous. We have also verified that the mappings $\beta$, $g\mapsto \Psi_{\theta_2\circ \theta_1(g)}^2$, and $g\mapsto\Psi_{\theta_3\circ \theta_2,\theta_1(g)}^3$ are continuous. Moreover, $\Psi_{\theta_1(g)}^1$ becomes the identity when $f$ is quasi-proper (cf.~Remark~\ref{Re:case quasi-proper}). Therefore, the mapping $\beta_s$ is also continuous provided that $f$ is quasi-proper. This completes the proof of (2) of Theorem~\ref{T:suff condi stability}. \section{Applications}\label{Se:application} In this section we will give two applications of Theorem~\ref{T:suff condi stability}. We first give an explicit example of strongly but not infinitesimally stable function. As we noted in the introduction, we could not obtain such an example relying on known results. We then show that a Nash function with some mild assumption on its gradient is stable (Theorem~\ref{T:stability Nash func}). \subsection{A strongly stable but not infinitesimally stable function} In this subsection we will prove the following theorem: \begin{theorem}\label{T:ex strongly but not inf stable func} The function $F:\mathbb{R}\to \mathbb{R}$ defined by $F(x) = \exp(-x^2)\sin x$ is strongly stable but not infinitesimally stable. \end{theorem} \begin{proof} As we briefly reviewed in Subsection~\ref{Se:stability maps}, a locally stable (Morse) function $f:\mathbb{R}\to \mathbb{R}$ is infinitesimally stable if and only if $f|_{\Sigma(f)}$ is proper, which is equivalent to the condition $Z(f|_{\Sigma(f)})=\emptyset$. Since quasi-properness of a function $f$ is equivalent to the condition $Z(f)\cap \Delta(f)=\emptyset$, it is enough to show the followings: \begin{enumerate} \item $F$ is a Morse function, \item $Z(F) = Z(F|_{\Sigma(F)}) = \{0\}$, \item $\Delta(F)$ does not contain $0$. \end{enumerate} To see them, we need the first and the second derivatives, which can be calculated as follows: {\allowdisplaybreaks \begin{align*} F'(x) &= \exp(-x^2)(-2x \sin x +\cos x), \\ F''(x) &= \exp(-x^2) \left\{(4x^2-3)\sin x -4x \cos x\right\}. \end{align*} } Thus $F'(x)$ is equal to $0$ if and only if $\tan x$ is equal to $1/2x$. Let $a_n\in \mathbb{R}$ be the $n$--th smallest positive solution of the equation $\tan x = 1/2x$. It is easy to see that $\Sigma(F)$ is equal to $\{\pm a_n\in \mathbb{R}~|~n\in \mathbb{Z}_{>0}\}$. We can further verify the following properties of the sequence $\{a_n\}_{n\geq 1}$: \begin{enumerate}[(A)] \item $\displaystyle n\pi < a_n < \frac{(2n+1)\pi}{2}$, in particular $\displaystyle \lim_{n\to \infty} a_n = \infty$, \item $\displaystyle \left|\sin a_n\right| >\left|\sin a_{n+1}\right| >0$ for any $n>0$. \end{enumerate} We can deduce from the condition (B) that $\left|F(\pm a_n)\right| = \left|F(\pm a_m)\right|$ if and only if $n=m$. Since $F$ is an odd function, we can conclude that $F|_{\Sigma}$ is injective. Suppose that $F''(a_n)$ were equal to $0$ for some $n>0$. The solution $a_n$ would satisfy the equality $\displaystyle \frac{4a_n}{4a_n^2-3}=\frac{1}{2a_n}$, but the equation $\displaystyle \frac{4x}{4x^2-3}=\frac{1}{2x}$ does not have a real solution. Hence each critical point of $F$ is non-degenerate, concluding that $F$ is a Morse function. We can deduce from the condition (A) on $\{a_n\}_{n>0}$ that $0$ is contained in $Z(F|_{\Sigma})$ (and also in $Z(F)$). On the other hand, since $\displaystyle \lim_{x\to \pm \infty}F(x)$ is equal to $0$, $0\in \mathbb{R}$ is the only improper point of $F$ (and that of $F|_{\Sigma}$). Lastly, by the condition (B) we can prove that $\Delta(F)$ does not contain $0$. \end{proof} \subsection{Stability of Nash functions} In this subsection we will discuss stability of Nash functions. The reader can refer to \cite{BochnakCosteRoy}, for example, for the definition and basic properties of Nash functions. We first show the following theorem: \begin{theorem}\label{T:stability Nash func} Let $f:\mathbb{R}^n\to \mathbb{R}$ be a locally stable Nash function, and $\nabla f :\mathbb{R}^n \to \mathbb{R}^n$ be the gradient of $f$. If $Z(\nabla f)$ does not contain the origin $0\in \mathbb{R}^n$, $f$ is stable. \end{theorem} \begin{proof} By Theorem~\ref{T:suff condi stability}, it is enough to see that $\Delta(f)$ is contained in $\tau(f)$. By the assumption, there exist positive numbers $\varepsilon, r>0$ such that $(\nabla f)^{-1}\left(\overline{B(\varepsilon)}\right)$ is contained in $B(r)$. We denote by $\mathcal{S}\subset T\mathbb{R}^n$ the horizontal distribution on $\mathbb{R}^n\setminus \{0\}$ defined by $\mathcal{S}_p = T_p\partial B(|p|)$ for $p\in \mathbb{R}^n\setminus \{0\}$. It is easy to check that the following subset is semi-algebraic for any $q\in \Delta(f)$ (note that the gradient $\nabla f$ is a Nash mapping (cf.~\cite[Proposition 2.9.1]{BochnakCosteRoy})): \[ \mathcal{T}_q = \left\{p\in B(r)^c \cap f^{-1}(q-\varepsilon/2,q+\varepsilon/2)~|~\mathcal{S}_p = \operatorname{Ker} df_p\right\}. \] Applying Proposition 2.9.10 of \cite{BochnakCosteRoy}, we can obtain finitely many mutually disjoint Nash submanifolds $\mathcal{T}_q^1,\ldots, \mathcal{T}_q^{j_q}\subset \mathbb{R}^n$ such that $\displaystyle \mathcal{T}_q = \bigcup_j \mathcal{T}_q^{j}$ and that $\mathcal{T}_q^j$ is diffeomorphic to an open cube $(0,1)^{\dim \mathcal{T}_q^j}$ for each $j$. For each $j$ we take a point $p_j \in \mathcal{T}_q^j$. Since $\mathcal{T}_q^j$ is a connected manifold, for any $p\in \mathcal{T}_q^j$ we can take a smooth embedding $c:[0,1]\to \mathcal{T}_q^j$ with $c(0)=p$ and $c(1)=p_j$. We then obtain the following estimate: {\allowdisplaybreaks \begin{align*} \left||p_j| - |p|\right|= & \left|\int_{0}^{1}\frac{d}{dt}\left(|c(t)|\right)dt\right| \\ = & \left|\int_0^1 \frac{c(t)\cdot c'(t)}{|c(t)|}dt\right|\\ = & \left|\int_0^1 \frac{1}{|c(t)|}\left(\frac{|c(t)|}{|\nabla f(c(t))|} \nabla f(c(t))\right)\cdot c'(t)dt\right| \\ = & \left|\int_0^1\frac{1}{|(\nabla f)(c(t))|}\cdot \frac{d}{dt}(f\circ c(t))dt\right| \\ \leq & \frac{1}{\varepsilon}\left|f(p_j)-f(p)\right| < 1. \end{align*} } Let $\displaystyle r' = \max\{|p_1|,\ldots,|p_{j_q}|\}+1$. The estimate above implies that $\mathcal{T}_q$ is contained in the ball $B(r')$. For any $q\in \Delta(f)$ and a compact subset $K\subset \mathbb{R}$, we take a positive number $R>r'$ so that $K$ is contained in $B(R)$. The set $B(R)^c\cap f^{-1}(q)$ does not contain a critical point of $f$. Furthermore, there exists a unique vector field $X$ on $f^{-1}(q-\varepsilon/2,q+\varepsilon/2)\cap B(R)^c$ such that $X_p$ is contained in $(\operatorname{Ker}(df_p)\cap \mathcal{S}_p)^\perp$, which is the orthogonal complement of $\operatorname{Ker}(df_p)\cap \mathcal{S}_p$ in $\mathcal{S}_p$, and $\displaystyle df_p(X_p)=\frac{d}{dt}$ for any $p\in f^{-1}(q-\varepsilon/2,q+\varepsilon/2)\cap B(R)^c$. We take a smooth function $\varrho:\mathbb{R}\to \mathbb{R}$ so that $\varrho(t)\equiv 1$ for $t\in (q-\varepsilon/4,q+\varepsilon/4)$ and $\varrho(t)\equiv 0$ for $t\not\in (q-\varepsilon/2,q+\varepsilon/2)$, and define a vector field $\tilde{X}$ on $B(R)^c$ as follows: \[ \tilde{X}_p = \begin{cases} \varrho(f(p)) X_p & (p\in f^{-1}(q-\varepsilon/2,q+\varepsilon/2)\cap B(R)^c) \\ 0 & (\mbox{otherwise}). \end{cases} \] Since the restriction $\tilde{X}|_{\partial B(R')}$ is a vector field on $\partial B(R')$ for any $R'\geq R$, there exists the integral curve $c_p:\mathbb{R} \to B(R)^c$ of $\tilde{X}$ with the initial value $p$ for any $p \in B(R)^c$. We define a smooth mapping \[ \Phi:(f^{-1}(q)\cap B(R)^c)\times (q-\varepsilon/4,q+\varepsilon/4)\to f^{-1}(q-\varepsilon/4,q+\varepsilon/4)\cap B(R)^c \] by $\Phi(q,t) = c_p(t)$. It is easy to see that $\Phi$ is a diffeomorphism and that $f\circ \Phi$ is the projection onto the second component. Thus $f$ is end-trivial at $q$. \end{proof} \begin{example}\label{Ex:modelMorse stable} For $k\in \{0,\ldots,n\}$ we define a function $G_k:\mathbb{R}^n\to \mathbb{R}$ as follows: \[ G_k(x_1,\ldots,x_n) = \sum_{i=1}^{k} x_i^2 -\sum_{j=k+1}^n x_j^2. \] It is obvious that $G_k$ is a Morse function (locally stable) for any $k$. Since the functions $G_0$ and $G_n$ is proper, we can deduce from the result of Mather \cite{MatherV} that they are (strongly) stable. However, $G_k$ for $0<k<n$ is \textit{not} proper, even not quasi-proper (especially not strongly stable). The function $G_k$ is defined by a polynomial, in particular it is a Nash function. Since the gradient $\nabla G_k$ is a diffeomorphism, the set of improper points $Z(\nabla G_k)$ is empty. We can therefore deduce from Theorem~\ref{T:stability Nash func} that $G_k$ is stable. \end{example} \begin{corollary}\label{T:stable after linear perturbation} Let $f:\mathbb{R}^n\to \mathbb{R}$ be a Nash function. There exists a Lebesgue measure zero subset $\Sigma\subset \mathbb{R}^n$ such that for any $a=(a_1,\ldots, a_n) \in \mathbb{R}^n\setminus \Sigma$ the function $f_a:\mathbb{R}^n\to \mathbb{R}$ defined as \[ f_a(x_1,\ldots,x_n) = f(x_1,\ldots,x_n) + \sum_{i=1}^{n}a_ix_i \] is stable. \end{corollary} \begin{proof} We can first deduce from \cite[Theorem 2]{Ichiki} that there exists a Lebesgue measure zero subset $\Sigma_1\subset \mathbb{R}^n$ such that $f_a$ is locally stable for any $a\in \mathbb{R}^n\setminus \Sigma_1$. Since the gradient $\nabla f$ is a Nash mapping, we can take finitely many semi-algebraic subsets $T_1,\ldots, T_l \subset \mathbb{R}^n$ with $\mathbb{R}^n = \bigcup_{i} T_i$ so that for each $i=1,\ldots,l$ there exist a semi-algebraic set $F_i$ and a semi-algebraic homeomorphism $\theta_i:f^{-1}(T_i)\to F_i\times T_i$ such that $f = p_2\circ \theta_i$ (see \cite[Theorem 9.3.2]{BochnakCosteRoy}). We define a subset $\Sigma_2\subset \mathbb{R}^n$ as follows: \[ \Sigma_2 = \bigcup_{\dim F_j >1} \{-a\in \mathbb{R}^n~|~ a\in T_j\}. \] Since the dimension of $F_i\times T_i$ (as a semi-algebraic set) is at most $n$, the dimension of $\Sigma_2$ is less than $n$. Thus $\Sigma_2$ has Lebesgue measure zero. Let $\Sigma$ be the union $\Sigma_1\cup \Sigma_2$, which has Lebesgue measure zero. For any $a\in \Sigma$, the function $f_a$ is locally stable. Furthermore, there exists $\varepsilon>0$ such that $(\nabla f)^{-1}(-a + \overline{B(\varepsilon)})$ is compact. In particular $-a$ is not contained in $Z(\nabla f)$. Thus $Z(\nabla f_a) = Z(\nabla f) +a$ does not contain the origin $0\in \mathbb{R}^n$. We can then deduce from Theorem~\ref{T:stability Nash func} that $f_a$ is stable. \end{proof}

aa0db35c7ce991b7eccccd49b5ff82dc666e6fdb

\section{Introduction} {\it Introduction.} It is well known that a laser beam can be focused via classical linear or nonlinear optical techniques, e.g., lens or self-Kerr effects~\cite{Boyd2008NO}. As the laser intensity decreases to the few-photon level, focusing can only be possible via linear optics since nonlinear effects become negligible. However, at the few-photon level, the quantum intensity fluctuations which leads to the laser beam width uncertainty becomes critically important. The uncertainty scales as $w_{0}/\sqrt{N}$ with $w_{0}$ and $N$ being the beam width and photon numbers respectively~\cite{Treps2003Science,Chille2015OE}, i.e., the weaker the laser beam becomes, the stronger the width uncertainty will be. Obviously, classical approaches based on linear optics can be only used to focus the beam but is unable to reduce the width uncertainty. It thus significantly limits its applications in quantum-based precision measurements where only very-weak light at the few-photon level is allowed in order to avoid damages, for example, in biological systems~\cite{Taylor2013NPhoton,Taylor2014PRX}. In recent years, multi-spatial-mode (MSM) quadrature squeezing, which explores the transverse spatial degree freedom of light, has received intensive investigations due to its promising applications in a variety of directions such as quantum entanglement and information~\cite{Gatti1999PRL,Treps2005PRA,Lassen2007PRL,Boyer2008Science,Marino2009Nature,Gatti2009PRL,Janousek2009NPhoton,Tasca2011PRA,Kang2012PRL,Edgar2012NComm,Pooser2014PRA,Kovlakov2017PRL}, ultra-sensitive measurement of nanometer displacement~\cite{Treps2004JOB,Pooser2015Optica}, quantum-enhanced metrology like sub-shot-noise or super-resolution quantum imaging imaging~\cite{Kolobov2000PRL,Treps2002PRL,Brambilla2008PRA,Lopaeva2013PRL} and detection of gravitational waves~\cite{LIGO2013NPhoton,Oelker2014OE,Toyra2017PRD}. In general, MSM squeezing involves a large number of squeezed spatial modes, implying localized spatial squeezing and thus reduction of local intensity fluctuation. A series of elegant experiments have been performed to demonstrate MSM squeezing in atomic system~\cite{Boyer2008PRL,Corzo2011OE,Chalopin2011OE,Gabriel2011PRL,Corzo2012PRL,Embrey2015PhysRevX,Du2017APL}. It is straightforward to envision that MSM squeezing would result in stronger suppression of local intensity fluctuations in the transverse plane~\cite{Lugiato1997JOSAB} and thus reduction of beam width uncertainty. Here we demonstrate a surprising effect enabled by MSM squeezing, which is, focusing of a very weak laser beam at the few-photon level and simultaneously remarkable reduction of beam width uncertainty due to the strong suppression of local intensity fluctuation. Our scheme is specifically explained in a conjugate four-wave mixing process in atomic gases, but should be also possible in other systems. \begin{figure}[b!] \centering \includegraphics[width=0.3\textwidth]{FWM.pdf} \caption{(Color online) Conjugate FWM process to realize beam focusing and reduction of beam width uncertainty, where two classical control fields $\Omega_{c1}$ and $\Omega_{c2}$, and a quantum probe $\hat{a}_{p}$ are applied. Experimentally, the four levels can be found from, for example, the D1 line of $^{87}$Rb atoms.} \label{ConjugateFWM} \end{figure} {\it Theoretical model.} As shown in Fig.~\ref{ConjugateFWM}, our scheme to realize beam focusing is implemented in a conjugate four-wave mixing process~\cite{Lukin1999PRL,Lu1998OL,McCormick2007OL,Jasperse2011OE,Boyer2008PRL,Boyer2008Science,Marino2009Nature,Embrey2015PhysRevX,Zhang2015SR,Du2017APL}. The interacting Hamiltonian can be written as \begin{align} \hat{H}=&-\frac{\hbar\mathcal{N} }{L}\int^{L}_{0}dz[\Delta\hat{\sigma}_{22} +\Delta_{c1}\hat{\sigma}_{33} +(\Delta+\Delta_{c2})\hat{\sigma}_{44} \nonumber\\[2ex] &+\Omega_{c1}\hat{\sigma}_{31} + \Omega_{c2}\hat{\sigma}_{42} + g_{32}\hat{a}\,\hat{\sigma}_{32} + g_{41}\hat{a}\,\hat{\sigma}_{41} + h.c.]\,, \end{align} where $\mathcal{N}$ is the number of atoms in the quantization volume of length $L$. $g_{j}$ is the coupling coefficient for the quantum probe $\hat{a}$ which is defined as $g_{j}=\mu_{j}\mathcal{E}_{p}/\hbar$ with $\mu_{j}$ being the dipole moment of the corresponding transition and $\mathcal{E}_{p}=\sqrt{\hbar\omega_{p}/(2\epsilon_0 V)}$ the electric field of a single probe photon ($j\in\{32,41\}$). Here for simplicity $g_{j}$ is assumed to be real and also we will set $g_{32}=g_{41}=g$ in the following. Here we have further simplified the collective atomic operators $\hat{\sigma}_{ij}(\bm{r},t)=\sum_{n}|i_{n}(t)\rangle\langle j_{n}(t)|\delta(\bm{r}-\bm{r}_{n})$ as $\hat{\sigma}_{ij}$ with $\delta(\bm{r}-\bm{r}_{n})$ being the Dirac delta function. $\Omega_{cj}$ ($j\in\{1,2\}$) are the Rabi frequencies of the classical control fields respectively. $\Delta_{c1}=\omega_{c1}-(\omega_{3}-\omega_{1}),\Delta_{c2}=\omega_{c2}-(\omega_{4}-\omega_{2})$ and $\Delta_{p}=\omega_{p}-(\omega_{3}-\omega_{2})$ are the detunings for the corresponding fields, and $\Delta=\Delta_{c1}-\Delta_{p}$ is the two-photon detuning between $\Omega_{c1}$ and $\hat{a}$, here we assume $2\omega_{p}=\omega_{c1}+\omega_{c2}$, leading to $\Delta_{p}=(\Delta_{c1}+\Delta_{c2}+\omega_{21}+\omega_{43})/2$ where $\omega_{ij}=\omega_{i}-\omega_{j}$. In the Hamiltonian, we have also assumed the two control fields are much stronger than the two quantum fields such that they can be considered as classical. Furthermore, since the control fields are chosen to be far-detuned from the atomic transition, their propagation in the medium would be the same as in vacuum where only free-space diffraction needs to be considered. Considering the continuous wave limit, the propagation equations for the quantum field $\hat{a}$ reads \begin{align} \label{propagation-eq} \left(\frac{\partial}{\partial z} - \frac{i}{2k_{p}}\nabla^2_{\perp} \right)\hat{a}(\bm{r}_{\perp},z) &= \frac{ig\mathcal{N}}{c}[\hat{\sigma}^{(1)}_{23}(\bm{r}) +\hat{\sigma}^{(1)}_{14}(\bm{r}) ]\,, \end{align} here $\nabla^2_{\perp}$ introduces the paraxial diffraction which would de-focus the probe in spatial domain. $\hat{\sigma}^{(1)}_{23}(\bm{r})$ and $\hat{\sigma}^{(1)}_{14}(\bm{r})$ which denote the atomic coherence are given by \begin{subequations} \begin{align} \hat{\sigma}^{(1)}_{23}(\bm{r}) &= g\chi_{l1}(\bm{r})\hat{a}(\bm{r}) + g\chi_{n1}(\bm{r})\hat{a}^{\dagger}(\bm{r}) + \hat{F}_{1}(\bm{r})\,,\\[2ex] \hat{\sigma}^{(1)}_{14}(\bm{r}) &= g\chi_{l2}(\bm{r})\hat{a}(\bm{r}) + g\chi_{n2}(\bm{r})\hat{a}^{\dagger}(\bm{r}) + \hat{F}_{2}(\bm{r})\,, \end{align} \end{subequations} $\chi_{j}(\bm{r})$ with $i\in\{l1,l2,n1,n2\}$ describes respectively the linear and nonlinear susceptibilities of the atoms whose exact expressions are usually complicated depending on the laser parameters and are given in the Supplement Material (SM). Furthermore, all $\chi_{j}(\bm{r})$ are now spatial-dependent, not only on the transverse coordinates $\bm{r}_{\perp}$ but also the propagation direction $z$ since we have considered a spatial-distributed control field $\Omega_{c1}(\bm{r})$. And $\hat{F}_{1}(\bm{r})$ and $\hat{F}_{2}(\bm{r})$ are the corresponding quantum noise terms respectively. In the following, we will consider the case when the laser parameters are tuned such that the linear and nonlinear absorptions are negligible, thus we may first drop the quantum noise terms. Note here we have assumed the phase-matching condition in the $z$ direction $k_{c1z}+k_{c2z}=2k_{pz}$ with $k_{jz}$ being the propagation wavevector of the field $j$~($j\in\{c1,c2,p\}$). Then the wave equation is modified to \begin{align} \label{propagation-eq} \left(\frac{\partial}{\partial\zeta} + \frac{i}{2}\nabla^2_{\bm{\xi}} \right)\hat{a}(\bm{\xi},\zeta) &= i\chi_{l}(\bm{\xi},\zeta)\hat{a}(\bm{\xi},\zeta) + i\chi_{n}(\bm{\xi},\zeta)\hat{a}^{\dagger}(\bm{\xi},\zeta)\,, \end{align} where we have rescaled the spatial coordinates as $\bm{\xi}=\bm{r}_{\perp}/S_{\perp}$ and $\zeta=z/S_{z}$ with $S_{z}=k_{p}S^{2}_{\perp}$, and $\chi_{j}=g^{2}\mathcal{N}S_{z}(\chi_{j1}+\chi_{j2})/c$ with $j\in\{l,n\}$ are real functions. Apparently, it is impossible to obtain an analytical solution for the wave equation~(\ref{propagation-eq}), even a numerical calculation turns out to be already very challenging due to the spatial-dependent susceptibilities. Nevertheless, we have managed to numerically solve Eq.~(\ref{propagation-eq}) based on the decomposition of the quantum field $\hat{a}(\bm{\xi},\zeta)$ into a complete set of orthogonal spatial modes \begin{align} \hat{a}(\bm{\xi},\zeta)=\sum_{j}\hat{a}_{j}(\zeta)u_{j}(\bm{\xi},\zeta) \end{align} where $a_{j}(\zeta)$ is the annihilation operator at propagation distance $\zeta$ for the $j$th spatial mode $u_{j}(\bm{\xi},\zeta)$ satisfying $\iint^{\infty}_{-\infty}d\bm{\xi}u^{*}_{j}(\bm{\xi},\zeta)u_{l}(\bm{\xi},\zeta)=\delta_{jl}$ and $\sum_{j}u^{*}_{j}(\bm{\xi},\zeta)u_{j}(\bm{\xi}^{'},\zeta)=\delta(\bm{\xi}-\bm{\xi}^{'})$. In principle, $\{u_{j}(\bm{\xi},\zeta)\}$ can be any complete set of functions satisfying the orthogonal relations. For the sake of simplicity, here $u_{j}(\bm{\xi},\zeta)$ is chosen as the eigenfunctions of the paraxial wave equation, i.e., $(\partial_{\zeta}-i\nabla^{2}_{\bm{\xi}}/2)u_{j}(\bm{\xi},\zeta)=0$ such that the diffraction term can be canceled out, leading to \begin{align} \label{mode-eq} \frac{d\hat{\bm{A}}(\zeta)}{d\zeta}=i\bm{M}(\zeta)\hat{\bm{A}}(\zeta) \end{align} where $\hat{\bm{A}}(\zeta)=\{\hat{a}_{0}(\zeta),\hat{a}_{1}(\zeta),\cdots,\hat{a}_{N-1}, \hat{a}^{\dagger}_{0}(\zeta),\hat{a}^{\dagger}_{1}(\zeta),\cdots,\\\hat{a}^{\dagger}_{N-1}\}$ with $N$ being the number of modes needing to be considered. And \begin{align} \bm{M}(\zeta)=\begin{bmatrix} \bm{C}(\zeta) & \bm{D}(\zeta)\\[2ex] -\bm{D}^{*}(\zeta) & -\bm{C}^{*}(\zeta) \end{bmatrix} \end{align} is the propagation matrix determining the output quantum field. $\bm{C}(\zeta)$ and $\bm{D}(\zeta)$ are defined as follows \begin{subequations} \begin{align*} c_{pq}(\zeta) =& \iint^{\infty}_{-\infty}d\bm{\xi} u^{*}_{p}(\bm{\xi},\zeta)\chi_{l}(\bm{\xi},\zeta)u_{q}(\bm{\xi},\zeta)\\[2ex] d_{pq}(\zeta) =& \iint^{\infty}_{-\infty}d\bm{\xi} u^{*}_{p}(\bm{\xi},\zeta)\chi_{n}(\bm{\xi},\zeta)u^{*}_{q}(\bm{\xi},\zeta) \end{align*} \end{subequations} with $p,q\in\{0,1,\cdots,N-1\}$, and we have $c_{pq}=c^{*}_{qp}$ and $d_{pq}=d_{qp}$ for real $\chi_{l}$ and $\chi_{n}$. Then the formal solution of Eq.~(\ref{mode-eq}) can be given as \begin{align} \hat{\bm{A}}(\zeta)=e^{i\int^{\zeta}_{0}d\zeta^{'}\bm{M}(\zeta^{'})}\hat{\bm{A}}(0)\,. \end{align} Furthermore, the effective Hamiltonian for Eq.~(\ref{mode-eq}) can be written as \begin{align} \label{quantum-eff-hamiltonian} \hat{H}_{\text{eff}}(\zeta)=\frac{1}{2}\sum^{N}_{p,q=1}c_{pq}(\zeta)\hat{a}^{\dagger}_{p}\hat{a}_{q}+d_{pq}(\zeta)\hat{a}^{\dagger}_{p}\hat{a}^{\dagger}_{q} + h.c. \end{align} $\hat{H}_{\text{eff}}$ clearly shows that each spatial mode is coupled to all modes including itself, where the first terms indicate photon redistributions between different modes due to the excitation of higher modes and account for classical physics like the optically induced waveguide effects, and the second terms introduce the quantum effects, i.e., quadrature squeezing for all modes and account for all the quantum effects that will be illustrated below including beam focusing, enhanced reduction of local intensity fluctuation, and reduction of beam width uncertainty. \begin{figure}[t!] \centering \includegraphics[width=0.48\textwidth]{field-dynamics.pdf} \caption{(Color online) The effect of beam focusing as a result of MSM squeezing. In (a) we show the full propagation dynamics of the quantum probe intensity $\langle\hat{a}^{\dagger}\hat{a}\rangle$ which are gradually focused and enhanced. The output probe is plotted in (b) for both the classical and quantum cases in order to exclude the classical effects. The unique feature of MSM squeezing is illustrated in both (c) and (d) where the squeezing of $\hat{a}_{p}$ is obtained versus the width and central position of the LO. Parameters are: the atomic density $n=3.0\times10^{17}~\text{m}^{-3}, L= 1.0~\text{cm}, S_{\perp}=w_{p}=0.1~\text{cm}, \Gamma_{32}=2\pi\times6.0~\text{MHz},\Omega^{(0)}_{c1}=3\Gamma_{32},\Omega_{c2}=10\Gamma_{32},\Delta_{c1}=41.4\Gamma_{32},\Delta_{c2}=-50\Gamma_{32}, \alpha_{0}= 0.2$.} \label{field-dynamics} \end{figure} Practically, it is convenient to choose the control field $\Omega_{c1}$ to be a fundamental Hermite-Gaussian beam, i.e., \begin{align*} \Omega_{c1}(x,z)= \frac{w_{c}\Omega^{(0)}_{c1}}{w_{c}(z)}e^{-\bm{r}^{2}_{\perp}/(2w^{2}_{c}(z))}e^{-ik_{c}\bm{r}^{2}_{\perp}/(4R^{2}_{c}(z))}e^{i\phi(z)} \end{align*} with the propagation dependent width $w_{c}(z)=w_{c}\sqrt{1+(z/z_{C})^2}$, radius of curvature $R_{c}(z)=(z^2+z^{2}_{c})/z$ and the Gouy phase $\phi(z)=\arctan(z/z_{C})$, here $z_{C}$ is the associated Rayleigh length. The choice of a Gaussian control beam will lead to a fully symmetric spatial distribution in the transverse plane for an input Gaussian probe, sufficing us to only consider the 2D~($x,z$) propagation dynamics. In order to realize beam focusing, here we chose the spatial size of $\Omega_{c1}(\bm{r}_{\perp},z)$ as $w_{c}=0.8w_{p}$ which is smaller than that of the probe $\hat{a}$. Furthermore, the input Gaussian probe is assumed to be in a coherent state, i.e., $\langle\hat{a}_{j}\rangle=\alpha_{0}\delta_{j0}$ with $\alpha_{0}$ being the amplitude. We have also chosen the Hermite-Gaussian mode basis for $u_{j}(\bm{\xi},\zeta)$ and set $N=40$ at which the numerical solutions have already converged. The numerical results are shown in Fig.~(\ref{field-dynamics}). Fig~\ref{field-dynamics}(a) plots the field intensity $\langle\hat{I}(x,z)\rangle=\sum_{jl}\langle\hat{a}^{\dagger}_{j}(z)\hat{a}_{l}(z)\rangle u^{*}_{j}(x,z)u_{l}(x,z)$ for a short distance $L=0.126z_{R}$ with $z_{R}=7.9\text{ cm}$ being the Rayleigh length for the quantum probe, it can be seen that the laser beam is gradually focused and amplified during propagation. To be more clearly, we have plotted the input and output probe in Fig.~\ref{field-dynamics}(b) where the output probe have been focused to a spatial width $\sqrt{\langle\hat{W}(L)\rangle}\simeq0.55w_{p}$, here the width is defined in terms of the spatial variance of the intensity distribution as follows according to Ref.~\cite{Chille2015OE} \begin{align} \label{width-definition} \hat{W}(\zeta) =&\frac{1}{\langle\int^{\infty}_{-\infty}\hat{I}(\xi,\zeta) d\xi\rangle } \int^{\infty}_{-\infty}f(\xi)\hat{I}(\xi,\zeta)d\xi\nonumber\\[2ex] =&\frac{1}{\sum_{j}\langle\hat{a}^{\dagger}_{j}\hat{a}_{j}\rangle } \sum_{jl}\hat{a}^{\dagger}_{j}\hat{a}_{l}\int^{\infty}_{-\infty}f(\xi)u^{*}_{j}(\xi,\zeta)u_{l}(\xi,\zeta)d\xi \end{align} where $f(\xi)$ is a measure function which here is chosen as $f(\xi)=2\xi^{2}$ such that $\sqrt{\langle\hat{W}(0)\rangle}=w_{p}$. It should be noted $\hat{W}(\zeta)$ has the dimension of an area but not of a length under this choice. In principle one can also consider different appropriate measures $f(\xi)$~\cite{Chille2015OE}. As mentioned above, the control field $\Omega_{c1}$ is taken to be smaller than the probe in beam size, meaning that parts of the probe lie outside the optical waveguide induced by $\Omega_{c1}$. Thus in the classical picture the probe should not be focused. In order to exclude the possibility that the beam focusing is indeed not induced by classical waveguide effects, we have also calculate the classical field propagation dynamics as shown by the red dashed line in Fig.~\ref{field-dynamics}(b). Evidently, the output field spreads and is totally distorted when only taking the classical dynamics into account, due to the excitations of higher-order spatial modes. However, the situation is essentially different in the quantum regime where all excitations are accompanied by quadrature squeezing. As demonstrated by the blue solid line, the quantum probe is considerably narrowed in width. Additionally, the probe is significantly amplified as illustrated by the magnificent factors for the input and classical output probes. The beam focusing can be understood directly in terms of local MSM squeezing in the transverse plane as further shown in Fig.~\ref{field-dynamics}(c) and (d) where we plot the squeezing as a function of the waist and central position of the local oscillator~(LO) with Gaussian distribution $f_{\text{LO}}(x_{0},x)\propto e^{(x-x_{0})^2/(2w^{2}_{l})}$ respectively. Here the degree of squeezing $S$ is defined as \begin{align} S(x_{0},\zeta)=10\log_{10}\frac{\langle\Delta\hat{P}^2(x_{0},\zeta)\rangle}{\langle\Delta\hat{P}^2(x_{0},0)\rangle} \end{align} with $\hat{P}(x_{0},\zeta) \propto i\int^{\infty}_{-\infty}dx[\hat{a}(x,\zeta)f_{\text{LO}}(x_{0},x)e^{-i\theta}-h.c.]$. As can be seen from Fig.~\ref{field-dynamics}(c), for $w_{l}=0.3w_{p}$ the squeezing for the output probe reaches the maximum at the central area, and then oscillates and eventually decreases gradually to 0, meaning stronger amplification in the probe center as compared to the two wings, and consequently the beam focusing. It should be emphasized here the beam focusing is not possible for single-mode squeezing where the beam width should remain as a constant as suggested by Eq.~(\ref{width-definition}). Furthermore, the spatial oscillation in squeezing can not be observed in single-mode squeezing where squeezing should decreases monotonically as LO shifts away from the probe center, and this is indeed due to the interference between squeezing of different spatial modes. In order to show the MSM nature of squeezing more clearly, we reduce the LO waist to $w_{l}=0.2w_{p}$ where stronger spatial oscillations can be seen. We also calculate the squeezing against the ratio between the spatial size of the LO and that of the probe also shown in Fig.~\ref{field-dynamics}(d). It can be seen that the degree of squeezing reduces rapidly as $w_{l}/w_{p}$. Again, depending on the spatial location of the LO, the squeezing degree will be different due to the interference between different modes as we explained above. \begin{figure}[t!] \centering \includegraphics[width=0.48\textwidth]{fluctuation-reduction.pdf} \caption{(Color online) The width uncertainty (a) and local intensity fluctuations (b) versus the amplitude of the coherent incident probe. Parameters are the same as in Fig.~\ref{field-dynamics} except for $\alpha_{0}=0.2e^{i\pi/2}$.} \label{fluctuation-reduction} \end{figure} The MSM nature of squeezing in the system does not only result in beam focusing, but also remarkable reduction of beam width uncertainty and local intensity fluctuations as depicted in Fig.~\ref{fluctuation-reduction}. In Fig.~\ref{fluctuation-reduction}(a), we calculate the relative reduction of beam width uncertainty $\sqrt{\langle\Delta\hat{W}^{2}\rangle}/\langle\hat{W}\rangle$ for the output probe as a function of the amplitude of the coherent incident probe $|\alpha|$. As a comparison, we also compute the width uncertainty for a single fundamental Hermitian-Gaussian mode which is in either coherent or squeezed state. The squeezed coherent state is chosen to be squeezed in amplitude which results in maximal reduction of width uncertainty~(see SM). As compared to the case of single-mode squeezing of -13.7 dB which is the strongest degree of squeezing that can be obtained in the case of $w_{l}=w_{p}$ in our results as sown Fig.~\ref{field-dynamics}(d), the relative reduction is always more pronounced in the case of MSM squeezing as shown by the blue solid line in Fig.~\ref{fluctuation-reduction}(a). In particular, for very weak incident probe at $|\alpha|=1$, the relative width uncertainty is significantly reduced to $47.9\%$ which is smaller than $77.4\%$ or $86.6\%$ for single-mode squeezed or coherent states respectively. One should be noted that the absolute reduction of width uncertainty in our result should be approximately doubled since the width of the output probe is reduced to $0.55w_{p}$. The underlying physics is that the MSM squeezing leads to a much stronger suppression of local intensity fluctuation in the central area of the probe which determines the width uncertainty, in contrast to the single-mode case as plotted in Fig.~\ref{fluctuation-reduction}(b), consequently a remarkable reduction in the width uncertainty which is a weighted summation of the local intensity fluctuation according to Eq.~(\ref{width-definition}) can be obtained. {\it Discussions and Conclusions.} Further enhancement on the beam focusing and reduction of width uncertainty can be expected by tuning the laser parameters including beam width, detuning and intensities of the control beam $\Omega_{c1}(x,z)$, as well as the atomic density. For example, the probe beam may be further tightly focused by employing a control $\Omega_{c1}(x,z)$ with much smaller spatial size, provided that the stronger diffraction of the small-sized $\Omega_{c1}(x,z)$ can be reduced. This would require to use atomic sample of short length, which in turn would require higher atomic density to get accountable amount of MSM squeezing which ensures stronger beam focusing and reduction of width uncertainty. However, in general, a reliable prediction will be very difficult to make, considering that a large number of spatial modes are involved and it is in general a many-body problem with ``time-dependent" interaction between them as indicated by Eq.~(\ref{quantum-eff-hamiltonian}). In our model, we have restricted ourself in the paraxial regime for proof-of-principle demonstration of simultaneous beam focusing and reduction of width uncertainty. In general, extension to the non-paraxial regime should be possible. However, the numerical calculation will be extremely complicated. As we have mentioned earlier, even in the paraxial regime the numerical simulations become already very challenging. We have demonstrated the unexpected and surprising effects of simultaneous beam focusing and remarkable reduction of width uncertainty via MSM squeezing in a conjugate FWM process. The beam focusing is achieved due to the MSM squeezing which leads to inhomogeneous spatial squeezing in the transverse plane. Furthermore, the considerable reduction of width uncertainty is due to the localized spatial squeezing which reduce significantly the local quantum fluctuations. Our results can be very useful in quantum metrology and precision measurements such as detection of very small displacement of particles in biological system where only weak quantum light is allowed. L. Zhang is grateful for the helpful discussions with Tao Peng, Zhenhuan Yi. We acknowledge the support of Office of Naval Research Grant No. N00014-16-1-3054 and Robert A. Welch Foundation Award No. A1261. \bibliographystyle{apsrev4-1}

50c139c80a7bec4d507d463e7c451ac76367efd8

\section{Introduction} The present paper is a contribution to the study of models of extensional properties in intensional type theories and is in particular concerned with W-types. The W-type constructor in Martin-L\"of type theory\xspace~\cite{Martin-Loef1982,Martin-Loef1984} produces an inductive type whose terms can be understood as well-founded trees. Alternatively, it may be viewed as the free term algebra for a, possibly infinitary, single-sorted signature. It may be used to construct in a uniform way several inductive types, like natural numbers or lists, it provides a predicative counterpart to the notion of well-ordering, and it may be used to give constructive justifications of certain theories of inductive definitions~\cite{Palmgren1992}. Furthermore, not only it is instrumental in constructing Aczel's model of Constructive Zermelo-Fraenkel set theory in Martin-L\"of type theory\xspace~\cite{Aczel1978}, where sets are well-founded trees labelled by (small) types, but it also allows to interpret the Regular Extension Axiom which adds general inductive definitions to CZF~\cite{Aczel1986}. Moerdijk and Palmgren identified a category-theoretic counterpart of W-types in initial algebras for polynomial endofunctors~\cite{MoerdijkPalmgren2000}. This notion has then been used in the context of predicative algebraic set theory to obtain models of constructive set theories and, more generally, for a model-theoretic analysis of predicative systems~\cite{MoerdijkPalmgren2002,Gambino2005,vdBerg2005, vdBergMoerdijk2008,vdBergMoerdijk2011,vdBergMoerdijk2012}. More recently, Swan has shown how to simulate the small object argument using (a generalisation of) W-types in a locally cartesian closed pretopos, in order to construct algebraic weak factorisation systems~\cite{Swan2018}. Most of these constructions and applications of W-types, however, require some extensionality to hold in the type theory. For example, in Dybjer's proof that fixed points of certain type operators may be represented by W-types~\cite{Dybjer1988,Dybjer1997}, the identity reflection rule plays an essential role. Let us point out that such type operators are those where the type variable only appears on the right of function types. Hence, together with~\cite{Aczel1986,Palmgren1992}, Dybjer's result inspired the semantic characterisation of W-types and initial algebras for polynomial endofunctors by Moerdijk and Palmgren. It is of no surprise then that the current category-theoretic formulation of W-types takes place in a locally cartesian closed (l.c.c.) category, i.e.\xspace the category-theoretic counterpart of extensional type theory~\cite{Seely1984,Hofmann1994}. Therefore, it is desirable to have a uniform construction of W-types for the great variety of quotient completions used for modelling extensional type theories into intensional ones, both from a type-theoretic standpoint, i.e.\xspace setoid models~\cite{Hofmann1997,BCP2003,Maietti2009,Palmgren2019}, and from a categorical one, namely exact completions~\cite{CarboniVitale1998,BCRS1998} and more general forms of quotient completions~\cite{MaiettiRosolini2013,MaiettiRosolini2016,vdBergMoerdijk2018}. At present, the author is aware of constructions of W-types for two such quotient completions. The first one is due to van den Berg~\cite{vdBerg2005} and it applies to exact completions of categories with finite limits. But the setoid model in intensional Martin-L\"of type theory\xspace is an instance of such completion \iff UIP holds in the underlying type theory~\cite{EmmeneggerPalmgren2017}. The second construction, instead, was formulated by Palmgren for the setoid model in intensional Martin-L\"of type theory\xspace and then adapted by Bressan to Minimalist Type Theory~\cite{Bressan2015}. It uses in an essential way recursion into a universe, hence it is not directly portable to a (predicative) categorical setting, like the internal language of a $ \Pi $W-pretopos, i.e.\xspace a locally cartesian closed pretopos with W-types. This paper makes a first step towards this general goal, by providing a ``small'' construction of initial algebras for polynomial endofunctors in the e-category of small setoids \catstdt. This construction is ``small'' in the sense that it does not require recursion into a universe and it is thus more suited for being reformulated in a categorical context. The initial observation is that, in order to deal with properties of an inductive type, it is useful to have a way to construct inductive predicates, i.e.\xspace families of inductive types. One possibility is to construct such predicates by recursion into a universe, like Palmgren did. Alternatively, we may consider the possibility of making the needed constructions \emph{primitive} in the theory. It turns out that the predicates needed in the construction of initial algebras are instances of a generalised form of W-type constructor, the so-called \emph{dependent W-type}, also known as tree type. Dependent W-types were introduced by Petersson and Synek~\cite{PeterssonSynek1989} to provide a constructor for general inductive data types, that is, families of mutually dependent inductive types (see also~\cite{NPS1990}). They generalise simple W-types in the sense that dependent ones give rise to free term algebras for multi-sorted signatures. Gambino and Hyland identified a category-theoretic counterpart in initial algebras for certain endofunctors, called \emph{dependent polynomial endofunctor}~\cite{GambinoHyland2004}. They also proved that, in an l.c.c.\ category, dependent W-types can be obtained from non-dependent ones, thus providing justification for an analogous result in type theory claimed in~\cite{PeterssonSynek1989}. An extensive study of (dependent) polynomial functors which includes the semantic of W-types, but goes well beyond that, is in~\cite{GambinoKock2013}. When trying to construct initial algebras in a predicative and intensional setting, two steps are more demanding: identifying the object for the initial algebra, and constructing the morphism of algebras witnessing initiality. Defining the algebra map, as well as showing commutativity and uniqueness for the morphism of algebras are usually conceptually simpler. In the case of setoids, the first step amounts to constructing a partial equivalence relation on the underlying W-type $ W $. In Palmgren's argument, this is done using recursion on the underlying W-type into a universe but, as we shall see, that relation is just an inductive family on the product $ W \times W $. For the construction of the morphism of algebras, a set-theoretic argument would consider a (well-founded) subtree relation and do transfinite recursion on it to obtain a function defined on all subtrees. An additional one step construction will then extend this to a function from all trees. Palmgren implemented this argument in type theory by constructing the transitive closure of a relation using recursion into a universe, and applying it to obtain the setoid of subtrees. In this construction though, the analogy between the recursion principle of W-types and initiality of the algebra is hidden behind a thick layer of set-theoretic machinery. Instead, we just consider the setoid of immediate subtrees $ \mathsf{ImS} \, w $, for a tree $ w : W $ in the underlying type of the (candidate) initial algebra $ \mathsf{s} $. These setoids are just the image factorisation of the branching functions of the tree $ w $, and form what we call a \emph{proof-irrelevant setoid family}, namely a family of setoids indexed by a setoid ($ W $ in this case), whose transport arrows between fibres do not depend on the proof of equality. Any algebra morphism from $ \mathsf{s} $ to an algebra $ a $ with underlying setoid $ A $ determines by restriction a family of functions $ k_w : \mathsf{ImS} \, w \Rightarrow A $ for $ w : W $, satisfying a coherence condition, namely that the family $ (k_w)_w $ is stable under transport of $ \mathsf{ImS} $. The converse is true as well: any coherent family determines a (unique) algebra morphism. One drawback of this approach is that the immediate subtree relation is not transitive. In particular, contrary to what happens with the setoid of all subtrees, there is no algebra map on $ \mathsf{ImS} \, w $. This means that, when trying to do recursion on $ w $ to obtain a function $ \mathsf{ImS} \, w \Rightarrow A $, we may not say that such a function is an algebra morphisms. In turn, this blocks the recursive step, since we no longer know whether the family of functions obtained by the inductive hypothesis is coherent. In order to tackle this issue, we exploit dependent W-types once more. Let us call an extensional function $ k : \mathsf{ImS} \, w \Rightarrow A $ \emph{recursively defined over $ w $} if it is obtained applying the above step to a coherent family $ (k_s : \mathsf{ImS} \, s \Rightarrow A)_s $, for $ s $ immediate subtree of $ w $, such that each of the functions $ k_s $ is also recursively defined. Since the functions $ k_s $ are over an immediate subtree, we may use dependent W-types to construct the type of recursively defined functions. Using the elimination principle of dependent W-types, we prove in \cref{lem:recdefuniqref} that any two recursively defined function over the same $ w $ are equal. This uniqueness is the key fact to prove initiality. Firstly, it ensures that the property of being recursively defined is preserved by the above recursion step, so that we obtain for each $ w : W $, a recursively defined function $ k_w \colon \mathsf{ImS} \, w \Rightarrow A $. Secondly, it implies that the setoid of algebra morphisms from $ \mathsf{s} $ to $ a $ is isomorphic to the setoid of families of recursively defined functions. With this characterisation, we may turn the family $ (k_w)_w $ of recursively defined functions into an algebra morphism $ k $. Uniqueness of recursively defined functions then yields uniqueness of $ k $. The next section contains a description of the type theory we shall be working with, and defines the basic concepts we shall be dealing with. \Cref{sec:initalgstd} contains the proof that setoids in Martin-L\"of type theory\xspace have initial algebras for polynomial endofunctors. It begins in \cref{ssec:exttrees} with the construction of the algebra $ \mathsf{s} $, \cref{ssec:imsubtrees} defines the family of immediate subtrees and describes the recursive step. The type of recursively defined functions in constructed in \cref{ssec:recdef}, where uniqueness of such functions is also proved. \Cref{ssec:charact} contains the characterisation of algebra morphisms as families of recursively defined functions, and \cref{ssec:init} concludes the proof of initiality. We conclude with some remarks on future work and possible extensions to the categorical setting. The formalisation of the present paper is available on the author's GitHub page~\cite{Emmenegger2018a}. \section{Type-theoretic setting} Although the formalisation is done in Coq, we only consider a fragment similar to the logical framework of Martin-L\"of type theory~\cite{NPS1990}. Hence we only assume existence of record types, $ \prod $-types and a universe \`a la Russel $ \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace $ which, in the formalisation is the type $ \mathsf{Set} $. In addition, we require the universe to be closed under $ \prod $-types, and to contain intensional $ \sum $-types, intensional identity types $ =_X $, the unit type \ttunitt and dependent W-types, whose rules are described in the next section. We interpret the logic according to propositions-as-types. However, we believe that our argument may be adapted to work also for other interpretations of logic, namely either through a type of propositions as in the Calculus of Constructions or in the Minimalist Type Theory, or using h-propositions as in Homotopy Type Theory. Application of function terms shall be denoted by juxtaposition, $ \equiv $ and $ \coloneqq $ shall denote judgemental equalities and definitions, respectively. In order to make judgements more readable, we shall drop occurrences of the decoding function $ \mathsf{T} $ and, for $ A : \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace $, write $ a : A $ instead of $ a : \mathsf{T} \, A $. \subsection{Dependent W-types} W-types provide a uniform way to construct inductively defined types. Similarly, dependent W-types allow to construct families of inductive types. Rules for W-types and dependent W-types in our type theory are in \cref{wrules,dwrules}, respectively. We shall often drop (some) subscripts if they are clear from the context. \begin{figure} \caption{Rules for W-types} \label{wrules} \vspace{5pt} \hrule \vspace{5pt} \[ \begin{prooftree} A : \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace \qquad B : A \to \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace \justifies \rule{0pt}{13pt} \wty[A,B] : \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace \using{\footnotesize \textsf{W-FORM}} \end{prooftree} \qquad \qquad \begin{prooftree} a : A \qquad f : B \, a \to \wty[A,B] \justifies \rule{0pt}{13pt} \mathsf{sup}_{A,B} \, a \, f : \wty[A,B] \using{\footnotesize \textsf{W-INTRO}} \end{prooftree} \] \bigskip \[ \begin{prooftree} C : \wty[A,B] \to \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace \qquad c : \prod_{(a : A)} \prod_{(f : B \, a \to \wty[A,B])}% \Bigg( \prod_{b : B \, a}% C\, (f \, b) \Bigg) \to C \, (\mathsf{sup}_{A,B} \, a \, f) \justifies \rule{0pt}{13pt} \wrec[A,B,C,c] : \prod_{w : \wty[A,B]} C \, w \using{\footnotesize \textsf{W-ELIM}} \end{prooftree} \] \bigskip \[ \begin{prooftree} a : A \qquad f : B \, a \to \wty[A,B] \justifies \rule{0pt}{13pt} \wrec[A,B,C,c] \, (\mathsf{sup}_{A,B} \, a \, f)% \equiv c \, a \, f \, (\lambda b.\wrec[A,B,C,c] \, (f \, b)) \using{\footnotesize \textsf{W-CONV}} \end{prooftree} \] \end{figure} \begin{figure} \caption{Rules for dependent W-types} \label{dwrules} \vspace{5pt} \hrule \vspace{5pt} \[ \begin{prooftree} \[ I : \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace \qquad A : I \to \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace \justifies \thickness=0pt B : \prod_{i : I} A \, i \to \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace \qquad% d : \prod_{(i : I)} \prod_{(a : A \, i)} B \, i \, a \to I \] \justifies \rule{0pt}{13pt} \dwty[I,A,B,d] : I \to \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace \using{\footnotesize \textsf{DW-FORM}} \end{prooftree} \] \bigskip \[ \begin{prooftree} i : I \qquad a : A \, i \qquad f : \prod_{b : B \, i \, a} \dwty[I,A,B,d] \, (d \, i \, a \, b) \justifies \rule{0pt}{13pt} \mathsf{dsup}_{I,A,B,d} \, i \, a \, f : \dwty[I,A,B,d] \, i \using{\footnotesize \textsf{DW-INTRO}} \end{prooftree} \] \bigskip \[ \resizebox{\columnwidth}{!}{% \begin{prooftree} \[ C : \prod_{i : I} \dwty[I,A,B,d] \, i \to \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace \justifies \thickness=0pt c : \prod_{(i : I)} \prod_{(a : A \, i)}% \prod_{(f : \prod_{b : B \, i \, a}% \dwty[I,A,B,d] \, (d \, i \, a \, b))}% \Bigg( \prod_{b : B \, i \, a} C \, (d \, i \, a \, b) \, (f \, b) \Bigg)% \to C \, i \, (\mathsf{dsup}_{I,A,B,d} \, i \, a \, f) \] \justifies \rule{0pt}{13pt} \dwrec[I,A,B,d,C,c] : \prod_{i : I} \prod_{w : \dwty[I,A,B,d] \, i}% C \, i \, w \using{\footnotesize \textsf{DW-ELIM}} \end{prooftree} }\] \bigskip \[ \begin{prooftree} i : I \qquad a : A \, i \qquad f : \prod_{b : B \, i \, a} \dwty[I,A,B,d] \, (d \, i \, a \, b) \justifies \rule{0pt}{13pt} \dwrec[I,A,B,d,C,c] \, i \, (\mathsf{dsup}_{I,A,B,d} \, i \, a \, f)% \equiv c \, i \, a \, f \,% (\lambda b.\dwrec[C,c](d \, i \, a \, b) \, (f \, b)) \using{\footnotesize \textsf{DW-CONV}} \end{prooftree} \] \end{figure} Useful ways to regard W-types are through the set-theoretic interpretation of type theory: in such a context W-types represent well-founded labelled trees or, equivalently, free term algebras for infinitary single-sorted signatures. In the first case the set $ A $ is the set of names for the nodes, for each $ a \in A $, the set $ B \, a $ consists of the branches out of the node with name $ a $. Trees are formed by providing a node $ a \in A $ and attaching other trees to the branches in $ B \, a $. This procedure is formally specified by functions $ f \colon B \, a \to \wty[A,B] $, which provide the instruction to attach the tree $ f(b) $ to the branch $ b \in B \, a $. In the second case, the set $ A $ contains the function symbols of the signature, while (the cardinality of) $ B \, a $ is the arity of symbol $ a $. Terms are build out of function symbols according to composition instructions specified by functions $ f \colon B \, a \to \wty[A,B] $. Similarly, dependent W-types may be seen as free term algebras for infinitary multi-sorted signatures: sorts are terms of type $ I $, terms in $ A \, i $ are function symbols with codomain sort $ i $ while the function $ d $ maps each tuple $ i,a,b $ into the sort of the $ b $-th argument of the function symbol $ a $. Notice that, in order for this procedure to produce a non-empty set of trees (resp.\ terms), there must be at least one node with no branches (resp. at least a constant). This is reflected in type theory: if we take $ B : A \to \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace $ to be the constant family with value the unit type, then we will not be able to produce a closed term of type $ \wty[A,B] $. In fact it is not difficult to see that, whenever $ B $ is a non-empty family, $ \wty[A,B] $ is type-theoretically equivalent to the empty type \ttemptyt. The meaning of non-empty can be understood either as $ \prod_{a : A} (B \, a \to \mathsf{0}} \newcommand{\ttemptyt}{$\ttemptym$\xspace) \to \mathsf{0}} \newcommand{\ttemptyt}{$\ttemptym$\xspace $ in plain Martin-L\"of type theory\xspace, or as the variant using propositional truncation $ \prod_{a : A} || B \, a || $, if available. Using elimination, we may define function terms $ \mathsf{n} : \wty[A,B] \to A $ and $ \mathsf{b} : \prod_{w : \wty[A,B]} B \, (\mathsf{n} \, w) \to \wty[A,B] $ such that $ \mathsf{n} \, (\mathsf{sup} \, a \, f) \equiv a $ and $ \mathsf{b} \, (\mathsf{sup} \, a \, f) \equiv f $. Similarly, for every $ i : I $ we have \[ \mathsf{dn}_i : \dwty \, i \to A \, i \qquad% \mathsf{db}_i : \prod_{(w : \dwty \, i)} \prod_{(b : B \, i \, (\mathsf{dn}_i \, w))}% \dwty \, (s \, i \, (\mathsf{dn}_i \, w) \, b), \] such that $ \mathsf{dn}_i \, (\mathsf{dsup} \, i \, a \, f) \equiv a $ and $ \mathsf{db}_i \, (\mathsf{dsup} \, i \, a \, f) \equiv f $. \subsection{The e-category of setoids} The notion of setoid allows the intensional type theory to represent extensional concepts. In fact, setoids provide a model of extensional type theory within the intensional one~\cite{Hofmann1997}. Recall that a setoid is informally defined as a type together with a type-theoretic equivalence relation on it. This can be made precise in various different ways, e.g.\xspace considering partial relations, or Prop-valued relations if a type of proposition is available, or instead h-sets and mere relations in the sense of Homotopy Type Theory. In our context we define a setoid as follows. \begin{defin} A \emph{setoid} $ X $ is a tuple $ (X_0, \approx_X, r_X, s_X, t_X) $ where $ X_0 : \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace $, $ \approx_X : X_0 \to X_0 \to \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace $ and \begin{gather*} r_X : \prod_{x : X_0} x \approx_X x, \qquad% s_X : \prod_{x, x' : X_0} x \approx_X x' \to x' \approx_X x, \\ t_X : \prod_{x, x', x'' : X_0}% x \approx_X x' \to x' \approx_X x'' \to x \approx_X x''. \end{gather*} The type of setoids is defined as a record on the types of $ X_0, \approx_X, r_X, s_X $ and $ t_X $ and it is denoted \catstdt. \end{defin} Any small type $ A : \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace $ gives rise to two setoids: the discrete one, with equality $ \lambda a,a'. a =_A a' $, and the codiscrete one, with equality $ \lambda a,a'.\mathsf{1}} \newcommand{\ttunitt}{$\ttunitm$\xspace $. Of course, the codiscrete setoid over any inhabited type is isomorphic to the discrete setoid on \ttunitt in the category of setoids defined below. When we want to regard a type as a setoid, we shall use the discrete equality unless otherwise stated. Functions between setoids are what one would expect. \begin{defin} Let $ X $ and $ Y $ be setoids. A function term $ f_0 : X_0 \to Y_0 $ is \emph{extensional} (with respect to\xspace the equalities of $ X $ and $ Y $) if there is a term of type \[ \text{ext}(f) \coloneqq \prod_{x, x' : X_0} x \approx_X x' \to f_0 (x) \approx_Y f(x). \] The setoid $ X \Rightarrow Y $ of \emph{extensional functions} from $ X $ to $ Y $ has the type of extensional functions $ \sum_{f_0 : X_0 \to Y_0} \text{ext}(f_0) $ as underlying type, and as equality the equivalence relation \[ f \approx_{X \Rightarrow Y} g \coloneqq \prod_{x : X_0} f_0(x) \approx_Y g_0(x). \] \end{defin} In fact, in the Coq implementation we found more convenient to define the type of extensional functions as a record rather than as a $ \sum $-type. However this makes no essential difference. In the rest of the paper we shall write $ x : X $, for a setoid $ X $, to mean $ x : X_0 $, and we shall often not distinguish between an extensional function $ f : X \Rightarrow Y $ and its underlying function term $ f_0 : X_0 \to Y_0 $. We shall denote by $ f \Rsh \alpha : f \, x \approx_Y f \, x' $ the proof of extensionality of $ f $ applied to $ \alpha : x \approx_X x' $. Occasionally, we shall also find it convenient to drop the subscript from the equality of a setoid. We do not expect these abuses of notation to lead to confusion. According to the idea of representing extensional concepts in intensional Martin-L\"of type theory\xspace using setoids, it is natural to define a (locally small) category \catat to be given by a type of objects $ \Obj_\abscatm{A}} \newcommand{\catat}{$\catam$\xspace $ and setoids of arrows $ \Hom_\abscatm{A}} \newcommand{\catat}{$\catam$\xspace : \Obj_\abscatm{A}} \newcommand{\catat}{$\catam$\xspace \to \Obj_\abscatm{A}} \newcommand{\catat}{$\catam$\xspace \to \conccatm{Std}} \newcommand{\catstdt}{$\catstdm$\xspace $, together with explicit function terms for identity and composition, where the latter has type \[ \prod_{a,b,c : \Obj_\abscatm{A}} \newcommand{\catat}{$\catam$\xspace} \Hom_\abscatm{A}} \newcommand{\catat}{$\catam$\xspace(b,c) \Rightarrow \Hom_\abscatm{A}} \newcommand{\catat}{$\catam$\xspace(a,b) \Rightarrow \Hom_\abscatm{A}} \newcommand{\catat}{$\catam$\xspace(a,c), \] and with identity and associativity axioms formulated using equalities of setoids. A functor between two categories \catat and \catbt consists of a function term $ F $ between their type of objects together with a term of type \[ \prod_{a,a' : \Obj_\abscatm{A}} \newcommand{\catat}{$\catam$\xspace} \Hom_\abscatm{A}} \newcommand{\catat}{$\catam$\xspace(a,a') \Rightarrow \Hom_\abscatm{B}} \newcommand{\catbt}{$\catbm$\xspace(F \, a, F \, a'). \] We shall denote the action of a functor $ F $ on an arrow $ \alpha : \Hom_\abscatm{A}} \newcommand{\catat}{$\catam$\xspace(a,a') $ as $ F_\alpha : \Hom_\abscatm{B}} \newcommand{\catbt}{$\catbm$\xspace(F \, a, F \, a') $. One may define a notion of natural transformation between functors and form the category $ \efun{\abscatm{A}} \newcommand{\catat}{$\catam$\xspace,\abscatm{B}} \newcommand{\catbt}{$\catbm$\xspace} $ of functors between \catat and \catbt. These formulations of the notions of category and functor in type theory are sometimes referred to as \emph{e-category} and \emph{e-functor}, in order to distinguish them from other possible formulations. However, since these are the only formulations that we shall consider in intensional type theory, we shall just say category and functor to mean e-category and e-functor. Within the type theory we may construct a category of setoids~\cite{PalmgrenWilander2014}, also denoted \catstdt, whose type of objects is \catstdt and whose setoid family of arrows is given by $ \lambda X,Y. X \Rightarrow Y $. Identity and composition are defined in the obvious way, the latter shall be denoted as $ g \circ f $. Since the universe of the underlying type theory contains $ \mathsf{1}} \newcommand{\ttunitt}{$\ttunitm$\xspace $ and is closed under $ \sum $ and $ \prod $-types, this category is locally cartesian closed~\cite{Hofmann1994,EmmeneggerPalmgren2017}. In \catstdt, we may define for $ f : B \Rightarrow A $ and $ a : A $ the \emph{setoid fibre} whose underlying type is $ \sum_{b : B} f(b) \approx_A a $ and whose equality is $ (b,\alpha) \approx (b',\alpha') \coloneqq b \approx_B b' $. This assignment gives rise to a setoid family over $ A $, i.e.\xspace a functor $ f^\# : \efun{A^\#, \conccatm{Std}} \newcommand{\catstdt}{$\catstdm$\xspace} $, where $ A^\# $ is the discrete category on the setoid $ A $: its type of objects is $ A_0 $ and, for $ a, a' : A $, its setoid of arrows from $ a $ to $ a' $ is the codiscrete setoid on the type $ a \approx_A a' $. Since $ A^\# $ is a groupoid, each term $ \alpha : a \approx_A a' $ gives rise to an isomorphism $ f^\#_\alpha : f^\# \, a \Rightarrow f^\# \, a' $. Furthermore, $ A^\# $ is posetal (but not skeletal), hence such isomorphism does not depend on equality terms, i.e.\xspace $ f^\#_{\alpha} \approx f^\#_{\alpha'} $ for $ \alpha, \alpha' : a \approx_A a' $, and $ f^\#_{\alpha} \approx id_{f^\# a} $ for any $ \alpha : a \approx_A a $. The converse is also true: any functor $ B : \efun{A^\#, \conccatm{Std}} \newcommand{\catstdt}{$\catstdm$\xspace} $ gives rise to an extensional function $ (\sum_{a : A} B \, a, \approx) \Rightarrow A $ whose underlying function term is the first projection and where equality on the domain is \[ (a,b) \approx (a',b') \coloneqq \sum_{\alpha : a \approx_A a'}% B_{\alpha} \, b \approx_{B \, a'} b'. \] For such a functor $ B $, we shall refer to its action on $ \alpha $ as transport along $ \alpha $. We shall also abbreviate $ B_{\alpha} \, b \approx_{B \, a'} b' $ as $ b \approx_{\alpha} b' $. It is of no surprise that this correspondence between extensional functions and setoid families gives rise to an equivalence of categories between the slice $ \conccatm{Std}} \newcommand{\catstdt}{$\catstdm$\xspace/A $ and the category of setoid families $ \mathsf{Fam} A \coloneqq \efun{A^\#, \conccatm{Std}} \newcommand{\catstdt}{$\catstdm$\xspace} $. \subsection{Polynomial functors and W-types} The category-theoretic analogue of W-types are initial algebras for polynomial endofunctors~\cite{MoerdijkPalmgren2000}. Let $ f \colon B \to A $ be an arrow in a locally cartesian closed category \catct, then we may define a functor $ P_f \colon \abscatm{C}} \newcommand{\catct}{$\catcm$\xspace \to \abscatm{C}} \newcommand{\catct}{$\catcm$\xspace $, the \emph{polynomial endofunctor} associated to $ f $, which maps an object $ X $ into $ \sum_A \prod_f (X \times B) $. An \emph{algebra} for $ P_f $ is given by an object $ X $ and an arrow $ s_X \colon P_f X \to X $, called \emph{algebra map}. and such an algebra is \emph{initial} if for any other algebra $ t_Y \colon P_f Y \to Y $, there is a unique $ h \colon X \to Y $ such that $ t_X \circ (P_f \, h) = h \circ s_X $. It is a well-known result by Lambek that the algebra map of an initial algebra is invertible. In extensional type theory with one universe, the (internal) category of small types and function terms is locally cartesian closed if the universe has \ttunitt, $ \sum $ and $ \prod $ types. Hence we may consider polynomial endofuctors for each function term $ f : B \to A $. An initial algebra is then given by the W-type of the family $ f^{-1} \coloneqq \lambda a. \sum_{b : B} f(b) =_A a $ with algebra map \[ (a, k) : \sum_{a : A} \Big( f^{-1}(a) \to \wty[A,f^{-1}] \Big)% \longmapsto \mathsf{sup}(a,k) : \wty[A,f^{-1}]. \] It is not difficult to see that initiality of $ (\wty[A,f^{-1}], \mathsf{sup}) $ amounts exactly to the recursion principle of $ \wty[A,f^{-1}] $~\cite{MoerdijkPalmgren2000}. In \catstdt, instead of defining polynomial functors associated to extensional functions, we prefer to work with polynomial functors defined from setoid families: in light of the equivalence between extensional functions and setoid families, this makes no difference. \begin{defin} Let $ B : \mathsf{Fam} A $ be a setoid family. The \emph{polynomial functor associated to $ B $} is defined on $ X : \conccatm{Std}} \newcommand{\catstdt}{$\catstdm$\xspace $ as \begin{gather*} \poly{B} \, X \coloneqq \Big( \sum_{a : A} (B \, a \Rightarrow X),\, \approx_{\poly{B} X} \Big),% \qquad% (a, k) \approx_{\poly{B} X} (a', k') \coloneqq \sum_{\alpha : a \approx a'}% k \approx k' \circ B_\alpha \\ (\poly{B} \, f ) (a, k) \coloneqq (a, f \circ k), \qquad \text{for } f: X \Rightarrow Y. \end{gather*} We shall say that an endofunctor on \catstdt is polynomial if it is naturally isomorphic to a polynomial functor associated to a setoid family $ B $. \end{defin} \section{Initial algebras in setoids} \label{sec:initalgstd} \subsection{The algebra of extensional trees} \label{ssec:exttrees} We proceed now to construct a setoid $ W $ and a $ \poly{B} $-algebra structure on it. Let $ \wty[A_0,B_0] $ be the W-type constructed on $ A_0 : \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace $ and the type family $ B_0 : A_0 \to \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace $, for $ A : \conccatm{Std}} \newcommand{\catstdt}{$\catstdm$\xspace $ and $ B : \mathsf{Fam} A $. A term in $ \poly{B} \, \wty[A_0,B_0] $ consists of a pair $ (a,k) $ where $ a : A_0 $ and $ k : B_0 \, a \to \wty[A_0,B_0] $. Hence $ \mathsf{sup} $ gives rise to a function term $ \poly{B} \, \wty[A_0,B_0] \to \wty[A_0,B_0] $. We then have to construct a partial equivalence relation $ \approx_W $ on $ \wty[A_0,B_0] $ such that this term will eventually be an isomorphism of setoids, that is, a relation such that \begin{align*} \mathsf{sup} \, a \, f \approx_W \mathsf{sup} \, a' \, f'% & \longleftrightarrow \sum_{\alpha : a \approx_A a'} f \approx f' \circ B_\alpha \\ & \longleftrightarrow \sum_{\alpha : a \approx_A a'}% \prod_{\substack{b,b'\\ b \approx_\alpha b'}} f \, b \approx_W f' \, b'. \end{align*} It is tempting to take this condition itself as the definition of $ \approx_W $, since on the right-hand side $ \approx_W $ occurs only on immediate subtrees. Indeed, this relation can be defined by recursion on the inductive type $ \wty[A_0,B_0] $ into the universe \ttunivt. There is however a more elementary construction, that does not require any elimination into a universe, and involve dependent W-types. Define a relation on $ \wty[A_0,B_0] $ as a dependent W-type with indices in $ \wty[A_0,B_0] \times \wty[A_0,B_0] $ as follows: let \begin{gather*} I \coloneqq \wty[A_0,B_0] \times \wty[A_0,B_0], \qquad% X \, (w, w') \coloneqq \mathsf{n} \, w \approx_A \mathsf{n} \, w',\rule{0pt}{1.5em} \qquad Y \, (w, w') \, \alpha \coloneqq% \sum_{b, b'} b \approx_\alpha b',\rule{0pt}{1.5em} \\ d \, (w, w') \, \alpha \, (b, b', \beta) \coloneqq% (\mathsf{b} \, w \, b, \mathsf{b} \, w' \, b') : I,\rule{0pt}{1em} \\ \end{gather*} the relation then is \[ \wper{B} \, w \, w' \coloneqq \dwty[I,X,Y,d] \, (w,w').\rule{0pt}{1.5em} \] \begin{lem} \label{lem:charWper} Let $ w, w' : \wty[A_0,B_0] $. Then $ \wper{B} \, w \, w' $ is inhabited \iff there are \[ \alpha : \mathsf{n} \, w \approx_A \mathsf{n} w' \qquad \text{and} \qquad% \phi : \prod \Big( z : \sum_{b,b'} b \approx_\alpha b' \Big)\,% \wper{B} \, (\mathsf{b} \, w \, (\textup{pr}_1 \, z)) \, (\mathsf{b} \, w' \, (\textup{pr}_2 \, z)). \] \end{lem} \begin{proof} If $ \wper{B} \, w \, w' $ is inhabited then $ \alpha $ and $ \phi $ are obtained from the (dependent) name and branches functions $ \mathsf{dn} $ and $ \mathsf{db} $. Conversely, when $ w \equiv \mathsf{sup} \, a \, k $ and $ w' \equiv \mathsf{sup} \, a' \, k' $, a term of $ \wper{B} \, w \, w' $ is given by $ \mathsf{dsup} \, (w,w') \, \alpha \, \phi $. \end{proof} \begin{prop} The type family $ \wper{B} : \wty[A_0,B_0]\to \wty[A_0,B_0] \to \mathsf{U}} \newcommand{\ttunivt}{$\ttunivm$\xspace $ is a partial equivalence relation, that is, the following types are inhabited: \begin{gather*} \prod_{w, w' : \wty[A_0,B_0]} \wper{B} \, w \, w' \to \wper{B} \, w' \, w, \\ \prod_{w, w', w'' : \wty[A_0,B_0]} \wper{B} \, w \, w' \to% \wper{B} \, w' \, w'' \to \wper{B} \, w \, w''. \end{gather*} \end{prop} \begin{proof} These terms are obtained from straightforward applications of the elimination rule for dependent W-types. Alternatively, one may use the previous Lemma and recursion on $ \wty[A_0,B_0] $. \end{proof} Notice that $ \wper{B} \, w \, w $ is inhabited \iff the branching function $ \mathsf{b} \, w : B_0 \, (\mathsf{n} \, w) \to \wty[A_0,B_0] $ is extensional in the sense that \[ b \approx_{B \, (\mathsf{n} \, w)} b' \to% \wper{B} \, (\mathsf{b} \, w \, b) \, (\mathsf{b} \, w \, b') \] is inhabited. In particular, $ \wper{B} $ is not in general reflexive. We shall say that a tree $ w : \wty[A_0,B_0] $ is \emph{extensional} if there is a term in $ \wper{B} \, w \, w $. We may form the setoid of extensional trees $ W \coloneqq (W_0, \approx_W) : \conccatm{Std}} \newcommand{\catstdt}{$\catstdm$\xspace $, where \[ W_0 \coloneqq \sum_{w : \wty[A_0,B_0]} \wper{B} \, w \, w \qquad \text{and} \qquad% (w,\_) \approx_W (w',\_) \coloneqq \wper{B} \, w \, w'.\rule{0pt}{1.5em} \] We shall often leave the proof of reflexivity in $ \wper{B} \, w \, w $ implicit, and write $ w : W $ to mean $ w : \wty[A_0,B_0] $ and $ w $ extensional. The name and branches term functions $ \mathsf{n} : \wty[A_0,B_0] \to A $ and $ \mathsf{b} : \prod_{w} B_0 \, (\mathsf{n} \, w) \to \wty[A_0,B_0] $ give rise to an extensional function $ \mathsf{n} : W \Rightarrow A $ and a family of extensional functions $ \mathsf{b} : \prod_{w : W} B \, (\mathsf{n} \, w) \Rightarrow W $ which we shall denote with the same symbols. This latter family is in fact itself extensional, since there is a term \[ \mathsf{extb} : \prod_{w, w' : : W} \prod_{\gamma : w \approx_W w'}% \mathsf{b} \, w \approx (\mathsf{b} \, w') \circ (B_{\mathsf{n} \Rsh \gamma}). \] In particular, immediate subtrees of extensional trees are themselves extensional. \begin{lem} \label{lem:charWeq} Let $ w, w' : W $. Then $ w \approx_W w' $ \iff there are \[ \alpha : \mathsf{n} \, w \approx_A \mathsf{n} w' \qquad \text{and} \qquad% \phi : \prod_{b : B \, (\mathsf{n} \, w)}% \mathsf{b} \, w \, b \approx_W (\mathsf{b} \, w' \, (B_{\alpha} \, b)). \] \end{lem} \begin{proof} One direction is proven applying $ \mathsf{n} \Rsh $ and $ \mathsf{extb} $. Conversely, suppose $ w \equiv \mathsf{sup} \, (a,f) $ and $ w' \equiv \mathsf{sup} \, (a',f') $ and let $ \alpha : a \approx a' $ and $\phi : \prod_{b} \wper{B} \, (f_0 \, b) \, (f'_0 \, (B_{\alpha} \, b)) $. Hence a proof of $ w \approx_W w' $ is given by $ \mathsf{dsup} \, \_ \, \alpha \, \psi $, where \[ \psi : \prod \Big( z : \sum_{b,b'} b \approx_\alpha b' \Big)\,% \wper{B} \, (f_0 \, (\textup{pr}_1 \, z)) \, (f'_0 \, (\textup{pr}_2 \, z)) \] is defined on $ (b,b',\beta) $ to be $ \phi \, b : f_0 \, b \approx f'_0 \, (B_\alpha \, b) $ concatenated with $ f' \Rsh \beta : f'_0 \, (B_\alpha \, b) \approx f'_0 \, b' $. \end{proof} We now show that the function term $ \mathsf{sup} : \prod_{a : A} \big( B \, a \to \wty[A_0,B_0] \big) \to \wty[A_0,B_0] $ gives rise to an extensional function $ \mathsf{s} : \poly{B} \, W \Rightarrow W $. \begin{lem} Let $ a : A $, $ f : B \, a \Rightarrow W $. Then there is $ \mathsf{s} \, (a,f) : W $. \end{lem} \begin{proof} Immediate from the observation that $ \mathsf{sup} \, a \, f $ is extensional \iff $ f : B_0 \to W_0 $ is extensional. We provide the required terms and leave the verification to the reader. Let $ f_0 := \textup{pr}_1 f : B_0 \, a \to \wty[A_0,B_0] $. Then $ w := \mathsf{sup} \, a \, f_0 : \wty[A_0,B_0] $ and \[ \mathsf{dsup} \, (w,w) \, (\rho \, a) \, \varepsilon'_f : \wper{B} \, w \, w, \] where $ \rho \, a $ is reflexivity on $ a $ and $ \varepsilon'_f : \prod \big( z : \sum_{b, b'} b \approx_{B \, a} b' \big)\,% \wper{B} \, (f_0 \, (\textup{pr}_1 \, z)) \, (f_0 \, (\textup{pr}_2 \, z)) $ is obtained from extensionality of $ f $. Hence \[ \mathsf{s} \, (a,f) \coloneqq (\mathsf{sup} \, a \, f_0,% \mathsf{dsup} \, \_ \, (\rho \, a) \, \varepsilon'_f) : W. \qedhere \] \end{proof} \begin{lem} The function term $ \mathsf{s} : \poly{B} \, W \to W $ is extensional. \end{lem} \begin{proof} To have two equal elements in the domain is to have $ a, a' : A $, $ f : B \, a \Rightarrow W $, $ f' : B \, a' \Rightarrow W $, $ \alpha : a \approx a' $ and $ \phi : f \approx f' \circ B_\alpha $. \Cref{lem:charWeq} yields the claim. \end{proof} We may now prove a necessary condition for the algebra map $ \mathsf{s} : \poly{B} \, W \Rightarrow W $ to be initial, namely that it is invertible. \begin{prop} There is $ \mathsf{us} : W \Rightarrow \poly{B} \, W $ such that $ \mathsf{s} \circ \mathsf{us} \approx id_W $ and $ \mathsf{us} \circ \mathsf{s} \approx id_{\poly{B} \, W} $. \end{prop} \begin{proof} The function $ \mathsf{us} $ maps a tree $ w : W $ into its name and branching function: \[ \mathsf{us} \, w \coloneqq (\mathsf{n} \, w, \mathsf{b} \, w) : \sum_{a : A} (B \, a \Rightarrow W). \] The terms $ \mathsf{n} \Rsh $ and $ \mathsf{extb} $ ensure its extensionality. The two equations follow unfolding the definitions of $ \mathsf{s} $ and $ \mathsf{us} $. \end{proof} \subsection{The family of immediate subtrees} \label{ssec:imsubtrees} Let $ C : \conccatm{Std}} \newcommand{\catstdt}{$\catstdm$\xspace $ and $ a_C : \poly{B} \, C \Rightarrow C $ be a $ \poly{B} $-algebra. Our aim is now to set up the machinery needed to construct the universal arrow $ W \Rightarrow C $. This shall involve formulating the recursive step and isolating properties that ensure its applicability recursively. \begin{defin} Let $ w : W $. The setoid of \emph{immediate subtrees of $ w $}, denoted $ \mathsf{ImS} \, u $, has $ B_0 \, (\mathsf{n} w) $ as underlying type, and \[ b \approx_{\mathsf{ImS} \, u} b' \coloneqq \mathsf{b} \, w \, b \approx_W \mathsf{b} \, w \, b' \] as equality. The assignment $ \mathsf{ImS}_\gamma \, s \coloneqq B_{\mathsf{n} \Rsh \gamma} \, s $, for $ s : \mathsf{ImS} \, w $ and $ \gamma : w \approx w' $, defines transport maps for $ \mathsf{ImS} $. Hence we obtain a setoid family $ \mathsf{ImS} : \mathsf{Fam} W $, the \emph{family of immediate subtrees}. \end{defin} Those familiar with the exact completion construction will see that, for each $ w : W $, the image factorisation of the branching function $ \mathsf{b} \, w $ takes the form \[ \xycenterm{ B \, (\mathsf{n} \, w) \ar@{=>}[dr]_{e_w \coloneqq (id,(\mathsf{b} \, w) \Rsh)\quad}% \ar@{=>}[rr]^{\mathsf{b} \, w} && W \\ & \mathsf{ImS} \, w. \ar@{=>}[ur]_{\quad m_w \coloneqq ((\mathsf{b} \, w)_0,id)} &} \] where $ e_ w $ and $ m_w $ denote the epi and mono arising from the factorisation of $ \mathsf{b} \, w $. The underlying function term of $ e_w $ is the identity, while its proof of extensionality is just extensionality of $ \mathsf{b} \, w $. The function $ m_w $ instead is the function term of $ \mathsf{b} \, w $ together with the identity function as proof of extensionality. The recursive step shall consist of constructing an extensional function $ \mathsf{ImS} \, w \Rightarrow C $ out of a family of extensional functions $ \mathsf{ImS} \, (\mathsf{b} \, w \, s) \Rightarrow C $, for $ s : \mathsf{ImS} \, w $. \begin{defin} Let $ w : W $. A family of extensional functions \[ F : \prod_{s : \mathsf{ImS} \, w} \mathsf{ImS} \, (\mathsf{b} \, w \, s) \Rightarrow C \] is \emph{coherent} if, for all $ s, s' : \mathsf{ImS} \, w $ and $ \sigma : s \approx_{\mathsf{ImS} \, w} s' $, $ F \, s \approx (F \, s') \circ \mathsf{ImS}_\sigma $. We shall say that two coherent families $ F $ and $ F' $ are equal if $ \prod_{s : \mathsf{ImS} \, w} F s \approx F' s $ and denote with $ \mathsf{CohMaps} \, w $ the setoid of coherent families of extensional functions. \end{defin} In fact $ \mathsf{CohMaps} $ is another setoid family over $ W $, whose transport function for $ \gamma : w \approx_W w' $ \[ \mathsf{CohMaps}_\gamma : \mathsf{CohMaps} \, w \to \mathsf{CohMaps} \, w' \] is defined on $ F $ and $ s : \mathsf{ImS} \, w' $ as $ (F \, (\mathsf{ImS}_{\gamma^{-1}} \, s)) \circ \mathsf{ImS}_{(\mathsf{extb} \, \gamma^{-1})}. $ Hence to have $ F \approx_\gamma F' $ is to have, for all $ s : \mathsf{ImS} \, w $, \[ F \, s \approx_{(\mathsf{ImS} \, (\mathsf{n} \, (\mathsf{b} \, w \, s)) \Rightarrow C)}% (F' \, (\mathsf{ImS}_\gamma \, s)) \circ \mathsf{ImS}_{(\mathsf{extb} \, \gamma)}. \] \begin{lem}[Recursive step] \label{lem:recstep} For every $ w : W $ and $ F : \mathsf{CohMaps} \, w $, there is \[ \mathsf{recst} \, w \, F : \mathsf{ImS} \, w \Rightarrow C, \] such that, for every $ \gamma : w \approx_W w' $, $ F : \mathsf{CohMaps} \, w $ and $ F' : \mathsf{CohMaps} \, w' $ there is \[ \mathsf{extrecst} \, \gamma : F \approx_\gamma F' \longrightarrow% \mathsf{recst} \, w \, F \approx (\mathsf{recst} \, w' \, F') \circ \mathsf{ImS}_{\gamma}. \] \end{lem} \begin{proof} The underlying function term of $ \mathsf{recst} \, w \, F $ is defined on $ s : \mathsf{ImS} \, w $ as \[ a_C \, (\mathsf{n} \, (\mathsf{b} \, w \, s), (F \, s) \circ e_{(\mathsf{b} \, w \, s)}). \] To see that it is extensional, let $ s, s' : \mathsf{ImS} \, w $ and $ \sigma : s \approx s' $. It is enough to show that there is $ \alpha : \mathsf{n} \, (\mathsf{b} \, w \, s) \approx \mathsf{n} \, (\mathsf{b} \, w \, s') $ such that $ F \, s \, b \approx F \, s' (B_{\alpha} \, b) $ for all $ b : B \, (\mathsf{n} \, (\mathsf{b} \, w \, s)) $. We may take $ \alpha \coloneqq \mathsf{n} \Rsh ((\mathsf{b} \, w \Rsh) \sigma) $, then the second equality follows from coherence of $ F $. Let now $ \gamma, F, F' $ be as above, and $ \phi : F \approx_\gamma F' $, $ s : \mathsf{ImS} \, w $. Applying extensionality of $ a_C $, it is enough to show \[ (\mathsf{n} \, (\mathsf{b} \, w \, s), (F \, s) \circ e_{(\mathsf{b} \, w \, s)}) \approx% (\mathsf{n} \, (\mathsf{b} \, w' \, (\mathsf{ImS}_\gamma \, s)),% F' \, (\mathsf{ImS}_\gamma \, s) \circ e_{(\mathsf{b} \, w' \, (\mathsf{ImS}_\gamma \, s))} ). \] For the first component we may take $ \alpha \coloneqq \mathsf{n} \Rsh (\mathsf{extb} \, \gamma) $. It remains to show that, for every $ s : \mathsf{ImS} \, $, \[ F \, s \approx_{(\mathsf{ImS} \, (\mathsf{n} \, (\mathsf{b} \, w \, s)) \Rightarrow C)}% (F' \, (\mathsf{ImS}_\gamma \, s)) \circ \mathsf{ImS}_{(\mathsf{extb} \, \gamma)}. \] For this it is enough to use $ \phi $. \end{proof} \subsection{The type of recursively defined maps} \label{ssec:recdef} We would like apply this construction recursively on $ w : W $, in order to get a term in $ \prod_w \mathsf{ImS} \, w \Rightarrow C $. In order to do this we need to make sure that coherence is preserved. However, the scope of the coherence condition is limited to functions defined on immediate subtrees of a given tree. We could say that coherence has a local character, as opposed to the global character of the commutativity condition $ a_C \circ \poly{B} \, h \approx h \circ \mathsf{s} $ which applies to functions defined on any tree. This makes the coherence condition not suited to be carried along through recursion. However, let us call an extensional function $ k : \mathsf{ImS} \, w \Rightarrow C $ \emph{recursively defined} if it is obtained applying $ \mathsf{recst} $ to a coherent family $ F $ and each of the functions $ F \, s $ is also recursively defined. Since the functions $ F \, s $ are over an immediate subtree, we may construct the type of all inductively defined functions on immediate subtrees using dependent W-types. For $ w : W $ and $ k : \mathsf{ImS} \, w \Rightarrow C $, let $ \mathsf{RecDef} \, w \, k $ be the dependent W-type on \begin{gather*} I \coloneqq \sum_{w : W} \mathsf{ImS} \, w \Rightarrow C, \qquad% X \, (w,k) \coloneqq \sum_{F : \mathsf{CohMaps} \, w} k% \approx \mathsf{recst} \, w \, F, \rule{0pt}{1.5em} \\ Y \, (w,k) \, (F,\_) \coloneqq \mathsf{ImS} \, w,\rule{0pt}{1.5em} \qquad% d \, (w,k_w) \, (F,\_) \, s \coloneqq (\mathsf{b} \, w \, s, F \, s).\rule{0pt}{1.5em} \end{gather*} From its names and branching functions, we obtain the following terms: \begin{gather*} \mathsf{rdfam} \, w \, k : \mathsf{RecDef} \, w \, k \to \mathsf{CohMaps} \, w \\ \mathsf{rdeq} \, w \, k : \prod_{(D : \mathsf{RecDef} \, w \, k)}% k \approx \mathsf{recst} \, w \, (\mathsf{rdfam} \, w \, k \, D), \\ \mathsf{rdcond} \, w \, k : \prod_{(D : \mathsf{RecDef} \, w \, k)}% \prod_{(s : \mathsf{ImS} \, w)}% \mathsf{RecDef} \, (\mathsf{b} \, w \, s) \, (\mathsf{rdfam} \, w \, k \, D \, s). \end{gather*} Recursively defined map are stable under transport. \begin{lem} \label{prop:recdeftransp} Let $ \gamma : w \approx_W w' $ and $ k : \mathsf{ImS} \, w \Rightarrow C $. Then \[ \mathsf{RecDef} \, w \, k \to \mathsf{RecDef} \, w' \, (k \circ \mathsf{ImS}_{\gamma^{-1}}). \] \end{lem} \begin{proof} This is proven by induction on $ D : \mathsf{RecDef} \, w \, k $ into the type \[ \prod_{(w' : W)} \prod_{(\gamma : w \approx w')}% \mathsf{RecDef} \, w' \, (k \circ \mathsf{ImS}_{\gamma^{-1}}). \] Since $ k $ is obtained applying $ \mathsf{recst} $ to $ F \coloneqq \mathsf{rdfam} \, w \, k \, D $, \cref{lem:recstep} implies \[ k \circ \mathsf{ImS}_{\gamma^{-1}} \approx \mathsf{recst} \, w' \, (\mathsf{CohMaps}_\gamma \, F). \] In order to apply $ \mathsf{dsup} $ it only remains to provide a branching function, that is, to show that for each $ s' : \mathsf{ImS} \, w' $, $ \mathsf{CohMaps}_\gamma \, F \, s' $ is recursively defined. This is precisely the inductive hypothesis applied to $ (\mathsf{extb} \, \gamma^{-1} \, s')^{-1} :% \mathsf{b} \, w \, (B_{\mathsf{n} \Rsh \gamma^{-1}} \, s') \approx \mathsf{b} \, w' \, s' $ and $ \mathsf{rdcond} \, w \, k \, D $. \end{proof} As it may be expected, recursively defined maps are unique. \begin{lem} \label{lem:recdefuniqref} Let $ w : W $, $ k : \mathsf{ImS} \, w \Rightarrow C $ and $ k' : \mathsf{ImS} \, w \Rightarrow C $. Then \[ \mathsf{RecDef} \, w \, k \to \mathsf{RecDef} \, w \, k' \to k \approx k'. \] \end{lem} \begin{proof} This is proven by induction on $ D : \mathsf{RecDef} \, w \, k $. It is enough to show that \[ \mathsf{recst} \, w \, (\mathsf{rdfam} \, w \, k \, D)% \approx \mathsf{recst} \, w \, (\mathsf{rdfam} \, w \, k' \, D'). \] Using $ \mathsf{extrecst} $ from \cref{lem:recstep} this reduces to show that, for every $ s : \mathsf{ImS} \, w $ \[ \mathsf{rdfam} \, w \, k \, D \, s \approx \mathsf{rdfam} \, w \, k' \, D' \, s. \] But this is precisely the inductive hypothesis. \end{proof} \begin{prop} \label{prop:recdefuniq} Let $ \gamma : w \approx_W w' $, $ k : \mathsf{ImS} \, w \Rightarrow C $ and $ k' : \mathsf{ImS} \, w' \Rightarrow C $. Then \[ \mathsf{RecDef} \, w \, k \to \mathsf{RecDef} \, w' \, k' \to k \approx k' \circ \mathsf{ImS}_\gamma. \] \end{prop} \begin{proof} Straightforward from the previous two Lemmas. \end{proof} We are now able to construct a family of extensional functions on immediate subtrees \begin{prop} \label{prop:recis} For every $ w : W $ there are \[ \mathsf{recImS} \, w : \mathsf{ImS} \, w \Rightarrow C \qquad \text{and} \qquad% \mathsf{recImSpf} \, w : \mathsf{RecDef} \, w \, (\mathsf{recImS} \, w). \] \end{prop} \begin{proof} The proof is by $ \wty $-elimination on $ w : \wty[A_0,B_0] $ into the type \[ \wper{B} \, w \, w \,\ \longrightarrow% \sum_{k : \mathsf{ImS} \, w \Rightarrow C} \mathsf{RecDef} \, w \, k. \] To simplify the exposition, we may assume without loss of generality that all the trees we shall be dealing with are extensional. This is the case since all (immediate) subtrees of an extensional tree are also extensional, as one can immediately see recalling that a tree is extensional \iff the branching function is extensional as a function $ B \, (\mathsf{b} \, w) \Rightarrow W $. We have an extensional tree of the form $ w \equiv \mathsf{sup} \, a \, f $, and the inductive hypothesis consists of a family \[ IH : \prod_{(s : \mathsf{ImS} \, w)}% \sum_{(k : \mathsf{ImS} \, (f \, s) \Rightarrow C)} \mathsf{RecDef} \, (f \, s) \, k. \] We work towards applying \cref{lem:recstep}. A coherent family is given by \[ F := \textup{pr}_1 \circ IH : \prod_{s : \mathsf{ImS} \, w}% \mathsf{ImS} \, (f \, s) \Rightarrow C. \] Its coherence follows from \cref{prop:recdefuniq} using $ \textup{pr}_2 \circ IH : \prod_{(s : \mathsf{ImS} \, w)}% \mathsf{RecDef} \, (f \, s) \, (F \, s) $. We thus may define \begin{gather*} \mathsf{recImS} \, w := \mathsf{recst} \, w \, F : \mathsf{ImS} \, w \Rightarrow C% \qquad \text{and} \qquad \\ \mathsf{recImSpf} \, w \coloneqq \mathsf{dsup} \, \_ \, (F,\rho \, \_) \, (\textup{pr}_2 \circ IH)% : \mathsf{RecDef} \, w \, (\mathsf{recImS} \, w). \qedhere \end{gather*} \end{proof} \subsection{Characterisation of algebra morphisms} \label{ssec:charact} Before proceeding to construct the universal arrow and show its commutativity and uniqueness, we provide in \cref{thm:commchar} a characterisation of algebra morphisms as those maps which are recursively defined. Consider the setoid $ \mathsf{Alg}_B(\mathsf{s},a_C) $ of algebra morphisms from $ \mathsf{s} $ to $ a_C $: \[ \sum_{h : W \Rightarrow C} h \circ \mathsf{s} \approx a_C \circ (\poly{B} \, h) \qquad \text{and} \qquad% (h,\_) \approx (h',\_) \coloneqq h \approx h', \] and the setoid of families of recursively defined maps $ \mathsf{RFam} $: \[ \sum \Big( F : \prod_w \mathsf{ImS} \, w \Rightarrow C \Big)% \prod_w \mathsf{RecDef} \, w \, (F w) \qquad \text{and} \qquad% (F,\_) \approx (F',\_) \coloneqq \prod_w F \, w \approx F' \, w. \] We shall prove that these setoids are isomorphic. Since any two families of recursively defined maps are equal because of \cref{lem:recdefuniqref}, the family $ \mathsf{recImS} $ from \cref{prop:recis} is the only inhabitant of $ \mathsf{RFam} $. From this observation and the isomorphism above, initiality will follow immediately. Given any function $ h : W \Rightarrow C $, we may consider the family of restrictions $ h|_w \coloneqq h \circ m_w : \mathsf{ImS} \, w \Rightarrow C $. \begin{lem} \label{lem:restisrecdef} Let $ h : W \Rightarrow C $ be an algebra morphism. For every $ w : W $, $ h|_w $ is recursively defined. \end{lem} \begin{proof} The proof is by $ \wty $-elimination on $ w $ into $ \mathsf{RecDef} \, w \, h|_w $. Let then $ w \equiv \mathsf{sup} \, a \, f $. In order to apply $ \mathsf{dsup} $, we first need to provide a coherent family $ F : \mathsf{CohMaps} \, w $ and a proof that $ h|_w \approx \mathsf{recst} \, w \, F $. The family consists of the restrictions of $ h $ to subtrees of $ w $: $ F \, s \coloneqq h \circ m_{(f \, s)} $. Its coherence follows easily from extensionality of $ h $ and $ \mathsf{extb} $. For $ s : \mathsf{ImS} \, w $ we have: \[ \begin{split} h|_w \, s% &\approx a_C \circ (\poly{B} \, h)\circ \mathsf{us} \, (f \, s) \\ &\approx a_C \, (\mathsf{n} \, (f \, s), h \circ \mathsf{b} \, (f \, s)) \\ &\approx a_C \, (\mathsf{n} \, (f \, s), h|_{(f \, s)} \circ e_{(f \, s)}) \\ &\approx \mathsf{recst} \, w \, (\lambda s.h|_{(f \, s)}) \, s. \end{split} \] The inductive hypothesis witnesses that each of the restrictions $ h|_{(f \, s)} $ is recursively defined. \end{proof} Hence there is a function, obviously extensional, \[ \mathsf{rest} : \mathsf{Alg}_B(\mathsf{s},a_C) \Rightarrow \mathsf{RFam} \] that maps an algebra morphism into its family of restrictions. In order to construct an inverse, let us say that a family $ F : \prod_w \mathsf{ImS} \, w \Rightarrow C $ is \emph{coherent} if, for every $ \gamma : w \approx w' $, $ F \, w \approx (F \, w') \circ \mathsf{ImS}_\gamma $. Any such family gives rise to an extensional function $ \mathsf{cmprh} \, F : W \Rightarrow \poly{B} \, C $ defined by \[ w : W \longmapsto% (\mathsf{n} \, w : A,\, (F \, w) \circ e_w : B \, (\mathsf{n} \, w) \Rightarrow C). \] Extensionality is immediate using $ \mathsf{n} \Rsh $ and coherence of $ F $. Further, using reflexivity on $ \mathsf{n} \, w $ and proof-irrelevance of the setoid family $ \lambda a. B \, a \Rightarrow C $ we obtain a term \[ \mathsf{extcmprh} : \Big( \prod_{w : W} F \, w \approx F' \, w \Big)% \longrightarrow \mathsf{cmprh} \, F \approx \mathsf{cmprh} F'. \] It follows from \cref{prop:recdefuniq} that families in $ \mathsf{RFam} $ are coherent, hence the assignment $ RF \mapsto a_C \circ (\mathsf{cmprh} \, (\textup{pr}_1 \, RF)) $ gives rise to an extensional function which, abusing notation, we shall denote \[ \lambda F. a_C \circ (\mathsf{cmprh} \, F) : \mathsf{RFam} \Rightarrow (W \Rightarrow C). \] Extensionality for this function is obtained from $ \mathsf{extcmprh} $ and $ a_C \Rsh $. The next two results show that $ a_C \circ (\mathsf{cmprh} \, F) $ is in fact an algebra morphism. \begin{lem} \label{lem:cmphrisrd} For every $ F : \mathsf{RFam} $ and every $ w : W $, the function $ (a_C \circ (\mathsf{cmprh} \, F))|_w : \mathsf{ImS} \, w \Rightarrow C $ is recursively defined. \end{lem} \begin{proof} In order to apply $ \mathsf{dsup} $ we first need a coherent family $ G : \mathsf{CohMaps} \, w $ and a proof of $ (a_C \circ (\mathsf{cmprh} \, F))|_w \approx \mathsf{recst} \, w \, G $. We may take $G \, s \coloneqq F \, (\mathsf{b} \, w \, s) $, coherence follows as before from \cref{prop:recdefuniq} and the fact that each $ F \, w' $ is recursively defined. The equation is just a matter of unfolding the definitions: on each $ s : \mathsf{ImS} \, w $ the two sides are judgementally equal. Since every function in the family $ F $ is recursively defined, we obtain a canonical witness of $ \mathsf{RecDef} \, w \, (a_C \circ (\mathsf{cmprh} \, F)) $. \end{proof} \begin{corol} \label{corol:cmphrisalg} For every $ F : \mathsf{RFam} $ \[ \mathsf{cmprh} \, F \approx (\poly{B} \, (a_C \circ (\mathsf{cmprh} \, F))) \circ \mathsf{us}. \] Hence the function $ a_C \circ (\mathsf{cmprh} \, F) : W \Rightarrow C $ is an algebra morphism. \end{corol} \begin{proof} For $ w : W $ we have \[ \begin{split} \mathsf{cmprh} \, F \, w% &\approx (\mathsf{n} \, w, (F \, w) \circ e_w) \\ &\approx (\mathsf{n} \, w, a_C \circ (\mathsf{cmprh} \, F) \circ m_w \circ e_w) \\ &\approx \poly{B} \, (a_C \circ (\mathsf{cmprh} \, F)) \, (\mathsf{n} \, w, \mathsf{b} \, w), \end{split} \] where the second equality uses proof-irrelevance of the setoid family $ \lambda a. B \, a \Rightarrow C $ and the fact just proved that both functions are recursively defined on $ w $, and so equal by \cref{lem:recdefuniqref}. \end{proof} We are now able to state our characterisation of algebra morphisms as families of recursively defined maps. \begin{theor} \label{thm:commchar} The two functions \[ \xycenterm[C=5em]{ \mathsf{Alg}_B(\mathsf{s},a_C) \ar@{=>}@<5pt>[r]^-{\mathsf{rest}} &% \mathsf{RFam} \ar@{=>}@<5pt>[l]^-{a_C \circ (\mathsf{cmprh} \_)} } \] are inverse to each other: for every $ h : \mathsf{Alg}_B(\mathsf{s},a_C) $ and $ F : \mathsf{RFam} $ \[ \prod_w \mathsf{rest} \, (a_C \circ (\mathsf{cmprh} \, F)) \, w \approx F \, w % \qquad \text{and} \qquad% a_C \circ (\mathsf{cmprh} \, (\mathsf{rest} \, h)) \approx h. \] \end{theor} \begin{proof} The first equality follows immediately from \cref{lem:recdefuniqref,lem:cmphrisrd}. The second one follows from the fact that $ h $ is an algebra morphism once we observe that, for every $ w : W $, \[ \begin{split} a_C \circ (\mathsf{cmprh} \, (\mathsf{rest} \, h))) \, w% &\equiv a_C \, (\mathsf{n} \, w, h|_w \circ e_w) \\ &\equiv a_C \circ (\poly{B} \, h) \circ \mathsf{us} \, w. \qedhere \end{split} \] \end{proof} \subsection{Initiality of $ \mathsf{s} $} \label{ssec:init} Since any two families of recursively defined maps are equal because of \cref{lem:recdefuniqref}, the family $ \mathsf{recImS} $ from \cref{prop:recis} is the only inhabitant of $ \mathsf{RFam} $. From this observation and the characterisation above, initiality of $ \mathsf{s} : \poly{B} \, W \Rightarrow W $ follows immediately. Unfolding the construction we see that the unique algebra morphism is given by \[ a_C \circ (\mathsf{cmprh} \, \mathsf{recImS}) : W \Rightarrow C, \] where we omitted the proof term witnessing that the maps in $ \mathsf{recImS} $ are recursively defined. \Cref{corol:cmphrisalg} implies that it is an algebra morphism and uniqueness follows from the fact that $ \mathsf{Alg}_B(\mathsf{s}, a_C) $ is isomorphic to the unit type. We have proven the following. \begin{theor} \label{thm:init} The category of setoids \catstdt has initial algebras for polynomial endofunctors, that is, the following type is inhabited: \[ \prod_{\Big( \substack{A : \conccatm{Std}} \newcommand{\catstdt}{$\catstdm$\xspace \\ B : \mathsf{Fam} \, A} \Big)} \prod_{\Big( \substack{C : \conccatm{Std}} \newcommand{\catstdt}{$\catstdm$\xspace \\ a_C : \poly{B} \, C \Rightarrow C} \Big)}% \sum_{(h : \mathsf{Alg}_B(\mathsf{s},a_C))} \prod_{(h' : \mathsf{Alg}_B(\mathsf{s},a_C))} h \approx h'. \] \end{theor} As already mentioned, a theorem by Gambino and Hyland ensures that a locally cartesian closed category has initial algebras for dependent polynomial endofunctors as soon as it has initial algebras for polynomial endofunctors~\cite{GambinoHyland2004}. Formulating this result for e-categories in Martin-L\"of type theory\xspace should pose no conceptual problems. Once this is done, \cref{thm:init} would yield as a corollary that \catstdt has dependent W-types as soon as the underlying universe in Martin-L\"of type theory\xspace has dependent W-types. In order to generalise the argument presented in this paper to general quotient completions, it may be better to first consider the notion of \emph{homotopy exact completion} introduced by van den Berg and Moerdijk~\cite{vdBergMoerdijk2018}. Indeed, such completion is probably the closest to the setoid construction described here among all quotient completions mentioned in the introduction. The first step in that direction would consist of defining a category-theoretic version of dependent W-types in intensional type theory as certain homotopy-initial algebra, along the lines of the work done in Homotopy Type Theory by Awodey, Gambino and Sojakova in~\cite{AGS2017}. Once this is done, we expect our argument to carry over without major obstacles. \section*{Acknowledgements} The work described in this paper would have not been possible without the support of my supervisor Erik Palmgren and, in particular, his Coq library on setoids and setoid families. This result was presented at the Workshop on Types, Homotopy Type Theory, and Verification, held at the Hausdorff Research Institute for Mathematics in Bonn in June 2018, and the paper was completed while I was hosted at the same Institute in July 2018. I thank the organisers of the workshop for giving me the opportunity to talk, the participants for valuable feedbacks and the Institute for excellent working conditions. Financial support from the Royal Swedish Academy of Sciences and the K\&A Wallenberg Foundation is also acknowledged. Paul Taylor's macros package was used to typeset prooftrees. \bibliographystyle{plain}

05837de5da61665fd465aa4330fa77639462bf7b

\section{Introduction} \label{intr} Deep learning techniques have achieved profound success in many challenging real-world applications, including image recognition~\cite{Krizhevsky:2012:ICD:2999134.2999257,DBLP:journals/corr/HeZR016}, speech recognition~\cite{38131,DBLP:journals/corr/abs-1303-5778}, and machine translation~\cite{DBLP:journals/corr/SutskeverVL14,DBLP:journals/corr/BahdanauCB14}, amongst many others~\cite{DBLP:conf/nips/GoodfellowPMXWOCB14,DBLP:journals/nature/SilverHMGSDSAPL16,DBLP:journals/corr/KipfW16}. A recently proposed such technique, MixUp~\cite{mixup17}, is a simple and yet very effective data-augmentation approach to enhance the performance of deep classification models. Through linearly interpolating random data sample pairs and their training targets, MixUp generates a synthetic set of examples and use these examples to augment the training set. In ~\cite{mixup17}, MixUp is shown to dramatically improve the predictive accuracy of the current art of deep neural networks. Despite its demonstrated effectiveness, the power of MixUp is mostly established empirically. To date, the working of MixUp has not been well explained. In addition, the mixing (i.e., interpolation) policies in MixUp are controlled by a global hyper-parameter $\alpha$, which needs to be tuned by trial and error on the data set. This on one hand makes MixUp inconvenient to use, and on the other hand, adds to the mystery what role such a parameter controls and how to tune it properly. The study of this current paper is motivated by developing a deeper understanding of MixUp, specifically pertaining to why it works. To that end, we formulate MixUp as a new form of regularization. Specifically, we categorize regularization schemes as {\em data-independent} regularization and {\em data-dependent} regularization. Data-independent regularization imposes constraints on the model without exploiting the structure (e.g., the distribution) of data; typical examples include penalizing various forms of the norm of the network parameters (e.g., weight decay~\cite{DBLP:conf/nips/HansonP88}) or dropout~\cite{Srivastava:2014:DSW:2627435.2670313}. Data-dependent regularization constrains the parameter space of the model in a way that depends on the structure of the data; previous examples of such schemes include, data augmentation schemes~\cite{Simard:1998:TIP:645754.668381,726791,SimonyanZ14a}, adversarial training schemes (e.g., \cite{Goodfellow-advExamples}), and some information bottleneck based regularization schemes (e.g., \cite{DVIB}). In this paper, we show that MixUp is also a data-dependent regularization scheme in the sense that the imposed constraints on the model explicitly exploit the data distribution. But different from all previous data-dependent regularization schemes, MixUp imposes its constraints by making use of the regions in the input space of the model that are outside of the data manifold. Identifying the constraints imposed by MixUp with a set of ``locally linear'' constraints, we call such a data-dependent regularization scheme a ``locally linear out-of-manifold'' regularization scheme. Under this perspective, we identify an intrinsic problem in MixUp, which we refer to as ``manifold intrusion''. Briefly, manifold intrusion occurs when a mixed example collides with a real example in the data manifold, but is given a soft label that is different from the label of the real example. Figure~\ref{fig:cifar100mess} (left) shows some examples of manifold intrusion. In the figure, the mixed images (top row) look close to a hand-written ``8'' (which indeed live in the data set) and yet MixUp assigns a soft-label that is not ``8''. For example, the soft-label of the top-left image is given as the half-half mix of labels ``1'' and ``5''. \begin{figure}[h] \caption{Left: Linearly interpolated images (top row) from the original images (bottom two rows). Right: Performance of MixUp on a reduced Cifar100 data set (reduced to containing 20\% of data samples) vs various values of $\alpha$. } \centering \includegraphics[width=3.3in]{conflict8.eps} \label{fig:cifar100mess} \end{figure} We explain in this work that manifold intrusion necessarily results in under-fitting and degradation of the model performance. This is confirmed by our experiments, where we see MixUp with an inappropriate hyper-parameter $\alpha$ actually deteriorates the model performance (Figure~\ref{fig:cifar100mess}, right). Indeed, with MixUp, one has to select hyper-parameter $\alpha$ carefully via trial and error, in order to enhance the base model. Built on our understanding and motivated by the potential manifold intrusion in MixUp, we then propose a generalized and adaptive version of MixUp, which we call AdaMixUp. On one hand, AdaMixUp extends the mixing policies in the standard MixUp from 2-fold to higher-fold; more importantly it adaptively learns good policy regions from the data and automatically avoids manifold intrusion. Our experiments on several image classification tasks show that AdaMixUp significantly improves the current art of deep classification models and outperforms the standard MixUp by a large margin. \section{Preliminaries} In a standard classification setting, let ${\cal X}$ denote the vector space in which each example $x$ lives. Throughout the paper, the elements of ${\cal X}$ will be treated as vectors and all vectors in this paper are taken as column vectors. Let ${\cal Y}$ denote the set of all class labels. The objective of our classification problem is to develop a classifier which assigns every example a label in ${\cal Y}$. It is worth noting however that not every $x\in {\cal X}$ is a valid example. That is, some $x$ cannot be associated to a label in ${\cal Y}$. Specifically, the set of all valid examples is only a subset of ${\cal X}$, which we refer to as the {\em data manifold}, or in short, the manifold, and denote it by ${\cal M}$. The following hypothesis is adopted for nearly every classification problem. \begin{hypo}[Basic Hypothesis] \label{ax:g} Each $x\in {\cal M}$ has a {\em unique} label in ${\cal Y}$. We use $g(x)$ to denote the label of $x$, where $g$ is a function mapping ${\cal M}$ onto ${\cal Y}$. We stress that on ${\cal X}$ outside ${\cal M}$, $g$ is not defined. \end{hypo} In the classification problem, we are given a subset ${\cal D}\subset {\cal M}$ as the training examples, where for each $x\in {\cal D}$, $g(x)$ is given. A distribution over a set of $m$ elements will be denoted by a vector in ${\mathbb R}^m$. It is well known that the set of all such distributions form the {\em probability simplex} in ${\mathbb R}^m$, which we will denote by ${\mathbb S}_m$. The set of all distributions on ${\cal Y}$ is obviously a probability simplex, but we will give it a separate notation ${\cal P}({\cal Y})$ for distinction. Specifically, each distribution in ${\cal P}({\cal Y})$ is regarded as a ``soft label'', whereas each distribution in other probability simplexes potentially serves as a set of coefficients, which will be used to mix training examples via a convex combination. For example, a Bernoulli distribution $[\alpha, 1-\alpha]^T$ as a distribution in ${\mathbb S}_2$ may potentially be used to mix two examples $x, x'\in {\cal X}$ by $\alpha x+ (1-\alpha) x'$. Thus the distributions in these simplexes will be referred to as a {\em mixing policy}, or simply {\em policy}. Any subset $\Lambda$ of these probability simplex with non-zero Lebesgue measure will then be called a {\em policy region}. A distribution or policy will be called {\em degenerate} if it is a distribution containing a single non-zero probability mass. Let ${\cal F}({\cal X}, {\cal Y})$ denote the space of all functions mapping ${\cal X}$ to ${\cal P}({\cal Y})$. A {\em discriminative classification model} (such as a neural network model) that outputs a predictive distribution over ${\cal Y}$ for an example $x\in {\cal X}$ can be characterized as a subset ${\cal H} \subset {\cal F}({\cal X}, {\cal Y})$。 Given the model ${\cal H}$, the objective of learning is then finding a member function $H\in {\cal H}$ which hopefully, for each $x\in {\cal M}$, returns a predictive distribution that puts highest probability on $g(x)$. Usually such an $H$ is found by optimizing a well-defined loss function ${\cal L}_{\cal D}(H)$ over all $H$ in the space ${\cal H}$ \begin{equation} \label{eq:basicOpt} \widehat{H}:=\arg\min_{H\in {\cal H}} {\cal L}_{\cal D}(H). \end{equation} We note that we subscript the loss with ${\cal D}$ to emphasize its dependency on the training data ${\cal D}$. \section{Regularization} It is well known that when the space ${\cal H}$ is large (for example, high dimensional), or when the model is too complex, the model tends to overfit and generalize poorly. The main techniques to cure such a problem is {\em regularization}. Regularization, in our view, may refer to any technique that effectively reduces the model capacity or reduces the space ${\cal H}$. In this paper, we wish to divide regularization into two categories: {\em data-dependent} and {\em data-independent}. {\em Data-independent regularization} is the usual notion of regularization. Typical such techniques directly impose certain constraint, say, $C(H)<A$, on the functions $H\in {\cal H}$. This turns the optimization problem (\ref{eq:basicOpt}) to a constrained optimization problem, which, under the Lagrangian formulation, can be expressed by \begin{equation} \label{eq:regOpt2} \widehat{H}:=\arg\min_{H\in {\cal H}} \left\{ {\cal L}_{\cal D}(H) + \beta C(H) \right\} \end{equation} For example, the weight-decay regularization is a data-independent regularization, which essentially imposes a L2-norm constraint on the model parameters. Dropout~\cite{Srivastava:2014:DSW:2627435.2670313} can be regarded as another example of such a regularizer, which can be understood as reducing the model capacity in a Bayesian style~\cite{DropoutAsBayesian} and which has an equivalent loss function similar to the form of (\ref{eq:regOpt2}). A key feature in data-independent regularization is that the training data do not enter the regularization term $C(H)$ in such a scheme. {\em Data-dependent regularization} imposes constraints or makes additional assumptions on ${\cal H}$ with respect to the training data ${\cal D}$. The most well known such techniques are data augmentation. Specifically, in a data augmentation scheme, another set ${\cal D}' \subset {\cal X}$, {\em believed to be inside ${\cal M}$} but beyond the training set ${\cal D}$, are introduced for training. For example, if ${\cal D}$ is a set of images, ${\cal D}'$ can be obtained by rotating each image in ${\cal D}$; the label $g(x)$ of an image $x$ in ${\cal D}'$ is taken as the label of the image $x$ before rotation. Data augmentation turns the optimization problem (\ref{eq:basicOpt}) into \begin{equation} \label{eq:dataAugment} \widehat{H}:=\arg\min_{H\in {\cal H}} \left\{ {\cal L}_{\cal D}(H) + {\cal L}_{{\cal D}'}(H) \right\} \end{equation} where ${\cal L}_{{\cal D}'}$ is the same loss function but on data ${\cal D}'$. Aligning (\ref{eq:dataAugment}) with (\ref{eq:regOpt2}), one can equivalently regard data augmentation as imposing a constraint on the space ${\cal H}$ (hence a regularization scheme), and such a constraint, depending on the data set ${\cal D}$, is precisely due to the hypothesis that ${\cal D}' \subset {\cal M}$ and $g(\cdot)$ is rotation-invariant. Another data-dependent regularization scheme is adversarial training (see, e.g., \cite{Goodfellow-advExamples}). In such a scheme, for each training example $x$, an adversarial example $x'$ is found, and it is assumed that $x'$ has the same label as $x$. The adversarial examples are then also used for training purpose. Although delicate procedures have been proposed to find adversarial examples, the found adversarial examples are used in the same way as the additional data set ${\cal D}'$ in data augmentation. In both data augmentation and adversarial training, the constraints imposed on the label function $g$ is in the data manifold; the constraints essentially specify how function $g$ should behave on additional points in the manifold ${\cal M}$. Next we will argue that MixUp may be viewed as another data-dependent regularization scheme, which introduces constraints on $g$ outside the manifold ${\cal M}$. \section{Locally Linear Out-of-Manifold Regularization} For each $y\in {\cal Y}$, let $\delta_y\in {\cal P}({\cal Y})$ be the single-point-mass (namely, degenerate) distribution on ${\cal Y}$ which puts all the probability mass on $y$. Now we associate with $g$ a function $G: {\cal X}\rightarrow {\cal P}({\cal Y})$ satisfying \begin{equation} \label{eq:G} G(x):=\delta_{g(x)}, {\rm for ~any} ~x\in {\cal M}. \end{equation} Obviously, on the data manifold ${\cal M}$, $G$ is simply a representation of $g$ in ${\cal F}({\cal X}, {\cal Y})$. Additionally any ${\cal G}$ satisfying the above equation is as perfect as $g$ for any valid example. Thus, any such function $G$ can be taken as a ``ground-truth'' classifier. Then the classification problem can be reformulated as constructing a model ${\cal H}\subset {\cal F}({\cal X}, {\cal Y})$ and finding a member $H\in {\cal H}$ that well approximates a function $G$ satisfying (\ref{eq:G}). Noting that the condition (\ref{eq:G}) on $G$ is only within ${\cal M}$, and it imposes no condition on the space ${\cal X}$ outside ${\cal M}$. However, the functions in ${\cal H}$ are defined on ${\cal X}$, well beyond ${\cal M}$. Thus, if one chooses to regularize the model in a data-dependent manner, there is a lots of degrees of freedom beyond ${\cal M}$ which one can exploit. We will show that the MixUp strategy presented in \cite{mixup17} can be seen as such a technique. To begin, for any subset ${\Omega}\subseteq {\cal X}$, we will use ${\Omega}^{(k)}$ to denote the set of all $k$-column matrices in which each column is a point (vector) in $\Omega$. This also defines notations ${\cal M}^{(k)}$ and ${\cal D}^{(k)}$; however with these latter two notations, we require each matrix in ${\cal M}^{(k)}$ and ${\cal D}^{(k)}$ satisfy an additional condition, namely, {\em the labels of the columns of the matrix must be all distinct}. Let $k$ be a given integer not less than 2. Let $\Lambda\subseteq {\mathbb S}_k$ be a policy region, and ${\bf X}\in {\cal X}^{(k)}$. We say that a function $F\in {\cal F}({\cal X}, {\cal Y})$ is {\em $\Lambda$-linear with respect to ${\bf X}$} if for any $\lambda \in \Lambda$ \[ F({\bf X}\lambda) = F({\bf X})\lambda \] where $F(\cdot)$, when having its domain extended to matrices, acts on matrix ${\bf X}$ column-wise. We also refer to such a notion of linearity as a {\em $k$-fold local linearity}. We note that if we allow $\Lambda$ to ${\mathbb R}^k$, then ${\mathbb R}^k$-linearity with respect to all ${\bf X}\in {\cal X}^{(k)}$ for every integer $k\ge 1$ implies the standard linearity. \subsection{Standard MixUp} \begin{hypo}[Standard MixUp Hypothesis] \label{hypo:standardMixUp} There exists a policy region $\Lambda \subset {\mathbb S}^2$ such that for any matrix ${\bf X}\in {\cal M}^{(2)}$, the function ${G}$ is $\Lambda$-linear with respect to ${\bf X}$. \end{hypo} We note that this hypothesis only assumes a fold-2 local linearity and a global choice of $\Lambda$. If Hypothesis \ref{hypo:standardMixUp} holds and one is given $\Lambda$, then one can implementing the hypothesis as a constraint on ${\cal H}$, by imposing on the optimization problem the following constraint. \centerline{ \shadowbox { \parbox[t]{7.5cm}{ Standard MixUp Constraint:\\ Every $H\in {\cal H}$ is $\Lambda$-linear w.r.t. every ${\bf X}\in {\cal D}^{(2)}$, } } } This gives rise a regularization scheme, which can be implemented by repeating the following process: draw a random pair data points\footnote{Although we have required in the definition of ${\cal D}^{(2)}$ that two points in matrix ${\bf X}$ should have different labels, this needs not to be rigorously followed in implementation. Relaxing this condition in fact is expected to decrease the chance of manifold intrusion. Similar comments apply to the AdaMixUp later.} in ${\cal D}$ to form ${\bf X}$, and draw a random $\lambda$ from $\Lambda$; take ${\bf X}\lambda$ as an additional training example and $G({\bf X})\lambda$ as its training target. This scheme, generating more training examples, is essentially the standard MixUp scheme \cite{mixup17}. In this paper, we however point out that one needs to caution with such a regularization scheme. \begin{lem} \label{lem:standardMixUpFail} Let $\Lambda$ be a policy region in ${\mathbb S}_2$ and ${\bf X} \in {\cal M}^{(2)}$. If there is a non-degenerate $\lambda \in \Lambda$ with ${\bf X}\lambda \in {\cal M}$, then Hypothesis \ref{hypo:standardMixUp} can not hold with this choice $\Lambda$\footnote{More precisely, in order to make the hypothesis fail, instead of requiring a single non-degenerate $\lambda$, we in fact require a subset of $\Lambda$ having non-zero measure. But we abbreviate it here for simplicity. Similar comments apply to Lemma. \ref{lem:generalMixUpFail}. \label{footnote:intrudeStdMixUp} }. \end{lem} \noindent {\em Proof:} The condition of the lemma, stated more precisely in Footnote \ref{footnote:intrudeStdMixUp} suggests that there is a region $\Lambda' \subseteq \Lambda$ with non-zero Lebesgue measure such that every $\lambda \in \Lambda'$ is non-degenerate and ${\bf X}\lambda \in {\cal M}$. The fact that ${\bf X}\lambda \in {\cal M}$ suggests, by Hypothesis \ref{ax:g}, that $g({\bf X}\lambda)\in {\cal Y}$, or ${\cal G} ({\bf X}\lambda)$ is a single-point-mass distribution for every $\lambda \in \Lambda'$. To induce a contradiction, suppose that Hypothesis \ref{hypo:standardMixUp} holds for $\Lambda$, namely, $G$ is $\Lambda$-linear with respect to every ${\bf X}\in {\cal M}^{(2)}$. Then ${\cal G}({\bf X}\lambda) = {\cal G}({\bf X}) \lambda$ for every $\lambda \in \Lambda' \subseteq \Lambda$. Since the two columns (points) in ${\bf X}$ have different labels, the two columns of $G({\bf X})$ must be two different one-hot vectors (degenerate distributions). But every $\lambda\in \Lambda'$ is non-degenerate; as a consequence, every $G({\bf X})\lambda$ must be non-degenerate. This contradicts the previous argument that ${\cal G} ({\bf X}\lambda)$ is a single-point-mass. Therefore Hypothesis \ref{hypo:standardMixUp} fails on a region in $\Lambda$ with non-zero measure. \hfill $\Box$ Geometrically, the failure of Hypothesis \ref{hypo:standardMixUp} in this lemma is due to that the mixing $x_1$ and $x_2$ by $\lambda$ causes the mixed point ${\bf X}\lambda$ to ``intrude'' into ${\cal M}$ and ``collide'' with a point, say, $x'$ in ${\cal M}$. We call such a phenomenon {\em manifold intrusion}. Note that in this case, ${\bf X}\lambda$ and $x'$ are in fact the same point, but they have different labels. Obviously, when such intrusion occurs, regularization using ${\bf X}$ and $\lambda$ will contradict with the original training data. This essentially induces a form of {\em under-fitting}. Figure \ref{fig:venn} depicts the effect of MixUp on constraining the space ${\cal H}$, where the region ${\cal G}$ denotes the space of functions in ${\cal F}({\cal X}, {\cal Y})$ satisfying (\ref{eq:G}). When the space ${\cal H}$ is large, there is a large region of ${\cal H}$ satisfying (\ref{eq:G}). This gives a large intersection of ${\cal H}$ and ${\cal G}$, each member of which is a perfect classifier on the training set but performing questionably on the testing set. Thus the model without MixUp tends to result in large prediction variances or over-fitting (left figure). When MixUp is applied and manifold intrusion does not occur (middle figure), the MixUp constraint effectively reduces the space ${\cal H}$. Since the MixUp constraint does not conflict (\ref{eq:G}) (at least explicitly), the reduced ${\cal H}$ would contain members compatible with (\ref{eq:G}). Thus the intersection of ${\cal G}$ and ${\cal H}$ reduces while remaining non-empty. Each member in the intersection is a perfect classifier on the training set and the mixed examples. But the significantly reduced intersection gives rise to significantly lower prediction variances on the testing set, thereby serving to regularize the model and reduce over-fitting. When the application of MixUp results in manifold intrusion (right figure), there is no member in ${\cal H}$ that satisfies both the MixUp constraint and the condition (\ref{eq:G}). The intersection of ${\cal H}$ and ${\cal G}$ then becomes empty. In this case, no member in ${\cal H}$ can fit both the original training data and the (intruding) mixed examples. This gives rise to under-fitting and high prediction biases. From the view point of restricting the space ${\cal H}$, one wishes to make $\Lambda$ as large as possible. But this lemma sets a limit on the possible $\Lambda$ that makes Hypothesis \ref{hypo:standardMixUp} hold. Due to this lemma, we need to assure that the $\Lambda$ we choose does not cause such an ``intrusion''. Let $\Lambda^*$ denote the largest policy region in ${\mathbb S}_2$ such that manifold intrusion does not occur. In the original MixUp, recall that inappropriate hyper-parameter result in no gain and only negative impact. We believe that much of this is due to manifold intrusion. On the other hand, the search of good hyper-parameter $\alpha$ in the original MixUp can be seen as searching for $\Lambda^*$ by trial and error. \begin{figure}[h] \caption{ \label{fig:venn} Effect of MixUp on constraining ${\cal H}$} \centering \includegraphics[width=3.31in]{regu.eps} \end{figure} In this paper, we generalize the standard MixUp to regularization with higher-fold local linearities, where $\Lambda^*$ (more precisely its generalization) is learned from the data. \subsection{Adaptive MixUp (AdaMixUp)} \begin{hypo}[Generalized MixUp Hypothesis] \label{hypo:generalMixUp} For any integer $k$ not less than $2$ and not greater than some $k_{\rm max}$, and for any matrix ${\bf X}\in {\cal M}^{(k)}$, there exists a policy region $\Lambda \subset {\mathbb S}^k$ such that the function $G$ is $\Lambda$-linear w.r.t. ${\bf X}$. Note that here $\Lambda$ depends on ${\bf X}$, which we will denote by $\Lambda({\bf X})$. \end{hypo} If this hypothesis holds, and if for every $k\in \{2, 3, \ldots, k_{\rm max}\}$ and every ${\bf X}\in {\cal M}^{(k)}$, $\Lambda({\bf X})$ in the hypothesis is given, then the hypothesis can be implemented as the following constraint on ${\cal H}$. \centerline{ \shadowbox { \parbox[t]{7.5cm}{ Generalized MixUp Constraint:\\ For every $k\in \{2, 3, \ldots, k_{\rm max}\}$, every ${\bf X}\in {\cal D}^{(k)}$ and every $\Lambda({\bf X})$, every $H\in {\cal H}$ is $\Lambda({\bf X})$-linear w.r.t. ${\bf X}$. } } } Imposing such a constraint in training provides another regularization scheme, which will be one ingredient of the proposed AdaMixUp scheme. We will present AdaMixUp in the next section, after developing its other ingredients. \begin{lem} \label{lem:generalMixUpFail} Let $k\in \{2, 3, \ldots, k_{\rm max}\}$ and ${\bf X}\in {\cal M}^{(k)}$. Suppose that $\Lambda$ is a policy region in ${\mathbb S}_k$. If there is a non-degenerate $\lambda\in \Lambda$ with ${\bf X}\lambda \in {\cal M}$, then Hypothesis \ref{hypo:generalMixUp} with $\Lambda({\bf X}) = \Lambda$ can not hold. \end{lem} This lemma parallels Lemma \ref{lem:standardMixUpFail} and can be proved similarly. Lemma \ref{lem:generalMixUpFail} similarly suggests that when the imposed constraint results in intrusion into the manifold ${\cal M}$, Hypothesis \ref{hypo:generalMixUp} necessarily fails. The lemma also sets a limit for each $\Lambda({\bf X})$. For each ${\bf X}$, we will denote the largest $\Lambda({\bf X})$ by $\Lambda^*({\bf X})$. The following result immediately follows. \begin{lem} For any ${\bf X}\in {\cal M}^{(2)}$, $\Lambda^* \subseteq \Lambda^*({\bf X})$. \end{lem} \noindent {\em Proof:} By the definitions of $\Lambda^*$ and $\Lambda^*({\bf X})$, $\Lambda^* = \bigcap_{{\bf X}\in {\cal M}^{(2)}} \Lambda^* ({\bf X})$. Thus $\Lambda^* \subseteq \Lambda^*({\bf X})$. \hfill $\Box$ In reality, one expects that $\Lambda^*$ is strictly contained in $\Lambda^*({\bf X})$. That is, even when we take $k_{\rm max}=2$, the Generalized MixUp Hypothesis imposes a stronger constraint than the Standard MixUp Hypothesis, and hence can provide a stronger regularization without causing under-fitting. When $k_{\rm max} > 2$, the Generalized MixUp Hypothesis imposes additional constraints, which further regularizes the model and improves generalization. \subsubsection{AdaMixUp} From here on, we stay in the Generalized MixUp regime, namely assuming Hypothesis \ref{hypo:generalMixUp} holds. We now explore the possibility of learning $\Lambda({\bf X})$ for each ${\bf X}$. Suppose that each $\Lambda^* ({\bf X})\subset {\mathbb S}_k$ can be sufficiently parametrized by $\pi_k({\bf X})$ in some space $\Pi$. Then $\pi_k(\cdot)$ can be regarded as a function mapping ${\cal M}^{(k)}$ to $\Pi$. Then learning each $\Lambda({\bf X})$ reduces to learning the function $\pi_k(\cdot)$. We now consider using a neural network to represent the function $\pi_k$. This network will be referred to as a {\em policy region generator}. With a slight abuse of notation, we will use $\pi_k({\bf X})$ to denote any candidate choice of $\Lambda^*({\bf X}) \subset {\mathbb S}_k$ generated by the network $\pi_k(\cdot)$. To train the networks $\{\pi_k\}$, consider the construction of another network $\varphi(\cdot)$. The network takes any $x\in {\cal X}$ as input and outputs a predictive distribution on whether the $x$ lies in manifold ${\cal M}$. This network, which we refer to as an {\em intrusion discriminator}, aims at distinguishing elements inside the manifold ${\cal M}$ and those outside. By assuring that the virtual examples obtained by mixing with the policies in $\pi_k({\bf X})$ are outside the manifold and the original examples are inside the manifold, the supervising signal trains both the networks $\{\pi_k\}$ and $\varphi$. More details are given below. For a given $x\in {\cal X}$, let $p(\cdot|x; \varphi)$ be the predictive distribution on $\{1, 0\}$ generated by network $\varphi$, where $1$ denotes $x$ is outside ${\cal M}$. The loss function for training $\{\pi_k\}$ and $\varphi$ is given by the following cross-entropy loss, which we call the ``intrusion loss'': \begin{equation} \begin{split} {\cal L}_{\rm intr}: = & \frac{1}{k_{\rm max}\!\!\!-1} \sum\limits_{k=2}^{k_{\rm max}}{\mathbb E}_{{\bf X}\sim {\cal D}^{(k)}, \lambda \sim \pi_k({\bf X})} \log p(1|{\bf X}\lambda; \varphi)\\ & + {\mathbb E}_{x \sim {\cal D}} \log p(0|x; \varphi) \end{split} \end{equation} This intrusion loss ${\cal L}_{\rm intr}$, depending on both $\pi_k$'s and $\varphi$ will be responsible for providing training signals for $\pi_k$'s and $\varphi$. Given $\pi_k$'s, let ${\cal D}'$ be the set of all the virtual examples ${\bf X}\lambda$ that can be obtained by drawing an ${\bf X}$ from ${\cal D}^{(k)}$ for some $k$ and mixing with some policy $\lambda$ in region $\pi_k({\bf X})$. Let the ``MixUp Loss'' ${\cal L}_{{\cal D}'}$ be the average cross-entropy loss for the virtual examples ${\bf X}\lambda$ in ${\cal D}'$ with respect to their respective soft labels $G({\bf X})\lambda$. Then we arrive at an overall loss function \begin{equation} \label{eq:lossAdaMixUp} {\cal L}_{\rm total}: = {\cal L}_{\cal D}(H) + {\cal L}_{{\cal D}'}(H, \{\pi_k\}) + {\cal L}_{\rm intr}(\{\pi_k\}, \varphi) \end{equation} Training the networks $H$, $\{\pi_k\}$ and $\varphi$ on ${\cal L}_{\rm total}$ then gives rise to the proposed AdaMixUp framework. Comparing the loss in (\ref{eq:lossAdaMixUp}) with the objective function in (\ref{eq:regOpt2}), one can equate ${\cal L}_{{\cal D}'}(H, \{\pi_k\}) + {\cal L}_{\rm intr}(\{\pi_k\}, \varphi)$ in (\ref{eq:lossAdaMixUp}) with regularization term $\beta C(H)$ in (\ref{eq:regOpt2}). Specifically, here ${\cal L}_{{\cal D}'}(H, \{\pi_k\})$ serves to regularize the model and ${\cal L}_{\rm intr}(\{\pi_k\}, \varphi)$ serves to prevent the model from ``over-regularization'', namely, intrusion or under-fitting. This justifies AdaMixUp (and the standard MixUp) as a regularization technique. The difference between AdaMixUp (\ref{eq:lossAdaMixUp}) and the standard regularization (\ref{eq:regOpt2}) is however two fold: first, unlike the standard regularization which is data-independent, AdaMixUp depends on the dataset; on the other hand, AdaMixUp contains parameters, which are adaptively learned from the dataset ${\cal D}$. Given that each $\pi_k$ has sufficient capacity and properly trained, in theory $\pi_k({\bf X})$ can approximate $\Lambda^*({\bf X})$ well. Then if the ${\varphi}$ also has sufficient capacity, the intrusion loss ${\cal L}_{\rm intr}$ can be driven to near zero. In practice however, one usually need to reduce the capacities of $\pi_k$'s and $\varphi$. Then the intrusion loss may be seen as an approximate measure of the extent to which the data allows for mix up without causing intrusion under the architectural choices of $\pi_k$'s and $\varphi$. \section{Implementation of AdaMixUp} \subsection{Implementation of $\{\pi_k\}$} Instead of constructing $k_{\rm max}-1$ networks $\pi_k$'s, we in fact implement these networks recursively using a single network $\pi_2$. For this purpose, we recursively parametrize $\Lambda({\bf X}) \in {\mathbb S}_k$, which we will rewrite as $\Lambda_k({\bf X})$ for clarity. For an ${\bf X}\in {\cal X}^{(k)}$, we may express it as $[{\bf X}^{(k-1)}, x_k]$, where ${\bf X}^{(k-1)}$ is the sub-matrix of ${\bf X}$ containing the first $k-1$ columns. Then， for $k>2$, we parametrize \begin{equation} \begin{split} \Lambda_k({\bf X}): = & \left\{ \left[ \begin{array}{c} \gamma \lambda\\ 1-\gamma \end{array} \right]: \right. \lambda \in \Lambda_{k-1}\left({\bf X}^{(k-1)}\right), \\ &\left. \left[ \begin{array}{c} \gamma\\ 1-\gamma \end{array} \right] \in \Lambda_2\left( [\lambda {\bf X}^{(k-1)}, x_k] \right) \right\} \end{split} \end{equation} This then allows the use of single network $\pi_2$ to express $\Lambda_k({\bf X})$ for all ${\bf X}$ and all $k$. Specifically, a Fold-2 mixing policy is parameterized by a single number $\gamma$. For each input ${\bf X}\in {\cal X}^{(2)}$, $\pi_2$ returns two positive values $\alpha$ and $\Delta$ with $\alpha+\Delta \le 1$, which specifies $\Lambda({\bf X})$ as the interval $(\alpha, \alpha+\Delta)$ as the range for $\gamma$. To this end, the last layer of the $\pi_2$ network is implemented as a softmax function, which generates a triplet of values $(\alpha, \Delta, \alpha')$ with sum of one, where the third element $\alpha'$ is discarded. Using network $\pi_2$, we generate a set of Fold-2 mixed examples by repeatedly drawing a pair of samples in ${\cal D}$ and mixing them using a random policy drawn from policy region generated by $\pi_2$ on the pair of samples. To generate mixed samples with an increased maximum fold number, we consider the original examples and previously generated mixed examples as the new training set ${\cal D}_{\rm new}$. We then repeat the following process: draw a pair of samples, one from ${\cal D}_{\rm new}$ and the other from ${\cal D}$; submit the pair to network $\pi_2$ to generate a policy region; draw a random policy from the region and obtain a new sample by mixing the pair of samples using the drawn policy. \subsection{Reparametrization Trick} To allow the gradient signal to back-propagate through the policy sampler, a reparametrization trick similar to that of ~\cite{DBLP:journals/corr/KingmaW13} is used. Specifically, drawing $\gamma$ from $(\alpha, \alpha+\Delta)$ is implemented by drawing $\epsilon$ from the uniform distribution over $(0, 1)$, then let $\gamma:=\Delta \cdot \epsilon + \alpha$. \input{fig_Mixup} \subsection{Implementation of $\varphi(\cdot)$} The network $\varphi$ is a neural network that classifies original examples from the mixed examples. In our implementation, $\varphi$ shares with the classifier network all but the final layers. The final layer of $\varphi$ is a simple logistic regression binary classifier. \subsection{Training} The network $\varphi$ and $\pi_2$ are trained jointly, where we iterate over minimizing ${\cal L}_{\cal D}+ {\cal L}_{{\cal D}'}$ and minimizing ${\cal L}_{\rm intr}$ in (\ref{eq:lossAdaMixUp}) via mini-batched SGD. A Fold-2 AdaMixUp is illustrated in Figure~\ref{schema}. \section{Experiments} \subsection{Data Sets and Experimental Settings} We evaluate AdaMixUp on eight benchmarks. \noindent {\em MNIST} is the popular digit (1-10) recognition dataset with 60,000 training and 10,000 test gray-level, 784-dimensional images. {\em Fashion} is an image recognition dataset having the same scale as MNIST, containing 10 classes of fashion product pictures. {\em SVHN} is the Google street view house numbers recognition data set with 73,257 digits (1-10) 32x32 color images for training, 26,032 for testing, and 531,131 additional, easier samples. We did not use the additional images. {\em Cifar10} is an image classification dataset with 10 classes, 50,000 training and 10,000 test samples. {\em Cifar100} is similar to CIFAR10 but with 100 classes and 600 images each. {\em Cifar10-S} and {\em Cifar100-S} are respectively Cifar10 and Cifar100 reduced to containing only 20\% of the training samples. {\em ImageNet-R} is the ImageNet-2012 classification dataset~\cite{RussakovskyDSKSMHKKBBF14} with 1.3 million training images, 50,000 validation images, and 1,000 classes. We follow the data processing approaches used in Mixup~\cite{mixup17}, except that the crop size is 100x100 instead of 224x224 due to our limited computational resources. We report both top-1 and top-5 error rates. In our experiments, the Intrusion Discriminator, Classifier, and Policy Region Generator have the same network architecture, and the first two share the same set of parameters. We test AdaMixUp on two types of baseline networks: a three layers CNN as implemented in~\cite{wu2016tensorpack} as the baseline for easier tasks MNIST and Fashion, and a ResNet-18 as implemented in~\cite{Zagoruykocode} for the other six more difficult tasks. All models examined are trained using mini-batched backprop, as specified in ~\cite{wu2016tensorpack} and ~\cite{Zagoruykocode}, for 400 epochs. Each reported performance value (accuracy or error rate) is the median of the performance values obtained in the final 10 epochs. \subsection{Predictive Performance} \label{acc} Our first experiment compares Fold-2 AdaMixUp (i.e., mixing image pairs) to two networks: the baseline networks (i.e., 3-layer CNN for MNIST and Fashion and ResNet-18 for the rest datasets) and MixUp on the baseline networks. We present the error rates obtained by the Baseline, MixUp, and AdaMixUp, along with the relative error reduction of AdaMixUp over the Baseline, in Table~\ref{tab:accuracy:resnet}. Table~\ref{tab:accuracy:resnet} indicates that the AdaMixUp outperforms both the Baseline and MixUp on all the eight testing datasets. Also, the relative error reduction of AdaMixUp over the Baseline is at least 5.7\%, and with a large margin in some cases. For example, for the SVHN and Cifar10 datasets, the relative error reduction achieved by AdaMixUp are over 30\%. Table~\ref{tab:accuracy:resnet} also suggests that, with the default $\alpha$ value, MixUp failed to improve the baseline systems' accuracy on four out of the eight datasets, namely the MNIST, Cifar10-S, Cifar100-S, and Imagenet-R datasets, as underlined in Table~\ref{tab:accuracy:resnet}, suggesting over-regularization or under-fitting. \begin{table}[h] \centering \resizebox{\columnwidth}{!}{ \begin{tabular}{l|r|r|r|c}\hline Data Set& Baseline&MixUp &Ada&Relative \\ & & &MixUp& Impro. (\%)\\ \hline mnist&0.52&\underline{0.57}&0.49&5.77\\ fashion&7.37&6.92&6.21&15.74\\ svhn&4.50&3.80&3.12&30.67\\ cifar10&5.53&4.24&3.52&36.35\\ cifar100&25.6&21.14&20.97&18.09\\ cifar10-S&7.68&\underline{7.88}&6.85&10.81\\ cifar100-S&28.47&\underline{29.39}&26.72&6.15\\ ImageNet-R top1&53.00&\underline{54.89}&49.17&7.22\\ ImageNet-R top5&29.41&\underline{31.02}&25.78&12.34\\ \hline \end{tabular}} \caption{Error rates (\%) obtained by the testing methods.} \label{tab:accuracy:resnet} \end{table} In Table~\ref{tab:accuracy:para}, we also present the $\alpha$ and $\Delta$ values learned by the Policy Region Generator, along with the losses of both the Intrusion Discriminator and the Classifier of the AdaMixUp method. Results in Table~\ref{tab:accuracy:para} indicate that the values of $\alpha$ and $\Delta$ vary for different datasets; the former ranges from 0.4 to 0.9 but the range for $\Delta$ is much smaller, with values between 0.010 and 0.035. Noting that the intrusion losses are fairly close to 0 on all the datasets, suggesting that these datasets well support Fold-2 AdaMixUp under the proposed structure of Policy Region Generator. \begin{table}[h] \centering \begin{tabular}{l|c|c|c|c}\hline Ada values&$\alpha$&$ \Delta$&${\cal L}_{\rm intr}$&${\cal L}_{\cal D} + {\cal L}_{{\cal D}'}$ \\ \hline mnist&0.497&0.010&0.002&0.201\\ fashion&0.511&0.011&0.003&0.290\\ svhn&0.924&0.011&0.002&0.116\\ cifar10&0.484&0.029&0.003&0.371\\ cifar100&0.484&0.034&0.002&0.679\\ cifar10-S&0.486&0.027&0.003&0.479\\ cifar100-S&0.482&0.035&0.007&0.712\\ ImageNet-R&0.493&0.004&0.038&2.382\\ \hline \end{tabular} \caption{Fold-2 AdaMixUp: Mean $\alpha$, $\Delta$, and training losses.} \label{tab:accuracy:para} \end{table} \subsection{Training Characteristics} Figure~\ref{fig:alpha} depicts the behavior of Policy Region Generator in Fold-2 AdaMixUp over training iterations. It appears that the Policy Region Generator initially explores a wide range of policy space before settling down at around 40K iterations (when the means of $\alpha$ and $\Delta$ both stabilize, at 0.48 and 0.03 respectively). Figure~\ref{fig:weights} shows the average of the drawn mixing policy $\gamma$ in Fold-2 AdaMixUp during training iterations and the corresponding classification performance of the model. The range of $\gamma$ appears to stabilize at around 40K iterations (left plot), consistent with the observation in Figure \ref{fig:alpha}. Interestingly, also at the same time, the classification error drops and begins to stabilize (right plot). \begin{figure}[h] \caption{The mean of $\alpha$ (left) and the mean of $\Delta$ (right) in Fold-2 AdaMixUp on Cifar10.} \centering \includegraphics[width=3.5in,height=1in]{alpha11.eps \label{fig:alpha} \end{figure} \begin{figure}[h] \caption{Fold-2 AdaMixUp on Cifar10: Mean of mixing policy $\gamma$ (left) and training/testing error rates (right).} \centering \includegraphics[width=3.42in]{weight7.eps \label{fig:weights} \end{figure} Figures~\ref{fig:localization2} shows some typical mixed images in Fold-2 AdaMixUp and their corresponding original images in MNIST. We note that these mixed images obviously distinguish themselves from the original images. Thus they do not intrude into the data manifold. \begin{figure}[h] \caption{Mixed images (top row) in Fold-2 AdaMixUp from the original images (bottom 2 rows) in MNIST.} \centering \includegraphics[width=3.3062in]{newlanguage2.eps \label{fig:localization2} \end{figure} \subsection{Impact of Mixing Fold} Table~\ref{tab:accuracy:nfold} presents the error rates obtained by Fold-3 and Fold-4 AdaMixUp on the seven testing datasets (excluding ImageNet due to limited computational resources). We see that increasing the mixing fold from 2 to higher values, the performance of AdaMixUp improves. This is due to stronger constraints are imposed to the model and hence results in stronger regularization. One however expects that imposing stronger constraints makes the model more difficult to fit and further increasing mixing fold may eventually lead to diminishing gain and even under-fitting. This is hinted by Table \ref{tab:accuracy:nfoldcost}, where the averages of learned mixing parameters and the training losses of the Fold-3 AdaMixUp are presented. In Table ~\ref{tab:accuracy:nfoldcost}, we see that the intrusion losses of Fold-3 AdaMixUp are much higher than those of Fold-2 AdaMixUp (Table \ref{tab:accuracy:para}). This implies that in higher-fold AdaMixUp, it is more difficult for the Policy Region Generator to carve a good policy region that is not regarded by the Intrusion Discriminator as causing intrusion. This difficulty is also suggested by the high $\alpha_2$ values in Table \ref{tab:accuracy:nfoldcost} (often close to 1). They indicate that in these cases, mixing only occurs ``slightly''. \begin{table}[h] \centering \begin{tabular}{l|c|c|c}\hline AdaMixUp& Fold-2&Fold-3&Fold-4 \\\hline mnist&0.49&0.42&0.41 \\ fashion&6.21&5.88&5.71\\ svhn&3.12&2.98&2.91\\ cifar10&3.52&3.25&3.21\\ cifar100&20.97&20.68&20.29\\ cifar10-S&6.85&6.14&5.96\\ cifar100-S&26.72&25.72&25.49\\ \hline \end{tabular} \caption{Error rates (\%) of Fold-3 and Fold-4 AdaMixUp.} \label{tab:accuracy:nfold} \end{table} \begin{table}[h] \centering \resizebox{\columnwidth}{!}{ \begin{tabular}{l|c|c|c|c}\hline Fold-3& $\alpha1$/$\Delta1$&$\alpha2$/$\Delta2$&${\cal L}_{\rm intr}$&${\cal L}_{\cal D} + {\cal L}_{{\cal D}'}$\\\hline mnist&0.533/0.013&0.662/0.013&0.002&0.337\\ fashion&0.517/0.015&0.678/0.019&0.024&0.327\\ svhn&0.637/0.011&0.943/0.020&0.028&0.482\\ cifar10&0.516/0.028&0.938/0.041&0.003&0.403\\ cifar100&0.121/0.010&0.959/0.015&0.004&0.305\\ cifar10-S&0.504/0.039&0.849/0.068&0.028&0.505\\ cifar100-S&0.639/0.011&0.943/0.018&0.023&0.542\\ \hline \end{tabular}} \caption{Fold-3 AdaMixUp: average training losses and policy region parameters. $(\alpha_1, \Delta_1)$: initial Fold-2 mixing parameter; $(\alpha_2, \Delta_2)$: further mixing parameter.} \label{tab:accuracy:nfoldcost} \end{table} \subsection{Sensitivity to Model Capacity} We vary the number of filters on each layer of the ResNet-18 with half quarter, quarter, half, and three-quarter of the original number of filters (denoted as base filter), and present the test error rates on Cifar100 in Figure~\ref{filtersize} (green curves). The green curves in Figure~\ref{filtersize} indicate that the Fold-2 AdaMixUp benefits from large number of filters deployed: the accuracy keeps improving while enlarging the number of filters. On the contrary, the ResNet-18 received dramatically accuracy improvement before the number of filters is half of the base but obtained no further improvement after. \begin{figure}[h] \caption{Error rates obtained by ResNet-18 and AdaMixUp on Cifar100, when varying the number of filters (green curves) and varying the size of the samples (red curves). } \label{filtersize} \centering \includegraphics[height=1.69in]{filtersize8.eps \end{figure} \subsection{Effect of Data Size} We down sample the Cifar100 data with 20\%, 40\%, 60\% and 80\% of the training samples, and present the test error rates obtained by Fold-2 AdaMixUp in Figure~\ref{filtersize} (red curves). The red curves in Figure~\ref{filtersize} indicate that more training samples help with both the ResNet-18 and AdaMixUp, but the accuracy gap between the two methods are widening while increasing the training sample size. \subsection{Benefit of the Intrusion Discriminator} Table~\ref{tab:accuracy:resnet:nointrusion} lists the test error rates obtained by the Fold-2 AdaMixUp with ResNet-18 on the Cifar10 and Cifar100 data, but without the Intrusion Discriminator. \begin{table}[h] \centering \begin{tabular}{l|c|c|c}\hline Data Set& Baseline&Ada&AdaMixUp w/o\\ &ResNet-18&MixUp&Intru. Discr.\\ \hline cifar10&5.53&3.52&3.83\\ cifar100&25.6&20.97&24.75\\ \hline \end{tabular} \caption{Error rates obtained by Fold-2 AdaMixUp without the Intrusion Discriminator on the Cifar10 and Cifar100.} \label{tab:accuracy:resnet:nointrusion} \end{table} Table~\ref{tab:accuracy:resnet:nointrusion} shows that excluding the Intrusion Discriminator component, the AdaMixUp method was still able to improve the accuracy over the baseline, but obtained much lower accuracy than that of including the Intrusion Discriminator unit. \subsection{Interpolating on Hidden Layer} We also evaluate an alternative structure for AdaMixUp, where mixing happens on the layer before the final softmax layer in ResNet-18. Our results suggest that the test errors of the Fold-2 AdaMixUp increased dramatically due to the high cost of the Intrusion Discriminator. The error rates increased from 20.97\% to 22.21\% and from 3.52\% to 4.94\%, respectively for Cifar100 and Cifar10. Notably, both have large intrusion loss of around 0.49, which may due to the fact that perceptual similarity for images in the hidden embedding space may not as distinguishable as that in the input space. As a result, the Policy Generator failed to find good mixing policies to avoid colliding into the data manifold. \section{Related Work} Common data augmentation methods have been designed based on substantial domain knowledge~\cite{726791,SimonyanZ14a,AmodeiABCCCCCCD15,abs170804896}, relied on specific network architectures~\cite{Gastaldi17,DBLP:journals/corr/abs170804552,ttbabs18020237}, or leveraged feedback signals to search the optimal augmentation strategies~\cite{DBLP:journals/corr/LemleyBC17,DBLP:journals/corr/abs180509501}. Our method excludes those requirements, and only leverages a simple linear interpolation for data augmentation. Our work closely relates to approaches linearly interpolating examples and labels~\cite{mixup17,articleDeVries,ManifoldMixUp,DBLP:journals/corr/abs171110284}. Nevertheless, these approaches depends on the correct user-predefined mixing policies. Also, their interpolation typically lies along the lines of sample pairs. On the contrary, our approach automatically learns the mixing policy regions and benefits from mixing multiple images. \section{Conclusions and Outlook} This work justifies the effectiveness of MixUp from the perspective of out-of-manifold regularization. We also identify the inherent problem with standard MixUp in causing manifold intrusion. This motivates us to develop AdaMixUp, which generalizes MixUp to higher-fold mixing policies and automatically learns the policy regions that avoid manifold intrusion. To us, this work is only the beginning of a rich research theme. The justification of MixUp and AdaMixUp we provide thus far is primarily qualitative. It is desirable to quantitatively characterize the generalization capability of AdaMixUp. Moreover, the out-of-manifold regularization perspective potentially opens a door to new regularization techniques. Beyond local linearization, we believe that there is a rich space of other possibilities. \bibliographystyle{aaai}

01618aa7d7145c8acf56c4169bbd2aa44be23e62

\section{Supplemental material} \onecolumngrid \newpage

cc588aeff1091b12c24a9b04a8ec8061baab91c3

\section{Background} \subsection{Prior art, current pitfalls and desiderata} The beginning of a principled probabilistic treatment of the spectral estimation problem can be attributed to E.T.~Jaynes, who derived the discrete Fourier transform using Bayesian inference \cite{jaynes1987bayesian}. Then, G.L.~Bretthorst proposed to place a prior distribution over spectra and update it in the light of observed temporal data, for different time series models \cite{bretthorst2013bayesian}. This novel approach, in the words of P.C.~Gregory, meant a \emph{Bayesian revolution in spectral analysis} \cite{gregory_revolution}. The so developed conceptual framework paved the way for a plethora of methods addressing spectral estimation as (parametric) Bayesian inference. In this context, by choosing a parametric model for time series with closed-form Fourier transform, a Bayesian treatment provides error bars on the parameters of such a model and, consequently, error bars on the parametric spectral representation, e.g., \cite{djuric_harmonic,turner_sahani,qi_minka_picard}. Within Bayesian nonparametrics, the increasing popularity and ease of use of Gaussian processes (GP, \cite{Rasmussen:2006}), enabled \cite{protopapas,hensman} to detect periodicities in time series by (i) fitting a GP to the observed data, and then (ii) analysing the so learnt covariance kernel, or equivalently, its power spectral density (PSD). Although meaningful and novel, this GP-based method has a conceptual limitation when it comes to nonparametric modelling: though a nonparametric model is chosen for the time series, the model for the PSD (or kernel) is still only parametric. Bayesian nonparametric models for PSDs can be traced back to \cite{choudhuri_2004}, which constructed a prior directly on PSDs using Bernstein polynomials and a Dirichlet process, and more recently to \cite{tobar:nonparametric,npr_15b}, which placed a prior on covariance kernels by convolving a GP with itself. Yet novel, both these methodologies produced intractable posteriors for the PSDs, where the former relied on Monte Carlo methods and the latter on variational approximations. The open literature is lacking a framework for spectral estimation that is: \begin{itemize} \item Nonparametric, thus its complexity grows with the amount of data. \item Bayesian, meaning that it accounts for its own uncertainty. \item Tractable, providing exact solutions at low computational complexity. \end{itemize} We aim to fulfil these desiderata by modelling time series and their spectra, i.e., Fourier transform, using Gaussian processes. A key consequence of using GPs is that missing/unevenly-sampled observations are naturally handled. \subsection{The Fourier transform} \label{sub:back_fourier} Let us consider a signal, e.g., a time series or an image, defined by the function $f: \mathcal{X} \mapsto{\mathbb R}$, where for simplicity we will assume $ \mathcal{X} ={\mathbb R}$. The spectrum of $f(t)$ is given by its Fourier transform \cite{kay:88} \begin{equation} F(\xi) = \fourier{f}(\xi) \triangleq \int_ \mathcal{X} f(t)e^{-j2\pi\xi t} {\text dt} \label{eq:fourier_transform} \end{equation} where $j$ is the imaginary unit and the frequency $\xi$ is the argument of the function $F(\cdot)$. Notice that for $F(\xi)$ to exist, $f(t)$ is required to be Lebesgue integrable, that is, $\int_ \mathcal{X} |f(t)|{\text dt}<\infty$. Observe that $F(\xi)$ is the inner product between the signal $f(t)$ and the Fourier operator $e^{-j2\pi\xi t} = \cos(2\pi\xi t) - j\sin(2\pi\xi t)$, therefore, the complex-valued function $F(\xi)$ contains the frequency content of the even part (cf. odd part) of $f(t)$ in its real part (cf. imaginary part). We also refer to the square absolute value $S(\xi)=|F(\xi)|^2$, which comprises the total frequency content at frequency $\xi$, as the power spectral density (PSD). Calculating the integral in eq.~\eqref{eq:fourier_transform} is far from trivial for general Lebesgue-integrable signals $f(t)$. This has motivated the construction of parametric models for SE that approximate $f(\cdot)$ by analytic expressions that admit closed-form Fourier transform such as sum of sinusoids \cite{lomb}, autoregressive processes \cite{Walker518} and Hermite polynomials. The proposed method will be inspired in this rationale: we will use a stochastic-process model for the signal (rather than a parametric function), to then apply the Fourier transform to such process and finally obtain a stochastic representation of the spectrum. A family of stochastic processes that admit closed-form Fourier transform is presented next. \subsection{Gaussian process priors over functions} \label{sub:back_gp} The Gaussian process (GP \cite{Rasmussen:2006}) is the infinite-dimensional generalisation of the multivariate normal distribution. Formally, the stochastic process $f(t)$ is a GP if and only if for any finite collection of inputs $\{t_i\}_{i=1}^N,\ N\in{\mathbb {N}}$, the scalar random variables $\{f(t_i)\}_{i=1}^N$ are jointly Gaussian. A GP $f(t)$ with mean $m$ function and covariance kernel $K$ will be denoted as \begin{equation} f(t)\sim{\mathcal {GP}}(m,K) \end{equation} where we usually assume zero (or constant) mean, and a kernel function $K(t,t')$ denoting the covariance between $f(t)$ and $f(t')$. The behaviour of the GP is encoded in its covariance function, in particular, if the GP $f(t)$ is stationary, we have $K(t,t')=K(t-t')$ and the PSD of $f(t)$ is given by $S(\xi) = \fourier{K(t)}(\xi)$ \cite{salomon}. The connection between temporal and frequency representations of GPs has aided the design of the GPs to have specific (prior) harmonic content in both parametric \cite{lazaro2010sparse,Wilson:2013,CSM,parra_tobar,hensman2016variational} and non-parametric \cite{tobar:nonparametric,npr_15b} ways. GPs are flexible nonparametric models for functions, in particular, for latent signals involved in SE settings. Besides their strength as a generative model, there are two key properties that position GPs as a sound prior within SE: first, as the Fourier transform is a linear operator, the Fourier transform of a GP (if it exists) is also a (complex-valued) GP \cite{icassp15,8307269} and, critically, the signal and its spectrum are jointly Gaussian. Second, Gaussian random variables are closed under conditioning and marginalisation, meaning that the exact posterior distribution of the spectrum conditional to a set of partial observations of the signal is also Gaussian. This turns the SE problem into an inference one with two new challenges: to find the requirements for the existence of the spectrum of a GP, and to calculate the statistics of the posterior spectrum given the (temporal) observations. \section{A joint generative model for signals and their spectra} \label{sec:contrib} The proposed model is presented through the following building blocks: (i) a GP model for the latent signal, (ii) a windowed version of the signal for which the Fourier transform exists, (iii) the closed-form posterior distribution of the windowed-signal spectrum, and (iv) the closed-form posterior power spectral density. \subsection{Definition of the local spectrum} \label{sec:local_spectrum} We place a stationary GP prior over $f(t)\sim{\mathcal {GP}}(0,K)$ and model the observations as evaluations of $f(t)$ corrupted by Gaussian noise, denoted by ${\mathbf y}=[y(t_i)]_{i=1}^N$. This GP model follows the implicit stationarity assumption adopted when computing the spectrum via the Fourier transform. However, notice that the draws of a stationary GP are not Lebesgue integrable almost surely (a.s.) and therefore their Fourier transforms do not exist a.s. \cite{royden1968real}. We avoid referring to the spectrum of the complete signal and only focus on the spectrum in the neighbourhood of a centre $c$. Then, we can then choose an arbitrarily-wide neighbourhood (as long as it is finite), or consider multiple centres $\{c_i\}_{i=1}^{N_c}$ to form a \emph{bank of filters}. We refer to the spectrum in such neighbourhood as the \emph{local spectrum} and define it through the Fourier transform as \begin{equation} F_c(\xi) \triangleq \fourier{f_c(t)} = \fourier{f(t-c)e^{-\alpha t^2}} \end{equation} where $f_c(t) = f(t-c)e^{-\alpha t^2}$ is a windowed version of the signal $f(t)$ centred at $c$ with width $1/\sqrt{2\alpha}$. Observe that since $f_c(t)$ decays exponentially for $t\rightarrow\pm\infty$, it is in fact Lebesgue integrable: \begin{equation} \int_{\mathbb R} |f_c(t)|{\text dt} = \int_{\mathbb R} |f(t-c)e^{-\alpha t^2}|{\text dt}<\max(|f|)\int_{\mathbb R} e^{-\alpha t^2}{\text dt}=\max(|f|)\sqrt{\frac{\pi}{\alpha}} <\infty\quad \text{a.s.} \end{equation} since the $\max(|f|)$ is finite a.s. due to the GP prior. As a consequence, the local spectrum $\fourierlocal{f(t)}$ exists and it is finite. The use of windowed signals is commonplace in SE, either as a consequence of acquisition devices or for algorithmic purposes (as in our case). In fact, windowing allows for a time-frequency representation, meaning that the signal does not need to be stationary but only piece-wise stationary, i.e., different centres $c_i$ might have different spectra. Finally, we clarify that the choice of a square-exponential window $e^{-\alpha t^2}$ obeys to tractability of the statistics calculated in the next section. A summary of the proposed generative model is shown in eqs.~\eqref{eq:GP}-\eqref{eq:windowed_F} and a graphical model representation is shown in fig.~\ref{fig:gen_mod}. \begin{align} \text{latent signal:}\quad f(t)&\sim{\mathcal {GP}}(0,K)\label{eq:GP}\\ \text{observations:}\quad y(t_i)&=f(t_i) + \eta_i, \eta_i\sim{\mathcal {N}}(0,\sigma_n^2),\forall i=1\ldots N,\label{eq:obs}\\ \text{windowed signal:}\quad f_{c} (t) &= e^{-\alpha t^2}f(t-c)\label{eq:windowed_f}\\ \text{local spectrum:}\quad F_c(\xi) &\triangleq \fourier{f_c(t)} =\fourier{f(t-c)e^{-\alpha t^2}} =\int_{\mathbb R} f(t-c)e^{-\alpha t^2}e^{-j2\pi\xi t} {\text dt} \label{eq:windowed_F} \end{align} \begin{figure}[H] \vspace{-0.75em} \centering \includegraphics[width=0.6\textwidth]{img/gen_mod}\\ \caption{Proposed model for a latent signal $f(t)$, observations $y(t)$, a windowed version $f_c(t)$ and local spectrum $F_c(\xi)$. We have considered $N$ observations and $C$ centres.} \label{fig:gen_mod} \end{figure} \subsection{The local-spectrum Gaussian process} \label{sub:spectrum_kernels} As a complex-valued linear transformation of $f(t)\sim{\mathcal {GP}}$, the local spectrum $F_c(\xi)$ is a complex-GP \cite{8307269,icassp15} and thus completely determined by its covariance and pseudocovariance \cite{mandic09} given by \begin{align} K_F(\xi,\xi') &= \E{F_c(\xi)F_c^*(\xi)} = \E{F_c(\xi)F_c(-\xi')}\label{eq:K_f}\\ P_F(\xi,\xi') &= \E{F_c(\xi)F_c(\xi')} = K_F(\xi,-\xi')\label{eq:P_f} \end{align} where the last identities in each line are due to the fact that the latent function $f(t)$ is real valued. Recall that we are ultimately interested in the real and imaginary parts of the local spectrum ($\Re F_c(\xi)$ and $\Im F_c(\xi)$ respectively) which are in fact real-valued GPs. However, we will calculate the statistics of the complex-valued $F_c(\xi)$ for notational simplicity, to then calculate the statistics of the real-valued processes $\Re F_c(\xi)$ and $\Im F_c(\xi)$ according to: \begin{align} \text{covariance}(\Re F_c(\xi))=K_{rr}(\xi,\xi') &= \tfrac{1}{2}(K_F(\xi,\xi')+K_F(\xi,-\xi'))\label{eq:Krr}\\ \text{covariance}(\Im F_c(\xi))=K_{ii}(\xi,\xi') &= \tfrac{1}{2}(K_F(\xi,\xi')-K_F(\xi,-\xi'))\label{eq:Kii}\\ \text{covariance}(\Re F_c(\xi),\Im F_c(\xi))=K_{ri}(\xi,\xi') &= K_{ir}(\xi,\xi') = 0.\label{eq:Kri} \end{align} The above expressions are due to the identity in eq.~\eqref{eq:P_f} and the fact that both $K_F(\xi,\xi')$ and $K_F(\xi,-\xi')$ are real-valued. The relationship between the covariance of a GP and the covariance of the spectrum of such GP is given by the following proposition \begin{proposition} The covariance of the local spectrum $F_c(\xi)$ of a stationary signal $f(t)\sim{\mathcal {GP}}(0,K(t))$ is given by \begin{equation}% K_F (\xi,\xi') = \sqrt{\frac{\pi}{2\alpha}} e^{-\frac{\pi^2}{2\alpha}(\xi-\xi')^2}\left({\mathcal {K}}(\rho)\ast \sqrt{\frac{2\pi}{\alpha}}e^{-\frac{2\pi^2}{\alpha}\rho^2}\right)\bigg|_{\rho= \frac{\xi+\xi'}{2}}\label{eq:prior_cov} \end{equation} where ${\mathcal {K}}(\xi)=\fourier{K(t)}(\xi)=\int_{{\mathbb R}} K(t)e^{-j2\pi\xi t} {\text dt}$ is the Fourier transform of the kernel $K$. Equivalently, as pointed out in eq.~\eqref{eq:P_f}, the pseudocovariance is given by replacing the above expression in $P_F(\xi,\xi')=K_F (-\xi,\xi')$. \end{proposition} See the proof in Section 1.1 of the supplementary material. Notice that the covariance of the local spectrum $K_F$ is a sequence of linear transformations of the covariance of the signal $K$ according to: (i) the Fourier transform due to the domain change, (ii) convolution with $ e^{-\frac{2\pi^2}{\alpha}\rho^2}$ due to windowing effect, and (iii) a smoothness factor $e^{-\frac{\pi^2}{2\alpha}(\xi-\xi')^2}$ that depends on the window width; this means that for wider windows the values of the local spectrum at different frequencies become independent. Critically, observe that each of the Gaussian functions in eq.~\eqref{eq:prior_cov} are divided by their normalising constants, therefore the norm of $K_F$ is equal to the norm ${\mathcal {K}}$, which is in turn equal to the norm of the covariance of the signal $K$ due to the unitary property of the Fourier transform. With an illustrative purpose, we evaluate $K_F$ for the $Q$-component spectral mixture (SM) kernel \cite{Wilson:2013} \begin{equation} K_{\text{SM}} (\tau)=\sum_{q=1}^Q\sigma^2_q \expo{-\gamma_q\tau^2}\cos(2\pi\theta^\top_q \tau) \label{eq:SM} \end{equation} the Fourier transform of which is known explicitly and given by \begin{align} {\mathcal {K}}_{\text{SM}}(\xi) = \sum_{q=1}^Q\sigma^2_q \sqrt{\frac{\pi}{\gamma_q}}\left(\frac{e^{-\frac{\pi^2}{\gamma_q}(\xi-\theta_q)^2}+e^{-\frac{\pi^2}{\gamma_q}(\xi+\theta_q)^2}}{2}\right) = \sum_{q=1}^Q\sum_{\theta=\pm\theta_q}\frac{\sigma^2_q}{2} \sqrt{\frac{\pi}{\gamma_q}}e^{-\frac{\pi^2}{\gamma_q}(\xi-\theta_q)^2}. \label{eq:SMF} \end{align} For this SM kernel, the covariance kernel of the local-spectrum process is (see supp. mat., \textsection1.2) \begin{align} K_F (\xi,\xi') = \sum_{q=1}^Q\sum_{\theta=\pm\theta_q} \frac{\sigma^2_q \pi}{2\sqrt{\alpha(\alpha+2\gamma_q)}} e^{-\frac{\pi^2}{2\alpha}(\xi-\xi')^2}e^{-\frac{2\pi^2}{\alpha+2\gamma_q} \left(\frac{\xi+\xi'}{2}-\theta_q\right)^2}. \label{eq:K_pectrum_SM} \end{align} With the explicit expression of $K_F$ in eq.~\eqref{eq:K_pectrum_SM} and the relationships in eqs.~\eqref{eq:K_f}-\eqref{eq:Kri}, we can compute the statistics of the real and imaginary parts of the local spectrum and sample from it. Fig.~\ref{fig:spec_cov_paths} shows these covariances and 3 sample paths revealing the odd and even properties of the covariances. \begin{figure}[H] \centering \includegraphics[width=\textwidth]{img/spec_cov_paths.pdf} \caption{Covariance and sample paths of the local-spectrum of a SM signal with $Q=1, \sigma_q=1, \gamma_q=5e-3 ,\theta_q=2.5,\alpha=5e-5$. Real (cf. imaginary) part shown in the left (cf. right) half.} \label{fig:spec_cov_paths} \end{figure} \subsection[The posterior density of the spectrum]{Joint samples and the conditional density $p(F_c(\xi)|{\mathbf y})$} \label{sub:posterior_spectrum} Although the joint distribution over the signal $f(t)$ and its local spectrum $ \mathcal{F} _c(\xi)$ is Gaussian, sampling directly from this joint distribution is problematic due to the deterministic relationship between the (complete and noiseless) signal $f$ and its local spectrum. We thus proceed hierarchically: we first sample ${\mathbf y}\sim {\mathcal {GP}}({\mathbf t};0,K)$, ${\mathbf y}\in{\mathbb R}^N$, and then $ \mathcal{F} _c(\xi)\sim p( \mathcal{F} _c|{\mathbf y})$, where the posterior is normally-distributed with mean and covariance given respectively by \begin{align} \E{F_c(\xi)|{\mathbf y}} &= K^\top_{{\mathbf y} F_c}({\mathbf t},\xi)K({\mathbf t},{\mathbf t})^{-1}{\mathbf y}\label{eq:post_mean}\\ \E{F_c^*(\xi)F_c(\xi')|{\mathbf y}} &= K_{F}(\xi,\xi') - K^\top_{{\mathbf y} F_c}({\mathbf t},\xi)K({\mathbf t},{\mathbf t})^{-1}K_{{\mathbf y} F_c}({\mathbf t},\xi)\label{eq:post_var} \end{align} where $K_{{\mathbf y} F_c}({\mathbf t},\xi)$ is presented in the next proposition. \begin{proposition} The covariance $K_{{\mathbf y} F_c}({\mathbf t},\xi)$ between the observations ${\mathbf y}$ at times ${\mathbf t}$ coming from a stationary signal $f(t)\sim{\mathcal {GP}}(0,K)$ and its local spectrum at frequency $\xi$ is given by \begin{equation}% K_{y F_c}(t,\xi) = \E{y^*_c(t) F_c(\xi)} ={\mathcal {K}}(\xi)e^{-j2\pi\xi t} \ast\sqrt{\frac{\pi}{\alpha}}e^{-\frac{\pi^2\xi^2}{\alpha}}\label{eq:xcov} \end{equation} where ${\mathcal {K}}(\xi)=\fourier{K(t)}(\xi)=\int_{{\mathbb R}} K(t)e^{-j2\pi\xi t} {\text dt}$ is the Fourier transform of the kernel $K$. \end{proposition} See the proof in Section 1.3 of the supplementary material. Notice that the convolution against $e^{-\frac{\pi^2\xi^2}{\alpha}}$ is also due to the windowing effect and that the norms of $K_{y F_c}$ and $K$ are equal. For the SM kernel, shown in eq.\eqref{eq:SM}, $K_{y F_c}$ becomes (details in supp. mat., \textsection1.4) \begin{align*} K_{y\mathcal{F}} =\sum_{q=1}^Q\sum_{\theta=\pm\theta_q} \frac{\sigma^2_q}{2\sqrt{\pi ( \tilde{\alpha} + \tilde{\gamma} _q)}} \expo{ -\frac{(\xi-\theta_q)^2}{ \tilde{\alpha} + \tilde{\gamma} _q}} \expo{ - \frac{\pi^2t^2}{L_q}} \expo{- j\frac{2\pi t}{L_q}\left(\frac{\theta_q}{ \tilde{\gamma} _q} + \frac{\xi}{ \tilde{\alpha} }\right) } \end{align*} where $ \tilde{\alpha} =\alpha/\pi^2$, $ \tilde{\gamma} _q=\gamma_q/\pi^2$ and $L_q=( \tilde{\alpha} ^{-1} + \tilde{\gamma} _q^{-1})^{-1}$. Fig.~\ref{fig:time_freq} shows this covariance together with joint samples of the signal and its spectrum (colour-coded). \begin{figure}[H] \centering \includegraphics[width=\textwidth]{img/time_freq.pdf} \caption{Hierarchical sampling. From left to right: Signal samples (solid) and window (dashed), covariance $K_{y\mathcal{F}}$ for the SM, real-part local-spectrum samples, imaginary-part local-spectrum. Parameters were $Q=1, \sigma_q=1, \gamma_q=2 ,\theta_q=2.5,\alpha=1$. Notice how $K_{y F_c}(t,\xi)$ vanishes as the frequency $\xi$ departs from $\theta_q$. } \label{fig:time_freq} \end{figure} We conclude this section with the following result. \begin{proposition} The power spectral density of a stationary signal $f(t)\sim{\mathcal {GP}}(0,K)$, conditional to a set of observations ${\mathbf y}$, is a $\chi^2$-distributed stochastic process and its mean is known in closed form. \end{proposition} This result follows from the fact the (posterior) real and imaginary parts of the spectrum are independent Gaussian process with explicit mean and covariance. This is a critical contribution of the proposed model, where the search for periodicities can be performed by optimising a closed-form expression which has a linear evaluation cost. \section{Discussion} We have proposed a nonparametric model for spectral estimation (SE), termed BNSE, and have shown that it admits exact Bayesian inference. BNSE builds on a Gaussian process (GP) prior over signals and its relationship to existing methods in the SE and GP literature has been illuminated from both theoretical and experimental perspectives. To the best of our knowledge, BNSE is the first nonparametric approach for SE where the representation of uncertainty related to missing and noisy observations is inherent (due to its Bayesian nature). Another unique advantage of BNSE is a nonparametric functional form for the posterior power spectral density (PSD), meaning that periodicities can be found through linear-cost optimisation of the PSD rather than by exhaustive search or expensive non-convex optimisation routines. We have shown illustrative examples and results using time series and exponential kernels, however, the proposed BNSE is readily available to take full advantage of GP theory to consider arbitrary kernels in multi-input, multi-output, nonstationary and even non-Gaussian applications. The promising theoretical results also open new avenues in modern SE, this may include novel interpretations of Nyquist frequency, band-pass filtering and time-frequency analysis. \section{Simulations} \label{sec:simulations} This experimental section contains three parts focusing respectively on: (i) consistency of BNSE in the classical sum-of-sinusoids setting, (ii) robustness of BNSE to overfit and ability to handle non-uniformly sampled noisy observations (heart-rate signal), and (iii) exploiting the functional form of the PSD estimate of BNSE to find periodicities (astronomical signal). \subsection{Identifying line spectra} \label{sub:line} Signals composed by a sum of sinusoids have spectra given by Dirac delta functions (or vertical lines) referred to as \emph{line spectra}. We compared BNSE against classic line spectra models such as MUSIC \cite{GJI:GJI347}, Lomb-Scargle \cite{lomb} and the Periodogram \cite{periodogram}. We considered 240 evenly-sampled observations of the signal $f(t)=10\cos(2\pi 0.5t)-5\sin(2\pi 1.0t)$ in the domain ${\mathbf t}\in[-10,10]$ corrupted by zero-mean unit-variance Gaussian noise. The window parameter was set to $\alpha=1/(2\cdot 50^2)$ for an observation neighbourhood much wider than the support of the observations, and we chose an SM kernel with rather permissive hyperparameters: a rate $\gamma=1/(2\cdot 0.05^2)$ and $\theta=0$ for a prior over frequencies virtually uninformative. Fig.~\ref{fig:cosines} shows the real and imaginary parts of the posterior local spectrum and the sample PSD against LS, MUSIC, and the Periodogram. Notice how BNSE recovered the spectrum with tight error bars and appropriate relative magnitudes. Additionally, from the PSD estimates notice how both BNSE and LS coincided with the periodogram and MUSIC at the peaks of the PSD. Finally, observe that in line with the structural similarities between BNSE and LS, they both exhibit the same lobewidths and that LS falls within the errorbars of BNSE. \begin{figure}[H] \centering \includegraphics[width=\textwidth]{img/cosines.pdf} \caption{Line spectrum estimates: BNSE is shown in red and its PSD is computed by first sampling form the real and imaginary parts of the posterior spectrum and then adding the square samples (LS: Lomb-Scargle and pgram: periodogram).} \label{fig:cosines}s \end{figure} \subsection{Discriminating between heart-rate signals} \label{sub:hr} We next considered two heart-rate signals from \url{http://ecg.mit.edu/time-series/}. The first one is known to have frequency components at the respiration rate of the subject, whereas the second one exhibits low-frequency energy which may be attributed to congestive heart failure \cite{Goldberger}. To show that the proposed method does not overfit to the spectrum of the training data, we used the first signal to train BNSE and then used BNSE to analyse the posterior PSD of the second signal. To make the experiment more realistic, we only used an unevenly-sampled 10\% of the data from the second (test) signal and considered the LS method with the entire (noiseless) signal as ground truth. Fig.~\ref{fig:heart_rate} shows the PSDs for both signals and methods. Observe that in both cases, BNSE's posterior PSD distribution includes the ground truth (LS), even for the previously-unseen test signal. Crucially, this reveals that BNSE can be used for SE beyond the training data to find critical harmonic features from noisy and limited observations. \begin{figure}[H] \centering \includegraphics[width=0.48\textwidth]{img/hr_train.pdf} \includegraphics[width=0.48\textwidth]{img/hr_test.pdf} \caption{PSD estimates for heart-rate time series. Notice how BNSE recovered the spectral content of the test signal from only a few noisy measurements} \label{fig:heart_rate} \end{figure} \subsection{Finding periodicities via efficient optimisation} \label{sub:optimisation} Lastly, we considered the sunspots dataset, an astronomical time series that is known to have a period of approximately $11$ years, corresponding to a fundamental frequency of $1/11\approx0.089$. Finding this period is challenging due to the nonstationarity of the signal. We implemented BNSE, Lomb-Scargle and a GP with spectral mixture kernel \cite{Wilson:2013} to find the fundamental frequency of the series. Satisfactory training for Lomb-Scargle and the SM kernel was not possible via gradient-based maximum likelihood (we used GPflow \cite{GPflow:2016}), even starting from the neighbourhood of the true frequency ($0.089$) or using minibatches. Our conjecture is that this is due to the fact that the sunspots series is neither strictly periodic nor Gaussian. We implemented BNSE with a lengthscale equal to one and $\theta=0$ for a broad prior over frequencies, and $\alpha=10^{-3}$ for a wide observation neighbourhood. Finally, the posterior mean of the PSD reported by BNSE was maximised using the derivative-free Powell method \cite{powell1964efficient} due to its non-convexity. Notice that optimising the PSD of BNSE with Powell has a computational cost that is linear in the number observations and dimensions, whereas maximising SM via maximum likelihood has a cubic cost in the observations. Fig.~\ref{fig:sunspots} shows the estimates of the PSD for BNSE and LS (recall that SM was not able to train) and their maxima, observe how the global maximum of the PSD estimated by BNSE is the true period $\approx0.089$. \begin{figure}[H] \centering \includegraphics[width=\textwidth]{img/sunspots.pdf} \caption{Finding periodicities via optimisation. Left: sunspots data. Right: PSDs estimates reported by BNSE (red) and LS (blue) with corresponding maxima in vertical dashed lines. The correct fundamental frequency of the series is approximately $1/11\approx0.089$.} \label{fig:sunspots} \end{figure} \section{Introduction} The need for frequency representation arises naturally in a number of disciplines such as natural sound processing \cite{turner:2010,ballenas}, astrophysics \cite{Huijse}, biomedical engineering \cite{ieee5605664} and Doppler-radar data analysis \cite{doppler}. When the signal of interest is known without uncertainty, the frequency representation can be obtained by means of the Fourier transform \cite{kay:88}. However, real-world applications usually only provide us with a limited number of observations corrupted by noise. In this sense, the main challenge in Spectral Estimation (SE) comes from the fact that, due to the convolutional structure of the Fourier transform, the uncertainty related to missing, noisy and unevenly-sampled data propagates across the entire frequency domain. In this article, we take a probabilistic perspective to SE, thus aiming to quantify uncertainty in a principled manner. Classical---yet still widely used---methods for spectral estimation can be divided in two categories. First, parametric models that impose a deterministic structure on the latent signal, which result in a parametric form for the spectrum \cite{GJI:GJI347,lomb, Walker518}. Second, nonparametric models that do not assume structure in the data, such as the periodogram \cite{periodogram} computed through the Fast Fourier Transform (FFT) \cite{1965-cooley}. Uncertainty is not inherently accounted for in either of these approaches, although one can equip parameter estimates with error bars in the first case, or consider subsets of training data to then average over the estimated spectra. Despite the key role of the frequency representation in various applications as well as recent advances in probabilistic modelling, the Bayesian machinery has not been fully exploited for the construction of rigorous and meaningful SE methods. In particular, our hypothesis is that Bayesian nonparametric models can greatly advance SE theory and practice by incorporating temporal-structure parameter-free generative models, inherent uncertainty representation, and a natural treatment of missing and noisy observations. Our main contribution is then to propose a nonparametric joint generative model for a signal and its spectrum, where SE is addressed by solving an exact inference problem. \section{Spectral estimation as Bayesian inference} \label{sub:spectral_estimation_as_bayesian_inference} Henceforth, the proposed method for Bayesian nonparametric spectral estimation will be referred to as BNSE. This section analyses BNSE in terms of interpretability, implementation, and connection with other methods. \subsection{Training and computational cost} \label{sub:cost} BNSE can be interpreted as fitting a continuous-input interpolation to the observations, computing the Fourier transform of the interpolation and finally average over all the possibly infinitely-many interpolations. Consequently, as our interpolation is a GP, both the Fourier transform and the infinite average can be performed analytically. Within BNSE, finding the appropriate interpolation family boils down to selecting the model hyperparameters, where the GP prior protects the model from overfitting \cite{Rasmussen:2006}. In this regard, the proposed BNSE can readily rely upon state-of-the-art training procedures for GPs and benefit from sparse approximations for computationally-efficient training. Finally, as the hyperparameters of the posterior spectrum are given by those of the GP in the time domain, computing the posterior local spectrum poses no additional computational complexity. \subsection{Model consistency and interpretation} \label{sub:model_intepretation} The problem of global (rather than local) SE can be addressed by choosing an arbitrarily-wide window. However, as pointed out in Section \ref{sec:local_spectrum} recall that the local-spectrum process is not defined for $\alpha\rightarrow 0$, since it turns into the sum of infinitely-many Gaussian RVs; in fact, note from eq.~\eqref{eq:prior_cov} that $\lim_{\alpha\rightarrow 0}K_F(\xi,\xi')=\infty$. Despite the lack of convergence for the posterior law of the spectrum when $\alpha\rightarrow 0$, let us only consider the point estimate as the posterior mean defined from eqs.~\eqref{eq:post_mean} and \eqref{eq:xcov} as \begin{equation} \E{ \mathcal{F} _c(\xi)|{\mathbf y}} = \left({\mathcal {K}}(\xi)e^{-j2\pi\xi t} \ast\sqrt{\frac{\pi}{\alpha}}e^{-\frac{\pi^2\xi^2}{\alpha}}\right)K({\mathbf t},{\mathbf t})^{-1}{\mathbf y} \label{eq:explicit_post_mean} \end{equation} Observe that we can indeed apply the limit $\alpha\rightarrow 0$ above, where the second argument of the convolution converges to a (unit-norm) Dirac delta function. Additionally, let us consider an uninformative prior over the latent signal by choosing $K({\mathbf t},{\mathbf t})=\mathbf{I}$, which implies ${\mathcal {K}}(\xi)= 1$. Under these conditions (infinitely-wide window and uninformative prior for temporal structure in the signal) the point estimate of the proposed model becomes the discrete-time Fourier transform. \begin{equation} \lim_{\alpha\rightarrow 0}\E{ \mathcal{F} _c(\xi)|{\mathbf y}} = e^{-j2\pi\xi {\mathbf t}}{\mathbf y} = \sum_{i=1}^N e^{-j2\pi\xi t_i}y(t_i). \end{equation} This reveals the consistency of the model and offers a clear interpretation of the functional form in eq.~\eqref{eq:explicit_post_mean}: the posterior mean of the local spectrum is a linear transformation of a whitened version of the observations that depends on the width of the window and the prior belief over frequencies. \subsection{Approximations for non-exponential covariances} \label{sub:non_exp} Though Sec.~\ref{sec:contrib} provides explicit expressions of the posterior local-spectrum statistics for the spectral mixture kernel \cite{Wilson:2013}, the proposed method is independent of the stationary kernel considered. For general kernels with known Fourier transform but for which the convolutions in eqs.~\eqref{eq:prior_cov} and \eqref{eq:xcov} are intractable such as the Sinc, Laplace and Matérn kernels \cite{duvenaud2014automatic}, we consider the following approximation for $\alpha$ sufficiently small \begin{align} K_F (\xi,\xi') &= \frac{\pi}{\alpha} e^{-\frac{\pi^2}{2\alpha}(\xi-\xi')^2}\left({\mathcal {K}}(\rho)\ast e^{-\frac{2\pi^2}{\alpha}\rho^2}\right) \bigg|_{\rho= \frac{\xi+\xi'}{2}} \approx \sqrt{\frac{\pi}{2\alpha}} e^{-\frac{\pi^2}{2\alpha}(\xi-\xi')^2}{\mathcal {K}}\left(\frac{\xi+\xi'}{2}\right) \label{eq:prior_cov_app} \\ K_{y_c \mathcal{F} _c}(t,\xi) &={\mathcal {K}}(\xi)e^{-j2\pi\xi t} \ast\sqrt{\frac{\pi}{\alpha}}e^{-\frac{\pi^2\xi^2}{\alpha}} \approx {\mathcal {K}}(\xi)e^{-j2\pi\xi t}\label{eq:xcov_approx} \end{align} where we approximated the second argument in both convolutions as a Dirac delta as in Sec.~\ref{sub:model_intepretation}. We did not approximate the term $\sqrt{\frac{\pi}{2\alpha}} e^{-\frac{\pi^2}{2\alpha}(\xi-\xi')^2}$ in eq.~\eqref{eq:prior_cov_app} since placing a Dirac delta outside a convolution will result on a degenerate covariance. We emphasise that this is an approximation for numerical computation and not applying the limit $\alpha\rightarrow 0$, in which case BNSE does not converge. \subsection{Proposed model as the limit of the Lomb-Scargle method} \label{sub:lomb_scargle} The Lomb-Scargle method (LS) \cite{lomb} is the \emph{de facto} approach for estimating the spectrum of nonuniformly-sampled data. LS proceeds by fitting a set of sinusoids via least squares to the observations and then reporting the estimated spectrum as the weights of the sinusoids. The proposed BNSE method is closely related to the LS method with clear differences: (i) we assume a probabilistic model (the GP) which allows for the spectrum to be stochastic, (ii) we assume a nonparametric model which expressiveness increases with the amount of data, (iii) BNSE is trained once and results in a functional form for $ \mathcal{F} _c(\xi)$, whereas LS needs to be retrained should new frequencies be considered, (iv) the functional form $ \mathcal{F} _c(\xi)$ allows for finding periodicities via optimisation, while LS can only do so through exhaustive search and retraining in each step. In Section 2 of the supplementary material, we show that the proposed BNSE model is the limit of the LS method when an infinite number of components is considered with a Gaussian prior over the weights. \subsubsection*{Acknowledgments} This work was funded by the projects Conicyt-PIA \#AFB170001 Center for Mathematical Modeling and Fondecyt-Iniciación \#11171165. \newpage {\small \bibliographystyle{unsrtnat}

ea92b6b80476f0f04badf682e32312dabba6b4c1

\section{Device details} \subsection{Tunnel field-effect transistor} The tunnel field-effect transistor (TFET) is a kind of metal-oxide-semiconductor field-effect transistor (MOSFET) that has an $n -type source and a $p$-type drain electrodes, effectively working as a gated PIN diode (a diode with an undoped intrinsic semiconductor region between a p$-type semiconductor and an $n$-type semiconductor region). Its channel, kept intrinsic for zero gate voltage $V_{G}$, can be tuned into $p$-($n -)~type for large enough positive (negative) $V_{G}$. A TFET is tuned on by gate-induced reduction of the PIN junction thickness, enabling steeper switchings than MOSFET. Recently, it has been intensively studied as a future low-power transistor element for large-scale integration \cit {Ionescu11}. Enhancement of the on-current is achieved by introducing deep impurities in the (relatively long-channel) TFETs, and this enhancement is ascribed to deep-level assisted resonant tunneling in the PIN junction. In order to electrically access a single deep impurity, and use its spin for a high-temperature qubit, we introduce deep impurities in a short-channel TFET. Tunneling transport through a deep impurity level as well as the gate tuning of the level are possible in short-channel TFET provided that the deep impurity is located appropriately in the channel. In contrast to a MOSFET, the impurity-electrode tunnel coupling can be in a reasonable range of the TFET for realistic channel lengths (several tens of nm), even when the deepest level is located in the middle of the band gap. Our TFET-based devices are fabricated with a process compatible with those for standard MOSFETs. Starting from silicon-on-insulator wafers, $n$-type electrodes (followed by $p$-type electrodes) are defined by standard ion implantations of shallow donors (shallow acceptors). Then, we lay both Al and N by the ion implantations in the whole area including source, channel, and drain, and follow by appropriate heat treatment. This process is known to form coupled Al-N impurity pairs in Si \cite{weber1980localized, sauer1984nitrogen, modavis1990aluminum, iizuka2015first}. We found this is crucial for introducing deep impurity levels to TFETs. Indeed, if we omit this process, no TFETs (including short-channel ones) show the quantum-dot-like transport as described below, but only conventional characteristics of TFETs. Finally, the gate electrodes are formed with standard high-k/metal gate technology. Some of our devices show characteristics similar to a double dot, which is formed when two quantum dots are connected in series between source/drain electrodes. Measurements of Coulomb diamonds, Coulomb conductance peaks, and their temperature dependence suggest the formation of multiple dots in the device, composed of a deep impurity with strong confinement ($>0.1$~eV, a Al-N deep level) and at least one satellite dot nearby the deep impurity with weaker confinement ($\sim 5$-$10$~meV), which is probably a shallow acceptor located close to the $p$-type drain electrode. Thus, interdot level detuning and interdot tunnel coupling are not tunable but nearly fixed. However, there is a little tunability for the tunnel coupling between the dots and the electrodes (leads). Because the space charge layer of the PIN structure depends on the $V_{SD}$, i.e. thicker for negative V_{SD}$ and vice versa. The tunnel coupling between the double dots and electrodes is smaller for negative $V_{SD}$ (and vice versa). We have checked the electron spin resonance (ESR) response of $I_{SD}$ for various sets of $(V_{SD},V_{G})$ in the range of $5$-$10$~mV, and identified the spin blockade region in the plane of $(V_{SD},V_{G})$. For the device used in the main text, we observed the ESR spectra with two resonant lines with $g$-factors equal to $2.3$ and $2.7$. The peak of the ESR with the $g -factor $2.7$ is weak. \subsection{High-temperature spin qubit} In Fig.~\ref{FigS1} we present the schematic of the single-electron tunneling cycle in the spin-blockade regime. This is the same scheme as Fig.~1(c) in the main text. Here we describe this in more detail. Consider the initial situation in (i) with one electron in the right dot. The localized energy level on the left (closer to the $n$-type source electrode) is the electron-number $N=1$ state of the deep impurity, while the right two energy levels are those of the $N=1$ and $N=2$ states (separated by the on-site Coulomb energy $E_{\mathrm{C}}\sim 10$~meV) of the shallow impurity, respectively. The Fermi energy of the $p$-type electrode sits between the two right states with a thermal broadening $k_{\mathrm{B}}T . The tunneling cycle occurs following the dark gray arrows, with equal probabilities for the (i)$\rightarrow $(ii) and (i)$\rightarrow $(ii') trajectories, and eventually is blocked at the parallel-spin state (ii'). Under the ESR condition of the right spin, the spin blockade is lifted, as shown in the inset, and the source-drain current is increased due to the newly opened cycle, as indicated with the light gray arrows: (i)$\rightarrow $(ii')$\rightarrow $(iii')$\rightarrow $(iii)$\rightarrow $(i). Note that the spin qubit energy levels have the separation $\Delta E\sim 9$~GHz$\cdot h $ at $B\sim 1$~T and the qubit transition is described as flipping of the spin at (ii')$\rightarrow $(iii'). In the spin blockade condition in our experiment, the energy levels of each dot are placed at a position that is not affected by the thermally excited Fermi distribution of the source/drain electrodes. Under these circumstances, spins of the double dot are initialized to $\uparrow \uparrow $ (up and up) or $\downarrow \downarrow $ (down and down). Here we drive the spin state to $\uparrow \downarrow $ or $\downarrow \uparrow $ by irradiating with a microwave $\pi $ pulse corresponding to one of the target spins (e.g., the second spin). This state is hybridized with the spin-triplet state $(\uparrow \downarrow +\downarrow \uparrow )$, thus lifting the spin blockade, then one electron is emitted to the drain electrode, and another electron is reloaded to the double dot, and again initialized to $\uparrow \uparrow $ or $\downarrow \downarrow $. Thus, the increase of $I_{SD}$ is a time-ensemble measure of the flipping rate of the target spin, from its initialized $\uparrow $ or $\downarrow $ state to the other state. In the calculation, we treat one $1/2$-spin and it is always initialized to the ground state $\downarrow $, and the occupation of the upper level $\uparrow $ is calculated. The coherence time $T_{2}^{\ast }$ of the spin is estimated to be $\sim 0.2$ $0.3$~$\mu $s from Rabi oscillations~\cite{Ono18}, which is consistent with the ESR line width of $4$~MHz. The relaxation time $T_{1}$ cannot be estimated from our spin blockade detection scheme. At least we know that T_{1}$ does not dominate the refreshment of the spin blockade and thus it is much longer than $T_{2}^{\ast }$. In our previous work \cite{Ono18}, the back plate was not grounded at the right hand side of Fig.~\ref{FigS1}(a). Thus, the AC electric field, rather than the AC magnetic field was applied to the impurity. So we concluded that the spin resonance observed in Ref.~\cite{Ono18} was an electric dipole spin resonance, rather than an electron spin resonance. In this work, we have improved this point and an AC magnetic field is applied, rather than an AC electric field. \begin{figure}[t] \begin{center} \includegraphics[width=0.8 \textwidth, keepaspectratio]{FigS1} \end{center} \par \caption{\textbf{Schematics of the g-factor modulation and ESR.} (a) The Zeeman energy of the spin is modulated in time by a square waveform. Note that the spin blockade condition is always kept under modulation. (b) Schematic of the ESR for the slow limit of the square-wave modulation, where the low-energy spin state is hit by a low ESR microwave frequency. (c) The same slow-modulation limit, where the high-energy spin state is resonant with the high-frequency ESR microwave. (d) High modulation-frequency limit, where the spin energy is motional-averaged and the ESR occurs for an averaged microwave frequency.} \label{FigS1} \end{figure} \subsection{Tuning the g-factor by the gate voltage} We have implanted Al-N coupled impurity pairs as the deep impurity. A large anisotropy of its $g$-factor (2.0-2.3 depending on the magnetic field direction) has been reported in Ref.~\cite{Ono18}. The anisotropy of the $g$-factor of Al-N centers has also been reported for bound excitons trapped to the Al-N center \cite{weber1980localized}. These results suggest that the wave function of the Al-N center is strongly localized and has axial symmetry, and its spin-orbit interaction is strong and anisotropic. In such system, the $g$-factor should also react with an electric field which modifies its wave function, and results in the observed gate-voltage dependence of the $g$-factor. However, the detailed nature of the deep impurity is unknown and its identification is a future research subject. Changing the gate voltage $V_{G}$ within the spin-blockade region changes the $g$-factor by about $1\%$ \cite{Ono18}. In such system the $g$-factor should also react with an electric field which modifies its wave function (Stark effect), and results in the observed gate-voltage dependence of the g $-factor. Figure~\ref{Fig2}(a) shows the ESR peak observed in the spin-blockade regime. The ESR linewidth, i.e., the inverse of the coherence time T_{2}^{\ast }$, is reasonably limited by the spin blockade lifetime as well as the natural abundance of $^{29}$Si \cite{pla2012single}. Changing the gate voltage $V_{G}$ within the spin blockade region changes the $g$-factor by about $1$\% due to the Stark effect [Fig.~\ref{Fig2}(b)] \cit {rahman2009gate}. Therefore, the device behaves as a spin qubit. Two ESR transitions with $g$-factors $g=2.3$ and $g=2.7$ for the two impurities are observed. Hereafter, we only focus on the ESR peaks at $g=2.3 . \begin{figure}[t] \begin{center} \includegraphics[width=0.95 \textwidth, keepaspectratio]{Fig2} \end{center} \caption{\textbf{ESR for the single-spin qubit with tunable }$\mathbf{g} \textbf{-factor.} (a)~The source-drain current $I_{SD}$ of the device at V_{SD}$ $=0.33$~V and $V_{G}=-0.36$~V as a function of the frequency $f$ for various magnetic field detunings $\Delta B$ from $B$ $=0.2755$~T, with fixed MW power of $-18$~dBm. ESR peaks with a linewidth of $4$~MHz are observed. (b)~$I_{SD}$ of the device at $V_{SD}=0.33$~V and $B=0.2755$~T versus the frequency~$f$ for various gate voltage detuning $\Delta V_{G}$ away from V_{G}$ $=-0.36$~V. The $g$-factor (thus the qubit energy) is increased for more positive $\Delta V_{G}$. The change of the $g$-factor is found to be linearly dependent on $\Delta V_{G}$. For $|\Delta V_{G}|>20$ mV, the ESR response is not observed, suggesting the $V_{G}$ is out of the SB region. Note that the ESR peak is superimposed on the background current, which is around 2.5 pA. Both in (a) and (b) all the upper curves are shifted vertically, for clarity.} \label{Fig2} \end{figure} \section{Theoretical description of the driven and modulated single spin} \subsection{Energy-level modulations} Consider a two-level system, described by the Hamiltonia \begin{equation} H(t)=\frac{B_{z}(t)}{2}\sigma _{z}+\frac{B_{x}(t)}{2}\sigma _{x} \label{H} \end{equation wit \begin{equation} B_{z}(t)/\hbar =\omega _{0}+\delta \cdot s(t), \label{w_z} \end{equation where we assume the amplitude to be small, i.e. $\delta \ll \omega _{0}$, an \begin{equation} B_{x}(t)/\hbar =2G\cos \omega t. \label{w_x} \end{equation (Here the factor $2$ is introduced so that the amplitude $G$ defines the Rabi frequency.) For the longitudinal-field modulation, we consider different possibilities: (i) sinusoidal modulation, (ii) asymmetric latching modulation, and (iii) ramp modulation. Below we will discuss these regimes in more detail. (i) The sinusoidal modulation is the one most often used, and it is given by \begin{equation} s^{(i)}(t)=\cos \Omega t=\cos 2\pi \tau , \end{equation where we introduced the dimensionless time \begin{equation} \tau =\frac{\Omega t}{2\pi }. \end{equation} (ii) Next we consider a modulation with asymmetric rectangular pulses with duty ratio $d$. This corresponds to a qubit latched in one of the two states, with fast switching between these states. We refer to this regime as \textquotedblleft latching modulation\textquotedblright\ \cite{Silveri15}. In this case we assume that the modulating function has two stages with equal areas under the curve: \begin{equation} s_{d}^{(ii)}(\tau )=\left\{ \begin{array}{c} 2d,\;\;\;\;\;\;\;\;0<\tau <1-d, \\ -2\left( 1-d\right) ,\;1-d<\tau <1 \end{array \right. \label{s_ii} \end{equation Here the factor $2$ is introduced so that this modulating function changes between $-1$ and $+1$ for the symmetric $50\%$ duty ratio: \begin{equation} s_{0.5}^{(ii)}(\tau )=\left\{ \begin{array}{c} 1,\;\;\;\;\;0<\tau <0.5, \\ -1,\;\;\;0.5<\tau <1 \end{array \right. \end{equation These two definitions can be written (with an insignificant shift of the time variable) as \begin{equation} s_{d}^{(ii)}(\tau )=2\theta \left( \cos 2\pi \tau +\cos \pi d\right) -2\left( 1-d\right) \end{equation and \begin{equation} s_{0.5}^{(ii)}(\tau )=\mathrm{sgn}\left( \cos 2\pi \tau \right) , \end{equation where $\mathrm{sgn}$\ is the sign function. (iii) Modulating with triangular pulses, or \textquotedblleft ramp modulation\textquotedblright , corresponds to \begin{equation} s^{(iii)}(\tau )=\left\{ \mathrm{\tau }\right\} , \end{equation where the curly brackets denote the fractional part. In all cases the modulation frequency is assumed to be small, \begin{equation} \Omega \ll \omega . \label{<<} \end{equation For this reason, the fast signal with frequency $\omega $ can be called \textquotedblleft driving\textquotedblright , while the slow signal with frequency $\Omega $ can be denoted as the \textquotedblleft energy-level modulation\textquotedblright . \subsection{Bloch equations and the rotating-wave approximation} With the Hamiltonian (\ref{H}) the qubit dynamics can be described by the Bloch equations (as e.g. in Ref.~\cite{Shevchenko05}) for the components of the density matrix $\rho =\frac{1}{2}\left( 1+X\sigma _{x}+Y\sigma _{y}+Z\sigma _{z}\right) $ \begin{equation} \left\{ \begin{array}{c} \dot{X}=-B_{z}Y-\Gamma _{2}X, \\ \dot{Y}=-B_{x}Z+B_{z}X-\Gamma _{2}Y, \\ \dot{Z}=B_{x}Y-\Gamma _{1}\left( Z-Z_{0}\right) \end{array \right. \label{Bloch} \end{equation Here the phenomenological parameters $\Gamma _{1}=T_{1}^{-1}$ and $\Gamma _{2}=T_{2}^{-1}$ are the relaxation rates with decoherence rate $\Gamma _{2} \frac{1}{2}\Gamma _{1}+\Gamma _{\phi }$, defined by the pure dephasing rate \Gamma _{\phi }$. Decoherence defines the relaxation of $X$ and $Y$ towards 0$, while the relaxation of the diagonal component $Z$ is defined by the Maxwell-Boltzmann distribution for the given effective temperature $T_ \mathrm{eff}}$, and it evolves towards $Z_{0}=\tanh \left[ \hbar \omega _{0}/\left( 2k_{\mathrm{B}}T_{\mathrm{eff}}\right) \right] $. It is often instructive to solve the Bloch equations analytically. There are several approaches, such as the adiabatic-impulse model and the rotating-wave approximation (RWA). We refer the interested reader to Refs.~\cit {Shevchenko10, Silveri15} and references therein for the adiabatic-impulse and other models, while the RWA calculations are presented below in detail. Based on the slowness of the energy-level modulation, Eq.~(\ref{<<}), we can make use of the RWA, following Refs.~[\onlinecite{Silveri15}] and~ \onlinecite{Shevchenko10}]. First, we make the unitary transformation \begin{equation} U_{1}=\exp \left( -i\omega \sigma _{z}t/2\right) , \end{equation which corresponds to moving to the rotating frame, to get rid of the fast time dependence. In the new representation, the Hamiltonian read \begin{equation} H_{1}=U_{1}^{\dag }HU_{1}-i\hbar U_{1}^{\dag }\dot{U}_{1}=\frac{\hbar }{2 \left[ \Delta \omega +f(t)\right] \sigma _{z}+\frac{\hbar G}{2}\sigma _{x}, \label{H_RWA} \end{equation where $f(t)=\delta \!\cdot \!s(t)$ and $\Delta \omega =\omega _{0}-\omega $. Next, in order to have the Hamiltonian conveniently written to solve the Bloch equations, we make another unitary transformation, \begin{equation} U_{2}=\exp \left[ -i\eta (t)\sigma _{z}/2\right] ,\;\;\;\;\;\eta (t)=\int_{0}^{t}dt^{\prime }f\left( t^{\prime }\right) . \end{equation We obtain a new Hamiltonia \begin{equation} H_{2}=\!\frac{\hbar \Delta \omega }{2}\sigma _{z}+\frac{\hbar }{2}G\left( e^{i\eta }\sigma _{+}+h.c.\right) , \label{H2} \end{equation with $\sigma _{+}=\frac{1}{2}\left( \sigma _{x}+i\sigma _{y}\right) $. Then the preparatory stage is finalized by the Fourier-series expansion \begin{equation} e^{i\eta }=\sum_{m=-\infty }^{\infty }\Delta _{m}e^{im\Omega t}, \label{expansion} \end{equation where the complex-valued amplitude is given by the inverse Fourier transform \begin{equation} \Delta _{m}=\frac{\Omega }{2\pi }\int\limits_{0}^{2\pi /\Omega }dte^{-im\Omega t}e^{i\eta (t)}=\int\limits_{0}^{1}d\tau \exp \left[ i\eta (\tau )-i2\pi m\tau \right] . \label{Delta} \end{equation Then the Hamiltonian become \begin{equation} H_{2}=\!\frac{\hbar \Delta \omega }{2}\sigma _{z}+\frac{\hbar G}{2 \sum_{m=-\infty }^{\infty }\left( \Delta _{m}e^{im\Omega t}\sigma _{+}+h.c.\right) . \end{equation} To solve the Bloch equations, for the moment we assume that the system is driven close to resonance, where the \textquotedblleft dressed energy distance\textquotedblright\ $\hbar \Delta \omega $ equals to the energy of k $ photons, $\hbar \Delta \omega \approx k\hbar \Omega $. Then we omit the \textquotedblleft fast-rotating\textquotedblright\ terms and leave only terms with $m=k$. With this, the r.h.s. of the Bloch equations does not contain any explicit time dependence. Then equating its l.h.s. to zero, we obtain the stationary solution. In particular, this gives the upper-level occupation probability, $P_{+}=\frac{1}{2}\left( 1-Z\right) $. Summing all possible resonant terms, we obtain the qubit upper-level occupation probabilit \begin{equation} P_{+}\left( \Delta \omega ,\frac{\delta }{\Omega }\right) =\frac{1}{2 \sum_{k=-\infty }^{\infty }\frac{G_{k}^{2}(\delta /\Omega )} G_{k}^{2}(\delta /\Omega )+\frac{\Gamma _{1}}{\Gamma _{2}}\left( \Delta \omega -k\Omega \right) ^{2}+\Gamma _{1}\Gamma _{2}}, \label{Pup} \end{equation where $G_{k}(x)=G\left\vert \Delta _{k}(x)\right\vert $. We emphasize, that for a complex-valued $\Delta _{k}$, what matters is its absolute value. \subsection{Calculations for different modulations} As shown in Eq.~(\ref{Pup}), in order to obtain the upper-level occupation probability, one has to calculate the functions $\Delta _{k}$. This is the subject of the present subsection. (i) For the sinusoidal modulation, we can make use of the Jacobi-Anger expansion, which read \begin{equation} \exp \left( ix\sin 2\pi \tau \right) =\sum_{m=-\infty }^{\infty }J_{m}(x)\;e^{im2\pi \tau }, \end{equation where $J_{m}(x)$ is the Bessel function of the first kind. Then, it is straightforward to see tha \begin{equation} \Delta _{m}^{(i)}(x)=J_{m}(x),\;\;\;\;x=\frac{\delta }{\Omega }. \label{i} \end{equation It is useful to recall here the asymptote \begin{equation} J_{m}(x)\approx \sqrt{\frac{2}{\pi x}}\cos \left[ x-\frac{\pi m}{2}-\frac \pi }{4}\right] . \end{equation} (ii) For the asymmetric latching modulation, by direct integration we obtain \begin{equation} \Delta _{d,m}^{(ii)}(x)=\frac{2}{\pi }\frac{x\sin \left[ \pi \left( 1-d\right) \left( m-2dx\right) \right] }{\left( m+2\left( 1-d\right) x\right) \left( m-2dx\right) }. \label{asymm} \end{equation As mentioned before, since only $\left\vert \Delta _{m}\right\vert $ matters in Eq.~(\ref{Pup}), we omitted factors with unit modulus. We will do this throughout. In particular, for the symmetric rectangular modulating signal with $d=0.5$, Eq.~(\ref{asymm}) gives \begin{equation} \Delta _{0.5,m}^{(ii)}(x)=\frac{2}{\pi }\frac{x}{m^{2}-x^{2}}\sin \left[ \frac{\pi }{2}\left( m-x\right) \right] . \label{symm} \end{equation} (iii) For the ramp modulation, we have \begin{equation} \Delta _{m}^{(iii)}(x)=\int_{0}^{1}d\tau \exp \left[ i\pi \left( x\tau ^{2}-2m\tau \right) \right] . \label{Dm_iii} \end{equation This can be rewritten in terms of the Fresnel integrals: \begin{eqnarray} \left\vert \Delta _{m}^{(iii)}(x)\right\vert ^{2} &=&\frac{1}{\pi x}\left[ C\left( \sqrt{\pi x}\left( 1-\frac{m}{x}\right) \right) +C\left( \sqrt{\pi x \frac{m}{x}\right) \right] ^{2}+ \notag \\ &&+\frac{1}{\pi x}\left[ S\left( \sqrt{\pi x}\left( 1-\frac{m}{x}\right) \right) +S\left( \sqrt{\pi x}\frac{m}{x}\right) \right] ^{2}, \end{eqnarray \begin{equation} S(y)=\int_{0}^{y}dt\sin t^{2},\;\;\;\;C(y)=\int_{0}^{y}dt\cos t^{2}. \end{equation Such rewriting allows to use asymptotic approximations. In particular, when \left\vert y\right\vert \gg 1 \begin{equation} S(y)\approx C(y)\approx \sqrt{\frac{\pi }{8}}\;\mathrm{sgn\,}y, \end{equation which gives \begin{equation} \left\vert \Delta _{m}^{(iii)}(x)\right\vert \approx 1/\sqrt{x}. \end{equation} \begin{figure}[t] \begin{center} \includegraphics[width= 0.98 \textwidth, keepaspectratio]{sinusoidal} \end{center} \caption{Radio frequency (RF) wave power dependence of Landau-Zener-St\"{u ckelberg-Majorana (LZSM) interference of the spin resonance signal. (a) Schematic measurement set up. Instead of modulating the g-factor by $V_{G}$, here we add the rf signal to the microwave signal with the power combiner. This set up is effectively equivalent to the set up of Fig.~1(a) because the rf signal is fed to the gate via a stray capacitance between the substrate and the gate. (b-e) the RF power dependence of the LZSM interference with fixed RF frequency of (b) 2 MHz, (c) 5 MHz, (d) 10 MHz, and (e) 20 MHz, respectively. (f-i) shows the corresponding calculations. For calculations the following parameters were used for all the graphs: $G/2\protect\pi =1 ~MHz, $\Gamma _{1}/2\protect\pi =0.2$~MHz,\ $\Gamma _{2}/2\protect\pi =1 ~MHz. } \label{Fig:sinusoidal} \end{figure} The above equations allow to plot interferograms. We generated the right panels of Figs.~2-4 in the main text and the interferograms below (lower panels of Figs.~\ref{Fig:sinusoidal} and~\ref{Fig:latching} and the right panel of Fig.~\ref{Fig:ramp}) employing Eq.~(\ref{Pup}). In addition, we used the $\Delta _{k}$ from Eqs.~(\ref{i}, \ref{asymm}, \ref{symm}, and \ref{Dm_iii}) for the sinusoidal, asymmetric and symmetric latching, and ramp modulations, respectively. For calculations, we took the parameters known from the experiment related to the driving and modulation as well as \omega _{0}$ and $\Gamma _{2}$, while the two unknown parameters, $G$ and \Gamma _{1}$, were used for fitting. We emphasize that the interference pictures are very sensitive to the relaxation and decoherence rate, so, having obtained the agreement between the experimental and theory interferograms, we can state that we have reliably obtained the relaxation and decoherence rates. Finally, we note that we have checked that the interferograms calculated analytically with these equations agree nicely with the ones calculated numerically by solving Eq.~(\ref{Bloch}). \begin{figure}[tbp] \begin{center} \includegraphics[width= 0.85 \textwidth, keepaspectratio]{latching} \end{center} \caption{ Amplitude dependence of the square-wave modulation. (a-c) Similar measurements as Fig.~3(c) in the main text with different amplitudes of the square-wave modulation, (a) 16 mV, (b) 24 mV, and (c) 40mV, respectively. (d-f) shows the corresponding calculations, using $\protect\delta /2\protec \pi =12$, $18$, $30$~MHz for (d)-(f), respectively.} \label{Fig:latching} \end{figure} \begin{figure}[t] \begin{center} \includegraphics[width= 0.85 \textwidth, keepaspectratio]{asymmetric} \end{center} \caption{ Duty-ratio dependence of the asymmetric square-wave modulations. (a-f)~Similar measurements as in Fig.~4(c) in the main text with different duty ratios of the asymmetric square-wave modulation, (a) 20\%, (b) 30\%, (c) 40\%, (d) 60\%, (e) 70\%, and (f) 80\%, respectively. } \label{Fig:asymmetric} \end{figure} \subsection{Limiting cases} Equation~(\ref{Pup}), together with the expression of $\Delta _{k}(x)$ [Eqs.~(\ref{i}-\ref{Dm_iii})], allows for an analytical treatment. Let us consider several illustrative limiting cases. First, let us consider the symmetric latching modulation, with $\Delta _{k}(x)$ given by Eq.~(\ref{symm}). For low modulating frequencies, \;\Omega \ll \delta $, we have $x=\delta /\Omega \gg 1$ and $\left\vert k\right\vert \gg 1$. As a result, from Eq.~(\ref{symm}) we obtain that, for any given value of $x$, the contribution comes from the two $k$-th terms with $k\simeq \pm x$, for which we obtain $\Delta _{k}(x)\simeq 1/2$. Inserting this result in Eq.~(\ref{Pup}) we find that the position of the resonances are at $\Delta \omega =k\Omega $, which, for $k=\pm x$, gives two lines a \begin{equation} \Delta \omega =\pm \delta . \label{two_lines} \end{equation At large modulating frequencies, $\Omega \gg \delta $, we have $x\ll 1$ and k=0$. This is because for non-zero $k$ we have $\Delta _{k}\sim x/k^{2}\longrightarrow 0$. As a consequence, the position of the resonance is defined by $\Delta \omega =k\Omega $, which, for the main peak, with $k=0 , gives a zero shift of the resonance line \begin{equation} \Delta \omega =0. \label{zero} \end{equation} With $\Delta _{k}\simeq 1/2$, we can estimate the frequency half-width at half-maximum $\Delta \omega _{\mathrm{HWHM}}$ by equating \frac{1}{2}P_{+}$ at $\Delta \omega =k\Omega $ and $P_{+}$ at $\Delta \omega =k\Omega +\Delta \omega _{\mathrm{HWHM}}$. This give \begin{equation} \Delta \omega _{\mathrm{HWHM}}^{2}=\Gamma _{2}^{2}+\frac{G^{2}}{4}\frac \Gamma _{2}}{\Gamma _{1}}. \end{equation This means that the minimal half-width is $\Gamma _{2}$ and it is increased by the driving amplitude $G$. Even more informative is the asymmetric latching modulation, with $d\neq 1/2 . For low modulating frequencies, $x\ll 1$, there are two characteristic values of $k$, defined from the denominator of Eq.~(\ref{asymm}). For k=2(1-d)x$, we have $\Delta _{k}\simeq d$ and $\Delta \omega =k\Omega $, so tha \begin{equation} \Delta \omega =2(1-d)\delta . \label{right} \end{equation For $k=-2dx$, we have $\Delta _{k}\simeq 1-d$, an \begin{equation} \Delta \omega =-2d\delta . \label{left} \end{equation We can see that the two terms, with different $k$, define the positive and negative shifts of different signs, Eqs.~(\ref{right}-\ref{left}), which for $d=1/2$ reduce to Eq.~(\ref{two_lines}). From Eq.~(\ref{Pup}) we can also define the heights of the two respective peaks, at $\Delta \omega =k\Omega $ and for given values of $k^{\prime }$s we obtai \begin{equation} P_{+}^{L}=\frac{1}{2}\frac{d^{2}}{d^{2}+\lambda },\qquad P_{+}^{H}=\frac{1}{ }\frac{(1-d)^{2}}{(1-d)^{2}+\lambda },\qquad \lambda =\frac{\Gamma _{1}\Gamma _{2}}{G^{2}}\text{.} \end{equation In this way, the asymmetric latching is defined by the $d$-dependent peaks at small modulating frequency, while for the large modulating frequency we again have $\Delta \omega =k\Omega $ with $k=0$, i.e. the zero frequency shift, as above in Eq.~(\ref{zero}), which is remarkably independent of $d$. From our formulas, we can also estimate the modulating frequency $\Omega $, at which transition from one regime (high-frequency one, with the interference fringes) to another regime (low-frequency one, with two resonance lines described by Eq.~(\ref{two_lines})) takes place. For this, we can estimate from Eq.~(\ref{Pup}) both the width of the $k$-th resonance and the distance between neighboring resonances. Let us define the transition frequency $\Omega _{\ast }$ as the one at which those two values become equal. Then we obtai \begin{equation} \Omega _{\ast }=2\sqrt{\Gamma _{2}^{2}+G_{k}^{2}\frac{\Gamma _{2}}{\Gamma _{1}}}\,\,\gtrsim \,\,2\Gamma _{2}\text{.} \end{equation When $G$ is small, this gives $\Omega _{\ast }\approx 2\Gamma _{2}$. For our parameters, with $\Delta _{k}\simeq 1/2$, this gives $\Omega _{\ast }/2\pi \approx 3$~MHz, in agreement with what we can see in both the simulations and the experiment in Fig.~\ref{Fig:latching}. \subsection{Interferograms} As predicted by the formula (\ref{Pup}), there are resonances (when the denominator has minima) and nodes (when the numerator tends to zero). The respective resonance lines interrupted by the nodes form interference fringes, containing important information about the system and its environment.~\cite{Shevchenko10} The overall upper-level occupation probability $P_{+}$ depends on both $\Delta \omega $ and $\delta /\Omega $. Thus, the fringes can be visualized by plotting the qubit upper-level occupation (in theory) or the source-drain current (in experiment) as a function of these parameters. One can choose either a pair $\Delta \omega $ and $\delta $ or $\Delta \omega $ and $\Omega $. In the main text and in the Supplemental Material we mostly follow the latter option, while the dependence on the rf power (the former option) is presented below in Fig. \ref{Fig:sinusoidal}. There are two models which are convenient to understand and describe the interference: the so-called adiabatic-impulse model and the RWA. As presented in Refs.~\cite{Shevchenko10, Silveri15}, the two models give results. Here, for a qubit with various energy-level modulations, we used the RWA. Let us now summarize several key features of the two models. If a two-level system is periodically driven, this can be described as an alternation of two processes: adiabatic evolution along the ground and excited qubit states most of the time, with sudden transitions between the two states, when they approach each other. The latter are known as Landau-Zener transitions, while the adiabatic evolution is described by the accumulation of the so-called St\"{u}ckelberg phase. This phase produces the interference. Since related phenomena were also considered in 1932 by Majorana, the overall picture is related to four names: Landau, Zener, S \"{u}ckelberg, and Majorana. Varying the system parameters, one can observe the alteration of the constructive and destructive interference. While the adiabatic-impulse model could provide an intuitive picture, in our case (of rf-modulation plus mw-driving), it is more informative to use the RWA. This was considered in detail in this Section, and more graphical results will be presented in the next Section. As we mention after Eq.~(4) in the main text, the dressed qubit is excited under the resonant condition, $\Delta \omega =k\Omega $. As can be seen from Eq.~(\ref{Pup}), the resonance lines are interrupted by zeros, when $\Delta _{k}(x)=0$. Indeed, there we have destructive St\"{u}ckelberg interference. \section{Details of experimental and calculation results for the modulated single spin} \subsection{Sinusoidal modulation} The radio frequency (RF) wave power dependence of the Landau-Zener-St\"{u ckelberg-Majorana (LZSM) interference of the spin-resonance signal is summarized in Fig.~\ref{Fig:sinusoidal}. The intervals between the satellite peaks are defined by the RF frequencies, and the heights of the main and satellite peaks follow Bessel functions as a function of the RF power. Note that a small and very slow drift of the ESR frequency ($\sim 20$ MHz per week) is observed for fixed $V_{G}$ that seems to depend on the filing condition of liquid helium of the cryostat, probably due to the small change of the position of the superconducting magnet. The effect of this slow drift is negligible during the $1$ hour measurement, but induces variations of the ESR frequency $f_{0}$ at $V_{G}$ $=-0.36$~V from $9.00$ to $9.01$~GHz. \subsection{Symmetric square-wave modulation} If the modulation is slow enough, then in the response there are two separate peaks situated at the two resonance frequencies corresponding to the two states. Increasing the modulation frequency, the coherent response is displayed as an averaged signal, situated at a frequency between the two resonance frequencies mentioned above, which is known as motional narrowing. One of the relevant time scales that sets the cross-over to motional averaging is the dynamical time scale associated with the difference in frequency of the two states that the system is modulated between. Another relevant time scale is the effective coherence time T_{2}^{\ast }$. The smaller time of these is the characteristic time for the crossover. We note that the $T_{2}^{\ast }$ of the qubit is limited not only by the nuclear spins but also by the lifetime (refresh time) of the spin blockade. Namely, the spin blocked state in the panel (ii') of Fig.~1(c) has a finite lifetime due to the higher-order tunneling pass that leads the state (ii') directly to the state (i). The amplitude dependence of the square-wave modulation is shown in Fig.~\re {Fig:latching}, which demonstrates that there are two characteristic frequencies. First, by increasing the modulating frequency, at $\Omega \approx \Omega _{1}=2\cdot 2\pi $~MHz the transient behavior with interference fringes start to appear. Our calculations, demonstrate that this characteristic frequency is defined by the decoherence, $\Omega _{1}=2\Gamma _{2}$ and it is independent of the modulating amplitude $\delta $. By further increasing the frequency, we can observe a kind of motional averaging, with one principal peak at $\Delta f=0$ \cite{Li13} replacing the two peaks at $f_{1,2}-f_{0}=\pm \delta $. The appearance of this peak depends on the amplitude $\delta $ and is independent on the decoherence rate; this happens at $\Omega \approx \Omega _{2}=\delta /2$ \cite{Li13}. \subsection{Asymmetric latching modulation} The duty ratio dependence of the asymmetric square-wave modulations is shown in Fig.~\ref{Fig:asymmetric}. Data for Fig.~4(h) and (i) in the main text are extracted from these, as well as from Fig.~3(c) in the main text for the 50\% duty ratio. \subsection{Ramp modulation} We have checked the effect of time reversal symmetry of the ramp waveform [Fig.~\ref{Fig:ramp}(a)]. It is nearly identical to Fig.~5(c) in the main text. Figure~\ref{Fig:ramp}(b) is the derivative, $dP_{+}/df$, of the Fig.~5(d) in the main text. Interference fringes with smaller wave length around the modulation frequency of 2 MHz are not clearly seen in the $P_{+}$ intensity plot [Fig.~5(d)]. For a more detailed study to check the time-reversal symmetry of the spin dynamics under a ramp modulation, it might be necessary to also reverse the magnetic field direction and microwave phase to properly implement the time-reversed process. \begin{figure}[t] \begin{center} \includegraphics[width= 0.67 \textwidth, keepaspectratio]{ramp} \end{center} \caption{ Ramp modulations. (a)~Measured ramp-modulation frequency dependence similar to Fig.~5(c) in the main text, but with inverted ramp waveform. (b)~Intensity plot of the derivative $dP_{+}/df$. Data is the same as in Fig.~5(d) in the main text. } \label{Fig:ramp} \end{figure} \nocite{apsrev41Control} \bibliographystyle{apsrev4-1}

2b067ac29ae43eb6f648faaf7f700d0c63e50cb7

\section{Introduction} \IEEEPARstart{T}{his} demo file is intended to serve as a ``starter file'' for IEEE journal papers produced under \LaTeX\ using IEEEtran.cls version 1.8b and later. I wish you the best of success. \hfill mds \hfill August 26, 2015 \subsection{Subsection Heading Here} Subsection text here. \subsubsection{Subsubsection Heading Here} Subsubsection text here. \section{Conclusion} The conclusion goes here. \appendices \section{Proof of the First Zonklar Equation} Appendix one text goes here. \section{} Appendix two text goes here. \section*{Acknowledgment} The authors would like to thank... \ifCLASSOPTIONcaptionsoff \newpage \fi \section{Introduction}\label{Sec:Intro} While the ever-increasing demand for wireless service makes the radio spectrum one of the most valuable and scarce resources for wireless communication, recent studies have shown that the spectrum is not efficiently utilized at some locations for certain times of the day \cite{spectrum2002spectrum}. Dynamic spectrum management is a new paradigm to manage the radio spectrum in a dynamic manner by allowing cognitive nodes to utilize the unused bandwidth \cite{wyglinski2008cognitive}. Cognitive radio systems are usually composed of legacy spectrum owners, primary users (PUs), and cognitive devices seeking to access the PU's spectrum, called secondary users (SUs). Generally, dynamic spectrum technologies are broadly categorized into the two categories of \emph{common model} and \emph{property-right model} \cite{jovicic2009cognitive}. Unlike the common model for spectrum sharing, where the primary users are oblivious to the presence of SUs, in the property-right model, the PUs can willingly lease a portion of their spectrum to the SUs in exchange for monetary benefits or physical compensations. This compensation could be in form of providing relaying service, energy harvesting or cooperative jamming for the PUs \cite{CR-Afghah2013,afghah2014cooperative,Simeone2008,Afghah_INFOCOM,afghahIJHCR}. The property-right model for spectrum sharing in exchange for relaying service, also known as 'cooperative spectrum leasing' has received much attention in the last years, as it offers a win-win solution for both licensed and unlicensed users. The primary users can benefit from such spectrum leasing by enhancing their quality of service (QoS), in particular when experiencing a poor channel condition, while the secondary users can obtain the chance of affordable spectrum access. Furthermore, the property-right model can result in less energy consumption for the unlicensed users compared to the spectrum sensing models where they need to constantly sense the PUs' spectrum looking for spectrum holes. In this work, we consider a general model of cognitive radio networks with co-existence of several types of SUs that operate in different modes. This network consists of two PU transceivers and two SU transceivers which desire to exchange their signals with the help of available secondary relays. It is assumed that two underlay SUs are allowed to share the radio spectrum with the PUs provided that their interference at the PUs' receivers remains below an acceptable threshold. In addition to the two SUs operating in the underlay model, we consider the existence of multiple SUs that are interested in obtaining the spectrum access in exchange for providing relaying service for the PUs, based on the property-right model. Such model can limit the potential undesired interference that can be caused to the PUs, as well as the level of interference among the underlay SUs. Enabling a cooperative spectrum leasing to other SUs through property-right model can extend the number of SUs that can get the chance of spectrum access while benefiting the PUs through the cooperative relaying. To further enhance the efficiency of radio spectrum utilization in this network, a two-way cooperative communication scenario is utilized in this network. In general, the two methods of time division broadcast (TDBC) and multiple access broadcast (MABC) are utilized in two-way communication scenarios \cite{gavili2015optimal,Ubaidulla2014,ma2013robust, nguyen2017wireless,li2017resource,iranpanah2016distributed,Pareek,Zhang,Alsharoa}. In spite of TDBC scheme, where the transceivers send their signals in different time-slots, in MABC protocol both transceivers transmit their signals simultaneously. Since the proposed cooperative spectrum sharing mechanism is designed for cases when the quality of the direct link for the primary users is low, we consider an MABC scheme as it outperforms the TDBC scheme in such conditions \cite{zaeri2012}. One of the main concerns regarding the implementation of spectrum sharing solutions is combating the interference caused by simultaneous transmission of SUs with the PUs. Spectrum sharing networks are also vulnerable to the presence of unfriendly interferers that, despite the compliance of SUs, are not designated to respect the QoS requirement for the PUs. Unlike the intentional interferers (jammers), who intend to disturb the PUs' communication, the unfriendly interferers degrade the PUs' performance due to simultaneous transmission without PUs' consent. A potential example of these interferers can be the sensing-based secondary users that may interfer with the PUs' communication due to false detection of the PUs' presence or synchronization imperfections. A cognitive radio network must account for such burdens, imposed by either intentional interferers or false detection errors in the sensing process. Therefore, we study a scenario for co-existence of two PUs, and two underlay SUs operating in a two-way relaying system with multiple secondary relays when multiple unfriendy interferers exist. When such interferers exist, the information related to their channel state information is not usually available to the cognitive radio network. This is due to the fact that there is no cooperation between the unfriendly interferers and the networks' centralized controller which can only obtain an imperfect knowledge of the CSI. Withal, this imperfection can be due to time delays or frequency offset between the reciprocal channels as well as inaccurate channel estimation \cite{wang2011robust}. In order to study the impact of the imperfect CSI of the interferers, we assume that the CSI of all other channels is perfectly known. This assumption can be easily justified due to the existing collaboration between the PUs and SUs and the secondary relays, where CSI could be directly fed back from node to node \cite{jovicic2009cognitive}. While no collaboration between the interferers and the primary and secondary users is imagined, other mechanisms can be used to estimate the CSI between them. For instance, this CSI can be measured by a band manager and be provided using finite bandwidth channels \cite{suraweera2010capacity}. Eventually, this mechanism will cause inaccuracy in the estimated CSI which should be considered in the design of dynamic spectrum sharing systems. To the best of our knowledge, this is the first work that considers the impact of multiple interferers and CSI uncertainty on beamforming in the context of underlay cognitive radio systems which allows both PUs and SUs to operate in a two-way relaying mode. The existing relay nodes can also obtain the chance of spectrum access in exchange for providing an amplify-and forward cooperative service based on property-right spectrum sharing model. The main contribution of this work is to find the optimal beamforming vector which maximizes the QoS for both PU and SU transceivers in the above-mentioned system. The optimization problem is formulated as finding the beamform vector of the relay nodes that maximizes the QoS for PUs and SUs in the presence of unfriendly interferers with imperfect CSI. We consider the most general scenario with respect to uncertainty in interferers' CSI knowledge, in which no information is available about the distribution of such CSIs or its stochastic parameters. In the proposed model, we only consider a limited bound on the uncertainty of the interferers' CSI knowledge and design a robust solution that accounts for the worst-case scenario. Here we study two cases, where in the first one a complete knowledge of CSI of the channels between the interferers and the PUs, the SUs, and the secondary relays is available. The feasibility of the pristine optimization problem is examined and closed form equations for the feasibility conditions are derived for this case. The feasibility condition leads us to provide an upper bound on the optimal PUs' SINR. Afterward, a solution for the SINR optimization problem is proposed. The bisection method is applied to obtain the optimal SINR which can be achieved by the PU and SU transceivers. In second case , we consider the impact of imperfect knowledge of interferers' CSI on the designed system and calculate a robust solution for the SINR optimization problem. Another mathematical contribution of this work is to obtain the closed-form formulation of the worst-case scenario for each constraint instead of using linear matrix inequality (LMI) approaches with additional variables. The rest of this paper is organized as follows: In Section \ref{sec:related}, an overview of some related works in the literature is presented. Section \ref{Sec:System} describes the system model. In Section \ref{Sec:PerfectCSI}, the SINR optimization problem is defined and solved with the assumption of perfect CSI knowledge for the unfriendly interferers. Subsequently, in Section \ref{Sec:ImperfectCSI}, we show how to obtain the solution for the SINR optimization problem if only an imperfect knowledge of interferers' CSI is available. Numerical results are provided in Section \ref{Sec:Simulation}; and Section \ref{Sec:Conclusion} draws the concluding remarks. \section{Related Works} \label{sec:related} In general, cooperative communication techniques have proven to significantly enhance the performance of wireless communication systems in terms of reducing the energy consumption, enhancing the transmission rate, and extending the connectivity, to only name a few \cite{Dai_Survey,Li_cooperative,Sami_Cooperative}. A growing body of literature has investigated different factors that play a key role in optimizing the performance of cooperative relaying systems including studying the impact of power allocation, relay selection, relaying modes and time allocation strategies among direct and cooperative communication \cite{afghah2013stochastic,Nam_Selection,Chen_2011}. Furthermore, several joint optimization techniques have been proposed with the goal of improving the network performance when looking at the combined effect of these factors \cite{Liu_joint,Mo_DF,Mo_AF,Ng_Joint,Su_ICC}. The authors in \cite{Mo_DF} and \cite{Mo_AF} aimed at generalizing the common assumption of equal time allocation between the source and relay nodes and designed optimum joint power and time allocation mechanisms to minimize the outage probability when only the statistical knowledge of CSI is available. Motivated by the results of cooperative relaying in wireless networks, the SUs have been deployed as relays in cognitive radio networks to enhance the QoS of PUs, particularly when the PUs experience a poor channel condition due to shadowing or sparse network coverage \cite{Simeone2008,CR-Afghah2013,Su_Active_2012}. Cooperative spectrum leasing solutions have recently received a considerable attention in cognitive radio networks as they offer a coordination mechanism between the licensed and licensed users for dynamic spectrum access. In this methods, the secondary users can obtain the chance of spectrum access in exchange for providing cooperative services for the spectrum owners when they face poor channel conditions \cite{CR-Afghah2013,afghah2014cooperative,Simeone2008,Afghah_INFOCOM,Jayaweera}. In \cite{Simeone2008}, a model for cooperative spectrum leasing among a primary user and a network of Ad-Hoc secondary users is presented in which the primary user can decide whether to lease a portion of its spectrum access time to the secondary users noting its channel quality. Moreover, a non-cooperative game theoretic model is defined to determine the optimum power allocation of the secondary users when they compete with one another to enhance their transmission rate over the assigned time for SUs' transmission. In \cite{Su_Active_2012}, the authors studied the cooperative spectrum leasing in heterogeneous Ad-Hoc networks and calculated the necessary condition on the channel quality between the primary user and the SUs to encourage primary users to participate in leasing. A cooperation protocol is proposed to maximize the transmission rate of secondary cognitive users for the given amount of spectrum released by the PU and their given power budget, where an equal time allocation is assigned for transmission of the primary and secondary users. Similar to any cooperative communication networks, the performance of the cooperative spectrum leasing techniques depend on several factors such as relay selection, deployed relaying methods, availability of global CSI, reliability of the secondary users, and presence of jamming or interference. Different relaying strategies, including decode and forward (DF), compress and forward (CF), and amplify and forward (AF) are investigated in literature \cite{kim2008comparison}. The AF relaying mode has been widely utilized in practical applications due to its simplicity, as the relay nodes are only required to amplify and phase steer, i.e., beamform, the received signal and rebroadcast it. Despite the DF and CF coding relaying techniques, where the relay nodes need to decode and re-encode the transmitter's message; in AF relaying mode, the relay nodes only amplify and forward the received signal. Therefore, AF involves lower complexity and are an appropriate relaying solution for cooperative spectrum sharing applications as selected in this paper since the SUs do not require to have the knowledge of PU's codebooks for relaying \cite{Truong,Su_Active_2012}. While in cooperative spectrum leasing models, it is assumed that the secondary relays follow the agreement among the users to only transmit their messages in their allocated time slots \cite{CR-Afghah2013,afghah2014cooperative,Simeone2008,Jayaweera,Korenda,Namvar}, it is likely that these users will deviate from this agreement and cause harmful interference for the spectrum owners. Such undesired interference can be also caused by other unlicensed users in the proximity of the primary users working based on other spectrum sharing schemes such as spectrum sensing. In these conditions, the information about the channel conditions among the interferers and the PUs is often unavailable due to the lack of coordination among these users. This calls for new models to provide robust solutions to combat such undesired interference when minimum amount of information is accessible about the interferers. In general, three different approaches are known to handle the difficulties imposed by uncertainty in a data set, which in our case is imperfection in the interferers' CSI, \cite{ben1998robust}: 1) stochastic programming (SP), 2) robust mathematical programming (RMP), and 3) robust counterpart (RC) approach (worst-case scenario). The SP is limited to the problems where the uncertainty is stochastic in nature, i.~e.~ can be modeled as random variables. In this approach, the system is robustly designed in such a way that the average of network constraints including interference level or minimum QoS thresholds are in a desired range \cite{suraweera2010capacity,zhang2013outage,jamal2012performance,ozcan2016energy,Zarakovitis2016,Mo_DF,Mo_AF}. One may think of two practical drawbacks for this approach. First, we need to be able to identify the underlying probability distributions of uncertainty in the data set. The other vital pragmatic drawback of this approach is that it is very likely that the constraints on the average will be violated. Similarly, in the RMP approach, the violation of the constraints can occur but with a penalty in the objective \cite{mulvey1995robust}. Hence, this cannot be a proper option in software defined radio (SDR) networks, where the constraints on PUs' QoS must be met accurately. Finally, RC approaches, first introduced by Ben-Tal and Nemrovski \cite{ben1998robust}, comply with the specified constraints in a problem by considering a worst-case scenario. To be more precise, a bounded region is assumed for the uncertainty on the variables that are not perfectly known, while the objective of RC method is to find a robust solution that is feasible over the whole uncertainty region. Commonly, a mathematical concept known as the linear matrix inequality (LMI) along with defining new variables is used to solve an RC problem such as robust optimization problems in underlay SDR systems with uncertainty in CSI \cite{li2014robust,Tian2016,CRBidirectional-zaeri2016,wang2011robust}. {\textit{Notations:}} A summary of the notation is provided in Table \ref{Table:notations}. \begin{table} \centering \caption{Notation}\label{Table:notations} \begin{tabular}{|c|c|} \hline Notation & Description \\ \hline \hline uppercase boldface letters & matrices \\ \hline lowercase boldface letters & vectors \\ \hline $(.)^*$ & conjugate of the complex scalar \\ \hline $(.)^T$ & transpose of a vector or matrix \\ \hline $(.)^H$ & Hermitian (conjugate) transpose of a vector or matrix \\ \hline $\|{\bf{a}}\|$ & Euclidean norm ($\|.\|_2$) of the vector ${\bf{a}}$ \\ \hline $\lambda_{max}\{{\bf{A}}\}$ & maximum eigenvalue of the matrix ${\bf{A}}$\\ \hline ${\bf{A}}_{{i,j}}$ or ${\bf{a}}_{{i}}$ & the ${\{i,j\}}^{\mbox{th}}$ or ${\{i\}}^{\mbox{th}}$ element of matrix ${\bf{A}}$ or vector ${\bf{a}}$\\ \hline ${\bf{A}} = {\rm{diag}}({\bf{a}})$ & diagonal matrix with ${\bf{A}}_{{i,i}} = {\bf{a}}_{{i}}$ \\ \hline ${\bf{a}} = {\rm{diag}}({\bf{A}})$ & vertical vector with ${\bf{a}}_{{i}} = {\bf{A}}_{{i,i}}$ \\ \hline $\textsl{blkdiag}({\bf{A}},{\bf{B}})$ & diagonal matrix with the elements of the vectors ${\bf{a}}$ and ${\bf{b}}$ in order\\ \hline ${\bf{A}} \bullet {\bf{B}}$ & $\sum_i{\sum_j{{\bf{A}}_{{i,j}}{\bf{B}}_{{i,j}}}}$ \\ \hline $\begin{array}{c} \\ \bar{i} \mbox{\;\;for\;} i = 1,2 \\ \end{array}$ & $\begin{array}{c} \\ \mbox{all except \;} i\\ \end{array}$ \\ \hline $\boldsymbol{\ell}_{N,j}$ & $N \times 1$ vector with $j$th element 1 and the rest 0\\ \hline \end{tabular} \end{table} \section{System Model}\label{Sec:System} In this section, the system and channel models considered in the paper are briefly outlined. A SDR network consists of two primary transceivers ($PU_1$ and $PU_2$) and two secondary transceivers ($SU_1$ and $SU_2$) is considered. The SUs are in a soft lease contract with the PUs, meaning that they obtained the permission to simultaneously access the PUs' spectrum based on an agreement \cite{sboui2015achievable}. Due to a low quality of the direct link between $PU_1$ and $PU_2$, the primary network is willing to employ several secondary relays for the sake of cooperative relaying advantages. An MABC two-way cooperative scenario is designed, where two PU transceivers, as well as two secondary transceivers, desire to exchange their signals with the help of $N_r$ relays, simultaneously. As a reward, the relays, called $R_1,R_2,\ldots,R_{N_r}$, will obtain access to the PU's spectrum during each communication cycle for a fixed portion of the time slot, as depicted in Fig.~ \ref{fig:MABC}. It is also assumed that the network is affected by $N_I$ unfriendly interferers $\{I_1,I_2, \ldots, I_{N_I}\}$ and that all wireless channels are reciprocal and frequency flat. \begin{figure} \centerline{\resizebox{!}{6.5cm}{\includegraphics{Fig1.png}}} \caption{(a) MABC phase I in which both the PUs and SUs transmit their signals, simultaneously, and the relays receive these signals in presence of interferers. (b) MABC Phase II, in which the relays broadcast an amplified and phase shifted version of the signals they received in Phase I. (c) As a reward, the relays transmit their signal to their own destinations in a portion of this time-slot. $T_1$ refers to a portion of time slot allocated to transmission of the PUs and the SUs, while $T_2$ is the portion of the time slot allocated to the relays' transmission as an award for their cooperative services.} \label{fig:MABC} \end{figure} We assume that, in a given time-slot, the antennas can only transmit or receive a signal but not both at the same time, i.~e.~ all antennas operate in half-duplex mode. A centralized controller is considered to provide the perfect CSI of the SDR system as well as to calculate the optimum beamforming vector of the relays, an assumption which has been considered in similar reported works \cite{CRBidirectional-zaeri2016,CRBidirectional-Shahram2015,CRBidirectional-Zaeri2012,CRBidirectional-Veria2010}. The parameters of the system model are summarized in Table \ref{Table:Summaryparameters}. \begin{table} \centering \caption{Summary of parameters}\label{Table:Summaryparameters} \begin{tabular}{|c|c|} \hline Parameters & Description \\ \hline \hline $PU_i$ & i'th primary user \\ \hline $SU_i$ & i'th Secondary users \\ \hline ${\bf{f}}_{P_i}$ & channel coefficient vectors between $PU_i$ and relays \\ \hline ${\bf{f}}_{S_i}$ & channel coefficient vectors between $SU_i$ and relays\\ \hline ${\bf{h}}_{P_i}$ & channel coefficient vectors between the interferers and $PU_i$ \\ \hline ${\bf{h}}_{S_i}$ & channel coefficient vectors between the interferers and $SU_i$ \\ \hline ${\bf{h}}_{I_l}$ & channel coefficient vectors between the interferers and l'th relay \\ \hline ${\bf{r}}$ & received vector signal at relays \\ \hline ${\bf{w}}$ & beamformer vector\\ \hline ${\bf{t}}$ & transmitted signal by relays\\ \hline $y_{P_i}$& received signal by $PU_i$\\ \hline $y_{S_i}$& received signal by $SU_i$\\ \hline $SINR_{P_i}$& SINR at $PU_i$\\ \hline $SINR_{S_i}$& SINR at $SU_i$\\ \hline \end{tabular} \end{table} We assume that the $N_r \times 1$ complex channel coefficient vectors ${\bf{f}}_{P_i}$, $i = 1,2$, are defined as: \begin{IEEEeqnarray}{rCl}\label{Def:fPiSr} {\bf{f}}_{P_i} &=& [f_{P_iR_1}, f_{P_iR_2}, \ldots, f_{P_iR_{N_r}}]^T, \end{IEEEeqnarray} where $f_{P_iR_j}$ is the instant reciprocal flat fading channel coefficient between $PU_i$ and the secondary relay $R_j$ for $i = 1,2$ and $j = 1,2,\ldots, N_r$. The $N_r \times 1$ complex channel coefficient vectors ${\bf{f}}_{S_i}$, $i = 1,2$, are denoted by: \begin{IEEEeqnarray}{rCl}\label{Def:fSiSr} {\bf{f}}_{S_i} &=& [f_{S_iR_1}, f_{S_iR_2}, \ldots, f_{S_iR_{N_r}}]^T, \end{IEEEeqnarray} where $f_{S_iR_j}$s, $i = 1,2$ and $j = 1,2,\ldots,N_r$ are the instant reciprocal flat fading channel coefficients between the transceiver $SU_i$ and the secondary relay $R_j$. Moreover, we assume that the $N_I \times 1$ complex channel coefficient vectors between the interferers and $PU_i$ and $SU_i$, $i=1,2$, are defined as: \begin{IEEEeqnarray}{rCl}\label{Def:hPiOrSi} {\bf{h}}_{P_i} &=& [h_{P_i{I_1}},h_{P_i{I_2}},\ldots,h_{P_i{I_{N_I}}}]^T \IEEEyesnumber \IEEEyessubnumber \\ {\bf{h}}_{S_i} &=& [h_{S_i{I_1}},h_{S_i{I_2}},\ldots,h_{S_i{I_{N_I}}}]^T, \IEEEyessubnumber \end{IEEEeqnarray} where $h_{P_i{I_l}}$ and $h_{S_i{I_l}}$ denote the instant reciprocal flat fading channel coefficients between the interferer $I_l$, $l = 1,\ldots,N_I$, and $PU_i$ and $SU_i$, $i = 1,2$, respectively. Also, the $N_r \times 1$ complex channel coefficient vectors between the secondary relays and the interferer $I_l$, $l = 1,2,\ldots,N_I$, are defined as: \begin{IEEEeqnarray}{rCl}\label{Def:hSrIi} {\bf{h}}_{I_l} &=& [h_{R_1I_l}, h_{R_2I_l}, \ldots,h_{R_{N_r}I_l}]^T, \end{IEEEeqnarray} where, $h_{R_jI_l}$s, $j = 1,2,\ldots,N_r$ and $l = 1,2,\ldots,N_I$, are the instant reciprocal flat fading channel coefficients between the secondary relay $R_j$ and the interferer $I_l$. The channel coefficients in the system model are summarized in Fig.~\ref{fig:Channels}. The complete knowledge of CSI between the interferers and the SDR users is not available and only imperfect CSI estimations of channels between the PUs and the interferers, the SUs and interferers, and the secondary relays and the the interferers denoted by $\hat{{\bf{h}}}_{P_i}$, $\hat{{\bf{h}}}_{S_i}$, $\hat{{\bf{h}}}_{I_l}$, $i = 1,2$ and $l=1,2,\ldots,N_I$, respectively is provided by the centralized controller. Mathematically, this assumption can be written as \cite{CRBidirectional-zaeri2016}: \begin{IEEEeqnarray}{rCl}\label{Equ:ChannelIntBounded} {{\bf{h}}_{P_i}} =& \hat{{\bf{h}}}_{P_i} + \nabla{\bf{h}}_{P_i} ,& \| \nabla{\bf{h}}_{P_i} \| \leq \epsilon_{P_i} \IEEEyesnumber \IEEEyessubnumber\\ {{\bf{h}}_{S_i}} =& \hat{{\bf{h}}}_{S_i} + \nabla{\bf{h}}_{S_i} ,& \| \nabla{\bf{h}}_{S_i} \| \leq \epsilon_{S_i} \IEEEyessubnumber\\ {{\bf{h}}_{I_l}} =& \hat{{\bf{h}}_{I_l}} + \nabla{\bf{h}}_{I_l} ,& \| \nabla{\bf{h}}_{I_l} \| \leq \epsilon_l \IEEEyessubnumber \end{IEEEeqnarray} where $\nabla{\bf{h}}_{P_i}$, $\nabla{\bf{h}}_{S_i}$ and $\nabla{\bf{h}}_{I_l}$ are the bounded uncertainty CSI vectors and $\epsilon_{P_i}$, $\epsilon_{S_i}$, and $\epsilon_l$ denote the maximum value of CSI estimation error. The key advantage of this model is that it does not rely on the knowledge of distribution of the estimation errors rather it only requires the maximum value of these errors \cite{CRBidirectional-zaeri2016}. \begin{figure} \centerline{\resizebox{!}{6.5cm}{\includegraphics{Fig2.png}}} \caption{Channel coefficient vectors between interferers and $SU_i$ (${\bf{h}}_{S_i}$), interferers and $PU_i$ (${\bf{h}}_{P_i}$), interferer $I_l$ and relays (${\bf{h}}_{I_l}$), $SU_i$ and relays(${\bf{f}}_{S_i}$), and $PU_i$ and relays(${\bf{f}}_{S_i}$).} \label{fig:Channels} \end{figure} In MABC protocol, the messages are exchanged in two time slots, where in the first one, the primary and secondary transceivers send their messages, $x_{P_i}$ and $x_{S_i}$, $i = 1,2$, simultaneously. The relays receive a linear combination of all primary and secondary transmitted signals as well as the external interference signals $x_{I_l}^{(1)}$, $l = 1,2,\ldots,N_I$. Each relay rebroadcasts a weighted version of the received signal (AF relaying) in the second time-slot. Each transceiver receives a linear combination of its own signal, the other primary or secondary transceivers and the external interference signals $x_{I_l}^{(2)}$, $l = 1,2,\ldots,N_I$, in this time-slot. Without loss of generality, we assume that {$\mathbf{E}\{|x_{P_i}|^2\} = \mathbf{E}\{|x_{S_i}|^2\} = \mathbf{E}\{|x_{I_l}^{(i)}|^2\} = 1$} for all $i = 1,2$ and $l = 1,\ldots,N_I$. Also, we assume that all messages from different sources or in different time-slots are independent. Either primary or secondary transceivers may extract the desired signal considering the full knowledge of the self-interference portion of the signal. The $N_r \times 1$ vector of the received signal at the secondary relay network, in the first time-slot, can be written as: \begin{IEEEeqnarray}{rCl}\label{Equ:SignalRelayRec} {\bf{r}} &=& \sum_{i=1}^{2}{\sqrt{P_{P_i}}{\bf{f}}_{P_i}x_{P_i}} + \sum_{i=1}^{2}{\sqrt{P_{S_i}}{\bf{f}}_{S_iR }x_{S_i}} + \sum_{l=1}^{N_I}{\sqrt{P_{I_l}}{\bf{h}}_{I_l}x_{I_l}^{(1)}} + \boldsymbol{\nu}, \end{IEEEeqnarray} where $P_{P_i}$, $P_{S_i}$ and $P_{I_l}$, $i = 1,2$ and $l = 1,2,\ldots,N_I$ denote the transmit power of primary transceivers, secondary transceivers and interferers, respectively. The $N_r \times 1$ vector $ \boldsymbol{\nu} \sim \mathcal{N}_{\mathcal{C}}(\textbf{0}, \sigma^2 {\bf{I}}_{N_r})$ represents the white Gaussian noise at the relays. Each relay multiplies its received signal by a complex number $w^*_j$, $j = 1,2,\ldots,N_r$ and rebroadcasts it in the second time-slot. By defining the $N_r \times 1$ beamforming vector ${\bf{w}} = [w_1,w_2,\ldots,w_{N_r}]^T$, the $N_r \times 1$ transmitted vector at the relays can be written as: \begin{IEEEeqnarray}{rCl}\label{Equ:SignalRelayTra} {\bf{t}} &=& {\bf{W}}^H {\bf{r}} \IEEEyesnumber\\ &=& \sum_{i=1}^{2}{\sqrt{P_{P_i}}{\bf{W}}^H{\bf{f}}_{P_i} x_{P_i}} + \sum_{i=1}^{2}{\sqrt{P_{S_i}}{\bf{W}}^H{\bf{f}}_{S_i} x_{S_i}} + \sum_{l=1}^{N_I}{\sqrt{P_{I_l}}{\bf{W}}^H {\bf{h}}_{I_l} x_{I_l}^{(1)}} + {\bf{W}}^H \boldsymbol{\nu}. \IEEEnonumber \end{IEEEeqnarray} The individual power consumption at each relay can be written as: \begin{IEEEeqnarray}{rCl}\label{Equ:Pi_individual} P_{r_j} &=& \mathbf{E}\{ |{\bf{t}}_{\{j\}}|^2 \} = ({\bf{W}}^H \mathbf{E}\{ {\bf{r}} {\bf{r}}^H \}{\bf{W}})_{\{j,j\}} \IEEEnonumber \\ &=& ({\bf{W}}^H\left\{ \sum_{i=1}^{2}{{P_{P_j}}{\bf{f}}_{P_iR}{\bf{f}}_{P_iR}^H} + \sum_{i=1}^{2}{{P_{S_i}}{\bf{f}}_{S_iR}{\bf{f}}_{S_iR}^H} + \sum_{l=1}^{N_I}{{P_{I_l}}{\bf{h}}_{I_l}{\bf{h}}_{I_l}^H} + \sigma^2 {\bf{I}}_{N_r}\right\}{\bf{W}})_{\{j,j\}}, \IEEEyesnumber \end{IEEEeqnarray} for $j = 1,\ldots,N_r$. After some mathematical manipulation and by using the fact that ${\bf{A}}{\bf{b}} = {\bf{B}}{\bf{a}}$ if ${\bf{a}}=diag({\bf{A}})$ and ${\bf{b}}=diag({\bf{B}})$ for arbitrary same-size vectors ${\bf{a}}$ and ${\bf{b}}$, (\ref{Equ:Pi_individual}) is simplified as: \begin{IEEEeqnarray}{rCl}\label{Equ:Pi_individual_simple} P_{r_j} = \xi_{r_j} |w_j|^2, \end{IEEEeqnarray} where \begin{IEEEeqnarray}{rCl}\label{Def:xi} \xi_{r_j} &=& \sum_{i=1}^{2}{{P_{P_i}}{|({\bf{f}}_{P_iR})_{\{j\}}|^2}} + \sum_{i=1}^{2}{{P_{S_i}}|({\bf{f}}_{S_iR})_{\{j\}}|^2} + \sum_{l=1}^{N_I}{{P_{I_l}}|({\bf{h}}_{I_l})_{\{j\}}|^2} + \sigma^2, \IEEEyesnumber \end{IEEEeqnarray} for $j = 1,\ldots,N_r$. The received signal at $PU_i$, $i=1,2$, in the second time-slot can be written as: \begin{IEEEeqnarray}{rCl}\label{Equ:SignalPrimaryRec} y_{P_i} &=& \sum_{k=1}^{2}{\sqrt{P_{P_k}}{\bf{w}}^H{\bf{F}}_{P_i}{\bf{f}}_{P_k} x_{P_k}} + \sum_{k=1}^{2}{\sqrt{P_{S_k}} {\bf{w}}^H {\bf{F}}_{P_i} {\bf{f}}_{S_k} x_{S_k}} \IEEEnonumber \\ &+& \sum_{l=1}^{N_I}{\sqrt{P_{I_l}}{\bf{w}}^H{\bf{F}}_{P_i}{\bf{h}}_{I_l} x_{I_l}^{(1)}} + \sum_{l=1}^{N_I}{\sqrt{P_{I_l}} h_{P_i{I_l}}x_{I_l}^{(2)}} + {\bf{w}}^H {\bf{F}}_{P_i}\boldsymbol{\nu} + \nu_{P_i},\IEEEyesnumber \end{IEEEeqnarray} where $\nu_{P_i} \sim \mathcal{N}_{\mathcal{C}}(0, \sigma^2)$ is the white Gaussian noise at the $PU_i$, ${\bf{w}}$ denotes the beamforming vector. Notation ${\bf{F}}$ represents a diagonal matrix with ${\bf{F}}_{{i,i}}=\bf_{{i}}$ as previously defined in Table \ref{Table:notations}. The received signal at $SU_i$, $i=1,2$, in the second time-slot can be written as: \begin{IEEEeqnarray}{rCl}\label{Equ:SignalSecondaryRec} y_{S_i} &=& \sum_{k=1}^{2}{\sqrt{P_{P_k}}{\bf{w}}^H {\bf{F}}_{S_i} {\bf{f}}_{P_k} x_{P_k}} + \sum_{k=1}^{2}{\sqrt{P_{S_k}}{\bf{w}}^H {\bf{F}}_{S_i} {\bf{f}}_{S_k} x_{S_k}} \IEEEnonumber \\ &+& \sum_{l=1}^{N_I}{\sqrt{P_{I_l}}{\bf{w}}^H {\bf{F}}_{S_i} {\bf{h}}_{I_l} x_{I_l}^{(1)}} + \sum_{l=1}^{N_I}{\sqrt{P_{I_l}} h_{S_i{I_l}}x_{I_l}^{(2)}} + {\bf{w}}^H {\bf{F}}_{S_i} \boldsymbol{\nu} + \nu_{S_i},\IEEEyesnumber \end{IEEEeqnarray} where $\nu_{S_i} \sim \mathcal{N}_{\mathcal{C}}(0, \sigma^2)$ is the white Gaussian noise at the $SU_i$. It is assumed that each primary and secondary transceiver can compute and subtract the self-interference part of the received signal. Therefore, the residual received signals at the primary transceiver can be written as:{\begin{IEEEeqnarray}{rCl}\label{Equ:SignalPrimaryRecRes} \tilde{y}_{P_j} &=& \underbrace{\sqrt{P_{P_{\bar{j}}}}{\bf{w}}^H{\bf{F}}_{P_j}{\bf{f}}_{P_{\bar{j}}R} x_{P_{\bar{j}}}}_{\mbox{Desired Signal}} + \sum_{i=1}^{2}{\sqrt{P_{S_i}}{\bf{w}}^H {\bf{F}}_{P_j} {\bf{f}}_{S_i} x_{S_i}} \IEEEnonumber\\ &+& \sum_{l=1}^{N_I}{\sqrt{P_{I_l}}{\bf{w}}^H {\bf{F}}_{P_j} {\bf{h}}_{I_l} x_{I_l}^{(1)}} + \sum_{l=1}^{N_I}{\sqrt{P_{I_l}} h_{P_j{I_l}}x_{I_l}^{(2)}} + {\bf{w}}^H {\bf{F}}_{P_j} \boldsymbol{\nu} + \nu_{P_j}, \;\;\; j = 1,2. \IEEEyesnumber \end{IEEEeqnarray}} Also the residual received signals at the secondary transceivers can be written as: \begin{IEEEeqnarray}{rCl}\label{Equ:SignalSecondaryRecRes} \tilde{y}_{S_j} &=& \sum_{i=1}^{2}{\sqrt{P_{P_i}}{\bf{w}}^H {\bf{F}}_{S_j}{\bf{f}}_{P_i} x_{P_i}} + \underbrace{\sqrt{P_{S_{\bar{j}}}}{\bf{w}}^H {\bf{F}}_{S_j} {\bf{f}}_{S_{\bar{j}}R} x_{S_{\bar{j}}}}_{\mbox{Desired Signal}} \IEEEnonumber \\ &+& \sum_{l=1}^{N_I}{\sqrt{P_{I_l}}{\bf{w}}^H {\bf{F}}_{S_j}{\bf{h}}_{I_l} x_{I_l}^{(1)}} + \sum_{l=1}^{N_I}{\sqrt{P_{I_l}} h_{S_j{I_l}}x_{I_l}^{(2)}} + {\bf{w}}^H{\bf{F}}_{S_j}\boldsymbol{\nu} + \nu_{S_j}, \;\;\; j = 1,2. \IEEEyesnumber \end{IEEEeqnarray} Let us define \begin{IEEEeqnarray}{rCl}\label{Def:k} {\bf{k}}_{P_iP_j} = {\bf{F}}_{P_i}{\bf{f}}_{P_j} \qquad , \qquad {\bf{k}}_{S_iS_j} = {\bf{F}}_{S_i}{\bf{f}}_{S_j} \qquad \mbox{and} \qquad {\bf{k}}_{S_iP_j} = {\bf{F}}_{S_i}{\bf{f}}_{P_j},\IEEEyesnumber \end{IEEEeqnarray} for $i,j = 1,2$, \begin{IEEEeqnarray}{rCl}\label{Def:Q} {\bf{Q}}_{P_j} &=& \sum_{i=1}^{2}{P_{S_i}{\bf{k}}_{S_iP_j}{\bf{k}}_{S_iP_j}^H}+\sigma^2 {\bf{F}}_{P_j} {\bf{F}}_{P_j}^H + \sum_{l=1}^{N_I}{P_{I_l}{{\bf{F}}_{P_j}{\bf{h}}_{I_l}{\bf{h}}_{I_l}^H{\bf{F}}_{P_j}^H}}, \;\;\; j = 1,2 \IEEEyesnumber \IEEEyessubnumber\\ {\bf{Q}}_{S_j} &=& \sum_{i=1}^{2}{P_{P_i}{\bf{k}}_{S_jP_i}{\bf{k}}_{S_jP_i}^H}+\sigma^2 {\bf{F}}_{S_j} {\bf{F}}_{S_j}^H + \sum_{l=1}^{N_I}{P_{I_l}{{\bf{F}}_{S_j}{\bf{h}}_{I_l}{\bf{h}}_{I_l}^H{\bf{F}}_{S_j}^H}}, \;\;\; j=1,2.\IEEEyessubnumber \end{IEEEeqnarray} and \begin{IEEEeqnarray}{rCl}\label{Def:zeta} \zeta_{P_i} = \sum_{l=1}^{N_I}{P_{I_l}|h_{P_i{I_l}}|^2} \qquad , \qquad \zeta_{S_i} = \sum_{l=1}^{N_I}{P_{I_l}|h_{S_i{I_l}}|^2}, \;\;\; i=1,2.\IEEEyesnumber \end{IEEEeqnarray} By using the above definitions and the residual signals (\ref{Equ:SignalPrimaryRecRes}) and (\ref{Equ:SignalSecondaryRecRes}), the SINRs at the primary and secondary transceivers can be written as: \begin{IEEEeqnarray}{rCl}{\label{Equ:SINRPerfect}} SINR_{P_i} &=& \frac{P_{P_{\bar{i}}}{\bf{w}}^H{\bf{k}}_{P_1P_2}{\bf{k}}_{P_1P_2}^H {\bf{w}}} { {\bf{w}}^H{\bf{Q}}_{P_i}{\bf{w}}+\zeta_{P_i}+\sigma^2}, \;\;\; i =1,2 \IEEEyesnumber \IEEEyessubnumber\\ SINR_{S_i} &=& \frac{P_{S_{\bar{i}}}{\bf{w}}^H{\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H{\bf{w}}} { {\bf{w}}^H{\bf{Q}}_{S_i}{\bf{w}}+\zeta_{S_i}+\sigma^2}, \;\;\; i =1,2. \IEEEyessubnumber \end{IEEEeqnarray} In the next section, the SINR optimization problem is discussed. \section{SINR's Optimization with perfect CSI}\label{Sec:PerfectCSI} The goal of this section is to find an optimal weight vector, ${\bf{w}}$ such that the SINRs in the primary and secondary network is maximized. This optimization problem can be represented as a Max-Min problem with the sense of maximizing the minimum value among primary and secondary's SINRs. Despite the majority of reported works, in which the optimal beamforming vector is found to guarantee a minimum QoS for only the PUs, here we provide a solution to assure the required QoS for both PUs and SUs. In fact, the SUs have already obtained access to the spectrum by reason of their soft-lease and deserve to have a minimum QoS. However, considering the priority of the PUs as the spectrum owners, a design parameter $ \mu \geq 1$ is defined as an expected ratio between SINRs for the PUs and SUs. Another assumption in this optimization problem is that the available individual power at the relays is limited. Therefore, the SINR maximization problem can be written as: \begin{IEEEeqnarray}{rCl}\label{Equ:OptimizationMaxMin} \max_{{\bf{w}}} \mbox{\;}& \min{\left\{\{SINR_{P_i}\}_{i = 1,2},\mu \{SINR_{S_i}\}_{i = 1,2}\right\}} & \IEEEeqnarraynumspace \IEEEyesnumber \\ & \mbox{Subject To:\;} ~~ P_{r_j} \leq P_l^{max},& \;\;\; l = 1,\ldots,N_r \IEEEyessubnumber. \end{IEEEeqnarray} By defining an auxiliary real variable $\gamma > 0$, the Max-Min problem (\ref{Equ:OptimizationMaxMin}) can be rewritten as: \begin{IEEEeqnarray}{rCl}\label{Equ:Optimizationwgamma} \max_{{\bf{w}}, \gamma>0} & \gamma & \IEEEyesnumber\\ \mbox{Subject To:}& SINR_{P_i} \geq \gamma, & \mbox{\;\;\;} i = 1,2\IEEEyessubnumber \\ &\mu SINR_{S_i} \geq \gamma, & \mbox{\;\;\;} i = 1,2\IEEEyessubnumber \\ & P_{r_j} \leq P_j^{max}, & \mbox{\;\;\;} j = 1,\ldots,N_r \IEEEyessubnumber. \end{IEEEeqnarray} By using (\ref{Equ:Pi_individual_simple}) and (\ref{Equ:SINRPerfect}), the optimization problem (\ref{Equ:Optimizationwgamma}) can be rewritten as: \begin{IEEEeqnarray}{rCl}\label{Equ:OptimizationwgammaExpand} \max_{{\bf{w}}, \gamma>0} & \gamma& \IEEEyesnumber\\ \mbox{Subject To:}&\frac{P_{P_{\bar{i}}}{\bf{w}}^H{\bf{k}}_{P_1P_2}{\bf{k}}_{P_1P_2}^H {\bf{w}}} { {\bf{w}}^H{\bf{Q}}_{P_i}{\bf{w}}+\zeta_{P_i}+\sigma^2} \geq \gamma, & \mbox{\;\;\;} i = 1,2 \IEEEyessubnumber \IEEEeqnarraynumspace\\ &\mu \frac{P_{S_{\bar{i}}}{\bf{w}}^H{\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H{\bf{w}}} { {\bf{w}}^H{\bf{Q}}_{S_i}{\bf{w}}+\zeta_{S_i}+\sigma^2} \geq \gamma, & \mbox{\;\;\;} i = 1,2 \IEEEyessubnumber \IEEEeqnarraynumspace\\ & \xi_{r_j} |{\boldsymbol{\ell}_{N_r,j}}^T {\bf{w}}|^2 \leq P_j^{max}, & \mbox{\;\;\;} j = 1,\ldots,N_r \IEEEyessubnumber. \IEEEeqnarraynumspace \end{IEEEeqnarray} In the following subsection we investigate the feasibility condition(s) of the optimization problem (\ref{Equ:OptimizationwgammaExpand}). \subsection{Feasibility Condition} \begin{lemma}\label{Lemma:OneEigen} Let ${\boldsymbol{\Delta}}$ denote a Positive Definite (PD) matrix, ${\bf{a}}$ represent a vector of the same length as size of ${\boldsymbol{\Delta}}$, t and c are positive scalars and ${\bf{x}}$ be the vector variable with the same length as ${\bf{a}}$, then ${\bf{a}}^H {\boldsymbol{\Delta}}^{-1}{\bf{a}} - t \geq 0$ is a feasibility condition for the following constraint: \begin{IEEEeqnarray}{rCl}\label{Equ:Lemma1Const} \frac{{\bf{x}}^H{\bf{a}}\ba^H{\bf{x}}}{{\bf{x}}^H{\boldsymbol{\Delta}}{\bf{x}} + c} \geq t. \end{IEEEeqnarray} \end{lemma} \begin{proof} The constraint (\ref{Equ:Lemma1Const}) can be rearranged and be written as: \begin{IEEEeqnarray}{rCl}\label{Equ:Lemma1ConstV1} &&{\bf{x}}^H({\bf{a}}\ba^H - t{\boldsymbol{\Delta}}){\bf{x}} \geq ct \end{IEEEeqnarray} An optimization becomes infeasible, i.~e.~ no feasible point for ${\bf{x}}$ were found, if the core matrix ${\bf{a}}\ba^H - t{\boldsymbol{\Delta}}$ was negative semi-definite. Also, if this matrix was not negative semi-definite, a vector ${\bf{x}}_p$ exists such a way that ${\bf{x}}_p^H({\bf{a}}\ba^H - t{\boldsymbol{\Delta}}){\bf{x}}_p > 0$, thus one may scale up ${\bf{x}}_p$ in a way that the inequality (\ref{Equ:Lemma1ConstV1}) is satisfied. Therefore, (\ref{Equ:Lemma1ConstV1}) is feasible if and only if the matrix ${\bf{a}}\ba^H - t{\boldsymbol{\Delta}}$ was not negative semi-definite. As a result of the above discussion, the infeasibility condition can be written as: \begin{IEEEeqnarray}{rCl}\label{Equ:Lemma1ConstNSD} &&{\bf{a}}\ba^H - t{\boldsymbol{\Delta}} \preceq 0. \end{IEEEeqnarray} Since the matrix ${\boldsymbol{\Delta}}$ is positive definite, the constraint (\ref{Equ:Lemma1ConstNSD}) is equivalent to ${\boldsymbol{\Delta}}^{\frac{1}{2}}({\boldsymbol{\Delta}}^{\frac{-1}{2}}{\bf{a}}\ba^H{\boldsymbol{\Delta}}^{\frac{-1}{2}} - t{\bf{I}}){\boldsymbol{\Delta}}^{\frac{1}{2}} \preceq 0$. Also, the matrix ${\boldsymbol{\Delta}}^{\frac{1}{2}}$ is positive definite, we can rewrite the above mentioned condition as ${\boldsymbol{\Delta}}^{\frac{-1}{2}}{\bf{a}}\ba^H{\boldsymbol{\Delta}}^{\frac{-1}{2}} - t{\bf{I}} \preceq 0$, or equivalently $\lambda_{max}\{{\boldsymbol{\Delta}}^{\frac{-1}{2}}{\bf{a}}\ba^H{\boldsymbol{\Delta}}^{\frac{-1}{2}} - t{\bf{I}} \} > 0$. The matrix ${\boldsymbol{\Delta}}^{\frac{-1}{2}}{\bf{a}}\ba^H{\boldsymbol{\Delta}}^{\frac{-1}{2}}$ is a rank one matrix. Hence, for $t > 0$, all of the eigenvalues of the matrix ${\boldsymbol{\Delta}}^{\frac{-1}{2}}{\bf{a}}\ba^H{\boldsymbol{\Delta}}^{\frac{-1}{2}} - t{\bf{I}}$ are equal to zero except ${\bf{a}}^H{\boldsymbol{\Delta}}^{-1}{\bf{a}} - t$. Therefore, the feasibility condition can be summarized as ${\bf{a}}^H{\boldsymbol{\Delta}}^{-1}{\bf{a}} - t > 0$. \end{proof} Using Lemma \ref{Lemma:OneEigen}, the feasibility conditions for the individual constraints in optimization problem (\ref{Equ:OptimizationwgammaExpand}) can be written as: \begin{IEEEeqnarray}{rCl}\label{Equ:feasiblecon} &P_{P_{\bar{i}}}{\bf{k}}_{P_1P_2}^H {{\bf{Q}}_{P_i}}^{-1} {\bf{k}}_{P_1P_2} - \gamma > 0,& \;\;\; i = 1,2 \IEEEyesnumber \IEEEyessubnumber\\ &\mu P_{S_{\bar{i}}}{\bf{k}}_{S_1S_2}^H {{\bf{Q}}_{S_i}}^{-1} {\bf{k}}_{S_1S_2} - \gamma > 0,& \;\;\; i = 1,2. \IEEEyessubnumber \end{IEEEeqnarray} Therefore, by using the feasibility conditions (\ref{Equ:feasiblecon}), an upper bound condition for $\gamma$ can be written as: \begin{IEEEeqnarray}{rCl}\label{Equ:Uppergamma} \gamma^{up} &=& \min{\left\{\{P_{P_{\bar{i}}}{\bf{k}}_{P_1P_2}^H {{\bf{Q}}_{P_i}}^{-1} {\bf{k}}_{P_1P_2}\}_{i=1,2 } , \mu \{P_{S_{\bar{i}}}{\bf{k}}_{S_1S_2}^H {{\bf{Q}}_{S_i}}^{-1} {\bf{k}}_{S_1S_2}\}_{ i=1,2}\right\}}. \IEEEyesnumber \end{IEEEeqnarray} The upper bound value (\ref{Equ:Uppergamma}) of $\gamma$ does not guarantee the feasibility of the optimization problem (\ref{Equ:OptimizationwgammaExpand}), since it was obtained from individual constraints. In other word, the union of the feasibility regions associated with each constraint in (\ref{Equ:OptimizationwgammaExpand}) may provide additional limitation on $\gamma$ value. Therefore, we provide a method to find the optimal solution of $\gamma$ in the next subsection. \subsection{Optimal SINR Solution}\label{Subsec:BisectionPerfect} In this subsection, a feasibility check bisection method is used to find the optimal value of $\gamma$. In this method, the optimization problem (\ref{Equ:OptimizationwgammaExpand}) will turn into the following feasibility check problem for a given value of $\gamma$: \begin{IEEEeqnarray}{rCl}\label{Equ:Findw} & \mbox{Find} \;\;\; {{\bf{w}}} \IEEEyesnumber \\ &\frac{P_{P_{\bar{i}}}{\bf{w}}^H{\bf{k}}_{P_1P_2}{\bf{k}}_{P_1P_2}^H {\bf{w}}} { {\bf{w}}^H{\bf{Q}}_{P_i}{\bf{w}}+\zeta_{P_i}+\sigma^2} \geq \gamma, &\mbox{\;\;\;} i =1,2 \IEEEyessubnumber \\ &\mu \frac{P_{S_{\bar{i}}}{\bf{w}}^H{\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H{\bf{w}}} { {\bf{w}}^H{\bf{Q}}_{S_i}{\bf{w}}+\zeta_{S_i}+\sigma^2} \geq \gamma, & \mbox{\;\;\;} i =1,2 \IEEEyessubnumber\\ & \xi_{r_j} |{\boldsymbol{\ell}_{N_r,j}}^T {\bf{w}}|^2 \leq P_j^{max}, \mbox{\;\;\;} & j = 1,\ldots,N_r. \IEEEyessubnumber \end{IEEEeqnarray} The idea is to find the optimal value of $\gamma$, i.e., $\gamma^{opt}$, we do not need to calculate the optimum vector ${\bf{w}}$. To do so, we start from an initial interval $(\gamma_0^{low},\gamma_0^{up})$ of $\gamma^{opt}$ where $\gamma_0^{up}$ is the initial upper bound of $\gamma^{opt}$, which is derived in (\ref{Equ:Uppergamma}) and $\gamma_0^{low}$ is the initial lower bound of $\gamma^{opt}$, which is zero. Then, at step $n$, by choosing $\gamma = \frac{1}{2}(\gamma_{n-1}^{low}+\gamma_{n-1}^{up})$ and checking the feasibility of problem (\ref{Equ:Findw}), the solution interval of $\gamma^{opt}$ will be updated as: \begin{IEEEeqnarray}{rCl}\label{Equ:bisectionStep} (\gamma_n^{low},\gamma_n^{up}) &=& \left\{\begin{array}{cc} (\frac{1}{2}(\gamma_{n-1}^{low}+\gamma_{n-1}^{up}),\gamma_{n-1}^{up}) & \mbox{, if (\ref{Equ:Findw}) feasible}\\ &\\ (\gamma_{n-1}^{low},\frac{1}{2}(\gamma_{n-1}^{low}+\gamma_{n-1}^{up})) & \mbox{, otherwise.} \end{array} \right. \IEEEeqnarraynumspace \end{IEEEeqnarray} This bisection method will be continued until a small enough range of $\epsilon$ for the solution interval of $\gamma^{opt}$ is achieved. It is worth mentioning that the bisection method increases the complexity order of our method by a factor of $\log_2\left(\frac{\gamma^{up}}{\epsilon}\right)$. In order to solve the feasibility check problem (\ref{Equ:Findw}), we rearrange the constraints in a quadratic format: \begin{IEEEeqnarray}{rCl}\label{Equ:FindwQuad} &\mbox{Find} \;\;\; {{\bf{w}}} \IEEEyesnumber \\ &{\bf{w}}^H (P_{P_{\bar{i}}} {\bf{k}}_{P_1P_2}{\bf{k}}_{P_1P_2}^H - \gamma {\bf{Q}}_{P_i}) {\bf{w}} \geq \gamma(\zeta_{P_i}+\sigma^2), & \mbox{\;\;\;} i =1,2 \IEEEyessubnumber \\ & {\bf{w}}^H (\mu P_{S_{\bar{i}}}{\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H - \gamma {\bf{Q}}_{S_i}){\bf{w}} \geq \gamma(\zeta_{S_i}+\sigma^2),& \mbox{\;\;\;} i =1,2 \IEEEyessubnumber \\ & \xi_{r_j} |{\boldsymbol{\ell}_{N_r,j}}^T {\bf{w}}|^2 \leq P_j^{max}, & \mbox{\;\;\;} j = 1,\ldots,N_r \IEEEyessubnumber. \end{IEEEeqnarray} It is observed that if ${\bf{w}}$ is in the feasible region of (\ref{Equ:FindwQuad}), then for any arbitrary real number $\theta$, the vector $\tilde{{\bf{w}}} = e^{j\theta} {\bf{w}}$ is also in the feasible region. Therefore, without loss of generality, we can assume that ${\bf{k}}_{P_1P_2}^H {\bf{w}}$ (or ${\bf{k}}_{S_1S_2}^H {\bf{w}}$) is a non-negative real number. By applying this assumption, the constraints (\ref{Equ:FindwQuad}a) are turned to: \begin{IEEEeqnarray}{rCl}\label{Equ:SOCPconst} &\sqrt{\frac{P_{P_{\bar{i}}}}{\gamma}}{{\bf{k}}_{P_1P_2}}^H {\bf{w}} \geq \sqrt{{\bf{w}}^H {\bf{Q}}_{P_i} {\bf{w}} +\zeta_{P_i}+\sigma^2}, & \mbox{\;\;\;} i =1,2 \IEEEyesnumber \IEEEyessubnumber\\ & \Re\{{\bf{k}}_{P_1P_2}^H {\bf{w}}\} \geq 0 \qquad , \qquad \Im\{{\bf{k}}_{P_1P_2}^H {\bf{w}}\} = 0 & \IEEEyessubnumber, \end{IEEEeqnarray} where (\ref{Equ:SOCPconst}a) constraints are complex second order (Lorentz) cone (SOC) \cite{boyd2004convex}. However, as an indirect conclusion from Lemma \ref{Lemma:OneEigen}, the matrices $\mu P_{S_{\bar{i}}}{\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H - \gamma {\bf{Q}}_{S_i}$, $i = 1,2$, are not positive definite and therefore the constraints (\ref{Equ:FindwQuad}) do not represent convex regions. In order to make these constraints set convex, we define an auxiliary matrix variable ${\boldsymbol{\Omega}} = {\bf{w}} {\bf{w}}^H$. In a quadratic optimization problem, we are allowed to use the matrix variable ${\boldsymbol{\Omega}}$ by adding the following counterpart constraints \cite{palomar2010convex}: \begin{IEEEeqnarray}{rCl}\label{Equ:OmegaPSD} {\boldsymbol{\Omega}} - {\bf{w}} {\bf{w}}^H \succeq 0 \end{IEEEeqnarray} and $rank({\boldsymbol{\Omega}}) = 1$. The positive semidefinite (PSD) condition (\ref{Equ:OmegaPSD}) represents a convex region for variables ${\boldsymbol{\Omega}}$ and ${\bf{w}}$ \cite{boyd2004convex}. Although, the rank one condition does not represent a convex region; however, one may ignore the rank constraint of $\Omega$ and solve the relaxed optimization problem. It is shown that if a relaxed problem (non ranked restricted) which is linear with respect to ${\boldsymbol{\Omega}}$, was feasible then the rank-one restricted version of that problem is also feasible \cite{palomar2010convex}. Therefore, we remove the rank constraint from our optimization problem. By using the matrix variable ${\boldsymbol{\Omega}}$, equation (\ref{Equ:FindwQuad}b) can be converted to: \begin{IEEEeqnarray}{rCl}\label{Equ:QOPconstOmega} &\gamma {\bf{w}}^H {\bf{Q}}_{S_i}{\bf{w}} - \mu P_{S_{\bar{i}}}{\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H \bullet {\boldsymbol{\Omega}} + \gamma(\zeta_{S_i}+\sigma^2) \leq 0 & \mbox{\;\;\;} i =1,2 \IEEEyesnumber \IEEEyessubnumber \\ &\Re\{({\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H) \bullet {\boldsymbol{\Omega}} \} \geq 0 \qquad , \qquad \Im\{({\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H) \bullet {\boldsymbol{\Omega}} \} = 0 & \IEEEyessubnumber \end{IEEEeqnarray} for $j=1,2$. In order to reduce the computational complexity order of the problem, the following lemma can be used to convert the quadratic region (\ref{Equ:QOPconstOmega}) into a SOC region. \begin{lemma}\label{Lemma:QOP2SOCP} For a vector ${\bf{a}}$ and positive real numbers $\alpha \geq 0$ and $\beta \geq 0$, the inequality ${\bf{a}}^H {\bf{a}} \leq \alpha \beta$ holds if and only if the inequality $\left\|\left[\begin{array}{c} \alpha - \beta \\ 2{\bf{a}} \end{array}\right]\right\| \leq \alpha + \beta$ holds \cite{palomar2010convex}. \end{lemma} \begin{proof} The lemma will be simply proved by noting that $\left\|\left[\begin{array}{c} \alpha - \beta \\ 2{\bf{a}} \end{array}\right]\right\|^2 = (\alpha - \beta)^2+4 \|{\bf{a}}\|^2$. \end{proof} By applying Lemma \ref{Lemma:QOP2SOCP} for the values of $\alpha = 1$, $\beta = \frac{\mu}{\gamma} P_{S_{\bar{i}}}{\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H \bullet {\boldsymbol{\Omega}} - \zeta_{S_i}-\sigma^2$ and ${\bf{a}} = {{\bf{Q}}_{S_i}}^{\frac{1}{2}}{\bf{w}}$, the constraints (\ref{Equ:QOPconstOmega}) will turned into the following SOC constraints: \begin{IEEEeqnarray}{rCl}\label{Equ:SOCconstOmega} &\left\|\left[\begin{array}{c} 1 - \frac{\mu P_{S_{\bar{i}}}}{\gamma} {\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H \bullet {\boldsymbol{\Omega}} +\zeta_{S_i}+\sigma^2 \\ 2{{\bf{Q}}_{S_i}}^{\frac{1}{2}}{\bf{w}} \end{array}\right]\right\| \leq 1 + \frac{\mu P_{S_{\bar{i}}}}{\gamma} {\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H \bullet {\boldsymbol{\Omega}} - \zeta_{S_i}-\sigma^2,& \;\;\; i = 1,2\IEEEyesnumber \IEEEyessubnumber\\ &\Re\{({\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H) \bullet {\boldsymbol{\Omega}} \} \geq 0 \qquad , \qquad \Im\{({\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H) \bullet {\boldsymbol{\Omega}} \} = 0. \IEEEyessubnumber \end{IEEEeqnarray} By using (\ref{Equ:SOCPconst}), (\ref{Equ:OmegaPSD}) and (\ref{Equ:SOCconstOmega}), the relaxed version of the problem (\ref{Equ:FindwQuad}) can be written as: \begin{IEEEeqnarray}{rCl}\label{Equ:FindwSOCP} & \mbox{Find} \;\;\; {{\bf{w}},{\boldsymbol{\Omega}}} \IEEEyesnumber \IEEEnosubnumber\\ &\sqrt{\gamma {\bf{w}}^H {\bf{Q}}_{P_i} {\bf{w}} + \zeta_i^p+\sigma^2} \leq \sqrt{\frac{P_{P_{\bar{i}}}}{\gamma}}{\bf{k}}_{P_1P_2}^H {\bf{w}}, & \mbox{\;\;\;} i =1,2 \IEEEnonumber \\ & \Re\{{\bf{k}}_{P_1P_2}^H {\bf{w}}\} \geq 0 \qquad \Im\{{\bf{k}}_{P_1P_2}^H {\bf{w}}\} = 0 & \IEEEnonumber \\ &\left\|\left[\begin{array}{c} 1 - \frac{\mu}{\gamma} P_{S_{\bar{i}}}{\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H \bullet {\boldsymbol{\Omega}} +\zeta_{S_i}+\sigma^2 \\ 2{{\bf{Q}}_{S_i}}^{\frac{1}{2}}{\bf{w}} \end{array}\right]\right\| \leq 1 + \frac{\mu}{\gamma} P_{S_{\bar{i}}}{\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H \bullet {\boldsymbol{\Omega}} - \zeta_{S_i} -\sigma^2, & \mbox{\;\;\;} i =1,2 \IEEEnonumber \\ &\Re\{({\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H) \bullet {\boldsymbol{\Omega}} \} \geq 0 \qquad ,\qquad \Im\{({\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H) \bullet {\boldsymbol{\Omega}} \} = 0 & \IEEEnonumber \\ & \sqrt{\xi_{r_j}}|{\boldsymbol{\ell}_{N_r,j}}^T {\bf{w}}| \leq \sqrt{P_j^{max}}& \mbox{\;\;\;} j = 1,\ldots, N_r \IEEEnonumber \\ & {\boldsymbol{\Omega}} - {\bf{w}} {\bf{w}}^H \succeq 0. \IEEEnonumber \end{IEEEeqnarray} The feasibility check problem can be solved by employing Second Order Cone Programming (SOCP) using cvx software \cite{cvx}. It is worth emphasizing that no rank deduction algorithm is required to obtain a rank one solution for ${\boldsymbol{\Omega}}$. In the next section, we define and solve the QoS's maximization problem when any imperfect interference CSI is available to the central controller. \section{SINR's Optimization with Imperfect Interferers CSI}\label{Sec:ImperfectCSI} In session \ref{Sec:PerfectCSI}, the problem of joint optimizing the SINR of the PUs and SUs considering the perfect knowledge of the external interferences is discussed. However, the assumption of perfect knowledge on the unfriendly interferers' CSI is not realistic in practical applications. We provide an optimization problem which is robust against all uncertainties in interferers' CSI. Let $\Psi$ denote the set of uncertainty regions of all interferers' CSI. By using (\ref{Equ:ChannelIntBounded}) we have: \begin{IEEEeqnarray}{rCl}\label{Def:Psi} \Psi &=& \left\{ \forall \left\{\{\nabla{\bf{h}}_{P_i}\}_{i = 1,2},\{\nabla{\bf{h}}_{S_i}\}_{i = 1,2},\{\nabla{\bf{h}}_{I_l}\}_{l = 1,\ldots,N_I} \right\} \left| \| \nabla{\bf{h}}_{P_i} \| \leq \epsilon_{P_i} \& \| \nabla{\bf{h}}_{S_i} \| \leq \epsilon_{S_i} \& \| \nabla{\bf{h}}_{I_l} \| \leq \epsilon_{l} \right.\right\}\mbox{\;} \IEEEeqnarraynumspace \IEEEyesnumber \end{IEEEeqnarray} By applying (\ref{Def:Psi}), the robust version of the optimization problem (\ref{Equ:Optimizationwgamma}) can be written as: \begin{IEEEeqnarray}{rCl}\label{Equ:OptimizationwgammaImperfect} \max_{{\bf{w}}, \gamma>0} & \gamma &\IEEEyesnumber \IEEEnosubnumber\\ \mbox{Subject To:}& SINR_{P_i}(\psi) \geq \gamma, & \mbox{\;\;\;}\forall \psi \in \Psi \mbox{\;\;\;} i=1,2 \IEEEnonumber \\ &\mu SINR_{S_i}(\psi) \geq \gamma, & \mbox{\;\;\;} \forall \psi \in \Psi \mbox{\;\;\;} i=1,2 \IEEEnonumber \\ & P_{r_j}(\psi) \leq P_j^{max}, & \mbox{\;\;\;} \forall \psi \in \Psi \mbox{\;\;\;} j = 1,\ldots,N_r \IEEEnonumber. \end{IEEEeqnarray} The robust optimization problem (\ref{Equ:OptimizationwgammaImperfect}) suggests common $\psi$ for all constraints. However, one may think of a separate uncertainty array $\psi$ for every set of constraints. In fact, Theorem 2.1 of \cite{ben1998robust} allows us to write a counterpart robust optimization problem of (\ref{Equ:OptimizationwgammaImperfect}) as: \begin{IEEEeqnarray}{rCl}\label{Equ:OptimizationwgammaImperfectCounterpart} \max_{{\bf{w}}, \gamma>0} & \gamma &\IEEEyesnumber \IEEEnosubnumber\\ \mbox{Subject To:}& SINR_{P_i}(\psi_{P_i}) \geq \gamma, & \mbox{\;\;\;}\forall \psi_{P_i} \in \Psi \mbox{\;\;\;} i=1,2 \IEEEnonumber \\ &\mu SINR_{S_i}(\psi_{S_i}) \geq \gamma, & \mbox{\;\;\;}\forall \psi_{S_i} \in \Psi \mbox{\;\;\;} i=1,2 \IEEEnonumber \\ & P_{r_j}(\psi_{r_j}) \leq P_j^{max}, & \mbox{\;\;\;}\forall \psi_{r_j} \in \Psi \mbox{\;\;\;} j = 1,\ldots,N_r \IEEEnonumber, \end{IEEEeqnarray} or equivalently: \begin{IEEEeqnarray}{rCl}\label{Equ:OptimizationwgammaImperfectworst} \max_{{\bf{w}}, \gamma>0} & \gamma \IEEEyesnumber \IEEEnosubnumber\\ \mbox{Subject To:}&\min_{\forall \psi_{P_i} \in \Psi} SINR_{P_i}(\psi_{P_i}) \geq \gamma, & \mbox{\;\;\;} i = 1,2 \IEEEeqnarraynumspace \IEEEyessubnumber\\ &\mu \min_{ \forall \psi_{S_i} \in \Psi}SINR_{S_i}(\psi_{S_i}) \geq \gamma, & \mbox{\;\;\;} i=1,2 \IEEEeqnarraynumspace \IEEEyessubnumber \\ &\max_{\forall \psi_{r_j} \in \Psi} P_{r_j}(\psi_{r_j}) \leq P_j^{max}, & \;\;\; j = 1,\ldots,N_r \IEEEyessubnumber . \end{IEEEeqnarray} In the next subsection, a closed-form format for all constraints of the optimization problem (\ref{Equ:OptimizationwgammaImperfectworst}) is provided. \subsection{Closed-Form Worst Case Imperfection}\label{Subsec:WorstClosedForm} By using the definitions (\ref{Equ:ChannelIntBounded}), (\ref{Def:Q}) and (\ref{Def:zeta}), the constraints (\ref{Equ:OptimizationwgammaImperfectworst}a-b) can be rewritten as: \begin{IEEEeqnarray}{rCl}\label{Equ:SINRConstworst} & \gamma \max_{\forall \psi_{P_i} \in \Psi} \left\{\sum_{l=1}^{N_I}{P_{I_l}}\| {\bf{w}}^H{\bf{F}}_{P_i}(\hat{{\bf{h}}_{I_l}}+ \nabla{\bf{h}}_{I_l})\|^2 + \|(\hat{{\bf{h}}_{P_i}} + \nabla{\bf{h}}_{P_i})^H{\bf{P}}_I^{\frac{1}{2}}\|^2 \right\} \IEEEnonumber\\ &\leq{\bf{w}}^H (P_{P_{\bar{i}}} {\bf{k}}_{P_1P_2}{\bf{k}}_{P_1P_2}^H - \gamma {\bf{T}}_{P_i}) {\bf{w}} - \gamma \sigma^2 \IEEEyesnumber \IEEEyessubnumber \\ & {\gamma} \max_{\forall \psi_{S_i} \in \Psi} \left\{\sum_{l=1}^{N_I}{P_{I_l}}\| {\bf{w}}^H{\bf{F}}_{S_i}(\hat{{\bf{h}}_{I_l}}+ \nabla{\bf{h}}_{I_l})\|^2 + \|(\hat{{\bf{h}}_{S_i}} + \nabla{\bf{h}}_{S_i})^H{\bf{P}}_I^{\frac{1}{2}}\|^2 \right\} \IEEEnonumber\\ &\leq{\bf{w}}^H (\mu P_{S_{\bar{i}}}{\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H - \gamma {\bf{T}}_{S_i}){\bf{w}} - \gamma \sigma^2 \IEEEyessubnumber \end{IEEEeqnarray} where ${\bf{P}}_I = [P_{I_1},P_{I_1},\ldots,P_{I_{N_I}}]^T$, and \begin{IEEEeqnarray}{rCl}\label{Def:R} {\bf{T}}_{P_i} &=& \sum_{j=1}^{2}{P_{S_j}{\bf{k}}_{S_jP_i}{\bf{k}}_{S_jP_i}^H}+\sigma^2 {\bf{F}}_{P_i} {\bf{F}}_{P_i}^H, \;\;\; i = 1,2 \IEEEyesnumber \IEEEyessubnumber\\ {\bf{T}}_{S_i} &=& \sum_{j=1}^{2}{P_{P_j}{\bf{k}}_{S_iP_j}{\bf{k}}_{S_iP_j}^H}+\sigma^2 {\bf{F}}_{S_i} {\bf{F}}_{S_i}^H, \;\;\; i = 1,2. \IEEEyessubnumber \end{IEEEeqnarray} Also, by using (\ref{Equ:ChannelIntBounded}) and (\ref{Equ:Pi_individual_simple}) the constraint (\ref{Equ:OptimizationwgammaImperfectworst}c) can be written as: \begin{IEEEeqnarray}{rCl}\label{Equ:PowerConstWorst} &\max_{\forall \psi_{r_j} \in \Psi}\{ \sum_{l=1}^{N_I}{{P_{I_l}}|{\boldsymbol{\ell}_{N_r,j}}^T(\hat{{\bf{h}}}_{I_l}+\nabla {{\bf{h}}}_{I_l})|^2}\} | {\boldsymbol{\ell}_{N_r,j}}^T {\bf{w}}|^2 \leq P_j^{max} - \chi_{r_j} | {\boldsymbol{\ell}_{N_r,j}}^T {\bf{w}}|^2\IEEEyesnumber \end{IEEEeqnarray} where, \begin{IEEEeqnarray}{rCl}\label{Def:chi} \chi_{r_j} =\sum_{i=1}^{2}{{P_{P_i}}|{\boldsymbol{\ell}_{N_r,j}}^T{\bf{f}}_{P_i}|^2} +\sum_{i=1}^{2}{{P_{S_i}}|{\boldsymbol{\ell}_{N_r,j}}^T {\bf{f}}_{S_i}|^2}+\sigma^2,\;\;\; j = 1,\ldots,N_r \end{IEEEeqnarray} In order to write the constraints (\ref{Equ:SINRConstworst}) and (\ref{Equ:PowerConstWorst}) in a closed format, we use two lemmas on norm vector inequalities. First: \begin{lemma}\label{Lemma:One} Let ${\bf{c}}$ and ${\bf{a}}$ denote complex vectors. If ${\bf{b}}$ is a vector whose norm is bounded by constant $\epsilon$, i.~e.~ $\|{\bf{b}}\| \leq \epsilon$, then $\|{\bf{c}}^H({\bf{a}} + {\bf{b}})\| \leq |{\bf{c}}^H{\bf{a}}| + \epsilon \|{\bf{c}}\|$ and the equality holds if and only if ${\bf{b}} = \frac{\epsilon}{\|{\bf{c}}\|}{\bf{c}} e^{j\angle{{\bf{c}}^H{\bf{a}}}}$. \end{lemma} \begin{proof} The proof is provided in \cite{CRBidirectional-zaeri2016} by using triangle inequality theorem and Cauchy-Schwarz inequality. \end{proof} By applying lemma \ref{Lemma:One} we have: \begin{IEEEeqnarray}{rCl}\label{Equ:UncertainwPSR} \max_{\forall \psi_{P_i} \in \Psi} &\| {\bf{w}}^H{\bf{F}}_{P_i}(\hat{{\bf{h}}}_{I_l}+ \nabla{\bf{h}}_{I_l})\| = |{\bf{w}}^H{\bf{F}}_{P_i}\hat{{\bf{h}}}_{I_l}|+ \epsilon_l \|{\bf{w}}^H{\bf{F}}_{P_i}\| \IEEEyesnumber \IEEEyessubnumber \\ \max_{\forall \psi_{S_i} \in \Psi} &\| {\bf{w}}^H{\bf{F}}_{S_i}(\hat{{\bf{h}}}_{I_l}+ \nabla{\bf{h}}_{I_l})\| = |{\bf{w}}^H{\bf{F}}_{S_i}\hat{{\bf{h}}}_{I_l}|+ \epsilon_l \|{\bf{w}}^H{\bf{F}}_{S_j}\| \IEEEyesnumber \IEEEyessubnumber \\ & \max_{\forall \psi^r_j \in \Psi} \mbox{\;\;\;} |{\boldsymbol{\ell}_{N_r,j}}^T(\hat{{\bf{h}}}_{I_j}+\nabla {{\bf{h}}}_{I_j})| = |{\boldsymbol{\ell}_{N_r,j}}^T \hat{{\bf{h}}}_{I_j}|+ \epsilon_{l} \IEEEyesnumber \IEEEyessubnumber \end{IEEEeqnarray} Next, \begin{lemma}\label{Lemma:Two} Let ${\boldsymbol{\Delta}}$ denote an invertible matrix, let ${\bf{a}}$ represent a vector and let ${\bf{b}}$ represent a vector whose norm is upper bounded by constant $\epsilon$, i.~e.~ $\|{\bf{b}}\| \leq \epsilon$. Then $\|({\bf{a}}+{\bf{b}})^H {\boldsymbol{\Delta}} \| \leq (1 + \frac{\epsilon}{\|{\bf{a}}\|})\|{\bf{a}}^H {\boldsymbol{\Delta}}\|$ and the equality holds if and only if ${\bf{b}} = \frac{\epsilon}{\|{\bf{a}}\|}{\bf{a}}$. \end{lemma} \begin{proof} The proof is provided in \cite{CRBidirectional-zaeri2016} by using triangle inequality theorem and Cauchy-Schwarz inequality. \end{proof} By defining: \begin{IEEEeqnarray}{rCl}\label{Def:kappa} \kappa_{p_i} = (1+\frac{\epsilon_{P_i}}{\| \hat{{\bf{h}}_{P_i}}\|})^2\| \hat{{\bf{h}}_{P_i}}^H{\bf{P}}_I^{\frac{1}{2}}\|^2 \qquad \mbox{and} \qquad \kappa_{s_i} = (1+\frac{\epsilon_{S_i}}{\|\hat{{\bf{h}}_{S_i}}\|})^2\| \hat{{\bf{h}}_{S_i}}^H{\bf{P}}_I^{\frac{1}{2}}\|^2, \;\;\; i = 1,2, \end{IEEEeqnarray} and by using Lemma \ref{Lemma:Two}, we have: \begin{IEEEeqnarray}{rCl}\label{Equ:UncertainPS} \max_{\forall \psi_{P_i} \in \Psi} &\|(\hat{{\bf{h}}_{P_i}} + \nabla{\bf{h}}_{P_i})^H{\bf{P}}_I^{\frac{1}{2}}\|^2 = \kappa_{p_i} \qquad \mbox{and} \qquad \max_{\forall \psi_{S_i} \in \Psi} & \|(\hat{{\bf{h}}_{S_i}} + \nabla{\bf{h}}_{S_i})^H{\bf{P}}_I^{\frac{1}{2}}\|^2 = \kappa_{s_i}, \;\;\; i = 1,2.\IEEEyessubnumber \end{IEEEeqnarray} As a result, the constraints (\ref{Equ:SINRConstworst}) and (\ref{Equ:PowerConstWorst}) can be rewritten as the following closed format: \begin{IEEEeqnarray}{rCl}\label{Equ:closedConst} & \gamma \left(\sum_{l=1}^{N_I}{P_{I_l}} (|{\bf{w}}^H{\bf{F}}_{P_i}\hat{{\bf{h}}}_{I_l}|+ \epsilon_l \|{\bf{w}}^H{\bf{F}}_{P_i}\|)^2 +\kappa_{p_i} \right) \leq {\bf{w}}^H (P_{P_{\bar{i}}} {\bf{k}}_{P_1P_2}{\bf{k}}_{P_1P_2}^H - \gamma {\bf{T}}_{P_i}) {\bf{w}} - \gamma \sigma^2, \;\;\; i = 1,2 \IEEEyesnumber \IEEEyessubnumber \\ & {\gamma} \left(\sum_{l=1}^{N_I}{P_{I_l}}(|{\bf{w}}^H{\bf{F}}_{S_i}\hat{{\bf{h}}}_{I_l}|+ \epsilon_l \|{\bf{w}}^H{\bf{F}}_{S_j}\|)^2+ \kappa_{s_i}\right) \leq {\bf{w}}^H (\mu P_{S_{\bar{i}}}{\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H - \gamma {\bf{T}}_{S_i}){\bf{w}} - \gamma \sigma^2, \;\;\; i = 1,2 \IEEEyessubnumber \\ & \kappa_{r_j} | {\boldsymbol{\ell}_{N_r,j}}^T {\bf{w}}|^2 \leq P_j^{max}, \;\;\; j = 1,\ldots, N_r\IEEEyessubnumber \end{IEEEeqnarray} where \begin{IEEEeqnarray}{rCl}\label{Def:kappar} \kappa_{r_j} = \sum_{l=1}^{N_I}{{P_{I_l}}(|{\boldsymbol{\ell}_{N_r,j}}^T \hat{{\bf{h}}}_{I_l}|+ \epsilon_{l})^2} + \chi_{r_j},\;\;\; j = 1,\ldots, N_r \end{IEEEeqnarray} The closed-form constraints are substituted in the optimization problem (\ref{Equ:OptimizationwgammaImperfectworst}). In the next subsection, we use SOCP method to obtain the optimal beamforming of our system model considering the imperfection of interferers' CSI. \subsection{Optimal SINR Solution}\label{Subsec:BisectionImPerfect} In this subsection, we provide a bisection method to find the optimal value of $\gamma$ when just imperfect interferers' CSI are known. By applying (\ref{Equ:closedConst}), the optimization problem (\ref{Equ:OptimizationwgammaImperfectworst}) is turned to a feasibility check problem of finding ${\bf{w}}$ with the constraints (\ref{Equ:closedConst}) for a given value of $\gamma$. The upper bound on $\gamma$ for the case when perfect knowledge of interferers' CSI is available was described in (\ref{Equ:Uppergamma}). It should be noted that the same upper bound applies for the case when only an imperfect knowledge of interferers' CSI is available. Therefore, the same bisection approach as is used in subsection \ref{Subsec:BisectionPerfect} can be applied for the imperfection case. Drawing to a close, we modify the constraints (\ref{Equ:closedConst}) to the SOCP format, so the discussed feasibility check problem is solved in an efficient way. Let us use auxiliary positive relay variables $\rho_{p_i}$ and $\rho_{s_i}$, $i = 1,2$ in such a way that $\|{\bf{w}}^H{\bf{F}}_{P_i}\| \leq \rho_{p_i}$ and $\|{\bf{w}}^H{\bf{F}}_{S_i}\| \leq \rho_{s_i}$. Also, we use the $N_I \times 1$ auxiliary positive real vector variables $\boldsymbol{\varrho}_{p_i} = [\varrho_{p_i,1}, \ldots , \varrho_{p_i,N_I}]^T$ and $\boldsymbol{\varrho}_{s_i} = [\varrho_{s_i,1}, \ldots , \varrho_{s_i,N_I}]^T$ for $i = 1,2$ in such a way that $|{\bf{w}}^H{\bf{F}}_{P_i}\hat{{\bf{h}}}_{I_l}| \leq \varrho_{p_i,l}$ and $|{\bf{w}}^H{\bf{F}}_{S_i}\hat{{\bf{h}}}_{I_l}| \leq \varrho_{s_i,l}$ for $l = 1,\ldots,N_I$. By using the above mentioned auxiliary variables the feasibility check problem for a given value of $\gamma$ can be written as: \begin{IEEEeqnarray}{rCl}\label{Equ:FindwImperfect} & \mbox{Find} \;\;\; { \{\rho_{p_i},\rho_{s_i},\boldsymbol{\varrho}_{p_i},\boldsymbol{\varrho}_{s_i}\}_{i = 1,2} ,{\bf{w}}} \IEEEyesnumber \\ & \gamma \left(\sum_{l=1}^{N_I}{P_{I_l}} ({\boldsymbol{\ell}_{N_I,l}}^T \boldsymbol{\varrho}_{p_i} + \epsilon_l \rho_{p_i})^2 +\kappa_{p_i}\right)\leq {\bf{w}}^H (P_{P_{\bar{i}}} {\bf{k}}_{P_1P_2}{\bf{k}}_{P_1P_2}^H - \gamma {\bf{T}}_{P_i}) {\bf{w}} - \gamma \sigma^2, & \mbox{\;\;\;} i = 1,2\IEEEyessubnumber \\ & {\gamma} \left(\sum_{l=1}^{N_I}{P_{I_l}}({\boldsymbol{\ell}_{N_I,l}}^T \boldsymbol{\varrho}_{p_i} + \epsilon_l \rho_{p_i})^2 + \kappa_{s_i} \right) \leq {\bf{w}}^H (\mu P_{S_{\bar{i}}}{\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H - \gamma {\bf{T}}_{S_i}){\bf{w}} - \gamma \sigma^2, & \mbox{\;\;\;} i = 1,2\IEEEyessubnumber \\ &\sqrt{\kappa_{r_j}} | {\boldsymbol{\ell}_{N_r,j}}^T {\bf{w}}| \leq \sqrt{P_j^{max}}, & \mbox{\;\;\;} j = 1,\ldots,N_r \IEEEyessubnumber \\ &\|{\bf{w}}^H{\bf{F}}_{P_i}\| \leq \rho_{p_i} \mbox{\;\;} \qquad , \qquad \|{\bf{w}}^H{\bf{F}}_{S_i}\| \leq \rho_{s_i} &\mbox{\;\;\;} i = 1,2\IEEEyessubnumber \\ &|{\bf{w}}^H{\bf{F}}_{P_i}\hat{{\bf{h}}}_{I_l}| \leq {\boldsymbol{\ell}_{N_I,l}}^T \boldsymbol{\varrho}_{p_i} \qquad , \qquad |{\bf{w}}^H{\bf{F}}_{S_i}\hat{{\bf{h}}}_{I_l}| \leq {\boldsymbol{\ell}_{N_I,l}}^T \boldsymbol{\varrho}_{s_i}, &\mbox{\;} i = 1,2 ~\mbox{and} ~ l = 1,\ldots,N_I ~ \IEEEeqnarraynumspace \IEEEyessubnumber. \end{IEEEeqnarray} All (\ref{Equ:FindwImperfect}c-e) constraints represent SOC regions. We use auxiliary vector variables $\boldsymbol{\varpi}_{p_i} = [{\varpi}_{p_i,1},\ldots,{\varpi}_{p_i,N_I}]^T$ and $\boldsymbol{\varpi}_{s_i} = [{\varpi}_{s_i,1},\ldots,{\varpi}_{s_i,N_I}]^T$, in the interest of clarity, in which ${\varpi}_{p_i,l} = {\boldsymbol{\ell}_{N_I,l}}^T \boldsymbol{\varrho}_{p_i} + \epsilon_l \rho_{p_i}$ and ${\varpi}_{s_i,l} = {\boldsymbol{\ell}_{N_I,l}}^T \boldsymbol{\varrho}_{s_i} + \epsilon_l \rho_{s_i}$ for $i = 1,2$ and $ l = 1,\ldots,N_I$. By assuming that ${\bf{k}}_{P_1P_2}^H {\bf{w}}$ is a positive real number, the constraints (\ref{Equ:FindwImperfect}a) can be turned to the following SOC constraint: \begin{IEEEeqnarray}{rCl}\label{Equ:SOCPconstImperfect} & \frac{P_{P_{\bar{i}}}}{\gamma} {\bf{k}}_{P_1P_2}^H {\bf{w}} \geq \sqrt{\boldsymbol{\varpi}_{p_i}^T {\bf{P}}_I \boldsymbol{\varpi}_{p_i} + {\bf{w}}^H {\bf{T}}_{P_i} {\bf{w}} +\kappa_{p_i} +\sigma^2},\;\;\; i = 1,2. \IEEEyesnumber. \end{IEEEeqnarray} Ultimately, by using Lemma \ref{Lemma:QOP2SOCP} and by defining matrices ${\bf{P}}_{s_i} = \textsl{blkdiag}({\bf{P}}_I,{\bf{T}}_{S_i})$, $i = 1,2$, the constraints (\ref{Equ:FindwImperfect}b) can be relaxed to the following SOC format: \begin{IEEEeqnarray}{rCl} & \left\| \left[\begin{array}{c} 1 - \frac{\mu P_{S_{\bar{i}}}}{\gamma} {\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H \bullet {\boldsymbol{\Omega}}+ \kappa_{s_i} +\sigma^2 \\ 2 {\bf{P}}_{s_i}^{\frac{1}{2}}\left[\begin{array}{c} \boldsymbol{\varpi}_{s_i} \\ {\bf{w}} \end{array} \right] \end{array}\right]\right\| \leq 1 + \frac{\mu P_{S_{\bar{i}}}}{\gamma} {\bf{k}}_{S_1S_2}{\bf{k}}_{S_1S_2}^H \bullet {\boldsymbol{\Omega}} - \kappa_{s_i} - \sigma^2, \;\;\; i = 1,2.\IEEEyesnumber \end{IEEEeqnarray} Therefore, the feasibility check problem (\ref{Equ:FindwImperfect}) can be effectively solved using the SOCP method. \section{Numerical Experiments}\label{Sec:Simulation} In this section, we present different simulation scenarios to illustrate the effectiveness of our proposed method in combating interference when the perfect information on CSIs of channels between the interferers and other users is not available. To do so, all channel coefficients were generated as complex Gaussian variables with zero mean and unit variance. The PU and SU transceivers' powers were assumed to be 0 dBm. For the sake of simplicity, we considered two unfriendly interferers that operate in the PUs' spectrum property, with the power of -1 dBm. Also, 10 relays were considered to cooperate with both PU and SU transceivers with a maximum individual power limit of 1 dBm. We consider 3 times priority for the primaries quality of service with respect to the secondaries, i.e., $\mu = 3$. The achievable rate at each receiver is defined by $R = \log_2(1 + SINR)$ and is plotted in Fig.~\ref{fig:SINRPUSUmu3Compare} versus the noise power ($\sigma^2$). The minimum achievable rate of $PU_1$ and $PU_2$ is shown by $R_P$ and plotted versus $\sigma^2$. As can be seen in the figure, $R_P$ decreased from 2 to 0 when noise power increased from -20 to 20 dBm. Also, by using the priority design parameter $\mu = 3$, the optimization problem forces the network to assign more resources for primary transmissions purposes. This limits the achievable rate for the secondary transmission as it is shown by allowable $R_S$ in the figure. As it can be seen in Fig. \ref{fig:SINRPUSUmu3Compare}, the optimization problem makes a restriction on maximum achievable rate of 1 when $\sigma^2$ is -20 dBm for $R_{S_1}$ and $R_{S_2}$. However, at this level of noise power, $R_{S_1}$ and $R_{S_2}$ achieve the rates of 0.4 and 0.6 dBm, respectively. \begin{figure} \centerline{\resizebox{!}{8.5cm}{\includegraphics{Fig3.png}}} \caption{Achievable rate for 1 Hz bandwidth for primary and secondary transceivers for fairness design parameter $\mu = 3$.} \label{fig:SINRPUSUmu3Compare} \end{figure} The impact of the interferers on the SDR system is illustrated in Fig.~ \ref{fig:SINRIntPower} for different transmission powers of these interferers. The designed parameter $\mu$ was considered to be 1, and achievable rate were obtained for different levels of interferers' power, while all other parameters were the same as before. Figure \ref{fig:SINRIntPower} shows that, by increasing the interferers' power from -2 dBm to 1 dBm, the rate decreased by 2.5 dB when the noise is too weak. However, this performance reduction is regulatable if noise variance is high. \begin{figure} \centerline{\resizebox{!}{8.5cm}{\includegraphics{Fig4.png}}} \caption{Achievable rate for 1 Hz bandwidth versus $P (\frac{\sigma_{{\bf{f}}}}{\sigma})^2$ for different interferers power.} \label{fig:SINRIntPower} \end{figure} We also investigated the effects of cooperation level of relays on interference mitigation. In this scenario, we considered 2 interferers and, the maximum power that each relay was assigned for relaying purposes, was varied from -2 to 2 dBm while all other parameters were the same as before. Figure \ref{fig:SINRRelayPower} shows the changes on the average $\gamma$ in a Monte-Carlo simulation versus the noise power for different maximum limitations on individual relay powers. As it is shown in this figure, by increasing the relay power limit, the diversity gain of the system increased. For instance, increasing relay power limit from -2 dBm to 2 dBm caused a 3 dB increase on SINR when the noise power is -20 dBm. \begin{figure} \centerline{\resizebox{!}{8.5cm}{\includegraphics{Fig5.png}}} \caption{Achievable rate for 1 Hz bandwidth versus $P (\frac{\sigma_{{\bf{f}}}}{\sigma})^2$ for different maximum limitation on individual relay powers.} \label{fig:SINRRelayPower} \end{figure} The previous numerical experiments were based on the assumption of perfect CSI availability for interferes. We examined our proposed robust method against uncertainties on interferers' CSI. A scenario where the interferers' CSI are known imperfectly is considered. We assumed the imperfection as a percentage of the estimated CSI and varied this percentage from $2\%$ to $10\%$. The optimal robust $\gamma$ is calculated by solving the robust optimization problem (\ref{Equ:FindwImperfect}). Figure \ref{fig:ImperfectPercentage} illustrates that the system will pay the cost by gaining less SINR if the knowledge accuracy is decreased. For example, the performance of the system decreased by 1 dB if the accuracy of the interferers' CSI decreases from 2 to 10 percent. \begin{figure} \centerline{\resizebox{!}{8.5cm}{\includegraphics{Fig6.png}}} \caption{Achievable rate for 1 Hz bandwidth versus $P (\frac{\sigma_{{\bf{f}}}}{\sigma})^2$ for different level of imperfection in knowledge of interferers' CSI.} \label{fig:ImperfectPercentage} \end{figure} The $\sigma^2$ value is fixed on 0 dB in Fig.~\ref{fig:ImperfectPercentageRelayP} and the achievable rate is plotted versus the changes in the individual relay powers for different levels on the interferers CSI imperfection. As it is mentioned before, the relays provide diversity gain to the system and the achievable rates for the transceivers in the system increase when the relays power increase. Besides, more uncertainty on the interferers CSI causes less performance for the system. As it can be seen from this figure, the achievable rate decreases from 1.25 to 1.05 when the imperfection on CSI increased between 5 and 15 percentages for a maximum available power of 4 dBm for the relays. \begin{figure} \centerline{\resizebox{!}{8.5cm}{\includegraphics{Fig7.png}}} \caption{Achievable rate for 1 Hz Bandwidth versus relays' power limit for different level of imperfection in knowledge of interferers' CSI.} \label{fig:ImperfectPercentageRelayP} \end{figure} To summarize, in the simulations, we showed the diversity gain on a two-way SDR-network in which the resources are fairly distributed among primary and secondary users. The effect of interferers in our cooperative model was investigated. Moreover, the interferers CSI uncertainties effect on the system performance was investigated and it was shown that our proposed method is robust against such uncertainties. It is worth mentioning that the convergence of the proposed method depends on the convergence of the bisection method which is in order of $O(\frac{1}{2^n})$ \cite{numerical_book}. While the bisection method may be slower than numerical methods, but it always converges to the solution. \section{Conclusion}\label{Sec:Conclusion} We propose a model for co-existence different types of secondary users in an SDR network. A system consists of two pairs of PU and SU transceivers, several SU relays and also several interferers is considered. The optimum beamforming solution is provided to maximize quality service in PU and SU transceivers. Moreover, the optimization problem is solved by considering the worst-case scenario when the knowledge on the interferers channels is imperfect. The simulation results show the performance of our proposed method and its robustness against uncertainties in interferers CSI. \bibliographystyle{IEEEtran

7db8d800592eadc5e586830aa89074b7cf89834e

\section{Annotation Methodology} \label{sec:annotation-methodology} We use the CI framework to annotate policy statements that describe contextual information exchanges. Our use of a CI flow-based abstraction is an important distinction from previous privacy policy annotation research, as it serves a useful semantic abstraction for checking privacy statements for more complex properties than previously attempted. For the remainder of the paper, we denote a privacy statement with a single set of CI parameters as an ``information flow.'' For example, we consider the following statement an information flow, or simply as a ``flow:'' \medskip\\ \noindent\fbox{% \parbox{0.45\textwidth}{% {\em {\bf \textcolor{ForestGreen}{We [Facebook]}}$^{\textcolor{red}{recipient}}$ also collect {\bf \textcolor{blue}{contact information}}$^{\textcolor{red}{attribute}}$ that {\bf \textcolor{orange}{you}}{$^{\textcolor{red}{sender}}$} provide {\bf \textcolor{Fuchsia}{if you upload, sync or import this information (such as an address book) from a device.}}$^{\textcolor{red}{TP}}$} }% } \medskip\\ This flow contains an explicit sender, recipient, attribute, and transmission principle. The subject parameter is not included, but is implicitly the consumer agreeing to the privacy policy. We use the following guidelines to identify CI parameters within individual flows for annotation: \begin{itemize} \item {\bf Sender.} Any entity (person, company, website, device, etc.) that transfers or shares the information. This may be a pronoun or a specific entity, such as ``Company A,'' ``strategic partners,'' or ``publisher.'' \item {\bf Recipient.} Any entity (person, company, website, device, etc.) that ultimately receives the information. This may be a pronoun or a specific entity, such as ``third party,'' ``developer,'' ``other users,'' or ``Company B and its affiliates.'' \item {\bf Transmission principle.} Any clause describing the ``terms and conditions under which [...] transfers ought (or ought not) to occur''~\cite{nissenbaum2010privacy}. This includes descriptions of how information may be used or collected. Examples include ``if the user gives consent,'' ``when an update occurs,'' or ``to perform specified functions.'' \item {\bf Attribute.} Any description of information type, instance, and/or example, such as ``date of birth,'' ``credit card number,'' ``photos,'' or, more generally, ``personal information.'' \item {\bf Subject.} Any subjects of the information exchanged in a flow. Subjects may be explicitly stated or implicitly described using pronouns and possessives. \end{itemize} We perform manual policy annotation using Document Type Definitions (DTD) markup and the Multi-document Annotation Environment~\cite{MAT} (Figure~\ref{fig:annotation_tool}). \begin{figure}[t] \centering \includegraphics[width=0.45\textwidth]{images/mpa_tool.png} \caption{Screenshot of the Multi-document Annotation Environment tool configured for CI privacy policy annotation. } \label{fig:annotation_tool} \end{figure} \section{Facebook Case Study}\label{sec:case_study} Recent revelations about the misuse of consumer data by Facebook and Cambridge Analytica~\cite{bloomberg_story} has rekindled the debate around users' privacy and informed consent on such platforms. Facebook claims~\cite{wpost} that they provide users with the right level of control to keep their information private. They also claim that consumers are well informed by the disclosure of information handling practices in the company's privacy framework. Assuming that this is indeed the case, i.e., ignoring the complexity and a sporadic evolution of Facebook controls~\cite{lipford2008understanding,liu2011analyzing,johnson2012facebook}, we see the Cambridge Analytica scandal as another example of how things can go wrong when consumers' privacy expectations are misaligned with privacy policy statements. Much of the problem stems from not having a coherent higher-level abstraction that can help reason about privacy policies. While talented legal scholars and professionals are trained to identify relevant privacy policy excerpts and mentally stitch them into coherent flows, so to speak, the average consumer is usually overwhelmed by the legal language of privacy policies~\cite{turow2008consumers}. Even experts themselves find some privacy policy statements confusing~\cite{reidenberg2015disagreeable}. Furthermore, research shows that consumers' tend to ``[project] the important factors to their privacy expectations onto the privacy notice''~\cite{martin2015privacy}. In other words, consumers implicitly fill in the blanks left by difficult-to-interpret policies, which inadvertently widens the gap between their expectations and actual company behaviors. As a result of public outcry~\cite{bloomberg_story}, Facebook has amended its privacy policy to include a more detailed account of its information sharing practices. It is therefore timely and instructive to apply our CI annotation technique to the previous and updated Facebook privacy policies in order to demonstrate the power of the method and highlight issues with both policy versions. \subsection{Analysis} Using the methodology described in Section~\ref{sec:annotation-methodology}, we manually annotated Facebook's previous privacy policy (data policy) as well as the official updated version\footnote{https://www.facebook.com/about/privacy/update}. The following sections demonstrate the range of analyses that can be performed using CI annotations but are not exhaustive. We anticipate a variety of additional analytic techniques building on these annotations in future work. \input{sections/compare-param-table.tex} \subsubsection{Comparison of CI parameters} We compared the number information flows prescribed by both previous and updated Facebook privacy policies (Figure~\ref{fig:FB_summary}) and the CI parameters they contain. We matched CI parameters across policies using fuzzy string matching~\cite{cohen2011fuzzywuzzy} with the following thresholds for each CI parameter: sender (70\%), attribute (65\%), recipient (70\%), and transmission principle (55\%). While the fuzzy string matching worked well, some corner cases required manual validation. We describe some notable differences between information flows in the previous and updated policies on a parameter-by-parameter basis as follows: \begin{figure}[t] \centering \includegraphics[width=0.45\textwidth]{general_stats.pdf} \vspace{-1.8em} \caption{Distribution of unique CI parameters identified in the previous and updated Facebook privacy policies.} \label{fig:FB_summary} \end{figure} \textbf{Sender.} The updated policy offers a more detailed account of the sources of information transfer. It elaborates on categories from the previous privacy policy and also includes several new senders, such as ``WhatsApp'', ``connected TV'', ``a business'' which were not specified in the previous policy. Not surprisingly, the most frequent senders in both policies are Facebook and the consumer (Table~\ref{tbl:param_freq_summary}). \textbf{Recipient.} Similarly to the sender parameter, the updated version introduces new recipients, such as ``people and businesses outside the audience that you shared with,'' ``content creators,'' ``page admin,'' ``Instagram business profiles,'' and ``companies that aggregate.'' As expected, the most common ``recipients'' in both versions are ``Facebook,'' and ``third party service, vendors, partners'' (Table~\ref{tbl:param_freq_summary}). \begin{table}[t] \centering \begin{tabular}{llp{4.5cm}} CI Param & Version & Instances (frequency) \\ \midrule Recipients & Previous & we [Facebook] (22), Third party service, vendors, partners (20) \\ & Updated & we [Facebook] (32), Third party service, vendors, partners (24) \\ &&\\ Senders & Previous & we [Facebook] (14), you (6) \\ & Updated & we [Facebook] (17), you (11) \\ &&\\ Attributes & Previous & information (8), information about you (2), information we have (2), non-personally identifiable information only (2) \\ & Updated & information (15), content (5), information about you (4), information that we have (4), public information (4), communications (2), shipping and contact details (2). \\ \end{tabular} \caption{The most frequent recipients, senders, and attributes mentioned in the previous and updated Facebook privacy policies.} \label{tbl:param_freq_summary} \end{table} \textbf{Attribute.} When describing the types of information being transferred or collected, the updated policy contains more attributes (179) than the the previous policy (86). However, we note that some attributes from the previous policy were omitted in the update. The updated policy does not mention ``user id'' (opting for ``username'' instead), or ``age range'' (instead providing the example ``\dots ad was seen by a woman between the ages of 25 and 34''). Generally, the updated policy describes new types of information and/or elaborates on information that was previously generic or abstract (Table~\ref{tbl:compared_params}). For example, the updated draft provides significantly more details about the type of content that is being collected about the user, including ``racial or ethnic origins,'' ``health,'' ``events attended,'' ``interests,'' ``religious views,'' ``general demographics,'' ``political views,'' ``trade union membership,'' and ``philosophical beliefs.'' Furthermore, the updated policy describes attributes not discussed in the previous policy, such as ``connected TVs,'' ``information about nearby Wi-Fi access points,'' ``beacons,'' and ``cell towers.'' \textbf{Transmission Principle.} When specifying conditions under which information transfer may be performed, the updated policy includes all conditions and information flow constraints in the previous policy. In addition, the updated policy also contains new transmission principles, such as ``whether or not you have a Facebook account or are logged in to Facebook,'' ``to recognise you in photos, videos and camera experiences,'' ``reshared or downloaded through APIs,'' ``to have lawful rights to collect, use and share your data before providing any data to us,'' and many others (Table~\ref{tbl:compared_params}). \textbf{Subject.} The subject of most flows in both policies is the consumer. We therefore do not include the subject parameter in our analysis. \subsubsection{Incomplete Information Flows} \label{sec:missing} Our analysis of the Facebook privacy policies finds many prescribed information flows with missing (non-specified) parameters (Figure~\ref{fig:FB_missing_params}). Failing to specify parameters introduces ambiguity, leaving consumers uninformed about company behavior. In the previous privacy policy, $45\%$ (19/42) of flows are missing one or more parameters. In the updated policy, this number increases to $68\%$ (49/72). \begin{figure}[t] \centering \includegraphics[width=0.45\textwidth]{missing_stats.pdf} \vspace{-1.4em} \caption{Percentage of incomplete information flows in Facebook's previous and updated privacy policies with missing CI parameters.} \label{fig:FB_missing_params} \end{figure} {\bf Missing Recipient.} Table~\ref{tbl:missing_rcpt} lists the flows from both policies with missing recipient parameter. The previous policy only has one flow without an explicit recipient while the updated policy has two. Not stating information recipients forces users to infer what entities will have access to their information from other sources, often leading to incorrect notions of company behavior~\cite{turow2018persistent, martin2016measuring}. Identifying the recipient can sometimes be difficult, as in the flow {\em``We are able to suggest that your friend tags you in a picture by comparing your friend's pictures to information we've put together from your profile pictures and the other photos in which you've been tagged.''} \begin{table}[t] \centering \begin{tabular}{p{6cm}c} Information Flow & Version \\ \midrule When you comment on another person's post or like their content on Facebook, that person decides the audience who can see your comment or like & Previous \\ &\\ You can choose to provide information in your Facebook profile fields or life events about your religious views, political views, who you are ``interested in'' or your health. This and other information (such as racial or ethnic origin, philosophical beliefs or trade union membership) could be subject to special protections under the laws of your country & Updated\\ &\\ For example, people can share a photo of you in a story or mention, tag you at a location in a post or share information about you in their posts or messages & Updated \\ \end{tabular} \caption{Information flows in the previous and updated Facebook privacy policies with missing recipient parameters.} \label{tbl:missing_rcpt} \end{table} \textbf{Missing Sender.} The sender parameter is not specified in 14 ($33\%$) flows in the previous policy nor in 33 ($45\%$) flows in the updated policy. Many of the statements with missing senders describe ``use-of-data,'' i.e., they inform the consumer how the collected information will be used but not from where it is collected. Missing senders can easily lead to misinterpretations and false privacy expectations. For example, the source of the information in the following statement is unclear: {\em ``We collect information about the people, Pages, accounts, hashtags and groups you are connected to and how you interact with them across our Products, such as people you communicate with the most or groups you are part of.''} \textbf{Missing Transmission Principle.} We identified 6 information flows in the previous policy where the transmission principle was missing. For example, the statement ``{\em We share information we have about you within the family of companies that are part of Facebook}'' does not specify under what conditions/constraints the information is being shared. Likewise, the statement ``{\em We also collect information about how you use our Services, such as the types of content you view or engage with or the frequency and duration of your activities. Things others do and information they provide}'' does not contain any transmission principles. These statements force consumers to guess when and for what reason information is collected. The updated policy contains even more (15) flows with missing transmission principles. Without a transmission principle, flows like ``{\em We also receive information about your online and offline actions and purchases from third-party data providers who have the rights to provide us with your information''} become ambiguous because it is not clear when or why this information is being collected. \subsubsection{CI Parameter Bloating} Our CI annotation analysis also identifies several flows in both previous and updated policies with multiple CI parameters of the same type. We refer to this phenomenon as CI parameter bloating. Parameter bloating adds to the cognitive effort required to isolate single information flows from privacy policy statements, because it is often not clear which combinations of parameters describe information flows that actually take place. Consider the following flow from the updated policy: \medskip \noindent\fbox{% \parbox{0.45\textwidth}{% {\bf \textcolor{Orange}{Advertisers, app developers and publishers}}$^{\textcolor{red}{senders}}$ can send \textcolor{ForestGreen}{us}$^{\textcolor{red}{recipient}}$ information {\bf \textcolor{Fuchsia}{through Facebook Business Tools that they use, including our social plug-ins (such as the Like button), Facebook Login, our APIs and SDKs or the Facebook pixel}}{$^{\textcolor{red}{TP}}$}. These partners provide information about {\bf your}{$^{\textcolor{red}{subject}}$} {\bf \textcolor{blue}{activities off Facebook including information about your device, websites you visit, purchases you make, the ads you see and how you use their services whether or not you have a Facebook account or are logged in to Facebook}}$^{\textcolor{red}{attributes}}$.}% } \medskip \noindent At first glance, the above privacy statement seems transparent and informative. It explicitly specifies the type of information that is being exchanged, between what actors and under what conditions. However, this is an example of CI parameter bloating. The prescribed information flow is overloaded with CI parameters. Note the many senders (advertisers, app developers and publishers) attributes (information about your device, websites you visit, purchases you make, the ads you see and how you use their services), and transmission principles (when you use Like, Facebook login, APIs, SDKs and through Facebook Pixel). How does the consumer reason about this information flow? Do all listed senders transfer all of these information types to Facebook or does each particular sender transmit a specific information type? Do flows with each sender/information pair occur under each listed TP or only specific ones? Even technically-savvy users will have difficulty reasoning about the many possible information flows with all combinations of each parameter type. We would like to emphasize that specifying multiple instances of the same parameter does not automatically lead to parameter bloating. Specifically, parameter bloating does not include instances where a single parameter is enumerated to clarify a given category, as in the following statement, which elaborates on several attributes: \medskip \noindent\fbox{% \parbox{0.45\textwidth}{% { {\em \textcolor{ForestGreen}{We}$^{\textcolor{red}{recipients}}$ collect \textcolor{blue}{information about how use our Products, such as types of content you view or engage with, the features you use, the actions you take, the people or accounts you interact with and the time, frequency and duration of your activities}}$^{\textcolor{red}{attributes}}$.}} }% \medskip Figure~\ref{fig:param_frequency} shows the number of CI parameters per flow in both current and updated policies. In the previous policy, there are 10 information flows that mention more than one recipient, with one information flow standing out, listing 10 potential recipients. Three flows mention more than one sender, and 16 flows mention multiple attributes, ranging from 2 to 18 attributes per flow. Multiple transmission principles appear in 16 flows, ranging from 2 to 5 TPs per flow. \begin{figure}[t] \centering \includegraphics[width=0.5\textwidth]{frequency.pdf} \caption{Number of CI parameters per flow in Facebook's previous (\textit{top}) and updated (\textit{bottom}) privacy policies. The previous policy had one flow with 18 attributes and the updated policy has one flow with 40 attributes that are omitted for readability.} \label{fig:param_frequency} \end{figure} The updated policy contains even more bloated flows. Multiple senders appears in 8 information flows (2 senders in 6 flows, 3 in 1 flow, and 4 in 1 flow). Multiple attributes occur in 36 flows ranging from 2 attributes in 18 flows to 40 attributes in a single flow. Nineteen of the flows include more than one recipient (2 recipients in 14 flows, 3 in 4 flows, and 7 in 1 flow). Finally, the number of flows with multiple transmission principles increased to 30, ranging from 2 TPs in 14 flows to 8 TPs in a single flow. Given that an average consumer today spends little to no time reading privacy policies, it is unreasonable to assume that the even the most privacy-concern citizen will dissect all possible combinations of this many multi-parameter flows. \subsubsection{Vague and Ambiguous Flows}\label{sec:vague} CI annotation analysis also allow us to identify information flows that use vague terminology as defined in~\cite{bhatia2016theory} (Table~\ref{tbl:catgeories_of_vaguness}). \begin{table*}[t] \centering \begin{tabular}{lp{5cm}p{8cm}} {\bf Category} & {\bf Definition} & {\bf Example Terms} \\ \midrule Conditionality & it is not clear what is the condition associated with information transfer & ``as needed'', ``as necessary'', ``as appropriate'', ``depending'', ``sometimes'', ``as applicable'', ``otherwise reasonably determined'', ``from time to time''\\ & & \\ Generalization & action or information types are too abstract or vague & ``typically", ``normally", ``often" , ``general", ``usually", ``generally", ``commonly ", ``among other things", ``widely", ``primarily", ``largely", ``mostly" \\ & & \\ Modality & Hard to estimate the possibility of occurrence & ``likely", ``may", ``can", ``could" ``would", ``might", ``could", ``possibly" \\ & & \\ Numeric Quantifier & Vague numeric quantifier & ``certain", ``most", "majority", "many", "some" "few" \\ \end{tabular} \vspace{8pt} \caption{Summary of four vagueness categories as defined in~\cite{bhatia2016theory} and associated example terms.} \label{tbl:catgeories_of_vaguness} \end{table*} Figure~\ref{fig:FB_vagueness} shows the percentage of flows in Facebook's previous and updated policies that use vague terminology. In both policies, ``modality'' vagueness dominates, occurring in close to 45\% of all flows. The updated policy does not represent a reduction in vague terminology from the previous version. Rather, the percentage of flows with vague terms remains the same. This supports our initial claim the updated data policy does not contribute to clarity. The widespread occurrence of flows with vague wording further supports the problem that privacy policies are too often ``obtuse and noncommittal [and] make it difficult for people to know what information a site collects and how it will be used''~\cite{turow2008consumers}. \begin{figure}[t] \centering \includegraphics[width=0.45\textwidth]{vagueness_stats.pdf} \vspace{-1.5em} \caption{Percentage of information flows in Facebook's previous and updated privacy policies with various categories of vague wording (categories defined in Table~\ref{tbl:catgeories_of_vaguness})}. \label{fig:FB_vagueness} \end{figure} \subsection{Summary} The updated Facebook privacy policy has twice as many information flows as the previous policy (Figure~\ref{fig:FB_summary}). However, more information flows does not necessarily equal less confusion. Our analysis shows that many of the newly introduced information flows are either incomplete, overloaded with CI parameters and/or use vague terms. Rather than fix fundamental issues in their privacy policy in the wake of the Cambrige Analytica scandal, Facebook seems to have opted to add more terms, entities, and conditions. While this may initially seem to provide additional information to the consumer, CI annotation analysis reveals that there are still many issues preventing users from interpreting clear information flows from these new details and from understanding how their data is being collected and shared. \section{Discussion} We present a CI annotation methodology to help researchers and regulators assess and evaluate privacy policies. This work is a stepping stone in a larger effort to improve readability and increase transparency in disclosure of information handling practices. While philosophical in origin, the theory of CI offers a practical framework to reason about privacy implications in a given context and therefore serves as a powerful tool for reasoning about privacy preserving efforts in technical fields. The notion of an appropriate information flow in the CI framework lends itself well to user data privacy policies; privacy statements are essentially prescribed by the policy information flows. Annotating privacy policies with CI parameter labels offers a way to apply a full-fledged formal theory of privacy to their analysis. Relevant stakeholders---consumers, legal scholars, and regulators---can perform qualitative, quantitative and normative analysis to find incomplete, vague and ambiguous privacy statements. This also enables leveraging other applications of the CI framework. For example, it is possible to compare which flows prescribed by the policy align or do not align with consumers privacy expectations~\cite{apthorpe2018discovering}. As privacy policies evolve, CI annotations assist comparative analyses of new updates to identify which information flows were amended, added or removed. These analyses will ideally help companies write more coherent and complete privacy policies by identifying privacy statements containing missing, vague and/or bloated CI parameters. Furthermore, we can use our CI annotation crowdsourcing methodology to produce a large corpus of privacy policies annotations and discover trends and patterns in the types of flows that are being prescribed by policies within and across industries. This corpus could also be used as a training set to build tools for automatically identifying CI flows and parameters in privacy policies. \section {Limitations and Future Work} We have identified the following opportunities for further research to improve and streamline the CI annotation process: First, privacy policies are not written to intentionally fit the CI framework. As discussed in Sections~\ref{sec:case_study}~and~\ref{sec:crowdsourcing}, privacy policy terms can be ambiguous, vague, compound and even missing. This complicates the task of annotating privacy policy text with CI parameter labels. Nevertheless, our crowdsourcing annotations showed promising results on a diverse privacy statements from privacy policies of 17 companies. In future work, we intend to continue validating the CI annotation approach on larger policy samples. Second, our annotation methodology deals only with statements describing information transfers. These statements comprise the majority of privacy policy text and lend themselves to the CI framework. However, other statements, such as those describing how long information is stored, when and how information is purged, and what features allow users to fine tune privacy settings, fall outside the reasoning of the CI framework. Annotating these statements will require additional methodologies to complement our approach. A blended technique, such as combining CI annotation for information transfer statements with more general tags like those used by the Usable Privacy Project \cite{sadeh2013usable}, could provide the rigor of our CI technique with the flexibility to account for the diversity of information included in privacy policies. \section{Conclusion} This paper presents a methodology for analyzing privacy policies using annotations based on the theory of contextual integrity~\cite{nissenbaum2010privacy}. We perform a case study annotation of pre- and post-GDPR Facebook privacy policies and demonstrate that CI offers a rigorous way to examine privacy statements. We find that Facebook's post-GDPR privacy policy describes more total information flows with more parameters than the pre-GDPR version, but the updates do not improve the percentage of flows that contain vague language, omit parameters, or include many parameters of the same type. These issues impede interpretability, preventing users from clearly understanding how their information is being collected and shared. To further scale our approach, we present a method for crowdsourcing CI annotation of privacy policies. We test this method on 48 excerpts from 17 policies with 141 Amazon Mechanical Turk workers. Resulting high-precision crowdsourced annotations indicate that CI annotation is an intuitive method for interpreting privacy policies and that crowdsourcing could be used to obtain a large corpus of annotated privacy policies for future analysis. \section{Related work} Privacy policies are notoriously hard to read. As a result, average users find them difficult to comprehend and correctly interpret. This leads to gaps between users' expectations and the stated policy~\cite{martin2015privacy}. The problem of privacy policy comprehension has been the focus of many previous studies. Some efforts focused on lexical~\cite{sathyendra2016automatic, evans2017evaluation} and semantic~\cite{sathyendra2017identifying} analysis of the privacy policies. Others works~\cite{Wilson2016} used crowdsourcing to provide annotations that allow users to more easily parse privacy policies and identify sections related to specific concerns, as well as to help researchers assess policies from different websites. The Usable Privacy Policy Project (UPPP)~\cite{sadeh2013usable} has recruited law students to hand-annotate 115 privacy policies with metadata tags such as ``first party collection/use,'' ``user choice/control,'' ``data retention,'' and ``data security.'' They then used the hand-labeled policies to train a machine learning algorithm that has annotated over 7,000 policies with the same metadata tags~\cite{wilson2016creation}. While extracting relevant paragraphs saves time for the interested reader, it does not provide a way of identifying issues with the policy itself. It is remains up to the reader to interpret the text. This tends to create gaps between privacy expectations and policy statements, especially when policy statements are ambiguous or incomplete~\cite{martin2015privacy}. Recent work has shown evidence that privacy policies often elide or obscure crucial contextual information that could help users formulate their privacy expectations. In 2016, Martin and Nissenbaum~\cite{martin2016measuring} showed that when confronted with a privacy-related scenario that was missing some contextual information, respondents mentally supplemented the information, essentially generating a different version of the scenario. Martin and Nissenbaum also conducted a survey of 569 respondents presented with 40 scenarios with random combination of contextual factors. The results showed that the ``context of information exchange -- how information is used and transmitted, the sender and receiver of the information -- all impact the privacy expectations of individuals''~\cite{martin2016measuring}. The importance of including contextual factors was also reported by Rao et al., in a 2016 study that compared users' privacy expectations with existing companies' practices~\cite{RaoUsenixExptn2016}. 240 participants were asked to state their expectations for the data collection, sharing, and deletion practices of 16 websites across finance, health, and dictionary categories. The results showed that users' privacy expectations depend on the type of website and the type of information being exchanged. For example, respondents expected medical data to be shared with a medical website, but not a financial website. These findings provide further evidence to support the importance of contextual factors in how individuals perceive privacy practices, motivating a contextual analysis of privacy policies to identify gaps which might result in mismatched privacy expectations. Another body of work has explored using crowdsourcing to annotate privacy policies, thereby splitting the cognitive load of understanding an individual policy over multiple workers. In 2016, Wilson et. al.,~\cite{Wilson2016} explored the feasibility of asking crowdworkers to answer questions on data collection practices. In the experiment, 218 crowdworkers were assigned the task of reading through 12 privacy policies and answering 9 questions about data collection, sharing, and deletion practices stated in the policies. To support their answers, respondents needed to annotate the relevant text in the privacy policies. The results showed that the answers of the crowdworkers agreed with those of skilled annotators over 80\% of the time. The results indicate that crowdsourcing can be used to identify paragraphs describing specific practices in privacy policies. Our results support this conclusion, but extend it to even more sophisticated annotations of individual components of contextual information flows described in privacy policies. \section{Introduction} Many online services operate by collecting and sharing users' information. To protect consumers, the U.S.~Federal Trade Commission (FTC) devised fair information practice principles (FIPPs) based on the ``notice and choice'' framework~\cite{federal1998privacy}. These principles, in concert with state regulations, require companies to notify consumers about their information collection and sharing practices through privacy policies. These privacy policies, which often include details about the type of information collected, the entities that receive or store the information, and the conditions governing data acquisition and handling, serve two main purposes: 1) informing consumers about data collection practices, which they can consider when deciding whether or not to use a service, and 2) offering regulators, such as the FTC, a way to audit online services for misleading privacy practices. As we write this paper, the European General Data Protection Regulation (GDPR)~\cite{euGDPR} is coming into effect, forcing companies to adapt their behavior and rewrite their privacy policies or face strict penalties. The changes are largely based on GDPR Articles 13, 14, and 15, which outline the details companies need to provide to consumers when collecting, processing and sharing their information. The regulation puts an emphasis on providing this information to the ``subject in a concise, transparent, intelligible and easily accessible form, using clear and plain language''~\cite{Article12GDPR}. As a result, consumers are receiving an avalanche of updated privacy policies as companies strive for GDPR compliance~\cite{GDPREmails}. However, just because the GDPR has pushed companies to update their privacy policies does not necessarily mean that these updated policies address the issues of previous versions. In this paper, we make a case for using the theory of contextual integrity (CI)~\cite{nissenbaum2010privacy} to annotate, assess, and compare information sharing practices disclosed in privacy policies, both within and across updates. We showcase this technique with a case study in which we use the CI framework to manually annotate Facebook's previous and updated privacy policies to identify the {\em senders, recipients} and {\em subjects} of information, information types {\em (attributes)}, and the conditions under which information may be transferred or collected ({\em transmission principles}). We then use these annotations to gain insight into the privacy policy and amendments. Our analysis shows that while the updated privacy policy includes statements that describe almost as twice as many information flows as the current policy, it fails to provide more clarity to the consumer. In many cases, the updated policy has more incomplete and ambiguous information flow statements than the current policy. Incomplete information flow statements (45\% of all statements in the current policy and 63\% in the updated policy) do not mention one or many information flow parameters. This allows readers to interpret the missing parameters according to their own expectations, which may not match the actual practices of the company. In contrast, some statements in both current and updated policies suffer from what we refer to as ``parameter bloating,'' i.e., they contain more than one instance of each CI parameter. This increases the cognitive load required for consumers to fully comprehend all possible information flows allowed by the statement. Finally, we identified privacy statements (over 50\% in both current and updated policies) that use vague and ambiguous language. To help streamline our approach beyond the Facebook case study, we present a methodology for crowdsourcing CI privacy policy annotations. We construct CI annotation as a Amazon Mechanical Turk (AMT) Human Intelligence Task (HIT) and compare crowdsourced annotations against ground-truth expert annotations. We test the annotation task on 48 excerpts of privacy policies from 17 companies with 141 AMT workers. The crowdsourced annotations have an average word-based precision score of $0.9$ across CI information flow parameters. This high precision indicates that CI annotation is both understandable and easily applicable by those with no prior exposure to CI, despite the often legalistic language employed by privacy policies. This provides further evidence that CI successfully expresses how most people intuitively reason about information privacy. Finally, the high precision of crowdsourced annotations indicates that crowdsourcing could be applied at scale to evaluate future privacy policy updates or to build a dataset for training a machine learning model to perform automatic CI annotations. In summary, this work makes the following contributions: \begin{enumerate} \item We present a method for annotating privacy policies using the contextual integrity framework. The use of a structured framework allows rigorous analysis of difficult privacy policy texts that is applicable to policies across companies and sectors. \item We describe a case study using CI annotation to analyze recent updates to Facebook's privacy policy, which identifies several issues with information flow descriptions across versions. \item We demonstrate that crowdsourcing can produce precise CI annotations of legalistic privacy policy excerpts for future CI annotation research at scale. \end{enumerate} \section{CI Primer} The theory of CI is based on two central premises: 1) privacy is defined as the appropriateness of information flows, which 2) is defined by contextual norms governing particular settings (contexts) in which information is transmitted \cite{nissenbaum2010privacy}. CI offers a template for describing information flows using 5-parameter tuples, which include specific actors ({\em senders}, {\em recipients}, and {\em subjects}) involved in the information flow, the type ({\em attribute}) of the information, and the condition ({\em transmission principle}) under which the information flow occurs. This combination of five parameters defines contexts which determine privacy norms. For example, while someone might consider sharing Fitbit\footnote{https://www.fitbit.com/home} data with their doctor, they might view the sharing of this same data with advertising or insurance companies as a privacy violation. The entire context, including recipient and information type, affects how we think about privacy. The CI framework was previously used as a lens for examining android permissions~\cite{wijesekera2015android}, online platform practices~\cite{hull2011contextual,zimmer2008privacy}, and examining GDPR regulations~\cite{guinchard2017contextual} themselves. In more recent efforts, CI was employed to capture individuals' privacy expectations, which can be then checked for inconsistencies or used to inform policymakers and manufacturers~\cite{apthorpe2018discovering,shvartzshnaider2016learning}. \section{Crowdsourcing CI Annotations}\label{sec:crowdsourcing} The ability to effectively crowdsource CI annotation would allow researchers to efficiently pursue two primary goals: 1) collect a large dataset of annotations in order to train a machine learning model to perform CI annotation automatically, and 2) perform a large-scale analysis of information flows across the privacy policies of many companies. This would provide a broad sense of information flow disclosure practices across the technology sector via many of the same analysis methods used in the Facebook case study. We have developed a crowdsourcing technique that poses CI annotation as an Amazon Mechanical Turk (AMT) Human Intelligence Task (HIT). We crowdsourced the annotation of 48 privacy policy excerpts, including 16 excerpts from the Google privacy policy circa October 2017 and 16 pairs of excerpts from the pre-GDPR and post-GDPR privacy policies of 16 well-known companies\footnote{Amazon, Fitbit, Indiegogo, LinkedIn, The New York Times, Mirosoft, Shapeways, Slack, Spotify, Steam, Stripe, Tinder, Twitter, Uber, WhatsApp, Yelp}. This choice of policy excerpts provides evidence that our crowdsourcing technique is effective within a single policy as well as for privacy policies across the technology sector. The excerpt pairs were selected as representative statements from the pre-GDPR policies of each company and the corresponding statements from the GDPR-compliant version of each policy updated in May 2018. The excerpts ranged from 21 to 113 words\footnote{Mean: 55 words/excerpt, SD: 23 words/excerpt} and from 1 to 4 sentences for a total of 2621 annotated words over 103 annotated sentences. We compared the crowdsourced annotations to ground-truth annotations from a CI expert. The crowdsourced annotations had an average word-based precision of 0.9 across CI information flow parameters, indicating that the crowdworkers understood the relatively complex notion of information flow parameters and were able to correctly identify them in real privacy policy text. These results show that crowdsourcing can be an effectual tool for CI annotation. We will release the crowdsourced annotations as a public dataset for further analysis upon publication. Sections \ref{sec:task-design}--\ref{sec:crowdsource-discuss} describe the design and evaluation of our CI annotation crowdsourcing method in more detail. \subsection{Annotation Task Design} \label{sec:task-design} We developed the annotation task as a Qualtrics \cite{qualtrics} survey deployed on AMT. The task was designed to optimize annotation accuracy while minimizing cost. \textbf{Consent and Instructions.} The first page of the annotation task is a consent form. Participants who do not consent are prevented from proceeding. The annotation task collects no personal information about crowdworkers and was approved by our university's Institutional Review Board. The task next presents annotation instructions (Appendix Figure~\ref{fig:task-instructions}), including a description of each information flow parameter that should be annotated (sender, attribute, recipient, and transmission principle) and an example annotated flow. The information flow parameter descriptions match those used by expert annotaters as described in Section~\ref{sec:annotation-methodology}. \textbf{Screening Questions.} Each crowdworker is asked to annotate (highlight and label) all words and phrases corresponding to CI information flow parameters in three privacy policy excerpts (Figure~\ref{fig:screener-questions}). These excerpts serve as screening questions to identify workers who are able to perform high-accuracy annotations. Workers whose annotations have at least a 0.7 word-based F\textsubscript{1} score (Section~\ref{sec:eval-metrics}) compared to ground-truth expert annotations on the first screening question (for which the correct answer is given) and either of the next two screening questions are allowed to proceed with the task. Workers whose annotations do not meet this accuracy threshold do not proceed. This helps limit the effect and cost of workers who do not understand the task or who attempt to ``cheat'' by performing minimal annotations (e.g., highlighting just the first word in each excerpt). \begin{figure}[t] \centering \includegraphics[width=0.48\textwidth]{images/screener-questions.png} \caption{Screening questions to identify AMT workers who are able to perform high accuracy annotations. The ground truth annotations are shown with sender in blue, recipient in green, attribute in red, and transmission principle in purple. Common English stopwords (except ``you,'' ``your,'' ``them,'' and ``we'') are not counted when comparing crowdworker annotations to the ground truth.} \label{fig:screener-questions} \end{figure} \textbf{Annotations.} Each worker who passes the screening questions is then asked to annotate 5 of the 48 excerpts of interest, although these could be replaced with arbitrary privacy policy excerpts for future research. The format of these annotation questions is equivalent to the screening questions (Figure~\ref{fig:screener-questions}). The instructions are also repeated at the top of the page for workers to refer to if they wish. Annotations of all excerpts from multiple workers are collected, analyzed, and processed into the final crowdsourced annotation for each privacy policy (Section~\ref{sec:maj-vote}). The task concludes with a field for optional open-ended comments if participants have anything they wish to communicate to the researchers. \subsection{Task Deployment} We first tested the annotation task on UserBob~\cite{userbob}, a usability-testing service where users narrate their experience while performing tasks. We collected seven UserBob responses. All UserBob workers completed the task in less than 15 minutes. We used the UserBob responses to adjust task instructions to ameliorate worker confusion. Performing such ``cognitive interviews'' is common practice in survey design and development \cite{sudman1997thinking}. We deployed the annotation task as a HIT on AMT using TurkPrime \cite{litman2017turkprime}, an online tool for researchers to easily manage AMT tasks. We limited the HIT to AMT workers in the United States with an HIT approval rating of 90--100\% and at least 100 HITs approved. 141 total workers accepted the HIT. Of these workers, 99 passed the screener questions. All 48 excerpts were annotated by between 7 and 12 workers (mean~10.2). AMT workers who did not pass the screening questions were automatically reimbursed \$0.25. AMT workers who passed the screening test and completed the entire annotation task were reimbursed \$1.50. Collecting all responses took approximately 4 hours. \subsection{Majority Vote Annotations} \label{sec:maj-vote} We are ultimately interested in acquiring the single highest-accuracy annotation for each privacy policy independent of individual workers. We therefore combine multiple annotations of each privacy policy excerpt into a ``majority vote'' annotation, which assigns each word in an excerpt to the CI parameter annotated by at least 50\% of the participants presented with that excerpt. If fewer than 50\% of workers labeled a word with the same parameter, then the word is given no label in the majority vote annotation. \subsection{Evaluation Metrics} \label{sec:eval-metrics} We had one of the authors perform expert ground truth annotations of all excerpts prior to seeing the crowdsourced results. We use the following evaluation metrics to compare the crowdsourced majority vote annotations to the expert annotations. \textbf{Parameter-based scoring.} We manually counted all instances of each CI parameter labeled in both the crowdsourced majority vote and expert annotations (true positives), in the expert annotation only (false negatives), and in the crowdsourced annotation only (false positives). We further categorized the false positives and false negatives to better understand crowdworker mistakes and how to improve the annotation task in future studies (Section~\ref{sec:errors}). \textbf{Word-based scoring.} We also applied a automated word-based scoring method that did not require manually comparing variable-length parameters and could be used to easily evaluate future large-scale CI annotation efforts. We first removed common English stopwords from all annotations to prevent variations in article or preposition highlighting from affecting annotation comparisons. We used the stopword list in Python NLTK library \cite{bird2009natural} less ``you,'' ``your,'' ``them,'' and ``we,'' as these pronouns could have been senders or recipients in the privacy policy excerpts. True positives are then words labeled by both the participant and the expert. False positives are words labeled by the participant only. False negatives are words labeled by the expert only. This allows us to calculate word-based precision, recall, and F\textsubscript{1} scores for each CI parameter and excerpt. Some CI parameters do not occur in every excerpt. If the expert did not label a particular parameter in an excerpt, participants' recalls were defined as 1 for the corresponding annotation. If a participant did not label a particular element in an excerpt, the participant's precision was defined as 1 for the corresponding annotation. These are standard definitions of precision and recall for edge cases. \subsection{Annotation Accuracy} Figure~\ref{fig:manual-scores} shows the counts of correctly and incorrectly annotated CI parameters across all excerpts from parameter-based scoring. The incorrect annotations are divided into categories to better understand the source of crowdworker errors. The crowdsourced majority vote annotations correctly labeled 43\% of the senders, 89\% of the attributes, 68\% of the recipients, and 60\% of the transmission principles across all excerpts. False negatives were by far the most common error, with the crowdsourced annotations missing 30\% of the senders, 9\% of the attributes, 21\% of the recipients, and 34\% of the transmission principles across all excerpts. Finally, false positive errors comprised 26\% of the senders, 2\% of the attributes, 11\% of the recipients, and 6\% of the transmission principles across all excerpts.\footnote{Percentages were rounded to the nearest whole value and may not add to 100\%} \begin{figure}[t] \centering \includegraphics[width=0.48\textwidth]{manual_scores_all.pdf} \caption{Parameter-based evaluation of crowdsourced majority vote annotations compared to expert ground truth. Correct (true positive) annotations are parameters labeled to match the expert annotation. Skipped (false negative) annotations are parameters only labeled by the expert. All other incorrect annotations (false positives) are described in Section~\ref{sec:errors}. Note that most errors are skipped parameters (false negatives), indicating that the crowdworkers understood the task, but that further work is needed to improve recall. } \label{fig:manual-scores} \end{figure} Figure~\ref{fig:turkers-maj-pr} shows the distributions of word-based precision and recall scores for the majority vote annotations across all excerpts and for each CI parameter. The average precision across all excerpts is 0.95 for attributes, 0.80 for senders, 0.89 for recipients, and 0.94 for a transmission principles. The corresponding average recall across all excerpts is 0.87 for attributes, 0.82 for senders, 0.83 for recipients, and 0.59 for transmission principles. \begin{figure}[t] \centering \includegraphics[width=0.48\textwidth]{pr_hist_all.pdf} \caption{Word-based precision and recall scores of majority vote crowdsourced annotations compared to expert ground truth for each CI element.} \label{fig:turkers-maj-pr} \end{figure} Overall, the high precision of the majority vote crowdworker annotations (by both parameter-based and word-based scoring methods) indicates that the majority of crowdworkers understood the CI annotation task, and were able to correctly identify and highlight CI parameters in short privacy policy excerpts. However, the many false negatives indicates that the framing of the task could potentially be improved to help crowdworkers avoid missing or intentionally skipping some parameters. \subsection{Evaluating Crowdworker Errors} \label{sec:errors} Analyzing the crowdsourced annotations raises the question ``What causes particular excerpts or CI parameters to be more difficult for crowdworkers to annotate than others?'' One intuitive explanation is that excerpts that are longer, more difficult to read, or contain more CI parameters are more difficult for crowdworkers to annotate. To test this hypothesis, we calculated Spearman correlations of the majority vote annotation word-based F\textsubscript{1} scores versus text length, Flesch-Kincaid Reading Ease \cite{kincaid1975derivation}, FOG Index \cite{kincaid1975derivation}, and number of CI parameters (Appendix Table~\ref{tbl:readability}). However, all of the resulting correlation coefficients had absolute values less than $0.5$, indicating no strong correlations with F\textsubscript{1} score. This suggests that crowdworker difficulties with certain excerpts or parameters are due to more nuanced factors than length or readability. We further investigate these factors by manually comparing the crowdsourced majority vote annotations to the expert annotations. We noticed that crowdworkers had more difficulty annotating {\em senders} and {\em recipients} than {\em attributes} and {\em transmission principles}. Attributes and transmission principles are generally nouns or verbs, occur in lists, and require less semantic parsing to identify. In contrast, senders and attributes are often pronouns that occur singly and require more complex sentence parsing to distinguish between them. More detailed analysis indicated that the 160 parameter-based annotation errors fall into four main categories. Each category has corresponding implications for crowdsourcing CI annotations. \subsubsection{Expert Errors} We identified 11 cases where the majority vote crowdsourced annotation was correct while the ``ground-truth'' expert annotation was incorrect. Most of these cases were due to the expert missing a one-word sender or recipient, e.g. ``we.'' We did not adjust recall or precision scores to reflect the incorrect expert annotations, as these judgments were made after, and could have been influenced by, viewing the crowdsourced annotations. However, the presence of these incorrect expert annotations demonstrates the non-triviality of the annotation task. \subsubsection{Skipped Parameters} The most common error occurred when the crowdworkers simply neglected to annotate some or all instances of a given parameter. These errors were the primary contributor to lowering recall scores without affecting precision. We identified 117 skipped parameter errors. There are three possible reasons why crowdworkers might have neglected to annotate all instances of each parameter: 1) the workers may have considered an excerpt and honestly thought that it didn't contain the parameter, 2) the workers may have intentionally skipped entire parameters, or 3) the workers may have found one or two instances of each parameter and then moved on to the next excerpt without double-checking to ensure that none were missed. This could be due to cognitive fatigue or the fact that crowdworkers are incentivized to finish the annotations as quickly as possible to optimize their hourly compensation rate. As an example of reason 1, consider the sentence {\em ``We collect information when you sync non-content like your email address book, mobile device contacts, or calendar with your account.''} Both the expert and the crowdworkers correctly labeled ``email address book,'' ``mobile device contacts,'' and ``calendar'' as attributes. However, the expert also labeled ``information'' as an attribute, while the majority vote annotation did not. This was marked as a false negative ``skipped parameter'' error, but ``information'' does not provide any specific details about the attribute, so it is understandable that the crowdworkers omitted this label. This specific skipped parameter error (``information'' not labeled as attribute) occurred in 6 of the annotated excerpts. Skipped errors could potentially reduced in future crowdsourcing tasks by using previous crowdworker annotations to provide ``hints'' for successive workers. For example, the number of parameters annotated by previous workers could be shown (likely as a range) to indicate approximately how many parameters the current worker should find. This would help address reason 3 for skipped errors above, providing a nudge for workers finding fewer parameters to continue searching for additional annotations. However, such hints would have to be carefully applied to prevent individual crowdworker errors from negatively influencing the collective annotation effort. \subsubsection{Ambiguous Parameters} Ambiguous parameter errors occurred when a CI parameter was mislabeled compared to the expert annotation, but the correct labeling is ultimately open to interpretation. Consider the sentence {\em ``If you want to take full advantage of the sharing features we offer, we might also ask you to create a publicly visible Google Profile, which may include your name and photo.''} In this sentence, ``publicly'' could be interpreted as a recipient, i.e. the public would receive the data in the Google Profile. However, ``publicly'' could also be interpreted as a transmission principle i.e. the flow is from ``you'' to your ``Google Profile'' and the condition on the flow is that it is public. The expert labeled ``publicly'' as a recipient, while the crowdsourced majority did not. We only identified 3 such ambiguous parameter errors, indicating that CI information flow descriptions map naturally to privacy policy texts. \subsubsection{Overlapping Parameters} Overlapping parameter errors occurred when a CI parameter was mislabeled compared to the expert annotation, but the text in question is part of two or more CI parameters simultaneously. We identified 16 overlapping parameter errors. Consider the excerpt {\em ``When you use our services or view content provided by Google, we automatically collect and store certain information in server logs.''} The first clause (before the comma) could be interpreted as a single transmission principle, but the ``you'' could also be a sender. Variations on this issue were the primary cause of false positive errors for the ``sender'' parameter, i.e. the expert annotated an entire clause as a transmission principle but the majority vote annotation instead labeled a single word in the clause as a sender. The presence of overlapping parameter errors is due to a tradeoff in our implementation of the CI annotation task. We chose to allow only one CI parameter annotation per word in each excerpt to simplify the task for workers. This tradeoff could be avoided in future work by asking each crowdworker to annotate only a single CI parameter type, simplifying the task from multi-class classification to binary classification. However, this would require more crowdworkers per policy and could lead to higher rates of false positives if crowdworkers aren't forced to discriminate between different parameters. \subsubsection{True Errors} True errors occurred when the crowdworkers unambiguously misannotated a CI parameter. Fortunately, true errors accounted for only 13 out of 160 total errors in the majority vote annotation. This implies that when a label makes it into the majority vote annotation (with sufficient workers contributing to the vote), it is very likely to be correct. The low frequency of true errors indicates that, with improvements to reduce the number of skipped parameter errors, crowdsourcing can be a high-accuracy method of obtaining CI annotations of privacy policies. \subsection{Summary} \label{sec:crowdsource-discuss} Our proof-of-concept experiment shows that crowdworkers with no prior exposure to CI are able to quickly understand and perform CI annotations of legalistic privacy policies. Labels which make it into a majority vote annotation compiled from several individual crowdworkers are very likely correct. This supports the notion that CI-style information flows are a natural way for people to think about privacy and thereby a useful framework for analyzing privacy policies and privacy policy updates. \section*{Appendix} \centering \includegraphics[width=0.9\textwidth]{images/task-instructions.png} \caption{CI annotation task instructions.} \label{fig:task-instructions} \end{figure*} \clearpage \begin{table*}[t] \centering \begin{tabular}{p{2.2cm}|l|l|l} \textbf{Statistic} & CI Parameter & \textbf{Corr. coeff.} & \textbf{p-value}\\ \hline Total \# words & Attribute & -0.03 & 0.82 \\ & Sender & 0.03 & 0.86 \\ & Recipient & -0.11 & 0.46 \\ & TP & -0.15 & 0.30\\ &&&\\ \# words labeled & Attribute & 0.07 & 0.62 \\ as CI parameters & Sender & 0.10 & 0.48 \\ by expert & Recipient & 0.01 & 0.96 \\ & TP & -0.02 & 0.89 \\ &&&\\ Flesch-Kincaid & Attribute & 0.14 & 0.35 \\ Reading Ease & Sender & 0.20 & 0.18 \\ & Recipient & 0.10 & 0.49 \\ & TP & -0.05 & 0.76 \\ &&&\\ FOG Index & Attribute & 0.15 & 0.32 \\ & Sender & 0.19 & 0.19 \\ & Recipient & 0.10 & 0.50 \\ & TP & -0.06 & 0.67 \\ \end{tabular} \vspace{6pt} \caption{Spearman correlations of majority vote annotation word-based F\textsubscript{1}~scores for each CI parameter versus various statistics of corresponding privacy policy excerpts.} \label{tbl:readability} \end{table*}

679d911f88af871b4b621ce148175c54c68de707

\section{Introduction} Consider an $n \times n$ array $A$ in which every cell $(i,j)$ contains a subset $A(i,j)$ of the symbols in $[n]=\{1, \dots,n \}$. If every cell contains at most $m$ symbols, and every symbol occurs at most $m$ times in every row and column, then $A$ is an {\em $(m,m,m)$-array}. Confirming a conjecture by H\"aggkvist \cite{Haggkvist}, it was proved in \cite{AndrenCasselgrenOhman} that there is a constant $c>0$ such that if $m \leq cn$ and $A$ is an $(m,m,m)$-array, then $A$ is {\em avoidable}; that is, there is a Latin square $L$ such that for every $(i,j)$ the symbol in position $(i,j)$ in $L$ is not in $A(i,j)$ (see also \cite{Lina, AndrenCasselgrenMarkstrom}). The purpose of this note is to prove an analogue of this result for Latin cubes of order $n=2^k$. In order to make this precise, we imagine a $3$-dimensional array having layers stacked on top of each other; we shall refer to such a $3$-dimensional array as a {\em cube}. Now, a cube has {\em lines} in three directions obtained from fixing two coordinates and allowing the third to vary. The lines obtained by varying the first, second, and third coordinates will be referred respectively as {\em columns}, {\em rows}, and {\em files}. The first, second, and third coordinates themselves will be referred to as the indices of the rows, columns, and files. A \textit{Latin cube} $L$ of order $n$ on the symbols $\{1,\dots,n\}$ is an $n \times n \times n$ cube such that each symbol in $\{1,\dots,n\}$ appears exactly once in each row, column and file. The symbol in position $(i,j,k)$ of $L$ is denoted by $L(i,j,k)$. Latin cubes have been studied by a number of authors, both with respect to enumeration and e.g. extension from partial cubes. An extensive survey of early results can be found in \cite{MckWan}. An $n \times n \times n$ cube where each cell contains a subset of the symbols in the set $\{1,\dots,n\}$ is called an \textit{$(m,m,m,m)$-cube (of order $n$)} if the following conditions are satisfied: \begin{itemize} \item[(a)] No cell contains a set with more than $m$ symbols. \item[(b)] Each symbol occurs at most $m$ times in each row. \item[(c)] Each symbol occurs at most $m$ times in each column. \item[(d)] Each symbol occurs at most $m$ times in each file. \end{itemize} Let $A(i,j,k)$ denote the set of symbols in the cell $(i,j,k)$ of $A$. If we simplify notation, and write $A(i,j,k)=q$ if the set of symbols in cell $(i,j,k)$ of $A$ is $\{q\}$, then a $(1,1,1,1)$-cube is a {\em partial Latin cube}, and a \textit{Latin cube} $L$ is simply a $(1,1,1,1)$-cube with no empty cell. Given an $(m,m,m,m)$-cube $A$ of order $n$, a Latin cube $L$ of order $n$ \textit{avoids} $A$ if there is no cell $(i,j,k)$ of $L$ such that $L(i,j,k) \in A(i,j,k)$; if there is such a Latin cube, then $A$ is \textit{avoidable}. Problems on extending partial Latin cubes have been studied for a long time, with the earliest results appearing in the 1970s \cite{Cruse}; in the more recent literature we have \cite{Bri, Bryant,KuhlDenley,DenleyOhman}. The more general problem of constructing Latin cubes subject to the condition that some symbols cannot appear in certain cells seems to be a hitherto quite unexplored line of research. Our main result is the following, which establishes an analogue of the main result of \cite{Lina}, which considered Latin squares, for Latin cubes. \begin{theorem} \label{maintheorem} There is a positive constant $\gamma$ such that if $t \geq 30$ and $m \leq \gamma 2^t$, then any $(m,m,m,m)$-cube $A$ of order $2^t$ is avoidable. \end{theorem} The restriction on the order of the cube is not believed to be necessary, but as for Latin squares, general orders are expected to require far more technical proof (unless some completely new method is invented). Our proof establishes this result for a small value of $\gamma$ which we believe to be far from the optimal one, much like the case for the similar results for Latin squares. We know from \cite{CulterOhman} that $\gamma \leq \frac{1}{3}$, since that is an upper bound for the corresponding result for Latin squares, and every $n\times n$ sub-array of an avoidable $(m,m,m,m)$-cube of order $n$ must be avoidable (in the sense that there is an $n\times n$ Latin square that avoids this array). An interesting question is how sparse an unavoidable $(m,m,m,m)$-cube can be if every square $n \times n$ sub-array is avoidable. \begin{problem} \label{problem} For how small $\gamma'=\frac{m}{n}$ does there exist an unavoidable $(m,m,m,m)$-cube $A$ of order $n$, where every square sub-array of order $n$ is avoidable for Latin squares? \end{problem} We note that $\gamma' \leq 1/2$, since there are unavoidable $(n/2,n/2,n/2,n/2)$-cubes of order $n$ that satisfies the condition in Problem \ref{problem}; such a cube can be obtained by dividing an $n\times n\times n$ cube into $8$ subcubes of equal order $n/2$, and putting symbols $1,\dots, n/2$ in all cells of two subcubes in ``opposite'' corners of the larger cube. We may also note that the main result of this paper, as well as the problem of extending partial Latin cubes, can be recast as list edge coloring problems on the complete $3$-uniform $3$-partite hypergraph $K^3_{n,n,n}$. Problems on extending partial edge colourings for ordinary graphs have been studied to some extent, see e.g. \cite{EGHKPS, GiraoKang} and the references given there, but similar problems for hypergraphs remain mostly unexplored. In Section 2 we give some definitions and preparatory lemmas, and in Section 3 we prove Theorem \ref{maintheorem}. \section{Definitions and properties of Boolean Latin cubes} In this section we give some definitions and collect essential properties of Boolean Latin cubes. Let $A$ be an $n \times n \times n$ cube. Given $i \in [n]$, \textit{row layer} $i$ in $A$ is a set of $n^2$ cells $\{(i,j^*,k^*) : j^* \in [n], k^* \in [n] \}$; given $j \in [n]$, \textit{column layer} $j$ in $A$ is a set of $n^2$ cells $\{(i^*,j,k^*) : i^* \in [n],k^* \in [n]\}$; given $k \in [n]$, \textit{file layer} $k$ in $A$ is a set of $n^2$ cells $\{(i^*,j^*,k) : i^* \in [n], j^* \in [n]\}$. As mentioned above, by fixing two coordinates and varying the third, we obtain rows, columns and files of a $n \times n \times n$ cube. Formally we define a row of such a cube $A$ as a set of cells $R_{i,k} = \{(i,j^*,k) : j^* \in [n]\}$, a column as the set $C_{j,k} = \{(i^*,j,k) : i^* \in [n]\}$, and files $F_{i,j} = \{(i,j,k^*) : k^* \in [n]\}$. \begin{definition} The {\em Boolean Latin square} of order $2^t$ is the Latin square with entries as in the addition table of $\mathbb{Z}^t_2$ with the elements of $\mathbb{Z}^t_2$ mapped to the integers $1,\dots,2^t$. \end{definition} A {\em $4$-cycle} in a Latin square $L$ is a set of four cells $\{(i_1,j_1), (i_1,j_2), (i_2,j_1), (i_2,j_2)\}$ such that $L(i_1,j_1)=L(i_2,j_2)$ and $L(i_1,j_2)=L(i_2,j_1)$. We note some important properties of Boolean Latin squares (cf. \cite{Lina}). \begin{property} \label{prop:cycles} Each cell in the $n \times n$ Boolean Latin square is in $n-1$ distinct $4$-cycles. Permuting the rows, the columns or the symbols does not affect the number of $4$-cycles that a cell is part of. \end{property} \begin{property} \label{pro4cyc} A $4$-cycle in the Boolean Latin square is uniquely determined by two cells; that is, if $C$ is a $4$-cycle and $(i_1,j_1), (i_1,j_2) \in C$, then $(i_2,j_1), (i_2,j_2) \in C$, where $i_2$ is the row such that $L(i_1,j_1) = L(i_2,j_2)$ and $L(i_1,j_2)= L(i_2,j_1)$. \end{property} \begin{property} \label{two4cycle} The intersection of two $4$-cycles is either empty, or it contains $1$ or $4$ cells. \end{property} Given an integer $t$, let $a_i$ ($1 \leq i \leq 2^t$) be the $i$th smallest element of $\mathbb{Z}^t_2$. (For example, with $t=2$, $a_1=00, a_2=01, a_3=10, a_4=11$.) We define the \textit{Boolean Latin cube} similarly as the Boolean Latin square. \begin{definition} \label{defBc} The {\em Boolean Latin cube} $B$ of order $n=2^t$ on the symbols $\{1,\dots,n\}$ is an $n\times n \times n$ Latin cube such that $B(i,j,k)=x$ with $a_x=a_i+a_j+a_k$ (addition in $\mathbb{Z}^t_2$) for all $1\leq i,j,k\leq n$. \end{definition} \begin{definition} A {\em $3$-cube} in a Latin cube $L$ is a set of eight cells $$\{(i_1,j_1,k_1), (i_1,j_2,k_1), (i_2,j_1,k_1), (i_2,j_2,k_1), (i_1,j_1,k_2), (i_1,j_2,k_2), (i_2,j_1,k_2), (i_2,j_2,k_2)\}$$ such that $$L(i_1,j_1,k_1)=L(i_2,j_2,k_1)=L(i_1,j_2,k_2)=L(i_2,j_1,k_2)$$ and $$L(i_1,j_2,k_1)= L(i_2,j_1,k_1)=L(i_1,j_1,k_2)= L(i_2,j_2,k_2).$$ \end{definition} Note that every row, column and file layer of the Boolean Latin cube is a Boolean Latin square. For the Boolean Latin cube we have the following analogue of Property \ref{prop:cycles}. \begin{property} Each cell in the Boolean Latin cube of order $n$ belongs to $n-1$ $3$-cubes. \end{property} \begin{proof} Consider an arbitrary cell $(i_1,j_1,k_1)$ of the Boolean Latin cube $B$ which belongs to a $4$-cycle $\mathfrak{c}_1=\{(i_1,j_1,k_1), $ $ (i_1,j_2,k_1), $ $ (i_2,j_1,k_1), $ $ (i_2,j_2,k_1)\}$ such that $B(i_1,j_1,k_1)=B(i_2,j_2,k_1)$ and $B(i_1,j_2,k_1)= B(i_2,j_1,k_1)$. There are $n-1$ $4$-cycles $\mathfrak{c}_1$ in file layer $k_1$ containing $(i_1,j_1,k_1)$, since by construction, the file layers of the Boolean Latin cube are isomorphic to Boolean Latin squares; this also holds for row and column layers. Now, by Property \ref{two4cycle}, the two cells $(i_1,j_1,k_1)$ and $(i_2,j_1,k_1)$ define a unique $4$-cycle $$\mathfrak{c}_2=\{(i_1,j_1,k_1), (i_2,j_1,k_1), (i_1,j_1,k_2), (i_2,j_1,k_2)\}$$ in the column layer $j_1$ such that $B(i_1,j_1,k_1)=B(i_2,j_1,k_2)$ and $B(i_2,j_1,k_1)= B(i_1,j_1,k_2)$. By Definition \ref{defBc}, $$a_{i_1}+a_{j_1}+a_{k_1}= a_{i_2}+a_{j_2}+a_{k_1} =a_{i_2}+a_{j_1}+a_{k_2}$$ and $$a_{i_1}+a_{j_2}+a_{k_1}= a_{i_2}+a_{j_1}+a_{k_1}= a_{i_1}+a_{j_1}+a_{k_2}.$$ Hence, we have $$a_{i_1}+a_{j_1}+a_{k_2}= a_{i_2}+a_{j_2}+a_{k_2} =a_{i_2}+a_{j_1}+a_{k_1}$$ and $$a_{i_1}+a_{j_2}+a_{k_2}= a_{i_2}+a_{j_1}+a_{k_2}= a_{i_1}+a_{j_1}+a_{k_1};$$ or, in other words, $$B(i_1,j_2,k_1)= B(i_2,j_1,k_1)=B(i_1,j_1,k_2)= B(i_2,j_2,k_2)$$ and $$B(i_1,j_1,k_1)=B(i_2,j_2,k_1)=B(i_1,j_2,k_2)=B(i_2,j_1,k_2).$$ This implies that $$\{(i_1,j_1,k_1), (i_1,j_2,k_1), (i_2,j_1,k_1), (i_2,j_2,k_1), (i_1,j_1,k_2), (i_1,j_2,k_2), (i_2,j_1,k_2), (i_2,j_2,k_2)\}$$ is a $3$-cube; and so each cell in the Boolean Latin cube belongs to $n-1$ $3$-cubes. \end{proof} \begin{property} \label{traversalsetproperty} Let $(i_1,j_1,k_1)$, $(i_2,j_2,k_2)$, $(i_3,j_3,k_3)$ be three cells in the Boolean Latin cube $B$ such that $(i_1 -i_2)(j_1-j_2)(k_1-k_2) \neq 0$, $(i_1 -i_3)(j_1-j_3)(k_1-k_3) \neq 0$ and $(i_2 -i_3)(j_2-j_3)(k_2-k_3) \neq 0$. If $(i_1,j_1,k_1)$ and $(i_2,j_2,k_2)$ both are in a $3$-cube $\mathcal{C}_1$, and $(i_1,j_1,k_1)$ and $(i_3,j_3,k_3)$ are in a $3$-cube $\mathcal{C}_2$, then $(i_2,j_2,k_2)$ and $(i_3,j_3,k_3)$ are in a $3$-cube $\mathcal{C}_3$. \end{property} \begin{proof} Assume $B(i_1,j_1,k_1)=x$, $B(i_2,j_2,k_2)=y$, $B(i_3,j_3,k_3)=z$. Since $\mathcal{C}_1$ and $\mathcal{C}_2$ are $3$-cubes, we have that $B(i_2,j_2,k_1)=B(i_1,j_1,k_1)=B(i_3,j_3,k_1)$, i.e, $a_{i_2} + a_{j_2} + a_{k_1} = a_{i_1} + a_{j_1} + a_{k_1} = a_{i_3} + a_{j_3} + a_{k_1}$. It follows that $a_{i_2} + a_{j_2}= a_{i_3} + a_{j_3}$, so $a_{i_2}+a_{j_2} + a_{k_2} = a_{i_3} +a_{j_3} + a_{k_2}$, which implies that $B(i_3,j_3,k_2)=B(i_2,j_2,k_2)=y$. Similarly, we have $B(i_3,j_2,k_3)=B(i_2,j_3,k_3)=B(i_2,j_2,k_2)=y$ and $B(i_3,j_2,k_2)=B(i_2,j_3,k_2)=B(i_2,j_2,k_3)=B(i_3,j_3,k_3)=z$, which implies that $(i_2,j_2,k_2)$ and $(i_3,j_3,k_3)$ are two cells of a $3$-cube $$\mathcal{C}_3=\{(i_2,j_2,k_2), (i_2,j_3,k_2), (i_3,j_2,k_2), (i_3,j_3,k_2), (i_2,j_2,k_3), (i_2,j_3,k_3), (i_3,j_2,k_3), (i_3,j_3,k_3)\}.$$ \end{proof} \begin{property} The intersection of two $3$-cubes in a Latin cube is either empty, or it contains $1$ or $8$ cells. \end{property} \begin{proof} Assume that the intersection of two given $3$-cubes contains at least $2$ cells. If these $2$ cells lie in a $4$-cycle of a layer of the Latin cube, then by Property \ref{two4cycle}, this $4$-cycle belongs to the intersection of two $3$-cubes. But each $4$-cycle defines a unique $3$-cube, which implies that the intersection of the two $3$-cubes contains $8$ cells. If not, these $2$ cells must have distinct row, column and file coordinates, so if we denote these two cells by $(i_1,j_1,k_1)$ and $(i_2,j_2,k_2)$, respectively, then $i_1 \neq i_2$, $j_1 \neq j_2$, $k_1 \neq k_2$. Hence, the intersection of the two $3$-cubes must be the $8$ cells $(i_1,j_1,k_1), $ $ (i_1,j_2,k_1), $ $ (i_2,j_1,k_1), $ $ (i_2,j_2,k_1),$ $ (i_1,j_1,k_2), $ $(i_1,j_2,k_2), $ $ (i_2,j_1,k_2), $ $ (i_2,j_2,k_2)$. \end{proof} \begin{definition} Given a $3$-cube $$\mathcal{C}=\{(i_1,j_1,k_1), (i_1,j_2,k_1), (i_2,j_1,k_1), (i_2,j_2,k_1), (i_1,j_1,k_2), (i_1,j_2,k_2), (i_2,j_1,k_2), (i_2,j_2,k_2)\}$$ in a Latin cube $L$, a \textit{swap on $\mathcal{C}$} (or simply a \textit{swap}) denotes the transformation $L \rightarrow L'$ which retains the content of all cells of $L$ except that if $$L(i_1,j_1,k_1)=L(i_2,j_2,k_1)=L(i_1,j_2,k_2)=L(i_2,j_1,k_2)=x_1$$ and $$L(i_1,j_2,k_1)= L(i_2,j_1,k_1)=L(i_1,j_1,k_2)= L(i_2,j_2,k_2)=x_2$$ then $$L'(i_1,j_1,k_1)=L'(i_2,j_2,k_1)=L'(i_1,j_2,k_2)=L'(i_2,j_1,k_2)=x_2$$ and $$L'(i_1,j_2,k_1)= L'(i_2,j_1,k_1)=L'(i_1,j_1,k_2)= L(i_2,j_2,k_2)=x_1.$$ \end{definition} \begin{property} \label{rcfblock} Consider an arbitrary column $\{(i_1,j_1,k_1),\dots, (i_n,j_1,k_1)\}$ of a Boolean Latin cube $B$ of order $n$. For any $k_2$ ($j_2$), there exists a unique $j_2$ ($k_2$), such that $B(x,j_1,k_1)=B(x,j_2,k_2)$ for every $x \in \{1,\dots,n\}$. \end{property} \begin{proof} For any $k_2$, we can choose $j_2$ satisfying $a_{j_2}=a_{j_1}+a_{k_1}-a_{k_2}$, and for any $j_2$, we can choose $k_2$ satisfying $a_{k_2}=a_{j_1}+a_{k_1}-a_{j_2}$. \end{proof} Evidently, all rows and files of a Boolean Latin cube have corresponding properties. \begin{property} \label{Srcfblock} Let $B$ be a Boolean Latin cube of order $n$, $\mathfrak{b}$ an arbitary symbol in $B$, and $S_1$ be the set of cells of $B$ in the first row layer which contain $\mathfrak{b}$. For any row layer $i$, the set of cells $S_i$ of $B$ in row layer $i$ which have the same column and file coordinates as cells in $S_1$ all contain the same symbol. \end{property} \begin{proof} Assume $(i,j_1,k_1) \in S_i$ and $B(i,j_1,k_1)=x$, and consider an arbitrary cell $(i,j_2,k_2) \in S_i$. By definition, there are two cells $(1,j_1,k_1)$ and $(1,j_2,k_2)$ such that $B(1,j_1,k_1) = B(1,j_2,k_2) = \mathfrak{b}$, that is, $a_1 + a_{j_1} + a_{k_1} = a_1 + a_{j_2} + a_{k_2}$. This implies that $a_i + a_{j_1} + a_{k_1} = a_i + a_{j_2} + a_{k_2}$, which means that $B(i,j_2,k_2)=B(i,j_1,k_1)=x$. Hence, all cells in $S_i$ contain the same symbol. \end{proof} Note that all column and file layers of $B$ have the same property. The following simple observation enables us to permute layers and symbols in a Latin cube. \begin{property} If $L$ is a Latin cube, then the cube obtained by permuting the row layers, the column layers, the file layers and/or the symbols of $L$ is a Latin cube. \end{property} For Boolean Latin cubes an even stronger property holds. If a Latin cube $L'$ is obtained from another Latin cube $L$ by permuting row/column/file layers and/or symbols of $L$, then we say that $L$ and $L'$ are {\em isomorphic}. Henceforth, all Latin cubes have order $n$. \begin{property} If $L$ is isomorphic to a Boolean Latin cube, then any cell of $L$ is in $n-1$ $3$-cubes. Moreover, Property \ref{traversalsetproperty}, \ref{rcfblock}, and \ref{Srcfblock} hold for $L$. \end{property} \bigskip In the following we shall define some sets of cells in Latin cubes that are isomorphic to Boolean Latin cubes. \begin{definition} Let $L$ be a Latin cube that is isomorphic to a Boolean Latin cube. A \textit{row block} of $L$ is a set of $n$ rows $R_{i,k}$ such that for every pair of rows $R_{i_1,k_1}=\{(i_1,j,k_1): j \in [n]\}$ and $R_{i_2,k_2} = \{(i_2,j,k_2): j \in [n]\}$ in this set, $B(i_1,x,k_1)=B(i_2,x,k_2)$ for every $x \in \{1,\dots,n\}$. It is obvious that there are $n$ row blocks in total. \textit{Column blocks} and \textit{file blocks} are defined similarly. \end{definition} \begin{property} \label{cellbelongsrowblock} If $$\mathcal{C}=\{(i_1,j_1,k_1), (i_1,j_2,k_1), (i_2,j_1,k_1), (i_2,j_2,k_1), (i_1,j_1,k_2), (i_1,j_2,k_2), (i_2,j_1,k_2), (i_2,j_2,k_2)\}$$ is a $3$-cube in a Latin cube $L$ that is isomorphic to a Boolean cube, then the two rows $R_{i_1,k_1}$ and $R_{i_2,k_2}$ are in the same row block, as are also the two rows $R_{i_2, k_1}$ and $R_{i_1, k_2}$. \end{property} Note that a similar property holds for columns blocks and file blocks. \begin{definition} If $L$ is a Latin cube that is isomorphic to a Boolean Latin cube, a \textit{transversal-set} $\mathfrak{t}$ of $L$ is a set of $n$ cells that satisfy the following \begin{itemize} \item no two cells in $\mathfrak{t}$ are in the same row/column/file; \item no two cells in $\mathfrak{t}$ contain the same symbol; \item for any two cells in $\mathfrak{t}$, there is a unique $3$-cube that contain these cells. \end{itemize} \end{definition} Note that by Property \ref{traversalsetproperty}, a transversal-set is well-defined, and every row block, column block and file block contains exactly $n$ disjoint transversal-sets. Based on Property \ref{Srcfblock}, we make the following definition. \begin{definition} A \textit{symbol-row block} of a Latin cube $L$ that is isomorphic to a Boolean Latin cube is a set $\mathfrak{s}$ of $n^2$ cells satisfying that \begin{itemize} \item all cells of $\mathfrak{s}$ that are in the same row layer contain the same symbol, and \item for every cell of $\mathfrak{s}$, there are $n-1$ other cells that have the same column and file coordinate. \end{itemize} \textit{Symbol-column blocks} and \textit{symbol-file blocks} are defined similarly. \end{definition} An intersection between a symbol-row block and a row layer (or a symbol-column block and a column layer, or a symbol-file block and a file layer) is called a \textit{symbol-set}. It is obvious that all cells in a symbol-set contain the same symbol, and that each row layer, column layer, file layer, symbol-row block, symbol-column block, and symbol-file block contains $n$ symbol-sets. \begin{definition} A \textit{symbol block} of a Latin cube $L$ is a set of $n^2$ cells such that all these cells contain the same symbol. \end{definition} Note that a Latin cube that is isomorphic to a Boolean Latin cube contains $n$ symbol blocks in total, and for each symbol block, there are three different ways to divide this symbol block to $n$ disjoint symbol-sets (group the symbol sets based on the row layers, the column layers or the file layers). \bigskip Given an $n\times n \times n$ cube $A$ where each cell contains a subset of the symbols in $\{1,\dots,n\}$, and a Latin cube $L$ of order $n$ that does not avoid $A$, we say that those cells $(i,j,k)$ of $L$ where $L(i,j,k) \in A(i,j,k)$ are \textit{conflict cells of $L$ with $A$} (or simply {\em conflicts} of $L$). An \textit{allowed $3$-cube} of $L$ is a $3$-cube $$\mathcal{C}=\{(i_1,j_1,k_1), (i_1,j_2,k_1), (i_2,j_1,k_1), (i_2,j_2,k_1), (i_1,j_1,k_2), (i_1,j_2,k_2), (i_2,j_1,k_2), (i_2,j_2,k_2)\}$$ in $L$ such that swapping on $\mathcal{C}$ yields a Latin cube $L'$ where none of $(i_1,j_1,k_1), $ $ (i_1,j_2,k_1), $ $ (i_2,j_1,k_1), $ $ (i_2,j_2,k_1),$ $ (i_1,j_1,k_2), $ $(i_1,j_2,k_2), $ $ (i_2,j_1,k_2), $ $ (i_2,j_2,k_2)$ is a conflict. \section{Proof of the main theorem} In this section we prove Theorem \ref{maintheorem}. Our basic proof strategy is similar to the one in \cite{Lina,AndrenCasselgrenOhman}; however, due to the extra dimension in a Latin cube, our arguments are considerably more involved and somewhat technical. Our starting point in the proof is the Boolean Latin cube; we permute its row layers, column layers, file layers and symbols so that the resulting Latin cube does not have too many conflicts with a given $(m,m,m,m)$-cube $A$. After that, we find a set of allowed $3$-cubes such that each conflict belongs to one of them, with no two of the $3$-cubes intersecting, and swap on those $3$-cubes. The proof of Theorem \ref{maintheorem} involves a number of parameters: $$\alpha, \gamma, \kappa, \epsilon, \theta,$$ and a number of inequalities that they must satisfy. For the reader's convenience, explicit choices for which the proof holds are presented here: $$\alpha = 1-38 \times 2^{-25}, \gamma=2^{-25}, \kappa= 6 \times 2^{-25}, \epsilon = 2^{-6}, \theta=2^{-12}.$$ By an example of an unavoidable $(\lfloor {\frac{n}{3}} \rfloor+1, \lfloor {\frac{n}{3}} \rfloor+1, \lfloor {\frac{n}{3}} \rfloor+1)$-arrays in \cite{CulterOhman}, the value of $\gamma$ for which Theorem \ref{maintheorem} holds cannot exceed $\frac{1}{3}$. Thus, since the numerical value of $\gamma$ for which the theorem holds is not anywhere near what we expect to be optimal, we have not put an effort into choosing optimal values for these parameters. Moreover, for simplicity of notation, we shall omit floor and ceiling signs whenever these are not crucial. We shall establish that our main theorem holds by proving two lemmas. \begin{lemma} Let $\alpha, \gamma, \kappa$ be constants and $n=2^t$ such that $$ \Big(7n^2 \dfrac{(\gamma n)^{\kappa n}}{(\kappa n)!} +3n^3\dfrac{{(2\gamma n)^{(1-\alpha-2\gamma)n/3}}}{((1-\alpha-2\gamma)n/3)!}\Big) < 1.$$ For any $(\gamma n, \gamma n, \gamma n, \gamma n)$-cube $A$ of order $n$ there is a quadruple of permutations $\sigma=(\tau_1, \tau_2, \tau_3, \tau_4)$ of the row layers, the column layers, the file layers and the symbols of the Boolean Latin cube $B$ of order $n$, respectively, such that applying $\sigma$ to $B$, we obtain a Latin cube $L$ satisfying the following: \begin{itemize} \item[(a)] No row in $L$ contains more than $\kappa n$ conflicts with $A$. \item[(b)] No column in $L$ contains more than $\kappa n$ conflicts with $A$. \item[(c)] No file in $L$ contains more than $\kappa n$ conflicts with $A$. \item[(d)] No symbol-set in $L$ contains more than $\kappa n$ conflicts with $A$. \item[(e)] No transversal-set in $L$ contains more than $\kappa n$ conflicts with $A$. \item[(f)] Each cell of $L$ belongs to at least $\alpha n$ allowed 3-cubes. \end{itemize} \end{lemma} \begin{proof} Let $X_a$, $X_b$, $X_c$, $X_d$, $X_e$ and $X_f$ be the number of permutations which do not fulfill the conditions $(a)$, $(b)$, $(c)$, $(d)$, $(e)$ and $(f)$, respectively. Let $X$ be the number of permutations satisfying the five conditions $(a)$, $(b)$, $(c)$, $(d)$, $(e)$ and $(f)$. There are $(n!)^4$ ways to permute the row layers, the column layers, the file layers and the symbols, so we have $$X \geq (n!)^4 - X_a - X_b - X_c - X_d - X_e - X_f.$$ We shall prove that $X$ is greater than $0$. \begin{itemize} \item To estimate $X_a$, assume that for any fixed permutation $(\tau_1,\tau_3,\tau_4)$ of the row layers, the file layers and the symbols, at most $N_a$ choices of a permutation $\tau_2$ of the column layers yield a quadruple $(\tau_1, \tau_2, \tau_3,\tau_4)$ of permutations that break condition $(a)$; so $X_a \leq n!n!n!N_a$. Let $R$ be a fixed row chosen arbitrarily; we count the number of ways a permutation $\tau_2$ of the column layers can be constructed so that $(a)$ does not hold on row $R$. Let $S$ be a set of size $\kappa n$ of column layers of $A$. There are $n \choose \kappa n$ ways to choose $S$. In order to have a conflict at cell $(i,j,k)$ of $R$, the column layers should be permuted in such a way that in the resulting Latin cube $L$, $L(i,j,k) \in A(i,j,k)$. Since $|A(i,j,k)| \leq \gamma n$, there are at most $(\gamma n)^{\kappa n}$ ways to choose which column layers of $B$ are mapped by $\tau_2$ to column layers in $S$ so that all cells on row $R$ that are in $S$ are conflicts. The rest of the column layers can be arranged in any of the $(n-\kappa n)!$ possible ways. In total this gives at most $${n \choose \kappa n}(\gamma n)^{\kappa n}(n-\kappa n)! = \dfrac{n!(\gamma n)^{\kappa n}}{(\kappa n)!}$$ permutations $\tau_2$ that do not satisfy condition $(a)$ on row $R$. There are $n^2$ rows in $B$, so we have $$N_a \leq n^2 \dfrac{n!(\gamma n)^{\kappa n}}{(\kappa n)!}$$ and $$X_a \leq n!n!n!N_a \leq n^2(n!)^4\dfrac{(\gamma n)^{\kappa n}}{(\kappa n)!}.$$ An analogous argument gives the same bound for $X_b$, $X_c$, so in total, we have that $$X_a + X_b + X_c \leq 3n^2(n!)^4\dfrac{(\gamma n)^{\kappa n}}{(\kappa n)!}.$$ \item To estimate $X_d$, assume that for any fixed permutation $(\tau_1,\tau_3, \tau_4)$ of the row layers, the file layers and the symbols, at most $N_d$ choices of a permutation $\tau_2$ of the column layers give a quadruple $(\tau_1, \tau_2, \tau_3,\tau_4)$ of permutations that break condition $(d)$; then $X_d \leq n!n!n!N_d$. Let $\mathfrak{b}$ be a fixed symbol chosen arbitrarily; we count the number of ways a permutation $\tau_2$ of the column layers can be constructed so that $(d)$ does not hold for $\mathfrak{b}$ in a given row layer. Let $R_L$ be a fixed row layer; there are $n$ cells containing $\mathfrak{b}$ in $R_L$ and these cells belong to $n$ different column layers since $B$ is a boolean Latin cube. Let $S$ be a set of size $\kappa n$ of column layers of $A$; there are $n \choose \kappa n$ ways to choose $S$. Since in $A$, each symbol occurs at most $\gamma n$ time in each row, there are at most $(\gamma n)^{\kappa n}$ ways to choose which column layers of $B$ are mapped by $\tau_2$ to column layers in $S$ so that all cells containing $\mathfrak{b}$ on row layer $R_L$ that are in $S$ are conflicts. The rest of the column layers can be arranged in any of the $(n-\kappa n)!$ possible ways. In total this gives at most $${n \choose \kappa n}(\gamma n)^{\kappa n}(n-\kappa n)! = \dfrac{n!(\gamma n)^{\kappa n}}{(\kappa n)!}$$ permutations $\tau_2$ such that in the resulting Latin cube $L$, symbol $\mathfrak{b}$ appears in more than $\kappa n$ conflicts in the row layer $R_L$. There are $n$ different row layers, $n$ different column layers and $n$ different file layers in $B$, so we deduce that there are at most $3n\dfrac{n!(\gamma n)^{\kappa n}}{(\kappa n)!}$ permutations $\tau_2$ that do not satisfy condition $(d)$ on symbol $\mathfrak{b}$. There are $n$ symbols in $B$, so we have $$N_d \leq 3n^2 \dfrac{n!(\gamma n)^{\kappa n}}{(\kappa n)!}.$$ and $$X_d \leq 3n^2 (n!)^4\dfrac{(\gamma n)^{\kappa n}}{(\kappa n)!}.$$ \item To estimate $X_e$, assume that for any fixed permutation $(\tau_1,\tau_2,\tau_3)$ of the row layers, the column layers, the file layers, at most $N_e$ choices of a permutation $\tau_4$ of the symbols give a quadruple $(\tau_1, \tau_2, \tau_3,\tau_4)$ of permutations that break condition $(e)$; so $X_e \leq n!n!n!N_e$. Let $T$ be a fixed transversal-set chosen arbitrarily; we count the number of ways a permutation $\tau_4$ of the symbols can be constructed so that $(e)$ does not hold on the set $T$. Let $S$ be a set of size $\kappa n$ of cells of $T$; there are $n \choose \kappa n$ ways to choose $S$. In order to have a conflict at cell $(i,j,k)$ of $T$, the symbols should be permuted in such a way that in the resulting Latin cube $L$, $L(i,j,k) \in A(i,j,k)$. Since $|A(i,j,k)| \leq \gamma n$, there are at most $(\gamma n)^{\kappa n}$ ways to choose which symbols of $B$ are mapped by $\tau_4$ to cells in $S$ so that all cells in $S$ are conflicts. The rest of the symbols can be arranged in any of the $(n-\kappa n)!$ possible ways. In total this gives at most $${n \choose \kappa n}(\gamma n)^{\kappa n}(n-\kappa n)! = \dfrac{n!(\gamma n)^{\kappa n}}{(\kappa n)!}$$ permutations $\tau_4$ that do not satisfy condition $(e)$ on the transversal-set $T$. There are $n^2$ transversal-sets in $B$, so we have $$N_e \leq n^2 \dfrac{n!(\gamma n)^{\kappa n}}{(\kappa n)!},$$ and so $$X_e \leq n!n!n!N_e \leq n^2(n!)^4\dfrac{(\gamma n)^{\kappa n}}{(\kappa n)!}.$$ \item To estimate $X_f$, assume that for any fixed permutation $(\tau_2,\tau_4)$ of the column layers and the symbols at most $N_f$ choices of a pair $(\tau_1, \tau_3)$ of the row layers and the file layers yield a quadruple $(\tau_1, \tau_2, \tau_3, \tau_4)$ of permutations that break condition $(f)$, then $X_f \leq n!n!N_f$. Let $(i_1,j_1,k_1)$ be an arbitrary fixed cell of $A$. There are $n^2$ ways to choose a row layer $i_x$ and a file layer $k_x$ so that $i_1=\tau_1(i_x)$ and $k_1=\tau_3(k_x)$; we fix such a row layer $i_x$ and file layer $k_x$. Moreover, each $3$-cube $\mathcal{C}$ containing $(i_1,j_1,k_1)$ is uniquely determined by the value of $j_2 \neq j_1$ where $(i_1,j_2,k_1) \in \mathcal{C}$; so a pair of permutations $(\tau_1, \tau_3)$ satisfy that the quadruple $(\tau_1, \tau_2, \tau_3, \tau_4)$ adds to $X_f$ if and only if there are more than $(1-\alpha)n$ choices for $j_2$ so that the swap along $\mathcal{C}$ is not allowed. We shall count the number of ways of choosing $(\tau_1, \tau_3)$ so that this holds. Let us first note that there are at most $2 \gamma n$ choices for $(i_x, k_x)$ that yield a $3$-cube $\mathcal{C}$ that is not allowed because of a conflict on row $i_1$ in file layer $k_1$; that is, after swapping on $\mathcal{C}$, we have a conflict cell on row $i_1$ in file layer $k_1$. This follows from the fact that there are $\gamma n$ choices for $j_2$ such that $A(i_1,j_2,k_1)$ contains $L(i_1,j_1,k_1)$, and since $|A(i_1,j_1,k_1)| \leq \gamma n$, we have $\gamma n$ choices for $j_2$ so that $L(i_1, j_2, k_1) \in A(i_1,j_1,k_1)$. So for a permutation $(\tau_1,\tau_3)$ to contribute to $N_f$, $(\tau_1,\tau_3)$ must be such that at least $(1-\alpha -2\gamma)n$ $3$-cubes containing the cell $(i_1,j_1,k_1)$ are not allowed because of restrictions on rows of $A$ that are distinct from row $i_1$ in file layer $k_1$. Since each $3$-cube $\mathcal{C}$ containing $(i_1,j_1,k_1)$ has cells from three other rows, this implies that at least $(1-\alpha-2\gamma)n / 3$ $3$-cubes cannot be allowed because of conflicts appearing in one of these rows. Let $N_{f_1}$ be the number of pairs of permutations $(\tau_1,\tau_3)$ such that at least $(1-\alpha-2\gamma)n / 3$ $3$-cubes containing $(i_1,j_1,k_1)$ cannot be allowed because swapping yields conflicts in cells in file layer $k_1$ that are not contained in row layer $i_1$. Let us first note that there are $(n-1)!$ ways to permute the remaining file layers of $B$. Consider a fixed permutation $\tau_3$ of the file layers; we count the number of permutations $\tau_1$ of the row layers such that the pair $(\tau_1,\tau_3)$ contributes to $N_{f1}$. Let $S$ be a set of columns, ($|S|=(1-\alpha-2 \gamma)n/3$), such that for every column $C_{j_2,k_2} \in S$, there is a unique $i_2$ satisfying that $$\mathcal{C}=\{(i_1,j_1,k_1), (i_1,j_2,k_1), (i_2,j_1,k_1), (i_2,j_2,k_1), (i_1,j_1,k_2), (i_1,j_2,k_2), (i_2,j_1,k_2), (i_2,j_2,k_2)\}$$ is a $3$-cube and this $3$-cube is not allowed because of conflicts arising in row $i_2$ in file layer $k_1$. There are $n -1 \choose (1-\alpha-2\gamma)n / 3$ ways to choose $S$. Fix a column $C_{j_2,k_2} \in S$; in column $j_1$ of file layer $k_1$ of $A$, there are at most $\gamma n$ cells containing $L(i_1,j_1,k_1)$ and in the column $j_2$ in file layer $k_1$ of $A$, there are at most $\gamma n$ cells containing $L(i_1,j_2,k_1)$, so there are up to $2\gamma n$ choices for $\tau^{-1}_1(i_2)$ in $B$ that would make $\mathcal{C}$ disallowed because of conflicts arising in rows distinct from $i_1$ in the file layer $k_1$. Since every column in $S$ yields a unique row index, $S$ determines $\tau_1$ on $(1-\alpha-2 \gamma)n/3$ row layers. The remaining row layers can be permuted in $(n-1-(1-\alpha-2\gamma)n / 3)!$ ways. This implies that the total number of permutations $\tau_1$ that yield at least $(1-\alpha-2\gamma)n / 3$ $3$-cubes that are not allowed because of conflicts appearing in file layer $k_1$ that are not contained in row layer $i_1$ is bounded from above by $${{n-1} \choose {(1 - \alpha - 2\gamma)n/3}} {(2\gamma n)^{(1-\alpha-2\gamma)n/3}} {(n-1-(1-\alpha-2\gamma)n / 3)!} = \dfrac{(n-1)! {(2\gamma n)^{(1-\alpha-2\gamma)n/3}}}{((1-\alpha-2\gamma)n/3)!}$$ Hence, $N_{f_1} \leq (n-1)! \dfrac{(n-1)! {(2\gamma n)^{(1-\alpha-2\gamma)n/3}}}{((1-\alpha-2\gamma)n/3)!}$. Let $N_{f_2}$ be the number of pairs of permutations $(\tau_1,\tau_3)$ such that at least $(1-\alpha-2\gamma)n / 3$ $3$-cubes containing $(i_1,j_1,k_1)$ are not allowed because swapping on them yields conflicts in rows contained in the row layer $i_1$ but not in file layer $k_1$. There are $(n-1)!$ ways to permute the remaining row layers of $B$. We consider a fixed permutation $\tau_1$ of the row layers and count the number of permutations $\tau_3$ of the file layers such that the pair $(\tau_1,\tau_3)$ contributes to $N_{f_2}$. Let $S$ be a set of files, ($|S|=(1-\alpha-2 \gamma)n/3$), such that for every file $F_{i_2,j_2} \in S$, there is a unique $k_2$ satisfying that $$\mathcal{C}=\{(i_1,j_1,k_1), (i_1,j_2,k_1), (i_2,j_1,k_1), (i_2,j_2,k_1), (i_1,j_1,k_2), (i_1,j_2,k_2), (i_2,j_1,k_2), (i_2,j_2,k_2)\}$$ is a $3$-cube and this $3$-cube is not allowed because of conflicts arising in cells in row layer $i_1$ that are not in file layer $k_1$. There are $n -1 \choose (1-\alpha-2\gamma)n / 3$ ways to choose $S$. Fix a file $F_{i_2,j_2} \in S$, in the file $F_{i_1,j_1}$ of $A$, there are at most $\gamma n$ cells containing $L(i_1,j_1,k_1)$ and in the file $F_{i_1,j_2}$ of $A$, there are at most $\gamma n$ cells containing $L(i_1,j_2,k_1)$, so there are up to $2\gamma n$ choices for $\tau^{-1}_3(k_2)$ in $B$ that would make $\mathcal{C}$ disallowed because of possible conflicts in row layer $i_1$ that are not in file layer $k_1$. As before, $S$ determines how $\tau_3$ acts on $(1-\alpha-2 \gamma)n/3)$ file layers, and the remaining file layers can be permuted in $(n-1-(1-\alpha-2\gamma)n / 3)!$ ways. This implies that the total number of permutations $\tau_3$ with not enough allowed $3$-cubes due to the fact that swapping yield conflicts in rows contained in the row layer $i_1$ but not in file layer $k_1$ is bounded from above by $${{n-1} \choose {(1 - \alpha - 2\gamma)n/3}} {(2\gamma n)^{(1-\alpha-2\gamma)n/3}} {(n-1-(1-\alpha-2\gamma)n / 3)!} = \dfrac{(n-1)! {(2\gamma n)^{(1-\alpha-2\gamma)n/3}}}{((1-\alpha-2\gamma)n/3)!}$$ Hence, $N_{f_2} \leq (n-1)! \dfrac{(n-1)! {(2\gamma n)^{(1-\alpha-2\gamma)n/3}}}{((1-\alpha-2\gamma)n/3)!}$. Let $N_{f_3}$ be the number of pairs of permutations $(\tau_1,\tau_3)$ such that at least $(1-\alpha-2\gamma)n / 3$ $3$-cubes $\mathcal{C}$ containing $(i_1,j_1,k_1)$ are not allowed because swapping on them yields conflicts in cells which lie in row and file layers distinct from $i_1$ and $k_1$, respectively. There are $(n-1)!$ ways to permute the remaining file layers of $B$. Consider a fixed permutation $\tau_3$ of the file layers; we count the number of permutations $\tau_1$ of the row layers such that the pair $(\tau_1,\tau_3)$ contributes to $N_{f_3}$. Let $S$ be a set of columns ($|S|=(1-\alpha-2 \gamma)n/3$), such that for every column $C_{j_2,k_2} \in S$, there is a unique $i_2$ satisfying that $$\mathcal{C}=\{(i_1,j_1,k_1), (i_1,j_2,k_1), (i_2,j_1,k_1), (i_2,j_2,k_1), (i_1,j_1,k_2), (i_1,j_2,k_2), (i_2,j_1,k_2), (i_2,j_2,k_2)\}$$ is a $3$-cube which is not allowed because of swapping yields conflicts in cells in row $i_2$ in file layer $k_2$. There are $n -1 \choose (1-\alpha-2\gamma)n / 3$ ways to choose $S$. Fix a column $C_{j_2,k_2} \in S$; in the column $C_{j_2,k_2}$ of $A$, there are at most $\gamma n$ cells containing symbol $L(i_1,j_1,k_1)$, and in the column $C_{j_1,k_2}$ of $A$, there are at most $\gamma n$ cells containing $L(i_1,j_2,k_1)$; so there are up to $2\gamma n$ choices for $\tau^{-1}_1(i_2)$ in $B$ that would make $\mathcal{C}$ disallowed because swapping yields conflicts in cells which lie in row and file layers distinct from $i_1$ and $k_1$, respectively. The set $S$ determines how $\tau_1$ acts on $(1-\alpha-2 \gamma)n/3$ row layers. The remaining row layers can be permuted in $(n-1-(1-\alpha-2\gamma)n / 3)!$ ways. This implies that the total number of permutations $\tau_1$ with too few allowed $3$-cubes because of conflicts arising in cells in row and file layers distinct from $i_1$ and $k_1$ is bounded from above by $${{n-1} \choose {(1 - \alpha - 2\gamma)n/3}} {(2\gamma n)^{(1-\alpha-2\gamma)n/3}} {(n-1-(1-\alpha-2\gamma)n / 3)!} = \dfrac{(n-1)! {(2\gamma n)^{(1-\alpha-2\gamma)n/3}}}{((1-\alpha-2\gamma)n/3)!}$$ Hence, $N_{f_3} \leq (n-1)! \dfrac{(n-1)! {(2\gamma n)^{(1-\alpha-2\gamma)n/3}}}{((1-\alpha-2\gamma)n/3)!}$. The Boolean Latin cube contains $n^3$ cells in total, so $$N_f \leq n^3(n^2 N_{f_1} + n^2 N_{f_2} + n^2 N_{f_3}) \leq 3n^3 (n!)^2 \dfrac{{(2\gamma n)^{(1-\alpha-2\gamma)n/3}}}{((1-\alpha-2\gamma)n/3)!}$$ and $$X_f \leq (n!)^2N_f \leq 3n^3 (n!)^4 \dfrac{{(2\gamma n)^{(1-\alpha-2\gamma)n/3}}}{((1-\alpha-2\gamma)n/3)!}$$ \end{itemize} Summing up, we conclude that $$X \geq (n!)^4 - 7n^2 (n!)^4\dfrac{(\gamma n)^{\kappa n}}{(\kappa n)!} - 3n^3 (n!)^4 \dfrac{{(2\gamma n)^{(1-\alpha-2\gamma)n/3}}}{((1-\alpha-2\gamma)n/3)!}$$ $$\geq (n!)^4 \Big(1 - 7n^2 \dfrac{(\gamma n)^{\kappa n}}{(\kappa n)!} - 3n^3\dfrac{{(2\gamma n)^{(1-\alpha-2\gamma)n/3}}}{((1-\alpha-2\gamma)n/3)!}\Big)$$ By assumption, $X$ is strictly greater than $0$. \end{proof} \begin{lemma} Let $L$ be a Latin cube that is isomorphic to a Boolean Latin cube, and let $A$ be an $(m,m,m,m)$-cube; both of order $n$. Furthermore, let $\alpha, \gamma, \kappa, \theta, \epsilon$ be constants, $n=2^t$ such that $\epsilon n \geq 3$ and $$\alpha n-21\kappa n -7\epsilon n-\dfrac{84\kappa}{\epsilon}n-\dfrac{21\theta}{\epsilon} n-\dfrac{80\kappa}{\theta} n-28>0$$ If $L$ has the following properties: \begin{itemize} \item[(a)] no row in $L$ contains more than $\kappa n$ conflicts with $A$; \item[(b)] no column in $L$ contains more than $\kappa n$ conflicts with $A$; \item[(c)] no file in $L$ contains more than $\kappa n$ conflicts with $A$; \item[(d)] no symbol-set in $L$ contains more than $\kappa n$ conflicts with $A$; \item[(e)] no transversal-set in $L$ contains more than $\kappa n$ conflicts with $A$; \item[(f)] each cell of $L$ belongs to at least $\alpha n$ allowed 3-cubes; \end{itemize} then there is a set of disjoint allowed $3$-cubes such that each conflict of $L$ belongs to one of them. Thus, by performing a number of swaps on $3$-cubes in $L$, we obtain a Latin cube $L'$ that avoids $A$. \end{lemma} \begin{proof} For constructing $L'$ from $L$, we will perform a number of swaps on $3$-cubes, and we shall refer to this procedure as \textit{$S$-swap}. We are going to construct a set $S$ of disjoint allowed $3$-cubes such that each conflict of $L$ with $A$ belongs to one of them. A cell that belongs to a $3$-cube in $S$ is called \textit{used} in $S$-swap. Since no row in $L$ contains more than $\kappa n$ conflicts with $A$, there are at most $\kappa n^3$ conflicts in $L$, which implies that the total number of cells that are used in $S$-swap is at most $8 \kappa n^3$. A row layer, a column layer, a file layer, a row block, a column block, a file block, a symbol block, a symbol-row block, a symbol-column block, or a symbol-file block is \textit{overloaded} if such a layer or block contains at least $\theta n^2$ cells that are used in $S$-swap; note that no more than $\dfrac{8\kappa n^3}{\theta n^2}=\dfrac{8\kappa}{\theta} n$ layers or blocks of each type are $S$-overloaded. A row, a column, a file, a transversal-set, or a symbol-set is \textit{overloaded} if this row, column, file, transversal-set or symbol-set contains at least $\epsilon n$ cells that are used in $S$-swap. Using these facts, let us now construct our set $S$ by steps; at each step we consider a conflict cell $(i_1,j_1,k_1)$ and include an allowed $3$-cube containing $(i_1,j_1,k_1)$ in $S$. Initially, the set $S$ is empty. So let us consider a each conflict cell $(i_1,j_1,k_1)$ in $L$; there are at least $\alpha n$ allowed $3$-cubes containing $(i_1,j_1,k_1)$. We choose an allowed $3$-cube $$\mathcal{C}=\{(i_1,j_1,k_1), (i_1,j_2,k_1), (i_2,j_1,k_1), (i_2,j_2,k_1), (i_1,j_1,k_2), (i_1,j_2,k_2), (i_2,j_1,k_2), (i_2,j_2,k_2)\}$$ that satisfies the following: \begin{itemize} \item[(1)] The row layer $i_2$, the column layer $j_2$, the file layer $k_2$, the row block containing the row $R_{i_2,k_1}$, the column block containing the column $C_{j_2,k_1}$, the file block containing the file $F_{i_1,j_2}$, the symbol-row block containing two cells $(i_1,j_2,k_1)$ and $(i_1,j_1,k_2)$, the symbol-column block containing two cells $(i_2,j_1,k_1)$ and $(i_1,j_1,k_2)$, the symbol-file block containing two cells $(i_1,j_2,k_1)$ and $(i_2,j_1,k_1)$, the symbol block containing symbol $L(i_1,j_2,k_1)$ are not overloaded. This eliminates at most $\dfrac{10 \times 8\kappa}{\theta} n=\dfrac{80\kappa}{\theta} n$ choices. With this strategy for including $3$-cubes in $S$, after completing the construction of $S$, every layer (or block) contains at most $4\kappa n^2 + (\theta n^2-1) +4$ cells that are used in $S$-swap. Hence, the number of overloaded rows (overloaded columns, overloaded files, overloaded transversal-sets or overloaded symbol-sets) in each layer (or block) is at most $\dfrac{4\kappa n^2 + \theta n^2 +3}{\epsilon n} \leq \dfrac{4\kappa + \theta}{\epsilon} n+1$. Note that here the statement ``each symbol block contains at most $\dfrac{4\kappa + \theta}{\epsilon} n+1$ overloaded symbol-sets'' is to be taken with respect to either row layers, column layers or file layers, i.e., when we consider the $n$ different symbol sets of a given symbol block belonging to $n$ different row layers (or $n$ different column layers or $n$ different file layers), the number of overloaded such symbol-sets is at most $\dfrac{4\kappa + \theta}{\epsilon} n+1$. \item[(2)] Some rows, columns, files, transversal-sets, symbol-sets are not overloaded as the following: \begin{itemize} \item[(2a)] The columns $C_{j_2,k_1}$, $C_{j_1,k_2}$, $C_{j_2,k_2}$ are not overloaded; this eliminates at most $\dfrac{12\kappa + 3\theta}{\epsilon} n+3$ choices since in the file layer $k_1$ (which contains the column $C_{j_2,k_1}$) and in the column layer $j_1$ (which contains the column $C_{j_1,k_2}$) and in the column block which contains the column $C_{j_1,k_1}$ (which also contains the column $C_{j_2,k_2}$), there are in total at most $\dfrac{4\kappa + \theta}{\epsilon} n+1$ overloaded columns. Similarly, we need that the rows $R_{i_2,k_1}$, $R_{i_1,k_2}$, $R_{i_2,k_2}$ and the files $F_{i_1,j_2}$, $F_{i_2,j_1}$, $F_{i_2,j_2}$ are not overloaded; this eliminates at most $\dfrac{24\kappa + 6\theta}{\epsilon} n+6$ choices. \item[(2b)] The transversal-set $\mathfrak{t}_1$ containing $(i_2,j_1,k_1)$ and $(i_1,j_2,k_2)$ is not overloaded; this eliminates at most $\dfrac{4\kappa + \theta}{\epsilon} n+1$ choices, since in the column block which contains the column $C_{j_1,k_1}$ (which also contains the transversal-set $\mathfrak{t}_1$), there are at most $\dfrac{4\kappa + \theta}{\epsilon} n+1$ overloaded transversal-sets. Similarly, we need that the transversal-set containing $(i_1,j_2,k_1)$ and $(i_2,j_1,k_2)$, and the transversal-set containing $(i_2,j_2,k_1)$ and $(i_1,j_1,k_2)$ are not overloaded; this eliminates at most $\dfrac{8\kappa + 2\theta}{\epsilon} n+2$ choices. \item[(2c)] The symbol-set $\mathfrak{s}_1$ containing $(i_2,j_1,k_2)$ and $(i_1,j_2,k_2)$ is not overloaded; this eliminates at most $\dfrac{4\kappa + \theta}{\epsilon} n+1$ choices, since in the symbol block which contains $(i_1,j_1,k_1)$ (which also contains the symbol-set $\mathfrak{s}_1$), there are at most $\dfrac{4\kappa + \theta}{\epsilon} n+1$ overloaded symbol-sets. Similarly, we need that the symbol-set containing $(i_2,j_2,k_1)$ and $(i_1,j_2,k_2)$, and the symbol-set containing $(i_2,j_2,k_1)$ and $(i_2,j_1,k_2)$ are not overloaded, this eliminates at most $\dfrac{8\kappa + 2\theta}{\epsilon} n+2$ choices. \item[(2d)] The symbol-set $\mathfrak{s}_2$ containing $(i_1,j_2,k_1)$ and $(i_2,j_1,k_1)$, and the symbol-set $\mathfrak{s}_3$ containing $(i_1,j_2,k_1)$ and $(i_2,j_2,k_2)$ are not overloaded. This eliminates at most $\dfrac{8\kappa + 2\theta}{\epsilon} n+2$ choices, since in the file layer $k_1$ (which contains the symbol-set $\mathfrak{s}_2$), and in the symbol-column block which contains $(i_1,j_1,k_1)$ (which also contains symbol-set $\mathfrak{s}_3$), there are at most $\dfrac{4\kappa + \theta}{\epsilon} n+1$ overloaded symbol-sets. Similarly, we need that the symbol-set containing $(i_1,j_1,k_2)$ and $(i_1,j_2,k_1)$, the symbol-set containing $(i_1,j_1,k_2)$ and $(i_2,j_2,k_2)$, the symbol-set containing $(i_2,j_1,k_1)$ and $(i_1,j_1,k_2)$, the symbol-set containing $(i_2,j_1,k_1)$ and $(i_2,j_2,k_2)$ are not overloaded. This eliminates at most $\dfrac{16\kappa + 4\theta}{\epsilon} n+4$ choices. \end{itemize} So in total, this eliminates at most $\dfrac{84\kappa + 21\theta}{\epsilon} n+21$ choices. Note that with this strategy for including $3$-cubes in $S$, after completing the construction of $S$, every row, column, file, transversal-set, and symbol-set contains at most $2 \kappa n +(\epsilon n-1)+2$ or $2 \kappa n +\epsilon n+1$ cells that are used in $S$-swap. \item[(3)] Except for $(i_1,j_1,k_1)$, none of the cells in $\mathcal{C}$ are conflicts or used before in $S$-swap. \begin{itemize} \item[(3a)] The cell $(i_2,j_1,k_1)$ is not conflict and has not been used before in $S$-swap; this eliminates at most $3\kappa n +\epsilon n+1$ choices since the column $C_{j_1,k_1}$ contains at most $\kappa n$ conflict cells and at most $2 \kappa n +\epsilon n+1$ cells that are used in $S$-swap. Similarly, we need that the cell $(i_1,j_2,k_1)$ and the cell $(i_1,j_1,k_2)$ are not conflicts and has not used before in $S$-swap; in total, this eliminates at most $6\kappa n +2\epsilon n+2$ choices. \item[(3b)] The cell $(i_1,j_2,k_2)$ is not conflict and has not been used before in $S$-swap. This eliminates at most $3\kappa n +\epsilon n+1$ choices, since in the symbol-set in row layer $i_1$ that contains the cell $(i_1,j_1,k_1)$, there are at most $\kappa n$ conflict cells and at most $2 \kappa n +\epsilon n+1$ cells that have been used in $S$-swap. Similarly, we need that the cell $(i_2,j_1,k_2)$ and the cell $(i_2,j_2,k_1)$ are not conflicts and has not been used before in $S$-swap; in total, this eliminates at most $6\kappa n +2\epsilon n+2$ choices. \item[(3c)] The cell $(i_2,j_2,k_2)$ is not conflict and has not been used before in $S$-swap. This eliminates at most $3\kappa n +\epsilon n+1$ choices since in the transversal-set containing the cell $(i_1,j_1,k_1)$, there are at most $\kappa n$ conflict cells and at most $2 \kappa n +\epsilon n+1$ cells that are used in $S$-swap. \end{itemize} So in total, this eliminates at most $21\kappa n +7\epsilon n+7$ choices. \end{itemize} It follows that we have at least $$\alpha n-21\kappa n -7\epsilon n-\dfrac{84\kappa}{\epsilon}n-\dfrac{21\theta}{\epsilon} n-\dfrac{80\kappa}{\theta} n-28$$ choices for an allowed $3$-cube $\mathcal{C}$ which contains $(i_1,j_1,k_1)$. By assumption, this expression is greater than zero, so we can conclude that there is a $3$-cube satisfying these conditions. Thus we may construct the set $S$ by iteratively adding disjoint allowed $3$-cubes such that each $3$-cube contains a conflict cell. After this process terminates, we have a set $S$ of disjoint $3$-cubes; we swap on all $3$-cubes in $S$ to obtain the Latin cube $L'$. Hence, we conclude that we can obtain a Latin cube $L'$ that avoids $A$. \end{proof}

f93be7af8ef04b1f2da9d37f8795e6ee87312b9c

\section{Introduction}\label{intro}} As data science has become a more solid field, theories and principles have developed to describe best practices. One such idea is `tidy data,' which defines a clean, analysis-ready format that informs workflows converting raw data through a data analysis pipeline \citep{Wickham2014}. Another idea is the grammar of graphics, describing how to map data values into a visualization \citep{Wilkinson2012}. These principles have been widely adopted, but do not address the problem of missing data. In particular, there is little guidance on how to handle missing values in a data analysis workflow. Most analysis and visualization software simply drop missing values when making a plot, although some (\pkg{ggplot2}) provide a warning (Figure \ref{fig:warning}). \begin{CodeChunk} \begin{figure} {\centering \includegraphics[width=0.75\linewidth]{tidy-missing-data-paper_files/figure-latex/warning-1} } \caption[How ggplot2 behaves when displaying missing values]{How ggplot2 behaves when displaying missing values. A warning message is displayed, but missing values are not shown in the plot}\label{fig:warning} \end{figure} \end{CodeChunk} The imputation literature focuses on ensuring valid statistical inference is made from incomplete data. This is approached chiefly through probabilistic modelling, assuming the mechanism of missing data is known to the analyst. It does not address how to explore, understand, and handle missing data structures and mechanisms. However, something must be known about the missing value structure to produce a complete dataset for analysis, whether by case- or variable-deletion, or with imputation. Decision tree models or latent group analysis can reveal structures or patterns of missingness \citep{Tierney2015, Barnett2017}, but are not definitive. While there are times where the missing data mechanism is obvious, determining this is typically not straightforward. To understand their structure and possible mechanisms, the analyst must explore the data with visualizations, summaries, and modelling, in an iterative fashion. While there have been many software tools for exploring missing data, they do not work together. The graphics literature provides several solutions for exploring missings visually by incorporating them into the plot in some way. For example, imputing values to be 10\% below the minimum to include all observations into a scatter plot \citep{Cook2007}, or treating missing values as an equal category of data \citep{Unwin1996}. The ideas from the graphics literature need to be translated into tidy data tools to integrate missing data handling in a data analysis pipeline. This paper is organized in the following way. The next section (\ref{background}) provides the background to tidy data principles and tools (\ref{tidy-data-concepts}) and missing data representations (\ref{missing-data-rep-dep}). Section \ref{existing-software} summarises existing software for handling missing data. The new work extending the tidy tools to facilitate exploring, visualizing, and imputing missing data is discussed in Section \ref{extensions}. Graphics (\ref{graphics}) and Numerical summaries (\ref{num-sum}) of missing values are then discussed. An application illustrating the use of the new methods is shown in Section \ref{case-study}. Section \ref{discussion} discusses strengths, limitations, and future directions. \hypertarget{background}{% \section{Background}\label{background}} \hypertarget{tidy-data-concepts}{% \subsection{Tidy data concepts and methods}\label{tidy-data-concepts}} Features of tidy data were formally described in \citet{Wickham2014}, and were discussed in terms of their importance for data science by \citet{Donoho2017}, and tools for data analysis. Tidy data is defined by \citet{r4ds} as: \begin{quote} \begin{enumerate} \def\arabic{enumi}.{\arabic{enumi}.} \tightlist \item Each variable must have its own column. \item Each observation must have its own row. \item Each value must have its own cell. \end{enumerate} \end{quote} Tidy data is easier to work with and analyze because the variables are in the same format as they would be put into modelling software. This helps the analyst works swiftly and clearly, closing up opportunities for errors. Tidy data principles are general, but are comprehensively implemented in the R programming language, so this paper focuses on R \citep{rcore}. Tidy tools require the same tidy data input and output. This consistency means multiple tools can be composed together into a sequence, allowing for rapid, elegant, and complex operations. Contrasting tidy tools are messy tools. These have tidy input but messy output. Messy tools slow down analysis by shifting the focus from analysis to transforming output so it is the right shape for the next step in the analysis. This makes the work at each step harder to predict, and more complex and difficult to maintain. This disrupts workflow, and invites errors. Tidy tools fall into three broad categories: data manipulation, visualization, and modelling. \hypertarget{tidy-data-manip}{% \subsubsection{Data manipulation}\label{tidy-data-manip}} Data manipulation is made input- and output-tidy with R packages \pkg{dplyr} and \pkg{tidyr} \citep{dplyr, tidyr}. These provide the five ``verbs'' of data manipulation: data reshaping, sorting, filtering, transforming, and aggregating. Data reshaping goes from long to wide formats; sorting arranges rows in a specific order; filtering removes rows based on a condition; transforming, changes existing variables or adds new ones; aggregating creates a single value from many values, say, for example, in computing the minimum, maximum, and mean. \hypertarget{tidy-vis}{% \subsubsection{Visualizations}\label{tidy-vis}} Visualization tools only have tidy data as their input, as the output is a graphic. The popular domain specific language \pkg{ggplot2} maps variables in a dataset to features (referred to as aesthetics) of a graphic \citep{ggplot2}. For example, a scatterplot can be created by mapping two variables to the x and y axes, and specifying a point geometry. \hypertarget{tidy-model}{% \subsubsection{Modelling}\label{tidy-model}} Modelling tools work well with tidy data, as they have a clear mapping from variables in the data to the formula for a model. For example in R, y regressed on x and z is: \code{lm(y ~ x + z)}. Modelling tools are input tidy, but their output is always messy - it is not in the right format for subsequent steps in analysis. For example, estimated coefficients, predictions, and residuals from one model cannot be easily combined with the output of another model. Messy models have been partially addressed with the \pkg{broom} package, which tidies up model outputs into a tidy data format for data analysis, and the developing \pkg{recipes} package, which helps make modelling input- and output-tidy \citep{recipes}. \hypertarget{the-tidyverse}{% \subsubsection{The tidyverse}\label{the-tidyverse}} Defining tidy data and tidy tools has resulted in a growing set of packages known collectively as the ``tidyverse'' \citep{tidyverse}. These are constructed to share similar principles in their design and behavior, and cover the breadth of an analysis - from importing, tidying, transforming, visualizing, modelling, to communicating \citep{r4ds, tidyverse, Tidyverse-Manifesto}. This has led to more tools for specific parts of analysis - from reading in data with \pkg{readr}, \pkg{readxl}, and \pkg{haven}, to handling character strings with \pkg{stringr}, dates with \pkg{lubridate}, and performing functional programming with \pkg{purrr} \citep{readr, readxl, haven, stringr, lubridate, purrr}. It has also led to a burgeoning of new packages for other fields following similar design principles, creating fluid workflows for new domains. For example, the \pkg{tidytext} \citep{tidytext} package for text analysis, the \pkg{tsibble} \citep{wang2020tsibble} package for time series data, and \pkg{tidycensus} \citep{tidycensus} for working with US census and boundary data. \hypertarget{tidy-formats-missing-data}{% \subsubsection{Tidy formats for missing data}\label{tidy-formats-missing-data}} Current tools for missing data are messy. Missing data tools can be used to perform imputations, missing data diagnostics, and data visualizations. However, these tools suffer the same problems as modelling: They use tidy input, but produce messy output - their output is challenging to integrate with other steps of data analysis. The complex, often multivariate nature of imputation methods also makes makes them difficult to represent. Visualization methods for missing data do not map data features to the aesthetics of a graphic, as in \pkg{ggplot2}, limiting expressive exploration. Taking existing methods from the missing data graphics literature, and translating and expanding them into tidy data and tidy tools would create more effective data visualizations. Defining these concepts allows the focus to be more general than just software, but rather, an extensible framework for tidy tools to explore missing data. \hypertarget{missing-data-rep-dep}{% \subsection{Missing data representation and dependence}\label{missing-data-rep-dep}} The convention for representing missingness is a \textbf{b}inary matrix, \(B\), for data \(y\) with \(i\) rows and \(j\) columns: \[ b_{ij} =\begin{cases} 1, & \text{if } y_{ij} \text{ is missing} \\ 0, & \text{if } y_{ij} \text{ is observed} \end{cases} \] There are many ways each value can be missing, we adopt the notation used in \citet{VanBuuren2012}. The information in \(B\) can be used to arrive at three categories of missing values: Missing completely at random (MCAR), missing at random (MAR), and missing not at random (MNAR). The distribution of missing values in \(b_{ij}\) can depend on the entire dataset, represented as \(Y = (Y_{obs}, Y_{miss})\). This relationship can be defined by the \emph{missing data model} \(Pr(b_{ij} | Y_{obs}, Y_{miss}, \psi)\), the probability of missingness is conditional on data observed, data missing, and some probability parameter of missingness, \(\psi\). This helps to precisely define categories of missing values. \textbf{MCAR} is where values being missing have no association with observed or unobserved data, that is, \(Pr(B = 1 | Y_{obs}, Y_{miss}) = Pr(B = 1 | \psi)\). Essentially, the probability of an observation being missing is unrelated to anything else, only the parameter \(\psi\), the overall probability of missingness. Although a convenient scenario, it is not actually possible to confirm, or clearly distinguish from MAR, as it relies on statements on data unobserved. In \textbf{MAR}, missingness only depends on data observed, not data missing, that is, \(Pr(B = 1 | Y_{obs}, Y_{miss}, \psi) = Pr(B = 1 | Y_{obs}, \psi)\). Some structure or dependence between missing and observed values is allowed, provided it can be explained by data observed, and some overall probability of missingness. In \textbf{MNAR}, missingness is related to values observed, and unobserved: \(Pr(B = 1 | Y_{obs}, Y_{miss}, \psi)\). This assumes conditioning on all observations: data goes missing due to some phenomena unobserved, including the structure of the missing data itself. This presents a challenge in analysis, as it is difficult to verify, and implies bias in analysis due to the unobserved phenomena. Visualizations can help assess whether data may be MCAR, MAR or MNAR. \hypertarget{existing-software}{% \section{Existing Software}\label{existing-software}} Methods for exploring, understanding, and imputing missing data are more accessible now than they have ever been. Values can be imputed with one value (single imputation), or multiple values (multiple imputation), creating \(m\) datasets. This section discusses existing software for single and multiple imputation, and missing data exploration. \hypertarget{imputation}{% \subsection{Imputation}\label{imputation}} \pkg{VIM} \citep{VIM} implements well-used imputation methods K nearest neighbors, regression, hot-deck, and iterative robust model-based imputation. These diverse approaches allows for imputing with semi-continuous, continuous, count, and categorical data. VIM identifies imputed cases by adding an indicator variable with a suffix \texttt{\_imp}. So \texttt{Var1} has a sibling column, \texttt{Var1\_imp}, with values TRUE or FALSE indicating imputation. \pkg{VIM} also has a variety of visualization methods, discussed in \ref{exploration}. \pkg{simputation} provides an interface to imputation methods from VIM, in addition providing hotdeck imputation, and the EM algorithm \citep{simputation, Dempster1977}. \pkg{simputation} provides a consistent formula interface for all imputation methods, and always returns a dataframe with the updated imputations. \pkg{Hmisc} \citep{Hmisc}, provides predictive mean matching, \pkg{imputeTS} \citep{imputeTS}, provides time series imputation methods, and \pkg{missMDA} \citep{missMDA}, imputes data using principal components analysis. Multiple imputation is often regarded as best practice for imputing values \citep{Schafer2002}, as long as appropriate caution is taken \citep{Sterne2009}. Popular and robust methods for multiple imputation include the \pkg{mice}, \pkg{Amelia}, and \pkg{mi} packages \citep{mice, amelia, mi}. \pkg{mice} implements the method of chained equations, using a variable-wise algorithm to calculate the posterior distribution of parameters to generate imputed values. The workflow in \pkg{mice} revolves around imputing data, returning completed data, and fitting a model and pooling the results. \pkg{Amelia} \citep{amelia} assumes data are multivariate normal, and samples from the posterior, and allows for incorporation of information on the values in a prior. It uses the computationally efficient (and parallelizable) Expectation-Maximization Bootstrap (EMB) algorithm \citep{Honaker2010}. \pkg{norm} \citep{norm}, provides multiple imputation using EM for multivariate normal data, drawing from methods in the NORM software \citep{schafer-norm}. \pkg{norm} does not provide a framework for tracking missing values, instead providing tools for making inference from multiple imputation. \pkg{mi} \citep{mi} also uses Bayesian models for imputation, providing better handling of semi-continuous values, and data with structural or perfect correlation. A collection of analysis models are also provided in \pkg{mi}, to work with data it has imputed. These include linear models, generalized linear models, and their corresponding Bayesian components. This approach promotes fluid workflow, with a similar penalty to tidying up model output, which is still messy. \hypertarget{imputation-summary}{% \subsubsection{Summary}\label{imputation-summary}} Each imputation method provides practical methods for different use cases, but most have different output structures, and do not have consistent interfaces in their implementation. This makes them inherently messy and challenging to integrate into an analysis pipeline. For example, combining different imputation methods from different pieces of software is not currently straightforward. \pkg{simputation} resolves some of these complications with a simple approach of a unified syntax for all imputation, and always returns a dataframe of imputed values. This reduces the friction of working with other tools, but comes at the cost of identifying imputed values. An ideal approach would use consistent, simple data structures that work with other analysis tools, and help track missing values. This would make imputation outputs tidy, streamlining subsequent analysis. \hypertarget{exploration}{% \subsection{Exploration}\label{exploration}} The primary focus of most missing data packages is making inferences, and exploring imputed values, not on exploring relationships in missing values, and identifying possible patterns. Texts covering the exploration phase of missing data have the same problem as with modelling: the input is tidy, but the output does not work with other tools \citep{VanBuuren2012}; this is inefficient. Methods for exploring missing values are primarily covered in literature on interactive graphics \citep{Swayne1998, Unwin1996, Cook2007}, and are picked up again in a discussion of a graphical user interface \citep{Cheng2015}. The missingness matrix \(B\) can be used to assess missing data dependence. It has been used in interactive graphics, dubbed a ``shadow matrix'', to link missing and imputed values to the data, facilitating their display \citep{Swayne1998}, focusing heavily on multivariate numeric data. This is an idea upon which this new work builds. The MANET (Missings Are Now Equally Treated) software \citet{Unwin1996} focussed on multivariate categorical data, with missingness explicitly added as a category. MANET also provided univariate visualizations of missing data using linked brushing between a reference plot of the missingness for each variable, and a plot of the data as a histogram or barplot. The MANET software is no longer maintained and cannot be installed. The approach of \citet{Swayne1998} in the software XGobi, further developed in ggobi \citep{Cook2007}, focussed on multivariate quantitative data. Missingness is incorporated into plots in ggobi by setting them to be 10\% below the minimum value. MissingDataGUI provides a Graphical User Interface (GUI) for exploring missing data structure, both numerically and visually. Using a GUI to explore missing data facilitates rapid insight into missingness structures. However, this comes as a trade off, as insights are not captured or recorded with a GUI, making it challenging to incorporate into reproducible analyses. This distracts and breaks analysis workflow, inviting mistakes. VIM (Visualizing and Imputing Missing Data) provides visualization methods to identify and explore observed, imputed, and missing values. These include spinograms, spinoplots, missingness matrices, plotting missingness in the margins of other plots, and other summaries. However, these visualizations do not map variables to graphical aesthetics, creating friction when moving through analysis workflows, making them difficult to extend to new circumstances. Additionally, data used to create the visualizations cannot be accessed, posing a barrier to further exploration. \pkg{ggplot2} removes missing values with a warning (Figure \ref{fig:warning}), and only incorporates missingness into visualizations when mapping a discrete variable with missings to a graph aesthetic. This has some limitations, shown in Figure \ref{fig:gg-box-na}, a boxplot visualization of school grade and test scores. If there are missings in a continuous variable like test score, \pkg{ggplot2} omits the missings and prints a warning message. However, if a discrete variable like school year has missing values, an NA category is created for school year, where scores are placed (Figure \ref{fig:gg-box-na}). \begin{CodeChunk} \begin{figure} {\centering \includegraphics[width=1\linewidth]{tidy-missing-data-paper_files/figure-latex/gg-box-na-1} } \caption[ggplot2 provides different visualizations depending on what type of data has missing values for data of student test scores in school year]{ggplot2 provides different visualizations depending on what type of data has missing values for data of student test scores in school year. (A) Data is complete; (B) Missings are only in year - an additional 'NA' boxplot is created; (C) Missings only in scores, no additional missingness information is shown; (D) Missings in both scores and year, additional missing information is shown. The missingness category is only shown when there are missings in categorical variables such as year (plots (B) and (D)). In (C), no missingness information is given on the graphic, despite there being missings in score, and a warning message is displayed about the number of missing values omitted.}\label{fig:gg-box-na} \end{figure} \end{CodeChunk} \hypertarget{extensions}{% \section{Tidy framework for missings}\label{extensions}} Applying tidyverse principles has the potential to clarify missing data exploration, visualization, and imputation. This section discusses how these principles are applied for data structures (\ref{data-structure}), common operations (verbs) (\ref{verbs}), graphics (\ref{graphics}), and data summaries (\ref{num-sum}). Care has been taken to make the names and design these features intuitive for their purpose, and is discussed throughout. \hypertarget{data-structure}{% \subsection{Data structure}\label{data-structure}} A data structure facilitating exploration of missing data needs to be simple to reason with and to transport with a data analysis, otherwise it will not be used. A useful template common in missing data literature, is the \(B\) matrix, where 0 and 1 indicate not missing and missing, respectively. This matrix was used to explore missing values in the interactive graphics library XGobi, called a ``missing value shadow'', or ``shadow matrix'', defined as a copy of the original data with indicator values of missingness. The shadow matrix could be interactively linked to the data. However, there are some limitations to the shadow matrix. Namely, the values 0 and 1 can be confusing representations of missing values, since it is not clear if 0 indicates an absence of observation, or the presence of a missing value. We propose a new form for tidy representation of missing data based on these ideas from past research. Four features are added to the shadow matrix to facilitate analysis of missing data, illustrated in Figure \ref{fig:nabularfig}. \begin{enumerate} \def\arabic{enumi}.{\arabic{enumi}.} \item \textbf{Missing value labels}. Simple labels for missing and not missing to clearly identify missing values for analysis and plotting. We propose ``NA'' and ``!NA''. (Figure \ref{fig:nabularfig}). This improves the 0 and 1 values in \(B\), which do not clearly identify missingness. These values follow a principle: ``\textbf{clarity of labelling}'' - the matrix's meaning is transparent, and anybody looking at these values could understand what they mean, which is not the case of binary values. Equally, these values could instead be ``missing'' or ``present''. \item \textbf{Special missing values}: Building on \textbf{missing value labels}, the values in the shadow matrix can be ``special'' missing values, indicated by ``NA\_\textless suffix\textgreater{}''. For example: \texttt{NA\_instrument} uses a short label, ``instrument'', indicating instrument error resulting in missing values. These could be also used to indicate imputations, (Figure \ref{fig:nabularfig}). \item \textbf{Coordinated names}: Variable names in the shadow matrix gain a consistent short suffix, "\_NA", keeping names coordinated throughout analysis (Figure \ref{fig:nabularfig}). It makes a clear distinction with \texttt{var\_NA} being a random variable of the missingness of a variable, \texttt{var}. This suffix is short and easy to remember during data analysis, and shifts the focus from the value of a variable, to its missingness state. \item \textbf{Connectedness}: Binding the shadow matrix column-wise to the original data creates a single, connected, \emph{nabular} data format, in sync with the data. It is useful for visualization, summaries, and tracking imputed values, discussed in more detail in \ref{nabular-data}. \end{enumerate} \hypertarget{nabular-data}{% \subsection{Nabular data}\label{nabular-data}} \emph{Nabular} data binds the shadow matrix column-wise to the original data. It is a portmanteau of \texttt{NA} and \texttt{tabular}. \emph{Nabular} data explicitly links missing values to data, keeping corresponding observations together and removing the possibility of mismatching records (Figure \ref{fig:nabularfig}). \emph{Nabular} data facilitates visualization and summaries by allowing the user to reference the missingness of a variable in a coordinated way: the missingness of \texttt{var}, as \texttt{var\_NA} during analysis. \emph{Nabular} data is a snapshot of the missingness of the data. This means when \emph{nabular} data are imputed, those imputed values can easily be identified in analysis (\texttt{var} vs \texttt{var\_NA}). \emph{Nabular} data is not unlike classical data formats with quality or flag variables associated with each measured variable, e.g.~Scripps CO2 data \citep{Keeling2005-scripps}, GHCN data \citep{Durre2008-ghcn}. \begin{CodeChunk} \begin{figure} {\centering \includegraphics[width=1\linewidth]{images/diagram} } \caption[The process of creating nabular data]{The process of creating nabular data. Data transformed to shadow matrix, where values are either not missing or missing: '!NA' or 'NA'. The shadow matrix can be converted to long form to create missingness summary plots. Nabular data is created by binding the columns of the data and shadow matrix. Special missing values (such as -99) are identified as special missings, and values imputed and tracked. Nabular data can be used to identify imputations and explore data values alongside missings, providing a useful format for missing data exploration and analysis.}\label{fig:nabularfig} \end{figure} \end{CodeChunk} Although a binary missingness matrix, \(B\), could be generated during analysis instead of using a \emph{nabular} data structure, there are key advantages to using \emph{nabular} data. Firstly, missing value labels \texttt{NA} and \texttt{!NA} are clearer than TRUE or FALSE. Secondly, special missing values cannot be added easily during analysis with a logical matrix. Finally, the logical matrix cannot capture which values are imputed if imputation has already taken place. Imputing values on \emph{nabular} data automatically tracks these imputations. Using additional columns to represent missingness information follows best practices for data organization, described in \citet{Ellis2017} and \citet{Broman2017}: (1) Keep one thing in a cell and (2) Describe additional features of variables in a second column. Here they suggest to indicate censored data with an extra variable called ``VariableNameCensored'', which would be TRUE if censored, otherwise FALSE. This information can now be represented in the shadow columns as special missing values. Encoding special missing values is achieved by defining logical conditions and suffixes. This is implemented with the \texttt{recode\_shadow} function in \pkg{naniar} (\ref{verbs-recode}). Special missing values are not a new idea, and have been implemented in other statistical programming languages, SPSS, SAS, and STATA. These typically represent missing values as a full-stop, \texttt{.}, and record special missing values as \texttt{.a} - \texttt{.z}. These special values from these languages break the tidy principle of one column having one type of value, as they record both the value, and the multivariate missingness state. \hypertarget{verbs}{% \subsection{Missing data operations}\label{verbs}} Common missing data operations can be considered verbs, in the tidyverse sense. For missing data, these include: \textbf{scan}, \textbf{replace}, \textbf{add}, \textbf{shadow}, \textbf{impute}, \textbf{track}, and \textbf{flag}. Data can be \textbf{scanned} to find possible missings not coded as \texttt{NA}. These values can then be \textbf{replaced} with \texttt{NA}. To facilitate exploration, summaries of missingness can be \textbf{added} as a summary column to the original data. The data can be augmented with the \textbf{shadow matrix} values, helping explore missing data, as well as facilitating the process of \textbf{imputing}, and \textbf{tracking}. Finally, unusual or specially coded missing values can be \textbf{flagged}. \hypertarget{verbs-search}{% \subsubsection{scan: Searching for common missing value labels}\label{verbs-search}} This operation is used to search the data for specific conventional representations of missings, such as, ``N/A'', ``MISSING'', \texttt{-99}. This is implemented in the function \texttt{miss\_scan\_count()}, which returns a table of occurrences of that value for each variable. A list of common NA values for numbers and characters, can be provided to help check for typical representations of missings. These are implemented in \texttt{naniar} as \texttt{common\_na\_numbers} and \texttt{common\_na\_strings}. \hypertarget{verbs-replace-with}{% \subsubsection{replace: Replacing values with missing values}\label{verbs-replace-with}} Once possible missing values have been identified, these values can be replaced. For example, a dataset could have the values \texttt{-99} meaning a missing value. \texttt{naniar} implements replacement with the function, \texttt{replace\_with\_na()}. Values -99 could be replaced in the \texttt{x} column with: \texttt{replace\_with\_na(dat\_ms,\ replace\ =\ list(x\ =\ -99))}. For operating on multiple variables, there are scoped variants for \texttt{replace\_with\_na}: \texttt{\_all}, \texttt{\_if}, and \texttt{\_at}. This means \texttt{replace\_with\_na\_all} operates on \textbf{all} columns, \texttt{replace\_with\_na\_at} operates \textbf{at} specific columns, and \texttt{replace\_with\_na\_if} makes a conditional change on columns \textbf{if} they meet some condition (such as \texttt{is.numeric} or \texttt{is.character}). \hypertarget{verbs-add-cols}{% \subsubsection{add: Adding missingness summary variables}\label{verbs-add-cols}} Understanding the missingness structure can be improved by adding summary information alongside the data. For example, in \citet{Tierney2015}, the proportion of missings in a row is used as the outcome in a model to identify variables important in predicting missingness structures. \texttt{naniar} implements a series of functions to add these missingness summaries to the data, starting with \texttt{add\_}. These are inspired by the \texttt{add\_count} function in \texttt{dplyr}, which adds count information for specified groups or conditions. \texttt{naniar} provides operations to add the number or proportion of missingness, the missingness cluster, or the presence of any missings, with: \texttt{add\_n\_miss()}, \texttt{add\_prop\_miss()}, \texttt{add\_miss\_cluster}, and \texttt{add\_any\_miss()}, respectively (Table \ref{tab:add-missing-info}). There are also functions for adding information about shadow values, and readable labels for any missing values with \texttt{add\_label\_shadow()} and \texttt{add\_label\_missings()}. \begin{CodeChunk} \begin{table} \caption{\label{tab:add-missing-info}Overview of the 'add' functions in naniar} \centering \begin{tabular}[t]{l|l} \hline Function & Adds column which:\\ \hline add\_n\_miss(data) & Contains the number missing values in a row\\ \hline add\_prop\_miss(data) & Contains the proportion of missing values in a row\\ \hline add\_miss\_cluster(data) & Contains the missing value cluster\\ \hline \end{tabular} \end{table} \end{CodeChunk} \hypertarget{verbs-nabular}{% \subsubsection{shadow: Creating nabular data}\label{verbs-nabular}} \emph{Nabular} data has the shadow matrix column-bound to existing data. This facilitates visualization and summaries, and allows for imputed values to be tracked. \emph{Nabular} data can be created with \texttt{nabular()}: \begin{CodeChunk} \begin{CodeInput} R> nabular(dat_ms) \end{CodeInput} \begin{CodeOutput} # A tibble: 5 x 6 x y z x_NA y_NA z_NA <dbl> <chr> <dbl> <fct> <fct> <fct> 1 1 A -100 !NA !NA !NA 2 3 N/A -99 !NA !NA !NA 3 NA <NA> -98 NA NA !NA 4 -99 E -101 !NA !NA !NA 5 -98 F -1 !NA !NA !NA \end{CodeOutput} \end{CodeChunk} \hypertarget{verbs-recode}{% \subsubsection{flag: Describing different types of missing values}\label{verbs-recode}} Unusual or spurious data values are often identified and \texttt{flagged}. For example, there might be special codes to mark an individual dropping out of a study, known instrument failure in weather instruments, or for values censored in analysis. \texttt{naniar} provides tools to encode these special types of missingness in the shadow matrix of \emph{nabular} data, using \texttt{recode\_shadow()}. This requires specifying the variable to contain the flagged value, the condition for flagging, and a suffix. This is then recoded as a new factor level in the shadow matrix, so every column is aware of all possible new values of missingness. For example, -99 could be recoded to indicate a broken machine sensor for the variable \texttt{x} with the following: \begin{CodeChunk} \begin{CodeInput} R> nabular(dat_ms) R+ recode_shadow(x = .where(x == -99 ~ "broken_sensor")) \end{CodeInput} \begin{CodeOutput} # A tibble: 5 x 6 x y z x_NA y_NA z_NA <dbl> <chr> <dbl> <fct> <fct> <fct> 1 1 A -100 !NA !NA !NA 2 3 N/A -99 !NA !NA !NA 3 NA <NA> -98 NA NA !NA 4 -99 E -101 NA_broken_sensor !NA !NA 5 -98 F -1 !NA !NA !NA \end{CodeOutput} \end{CodeChunk} \hypertarget{verbs-impute}{% \subsubsection{impute: Imputing values}\label{verbs-impute}} \texttt{naniar} does not reinvent the wheel for imputation, instead working with existing methods. However, \texttt{naniar} provides a few imputation methods to facilitate exploration and visualization: \texttt{impute\_below}, \texttt{impute\_mean}, and \texttt{impute\_median}. While useful to explore structure in missingness, they are not recommended for use in analysis. \texttt{impute\_below} imputes values below the minimum value, with some controllable jitter (random noise) to reduce overplotting. Similar to \texttt{simputation}, each \texttt{impute\_} function returns the data with values imputed. However, \texttt{naniar} does not use a formula syntax, instead each function has ``scoped variants'' \texttt{\_all}, \texttt{\_at} and \texttt{\_if} as in \ref{verbs-replace-with}. \texttt{impute\_} functions with no scoped variant, (\texttt{impute\_mean}), will work on a single vector, but not a data.frame. One challenge with this approach is imputed value locations are not tracked. This issue is resolved with \code{nabular} data covered in Section \ref{verbs-track}. \hypertarget{verbs-track}{% \subsubsection{track: Shadow and impute missing values}\label{verbs-track}} To evaluate imputations they need to be tracked. This is achieved by first using \texttt{nabular} data, then imputing, and imputed values can then be referred to by their shadow variable, \texttt{\_NA} (Figure \ref{fig:track-impute-example}). The code below shows the track pattern, first using \texttt{nabular}, then imputing with \texttt{impute\_lm}. \texttt{label\_shadow} then adds a label to facilitate identifying missings: \begin{CodeChunk} \begin{CodeInput} R> aq_imputed <- nabular(airquality) R+ as.data.frame() R+ simputation::impute_lm(Ozone ~ Temp + Wind) R+ simputation::impute_lm(Solar.R ~ Temp + Wind) R+ add_label_shadow() R> R> head(aq_imputed) \end{CodeInput} \begin{CodeOutput} Ozone Solar.R Wind Temp Month Day Ozone_NA Solar.R_NA Wind_NA Temp_NA 1 41.00000 190.0000 7.4 67 5 1 !NA !NA !NA !NA 2 36.00000 118.0000 8.0 72 5 2 !NA !NA !NA !NA 3 12.00000 149.0000 12.6 74 5 3 !NA !NA !NA !NA 4 18.00000 313.0000 11.5 62 5 4 !NA !NA !NA !NA 5 -11.67673 127.4317 14.3 56 5 5 NA NA !NA !NA 6 28.00000 159.5042 14.9 66 5 6 !NA NA !NA !NA Month_NA Day_NA any_missing 1 !NA !NA Not Missing 2 !NA !NA Not Missing 3 !NA !NA Not Missing 4 !NA !NA Not Missing 5 !NA !NA Missing 6 !NA !NA Missing \end{CodeOutput} \end{CodeChunk} Multiple missing or imputed values can be mapped to a graphical element in ggplot2 by setting the \texttt{color} or \texttt{fill} aesthetic in ggplot to \texttt{any\_missing}, a result of the \texttt{add\_label\_shadow()} function. (Figure \ref{fig:track-impute-example}). Imputed values can also be compared to complete case data, grouping by \texttt{any\_missing}, and then summarizing, similar to other dplyr summary workflows shown below in Table \ref{tab:impute-summary-out}, showing similarities and differences in imputation methods. \begin{CodeChunk} \begin{figure} {\centering \includegraphics[width=1\linewidth]{tidy-missing-data-paper_files/figure-latex/track-impute-example-1} } \caption[Scatterplot (A) and density plots (B and C) of ozone and solar radiation from the airquality dataset containing imputed values from a linear model]{Scatterplot (A) and density plots (B and C) of ozone and solar radiation from the airquality dataset containing imputed values from a linear model. Imputed values are colored green, and data values orange. Imputed values are similar, but slightly trended to the mean.}\label{fig:track-impute-example} \end{figure} \end{CodeChunk} \begin{CodeChunk} \begin{CodeInput} R> aq_imputed R> group_by(any_missing) R> summarise_at(.vars = vars(Ozone), R> .funs = lst(min, mean, median, max)) \end{CodeInput} \end{CodeChunk} \begin{CodeChunk} \begin{table} \caption{\label{tab:impute-summary-out}Summarising values of imputed vs non imputed values. Comparing imputed values (denoted as 'Missing', since they were previously missing), the mean and median values are similar, but the minimum and maximum values are very different.} \centering \begin{tabular}[t]{l|r|r|r|r} \hline any\_missing & min & mean & median & max\\ \hline Missing & -16.86418 & 41.22494 & 45.4734 & 78\\ \hline Not Missing & 1.00000 & 42.09910 & 31.0000 & 168\\ \hline \end{tabular} \end{table} \end{CodeChunk} \hypertarget{graphics}{% \section{Graphics}\label{graphics}} Missing values are often ignored when plotting data - which is why the data visualization software, MANET, was named and is an acronym corresponding to ``Missings Are Now Equally Treated'' \citep{Unwin1996}. However, plots can help to identify the type of missing value patterns, even those of MCAR, MAR or MNAR. Here we summarise how to systematically explore missing patterns visually, and define useful plots to make, relative to the nabular data structure. \hypertarget{overviews}{% \subsection{Overviews}\label{overviews}} The first step is to get an overview of the extent of the missingness. Overview visualizations for variables and cases are provided with \texttt{gg\_miss\_var} and \texttt{gg\_miss\_case} (Figure \ref{fig:gg-miss-case-var}A), drawing attention to the amount of missings, and ordering by missingness. The ``airquality'' dataset (from base R), is shown, and contains daily air quality measurements in New York, from May to September, 1973. We learn from figure \ref{fig:gg-miss-case-var}A-B, that two variables contain missings, approximately one third of observations have one missing value, and a tiny number of observations are missing across two variables. These overview plots are created from the shadow matrix in long form (Figure \ref{fig:nabularfig}). Numerical statistics can also be reported (Section \ref{num-sum}). \begin{CodeChunk} \begin{figure} {\centering \includegraphics[width=0.9\linewidth]{tidy-missing-data-paper_files/figure-latex/gg-miss-case-var-1} } \caption[Graphical summaries of missingness in the airquality data]{Graphical summaries of missingness in the airquality data. Missings invariables (A) and cases (B), and for a birds eye view with missingness as a heatmap in (C), and with clustering applied (D). There are missing values in Ozone and Solar.R, with Ozone having more missings. Not many cases have two missings. Most missingness is from cases with one missing value. The default output (C) and ordered by clustering on rows and columns (D). These overviews are made possible using the shadow matrix in long form. There are only missings in ozone and solar radiation, and there appears to be some structure to their missingness.}\label{fig:gg-miss-case-var} \end{figure} \end{CodeChunk} The shadow matrix can be put into long form, allowing both the variables and cases to be displayed using a heatmap style visualization, with \texttt{vis\_miss()} from \pkg{visdat} \citep{visdat} (Figure \ref{fig:nabularfig}). This also provides numerical summaries of missingness in the legend, and for each column (Figure \ref{fig:gg-miss-case-var}C-D). Clustering can be applied to the rows, and columns arranged in order of missingness (Figure \ref{fig:gg-miss-case-var}D). Similar visualizations are available in other packages such as \pkg{VIM}, \pkg{mi}, \pkg{Amelia}, and \pkg{MissingDataGUI}. A key improvement is \texttt{vis\_miss()} orients the visualization analogous to a regular data structure: variables form columns and are named at the top, and each row is an observation. Using \pkg{ggplot2}, as the foundation, makes the plot easily customizable. \begin{CodeChunk} \begin{figure} {\centering \includegraphics{tidy-missing-data-paper_files/figure-latex/airquality-upset-1} } \caption[The pattern of missingness in the airquality dataset shown in an upset plot]{The pattern of missingness in the airquality dataset shown in an upset plot. Only Ozone and Solar.R have missing values, and Ozone has the most missing values. There are 2 cases where both Solar.R and Ozone have missing values.}\label{fig:airquality-upset} \end{figure} \end{CodeChunk} The number of times observations are missing together can be visualized using an ``upset plot'' \citep{Conway2017}. An alternative to a Venn diagram, an upset plot shows the size and features of overlapping sets, and scales well with more variables. An upset plot can be constructed from the shadow matrix, as shown in Figure \ref{fig:airquality-upset} which displays the overlapping counts of missings in the airquality data. The bottom right shows the combinations of missingness, the top panel shows the size of these combinations, and the bottom left shows missingness in each variable. This provides similar information to Figure \ref{fig:gg-miss-case-var}A-D, but more clearly illustrating overlapping missingness, where 2 cases are missing together in variables Solar.R and Ozone. \hypertarget{univariate}{% \subsection{Univariate}\label{univariate}} Missing values are typically not shown for univariate visualizations such as histograms or densities. Two ways to use \emph{nabular} data to present univariate data with missings are discussed. The first imputes values below the range to facilitate visualizations. The second displays two plots of the same variable according to the missingness of a chosen variable. \textbf{Imputing values below the range}. To visualize the amount of missings in each variable, the data is transformed into \emph{nabular} form, then values are imputed below the range of data using \texttt{impute\_below\_all}. Figure \ref{fig:impute-shift-histogram}A shows a histogram of Ozone values on the right in green, and the histogram of missing ozone values on the left, in orange. The missings in Ozone are imputed and \texttt{Ozone\_NA} is mapped to the fill aesthetic. Nabular data facilitates adding counts of missingness to a histogram, allowing examination of a variables' distribution of values, and also the magnitude of missings. \textbf{Univariate split by missingness}. The distribution of a variable can be shown according to the missingness of another variable. The shadow matrix part of nabular is used to handle the faceting, and color mapping. Figure \ref{fig:impute-shift-histogram} shows the values of temperature when ozone is present, and missing, using a faceted histogram (B), and an overlaid density (C). This shows how values of temperature are affected by the missingness of ozone, and reveals a cluster of low temperature observations with missing ozone values. This type of plot can facilitate exploring missing data distributions. For example, we would expect if data were MCAR, for values to be roughly uniformly missing throughout the histogram or density. \begin{CodeChunk} \begin{figure} {\centering \includegraphics[width=0.75\linewidth]{tidy-missing-data-paper_files/figure-latex/impute-shift-histogram-1} \includegraphics[width=0.75\linewidth]{tidy-missing-data-paper_files/figure-latex/impute-shift-histogram-2} \includegraphics[width=0.75\linewidth]{tidy-missing-data-paper_files/figure-latex/impute-shift-histogram-3} } \caption[Univariate summaries of missingness]{Univariate summaries of missingness. (A) A histogram using nabular data to show the values and missings in ozone. Values are imputed below the range to show the number of missings in ozone and colored according to missingness of ozone (`Ozone\_NA`). There are about 35 missings in Ozone. Panel C shows temperature according to missingness in ozone from in the airquality dataset. A histogram of temperature facetted by the missingness of ozone (B), or a density of temperature colored by missingness in ozone (C). These show a cluster of low temperature observations with missing ozone values, but temperature is otherwise similar.}\label{fig:impute-shift-histogram} \end{figure} \end{CodeChunk} \hypertarget{bivariate}{% \subsection{Bivariate}\label{bivariate}} To visualize missing values in two dimensions the missing values can be placed in plot margins, by imputing values below the range of the data. Using \emph{nabular} data identifies imputed values, and color makes missingness pre-attentive \citep{treisman1985}. The steps of imputing and coloring are combined into \texttt{geom\_miss\_point()}. Figure \ref{fig:geom-miss} shows a mostly uniform spread of missing values for Solar.R and Ozone. As \texttt{geom\_miss\_point()} is a defined \pkg{ggplot2} geometry, it works with features such as faceting and mapping other variables to graphical aesthetics. \begin{CodeChunk} \begin{figure} {\centering \includegraphics{tidy-missing-data-paper_files/figure-latex/geom-miss-1} } \caption[Scatterplots with missings displayed at 10 percent below for the airquality dataset]{Scatterplots with missings displayed at 10 percent below for the airquality dataset. Scatterplots of ozone and solar radiation (A), and ozone and temperature (B). There are missings in ozone and solar radiation, but not temperature.}\label{fig:geom-miss} \end{figure} \end{CodeChunk} \hypertarget{multivariate}{% \subsection{Multivariate}\label{multivariate}} Parallel coordinate plots can help to visualize missingness beyond two dimensions. They transform variables to the same scale, ranging between 0 and 1. The \texttt{oceanbuoys} dataset from \pkg{naniar} is used for this visualization, containing measurements of sea and air temperature, humidity, and east west and north south wind speeds. Data was collected in 1993 and 1997, to understand and predict El Niño and El Niña. Figure \ref{fig:parallel-cord-plot} is a parallel coordinate plot of \texttt{oceanbuoys}, with missing values imputed to be 10\% below the range, and values colored according to whether humidity was missing (\texttt{humidity\_NA}). Figure \ref{fig:parallel-cord-plot} shows humidity is missing at low air and sea temperatures, and humidity is missing in one year, and one location. \begin{CodeChunk} \begin{figure} {\centering \includegraphics[width=1\linewidth]{tidy-missing-data-paper_files/figure-latex/parallel-cord-plot-1} } \caption[Parallel coordinate plot shows missing values imputed 10\% below range for the oceanbuoys dataset]{Parallel coordinate plot shows missing values imputed 10\% below range for the oceanbuoys dataset. Values are colored by missingness of humidity. Humidity is missing for low air and sea temperatures, and is missing for one year and one location. }\label{fig:parallel-cord-plot} \end{figure} \end{CodeChunk} \hypertarget{num-sum}{% \section{Numerical summaries}\label{num-sum}} This section describes approaches to summarizing missingness, and an implementation in \pkg{naniar}. Numerical summaries should be easy to remember with consistent names and output, returning either a single number \ref{single-num-sum}, or a dataframe \ref{sum-tab-missings}, so they integrate well with plotting and modelling tools. How these work with other tools in an analysis pipeline is shown in \ref{num-sum-w-group}. \hypertarget{single-num-sum}{% \subsection{Single number summaries}\label{single-num-sum}} The overall number, proportion, or percent of missing values in a dataset should be simple to calculate. \texttt{naniar} provides the functions \texttt{n\_miss}, \texttt{prop\_miss} and \texttt{pct\_miss}, as well as their complements. Summaries for variables and cases are made by appending \texttt{\_case} or \texttt{\_var} to these summaries. An overview is shown in Table \ref{tab:n-prop-pct-miss-complete}. \begin{CodeChunk} \begin{table} \caption{\label{tab:n-prop-pct-miss-complete}Single number summaries of missingness and completeness of the airquality dataset. The functions follow consistent naming, making them easy to remember, and their use clear.} \centering \begin{tabular}[t]{l|r|l|r} \hline Missing Function & missing value & Complete function & complete value\\ \hline n\_miss & 44.00 & n\_complete & 874.00\\ \hline prop\_miss & 0.05 & prop\_complete & 0.95\\ \hline pct\_miss & 4.79 & pct\_complete & 95.21\\ \hline pct\_miss\_case & 27.45 & prop\_complete\_case & 72.55\\ \hline pct\_miss\_var & 33.33 & pct\_complete\_var & 66.67\\ \hline \end{tabular} \end{table} \end{CodeChunk} \hypertarget{sum-tab-missings}{% \subsection{Summaries and tabulations of missing data}\label{sum-tab-missings}} Presenting the number and percent of missing values for each variable, or case, provides a summary usable in data handling, or in models to inform imputation. For example, potentially dropping variables, or deciding to include others in an imputation model. Another useful approach is to tabulate the frequency of missing values for each variable or case; that is, the number of times there are zero missings, one missing, two, and so on. These summaries and tabulations are shown for variables in Tables \ref{tab:miss-var-summary} and \ref{tab:miss-var-table}, and implemented with \texttt{miss\_var\_summary} and \texttt{miss\_var\_table}. Case-wise (row-wise) summaries and tabulations are implemented with \texttt{miss\_case\_summary} and \texttt{miss\_case\_table}. These summaries order rows by the number of missings (\texttt{n\_miss}), to show the most missings at the top. The number of missings across a repeating span, or finding ``runs'' or ``streaks'' of missings in given variables can also be useful for identifying missingness patterns, and are implemented with \texttt{miss\_var\_span}, and \texttt{miss\_var\_run}. \begin{CodeChunk} \begin{table} \caption{\label{tab:miss-var-summary}\texttt{miss\char`_var\char`_summary} provides the number and percent of missings in each variable in airquality. Only ozone and solar radiation have missing values.} \centering \begin{tabular}[t]{l|r|r} \hline variable & n\_miss & pct\_miss\\ \hline Ozone & 37 & 24.2\\ \hline Solar.R & 7 & 4.6\\ \hline Wind & 0 & 0.0\\ \hline Temp & 0 & 0.0\\ \hline Month & 0 & 0.0\\ \hline Day & 0 & 0.0\\ \hline \end{tabular} \end{table} \end{CodeChunk} \begin{CodeChunk} \begin{table} \caption{\label{tab:miss-var-table}\texttt{miss\char`_var\char`_table} tabulates the amount of missing data in each variable in airquality. This shows the number of variables with 0, 7, and 37 missings, and the percentage of variables with those amounts of missingness. There are few missingness patterns.} \centering \begin{tabular}[t]{r|r|r} \hline n\_miss\_in\_var & n\_vars & pct\_vars\\ \hline 0 & 4 & 66.7\\ \hline 7 & 1 & 16.7\\ \hline 37 & 1 & 16.7\\ \hline \end{tabular} \end{table} \end{CodeChunk} \hypertarget{num-sum-w-group}{% \subsection{Combining numerical summaries with grouping operations}\label{num-sum-w-group}} It is useful to explore summaries and tabulations within groups of a dataset. \texttt{naniar} works with \texttt{dplyr}'s \texttt{group\_by} operator to produce grouped summaries, which work well with the ``pipe'' operator. The code and table below \ref{tab:group-miss-var-summary} show an example of missing data summaries for airquality, grouped by month. \begin{CodeChunk} \begin{table} \caption{\label{tab:group-miss-var-summary}\texttt{miss\char`_var\char`_summary} combined with \texttt{group\char`_by} provides a grouped summary of the missingness in each variable, for each Month of the airquality dataset. Only the first 10 rows are shown. There are more ozone missings in June than May.} \centering \begin{tabular}[t]{r|l|r|r} \hline Month & variable & n\_miss & pct\_miss\\ \hline 5 & Ozone & 5 & 16.1\\ \hline 5 & Solar.R & 4 & 12.9\\ \hline 5 & Wind & 0 & 0.0\\ \hline 5 & Temp & 0 & 0.0\\ \hline 5 & Day & 0 & 0.0\\ \hline 6 & Ozone & 21 & 70.0\\ \hline 6 & Solar.R & 0 & 0.0\\ \hline 6 & Wind & 0 & 0.0\\ \hline 6 & Temp & 0 & 0.0\\ \hline 6 & Day & 0 & 0.0\\ \hline \end{tabular} \end{table} \end{CodeChunk} \hypertarget{case-study}{% \section{Application}\label{case-study}} This section shows how the methods described so far are used together in a data analysis workflow. We analyse a case study of housing data for the city of Melbourne from January 28, 2016 to March 17, 2018. The data was compiled by scraping weekly property clearance data \citep{Kaggle-2018-data}. There are 27,247 properties, and 21 variables in the dataset. The variables include real estate type (town house, unit, house), suburb, selling method, number of rooms, price, real estate agent, sale date, and distance from the Central Business District (CBD). The goal in analyzing this data is to accurately predict Melbourne housing prices. The data contains many missing values. As a precursor to building a predictive model, this analysis focuses on understanding the patterns of missingness. \hypertarget{case-study-explore-pattern}{% \subsection{Exploring patterns of missingness}\label{case-study-explore-pattern}} Figure \ref{fig:housing-miss-case-var}A shows 9 variables with missing values. The most missings are in ``building area'', followed by ``year built'', and ``land size'', with similar amounts of missingness in ``Car'', ``bathroom'', ``bedroom2'', ``longitude'', and ``latitude.'' Figure \ref{fig:housing-miss-case-var}B reveals there are up to 50\% missing values in cases, and the majority of cases have more than 5\% values missing. The variables ``building area'' and ``year built'' have more than 50\% missing data, and so could perhaps be omitted from subsequent analysis, as imputed values are likely to be spurious. Three missingness clusters are revealed by visualizing missingness in the whole dataset, clustering and arranging the rows and columns of the data (Figure \ref{fig:applic-vis-miss}). \begin{CodeChunk} \begin{figure} {\centering \includegraphics[width=1\linewidth]{tidy-missing-data-paper_files/figure-latex/housing-miss-case-var-1} } \caption[The amount of missings in variables (A) and cases (B) for Melbourne housing data]{The amount of missings in variables (A) and cases (B) for Melbourne housing data. (A) Build area and year built have more than 50\% missing, and car, bathroom, bedroom2 and longitude and latitude have about 25\% missings. (B) Cases are missing 5 - 50\% of values. The majority of missingness is in selected cases and variables.}\label{fig:housing-miss-case-var} \end{figure} \end{CodeChunk} \begin{CodeChunk} \begin{figure} {\centering \includegraphics[width=0.85\linewidth]{tidy-missing-data-paper_files/figure-latex/applic-vis-miss-1} } \caption[Heatmap of clustered missingness for housing data reveals structured missingness]{Heatmap of clustered missingness for housing data reveals structured missingness. Three groups of missingness are apparent. At the top: building area to longitude; the middle: building area and year built; the end: building area, year built, and landsize.}\label{fig:applic-vis-miss} \end{figure} \end{CodeChunk} Figure \ref{fig:housing-upset} shows missingness patterns with an \texttt{upset} plot \citep{Conway2017}, displaying 8 intersecting sets of missing variables. Two patterns stand out: two, and five variables missing, providing further evidence of the missingness patterns seen in Figures \ref{fig:housing-miss-case-var} and \ref{fig:housing-upset}. \begin{CodeChunk} \begin{figure} {\centering \includegraphics[width=1\linewidth]{tidy-missing-data-paper_files/figure-latex/housing-upset-1} } \caption[An upset plot of 8 sets of missingness in the housing data]{An upset plot of 8 sets of missingness in the housing data. Missingness for each variable is shown on the bottom left. Connected dots show co-occurences of missings in variables. Two missingness patterns are clear, year built and building area, and lattitude through to building area.}\label{fig:housing-upset} \end{figure} \end{CodeChunk} Tabulating the number of missings in variables in Table \ref{tab:housing-miss-var-case-table} (left) shows three groups of missingness. Tabulating missings in cases (Table \ref{tab:housing-miss-var-case-table} (right)) shows six patterns of missingness. These overview plots lead to the removal of two variables from with more than 50\% missingness from analysis: ``building area'' and ``year built''. \begin{CodeChunk} \begin{table} \caption{\label{tab:housing-miss-var-case-table}Tabulating missingness for variables (left) and cases (right) to understand missingness patterns. 14 variables have 0 - 3 missings, 6 variables have 6000 - 9000 missings, and 2 variables have 15 - 16,000 missings. About 30\% of cases have no missings, 45\% of cases have 1 - 6 missings, and about 23\% of cases have 8 or more missings. There are different patterns of missingness in variables and cases, but they can be broken down into smaller groups.} \centering \begin{tabular}[t]{r|r|r} \hline n\_miss\_in\_var & n\_vars & pct\_vars\\ \hline 0 & 10 & 47.6\\ \hline 1 & 2 & 9.5\\ \hline 3 & 1 & 4.8\\ \hline 6254 & 2 & 9.5\\ \hline 6441 & 1 & 4.8\\ \hline 6447 & 1 & 4.8\\ \hline 6824 & 1 & 4.8\\ \hline 9265 & 1 & 4.8\\ \hline 15163 & 1 & 4.8\\ \hline 16591 & 1 & 4.8\\ \hline \end{tabular} \centering \begin{tabular}[t]{r|r|r} \hline n\_miss\_in\_case & n\_cases & pct\_cases\\ \hline 0 & 8887 & 32.6\\ \hline 1 & 3237 & 11.9\\ \hline 2 & 7231 & 26.5\\ \hline 3 & 1370 & 5.0\\ \hline 4 & 79 & 0.3\\ \hline 5 & 8 & 0.0\\ \hline 6 & 203 & 0.7\\ \hline 8 & 6229 & 22.9\\ \hline 9 & 2 & 0.0\\ \hline 11 & 1 & 0.0\\ \hline \end{tabular} \end{table} \end{CodeChunk} \hypertarget{case-study-explore-for-imp}{% \subsection{Exploring missingness patterns for imputation}\label{case-study-explore-for-imp}} Using information from \ref{case-study-explore-pattern}, the following variables are explored for features predicting missingness: ``Land size'', ``latitude'', ``longitude'', ``bedroom2'', ``bathroom'', ``car'', and ``land size''. Missingness structure is explored by clustering the missing values into groups. Then, a classification and regression trees (CART) model is applied to predict these missingness clusters using the remaining variables \citep{Tierney2015, Barnett2017}. Two clusters of missingness are identified and predicted using all variables in the dataset with the CART package \texttt{rpart} \citep{rpart}, and plotted using the \texttt{rpart.plot} package \citep{rpart-plot}. Importance scores reveal the following variables as most important in predicting missingness: ``rooms'', ``price'', ``suburb'', ``council area'', ``distance'', and ``region name''. These variables are important for predicting missingness, so are included in the imputation model. \begin{CodeChunk} \begin{figure} {\centering \includegraphics[width=0.9\linewidth]{tidy-missing-data-paper_files/figure-latex/rpart-plot-1} } \caption[Decision tree output predicting missingness clusters]{Decision tree output predicting missingness clusters. Type of house, year quarter, and year were important for predicting missingness cluster. The cluster with the most missingness was for quarters 1 and 4, for 2017 and 2018. Type of house, year, and year quarter are important features related to missingness structure.}\label{fig:rpart-plot} \end{figure} \end{CodeChunk} \hypertarget{case-study-imp-diagnosis}{% \subsection{Imputation and diagnostics}\label{case-study-imp-diagnosis}} \pkg{simputation} is used to implement two imputations: simple linear regression and K nearest neighbors. Values are imputed stepwise in ascending order of missingness. The track missings pattern is applied (described in \ref{verbs-track}), to assess imputed values. Imputed datasets are compared on their performance in a model predicting log house price for 4 variables (Figure \ref{fig:imputed-by-model}). Compared to KNN imputed values, the linear model imputed values closer to the mean. \begin{CodeChunk} \begin{figure} {\centering \includegraphics[width=0.95\linewidth]{tidy-missing-data-paper_files/figure-latex/imputed-by-model-1} } \caption[Boxplots of complete case data, and data imputed with KNN or linear model for different variables]{Boxplots of complete case data, and data imputed with KNN or linear model for different variables. (A) number of bedrooms, (B) number of bathrooms, (C) number of carspots, and (D) landsize (on a log10 scale). KNN had similar results to complete case, and linear model had a lower median for cars and fewer extreme values for bedrooms.}\label{fig:imputed-by-model} \end{figure} \end{CodeChunk} \hypertarget{case-study-assess-model}{% \subsection{Assessing model predictions}\label{case-study-assess-model}} Coefficients of the linear model of log price vary for room for different imputed datasets (Figure \ref{fig:tidy-coefs}). Notably, complete cases underestimate the impact of room on log price. A partial residual plot (Figure \ref{fig:partial-resid}) shows there is not much variation amongst the models from the differently imputed datasets. \begin{CodeChunk} \begin{figure} {\centering \includegraphics[width=1\linewidth]{tidy-missing-data-paper_files/figure-latex/tidy-coefs-1} } \caption[The coefficient estimate for the number of rooms varies according to the imputed dataset]{The coefficient estimate for the number of rooms varies according to the imputed dataset. Complete case dataset produced lower coefficients, compared to imputed datasets}\label{fig:tidy-coefs} \end{figure} \end{CodeChunk} \begin{CodeChunk} \begin{figure} {\centering \includegraphics[width=1\linewidth]{tidy-missing-data-paper_files/figure-latex/partial-resid-1} } \caption[Partial residual plot for each data set, complete cases (cc), and imputed with KNN (knn) or linear model (lm)]{Partial residual plot for each data set, complete cases (cc), and imputed with KNN (knn) or linear model (lm). These are plotted as hex bins, colored according to the number of points in a given hexagon. Brighter colors mean more points. Compared to complete cases, imputed data has more points clustered around zero.}\label{fig:partial-resid} \end{figure} \end{CodeChunk} \hypertarget{case-study-summary}{% \subsection{Summary}\label{case-study-summary}} The \texttt{naniar} and \texttt{visdat} packages implement the methods discussed in the paper, building on existing tidy tools and strike a compromise between automation and control, making analysis efficient, readable, but not overly complex. Each tool has clear intent and effects - summarising, plotting or generating data or augmenting data in some way. This not only reduces repetition and typing in an analysis, but most importantly, allows for clear expression of intent, making exploration of missing values fluent. \hypertarget{discussion}{% \section{Discussion}\label{discussion}} This paper has described new methods for exploring, visualizing, and imputing missing data. The work was motivated by recent developments of tidy data, and extends them for better missing value handling. The methods have standard outputs, function arguments, and behavior. This provides consistent workflows centered around data analysis that integrate well with existing imputation methodology, visualization, and modelling. The \emph{nabular} data structures discussed in the paper are simple by design, to promote flexibility. They could be used to create different visualizations than were shown in the paper. The analyst can use the data structures to decide on appropriate visualization for their problem. The data structures could also be used to support interactive graphics, in the manner of MANET and ggobi. Linking the plots (via linked brushing) could facilitate exploration of missingness, and could be implemented with \pkg{plotly} \citep{plotly} for added interactivity. Animating between different sets of imputed values could also be possible with packages like \pkg{gganimate} \citep{gganimate}. Other data structures such as spatial data, time series, networks, and longitudinal data would be supported by the inherently tabular, \emph{nabular} data, if they are first structured as wide tidy format. Large data may need special handling, and additional features like efficient storage of purely imputed values and lazy evaluation. Special missing value codes could be improved by creating special classes, or expanding low level representation of \texttt{NA} at the source code level. The methodology described in this paper can be used in conjunction with other approaches to understand multivariate missingness dependencies (e.g.~decision trees \citep{Tierney2015}, latent group analysis \citep{Barnett2017}, and PCA \citep{FactoMineR}). Evaluating imputed values using a testing framework like \citet{VanBuuren2012} is also supported. The approach meshes with the dynamic nature of data analysis, allowing the analyst to go from raw data to model data in a fluid workflow. \hypertarget{acknowledgements}{% \section{Acknowledgements}\label{acknowledgements}} The authors would like to thank Miles McBain, for his key contributions and discussions on the \texttt{naniar} package, in particular for helping implement \texttt{geom\_miss\_point}, and for his feedback on ideas, implementations, and names. We also thank Colin Fay for his contributions to the \texttt{naniar} package, in particular for his assistance with the \texttt{replace\_with\_na} functions. We also thank Earo Wang and Mitchell O'Hara-Wild for the many useful discussions on missing data and package development, and for their assistance with creating elegant approaches that take advantage of the tidy syntax. We would also like to thank those who contributed pull requests and discussions on the \texttt{naniar} package, in particular Jim Hester and Romain François for improving the speed of key functions, Ross Gayler for discussion on special missing values, and Luke Smith for helping \texttt{naniar} be more compliant with \texttt{ggplot2}. We would also like to thank Amelia McNamara for discussions on the paper. \renewcommand\refname{References}

cf0dbf0730882b0e6fbaf98282c59ad5dcd92dc2

\section{Introduction} The Internet is perhaps the world's most popular medium to exchange any kind of information. Common examples are streaming platforms, file sharing services or social media networks. Such applications are often maintained by overlay networks, called overlays for short. An overlay is a computer network that is built atop another network, the so-called underlay. In an overlay, nodes that may not be directly connected in the underlay can create virtual links and exchange messages if they know each others' addresses. The resulting links then represent a path in the underlying network, perhaps through several links. With increasing size of the network, there are several obstacles in designing these overlays. First of all, errors such as node or link failures are inevitable. Thus, there is a need for protocols that let the system recover from these faults. This can be achieved through self-stabilization, which describes a system's ability to reach a desired state from \textsl{any} initial configuration. Since its conception by Edsger W. Dijkstra in 1975, self-stabilization has proven to be a suitable paradigm to build resilient and scalable overlays that can quickly recover from changes. There is a plethora of self-stabilizing protocols for the formation and maintenance of overlay networks with a specific topology. These topologies range from simple structures like line graphs and rings \cite{onus2007linearization} to more complex overlay networks with useful properties for distributed systems \cite{Feldmann,Richa2011,corona,Skip+}. These overlays usually minimize the diameter while also maintaining a small node degree, usually at most logarithmic in the number of nodes. However, the aforementioned overlay protocols are often not concerned with path lengths in the underlay. This is remarkable, since for many use cases these path lengths and the resulting latency are arguably more important than the diameter. In this paper, we work towards closing this gap by proposing a self-stabilizing protocol that forms and maintains an overlay that resembles the Minimum Spanning Tree (MST) implied by the distances between nodes in the underlying network. In particular, we model these distances as a tree metric, i.e., as the length of the unique shortest path between two overlay nodes in a weighted tree. We chose this type of metric because one can find weighted trees in many areas of networking. In the simplest case, the physical network that interconnects the overlay nodes resembles a tree. This is often the case in data centers. Here, the servers are the tree's leaves while the switches are the tree's intermediate nodes (cf. \cite{Arregoces03,Leiserson85,Al-Fares08}). Therefore, we can define a tree metric directly on the paths in this physical infrastructure. Of course, not all physical networks are strictly structured like trees, and may instead contain cycles. However, for small networks there are practical protocols that explicitly reduce the network graph to a tree for routing purposes \cite{ieee_802_1d,Perlman85}. These protocols are executed directly on the network appliances and exclude certain physical connections, such that the remaining connections form a spanning tree. Thus, we can define a tree metric based this tree. Last, in large-scale networks like the internet neither the physical network nor the routing paths strictly resemble trees. However, there is strong evidence that even these large-scale networks can be closely approximated by or embedded into weighted trees by assigning them virtual coordinates (cf. \cite{Adcock13,Abu-Ata14,Montgolfier11,Shavitt08}). Thus, we can define a tree metric based on the shortest paths in such an embedding. In summary, tree metrics promise to be a versatile abstraction for many kinds of real-world networks. \subsection{Model \& Definitions} \label{sec:model} We consider a distributed system based on a fixed set of nodes $V$. Each node $v \in V$ represents a computational unit, e.g., a computer, that possesses a set of local variables and references to other nodes, e.g, their IP addresses. These references are immutable and cannot be corrupted. If clear from the context, we refer to the reference of some node $w \in V$ simply as $w$. Further, each node in $V$ has access to a \emph{tree metric} $d_T: V^2 \rightarrow \mathbb{R}^+$ that assigns a weight to each possible edge in the overlay. In particular, the function $d_T$ returns the weighted length of the unique shortest path between two nodes in the weighted tree $T := (V_T,E_T,f)$ with $f: V^2 \rightarrow \mathbb{R}^+$ and $V_T \supseteq V$. A node $v \in V$ can check the distance $d_T(v,w)$ only if it has a reference to $w \in V$ in its local variables. Furthermore, it can check the distance $d_T(u,w)$ of all nodes $u,w \in V$ in its local variables. Throughout this paper, we refer to the metric space $(V,d_T)$ also as a tree metric for ease of description. Sending a message from a node $u$ to another node $v$ in the overlay is only possible if $u$ has a reference to $v$. All messages for a single node are stored in its so-called channel and we assume \emph{fair message receipt}, which means each message will \emph{eventually} be received. In particular, we do \emph{not} assume FIFO-delivery, i.e., the messages may be received and processed in any order. We assume that each node runs a protocol that can perform computations on the node's local variables and send them within messages to other nodes. To formalize the protocol's execution, we use the notion of \textsl{configurations}. A configuration $c$ contains each node's internal state, i.e., its assignment of values to its local variables, its stored references, and all messages in the node's channel. We denote $C$ to be the set of all possible configurations. Further, a \emph{computation} is an infinite series of configurations $(c_t, c_{t+1}, \dots)$, such that $c_{i+1}$ is a succeeding configuration of $c_{i}$ for $i \geq t$ according to the protocol. In each step from $c_i$ to $c_{i+1}$, the following happens: One node $v \in V$ is activated and an arbitrary (possibly empty) set of messages from $v$'s channel is delivered to $v$. Once activated, the node will execute its protocol and processes all messages delivered to it. As we do not specify which node is activated and which messages get delivered, there are maybe several possible succeeding configurations $c' \in C$ for any configuration $c$. Last, we assume \emph{weakly fair execution}, which means that each node is eventually activated. Other than that, we place no restriction on the activation order. Given a subset $C' \subseteq C$, we say that the system \textit{reaches} $C'$ from $c_t$ if \textit{every} computation that starts in configuration $c_t$ eventually contains a configuration $c_{t'} \in C'$. Note that this does not imply that any succeeding configuration of $c_{t'}$ is in $C'$ as well. Based on this notion of configurations, we can now define self-stabilization. A protocol is self-stabilizing concerning a set of legal configurations $L \subseteq C$ if starting from any initial configuration $c_0 \in C$ each computation will eventually reach $L$ (Convergence) and every succeeding configuration is also in $L$ (Closure). Formally: \begin{definition}[Self-Stabilization] \label{def:self-stabilization} A protocol $\mathcal{P}$ is self-stabilizing if it fulfills the following two properties. \begin{enumerate} \item (Convergence) Let $c_0 \in C$ be \textsl{any} configuration. Then every computation that starts in $c_0$ will reach $L$ in finitely many steps. \item (Closure) Let $c_t \in L$ be \textsl{any} legal configuration. Then \textsl{every} succeeding configuration of $c_t$ is legal as well. \end{enumerate} \end{definition} Throughout this paper we distinguish between two kinds of edges in each configuration $c \in C$. We call an edge $(v,w) \in V^2$ \emph{explicit} if and only if $v$ has a reference to $w$ stored in its local variables. Otherwise, if the reference is in $v$'s channel, we call the edge \emph{implicit}. Based on this definitions, we define the directed graph $G_c := (V, E_c^X \cup E^T_c)$ where the set $E^X_c \subseteq V^2$ denotes the set of explicit edges and $E_c^T \subseteq V^2$ denotes the set of implicit edges. Further, the undirected graph $G_c^* := (V,E^*_c)$ arises from $G_c$ if we ignore all edges' direction and whether they are implicit or explicit. \subsection{Our Contribution} \label{sec:protocol_tree} Our main contribution is \textsc{BuildMST}\xspace, a self-stabilizing protocol that forms and maintains overlay representing the MSTs of all connected components of $V$. An MST is a set of edges that connects a set of nodes and minimizes the sum of the edges' weights given by the underlying metric. Because of this minimality, it can serve as a building block for more elaborate topologies. Note that in our model it is not always possible to construct the MST of all nodes, even if it is unique. To exemplify this, consider an initial configuration $c_0$ where $G^*_{c_0}$ is \emph{not} connected. Then two nodes from two different connected components of $G^*_{c_0}$ can never communicate with each other and create edges because they cannot learn each other's reference. This was remarked in \cite{corona}. In this case, it is impossible to construct an MST for all nodes as no protocol can add the necessary edges. Instead one can only construct the MST of all initially connected components, i.e., a Minimum Spanning Forest. Formally an MST is defined as follows. \begin{definition}[Minimum Spanning Tree] Let $G := (V,E)$ be a graph and $f: E \to \mathbb{R}^+$ a weight function, then the Minimum Spanning Tree $MST(G,f) \subseteq E$ is a set of edges, such that: \begin{enumerate} \item $(V,MST(G,f))$ is a connected graph, and \item $\sum_{e \in MST(G,f)} f(e)$ is minimum. \end{enumerate} For the special case of $E := V^2$, i.e., the MST over all possible edges, we write $MST(V,f)$ for short. \end{definition} In this paper, we will only consider metrics with distinct distances for each pair of nodes. Otherwise the MST may not be unique for a metric space $(V,d_T)$. If we had edges with equal distances, we would need to employ some mechanism of tie-breaking, e.g., via the nodes' identifiers. In the following, we define the set $\mathcal{L}_{MST} \subset C$ of legal configurations for \textsc{BuildMST}\xspace. We regard all configurations $c \in \mathcal{L}_{MST}$ as legal in which the explicit edges form the MST of each connected component in $c$. Further, a legal configuration may contain arbitrarily many implicit edges as long as they are part of an MST. Formally: \begin{definition}[Legal Configurations $\mathcal{L}_{MST}$] \label{def:legal_rng} Let $(V,d_T)$ be a tree metric and $c \in C$ be a configuration. Further denote $G_1, \dots, G_k$ as the connected components of $G^*_{c}$. Then the set of legal configurations $\mathcal{L}_{MST}$ is defined by the following two conditions: \begin{enumerate} \item A configuration $c \in \mathcal{L}_{MST}$ contains an explicit edge $(v,w) \in E^X_{c} $ if and only if there is component $G_i:=(V_i,E_i)$ with $\{v,w\} \in MST(V_i,d_T)$. \item A configuration $c \in \mathcal{L}_{MST}$ contains an implicit edge $(v,w) \in E^T_{c} $ only if there is a component $G_i:=(V_i,E_i)$ with $\{v,w\} \in MST(V_i,d_T)$. \end{enumerate} \end{definition} \section{Related Work} \label{sec:related-work} There are several self-stabilizing protocols for constructing spanning trees in a fixed communication graph, e.g., \cite{Blin09,Gheorge97,BlinSSS10,Higham01,Blin10,Korman}. These works do not consider a model where nodes can create arbitrary overlay edges. Instead, each node has a \emph{fixed} set of neighbors and chooses a subset of these neighbors for the tree. Furthermore, the communication graph in all these works is modeled as an arbitrary weighted graph instead of a tree. The fastest protocol given in \cite{Blin10} constructs an MST in $\mathcal{O}(N^2)$ rounds where $N$ is the number of nodes. Note that \cite{Korman} proves the existence of a protocol that converges in $\mathcal{O}(N)$ rounds but does not present and rigorously analyze an actual protocol. As stated in the introduction, these protocols can be used in the underlying network to construct a tree metric for our protocol. In the area of topological self-stabilization of overlay networks, there is a plethora of works that consider different topologies like line graphs \cite{onus2007linearization}, De-Bruijn-Graphs \cite{Feldmann,Richa2011}, or Skip-Graphs \cite{Skip+,corona}. Besides these results that do not take the underlying network into account, there are also efforts to build a topology based on a given metric. An interesting result in this area is a protocol for building the Delaunay Triangulation of two-dimensional metric space by Jacob et al. \cite{JACOB2012137}. This work bears several similarities with ours. In particular, the Delaunay Triangulation is a superset of the metric's MST and shares some of the properties we present in Section \ref{sec:preliminaries}. Also their protocol $D_{STAB}$ is very similar to our protocol \textsc{BuildMST}\xspace. Recently Gmyr et al.\ proposed a self-stabilizing protocol for constructing an overlay based on an arbitrary metric \cite{Gmyr2016}. Instead of building a spanning tree, their goal is to build an overlay in which the distance between two nodes is exactly the distance in the underlying metric. In particular, their algorithm is also applicable to a tree metric. However, note that for a tree metric the number of edges in the resulting overlay can be as high as $\Theta(N^2)$. Last, there are several non-self-stabilizing approaches for creating underlay-aware overlays, e.g., \cite{Gross12,Rowstron2001,abraham2004land,plaxton1999}. With their often-cited work in \cite{plaxton1999}, Plaxton et al.\ introduced these so-called location-aware overlays. The authors present an overlay for an underlay modeled by a growth-bounded two-dimensional metric. This means that the number of nodes within a fixed distance of a node only grows by a factor of $\Delta \in \mathbb{R}^+$ when doubling the distance. Their overlay has a polylogarithmic degree and the length of the routing paths approximate the distances in the underlying metric by a polylogarithmic factor. In \cite{abraham2004land} Abraham et al.\ extended on \cite{plaxton1999} and proposed an overlay for growth-bounded metrics where the latter is reduced to a factor of $1+\epsilon$. Here, $\epsilon \in \mathbb{R}^+$ is a parameter that can be set to an arbitrarily small value. The resulting overlay's degree depends on $\epsilon$ and is not analyzed in detail. \section{Preliminaries} \label{sec:preliminaries} In this section, we present some useful properties of tree metrics and their MSTs that will help us in designing and analyzing our protocol. Therefore, we introduce the notion of \emph{relative neighbors}. Two nodes $v,w \in V$ are relative neighbors with regard to a metric $d_T$ if there is no third node that is closer to either of them, i.e., it holds $\nexists u \in V: \left(d_T(u,v) < d_T(v,w)\right) \wedge \left(d_T(u,w) < d_T(v,w) \right)$. Throughout this paper we write $u \prec (v,w)$ as shorthand for $\left(d_T(u,v) < d_T(v,w)\right) \wedge \left(d_T(u,w) < d_T(v,w)\right)$. Relative neighbors have been defined and analyzed for a variety of metrics (cf. \cite{Toussaint80,Jaromczyk92,Supowit83}), but they prove to be especially useful in the context of tree metrics. In particular, they allow nodes to form and maintain an MST based on local criteria. This fact is stated by the following lemma: \begin{lemma} \label{lemma:equivalence} Let $(V,d_T)$ be a tree metric, then the following two statements hold: \begin{enumerate} \item $\{v,w\} \in MST(V,d_T) \Longrightarrow \nexists u \in V: \, u \prec (v,w)$ \item $\{v,w\} \not\in MST(V,d_T) \Longrightarrow \exists u \in V: \, \big( u \prec (v,w) \wedge \{v,u\} \in MST(V,d_T) \big)$ \end{enumerate} \end{lemma} In the following, we will outline the proof and thereby present some helpful lemmas, which we will reuse in Section \ref{sec:analysis}. First, we note that the lemma's first statement is generally true for all metrics (cf. \cite{Supowit83}). Thus, it remains to show the second statement. We begin the proof with a useful fact that will be at the core of many proofs in this paper. \begin{lemma} \label{thm:root} Let $(V,d_T)$ be a tree metric. Further let $u,v,w,r \in V$ be four nodes, s.t. \[ d_T(u,r) < d_T(w,r) \,\wedge\, d_T(v,r) < d_T(w,r) \] Then it either holds $u \prec (v,w)$ or $v \prec (u,w)$ (and in particular not $w \prec (u,v)$). \end{lemma} \begin{proof} Let $T:=(V_T,E_T,w)$ be the tree which implies the metric $d_T$. Further denote the unique path shortest between two nodes $s,t \in V_T$ in $T$ as $P_T(s,t)$. Last, let $\varphi \in P_T(u,v) \cap P_T(u,r) \cap P_T(v,r) $ be a node that lies on all three unique shortest paths between the nodes $u,v$ and $r$. Note that $\varphi$ is also called \emph{median} of $u,v$ and $r$ and is unique in a tree. First, we show that \[ d_T(u,\varphi) < d_T(w,\varphi) \,\,\, \wedge \,\,\, d_T(v,\varphi) < d_T(w, \varphi) \] Assume for contradiction that $d_T(u,\varphi) > d_T(w,\varphi)$. From the triangle inequality we can follow that $d_T(w,r) \leq d_T(w,\varphi) + d_T(\varphi,r)$. If we combine these two inequalities, we deduce \[ d_T(w,r) \leq d_T(w,\varphi) + d_T(\varphi,r) < d_T(u,\varphi) + d_T(\varphi,r) = d_T(u,r) \] This would be a contradiction to our initial assumption that $d_T(u,r) < d_T(w,r)$. Therefore, it must hold $d_T(u,\varphi) < d_T(w,\varphi)$. The proof for $d_T(v,\varphi) < d_T(w, \varphi)$ is analogous and thus, our claim holds. Second, we prove that it holds $\varphi \in P_T(v,w)$ or $\varphi \in P_T(u,w)$. Assume for the sake of contradiction that neither $\varphi \in P_T(v,w)$ nor $\varphi \in P_T(u,w)$. Then there is a path from $u$ to $v$ via $w$ that does not contain $\varphi$. This is a contradiction to the fact that there is only one simple path $P_T(u,v)$ between $u$ and $v$ and per definition it holds $\varphi \in P_T(u,v)$. Now distinguish between the two cases we have just shown: \begin{enumerate} \item If $\varphi \in P(v,w)$, the following inequality must hold. \begin{align*} d_T(u,v) = d_T(u,\varphi) + d_T(\varphi,v) < d_T(w,\varphi) + d_T(\varphi,v) = d_T(v,w) \end{align*} This follows from the fact that $d_T(u,\varphi) < d_T(w,\varphi)$. Since $d_T(v,w) < d_T(u,v)$ is one of the two requirements for $w \prec (u,v)$, it cannot hold in this case. \item Otherwise, if $\varphi \in P(u,w)$, it holds \begin{align*} d_T(u,v) = d_T(u,\varphi) + d_T(\varphi,v) < d_T(u,\varphi) + d_T(\varphi,w) = d_T(u,w) \end{align*} This follows from the fact that $d_T(v,\varphi) < d_T(w,\varphi)$. Since $d_T(u,w) < d_T(u,v)$ is required for $w \prec (u,v)$, it cannot hold in this case either. \end{enumerate} Hence, it must hold $u \prec (v,w)$ or $v \prec (u,w)$, which was to be shown. \end{proof} Using Lemma \ref{thm:root} we can show the following. \begin{lemma} \label{lemma:greedy-routing} Let $(V,d_T)$ be a tree metric and $v,w \in V$ two of its nodes. Further, let $v_0, \dots, v_k \in V$ be the unique path from $v_0 := v$ to $v_k := w$ in the MST. Then it holds: \[ d_T(v_{i},v) < d_T(v_{i+1},v) \,\,\, \forall v_i \in (v_0, \dots, v_{k-1}) \] \end{lemma} \begin{proof}[Lemma \ref{lemma:greedy-routing}] Assume for the sake of contradiction that the lemma does not hold. Let $v_i$ be the first node on a path to $v$ for which it instead holds $d_T(v_{i}, v) > d_T(v_{i+1},v)$. Note, that it cannot hold $d_T(v_{i}, v) = d_T(v_{i+1},v)$ because we assume pairwise distinct distances. Since $d_T(v,v_0) = 0$ it must hold that $i \geq 1$. Therefore $v_{i-1}$ is well-defined and it must hold $d_T(v_{i-1}, v) < d_T(v_{i},v)$ because $v_i$ is the first node that is further away from $v$ than its successor. Combining these two facts yields: \[ d_T(v_{i-1},v) < d_T(v_{i},v) \,\,\, \wedge \,\,\, d_T(v_{i+1},v) < d_T(v_{i},v) \] Following Lemma \ref{thm:root} it must therefore either hold $v_{i-1} \prec (v_{i},v_{i+1})$ or $v_{i+1} \prec (v_{i-1},v_{i})$. In particular that means, it holds either hold $d_T(v_{i-1}, v_{i+1}) < d_T(v_{i},v_{i+1})$ or $d_T(v_{i-1}, v_{i+1}) < d_T(v_{i-1},v_{i})$. In the following we assume the latter since both cases are analogous. We will now show that we can improve the MST by swapping $\{v_{i-1}, v_{i+1}\}$ for $\{v_{i-1},v_{i}\}$, which is a contradiction. If we remove $\{v_{i-1},v_{i}\}$ from $MST(V,d_T)$ we divide the tree into two subtrees $T_{i-1}$ and $T_i$, which contains $v_{i-1}$ and $v_{i}$, respectively. Further, it holds that $v_{i+1}$ is in $T_i$ because it connected to $v_i$ via the edge $\{v_{i},v_{i+1}\}$. Thus, the edge $\{v_{i-1}, v_{i+1}\}$ also connects $T_{i-1}$ and $T_i$ and has lower weight than $\{v_{i},v_{i-1}\}$. That means, we can improve $MST(V,d_T)$, which is a contradiction. Therefore, there cannot be such a first deviator $v_i$ and the lemma must hold. \end{proof} In the remainder, we conclude the proof for Lemma \ref{lemma:equivalence}. Therefore, let $v,w \in V$ be two nodes with $\{v,w\} \not\in MST(V,d_T)$. Further, let $u \in V$ be the first node of the path $P_{vw}$ from $v$ to $w$ in the MST. Such a node must exist because there is no direct edge between $v$ and $w$ in the MST. Note that $P_{vw}$ contains the same nodes as a path $P_{wv}$ from $w$ to $v$ but in reverse order. Thus, we can apply Lemma \ref{lemma:greedy-routing} in "both directions". That means, the node $u$ with $\{v,u\} \in MST(V,d_T)$ must be closer to $w$ than $v$, but also closer to $v$ than its successor in $P_{wv}$. A simple induction then yields that $u \prec (v,w)$. Since by definition it holds $\{v,u\} \in MST(V,d_T)$, this proves the lemma. \section{Protocol} \label{sec:protocol} In this section, we describe our protocol \textsc{BuildMST}\xspace, which forms and constructs an overlay according to Definition \ref{def:legal_rng}. Intuitively, the protocol works as follows: Upon activation, a node $v \in V$ checks, which of its current neighbors are relative neighbors. All nodes that fulfill the property are kept in the neighborhood. All others are delegated in a greedy fashion. This idea resembles that of the protocols in \cite{JACOB2012137} and \cite{onus2007linearization}, where essentially the same technique is used for different underlying metrics, i.e., the two-dimensional plane and a line. \begin{lstlisting}[caption={\textsc{BuildMST}\xspace},label=alg:BuildDG,captionpos=b,float, mathescape=true, belowskip=-\baselineskip, frame = single] Upon activation a node $v \in V$ performs: for all $w \in N_v$ if $\exists u \in N_v: \, u \prec (v,w)$ $N_u \longleftarrow N_u \cup \{w\}$ #$v$ delegates $w$ to $u$ $N_v \longleftarrow N_v \setminus \{w\}$ else $N_w \longleftarrow N_w \cup \{v\}$ #$v$ introduces itself \end{lstlisting} The pseudocode for this protocol is given in Figure \ref{alg:BuildDG}. Therein, each node $v \in V$ only maintains a single variable $N_v \subseteq V$. This is a set that contains all currently stored references to other nodes. It contains each entry only once and multiple occurrences of the same reference are merged automatically. With each activation, a node iterates over all nodes in $w \in N_v$ and checks whether to delegate $w$ or to introduce itself. In this context, a delegation means that $v$ sends a reference of $w$ to $u$ and then deletes the reference to $w$ from $N_v$. The protocol assures that a node $v$ delegates $w$ to $u$, if and only if it holds $u \prec (v,w)$. Otherwise $v$ introduces itself to $w$, which means that it sends a reference of itself to $w$. Note, that the primitives of introduction and delegation preserve the system's connectivity (cf. \cite{corona}). In the pseudocode introductions and delegations are indicated by statements of the form $N_u \longleftarrow N_u \cup \{w\}$. This notation is used for convenience. It describes that the executing node $v$ sends a message containing a reference of $w$ to $u$. The variable $N_u$ is not directly changed and $w$ is only added in some later configuration when $u$ is activated and the message is delivered to $u$. A graphical example of the protocol's computations can be seen in Figure~\ref{fig:protocol}. \begin{figure} \begin{subfigure}[t]{.49\textwidth} \centering \begin{tikzpicture}[every tree node/.style={draw,circle, minimum size=1.5em}, level distance=1cm,sibling distance=1.2cm, edge from parent path={(\tikzparentnode) -- (\tikzchildnode)}] \tikzset{level 1/.style={level distance=0.905cm, sibling distance=0.16cm}} \tikzset{level 2/.style={level distance=0.7cm, sibling distance=1.5cm}} \tikzset{level 3/.style={level distance=0.6cm, sibling distance=0.32cm}} \Tree [.\node[opacity=1] (w) {$w$}; % \edge[opacity=0.8] node[auto=left] {$6$};[.\node[opacity=0.8, dotted] (phi) {}; \edge[opacity=0.8] node[auto=right] {$1$}; [.\node (v) {$u$}; ] \edge[opacity=0.8] node[auto=left] {$5$}; [ .\node (u) {$v$}; ] ] % ] \path (u) edge[->, line width=1pt, bend left, color=red] node[auto=left] (t) {$6$} (v); \path (u) edge[->, line width=1pt, bend right, color=red] node[auto=right] (t) {$11$} (w); \end{tikzpicture} \caption{An example configuration: $v$ has $u$ and $w$ in its local memory. Note that $u$ and $w$ are neighbors of each other in the \textsc{MST}\xspace.} \end{subfigure} ~ \begin{subfigure}[t]{.49\textwidth} \centering \begin{tikzpicture}[ every tree node/.style={draw,circle, minimum size=1.5em}, level distance=1cm,sibling distance=1.2cm, edge from parent path={(\tikzparentnode) -- (\tikzchildnode)}] \tikzset{level 1/.style={level distance=0.905cm, sibling distance=0.16cm}} \tikzset{level 2/.style={level distance=0.7cm, sibling distance=1.5cm}} \tikzset{level 3/.style={level distance=0.6cm, sibling distance=0.32cm}} \Tree [.\node[opacity=1] (w) {$w$}; % \edge[opacity=0.8] node[auto=left] {$6$};[.\node[opacity=0.8, dotted] (phi) {}; \edge[opacity=0.8] node[auto=right] {$1$}; [.\node (v) {$u$}; ] \edge[opacity=0.8] node[auto=left] {$5$}; [ .\node (u) {$v$}; ] ] % ] \path (u) edge[->,line width=1pt, bend left, color=red] node[auto=left] (t) {$6$} (v); \path (v) edge[->, line width=1pt, color=red, dashed] node[auto=left] (t) {$6$} (u); \path (v) edge[->, line width=1pt, bend left, color=red, dashed] node[auto=left] (t) {$7$} (w); \end{tikzpicture} \caption{The succeeding configuration: $v$ has delegated $w$ to $u$ and introduced itself to $u$.} \end{subfigure} \caption[An example of the execution of \textsc{BuildMST}\xspace]{An example of the protocol's execution. The black edges are part of the underlying tree. Red edges denote the overlay's edges. The dotted edges are implicit, i.e., the references are still the node's channel. Solid edges are explicit,i.e., the references are in the node's memory. The numbers denote the edges' weights.} \label{fig:protocol} \end{figure} \section{Analysis} \label{sec:analysis} In this section we rigorously analyze \textsc{BuildMST}\xspace. We prove the protocol's correctness with regard to Definition \ref{def:self-stabilization} and the set of configurations given in Definition \ref{def:legal_rng}. Furthermore, we bound the protocol's convergence time. The main result of this section is that \textsc{BuildMST}\xspace is indeed a self-stabilizing protocol as stated by the following theorem: \begin{theorem} \label{theorem:self_stabil} Let $(V,d_T)$ be a tree metric. Then \textsc{BuildMST}\xspace is a self-stabilizing protocol that constructs an overlay with regard to $\mathcal{L}_{MST}$. \end{theorem} In this section, we will concentrate on initial configurations $c_0 \in C$ where $G^*_{c_0}$ is connected. Since two nodes from different components can \emph{never} communicate with each other (cf. \cite{corona}), the result can trivially be extended to all initial configurations. Our proof's structure is as follows. First, we will show that eventually the system will contain all edges of $MST(V,d_T)$ and also keeps them in all subsequent configurations. This will be the major part of this section. Then we show that all remaining edges that are not part of the MST but may still be part of a configuration will eventually vanish. This proves the protocol's convergence. Last, we prove that once the system is in a legal configuration, the set of explicit edges does not change and no more edges that are not part of the MST are added. This shows the protocol's closure. Over the course of this section we will refer to all edges $e \in MST(V,d_T)$ as \textit{valid} edges. We call all other edges \textit{invalid}. We begin by showing that the system eventually reaches a configuration that contains all valid edges. For the proof, we assign a potential to each configuration $c \in C$. As the potential, we choose the weight of the minimum spanning tree that can be constructed from all implicit and explicit edges in the configuration if we ignore their direction, i.e., we consider the MST of $G_c^*$. Since $G_c^*$ is simply an undirected, weighted graph with unique edge weights, it must have a unique minimum spanning tree if it is connected. This fact is a well-known result in graph theory. The potential is formally defined as follows: \begin{definition}[Potential] \label{def:potential} Let $c \in C$ be a configuration and $\mathcal{M}_c := MST(E_c^*,d_T)$ the minimum spanning tree of $G_c^* := (V,E_c^*)$, then the potential $\Phi: C \rightarrow \mathbb{R}^+$ is defined as $\Phi(c) := \begin{cases} \sum_{e \in \mathcal{M}_c} d_T(e)& \mbox{if $G_c^*$ is connected}\\ \infty&\mbox{else} \end{cases}$ \end{definition} The weight of the globally optimal minimum spanning tree $MST(V,d_T)$ that considers all edges provides a lower bound for the potential. Therefore, it cannot decrease indefinitely. In the following, we show that the potential decreases monotonically and once the system reached a configuration with minimum potential it will eventually contain all valid edges. First, we show that the potential can not increase. \begin{lemma} \label{lemma:potential_decreasing} Consider an execution of \textsc{BuildMST}\xspace and let the system be in configuration $c \in C$. Further, let $c'$ be an arbitrary succeeding configuration of $c$. Then it holds $\Phi(c') \leq \Phi(c)$. \end{lemma} \begin{proof} To simplify notation let $E$ and $E'$ be the set of all edges in $G_c^*$ and $G_{c'}^*$ respectively. In the following, we will show that we can only construct equally good or better spanning trees from the edges in $E'$. Per definition, exactly one node $v \in V$ is activated in the transition from $c$ to $c'$. This node then executes the for-loop given in the pseudocode in Listing \ref{alg:BuildDG}. Let $v$ be the node that is activated and $\{v,w\} \in E$ be an edge that is delegated removed from $E$ during its activation, i.e., $v$ delegates $w$ to some node $u$. As a result of the delegation, the configuration $c'$ contains the (implicit) edge $(u,w) \in E^T_{c'}$ and thus $E'$ contains the edge $\{u,w\} \in E'$. This allows us to view the delegation as swapping edge $\{v,w\}$ for $\{u,w\}$. In the following we observe the swaps $(e_1,e'_1), \dots, (e_k,e'_k)$, such that $e_i \in E$ is swapped for $e'_i \in E'$ in the transition from $c$ to $c'$. The order in which we observe these swaps must be consistent with the protocol. That means that two delegations must appear in the same order as they could in the for-loop, i.e., $v$ can only delegate to node whose reference's are still in its local memory. Next, we define $E_0, \dots E_k \subseteq V^2$ with $E_0 := E$ and $E_i := E_{i-1} \setminus \{e_i\} \cup \{e'_i\}$ for $i>0$ as the edge sets resulting from these swaps. As the proof's main part we inductively show that each $MST(E_i, d_T)$ with $i \in \{1,\dots,k\}$ has a lower or equal weight than $MST(E_{i-1}, d_T)$. For this, we distinguish between two cases. First, if $e_i \not\in MST(E_{i-1}, d_T)$, the spanning tree is not affected by the swap and thus the weight remains equal. Second, if $e_i \in MST(E_{i-1}, d_T)$, we must show that we can construct an equally good spanning tree in $E_i$. For this, consider $\mathcal{M}_i := MST(E_{i-1}, d_T) \setminus \{e_i\} \cup \{e'_i\}$. Note $\mathcal{M}_i$ and $MST(E_{i-1}, d_T)$ only differ in the edges $e_i := \{v,w\}$ and $e'_i := \{u,w\}$. For the delegation of $w$ to $u$ it must have held $u \prec (v,w)$ and thus $d_T(u,w) < d_T(v,w)$. Therefore, $\mathcal{M}_i$ has lower weight than $MST(E_{i-1}, d_T)$. It remains to show that $\mathcal{M}_i$ is a connected spanning tree for $V$. Further denote $T_v$ and $T_w$ as the subtrees of $MST(E_{i-1}, d_T)$ connected by $\{v,w\}$. To prove that $\mathcal{M}_i$ is a spanning tree, we must show that $\{u,w\}$ connects $T_v$ and $T_w$, i.e., it holds $u \in T_v$. Suppose for contradiction that $u \in T_w$. Then the path from $v$ to $u$ in $MST(E_{i-1}, d_T)$ contains the edge $\{v,w\}$. Further, note that $E_{i-1}$ must have contained the edge $\{v,u\}$ because $v$ cannot delegate any node to $u$ without having a reference to $u$ itself. Therefore, the edges $\{v,w\}$ and $\{v,u\}$ are \emph{both} part of $E_i$ and \emph{both} connect $T_v$ and $T_w$. Now consider that $\{v,u\}$ is shorter than $\{v,w\}$, because a delegation requires $u \prec (v,w)$ and thus $d_T(v,u) < d_T(v,w)$. Hence $MST(E_{i-1}, d_T)$ could be improved by swapping $\{v,w\}$ for $\{v,u\}$. This is a contradiction because $MST(E_{i-1}, d_T)$ is a minimum spanning tree. Therefore $u \in T_v$ and the edge $\{u,w\}$ connects $T_v$ and $T_w$. Thus, $\mathcal{M}_i$ is a spanning tree that can be constructed solely from edges in $E_i$. Further, it has a lower or equal weight than $MST(E_{i-1}, d_T)$. The lemma then follows by a simple induction. \end{proof} It remains to show that the potential actually decreases until it reaches the minimum. That means, we need to show that there cannot be a configuration with suboptimal potential where no more delegations that decrease the potential occur. Note that the proof of Lemma \ref{lemma:potential_decreasing} tells us that the potential decreases if an edge $\{v,w\} \in \mathcal{M}_c$ is delegated. Therefore, we first show that in each suboptimal spanning tree there is a node that can potentially detect an improvement. \begin{lemma} \label{lemma:locally_checkable} Let the system be in configuration $c \in C$, s.t. the potential $\Phi(c)$ is not minimum. Then there must exist nodes $u,v,w \in V$, such that \[ \big( u \prec (v,w) \big) \,\, \wedge \,\, \big(\{v,u\} \in \mathcal{M}_c\big) \,\, \wedge \,\, \big(\{v,w\} \in \mathcal{M}_c\big) \] \end{lemma} \begin{proof} Let $\mathcal{M}_c$ be the minimum spanning tree of a configuration $c$. Since the potential is suboptimal, there must be two nodes $v,w \in V$ with $\{v,w\} \in MST(V,d_T)\setminus\mathcal{M}_c$. Since $\mathcal{M}_c$ is connected, there is a path $v:=v_0, v_1, \dots, v_k:=w$ from $v$ to $v_k$ in $\mathcal{M}_c$. Now consider $v_1$. According to Lemma \ref{lemma:equivalence} it cannot hold $v_1 \prec (v,w)$ because $\{v,w\} \in MST(V,d_T)$. Thus, it holds $d_T(v,w) < d_T(v_1,w)$ or $d_T(v,w) < d_T(v_1,v)$. Now we distinguish between two cases: \begin{enumerate} \item Assume, it holds $d_T(v,w) < d_T(v_1,w)$. Next, consider that it holds $d_T(v_{k-1},w) > d_T(v_k,w)$ because no node can be closer to $w = v_k$ than $w$ itself. Thus, there must be a first node $v_i$ on the path with $d_T(v_i,w) > d_T(v_{i+1},w)$. Since $d_T(v,w) < d_T(v_1,w)$ it further holds that $i \geq 1$. Therefore, the node $v_{i-1}$ is well-defined and it must hold $d_T(v_{i-1}, w) < d_T(v_{i},w)$ because $v_{i+1}$ is the first node that is closer to $w$ than its successor. Hence, it holds $\big(d_T(v_{i-1},w) < d_T(v_{i},w)\big)$ and $\big(d_T(v_{i+1},w) < d_T(v_{i},w)\big)$ Following Lemma \ref{thm:root} it follows that either $v_{i-1} \prec (v_{i},v_{i+1})$ or $v_{i+1} \prec (v_{i-1},v_{i})$. Since in both cases all of the involved edges are part of $\mathcal{M}_c$, the lemma follows. \item Assume, it holds $d_T(v,w)<d_T(v_1,v)$. Then there must be node $v_i$, such that $d_T(v_i,v) > d_T(v_{i+1},v)$. Otherwise a simple induction from $v_0$ to $v_k$ would yield that $d_T(v,w)<d_T(v_k,v)$. Since $v_k = w$ this is a contradiction. The rest of the proof is analogous to the previous case. For the first deviator $v_i$ it holds $d_T(v_i,v) > d_T(v_{i+1},v)$ and $d_T(v_i,v) > d_T(v_{i-1},v)$ and thus, we can apply Lemma \ref{thm:root} to conclude the proof. \end{enumerate} \end{proof} Lemma \ref{lemma:locally_checkable} only made assumptions about edges in $G^*_c$ and did not consider the actual edges. Since each node only has access to its local references, node $v$ can only perform a delegation if it ever has explicit references to $u$ and $w$. In the following lemma, we will see that if the potential does not decrease, a node will eventually have the references in local memory. \begin{lemma} \label{lemma:eventually_explicit} Let the system be in configuration $c \in C$ and let $\mathcal{M}_c$ be the minimum spanning tree of $c$. If the potential does not decrease, then the following two statements hold: \begin{enumerate} \item Every computation that starts in $c$ will reach a set $C_c \subset C$, such that \[ \forall c^* \in C_c: \, \big( \{v,w\} \in \mathcal{M}_{c} \Rightarrow (v,w) \in E^X_{c^*} \big) \] \item \textsl{Every} succeeding configuration of $c^* \in C_c$ is in $C_c$ as well \end{enumerate} \end{lemma} \begin{proof} Recall from the proof of Lemma \ref{lemma:potential_decreasing} that if \textsl{any} edge is removed from $\mathcal{M}_c$, then the potential decreases. If we assume that the potential does not decrease, no edge is ever removed from $\mathcal{M}_c$. Now fix an edge $\{v,w\} \in \mathcal{M}_c$. This edge exists because $(v,w) \in E^X_c \cup E^T_c$ or $(w,v) \in E^X_c \cup E^T_c$. In the following we assume that $(v,w) \in E^X_c \cup E^T_c$, the other case is analogous. In order to prove the lemma we must show that if $(v,w) \in E^T_{c}$ it will become explicit and there must eventually be an explicit edge $(w,v) \in E^X_{c'}$ in some later configuration $c' \in C$. Further, none these edges is ever delegated as long as the potential does not decrease. \begin{enumerate} \item If $(v,w) \in E_c^T$ is implicit in $c$, it will eventually be delivered to $v$. Since the potential does not decrease, $v$ never performs a delegation of an edge that is part of a minimum spanning tree. Thus, the system must reach a configuration $c'$ with $(v,w) \in E^X_{c'}$. By the same argument, any subsequent configuration $c' \in C$ with the same potential must also contain $(v,w) \in E^X_{c'}$. \item If $(v,w) \in E^X_c$ is explicit, then the edge $(w,v) \in E^T_{c'}$ will eventually be added in some later configuration $c'$ when $v$ is activated. This happens, because we assume, that no edge is delegated. If $v$ does not delegate $w$, it introduces itself upon its activation. Thus, it adds an implicit edge $(w,v)$ that will eventually become explicit. \end{enumerate} In conclusion: For each edge $\{v,w\} \in \mathcal{M}_c$ in a configuration $c$, there will eventually be a configuration $c'$ with edges $(v,w),(w,v) \in E^X_{c'}$. Since no edge of $\mathcal{M}_c$ is ever delegated as long as the potential is fixed, these explicit edges stay part of all subsequent configurations if the potential does not decrease. \end{proof} Using this fact we can finally show that the following holds: \begin{lemma}[Convergence I] \label{lemma:convergenceI} The following two statements hold: \begin{enumerate} \item Every computation will reach a set $C_{MST} \subset C$, such that \[ \forall c' \in C_{MST}: \,\,\, \big(\{v,w\} \in MST(V,d_T) \Rightarrow (v,w) \in E^X_{c'}\big) \] \item \textsl{Every} succeeding configuration of $c' \in C_{MST}$ is in $C_{MST}$ as well. \end{enumerate} \end{lemma} \begin{proof} The proof is structured in two parts: First, we show that the system eventually reaches a configuration $c_{min} \in C$ with minimum potential from any initial configuration $c \in C$. Second, we elaborate eventually the system will contain all valid edges once it is in $c_{min}$ and argue, why all subsequent configurations must also contain all valid edges. \begin{enumerate} \item Let $c \in C$ be an arbitrary configuration with suboptimal potential. Further, let $\mathcal{M}_c$ be a minimum spanning tree of that configuration. Following Lemma \ref{lemma:locally_checkable}, there must be nodes $u,v,w \in V$ with edges, such that \[ u \prec (v,w) \,\, \wedge \,\, \{v,w\} \in \mathcal{M}_c \,\, \wedge \,\, \{v,u\} \in \mathcal{M}_c \] Since the potential does not decrease, we can apply Lemma \ref{lemma:eventually_explicit}. Thus, there will eventually be a configuration $c' \in C$ with \[ (v,u) \in E^X_{c'} \,\, \wedge \,\, (v,w) \in E^X_{c'} \] This causes $v$ to eventually delegate $w$ to $u$ and decrease the potential (if the potential does not decrease otherwise before). Since the potential is lower bounded by the weight of $MST(V,d_T)$ that by definition cannot decrease, the system must eventually reach a configuration $c_{min} \in C$ with minimum potential. \item If the system is in a configuration $c_min \in C$, the potential cannot decrease further. Thus, we can apply Lemma \ref{lemma:eventually_explicit} and eventually the system reaches a configuration $c' \in C_{MST}$, such that: \[ \{v,w\} \in MST(V,d_T) \Rightarrow (v,w) \in E^X_{c'} \] Hence, the configuration $c'$ contains all valid edges. Further, Lemma \ref{lemma:eventually_explicit} states that these edges are not delegated as long as the potential does not decrease. Since the potential is minimum, it can never decrease and therefore the statement follows. \end{enumerate} Hence, starting in any configuration $c \in C$ the system will eventually reach a configuration $c' \in C_{MST}$ with all valid edges. Further, all subsequent configurations of $c'$ are in $C_{MST}$ as well. This was to be shown. \end{proof} This concludes the first part of the convergence proof. Now we know that the system eventually converges to a superset of the \textsc{MST}\xspace. It remains to show that eventually all invalid edges will vanish. \begin{lemma}[Convergence II] \label{lemma:convergenceII} The following two statements hold: \begin{enumerate} \item Eventually each computation will reach a set of configurations $C' \subset C$, such that \[ \forall c \in C': \,\,\, \big(\{v,w\} \not\in MST(V,d_T) \Rightarrow \{v,w\} \not\in E^*_{c}\big) \] \item \textsl{Every} succeeding configuration of $c' \in C'$ is in $C'$ as well. \end{enumerate} \end{lemma} \begin{proof} For this proof, we will again employ a potential function. The potential of a configuration $c \in C$ is the weight of the longest invalid edge. Formally: \[ \tilde{\Phi}(c) := \begin{cases} \max_{e \in E^*_{c}\setminus MST(V,d_T) } d_T(e)& \textit{if } E^*_{c}\setminus MST(V,d_T) \neq \emptyset\\ 0& \textit{else}\\ \end{cases} \] If this potential is $0$, there are no invalid edges left. This trivially follows from the fact that all distances are greater than zero. Just as with the other potential, we will show that this potential (1) never increases and (2) will decrease as long as it is not minimum. \begin{enumerate} \item $\tilde{\Phi}(c)$ \textsl{cannot increase.}\\ For the proof let $c \in C$ be an arbitrary configuration and $c' \in C$ be any succeeding configuration of $c$. To prove the assumption, we show that the protocol never adds an invalid edge that is longer than any existing edge. Let $v \in V$ be the node that is activated in the transition from $c$ to $c'$ and let $w \in V$ be an explicit neighbor of $v$ in $G_c$. Then $v$ performs one of the following two actions that add new edges to the system: \begin{enumerate} \item If $v$ introduces itself to $w$, it adds the implicit edge $(w,v) \in E^T_{c'}$ to the system. Since the edge $(v,w) \in E^X_c$ with $d_T(v,w) = d_T(w,v)$ is already present, this cannot raise the potential. \item If $v$ delegates $w$ to some node $u \in V$, then it adds the implicit edge $(u,w) \in E^T_{c'}$ to system if it was not already present before. Since for delegation it must hold that $d_T(u,w)<d_T(v,w)$ for the existing edge $(v,w) \in E^X_c$, the new edge cannot raise the potential. \end{enumerate} Thus, it holds $\tilde{\Phi}(c') \leq \tilde{\Phi}(c)$. \item $\tilde{\Phi}(c)$ \textsl{will eventually decrease if $\tilde{\Phi}(c)>0$.}\\ Let $c \in C$ be an arbitrary configuration and $\{v,w\} \in E^*_{c}$ an invalid edge in $c$ with $\tilde{\Phi}(c) = d_T(v,w)$. Since $\{v,w\}$ is oblivious of the true edge's direction, both $(v,w)$ and $(w,v)$ could be part of the configuration. Since the proof is analogous for both edges, we will only consider $(v,w)$ and show that all instances of this edge will eventually be delegated. First, consider the case that $v$ has an explicit edge to $w$. Since we assume the system is in a configuration that contains all edges in $MST(V,d_T)$, we can use Lemma \ref{lemma:equivalence}. According to the Lemma, there must be a node $u \in V$ with an explicit edge $(v,u) \in E^X_c$ and $u \prec (v,w)$. Thus, $v$ will delegate $w$ to $u$ upon activation and add the edge $(u,w)$ with $d_T(u,w)<d_T(v,w)$. Second, consider the case that $(v,w) \in E^T_c$ is implicit. For the proof, we need to mind that there can be multiple instances of the reference to $w$ in $v$'s channel. The potential will only sink once all of these instances are gone. Therefore let $\theta_v$ be the number of references to $w$ in $v$'s channel. In the following, we will show that $\theta_v$ decreases to $0$. Note that $\theta_v$ can only be raised if some node $u \in V$ delegates $w$ to $v$ or $w$ introduces itself. A delegation always implies that some node $u$ has a reference to $w$ and it holds $d_T(u,w) > d_T(v,w)$. In that case, there exists an invalid edge $\{u,w\} \in E^*_c$, which is longer than $\{v,w\}$. This is impossible because $\{v,w\}$ is by assumption the longest invalid edge. Hence, $\theta_v$ may only increase if $w$ introduces itself. To do this, there must be an explicit edge $(w,v)$. However, we can apply the same argumentation as above for $(v,w)$ and see that $w$ must delegate its reference of $v$ to some other node $u' \in V$ instead of introducing itself. In summary, the protocol never increases $\theta_v$ and thus it can only decrease if a reference is delivered to $v$. Since this eventually happens to every reference, the system will reach a configuration with no references of $w$ in $v$'s channel. \end{enumerate} Hence, the potential will eventually reach $0$ and no more invalid edges are left. Furthermore, no more invalid edges can ever be added as this would increase the potential. \end{proof} Thus, we have shown that starting from any weakly connected initial configuration $c \in C$ the system will converge to a superset of the \textsc{MST}\xspace and eventually to a legal configuration. This is the combined result of Lemmas \ref{lemma:convergenceI} and \ref{lemma:convergenceII}. To complete the proof we must show that the system once it is legal never leaves the set of legal configurations. Formally: \begin{lemma}[Closure] \label{lemma:closure} Let the system be in a legal configuration $c \in \mathcal{L}_{MST}$, then every succeeding configuration $c' \in C$ is also legal. \end{lemma} However, the lemma is a direct corollary of Lemmas \ref{lemma:convergenceI} and \ref{lemma:convergenceII}. Hence, \textsc{BuildMST}\xspace is self-stabilizing with regard to Definition \ref{def:self-stabilization}. This proves Theorem \ref{theorem:self_stabil} and concludes the analysis of the protocol's correctness. It remains to analyze how many steps are needed until a legal configuration is reached. Therefore, we adapt the notion of \emph{asynchronous rounds} from \cite{Dolev00}. Each computation can be divided into rounds $R_0, \dots, R_t$ with $t \to \infty$, such that each round $R_i$ consists of a finite sequence of consecutive configurations. Let $c_i$ be the first configuration of $R_i$, then the rounds in the first configuration, such that: \begin{enumerate} \item For each $v\in V$, all messages that are in $v$'s channel in configuration $c_i$ have been delivered at any of $v$'s activations in this round. \item All nodes have been activated at least once. \end{enumerate} Since we assume weakly fair action execution and fair message receipt rounds are well-defined. Using this definition, we can show the following. \begin{theorem} \label{lemma:complexity} \textsc{BuildMST}\xspace needs $\mathcal{O}(N^2)$ asynchronous rounds to converge to a legal configuration. \end{theorem} \begin{proof} For the proof, we show that (1) it takes at most $\mathcal{O}(N^2)$ rounds until all valid edges are added and (2) it takes another $\mathcal{O}(N^2)$ rounds for all invalid edges to vanish. \begin{enumerate} \item Consider the potential function in Definition \ref{def:potential} and note that it can lower at most $\mathcal{O}(N^2)$ times. In the following we show that the potential will reduce at least every $4$ rounds. Therefore, assume that the system is in configuration $c \in C$ which is part of round $R_i$. According to Lemma \ref{lemma:locally_checkable}, there are nodes $u,v,w \in V$ with $\{v,u\} \in \mathcal{M}_c$ and $\{v,w\} \in \mathcal{M}_c$ such that the potential is reduced if $v$ delegates $w$ to $u$. If $v$ has the references to both these nodes in its local memory, it will perform the delegation upon its next activation. This will happen at latest in round $R_{i+1}$ because each node must be activated at least once every round. If $v$ does not have the references in its local memory, it will eventually receive them both (cf. Lemma \ref{lemma:eventually_explicit}) if the potential does not decrease. Assume w.l.o.g. that $w$ is the the latter of the two nodes whose reference is delivered to $v$. In the following, we will bound the number of rounds until the system is a configuration $c' \in C$ with $(v,w) \in E^X_{c'}$ if potential the potential does not reduce. Therefore, we make the following observations: \begin{enumerate} \item If $(v,w) \in E_c^T$ is implicit in some round $R_{j}$, it will eventually be explicit in round $R_{j+1}$. \item If $(w,v) \in E^X_c$ is explicit in some round $R_j$, then the implicit edge $(v,w) \in E^T_{c'}$ will be added in round $R_{j+1}$ or the potential decreases. Note that each node must be activated at least once every round. If $w$ delegates $v$, then the potential decreases because $\{v,w\} \in \mathcal{M}_c$. If $v$ does not delegate $w$, it introduces itself. Thus, $w$ adds the implicit edge $(v,w)$ in round $R_{j+1}$. \item If $(v,w) \in E_c^T$ is implicit in some round $R_j$, it becomes explicit in round $R_{j+1}$. This happens because we assume that each message will be delivered within one round. \end{enumerate} By looking at the possible combination of these cases, we can see that the potential reduces at latest in round $R_{i+4}$. Together with the fact that the potential can reduce at most $\mathcal{O}(N^2)$ times, our statement follows. \item Consider a configuration that already contains all valid edges and let the corresponding round be $R_i$. We will now show that the potential defined in the proof of Lemma \ref{lemma:convergenceII} reduces after at most two rounds. Therefore, consider the longest invalid edge $\{v,w\} \not\in MST(V,d_T)$ that is present in first configuration of round $R_i$. Until the end of round $R_i$ all instances of the edge became explicit if they were not already. Thus, in the first configuration of round $R_{i+1}$, the according references are in the local memory of $v$ or $w$ and there are no more implicit instances of the edge. The proof of Lemma \ref{lemma:convergenceII} also suggests that no further implicit instances are ever added. Thus, it remains to show that all explicit instances are removed in $R_{i+1}$. In the following, assume that a reference of $w$ is in $v$'s memory. The other case is analogous. According to Lemma \ref{lemma:equivalence} there is a node $u \in N_v$ with $u \prec (v,w)$. That means, $v$ delegates $w$ to another node once it is activated in round $R_{i+1}$. Thus, at the end of round $R_{i+1}$ all instances of $\{v,w\}$ are gone and the potential must decrease. Together with the fact that the potential can decrease at most $\mathcal{O}(N^2)$ times, our statement follows. \end{enumerate} Thus, after $\mathcal{O}(N^2)$ rounds the system contains all valid and no invalid edges and therefore is in a legal configuration. \end{proof} \section{Conclusion \& Outlook} In this work, we focused on designing and analysing self-stabilizing overlay networks that take into account the underlay. For the tree metric we considered, it turns out that there is an extremely simple protocol for MST construction that naturally follows from some general properties of MSTs in such tree metrics (notice the close relation between Lemma~\ref{lemma:equivalence} and the protocol). Considering different kinds of underlays (such as planar graphs or graphs with bounded growth) as well as other types of overlays than a minimum spanning tree may be possible next steps. Of course, the high upper bound on the running time of our algorithm naturally raises the question whether a better running time can be achieved by a more sophisticated algorithm or a refined analysis. Thus, improving on our results may also be a possible next step.

0c239f9283c3aa8ccc178013ea885ee2592b6f57

\section{Introduction} The chemical evolution of the raw materials inside molecular clouds when new stars are born raises many important and interesting questions. Answers to these questions should help us to understand star formation processes or even the origin of life better. Therefore, much work has been done to address these questions. The chemical evolution is coupled with the physical processes of star formation because chemical reactions are sensitive to physical conditions. On the other hand, chemistry in turn affects line cooling which will impact the gas temperatures and hence the dynamics of molecular cloud evolution. Currently, low-mass star formation is considered to proceed as follows. A prestellar core collapses by self gravity. Initially the collapsing core is almost isothermal, and thus cold ($\sim 10$ K), but the contraction heating eventually overwhelms the radiative cooling in the central region, which becomes adiabatic. Because the temperature and gas pressure increase, the contraction is decelerated so that a hydrostatic core is formed, which is called the first hydrostatic core or the first Larson core \citep{Larson1969}. When the temperature of the central region increases up to $\sim 2000$ K, the hydrogen molecules are dissociated to neutral atoms, which are then ionized to be protons and electrons. These endothermic reactions make the first core unstable. The second collapse ends up with the second hydrostatic core, i.e. the protostar. The dense gas harboring a protostar is called a protostellar core ~\citep{Andre2000}, which continues to infall onto the protostar. Radiation from the protostar and accretion shock heat the protostellar core. So far, physical models of collapsing cores used for astrochemical models simulations vary in degrees of complexity from simple 0-D~\citep{Garrod2006} models to more complicated 1-D radiation hydrodynamic (RHD) models~\citep{Aikawa2008}, to 2-D models~\citep{Weeren2009}, and 3-D RHD models~\citep{Furuya2012} and finally to full 3-D radiation-magnetohydrodynamic models~\citep{Hincelin2013, Yoneda2016}. The results of these studies significantly enhance our knowledge about the molecular evolution that occurs when molecular cores collapse to form stars. However, there are still at least two questions that have not been well answered. COMs, which are defined to be carbon-containing molecules with at least 6 atoms~\citep{Herbst2009}, have been observed toward many astronomical sources including low-mass protostars \citep{Ceccarelli2007}. The first question is how these COMs are formed. The abundances of some terrestrial COMs such as dimethyl ether (CH$_3$OCH$_3$) predicted by theoretical models are often more than one order of magnitude lower than the observed values toward low-mass protostars~\citep{Aikawa2008}. The second question is how the volatile species diffuse in the ice mantle and desorb into the gas phase in warming-up cores. Laboratory experiments show that not all solid CO can sublime at 20 K~\citep{Collings2003, Sandford1988}. However, most astrochemical models used for collapsing cores are the simple two-phase models. In these models, no distinction is made between the active surface layer and the bulk ice, which should be less active, so that all solid CO molecules sublime at temperature around 20 K. It should be noted that the diffusion and desorption of CO are linked to the question in COMs. In cloud cores, CO is the major carbon carrier. In some protostellar cores, however, the abundance of sublimated CO is lower than its canonical value~\citep{Fuente2012}. So far, the interpretation is either that CO is locked in water ice, or a fraction of CO is converted to COMs. There are more advanced astrochemical models used for collapsing cores than the two-phase model. Three-phase models are able to distinguish the active layer from the ice mantle~\citep{Hasegawa1992}; however, the application of three-phase models to the chemical simulation of collapsing cores has only achieved limited success. The three-phase models used by \citet{Furuya2015}, \citet{Taquet2012}, \citet{Vasyunin2013} and \citet{Vasyunin2017} assume completely inert ice mantles, so the volatile species are locked in the bulk ice and cannot desorb until the water ice sublime. The three-phase model in \citet{Garrod2013} introduces the swapping mechanism of species in ice mantles, so that these species are able to diffuse; CO molecules are thereby able to diffuse out of the ice mantles and then sublime at temperatures above 20 K. However, no CO molecules can be trapped for a long period of time inside the ice mantles in the three-phase model with swapping mechanism, which contradicts the laboratory experiments~\citep{Garrod2013}. Improved surface astrochemical models are required to answer the above two questions better. Theoretical models and laboratory experiments show that gas-phase reactions are not efficient enough to produce some terrestrial COMs such as methyl formate~\citep{Horn2004, Geppert2006, Garrod2006} in warm regions although they may be more successful in cold regions~\citep{Balucani2015}. COMs such as methyl formate are believed to be formed on warm grain surfaces by the recombination of radicals, which are products of ice mantle photodissociation reactions, when the temperatures of dust grains are between 20 and 40 K~\citep{Garrod2006}. Thus, more efficient production of COMs must rely on better surface astrochemical models. Progress has been made to solve the problem of COMs formation. In particular, the production of COMs in the three-phase model of \citet{Garrod2013} is much more efficient than that in the two-phase models. However the exponential decay of photodissociation reactions with depth into ice mantles is not considered, thus COM formation via radical-radical recombination may be overestimated. In a four-layer model~\citep{Kalvans2017}, the whole ice mantle is divided into 4 layers and the shielding of species buried in ice mantle by outer layers is introduced. But the four-layer model is only a crude approximation; photodissociation reaction rates in the ice mantles should decay monolayer by monolayer. By modeling the layered structure of ice, we can also distinguish the trapped CO molecules and CO molecules that can diffuse inside ice mantles, which could provide a better answer to the second question. \citet{Fayolle2011} suggested an extended three-phase model in which a mantle-surface diffusion term is introduced and not all mantle species can participate the mantle-surface circulation. Thus, trapped species and species that can diffuse inside ice mantles are distinguished. However, to the best of our knowledge, the extended three-phase model has only been used to fit experiments~\citep{Fayolle2011}. A numerical method to simulate surface chemistry is another issue. It has been found that if the average population of reactive species on dust grains is much less than one, significant error may occur if the rate equation approach is used for numerical simulation~\citep{Biham2001}, a problem known as finite size effect. This problem has been extensively studied in the past 20 years~\citep{Cuppen2013} and a few numerical methods have been suggested to replace the rate equation approach in order to solve it. The modified rate equation method is the most efficient method, but, it also is the least rigorous approach~\citep{Garrod2008}. To the best of our knowledge, only the modified rate equation has been adopted for hydrodynamic-chemical simulations~\citep{Furuya2015} while a more rigorous approach, the macroscopic Monte Carlo (MC) mechanism or Gillespie algorithm~~\citep{Gillespie1976, Vasyunin2009}, has only been adopted to simulate the chemical evolution of simple 0-D collapsing cores~\citep{Vasyunin2013}. Recently, \citet{Chang2014} reported a full gas-grain reaction network simulation with an updated surface astrochemical model, which distinguishes the trapped species and species that can diffuse inside ice mantles. The simulation was performed by the microscopic MC method~\citep{Chang2005}, which follows the motion of every molecule on a grain. The species trapped in ice mantles are called normal species while the species that can diffuse inside ice mantles are known as interstitial species. \citet{Chang2014} found that many radicals are formed and then frozen inside mantles under physical conditions that pertain to cold cores. The radicals that accumulate in the cold core stage should then be able to recombine to form COMs when the temperatures of cores increase. It is particularly interesting to determine if COMs formed by this mechanism are able to explain the observed abundances of COMs toward protostars. Unfortunately, while microscopic MC is more rigorous than the macroscopic MC, its computational cost is very expensive when the temperatures of dust grains are higher than 15 K. Thus, this more rigorous approach can only be used for the chemical simulations of cold cores~\citep{Chang2014}. In this paper, we present a macroscopic MC chemical simulation with a 1-D RHD model. Our major purpose is to go one step further than earlier approaches to explain the COM formation during the collapse of molecular cloud cores. We also pay attention to the CO sublimation problem. We use a surface astrochemical model that is similar to that in \citet{Chang2014} except that we use the macroscopic MC approach in order to reduce the computational cost. One strength of our approach is that as in the microscopic approach, the trapped species are also well distinguished from species that can diffuse inside ice mantles. As in \citet{Chang2014}, abundant radicals can also be formed inside the ice mantle when dust grains are about 10 K in our updated model. The physical model used is from \citet{Masunaga1998} and \citet{Masunaga2000}, which has been used for simple two-phase hydrochemical simulations~\citep{Aikawa2008, Aikawa2012}. The remainder of our paper is organized as follows. The physical model, numerical methods and chemical model are discussed in Section \ref{models} while we present the simulation results of our models in Section \ref{res}. We compare model results with observations in Section \ref{comp}. Discussions and conclusions follow in Section \ref{diss}. \section{Models}\label{models} \subsection{Physical Model} \citet{Masunaga1998} and \citet{Masunaga2000} constructed a 1-D spherical RHD model for low mass star formation. \citet{Aikawa2008} and \citet{Aikawa2012} calculated the molecular evolution in infalling fluid parcels in this 1-D RHD model, which we adopt in the present work. We refer to \citet{Aikawa2008} , \citet{Aikawa2012}, \citet{Masunaga1998} and \citet{Masunaga2000} for details of the physical models. We only briefly explain them below. Initially, the number density of H nuclei at the center of a prestellar core is $\sim6 \times 10^{4}$ cm$^{-3}$. The prestellar core is assumed to be in hydrostatic equilibrium for a period of $10^6$ yr. We assume the temperature of the prestellar core before collapsing is 10 K. Typical lifetime of prestellar cores and triggering of the core contraction are still open questions in star formation studies. Decay of turbulence is one possibility. Then the typical lifetime of the core before the collapse is sound crossing time $\sim~10^6$ yr (a few 0.1 pc divided by $10^4$ cm/s). Alternatively, the timescale is set by the rate of mass accretion from the diffuse warm gas to form the filamentary molecular clouds; recent observational studies found that the star forming cores exist in filaments which have the column density higher than a critical value~\citep{Andre2014}. Theoretical calculations suggest that the timescale of cloud (filament) formation is $10^7$ yr, but it includes low (column) density phase which is much longer than the dense phase~\citep{Inutsuka2015}. Since the original hydrodynamics model we adopted~\citep{Masunaga2000} investigates only the collapse phase, we made a simple assumption that the core was supported against gravitational collapse by turbulence and/or magnetic fields for 1 Myr before the collapse starts. So, at the time $t_1 = 10^6$ yr, the prestellar core starts a contraction to form the first core and then a protostar. The protostar is formed at the time $t_2 = 10^6 + 2.5 \times 10^{5} $ yr while the first core is formed at around time $t_3 = t_2- 5.6\times 10^2$ yr. After the protostar is formed, the physical model continues to evolve for another $9.3 \times 10^{4}$ yr, thus the final time is t$_{\textrm{final}}$ = $ 10^6 + 2.5 \times 10^{5} + 9.3 \times 10^{4}$ yr. The model gives the physical conditions of each falling fluid parcel as a function of time. Among these physical conditions, the temperatures of gas and dust grains, density and visual extinction ($A_v$) are important to study the chemistry. Temperatures of gas and dust are set to be equal for simplicity; the effect of thermal decoupling of dust and gas are discussed in Section \ref{diss}. We studied the chemical evolution of 13 fluid parcels which initially are at $r \sim 4\times 10^4$ AU. Fig. \ref{Fig1} shows the radial distribution of temperatures and H nuclei density at t$_{\textrm{final}}$. The temporal evolution of temperature, visual extinction and H nuclei density of selected fluid parcels are shown in Fig.~\ref{Fig2}. \begin{figure} \centering \resizebox{12cm}{8cm}{\includegraphics{Fig1.eps}} \caption{ Radial distribution of temperature and H nuclei density at t$_{\textrm{final}}$. } \label{Fig1} \end{figure} \begin{figure} \centering \resizebox{8cm}{8cm}{\includegraphics{Fig2_1.eps}} \resizebox{8cm}{8cm}{\includegraphics{Fig2_2.eps}} \resizebox{8cm}{8cm}{\includegraphics{Fig2_3.eps}} \caption{ The temperature, visual extinction and H nuclei density as a function of time for the fluid parcels that reach 2.5, 500 and 8000 AU at t$_{\textrm{final}}$. Physical conditions at $0<t<10^6$ yr are the same as those at $10^6$ yr. } \label{Fig2} \end{figure} \subsection{Chemical Models} The chemical reaction network used in this work is based on the network in \citet{Hincelin2011} with modifications, which will be discussed in Section \ref{network}. We use the standard dust grain size $r_d = 0.1 \mu$m where $r_d$ is the radius of dust grains. There are $N_s=10^6$ binding sites per monolayer on grain surfaces. The gas-to-dust number ratio is kept at $10^{12}$. We use the standard cosmic-ray ionization rate of H$_2$, $\zeta=1.3 \times 10^{-17}$ s$^{-1}$. The surface diffusion barrier $E_b$ is fixed at 50\% of the desorption energy $E_D$. We use the initial low-metal abundances for gas phase species~\citep{Semenov2010}, which are shown in Table \ref{table1}. Moreover, we assume that initially all species are in gas phase. Below, we first discuss our surface and bulk ice models and then move on to the reaction networks. \begin{table} \caption{Initial Gas Phase Abundances} \label{table1} \begin{tabular}{lc} \hline \hline Species & Fractional Abundance w.r.t. $n_{\rm H}$ \\ \hline He & $9.0 \times 10^{-2}$ \\ e$^-$ & $1.2 \times 10^{-4}$ \\ H$_2$ & $5.0 \times 10^{-1}$ \\ C$^{+}$ & $1.2 \times 10^{-4}$\\ N & $7.6 \times 10^{-5}$\\ O & $2.6 \times 10^{-4}$\\ S$^{+}$ & $8.0 \times 10^{-8}$\\ Si$^{+}$ & $8.0 \times 10^{-9}$\\ Na$^{+}$ & $2.0 \times 10^{-9}$\\ Mg$^{+}$ & $7.0 \times 10^{-9}$\\ Fe$^{+}$ & $3.0 \times 10^{-9}$\\ P$^{+}$ & $2.0 \times 10^{-10}$\\ Cl$^{+}$ & $1.0 \times 10^{-9}$\\ \hline \end{tabular} \end{table} \subsubsection{Surface and Bulk Ice Models} We first briefly review the multiphase model introduced by \citet{Vasyunin2013}, because our new model of ice mantle is based on the prior model. The multiphase model allows not only the layer-by-layer growth of ice mantles but also their layer-by-layer desorption. In this model, only the top 4 monolayers are active; i.e., all surface reactions occur only in these 4 monolayers while the bulk ice in the deep layers is completely inert. Species inside bulk ice are frozen, so not able to diffuse to react with other species. FUV photons are not allowed to penetrate into the ice bulk in this model, so, cannot photodissociate the ice bulk species. The multiphase model used a macroscopic MC approach for simulations. We modify this multiphase model so that species inside the bulk ice are subject to photodissociation and the photofragments diffuse inside the bulk ice mantle. Fig. \ref{Fig3} is a schematic diagram of our surface and bulk ice model. We consider photon penetration, which generates photofragments in the ice mantle, and the diffusion of photofragments in the new multiphase model. If photons are not allowed to penetrate into the bulk ice mantle, our model is the same as the multiphase model of \citet{Vasyunin2013} since the mobile species are generated only by photons in the bulk ice in our model. Our model can be viewed as a macroscopic version of the surface and bulk model in \citet{Chang2014}. We refer the reader to \citet{Vasyunin2013} and \citet{Chang2014} for more details of the multiphase model and the microscopic surface and bulk model. Our macroscopic surface and bulk models have the following constraints and rules: \begin{enumerate} \item An ice mantle is made of surface active layers and a bulk of ice underneath the surface layers. The bulk ice is not inert but partly active as described below. We thus call it partially active ice layers. \item There are two types of binding sites, normal sites and interstitial sites, inside the ice mantle~\citep{Chang2014}. The normal binding sites are the potential minima. Species that occupy normal sites are normal species. There are also weaker potential minima, known as subminima, which are called interstitial binding sites, in the partially active ice layers. Interstitial species occupy the interstitial binding sites. All binding sites in the active layers are normal sites while interstitial binding sites only exist in the bulk of ice. There are $N_s$ normal binding sites in each active layer and bulk ice layer. The number of interstitial binding sites in each bulk ice layer is also $N_s$. \item A bulk ice mantle is made of multiple monolayers of normal species and interstitial species that are uniformly distributed. Normal species inside the bulk ice mantle are trapped (i.e. cannot diffuse) while interstitial species are able to diffuse. \item The active layers are made of the topmost four monolayers which can only be occupied by normal species, so the maximum number of normal species in the active layers is $4N_s$. When gas phase species accrete onto grain surfaces, they become normal species in the active layers. The normal species in the active layers can diffuse and react with each other. \item Surface chemical reactions and processes, including accretion, desorption and other reactions in the active layers, are able to change the population of the normal species in the active layers. If the total population of normal species in the active layers is more than $4N_s$, normal species in the active layers are transformed into a normal species in the partially active layer based on the procedure in \citet{Vasyunin2013}. Similarly, if the total population of normal species in the active layers is less than $4N_s$, normal species in the partially active ice mantle are transformed to normal species in the active layers \citep{Vasyunin2013}. Surface photodissociation reactions in the active layers are treated in the same way as that in \citet{Vasyunin2013}. \item We follow how photons penetrate into the partially active ice layers and photodissociate bulk ice species in a manner similar to ~\citet{Chang2014}. The fluxes of the external and the cosmic ray induced far-ultraviolet (FUV) photons are given as $F_{FUV1}=F_0 e^{-\gamma A_v}$ and $F_{FUV2}=G_0^{'} F_0$ respectively, where $F_0=10^8$ cm$^{-2}$ s$^{-1}$ is the standard interstellar radiation field, $G_0^{'}=10^{-4}$ is the scaling factor for cosmic ray induced photons~\citep{Shen2004} and $\gamma \sim 2$ measures the UV extinction relative to visual extinction \citep{Oberg2007,Roberge1991}. The rates of the external and the cosmic ray induced photon bombardment are $\pi r^2 F_{FUV1}$ and $\pi r^2 F_{FUV2}$ respectively. FUV photons penetrate into the bulk ice layer-by-layer. At each monolayer, we randomly select a species, which is either a normal species or an interstitial species and then decide whether this species can be photodissociated or not. However, the probability to photodissociate ice species other than water by photons is poorly known. Therefore, following \citet{Chang2014}, we calculate the probability that the i-th surface species is photodissociated as $P_i = P_0 \frac{\sum k_{FUV, mi}}{\sum k_{FUV, m0}}$ where $k_{FUV, m0}$ and $k_{FUV, mi}$ are the photodissociation rate coefficients of water and the i-th surface species in the m-th product channel respectively. Both summations are over all product channels for each species. The probability to photodissociate a monolayer of water was estimated to be $P_0 =0.007$~\citep{Andersson2008}. We generate a random number $X_1$ that is uniformly distributed within (0, 1). If $X_1 < P_i$, the selected species is photodissociated, otherwise the FUV photon penetrates into the next monolayer. We put the photodissociation products in normal sites or interstitial sites based on rules 2 and 7. The exponential decay of photodissociation rates in ice mantles is mimicked well~\citep{Chang2014}. \item Interstitial species are generated when FUV photons photodissociate normal species in the partially active ice mantle. Photodissociation products of a normal species can include interstitial species, and thus, the normal sites which the photodissociated normal species occupied can become empty. Assuming there are $N_{nh}$ and $N_{ih}$ empty normal and interstitial binding sites respectively in a monolayer, the probability that a photodissociation product occupies a normal binding site is, $p_n = \frac{N_{nh}}{N_{nh}+ \alpha N_{ih}}$, where $\alpha$ is a parameter that differentiates the normal binding sites and interstitial binding sites. The smaller the $\alpha$ value is, the more likely photodissociation products occupy normal sites. We assume that species occupying the normal sites are always no fewer than those in interstitial sites in the ice mantles because the normal binding sites are the potential minima while the interstitial binding sites are the potential subminima. The value of $\alpha$ is therefore $\leq 1$. If $\alpha = 1$, the potential minima is the same as the potential subminima, which is an extreme case. On the other hand, if $\alpha = 0$, photodissociation products are prohibited from occupying interstitial sites. Photodissociation reactions increase the population of species, and therefore, there are not enough normal sites for the extra species to occupy if $\alpha =0$. So, we limit the range of $\alpha$ to be $0< \alpha \leq 1$. However, $\alpha$ is poorly constrained other than the discussions above. We set $\alpha=0.01, 0.5$ in our models to test the influence of its value on simulation results. In the MC simulations, a random number X, which is uniformly distributed within 0 and 1, is generated. If $X< p_n$, the photodissociation product occupies normal sites. Otherwise, it occupies interstitial sites. \item Interstitial species are able to diffuse into the active layers from the partially active bulk of ice. Assuming the bulk diffusion barrier is $E_{b2}$, the rate that an interstitial species hops from one interstitial site to another one is $k_{hop}= \nu \exp(-\frac{E_{b2}}{T})$, where $T$ is the grain surface temperature and $\nu=10^{12}$ s$^{-1}$ is the trial frequency. See below the discussion about the bulk diffusion barriers. Two assumptions are made in order to derive the rate at which interstitial species diffuse into the active layers. First, interstitial species can hop toward 6 different directions~\citep{Chang2014} while only the hopping toward the active layers transforms an interstitial species into a normal species in the active layers. Second, only the interstitial species that are in the monolayer which is next to the active layers can diffuse into the active layers. The rate at which the jth interstitial species diffuse into the active layers is thus, \begin{equation} r_{diff_j} = \frac{k_{hop_j}}{6} \frac{N_j}{N_L}, \label{diffuse} \end{equation} where $N_L$ is the number of monolayers of normal species in the partially active ice mantle, $N_j$ is the population of the jth interstitial species and $k_{hop_j}$ is the hopping rate of the jth interstitial species. In the active layers, the population of normal species increases when interstitial species diffuse into these layers. We follow the rules in \citet{Vasyunin2013} to move one normal species in the active layers to the partially active ice mantle if there are more than $4N_s$ species in the active layers. \item Interstitial species are able to react with other interstitial species or normal species in the partially active ice mantle. The reaction rate between two interstitial species can be derived in a manner similar to that for surface reactions. The rate coefficient of a reaction between two surface species is based on the assumption that two species must on average visit all sites on grain surfaces before they encounter each other~\citep{Hasegawa1992}. We assume that two interstitial species must also on average visit all interstitial sites in the partially active ice mantle before they encounter each other. So, the rate of the reaction between two different interstitial species is, \begin{equation} r_{ij} = \frac{\kappa_{ij}}{N_s N_L} (k_{hop_i} + k_{hop_j})N_i N_j, \label{int_int} \end{equation} where $N_L$ is the number of monolayers in the partially active ice mantle, $N_i$ and $N_j$ are the population of the ith and jth interstitial species respectively, $\kappa_{ij} \leq 1$ is a factor for the reaction activation energy~\citep{Semenov2010,Chang2007,Garrod2011}, and $k_{hop_i}$ and $k_{hop_j}$ are the hopping rates of the ith and jth interstitial species respectively. The factor $\kappa_{ij}$ is not a simple exponential function to overcome the barrier; we consider the tunneling and competition between diffusion and reaction. The rate of the reaction between two identical species i is \begin{equation} r_{ii} = \frac{\kappa_{ii}}{2N_s N_L} (k_{hop_i} + k_{hop_i})N_i^2. \label{int_int_ii} \end{equation} Similarly, because all normal species in the partially active ice mantle are frozen and only interstitial species can diffuse in the partially active ice mantle, the rate of the reaction between an interstitial species i and a normal species n is, \begin{equation} r_{in} = \frac{\kappa_{in}}{N_s N_L} k_{hop_i} N_i N_n, \label{int_nor} \end{equation} where $N_n$ is the population of the nth normal species. When interstitial species react with normal species, the product becomes normal species if there is only one product. Otherwise, the product with the largest diffusion barrier becomes a normal species while all other products become interstitial species. \end{enumerate} We simulate 5 different models as shown in Table~\ref{table2}. Hereafter, our new surface and bulk model are called the new multiphase model. In models MC1, MC2 and MC3, we adopt the new multiphase model. Normal species are frozen in partially active ice mantles while the bulk diffusion barrier $E_{b2}$ for interstitial species has been poorly constrained so far. Following ~\citet{Chang2014}, we set $E_{b2}= 0.7 E_D$ in models MC1 and MC2 while in model MC3, we set $E_{b2}= 1.0 E_D$. The desorption energy of species in the active layers are taken from~\citet{Garrod2006}. The parameter $\alpha$ is even less constrained. In model MC1, $\alpha$ takes a value of 0.01 assuming photodissociation products in the partially active ice mantle can hardly occupy interstitial sites while in models MC2 and MC3, $\alpha$ takes a value of 0.5 so that much more photodissociation products can occupy interstitial sites. In model MC4, we use the basic multiphase model proposed by \citet{Vasyunin2013} while MC5 is a two-phase model. The basic multiphase model is similar to the three phase models proposed by~\citet{Taquet2012} and \citet{Furuya2015} in that the ice mantle consists of chemically active surface layers and inert ice bulk mantle or photofragments are assumed to recombine immediately in the bulk ice mantle. In model MC5, all surface species are in the active layers and there is no inert or partially active ice bulk. The photodissociation rates of surface species may be modified so that the exponential decay of photodissociation rates with depth into ice mantle is included in two-phase models. This more rigorous treatment of surface photodissociation reactions is not considered in most two-phase astrochemical models. Therefore, we do not take into account the exponential decay of photodissociation reaction rates in two-phase models in this paper for simplicity. Models MC4 and MC5 are simulated for the purpose of comparison with models MC1, MC2, MC3. \begin{figure} \centering \resizebox{7cm}{6cm}{\includegraphics{Fig3.eps}} \caption{ Schematic diagram of the surface and bulk model. The yellow area indicates the gas phase. The blue small balls represent the active layer molecules, which are free to diffuse and undergo chemical reactions. The green small balls represent the normal species in the partially active ice mantle. They are frozen in ice mantle, thus cannot diffuse. The red small balls represent the interstitial species. They can diffuse and react with normal or interstitial species. Photons can penetrate into ice mantle and photodissociate species to generate interstitial species. } \label{Fig3} \end{figure} \begin{table} \caption{Parameters for Different Models} \label{table2} \begin{tabular}{cccc} \hline \hline Model & $E_{b2}$/$E_D$ & $\alpha$ & Surface and Bulk Model \\ \hline MC1 & 0.7 & 0.01 & new multiphase \\ MC2 & 0.7 & 0.5 & new multiphase \\ MC3 & 1 & 0.5 & new multiphase\\ MC4 & - & - & basic multiphase\\ MC5 & - & - & two-phase \\ \hline \end{tabular} \end{table} \subsubsection{Reaction Network}\label{network} Radicals CH$_3$O and CH$_2$OH are not distinguished in the original reaction network~\citep{Hincelin2011}. A major modification is to distinguish these two radicals. Photodissociation reactions are the major formation pathways of CH$_2$OH and CH$_3$O~\citep{Chang2016}. In the active layers or the partially active bulk ice mantle, both CH$_3$O and CH$_2$OH can be formed by photodissociation of methanol while CH$_2$OH can also be produced by the hydrogenation of H$_2$CO. Photodissociation of methanol can also produce CH$_3$. The methanol photodissociation branching ratios to produce CH$_3$O, CH$_2$OH and CH$_3$ are 20:20:60~\citep{Garrod2008}. Although the hydrogenation of both surface or bulk CH$_3$O and CH$_2$OH forms methanol, they can react with other radicals to form different COMs as in the following reactions: \begin{equation} \begin{array}{lllll} \rm{CH}_3\rm{O} & + & \rm{HCO} & \rightarrow & \rm{HCOOCH}_3 \\ \rm{CH}_3\rm{O} & + & \rm{CH}_3 & \rightarrow & \rm{CH}_3\rm{OCH}_3 \\ \rm{CH}_2\rm{OH} & + & \rm{HCO} & \rightarrow & \rm{CH}_2\rm{OHCHO} \\ \rm{CH}_2\rm{OH} & + & \rm{CH}_3 & \rightarrow & \rm{C}_2\rm{H}_5\rm{OH}. \\ \end{array} \end{equation} Both glycolaldehyde (CH$_2$OHCHO) and methoxy (CH$_3$O) are new species in the reaction network. Relevant glycolaldehyde reactions are included in the reaction network~\citep{Garrod2015}. We assume that gas phase CH$_3$O can participate in all reactions in which gas phase CH$_2$OH participates. The rate coefficients of the reactions between gas phase CH$_3$O or CH$_2$OH and other species are assumed to be the same. The binding energies of methoxy and glycolaldehyde are not available in~\citet{Garrod2006}, but are set to be 5084 K and 6295 K respectively because we assume the binding energies of isomers are the same in our models. Table~\ref{table22} shows the desorption energies, $E_D$ of selected surface species. The encounter desorption mechanism is used in this work~\citep{Hincelin2015}. We include the reaction gH$_2$ + gH$_2$ $\rightarrow$ gH$_2$ + H$_2$ in the active layers where the letter g designate species in these layers. As the density of molecular clouds increases, too many gH$_2$ molecules are frozen on grain surfaces without the encounter desorption mechanism, which is not physical. The encounter desorption mechanism takes into account the fact that the desorption energy of gH$_2$ on another gH$_2$ molecule is much smaller than that of gH$_2$ on water ice, which significantly increases the desorption of gH$_2$. \begin{table} \caption{The Desorption Energies of Selected Species} \label{table22} \begin{tabular}{cc} \hline \hline Species & $E_D$(K) \\ \hline H & 450 \\ O & 800\\ CO & 1150 \\ CO$_2$ & 2575 \\ N$_2$ & 1000 \\ HCO & 1600 \\ H$_2$CO & 2050 \\ CH$_3$ & 1175 \\ CH$_2$OH & 5084 \\ CH$_3$O & 5084 \\ HCOOCH$_3$ & 6295 \\ CH$_2$OHCHO & 6295 \\ \hline \end{tabular} \tablecomments{Taken from~\citet{Garrod2006}. The desorption energies of isomers are assumed to be the same.} \end{table} Finally, the CO hydrogenation reactions on grain surface to form methanol have been well studied in laboratory experiments. The microscopic MC method has been used to fit the activation barriers for the reactions CO + H $\rightarrow$ HCO and H$_2$CO + H $\rightarrow$ H$_3$CO~\citep{Fuchs2009}, in which the competition mechanism~\citep{Chang2007,Chang2012,Garrod2011} can be naturally included. The competition mechanism and the reaction barriers fitted by \citet{Fuchs2009} are used to calculate the rates for surface reactions H$_2$CO + H and CO + H in this work. However, the reaction rates are only fitted at fixed temperatures, T = 12 K, 13.5 K, 15.0 K and 16.5 K~\citep{Fuchs2009} respectively. In order to perform astrochemical simulations, we assume that the reaction rates are constant at temperature ranges, 10.0 K $\leq$ T $\leq$ 12 K, 12 K $<$ T $\leq$ 13.5 K, 13.5 K $<$ T $\leq$ 15.0 K and T $\geq$ 15 K. For T $\leq$10 K, the factor $\kappa_{ij}$ is assumed to be constant. For all other reactions with barriers on grain surfaces, the rate coefficients are calculated based on tunneling as in the original reaction network~\citep{Hincelin2011}. In total, there are 663 species and 6370 chemical reactions in our chemical models ; among them, 464 species and 4667 reactions are in the gas phase. \subsection{Numerical Methods} An accelerated Gillespie algorithm (QSSA1), which is based on the regular Gillespie algorithm, is used in this work~\citep{Chang2017}. We first briefly introduce the regular Gillespie algorithm, and then the accelerated Gillespie algorithm method. In order to perform simulations with the Gillespie algorithm, we isolate a cell of gas containing one dust grain~\citep{Vasyunin2009}. There are $10^{12}$ H nuclei in the cell of gas because we assume the gas-to-dust number ratio is $10^{12}$. The initial population of species a is $fr_a 10^{12}$ where $fr_a$ is the initial fractional abundance of species a. The reaction rates of all chemical reactions in gas phase and on dust grains must be calculated. For the m-th bimolecular gas-phase or surface reaction with two different species a and b, the reaction rate is, $r_m = k_m N_a N_b$, where $k_m$ is the rate coefficient of the m-th reaction, $N_a$ is the population of species a and $N_b$ is the population of species b. If the reactants in the n-th bimolecular reaction are the identical species a, the reaction rate is $r_n = 0.5 k_n N_a(N_a-1)$. For the m-th unimolecular reaction with reactant a, the reaction rate is $r_m = k_m N_a $. The numerical implementation of the regular Gillespie algorithm is the following. First, the total reaction rate of all reactions is calculated by $r_{\textrm{total}} = \sum_m r_m$ where m is for all reactions in the chemical system. The time step is calculated by $\Delta t = -ln(X)/r_{total}$ where $X$ is a random variable uniformly distributed in the range between 0 and 1. Assuming the current time is $t_0$, a reaction will fire at the time $t_{\textrm{next}} = t_0 + \Delta t$. The n-th reaction, which fires at $t_{\textrm{next}}$, satisfies the equation $\sum_{m=0}^{m=n-1} r_m < r_{\textrm{total}}Y \leq \sum_{m=0}^{m=n} r_m$ where $Y$ is another random variable uniformly distributed between 0 and 1. The populations of all reactants involved in the n-th reaction are updated at the time $t_{\textrm{next}}$. Immediately after the n-th reaction fires, the reaction rates of all reactions in which the population of reactants changes are also updated at the time $t_{\textrm{next}}$. The processes are repeated until $t_{\textrm{next}}$ reaches the final time. We used the accelerated Gillespie algorithm instead of the regular Gillespie algorithm to perform simulations in this work because H$_2$ accretion and desorption consume significant amounts of CPU time if the regular Gillespie algorithm is used. The accelerated Gillespie algorithm is based on quasi-steady-state assumption~\citep{Rao2003}. The implementation of this Gillespie algorithm is similar to that of the regular Gillespie algorithm. The difference is that the conditional expectation of reaction rates are used to calculate the time step $\Delta t$ and the selection of the n-th reaction which fires at $t_{\textrm{next}}$. We treat surface H$_2$ as a transient species in the accelerated Gillespie algorithm~\citep{Chang2017}. The conditional reaction rates of all reactions that do not involve surface and gas phase H$_2$ are the same as the reaction rates in the regular Gillespie algorithm. However, the conditional rates of reactions that involve gas phase and surface H$_2$ are calculated using a different algebraic expression, in which the accretion of gas phase H$_2$, thermal desorption and encounter desorption of surface H$_2$ do not explicitly appear in the gas-grain reaction network, so that the simulations are significantly accelerated. We refer to \citet{Chang2017} for details of the accelerated Gillespie algorithm. Finally, when the temperature of the fluid parcel exceeds 200 K, the molecular evolution of fluid parcels is dominated by gas phase chemistry. We thus stop the accretion of all species in order to further increase the simulation efficiency. We use the Taurus High Performance Computing system of Xinjiang Astronomical Observatory for the simulations in this work. The computational cost varies depending on the fluid parcel and the chemical model. It takes no more than 3 days to simulate the molecular evolution of each fluid parcel for each chemical model. \section{Results}\label{res} \subsection{Evolution of Granular and Gas-Phase Species in Fluid Parcels} Fig.~\ref{Fig4} shows the temporal variation of selected radicals JCH$_3$O, JCH$_3$, JCH$_2$OH and JOH before the contraction of the molecular cloud in the fluid parcel that reaches r= 2.5 AU at the final time in models MC1, MC2, MC3, MC4 and MC5. The letter J designates granular species, which are the species on dust grains. In models MC1, MC2 and MC3, granular species Ji includes species gi, interstitial species i (Ii) and species i locked in the normal sites (Ki). In model MC4, granular species Ji includes species gi and species i in the inert ice bulk while in model MC5 granular species Ji is equivalent to species gi. We can see that the abundances of these radicals are almost the same before around $7\times 10^3$ yr in all models. This is because there are fewer than 4 monolayers of granular species on dust grains before about this time, so all granular molecules in models MC1, MC2, MC3 and MC4 are in the active layers, so that surface chemistry in models MC1, MC2 and MC3 and MC4 is the same as that in MC5. After $7\times 10^3$ yr, the differences of granular radical abundances among models increase dramatically. In models MC1, MC2 and MC3, radicals can be generated by photodissociation reactions or recombination reactions and then buried in the partially active ice bulk. Interstitial species can also diffuse to enter the active layers. Among them, IH atoms have the highest diffusion rate. Radicals that diffuse much slower than IH or these that are frozen in the normal sites tend to accumulate in the ice mantle. On the other hand, some recombination reactions in the bulk of ice can also produce radicals. For instance, IH can recombine with IO or KO to form IOH or KOH. So, granular radical abundances increase quickly after $7\times 10^3$ yr. The rapid increase of granular radical abundances in models MC1, MC2 and MC3 is similar to that in the microscopic Monte Carlo models~\citep{Chang2014}. The differences in granular radical abundances among the new multiphase models MC1, MC2 and MC3 are not significant, because radicals are mainly formed via photodissociation of ice molecules. Photofragments of major icy species such as JO, JOH and JCH$_3$ are the dominant granular radicals. The population of all radicals account for 20-30\% of the total population of granular species in models MC1-3. In the two-phase model MC5, on the other hand, radicals produced on grain surfaces quickly react with newly accreted species such as JH; thus, granular radicals cannot accumulate in model MC5. At the time of $10^6$ yr, the abundances of granular radicals in models MC1, MC2 and MC3 are more than four orders of magnitude higher than those in model MC5. The abundances of granular radicals in the basic multiphase model MC4 are typically higher than those in model MC5 but much lower than those in our multiphase models. At $10^6$ yr, the abundance of JCH$_3$O in model MC4 is more than five orders of magnitude lower than those in our new multiphase models. The abundances of JCH$_3$, JCH$_2$OH and JOH in model MC4 are more than a factor of 4 higher than those in model MC5 but a few orders of magnitude less than that in our new multiphase models. The larger granular radical abundances in model MC4 than model MC5 occurs because surface radicals can be buried in the inert ice mantle if these radicals do not react with other species in the active layers. Fig.~\ref{Fig5} shows the temporal evolution of selected gas-phase species before the start of contraction in the fluid parcel that reaches r= 2.5 AU at $t_{\textrm{final}}$. The visual extinction of the fluid parcel is around 4 before the start of contraction, thus the FUV radiation is intense enough for the photodesorption to play important roles in the production of gas-phase H$_2$CO and CH$_3$OH in our models. On the other hand, thermal desorption is not important before the start of contraction because the temperature is fixed at 10 K. We can see that the abundances of gaseous H$_2$CO and CH$_3$OH predicted by the basic and new multiphase models are similar because photons can only desorb species in the active layers, and because the population of gH$_2$CO and gCH$_3$OH are similar in these models. Before $2\times 10^4$ yr, the abundances of gaseous H$_2$CO and CH$_3$OH predicted by the two-phase model are also similar to those by models MC1-4 because all granular species are in the active layers in models MC1-4. After $2\times 10^4$ yr, the two-phase model predicts higher H$_2$CO abundance because all JH$_2$CO can be photodesorbed in the two-phase model while only gH$_2$CO can be desorbed in models MC1-4. The two-phase model also predicts higher CH$_3$OH abundance than other models do between $2\times 10^4$ and $5 \times 10^5$ yr because the abundance of JCH$_3$OH in the two-phase model is higher than gCH$_3$OH abundances in models MC1-4. However, after $5 \times 10^5$ yr all models predict similar CH$_3$OH abundances because the abundance of JCH$_3$OH in the two-phase model and gCH$_3$OH abundances in models MC1-4 are similar. Fig.~\ref{Fig6} shows the temporal evolution from the start of contraction of abundances for the above selected granular radicals in the fluid parcel that reaches r=2.5 AU at $t_{\textrm{final}}$ in all models. The temporal evolution of the temperature and density for the fluid parcel is also plotted as well. During the contraction, the temperature of the fluid parcel first slightly decreases and then quickly increases after around $1.26\times 10^6$ yr. We can see that granular radical abundances as a function of time vary among models. In the two-phase model MC5, granular radicals quickly disappear after $1.26\times 10^6$ yr because these granular radicals are able to diffuse to react with other radicals or desorb into gas phase. Almost all granular radicals in model MC4 are buried in the completely inert ice mantle before most water molecules desorb at $\sim 1.34\times 10^6$ yr, so the granular radical abundances do not change much before this time. After most water molecules desorb, the ice mantle no longer exists and granular radicals quickly disappear in model MC4. The temporal evolution of granular radicals in the new multiphase models MC1, MC2 and MC3 is more complicated. Before $1.26\times 10^6$ yr, when the temperature of the fluid parcel is $\lesssim$ 10 K, radicals can still be produced in the partially active ice mantle, so the abundances of granular radicals slightly increase. When the temperatures quickly increase after $ 1.26\times 10^6$ yr, granular radicals are able to diffuse inside the partially active ice mantle and more interstitial species can react with other species or diffuse into the active layers. The recombination of granular radicals obviously can reduce the abundances of these radicals, but may also produce other granular radicals. We can see that the JCH$_2$OH abundance first increases after $1.26\times 10^6$ yr and then remains almost constant because JCH$_2$OH can be produced by the recombination of species in the ice mantle by reactions such as JCH$_3$ + JO $\rightarrow$ JCH$_2$OH, which is barrierless in our reaction network. The production rate exceeds the destruction rate for JCH$_2$OH inside the ice mantle, so the abundance of JCH$_2$OH initially increases with temperature in models MC1, MC2 and MC3. However, after $1.34\times 10^6$ yr, when most water molecules desorb from grain surfaces, JCH$_2$OH quickly disappears in our new multiphase models as in the basic multiphase model MC5. The production rates of JCH$_3$O, JCH$_3$ and JOH are less than the loss rates of these radicals inside ice mantles, so their abundances decrease after $1.26\times 10^6$ yr in our new multiphase models. Significant amounts of COMs are produced by granular radicals recombination inside the partially active ice mantle in models MC1, MC2 and MC3 as shown in Fig.~\ref{Fig7}. As in Figure \ref{Fig6}, the temporal evolution of the temperature and density for the fluid parcel is also plotted in Fig.~\ref{Fig7}. We can see that the abundance of JHCOOCH$_3$ in models MC2 and MC3 is almost two orders of magnitude higher than that in model MC5 around the time of $1.33\times 10^6$ yr. Methyl formate molecules are formed by the recombination of the two radicals JHCO and JCH$_3$O. In the two-phase model MC5, radical abundances are much lower than in our new multiphase models, so that the production rate of COMs is much lower than that in models MC2 and MC3. We can also see that the efficient production of JHCOOCH$_3$ occurs earlier in model MC2 than in model MC3 because the bulk diffusion barriers in model MC2 are lower, so interstitial radicals start to diffuse at lower temperatures before the temperatures of the fluid parcel further increases in model MC2. On the other hand, we can see that the $\alpha$ value can also significantly influence the production of granular COMs. The abundance of JHCOOCH$_3$ in model MC1 is about one order of magnitude lower than in models MC2 and MC3. Because $\alpha$ is smaller in model MC1, radicals produced by photodissociation reactions are less likely to occupy interstitial sites than in models MC1 and MC2. Since only interstitial species can diffuse to react with other species in our new multiphase models, the production of COMs in model MC1 is less efficient. The binding sites where granular COMs are formed also varies among models. Because the products of the reactions by two interstitial radicals still occupy interstitial sites in our new multiphase models, there are more interstitial COMs in models MC2 and MC3 than in model MC1. The ratio of the population of interstitial JHCOOCH$_3$ to the population of normal JHCOOCH$_3$ is about 1:4, 2:1 and 2:1 in models MC1, MC2 and MC3 respectively at $1.34\times 10^6$ yr, when the temperature is about 44 K. Formation of JHCOOCH$_3$ in model MC4 is less efficient than that in other models, because only radicals in the active layers can diffuse to recombine and form JHCOOCH$_3$ while buried radicals in the inert ice mantle cannot diffuse. Similarly, we can explain the difference in the production of JCH$_3$OCH$_3$, JC$_2$H$_5$OH and JCH$_2$OHCHO inside the ice bulk mantle among the models. Fig.~\ref{Fig8} shows the temporal evolution of major granular species, temperature and density from the start of contraction to $t_{\textrm{final}}$ in the fluid particle that reaches r=2.5 AU in all models. Significant amounts of products formed by photodissociation of water molecules cannot recombine to form water again in models MC1, MC2 and MC3, which agrees with results calculated by more rigorous microscopic models~\citep{Chang2014} and molecular dynamic simulations~\citep{Andersson2008}. In the two-phase model MC5, products of the water photodissociation can easily recombine to form water again. In model MC4, because photons cannot penetrate into the ice mantle, water inside the ice mantle cannot be photodissociated. Consequently, the abundances of JH$_2$O in models MC4 and MC5 are higher than in models MC1, MC2 and MC3. Unlike water, the photodissociation products of granular methanol cannot easily recombine to form methanol again on dust grains in model MC5, so the methanol ice abundance in model MC5 is the lowest. The abundance of granular methanol in model MC4 is the highest because methanol in the ice mantle cannot be photodissociated in this model. In the new multiphase models, the exponential decay of photodissociation reactions in the ice mantle is well reproduced, and therefore granular methanol abundances in models MC1, MC2 and MC3 are higher than in model MC5. Granular carbon dioxide is mainly produced by surface reactions: JO + JHCO $\rightarrow$ JCO$_2$ + JH and JCO + JOH $\rightarrow$ JCO$_2$ + JH. At around 10 K, because neither JCO nor JOH can diffuse quickly on grain surfaces, JO + JHCO $\rightarrow$ JCO$_2$ + JH is the dominant reaction to produce JCO$_2$. The photodissociation of JH$_2$CO or methanol contributes significantly to the production of JHCO. As with methanol, photodissociation of JH$_2$CO occurs more frequently in two-phase models than in the new and basic multiphase models. Therefore, the abundance of JCO$_2$ in model MC5 is higher than the other models. Fig.~\ref{Fig8} also shows that the abundances of JCO initially decrease as temperatures of the fluid parcel increase and then remain almost constant until most water molecules sublime in models MC1, MC2 and MC3. The abundance of JCO initially decreases because ICO diffuse into the active layers and then sublime. After all ICO and gCO molecules sublime, significant amounts of KCO are still locked in the normal sites in models MC1, MC2 and MC3, thus the abundances of JCO remain almost constant until most water molecules sublime. We calculate $p=N_{JCO}^{30K}/N_{JCO}^{20K}$ where $N_{JCO}^{30K}$ and $N_{JCO}^{20K}$ represent the population of JCO when the temperatures of the fluid parcel are 30 K and 20 K respectively. We find $p=0.87,~ 0.60,~ 0.59$ and $0.90$ in models MC1, MC2, MC3 and MC4 respectively while almost all JCO molecules sublime in model MC5 at 30 K. Similarly, ICO$_2$ and gCO$_2$ molecules sublime at lower temperatures and then KCO$_2$ sublime when most water molecules sublime. However, the time interval between the sublimation of ICO$_2$ and KCO$_2$ is only about $10^3$ yr because the temperature of the fluid parcel increases quickly after ICO$_2$ and gCO$_2$ start to sublime. Thus, JCO$_2$ molecules disappear almost at the same time in all models. \begin{figure} \centering \resizebox{12cm}{12cm}{\includegraphics{Fig4.eps}} \caption{ Selected granular radical abundances as a function of time in the fluid parcel that reaches r=2.5 AU at $t_{\textrm{final}}$ before the contraction of the molecular cloud in models MC1, MC2, MC3, MC4 and MC5. The temperature is fixed at 10 K. } \label{Fig4} \end{figure} \begin{figure} \centering \resizebox{12cm}{12cm}{\includegraphics{Fig5.eps}} \caption{ Selected gas-phase species abundances as a function of time in the fluid parcel that reaches r=2.5 AU at $t_{\textrm{final}}$ before the contraction of the molecular cloud in models MC1, MC2, MC3, MC4 and MC5. The temperature is fixed at 10 K. } \label{Fig5} \end{figure} \begin{figure} \centering \resizebox{12cm}{12cm}{\includegraphics{Fig6.eps}} \caption{ The temporal variation of selected granular radical abundances after the molecular clouds start to contract in the fluid parcel that reaches r=2.5 AU at $t_{\textrm{final}}$ in models MC1, MC2, MC3, MC4 and MC5. The temporal evolution of the temperature and density for the fluid parcel is also plotted. } \label{Fig6} \end{figure} \begin{figure} \centering \resizebox{12cm}{12cm}{\includegraphics{Fig7.eps}} \caption{ The temporal variation of selected COM abundances in the fluid parcel that reaches r=2.5 AU at $t_{\textrm{final}}$ after the temperatures of the fluid parcel starts to increase in models MC1, MC2, MC3, MC4 and MC5. The temporal evolution of the temperature and density for the fluid parcel is also plotted. } \label{Fig7} \end{figure} \begin{figure} \centering \resizebox{12cm}{12cm}{\includegraphics{Fig8.eps}} \caption{ The temporal variation of major granular species, JH$_2$O, JCO$_2$, JCO and JCH$_3$OH in the fluid parcel that reaches r=2.5 AU at $t_{\textrm{final}}$ after the molecular clouds start to contract in models MC1, MC2, MC3, MC4 and MC5. The temporal evolution of the temperature and density for the fluid parcel is also plotted. } \label{Fig8} \end{figure} \subsection{Distribution of Molecules} Fig. \ref{Fig9} and Fig. \ref{Fig10} show, respectively, the radial distributions of water, methanol, CO, H$_2$CO, CO$_2$ and N$_2$ on grain and in gas at $t_{\textrm{final}}$. The molecular evolution of 13 fluid parcels was calculated for each model in order to derive the radial distributions. Fig. \ref{Fig9} shows that abundant water ice exists in the outer regions ($r>125$ AU) where the temperatures are below the sublimation temperature of water ice regardless of models. The distribution of CO ice varies among model. In model MC5, most CO ice sublimes inside the radius of 3000 AU. However, as discussed earlier, CO ice sublime via two processes in the basic and new multiphase models; significant amount of CO ice still exists between 3000 AU and 125 AU because CO molecules locked in the normal sites cannot desorb until the water ice mantle sublimes. Interstitial CO molecules are able to desorb when the temperature is higher than $\sim~20$ K so that JCO abundance drops slightly between 125 AU and 3000 AU in models MC2 and MC3. The distribution of granular CO$_2$, H$_2$CO, and N$_2$ is similar to that of CO ice in each model. The desorption energy of methanol is close to that of water. Thus, methanol ice and water ice disappear in almost the same inner regions. The abundances of gas phase water, methanol, CO, CO$_2$ and N$_2$ typically increase in all models in regions where these species sublime. The radial distribution of gas phase H$_2$CO clearly shows that JH$_2$CO sublime via two processes. The gas phase H$_2$CO abundance increases inward in two steps in models MC1, MC2, MC3 and MC4 because gH$_2$CO and IH$_2$CO molecules desorb first followed by the desorption of KH$_2$CO molecules. We can also see that the gas-phase H$_2$CO abundance typically drops in inner regions with $r<20$ AU because of the gas phase destruction of H$_2$CO. Moreover, gas-phase H$_2$CO abundance drops inward more steeply in models MC1, MC2, MC3 and MC4 than in model MC5. Gas-phase H$_2$CO can be destructed by the reaction with CH$_5$O$^{+}$, which can be produced by the reactions between gas-phase methanol and other ions. The abundance of methanol ice in models MC1, MC2, MC3 and MC4 are much higher than in model MC5, so more H$_2$CO is destructed in regions where the temperatures are high enough to desorb methanol ice. \begin{figure} \centering \resizebox{15cm}{15cm}{\includegraphics{Fig9.eps}} \caption{ Radial distributions of selected granular species in the protostellar core in different models at t$_{\textrm{final}}$. } \label{Fig9} \end{figure} \begin{figure} \centering \resizebox{15cm}{15cm}{\includegraphics{Fig10.eps}} \caption{ Radial distributions of selected gas-phase species in the protostellar core in different models at t$_{\textrm{final}}$. } \label{Fig10} \end{figure} Fig.~\ref{Fig11} shows the radial distribution of gas phase N$_2$H$^{+}$ and HCO$^{+}$ abundances at $t_{\textrm{final}}$. The reaction H$_3^{+}$ + N$_2$ $\rightarrow$ N$_2$H$^{+}$ + H$_2$ is the major formation route of N$_2$H$^{+}$ while its destruction route is N$_2$H$^{+}$ + CO $\rightarrow$ HCO$^{+}$ + N$_2$. So CO sublimation can decrease the abundance of N$_2$H$^{+}$ while N$_2$ sublimation increases its abundance. We can see that the abundance of N$_2$H$^{+}$ drops inward in regions where both JN$_2$ ice and JCO ice sublime from grain surfaces in model MC5. Because a significant amount of KCO molecules are locked in the ice mantles in models MC1, MC2, MC3 and MC4, N$_2$H$^{+}$ abundance drops inward much less steeply. The main destruction pathways for HCO$^{+}$ are its reactions with HCN or H$_2$CO after these two molecules sublime. Since significant amounts of KH$_2$CO and KHCN are locked in the ice bulk in models MC1, MC2, MC3 and MC4, HCO$^{+}$ abundance also drops much less steeply in these four models than in model MC5. \begin{figure} \centering \resizebox{10cm}{12cm}{\includegraphics{Fig11.eps}} \caption{ Radial distributions of simple gas-species in the protostellar core of different models at t$_{final}$. } \label{Fig11} \end{figure} Fig.~\ref{Fig12} shows the radial distributions of selected granular COM abundances at $t_{\textrm{final}}$. If the desorption energy of a COM is larger than that of water ice, the COM can sublime only in regions where the temperature is high enough to desorb the COM. However, if the desorption energy of the COM is smaller than that of water ice, the COM sublime in two different regions as discussed earlier. The COM abundances vary in outer regions in different models. The CH$_3$OCH$_3$ and C$_2$H$_5$OH ices extend to 8000 AU because JCH$_3$ can diffuse in ice mantles to form these species at around 8000 AU where the temperature is sightly above 16 K in models MC1 and MC2. Because the bulk diffusion barriers of interstitial radicals in model MC3 are larger than those in model MC1 and MC2, JCH$_3$OCH$_3$ and JC$_2$H$_5$OH are formed in warmer regions (~3000 AU) in model MC3. Little JCH$_3$OCH$_3$ and JC$_2$H$_5$OH are produced in models MC4 and MC5, so JCH$_3$OCH$_3$ and JC$_2$H$_5$OH only exist in a narrow range. Similarly, we can explain the distributions of HCOOCH$_3$ ice and CH$_2$OHCHO ice. Because CH$_3$CN ice is mainly produced by surface hydrogenation reactions at around 10 K in the basic and new multiphase models, CH$_3$CN ice extends to 8000 AU in models MC1, MC2, MC3 and MC4. However, in the two-phase model MC5, the CH$_3$CN ice abundance at r=8000 AU is much lower than that in the other models because of the absence of the exponential decay with depth of photodissociation for CH$_3$CN ice. The peak of the CH$_3$CN ice abundance occurs around 250 AU where Av is large enough to attenuate FUV radiation so that only a small amount of CH$_3$CN ice can be photodissociated while CH$_3$CN ice can also be efficiently formed by the recombination of JCN and JCH$_3$~\citep{Garrod2008}. A small amount of HCOOH ice also exists around 8000 AU because HCOOH can be produced in the gas phase and deposited on dust grains. The radial distributions of selected gas phase COMs abundances at $t_{\textrm{final}}$ are shown in Fig.~\ref{Fig13}. The abundances of the COMs significantly increase in the inner regions where the ice mantle disappears for two reasons. First, the sublimation of COMs that are mainly formed on dust grains can increase their abundances in gas phase. Second, ice mantle species other than COMs also produce COMs in the gas phase when they sublime. \begin{figure} \centering \resizebox{12cm}{12cm}{\includegraphics{Fig12.eps}} \caption{ Radial distributions of selected granular COM abundances at $t_{\textrm{final}}$. } \label{Fig12} \end{figure} \begin{figure} \centering \resizebox{12cm}{12cm}{\includegraphics{Fig13.eps}} \caption{ Radial distributions of selected gas phase COM abundances at $t_{\textrm{final}}$. } \label{Fig13} \end{figure} Finally, Fig.~\ref{Fig14} shows the radial distributions of the fraction of empty normal sites (F$_{\textrm{empty}}$) and the fraction of interstitial species (F$_{\textrm{inter}}$) predicted by the new multiphase models at t = t2 + $9.279 \times 10^{4}$ yr when the temperature of the inner most fluid parcel is 100 K. The fraction of empty normal sites is defined as, $F_{\textrm{empty}} = {\textrm{Total}}_{\textrm{empty}}/(N_L N_s) $ where ${\textrm{Total}}_{\textrm{empty}}$ is the population of empty normal sites in the bulk of ice. Similarly, $F_{\textrm{inter}} = {\textrm{Total}}_{\textrm{inter}}/N_L N_s $, where ${\textrm{Total}}_{\textrm{inter}}$ is the population of all interstitial species in the bulk of ice. Because more photodissociation fragments occupy interstitial binding sites in models with larger $\alpha$ value, both F$_{\textrm{empty}}$ and F$_{\textrm{inter}}$ are larger in models MC2 and MC3 than in model MC1. On the other hand, because more interstitial species can diffuse into the active layers or participate reactions inside the bulk of ice as the temperature increases, F$_{\textrm{inter}}$ drops quickly inward in all models. On the other hand, F$_{\textrm{empty}}$ and F$_{\textrm{inter}}$ typically are slightly larger in outer regions in all models because of stronger radiation in these regions. \begin{figure} \centering \resizebox{12cm}{12cm}{\includegraphics{Fig14.eps}} \caption{ Radial distributions of F$_{\textrm{empty}}$ and F$_{\textrm{inter}}$ predicted by the new multiphase models at t= t2 + $9.279 \times 10^{4}$ yr. MC1empty, MC2empty and MC3empty represent F$_{\textrm{empty}}$ in models MC1, MC2 and MC3 respectively while MC1inter, MC2inter and MC3inter represent F$_{\textrm{inter}}$ in models MC1, MC2 and MC3 respectively. } \label{Fig14} \end{figure} \section{Comparison With Observations and Previous Models}\label{comp} We compare our theoretical results with observations and the rate equation two-phase (RETP) model~\citep{Aikawa2008} results in this section. Most COMs were found within a few arcseconds from the center of the core~\citep{Kuan2004}. On the other hand, most gaseous COM abundances do not vary much at $r\leq 100$ AU in our model results. Therefore, as in \citet{Aikawa2008}, the abundances of gaseous COMs in the fluid parcel that reaches 30.6 AU at $t_{\textrm{final}}$ are compared with the gaseous observations. The temperature and density of hydrogen nuclei of the fluid parcel are about 198 K and 1.49$\times 10^8$ cm$^{-3}$, respectively at $t_{\textrm{final}}$. The granular species exist at outer regions where water ice is abundant ($r>100$ AU). Therefore, we calculate the average abundances of granular species relative to water ice in outer 7 fluid parcels in which water ice cannot sublime at $t_{\textrm{final}}$ to compare with observations of ices. The average abundance of granular species Ji relative to water ice, $\bar{N}(Ji)$, is calculated as $\bar{N}(Ji) = \frac{\sum_{m=1}^{m=7} N(Ji)_m}{\sum_{m=1}^{m=7} N(JH_2O)_m}$, where $N(Ji)_m$ and $N(JH_2O)_m$ are the population of species Ji and JH$_2$O in m-th fluid parcel respectively. The first fluid parcel is the outermost one and m increases as the distance from the fluid parcel to the center of the protostar decreases. \citet{Bergner2017} recently observed 16 low-mass protostars using the IRAM 30 m telescope, and reported the median abundances of a few COMs. \citet{Bergner2017} did not observe the abundance of glycolaldehyde, but its abundance toward low mass protostar IRAS 16293-2422 was recently reported by\citet{Jorgensen2016}. Table~\ref{table33} summarizes the median COM abundances with respect to methanol derived by the observations and our models. Overall, the new multiphase models agree with observations better than the basic multiphase model and two phase models. Model MC1 is the only model that can reproduce the observed abundances of all COMs within one order of magnitude. All COM abundances other than that of glycolaldehyde are well predicted by models MC2 and MC3 while models MC4 severely underestimate the abundances of COMs other than CH$_3$CN. Models MC5 and RETP are both simple two-phase models, but COM abundances predicted by model RETP can be more than one order of magnitude higher than that by model MC5. One possible reason is the different reaction networks used in simulations. The discrepancy might also be explained by the finite size effect~\citep{Vasyunin2009}. Comparing with observations, both models severely underestimate the abundances of HCOOCH$_3$ and CH$_3$OCH$_3$. \begin{table} \caption{Comparison of Observational COM abundances with Theoretical Results} \label{table33} \begin{tabular}{cccccc} \hline \hline Model & $^b$HCOOCH$_3$ & $^b$CH$_3$OCH$_3$ & $^b$CH$_3$CN & $^b$CH$_3$CHO & $^c$CH$_2$OHCHO\\ \hline Observations & 3.1(-2) & 5.3(-2) & 4.8(-3) & 4.4(-2) & 3.4(-3) \\ MC1 & 3.1(-3) & 1.8(-2) & 2.0(-3) & 1.3(-2) & 5.4(-3) \\ MC2 & 2.0(-2) & 1.2(-2) & 6.5(-3) & 1.7(-2) & {\bf 5.0(-2)} \\ MC3 & 2.2(-2) & 9.1(-3) & 5.5(-3) & 1.8(-2) & {\bf 6.1(-2)} \\ MC4 & {\bf 6.5(-5)} & {\bf 2.2(-5)} & 3.0(-3) & {\bf 6.7(-7)} & {\bf 2.4(-5)} \\ MC5 & {\bf 2.0(-3)} & {\bf 3.0(-6)} & {\bf 3.1(-4)} & {\bf 1.5(-4)} & 2.0(-3)\\ RETP$^{a}$ & {\bf 6.0(-4)} & {\bf 1.2(-4)} & 1.0(-2) & & \\ \hline \end{tabular} \medskip{\protect\\ Notes.\protect\\ ${a}$ \citet{Aikawa2008}.\\ ${b}$ Observational results are median values toward 16 low-mass protostars~\citep{Bergner2017}.\\ ${c}$ The observed abundance of CH$_2$OHCHO is toward IRAS 16293-2422~\citep{Jorgensen2016}.\\ All abundances are with respect to methanol. a(-b) means a$\times 10^{-b}$. Boldface indicates more than one order of magnitude disagreement between model and observations. } \end{table} Table~\ref{table4} shows major granular species abundances predicted by different models and the observed ice abundances toward low-mass protostars by~\citet{Oberg2011}. All the abundances are expressed in percentages with respect to the water ice. We can see that the abundances of CO$_2$ are much lower than the observed value. Moreover, the CO$_2$ abundance in the model MC5 is the highest. The reaction CO + OH $\rightarrow$ CO$_2$ + H is the most efficient reaction to convert CO to CO$_2$ on grain surfaces when the temperature is higher than 12 K~\citep{Garrod2013}. However, most granular species are formed when the temperature of the core is around 10 K, so the conversion from CO to CO$_2$ on grain surfaces is not efficient in our models. The chain reaction mechanism ~\citep{Chang2007, Chang2012} allows surface OH to react with surface CO immediately after surface OH is produced in ice. So, if it is introduced in surface and bulk models, the reaction CO + OH $\rightarrow$ CO$_2$ + H can proceed efficiently at around 10 K. However, rigorous treatment of the chain reaction mechanism requires the microscopic Monte Carlo method, which is much more computationally expensive. On the other hand, the chain-reaction mechanism can be approximately implemented by the rate equation approach~\citep{Garrod2011}. Following \citet{Garrod2011}, we implement the chain-reaction mechanism in models MC1, MC2 and MC3. The abundances of CO$_2$ ice increase to more than 10\% relative to water ice with the chain-reaction mechanism. The abundance of JCO predicted by model MC5 is the lowest among all models because JCO cannot be trapped in ice mantle in the two-phase model. We can also see that granular methanol abundances are overestimated in the basic and new multiphase models. Because of the exponential decay of the rate of photodissociation reactions with depth in the new multiphase models, methanol molecules are less likely to be photodissociated in models MC1, MC2 and MC3 than in model MC5. Therefore, granular methanol abundances in models MC1, MC2 and MC3 are higher than that in model MC5. Moreover, the ice mantle is completely inert in model MC4, thus the methanol ice abundance in model MC4 is even higher than that in the new multiphase models. \citet{Pontoppidan2003} found that the methanol ice abundances are more than 15\% relative to water ice toward 3 low-mass protostars, which agrees with our new multiphase model results. The different formation history of protostars, for instance the time scale before collapse, may be able to explain the discrepancy. Methane and NH$_3$ abundances predicted by theoretical models agree reasonably well with observations. \begin{table} \caption{Comparison of Model and Observational Granular Species Abundances} \label{table4} \begin{tabular}{c c c c c c c} \hline Model &JCO &JCO$_2$ & JH$_2$CO &JCH$_3$OH &JCH$_4$ & JNH$_3$\\ \hline \multirow{1}{*}{MC1} &65.5 &0.59 &10.8 &22.6 &7.2 & 11.4 \\ \hline \multirow{1}{*}{MC2} &45.3 &0.99 &10.7 &24.4 &6.8 & 11.7\\ \hline \multirow{1}{*}{MC3} &46.9 &0.60 &12.1 &24.6 &6.9 & 11.6\\ \hline \multirow{1}{*}{MC4} &11.6 &0.17 &15.2 &33.8 &7.6 & 25.0 \\ \hline \multirow{1}{*}{MC5} &10.0 &2.0 &9.4 &5.8 &1.1 & 13.4 \\ \hline \multirow{1}{*}{Observation } &29 &29 & -- & 3 & 5 & 5 \\ \hline \end{tabular} \tablecomments{The abundances of major ice mantle components are percentages with respect to the water ice. Observational results are from ~\citet{Oberg2011}. } \end{table} It was recently found that there are abundant O$_2$ molecules (O$_2$/H$_2$O in the range of 1\%-10\%, mean value $\sim$ 4\%) in the coma of comet 67P/Churyumov-Gerasimenko(67/C-G)~\citep{Bieler2015}. \citet{Mousis2016} argued that O$_2$ can be formed radiolysis of water ice in parental molecular clouds and then trapped in clathrates in the solar nebula stage. \citet{Taquet2016}, on the other hand, solved detailed gas-grain reaction networks and showed that primordial (i.e. interstellar) O$_2$ ice in dark molecular clouds can be as abundant as observed in 67/C-G, if the dark clouds are rather dense and warm. However, it was found that the JO$_2$/JH$_2$O ratio is well below 0.01 in most fluid parcels during the formation of protostellar disks and the high JO$_2$/JH$_2$O ratio only exists in the upper layers of disks\citep{Taquet2016}. Since the bulk ice mantle is partially active in our models while it is completely inert in \citet{Taquet2016}, it is worth presenting how O$_2$ ice evolve in our new multiphase models. Figure~\ref{Fig15} shows the radial distribution of O$_2$ ice relative to water ice at $t_{\textrm{final}}$. We can see that our new multiphase models predict that the abundance of O$_2$ ice is around 10\% of water ice in the outermost fluid parcel, whose temperature is around 16 K and is about 8000 AU from the center of the protostar. The abundance of O$_2$ ice drops inward to around 3\% of water ice abundance at r=125 AU. Although our results cannot be directly compared with that of~\citet{Taquet2016}, because our core model is spherical, relatively abundant O$_2$ ice survives down to the water sublimation radius in our new multiphase models. On the other hand, the abundances of O$_2$ ice in models MC4 and MC5 are much lower than 4\%. \begin{figure} \centering \resizebox{12cm}{12cm}{\includegraphics{Fig15.eps}} \caption{ Radial distributions of O$_2$ ice abundances at $t_{\textrm{final}}$. The abundances are relative to water ice. } \label{Fig15} \end{figure} In recent years, COMs are detected also in prestellar cores~\citep{Bacmann2012}. It motivated intense studies on grain-surface reactions, non-thermal desorption, and gas-phase formation of COMs at cold temperatures ($\sim$ 10 K)~\citep{Vasyuninherbst2013,Chang2016,Vasyunin2017,Chen2018}. It would thus be useful to compare our model results before the gravitational contraction with the observations of prestellar cores. The fluid parcels in our model are, however, located at the radius of $\sim 10^4$ AU before the gravitational contraction, so that they fall to the central region of the protostellar core. We thus compare our gas-phase abundances before the contraction with the observations of dense cloud TMC-1, rather than prestellar cores. The observed fractional abundances of gas-phase H$_2$CO and CH$_3$OH toward the dense cloud TMC-1 are $5\times 10^{-8}$ and $3\times 10^{-9}$ respectively~\citep{Smith2004}. On the other hand, \citet{Soma2018} recently found that H$_2$CO and CH$_3$OH abundances could be similar. Fig. \ref{Fig5} shows that models MC1-4 either underestimate H$_2$CO or CH$_3$OH abundances while the two-phase model can reproduce the observed abundances of H$_2$CO and CH$_3$OH well at the time later than $1\times 10^5$ yr. So, more efficient desorption mechanism such as reactive desorption should be introduced into models MC1-4 in order to reproduce the observed abundances of CH$_3$OH toward TMC-1~\citep{Garrod2007}. On the other hand, models MC1-4 can predict that abundances of gas-phase H$_2$CO and CH$_3$OH are similar at the time $1\times 10^6$ yr while difference of H$_2$CO and CH$_3$OH abundances in the two-phase model is usually around one order of magnitude. The visual extinction of fluid parcels in our model are around 4 before contraction, which is much lower than that under standard physical conditions (Av=10). FUV photons can not only desorb molecules out of grains, but also destruct these molecules in gas phase and ice mantle. We perform test simulations of all chemical models using the standard physical conditions in order to see if the lower visual extinction makes any difference to the gas-phase H$_2$CO and CH$_3$OH abundances. In the test simulations, the visual extinction is set to be 10, the density of H nuclei is $2\times 10^4$ cm$^{-3}$ while the temperature is 10 K. We found that the abundances of H$_2$CO and CH$_3$OH at the time $2\times 10^5$ yr using the standard physical conditions and in the fluid parcel differ by less than a factor of two in all models. At the time $1\times 10^6$ yr, a factor of two or less difference is introduced into the abundances of these two species in models MC1-4 and H$_2$CO abundance in the two-phase model by the lower visual extinction. However, at the later time, CH$_3$OH and H$_2$CO abundances in the two-phase model using the standard physical conditions are similar, which is more than one order of magnitude larger than CH$_3$OH abundance in the fluid parcel. \section{Conclusions and Discussions}\label{diss} In this work, we study gas-grain chemistry in a collapsing core during star formation using a new multiphase astrochemical approach. Our emphasis is on the formation of complex organic molecules. We also pay attention to the gas-ice balance of CO. The physical structure of the collapsing core is based on a 1-D RHD model. Our new multiphase model results are compared with older two-phase and basic multiphase results. The abundances of COMs in our new multiphase models are found to be more than one order of magnitude higher than in two-phase models. This difference can be explained by two reasons. First, because of the exponential decay with depth of photodissociation reaction rates, COMs buried in the ice mantle in the new multiphase models are less likely to be photodissociated than in the two-phase model. Second, abundant radicals accumulate in the ice mantle when the temperature is around 10 K in the new multiphase models. The accumulated radicals diffuse inside ice mantle and react with each other to form COMs when the temperature increases. In the basic multiphase model, the abundances of COMs such as methanol, which are mainly formed by hydrogenation reactions at 10 K, are higher than in the new multiphase models. However, the abundances of COMs, that are mainly formed by radical recombination at higher temperatures, are much lower in the basic multiphase model than in the new multiphase models because radicals cannot accumulate in the bulk of ice mantle in the basic multiphase model. The desorption of CO in the new multiphase models and in the two-phase model is different. Because significant amounts of CO are locked in the ice bulk mantle, solid CO molecules sublime via two processes in the new and basic multiphase model. First, solid CO occupying interstitial binding sites or in the active layers desorbs when the temperature reaches the sublimation points. Second, solid CO sublimes together with water ice when the temperature is high enough to desorb water ice. The amounts of CO locked in the ice mantles depend on the $\alpha$ value, which is a parameter that determines the probability photofragments enter normal sites to be locked in the ice mantle. The larger the $\alpha$ value is, the fewer CO molecules are locked in the ice mantle. Because dust grains are covered with ice mantles at $T<100$ K, significant amounts of solid CO still exist at $T<100$ K. We introduce a new COM formation scenario applicable during star formation. COMs are believed to be formed when the temperature of a collapsing core increases and radicals, which are generated by photodissociation reactions, recombine with each other~\citep{Garrod2006,Aikawa2008}. Almost all radicals that form COMs in the two-phase model were generated when the temperature of the core increases~\citep{Garrod2006}. In our multiphase model, most radicals that recombine to form COMs have a primordial origin when the temperature is around 10 K before the core starts to collapse. The abundances of COMs such as CH$_3$OCH$_3$ in the two-phase model are much lower than observations, which suggests that more radicals need to be generated in ice so that these radicals can recombine to form more COMs. In our new multiphase model, radicals formed in the prestellar phase can be stored, thus more COMs can be formed upon warm-up. Our COM formation scenario is also consistent with laboratory experiments, at least qualitatively. For example, \citet{Butscher2016} showed that abundant free radicals can be stored in ice mantles by photolysis of water rich ice mixture at around 13 K while COMs can be formed by radical recombination when the temperature increases. Moreover, it was found that abundant free radicals can be frozen in the ice mantle which is photolysed by UV radiation at around 10 K~\citep{Schutte1991}. The recombination of free radicals can even lead to chemical explosion if there are enough free radicals in the ice mantle~\citep{Schutte1991}. Our new multiphase models are also consistent with laboratory experiments of CO desorption~\citep{Collings2003, Sandford1988}. A fraction of CO is desorbed around 20 K and the rest is locked in water ice and thus desorbed together with water ice at $\gtrsim$ 100 K. The abundant O$_2$ found in the coma of comets must also be locked in ice if these O$_2$ molecules have a primordial origin, because unlocked O$_2$ sublime at a much lower temperature. Our new multiphase models keep the O$_2$/water ratio as high as $\gtrsim$ 3\%, around the water ice sublimation region. It is interesting that COM formation and CO desorption, which are two independent questions, are linked together in our new multiphase models. The population of species unlocked in ice mantles, which can diffuse inside ice mantles before water ice sublime, play a central role in answering these two questions. Finally, we discuss the limitations of our modeling. First, the gas and dust temperatures are assumed equal. However, the gas densities of fluid parcel beyond 3000 AU are low enough for the gas and dust temperatures to be decoupled. In order to find out the effect of temperature decoupling on the chemical evolution of fluid parcels, we simulate a test model in which the dust temperatures are 2 K lower than the gas temperatures when the gas temperatures are higher than 12 K. We still assume that the gas temperatures and dust temperatures are equal when the gas temperatures are lower than 12 K because surface chemistry on dust grains is not dependent on temperatures when the dust temperatures are between 8 K and 12 K~\citep{Garrod2011}. The chemical model used in the test model is the same as that in model MC2. We only simulate the molecular evolution of the fluid parcel that reaches 4000 AU and 8000 AU (the outermost fluid parcel) at t$_{\textrm{final}}$. We found that the test model and model MC2 predict similar abundances of all species in the fluid parcel that reaches 4000 AU at t$_{\textrm{final}}$. Moreover, other than granular COMs that are formed by the recombination of two radicals, species abundances in the outermost fluid parcel predicted by the test model are also similar to these by model MC2. However, the abundances of granular COMs which are formed by the recombination of two radicals in the test model may be much lower than in model MC2 because the dust temperatures in the test model are not high enough so that radicals cannot diffuse and recombine to form COMs in ice. For instance, model MC2 predicts that the fractional abundance of JCH$_3$OCH$_3$ in the outermost fluid at t$_{\textrm{final}}$ is around $10^{-8}$ while JCH$_3$OCH$_3$ molecules can hardly be formed in the same fluid parcel in the test model because of the lower dust temperatures. Moreover, because the temperatures of inner fluid parcels (r $<$ 4000 AU) are not lower than the fluid parcel that reaches 4000 AU at t$_{\textrm{final}}$, the test model and model MC2 should predict similar abundances of all species in these inner fluid parcels. Another issue is the binding energy of atomic oxygen. In this work, the binding energy of atomic oxygen is set to be 800 K, which is commonly used in astrochemical modeling. However, \citet{He2015} recently suggested a much higher value, 1660 K. In order to investigate how the higher binding energy of atomic oxygen affect our simulation results, we run a test model in which the binding energy of atomic oxygen is set to be 1660 K. The chemical model used is model MC2. Hereafter, we call this test model TMC2. We simulate model TMC2 for all fluid parcels. We found that in the pre-collapse phase, the majority of species are not much affected by the binding energy of atomic oxygen, except for the abundances of a few species whose production depends on the diffusion of O atoms. At the time 10$^6$ yr, the abundance of ICO$_2$ in model TMC2 is more than one order of magnitude lower than model MC2 in all fluid parcels, because the recombination of IO and IHCO is one of the major reaction to form ICO$_2$ at 10 K. The abundance of interstitial O$_2$ is also reduced in model TMC2 because most IO$_2$ is formed by the recombination of IO and IO. At 10$^6$ yr, the abundances of IO$_2$ are around $2\times 10^{-6}$ in all fluid parcels in model MC2 while few IO$_2$ molecules are produced in model TMC2. The abundances of JCO$_2$ and JO$_2$ also drop by about a factor of two and one order of magnitude respectively in model TMC2 in all fluid parcels. At t$_{\textrm{final}}$, the abundances of all species predicted in model TMC2 do not differ much from these in model MC2 in the inner fluid parcels (r $\leq$ 4000 AU). However, the abundance of JO$_2$ in model TMC2 is about two orders of magnitude lower than in MC2 in the outermost fluid parcels (T $\sim$ 16 K) at t$_{\textrm{final}}$. Because the binding energy of atomic oxygen can strongly affect the formation of a few granular species such as JO$_2$ when the grain temperature is low (T $\leq$ 16 K), more work is needed to measure this binding energy more accurately. Our new multiphase models have two strengths. First, the exponential decay of photodissociation reactions with depth into ice mantles is well mimicked. Second, species that are locked in the ice mantle and these species that can diffuse in the ice mantle are well distinguished. To the best of our knowledge, few current astrochemical models have the first strength. On the other hand, the computational cost to simulate the new multiphase models is much lower than that required by the model introduced by \citet{Chang2014} because we use a macroscopic approach. We focus on the formation of COMs in this paper, but our new multiphase models warrant further investigation. For example, the warm carbon-chain chemistry (WCCC) is triggered by the sublimation of CH$_4$ from ice mantle~\citep{Sakai2008}, thus, the gas-ice balance of CH$_4$ is crucial for the WCCC. The gas-ice balance of volatile species in the new multiphase models will be discussed in detail in a subsequent paper~\citep{Wang2018}. Q. Chang is a research fellow of the One-Hundred-Talent project of the Chinese Academy of Sciences. This work was funded by The National Natural Science foundation of China under grant 11673054. YA wishes to acknowledge the support of the JSPS through KAKENHI Grant 16K13782, 16H00931 and 18H03718. We thank Eric Herbst for a critical reading of the manuscript and helpful comments. We thank Ling Liao and Yao Wang for helpful discussions. We thank Fangfang Li and Long-Fei Chen for proofreading of the manuscript. We thank our referee's constructive comments to improve the quality of the manuscript. The Taurus High Performance Computing system of Xinjiang Astronomical Observatory was used for the simulations.

3c31773dbd4b751467e1f152bd69e894c1009775

\section{Introduction} The interaction of electrons and spin excitations plays a fundamental role for a wide variety of phenomena. For example, spin-polarized currents depolarize due to their interaction with magnons~\cite{zutic04}, the characteristic temperature dependence of the tunneling magnetoresistance (TMR) \cite{macdonald98} is determined by the electron scattering by magnons, and in nanospintronics the spin and charge currents that flow through nanostructures can be strongly affected by the electron-magnon interaction \cite{balashov06,balashov08,schweflinghaus14,schweflinghaus16}. Moreover, it is speculated that the electron-magnon interaction is the origin of the superconductivity in Fe pnictides \cite{dagotto94,scalapino95,sasmal08,dahm09}. The elementary ferromagnets Fe, Co, and Ni are suitable model systems to study the electron-magnon interaction. They form a class of intermediately correlated materials in which localized $d$ states close to the Fermi level are embedded in a free-electron-like band structure. While density-functional theory (DFT), employing either the local-spin-density (LSDA) or the generalized gradient approximation (GGA) to the exchange-correlation potential, is able to capture the ground-state properties of these materials, it fails to yield accurate excited-state properties. For example, the exchange splitting of these materials is often overestimated within DFT calculations. Nickel is an extreme case in that the exchange splitting is almost a factor of two too large. A strongly spin-dependent band broadening is observed in angle-resolved~\cite{monastra02} and two-photon photoemission experiments \cite{knorren00}, which indicates that the electron-magnon interaction plays an important role in $3d$ ferromagnets. In addition, recent angle-resolved photoemission spectroscopy (ARPES) experiments reveal pronounced differences in the quasiparticle dispersion compared to DFT. ARPES measurements, for example, have exhibited anomalous kinks in the band dispersion of iron surface states \cite{schaefer04} and nickel bulk states \cite{hofmann09}. These kinks appear at binding energies much larger than what one would expect for a phonon-mediated band renormalization. It has been speculated that these band anomalies are a footprint of electron-magnon scattering because the energetic position of the kinks corresponds to typical magnon energies in the materials. We will later show that a similar band anomaly in iron, which was observed in ARPES spectra \cite{Mlynczak19} at a comparatively large binding energy of 1.5 eV, can indeed be explained within our theory. (In this particular case, the theoretical prediction predated the experimental observation.) The appearance of kinks in the electronic band dispersion and the strong band broadening of the quasiparticle peaks clearly go beyond the scope of a static mean-field theory such as DFT and calls for a genuine many-body description. One such method is the dynamical mean-field theory (DMFT), which maps the interacting many-body system onto an Anderson impurity model~\cite{georges96,kotliar06}. DMFT relies on the choice of model parameters, the effective intra-atomic interaction parameters $U$ and $J$ as well as the intra- and inter-atomic hopping parameters. In applications to real materials, the latter are usually taken from a L(S)DA mean-field solution. An early implementation~\cite{katsnelson99} of this LSDA+DMFT method showed that majority hole states are strongly damped in iron. This result has been confirmed by Grechnev~\textit{et al.}~\cite{grechnev07} who studied Fe, Co, and Ni with the same approach. They found a strong damping of the majority quasiparticle states and, in addition, a shallow satellite feature below the bottom of the $d$ bands. A comparison of ARPES spectra to LSDA+DMFT calculations~\cite{sanchez-barriga09,sanchez-barriga10,sanchez-barriga12} revealed that the agreement between experiment and theory is considerably improved with respect to LSDA, but still the linewidths and the effective masses tend to be underestimated compared to experiment. So far, no evidence of band dispersion anomalies has been found in the theoretical studies. While allowing for an efficient treatment of many-body effects, the DMFT method suffers from some badly controlled approximations. Above all, the usage of the impurity model essentially amounts to neglecting the momentum dependence of the many-body scattering processes. Furthermore, the choice of the model parameters introduces an element of arbitrariness. Often they are fitted to experiment, which limits the predictive power of the method. Finally, since LSDA already contains the electron-electron interaction in an approximate way, it is necessary to apply a double-counting correction, for which no unique definition exists. In this work, we employ an alternative description that avoids the usage of a model. The single-particle wave functions and propagators are allowed to extend over the whole (infinite) crystal. In this way, the momentum dependence is retained, and there is no need for model parameters nor a double-counting correction. Many-body perturbation theory (MBPT) is employed to construct an approximation to the electronic self-energy $\Sigma(\mathbf{r},\mathbf{r}';\omega)$. This effective scattering potential describes many-body exchange and correlation scattering processes that an electron or hole experiences as it propagates through a many-electron system. In other words, the self-energy connects the non-interacting mean-field system to the real interacting system. The solution of the Dyson equation then yields the single-particle Green function of the interacting system, and the imaginary part of which is directly related to the photoemission spectra~\cite{onida02}. The most popular approximation to the electronic self-energy is the $GW$ approximation, which has been shown to yield accurate band structures for a wide range of materials. For example, it is known to have a strong effect on the band gaps of semiconductors and insulators \cite{Aulbur99}, which are corrected from their (usually underestimated) DFT values towards experiment. The $GW$ method has also been applied to the elementary ferromagnets, where it partly cures the shortcomings of LSDA. While the $3d$ band width is typically overestimated in LSDA, the $GW$ approximation applied to iron~\cite{yamasaki03} and nickel~\cite{aryasetiawan92,friedrich10} yields results in better agreement to the experimental values~\cite{eastman80,hoechst77,himpsel79,eberhardt80,heimann81,kirby85,hoechst76,hoechst76a}. The explicit electron-electron scattering of the $GW$ self-energy gives rise to a lifetime broadening of the bands, which is, however, too small to explain the broadening seen in photoemission experiments, indicating that the $GW$ self-energy misses some important scattering processes in these materials. \iffalse and, in addition, it does not reproduce the satellite structure present in the electronic spectrum of {\color{red} iron and} nickel. Aryasetiawan~\cite{aryasetiawan92} argues that the overestimation of the exchange splitting and the missing satellite structure are related features that need a treatment beyond the $GW$ approximation taking into account particle-particle scattering within a $T$-matrix approach. Springer \textit{et al.~}\cite{springer98} have applied an \textit{ab initio} $GW$+$T$-matrix theory to nickel. Indeed they found the satellite structure due to multiple hole-hole scattering processes and a significant reduction of the exchange splitting. Zhukov \textit{et al.~}\cite{zhukov04,zhukov06} have applied a $GW$+$T$-matrix theory to study excited electrons in iron and nickel. They take into account spin-flip and non-spin-flip processes within the multiple electron-electron and electron-hole scattering channel. While in iron the electron-hole spin-flip scattering strongly reduces the electron lifetime, this effect appears to be weak in nickel. \fi In this work, we derive from the Hedin equations~\cite{hedin65} a first-principles self-energy approximation that describes the scattering of electrons and magnons. The Hedin equations are a set of integro-differential equations that, if solved self-consistently, would yield, in principle, the exact self-energy for a many-electron system. A full self-consistent solution is not possible in practice, but the Hedin equations can be used to derive approximations to the self-energy. For example, the $GW$ approximation results from a single cycle through the equations starting from $\Sigma=0$. By iteration of the Hedin equations, one systematically generates more and more higher-order self-energy diagrams. In this way, we identify and select those scattering diagrams that are relevant for the coupling of electrons to many-body spin excitations. Then, summing these ladder diagrams to all orders in the interaction yields the $GT$ approximation, where $T$ stands for the magnon propagator, which describes the correlated motion of an electron-hole pair with opposite spins. It is shown that the lowest-order diagram of the $T$ matrix is of third order in $W$, which renders double-counting corrections with $GW$ unnecessary. The $GT$ approximation has a similar mathematical form as the $GW$ self-energy in that it is given by a product of the single-particle Green function $G$ and the effective magnon propagator $T$. The implementation is, however, complicated by the fact that this $T$ matrix is a four-point quantity: It depends on four points in space and results from a solution of a Bethe-Salpeter equation. A numerical implementation has been realized by employing a basis set of Wannier functions, which allow for an efficient truncation of the four-point quantity in real space. \cf{It should be noted that the present first-principles approach is based on a formulation at absolute zero, whereas DMFT approaches usually require the assumption of a finite electronic temperature.} \iffalse In this paper we propose a self-energy correction within MBPT relying on the multiple-scattering $T$-matrix to account for the scattering of electrons and magnons. The self-energy, which we call $GT$ self-energy in the present paper, can be combined with the $GW$ approximation \cite{hedin65} without the need for a double-counting correction. \fi The theoretical foundation of the $GT$ self-energy is sketched in Sec.~\ref{sec::Theory}. \cf{Section \ref{sec:Goldstone} discusses the self-energy contribution of electronic scattering with acoustic magnons (Goldstone mode) in the long-wavelength limit.} The implementation of the $GT$ self-energy within the \textsc{spex} code \cite{friedrich10} is described in Sec.~\ref{sec::Implementation}. \cf{The important topic of the violation of causality is discussed in detail in Sec.~\ref{sec:causality}.} We analyze the many-body renormalization of the band structure of the bulk elementary ferromagnets iron, cobalt, and nickel in Sec.~\ref{sec::Calculations}. The coupling of electrons to spin excitations leads to a pronounced spin-dependent lifetime broadening of the quasiparticle states. The lifetime broadening, which is particularly strong in iron, can lead to a complete loss of the quasiparticle character in certain energy regions of the electronic spectrum -- with energy gaps appearing in electronic bands, effectively cutting them in two. Furthermore, we find a band anomaly at higher binding energy in iron, in agreement with a very recent ARPES study. Section~\ref{sec::Conclusions} gives a summary. \section{Theory} \label{sec::Theory} Our goal is to construct a self-energy that describes the many-body renormalization due to the scattering of electrons and magnons. Here, we should understand the term \emph{magnon} to comprise not only the collective spin-wave excitations but also the single-particle Stoner excitations, which provide a decoherence channel, through which the spin waves acquire a finite lifetime. In our formulation, the two types of spin excitations are treated on the same footing. They are intimately coupled to each other (two sides of the same coin) and cannot, as a matter of principle, be treated separately. In this sense, the self-energy will describe the dressing of a propagating particle (electron or hole) through the creation and annihilation of collective spin waves and single-particle Stoner excitations alike. We note that the spin-orbit coupling is neglected. \begin{figure} \includegraphics[width=1.\columnwidth]{SelfEnergyExpansion.pdf} \caption{\label{fig::SelfEnergyExpansion} Feynman diagrams of the Hedin self-energy expansion up to the third order in the screened interaction $W$, which is denoted by a wiggly line. The Green function is denoted by an arrow.} \end{figure} Figure \ref{fig::SelfEnergyExpansion} shows the self-energy expansion in Feynman diagrams up to third order in the screened interaction $W$ as derived from the Hedin equations. Cutting the expansion at the first order yields the $GW$ self-energy, an approximation commonly used in computational condensed-matter physics. Despite its success, however, it lacks a number of many-body scattering effects, for example, particle-particle (electron-electron and hole-hole) and electron-hole scattering as well as higher-order exchange processes. To account for these, the $GW$ method has been combined \cite{springer98,zhukov04,romaniello12} with the $T$-matrix theory, in which the self-energy is expanded in terms of infinitely many ladder diagrams, describing the correlated motion of two particles (electrons or holes). This $GW$+$T$ approach is motivated phenomenologically: While $GW$ describes the correlation of itinerant $s$ and $p$ states, the ladder diagrams account for correlation effects taking place in localized states. However, as the $T$-matrix self-energy contains the Hartree ``tadpole'' diagram and the direct term of the second Born approximation, which is also contained in the $GW$ self-energy \footnote{This is true if the diagrams are defined in terms of the free Green function (excluding the electron-electron term of the Hamiltonian) and the bare interaction or the effective Hubbard $U$ interaction. We will later use the Kohn-Sham Green function and the fully screened interaction $W$, in which case the two diagrams (omitted in the present formulation) would contain double-counting errors in themselves.}, a double-counting correction would be required in this approach~\cite{springer98}. Instead, we come back to the expansion shown in Fig.~\ref{fig::SelfEnergyExpansion}. Apart from the $GW$ self-energy, this expansion consists of the second- and third-order screened exchange diagrams and two direct diagrams (the ones of third order in the first row) that also appear in the $T$-matrix approach. These two diagrams have the characteristic form of ladder diagrams where the rungs of the ladder correspond to the screened interaction and the rails correspond to single-particle Green functions. They describe the correlated motion of an electron-hole pair and a particle pair (electron-electron or hole-hole), respectively. If the electron and hole are of opposite spins, the corresponding propagator can be seen as part of a solution of the Bethe-Salpeter equation (BSE) for the transverse magnetic response function~\cite{aryasetiawan99,sasioglu10,friedrich14,friedrich18}. The full BSE solution would comprise ladder diagrams of all orders, including the first and second as well as the infinitely many higher-order diagrams. For example, the collective spin-wave excitations arise from a resummation of ladder diagrams up to infinite order. Transversal spin waves are low-energy excitations (going down to zero energy for vanishing momentum in the absence of spin-orbit coupling). The corresponding self-energy diagrams are therefore expected to yield the principal low-energy scattering contribution in ferromagnets. The particle-particle diagrams, on the other hand, have been shown to be responsible for the appearance of satellite features, e.g., the $6\,\textrm{eV}$ satellite in nickel~\cite{springer98}. The electron-hole diagrams without a spin flip (electron and hole have the same spin) couple to density and longitudinal spin fluctuations, which are of much higher energy (in the order of the plasma frequency) than the transversal spin waves. The other diagrams shown in Fig.~\ref{fig::SelfEnergyExpansion} cannot describe a coupling to spin fluctuations because, in the absence of spin-orbit coupling, they have a single continuous Green-function line with a unique spin quantum number. For these reasons, we have implemented the electron-hole self-energy ladder diagrams, starting at the third order as prescribed by the Hedin self-energy expansion. The summation is to all orders in $W$ (the third-order diagram generates the fourth-order diagram in the next cycle of the Hedin equations and so on) but excludes the first- and second-order diagrams, which avoids the double-counting problem mentioned above. Iterating the Hedin equations thus yields, in a natural way, a combination of the $GW$ and the $T$-matrix self-energy that is free from double-counting errors. \cf{At this point, we should mention that an implementation starting from the third-order diagram was already considered in Ref.~\onlinecite{springer98} but dismissed because the authors had realized that the renormalized Green function violated a causality condition: The imaginary part of the Green function showed an incorrect change of sign at large (absolute) energies. We observe the same behavior and discuss below in detail that the sign change is not a numerical artifact but, in fact, expected to occur for the presently chosen self-energy diagrams. However, we argue that this violation of causality can be accepted as the sign change is observed in an energy region, which is not in the focus of the present paper. In particular, it is expected that other self-energy diagrams, neglected in the present work, will play a more important role there and should eventually act to restore the correct sign.} \begin{figure} \includegraphics[width=0.85\columnwidth]{GW_GT.eps} \caption{\label{GW_GT}$GT$ diagram (right) drawn to resemble the $GW$ diagram (left). The $T$ matrix is a magnon propagator. The $GT$ self-energy thus describes the emission and re-absorption of magnons in much the same way as $GW$ describes the corresponding processes for plasmons. Note that, after emission of the magnon, the spin of the propagating particle has flipped, since the magnon carries a total spin of unity. } \end{figure} From a physical point of view, the self-energy diagrams, which form what we call the $GT$ self-energy in the following, describe the emission and re-absorption of magnons. Here, the magnons should again be understood to comprise both spin waves and Stoner excitations. In this sense, the $T$ matrix can be viewed as an effective interaction that acts through the exchange of magnons in much the same way as $W$ includes the exchange of plasmons. To illustrate this analogy, we represent the self-energy diagram in a way that resembles the $GW$ diagram in Fig.~\ref{GW_GT}. However, we will see that the $T$ matrix is a much more complex quantity than $W$ as it is a four-point function, i.e., it depends on four points in space (and time), whereas $W$ is a two-point function. The $GT$ self-energy is expected to yield an important scattering channel for electrons in magnetic materials. We should mention already at this point that for the present study we omit the explicit evaluation of the $GW$ self-energy for simplicity and employ a corrected LSDA solution instead. This solution is made to fulfill the Goldstone condition for the spin excitations and has been shown~\cite{mueller16} to resemble the self-consistent solution obtained with the Coulomb-hole screened-exchange (COHSEX) self-energy, the static limit of $GW$, \cf{and to bring the exchange splittings closer to experiment.} \begin{figure} \includegraphics[width=1.\columnwidth]{GTSelfEnergy.pdf} \caption{\label{fig::GTSelfEnergy} Feynman diagram of the electron-magnon self-energy in the ladder approximation. Notation as in Fig.~\ref{fig::SelfEnergyExpansion}. It is assumed that the Green functions that are solely connected by the screened interaction have opposite spins.} \end{figure} The $GT$ self-energy \begin{equation}\label{eq::ElMagSelfEnergy} \Sigma^{\sigma}(13) = -\textrm{i} \int \textrm{d}2\,\textrm{d}4\,\,T^{\sigma\sigma^{\prime}}(12,34) G^{\sigma^{\prime}}(42) \end{equation} with the spin indices $\sigma$ and $\sigma^\prime=-\sigma$ is written in terms of the interacting time-ordered single-particle Green function $G^{\sigma}$ and an electron-hole $T$ matrix. We have used the short-hand notation $1=(\mathbf{r}_{1},t_{1})$, and atomic units are used throughout. Figure \ref{fig::GTSelfEnergy} shows a representation in Feynman diagrams. In the ladder approximation, the multiple-scattering $T$ matrix is the solution of the BSE \begin{eqnarray}\label{eq::TmatrixBSE} &&T^{\sigma\sigma^{\prime}}(12,34) = \int \textrm{d}5\,\textrm{d}6\,\, W(12) K^{\sigma\sigma^{\prime}}(12,56) \nonumber\\ &&\times\left\{ W(56)K^{\sigma\sigma^{\prime}}(56,34) W(34) + T^{\sigma\sigma^{\prime}}(56,34) \right\} \end{eqnarray} with the screened interaction $W(12)$ in the random-phase approximation (RPA) and the free electron-hole pair propagator \begin{equation} K^{\sigma\sigma^{\prime}}(12,34) = \textrm{i} G^{\sigma}(13) G^{\sigma^{\prime}}(42).\label{K} \end{equation} Equation (\ref{eq::TmatrixBSE}) defines the $T$ matrix to contain the ladder diagrams to all orders starting from the third order in $W$. \section{Goldstone contribution}\label{sec:Goldstone} As already mentioned, the $T$ matrix is closely related to the transverse magnetic response function $R^{+-}(12)=\delta m^+(1) / \delta B^+(2)$, which gives the linear response of the spin density with respect to an externally applied circularly polarized magnetic $B$ field. For simplicity, we omit ``$+$'' and ``$-$'' denoting the circular polarizations in the following. We now generalize the spin density to the spin density matrix $m(12)$ and the magnetic field to a non-local field $B(34)$. Then, the four-point response function $R(12,34)=\delta m(12)/\delta B(34)$ can be written as the solution of a BSE (see, e.g., Ref.~\onlinecite{friedrich18} \footnote{$R=-2R^{(4)}$ in the notation of Ref.~\onlinecite{friedrich18}.}). In matrix notation, the solution reads $R=-2(\mathds{1}-KW)^{-1}K$ with the free electron-hole pair propagator of Eq.~(\ref{K}), which has poles at the energies of the unrenormalized Stoner excitations. The inverse $(\mathds{1}-KW)^{-1}$ renormalizes these Stoner excitations, but it also gives rise to the collective spin-wave modes. Without it, one would simply obtain $-2K$, the transverse magnetic response function of the non-interacting mean-field system. An important special case is the limit $\mathbf{k}\rightarrow\mathbf{0}$. In this limit (and in the absence of spin-orbit coupling), a static $B$ field that is orthogonal to the spin polarization is able to rotate the spins collectively towards the field direction even if $B$ is infinitesimally small. This leads to a spin-wave delta peak in $R$ at $\mathbf{k}=\mathbf{0}$ and $\omega=0$. For finite $\mathbf{k}$, the spin-wave dispersion can be shown \cite{moriya85} to be quadratic in $k$, which, in a more general sense, corresponds to the Goldstone mode expected to arise in the case of a spontaneously broken symmetry \cite{nambu60,goldstone61} such as ferromagnetism. The delta peak at $\omega=0$ results from the infinite summation of ladder diagrams and will thus also appear in the $T$ matrix, where it is likely to pose a numerical problem in the evaluation of the self-energy. Therefore, we discuss its contribution to Eq.~(\ref{eq::ElMagSelfEnergy}) in the following. As a first step, we have to characterize the Goldstone delta peak in a more detailed way. As already mentioned, an externally applied infinitesimal transversal $B$ field can rotate the electron spins macroscopically. In order for this rotation to be rigid (collective rotation of the spins), the linear transversal change of the ground-state spin density matrix $m(12)$ must be proportional to $m(12)$ itself, $\delta m\propto m$ (intercept theorem). Thus, integrating the response function $R$ with the infinitesimal $B$ field should give a finite response proportional to $m$. Conversely, the inverse of $R$ should have a vanishing eigenvalue with the eigenvector $m$, i.e., $(\mathds{1}-KW)m=0$. We now sketch a proof that this is fulfilled if the Coulomb-hole screened-exchange (COHSEX) self-energy is taken as the starting point. For brevity, we use a simplified notation. The static COHSEX self-energy $\Sigma^\sigma(12)=-W(12)n^\sigma(12)+\Sigma_\mathrm{c}(1)\delta(12)$ with the density matrix $n^\sigma(12)$ consists of the screened exchange (Hartree-Fock exchange potential with a statically screened Coulomb potential) and a static Coulomb-hole self-energy, which has the form of a local potential. When separating off the spin-independent part [$(\Sigma^{\uparrow}+\Sigma^{\downarrow})/2$], the remaining spin-dependent part can formally be interpreted as arising from a nonlocal magnetic field $B(12)=-W(12)m(12)/2$, which gives rise to the spin polarization in the COHSEX mean-field solution. Now we use the simple fact that rigidly rotating this $B$ field will rotate the magnetization in the same way, implying the proportionalities $\delta B\propto B$ and $\delta m\propto m$ with $\delta B/B=\delta m/m$. With the magnetic response function $-2K$ of the mean-field system, corresponding to the formula for $R$ without the inverse, we can formulate this magnetic response as $-2K\delta B=\delta m$ (in matrix notation), and, with the above relationship for $\delta B$ and $\delta m$, we thus have $-2KB=m$. Then, inserting $B=-Wm/2$ finally yields $KWm=m$. This proves that $m$ is an eigenvector of $\mathds{1}-KW$ with vanishing eigenvalue, a claim that we have previously \cite{mueller16} substantiated with arguments of theoretical consistency. There is another more detailed proof, which is presented elsewhere \cite{friedrich18}. To continue, we have to define a basis in which the quantities are to be expanded. We will later use a Wannier product basis, but, for the present purpose, it is easier to consider products of single-particle wave functions $\{ \varphi_{\mathbf{k}m}^{\uparrow}(1) \varphi_{\mathbf{k}'m'}^{\downarrow*}(2)\}$ as a basis, instead. The Goldstone condition $KWm=m$ gives $m$ as an eigenvector in this basis. With $m(12)=n^\uparrow(12)-n^\downarrow(12)$ and \begin{equation} n^\sigma(12)=\frac{1}{N}\sum_\mathbf{k}\sum_m f^\sigma_{\mathbf{k}m} \varphi^\sigma_{\mathbf{k}m}(1)\varphi^{\sigma*}_{\mathbf{k}m}(2) \end{equation} with the number $N$ of $\mathbf{k}$ points and the occupation numbers $f^\sigma_{\mathbf{k}m}$, we obtain after expansion of $\varphi_{\mathbf{k}m}^\sigma(1)$ in terms of $\{\varphi^{-\sigma}_{\mathbf{k}m'}(1)\}$ \begin{equation} m(12)=\frac{1}{N}\sum_{\mathbf{k}}\sum_{m,m'}(f_{\mathbf{k}m'}^{\uparrow}-f_{\mathbf{k}m}^{\downarrow}) \langle\varphi_{\mathbf{k}m}^{\downarrow}|\varphi_{\mathbf{k}m'}^{\uparrow}\rangle \varphi_{\mathbf{k}m}^{\downarrow}(1)\varphi_{\mathbf{k}m'}^{\uparrow*}(2)\,.\label{m-expand} \end{equation} According to this expression, we can immediately conclude (because of the factor $f_{\mathbf{k}m}^{\uparrow}-f_{\mathbf{k}m'}^{\downarrow}$) that the nonzero eigenvector elements are those where one of the states $(\mathbf{k}m\uparrow)$ or ($\mathbf{k}m'\downarrow$) is occupied and the other is unoccupied. If the spin-up and spin-down states are not too different from each other, we can also see that the combination $(\mathbf{k}m\uparrow)$ occupied and ($\mathbf{k}m'\downarrow$) unoccupied gives a sizable contribution, while the other term is negligible (because of the small overlap) These considerations are helpful in the discussion of the Goldstone-mode contribution to the $GT$ self-energy \begin{equation} \label{convol} \Sigma^{\sigma}(\omega)=-\frac{\mathrm{i}}{2\pi}\int_{-\infty}^\infty T^{\sigma\sigma'}(\omega')G^{\sigma'}(\omega-\omega')\mathrm{d}\omega'\,, \end{equation} here written in simplified notation and Fourier space. The frequency integral can be evaluated along a closed contour from $-\infty$ to $\infty$ and back to $-\infty$ along an infinite semiarc in the upper or lower complex half-plane. The latter does not contribute because both the $T$ matrix and the Green function behave asymptotically as $\omega^{-1}$. When writing $G$ and $T$ as sums over poles [e.g., see Eq.~(\ref{Lehmann0})], the contribution of any pair of poles can be evaluated with the residue theorem, and it vanishes unless the poles are both below or both above the real frequency axis (note the negative sign of $\omega'$ in the Green function). Let us consider the spin-down case first. The $T^{\downarrow\uparrow}$ matrix has the Goldstone delta peak infinitesimally below the real frequency axis, so its contribution is proportional to $1/(\omega-\epsilon_\mathrm{M}+\mathrm{i}\eta)$ with $\epsilon_\mathrm{M}=0$. Then, in the frequency convolution of Eq.~(\ref{convol}), only the poles of the Green function that are also below the real axis contribute, which, according to Eq.~(\ref{Lehmann0}), are those of the unoccupied states ($\mathbf{k}m'\uparrow$), resulting in self-energy poles at $\epsilon_{\mathbf{k}m'}^{\uparrow}+\epsilon_\mathrm{M}=\epsilon_{\mathbf{k}m'}^{\uparrow}$. According to Eq.~(\ref{m-expand}) and the discussion thereafter, the unoccupied states ($\mathbf{k}m'\uparrow$) must be combined with occupied states ($\mathbf{k}m\downarrow$) to form the eigenvector of the $T$ matrix that is related to the Goldstone peak. The corresponding eigenvector elements are exactly those that have been discussed above to be very small. Furthermore, because of the combination of states, only the self-energy matrix elements that couple occupied states can get a non-zero contribution from this eigenvector, for example, the diagonal element $\Sigma_{\mathbf{k}m}^\downarrow(\omega)=\langle\varphi_{\mathbf{k}m}^{\downarrow}|\Sigma^{\downarrow}(\omega)|\varphi_{\mathbf{k}m}^{\downarrow}\rangle$. The contribution would consist of poles located at frequencies larger than the Fermi energy, thus in an energy region that is less relevant for the renormalization of the occupied states ($\mathbf{k}m\downarrow$). Analogous arguments apply to the spin-up case: The $T^{\uparrow\downarrow}$ matrix has its Goldstone delta peak infinitesimally above the real-frequency axis. So, it is the Green-function poles related to the occupied states ($\mathbf{k}m\downarrow$) that now give rise to self-energy poles at $\epsilon_{\mathbf{k}m}^{\downarrow}$, the other poles do not contribute. These states have to be combined with the unoccupied states ($\mathbf{k}m'\uparrow$)---the same combination as before---to yield a contribution to the diagonal element $\Sigma_{\mathbf{k}m'}^\uparrow(\omega)=\langle\varphi_{\mathbf{k}m'}^{\uparrow}|\Sigma^{\uparrow}(\omega)|\varphi_{\mathbf{k}m'}^{\uparrow}\rangle$ with poles at frequencies below the Fermi energy. Again, these poles do not play an important role for the renormalization of the unoccupied states ($\mathbf{k}m'\uparrow$). The Goldstone contribution to the $GT$ self-energy is thus expected to be small. Of course, the spin-wave peaks at finite $\mathbf{k}$ and $\omega$ can contribute sizably to the self-energy, and we will see that these collective excitations can give rise to anomalies in the electronic band dispersions. \section{Implementation} \label{sec::Implementation} We have implemented the $GT$ self-energy into the $GW$ code \textsc{spex}~\cite{friedrich10}. The self-consistent DFT calculations are carried out with the \textsc{fleur} code~\cite{fleur}. Both codes rely on the all-electron full-potential linearized augmented-plane-wave method, which provides an accurate basis set for the representation of both the itinerant $s$ and $p$ states as well as the localized $d$ states. Solving the BSE [Eq.~(\ref{eq::TmatrixBSE})] leads to an infinite series of ladder diagrams. This series appears also in the diagrammatic expansion of the transverse magnetic response function (then starting from the first order in $W$), for which we reported an implementation in the Refs.~\onlinecite{sasioglu10}-\onlinecite{mueller16}. The implementation exploits the fact that the Hamiltonian is not explicitly time dependent. In addition, the RPA screened interaction $W(\mathbf{r},\mathbf{r}';\omega)$ is approximated by its static limit $W(\mathbf{r},\mathbf{r}')=W(\mathbf{r},\mathbf{r}';0)$. As a result, the self-energy [Eq.~(\ref{eq::ElMagSelfEnergy})] involves only a single frequency integration. Furthermore, we restrict ourselves to calculating only the diagonal elements $\Sigma^\sigma_{\mathbf{q}m}=\langle \varphi_{\mathbf{q}m}^{\sigma}| \Sigma^{\sigma}(\omega)|\varphi_{\mathbf{q}m}^{\sigma}\rangle$ of the self-energy \begin{eqnarray} \Sigma^\sigma_{\mathbf{q}m}(\omega)= -\frac{\textrm{i}}{2\pi} \int \textrm{d}^3 r_1\, \textrm{d}^3 r_2\, \textrm{d}^3 r_3\, \textrm{d}^3 r_4 \,\, \varphi_{\mathbf{q}m}^{\sigma *}(\mathbf{r}_1) \varphi_{\mathbf{q}m}^{\sigma}(\mathbf{r}_{3}) \nonumber\\ \times\int_{-\infty}^{\infty} T^{\sigma\sigma^{\prime}}(\mathbf{r}_{1},\mathbf{r}_{2};\mathbf{r}_{3},\mathbf{r}_{4};\omega^{\prime}) G^{\sigma^{\prime}}(\mathbf{r}_{2},\mathbf{r}_{4};\omega-\omega^{\prime} \textrm{d}\omega^{\prime}\,\, \label{GTimp} \end{eqnarray} with the momentum $\mathbf{q}$, the band index $m$, and the spin $\sigma$ of the Bloch state $\varphi_{\mathbf{q}m}^{\sigma}(\mathbf{r})$. The short-range behavior of the screened interaction $W$ in metals allows an on-site approximation to be employed \cite{sasioglu10,friedrich14,mueller16,friedrich18}: The electron and hole are assumed to occupy the same lattice site when they interact with each other. To realize a lattice-site resolution, we formulate the theory with the help of Wannier functions \begin{equation} w_{\mathbf{R} n}^{\sigma} (\mathbf{r}) = \frac{1}{N}\sum_{\mathbf{k}} e^{-\textrm{i}\mathbf{k}\mathbf{R}} \sum_{m} U_{\mathbf{k}m,n}^{\sigma} \varphi_{\mathbf{k}m}^{\sigma}(\mathbf{r}), \end{equation} where $\mathbf{R}$ is the lattice site, $n$ is an orbital index, and $N$ is the number of $\mathbf{k}$ points. The complex expansion coefficients $U_{\mathbf{k}m,n}^{\sigma}$ of the $n$th Wannier orbital are given by projection of the eigenstates onto a muffin-tin basis function with suitable orbital character and a subsequent L\"owdin orthonormalization \cite{freimuth08}. The Wannier representation of the $T$ matrix reads \begin{eqnarray} T^{\sigma\sigma^{\prime}}(\mathbf{r}_{1},\mathbf{r}_{2};\mathbf{r}_{3},\mathbf{r}_{4};\omega) = \sum_{\mathbf{R},\mathbf{R}^{\prime}} w_{\mathbf{R} n_{1}}^{\sigma} (\mathbf{r}_{1}) w_{\mathbf{R} n_{2}}^{\sigma^{\prime} *} (\mathbf{r}_{2}) \nonumber \\ \times T^{\sigma\sigma^{\prime}}_{\mathbf{R}n_{1}\mathbf{R}n_{2},\mathbf{R}^{\prime}n_{3}\mathbf{R}^{\prime}n_{4}}(\omega) w_{\mathbf{R}^{\prime} n_{3}}^{\sigma *} (\mathbf{r}_{3}) w_{\mathbf{R}^{\prime} n_{4}}^{\sigma^{\prime} } (\mathbf{r}_{4}). \end{eqnarray} Due to the lattice periodicity, the coefficients $T^{\sigma\sigma^{\prime}}_{\mathbf{R}n_{1}\mathbf{R}n_{2};\mathbf{R}^{\prime}n_{3}\mathbf{R}^{\prime}n_{4}}(\omega)$ depend only on the difference vector $\Delta\mathbf{R}=\mathbf{R}-\mathbf{R'}$. A lattice Fourier transformation then yields \begin{equation} T^{\sigma\sigma'}_{n_1n_2,n_3n_4}(\mathbf{k},\omega)=\sum_{\Delta\mathbf{R}}T^{\sigma\sigma'}_{n_1n_2,n_3n_4;\Delta\mathbf{R}}(\omega)e^{-\mathrm{i}\mathbf{q}\Delta\mathbf{R}}. \end{equation} Employing the Lehmann representation of the Green function \begin{equation}\label{Lehmann0} G^{\sigma}(\mathbf{r},\mathbf{r}^{\prime};\omega) = \frac{1}{N} \sum_{\mathbf{k}}\sum_{m} \frac{\varphi_{\mathbf{k}m}^{\sigma}(\mathbf{r}) \varphi_{\mathbf{k}m}^{\sigma *}(\mathbf{r}^{\prime})} {\omega - \epsilon_{\mathbf{k}m}^{\sigma} + \textrm{i}\eta\,\textrm{sgn}(\epsilon_{\mathbf{k}m}^{\sigma}-\epsilon_{\textrm{F}})} \end{equation} with the energies $\epsilon_{\mathbf{k}m}^{\sigma}$ of the corresponding Bloch states, the Fermi energy $\epsilon_{\textrm{F}}$, and a positive infinitesimal $\eta$, one obtains for the diagonal element of the $GT$ self-energy \begin{eqnarray}\label{eq::DiagonalSelfEnergyRealFrequency} \Sigma_{\mathbf{q}m}^{\sigma}(\omega) &=& -\frac{\textrm{i}}{2\pi} \int_{-\infty}^{\infty}\textrm{d}\omega^{\prime}\;\sum_{\mathbf{k}} \sum_{n_{1},n_{2},n_{3},n_{4}} T_{n_{1}n_{2},n_{3}n_{4}}^{\sigma\sigma^{\prime}}(\mathbf{k},\omega^{\prime}) \nonumber \\ &\times& \sum_{m'} \frac{ U_{\mathbf{q}m,n_{1}}^{\sigma} U_{\mathbf{q}-\mathbf{k}m',n_{2}}^{\sigma^{\prime} *} U_{\mathbf{q}m,n_{3}}^{\sigma *} U_{\mathbf{q}-\mathbf{k}m',n_{4}}^{\sigma^{\prime}} } { \omega-\omega^{\prime}-\epsilon_{\mathbf{q}-\mathbf{k}m'}^{\sigma^{\prime}} + \textrm{i}\eta\,\textrm{sgn}(\epsilon_{\mathbf{q}-\mathbf{k}m'}^{\sigma^{\prime}}-\epsilon_{\textrm{F}}) }, \end{eqnarray} where the summation over $m'$, in principle, runs over the infinitely many single-particle eigenstates, but in the present case the summation is effectively restricted to the ones which have been used for the construction of the Wannier basis. The $T$ matrix exhibits poles at the spin excitation energies along the real-frequency axis. The collective spin-wave excitations produce particularly strong poles at low energies, which complicates a straightforward frequency integration. The integration along the real-frequency axis can be avoided by employing the method of analytic continuation \cite{rojas95,rieger99}, which is one of the two methods implemented. To this end, Eq.~(\ref{eq::DiagonalSelfEnergyRealFrequency}) is analytically continued to the imaginary axis and evaluated there for a set of imaginary frequencies forming a mesh along this axis. This is achieved by replacing $\omega \to \textrm{i}\omega$ and $\omega' \to \textrm{i}\omega'$, formally changing the prefactor $-\textrm{i}/(2\pi)$ to $1/(2\pi)$. In this way, one avoids the real-frequency axis where the quantities show strong variations. Along the imaginary axis, the functions are much smoother, which enables an accurate sampling and interpolation of the functions with relatively coarse frequency meshes, also simplifying the frequency convolution of Eq.~(\ref{eq::DiagonalSelfEnergyRealFrequency}). At the end of the calculation, the self-energy has to be analytically continued back to the real-frequency axis. We employ Pad\'{e} approximants for the $T$ matrix (Thiele's reciprocal difference method \cite{baker10}), which allows an analytic frequency convolution with $G$, and also for the self-energy, yielding the self-energy on the whole complex frequency plane. In this approach, the self-energy is effectively represented by a sum over poles with analytically determined complex positions and weights. However, it is well known that this extrapolation can lead to spurious features in the self-energy if one of the effective poles happens to lie close to the real frequency axis. \cf{ Therefore, we have also implemented the contour deformation technique to evaluate the self-energy. This method is more accurate than analytic continuation but also requires more computation time. Details of both implementations will be presented elsewhere \cite{friedrich19} together with a new tetrahedron method, with which the two methods yield nearly identical results. There, we also discuss a third method that could be described as a hybrid of the two. The contour deformation technique evaluates the frequency integral of Eq.~(\ref{eq::DiagonalSelfEnergyRealFrequency}) explicitly, yielding the self-energy directly for real frequencies. The integration path, however, does not run along the real frequency axis from $-\infty$ to $\infty$ but along a deformed integration contour: from $-\mathrm{i}\infty$ to $\mathrm{i}\infty$ and, at halfway, taking a \emph{rectangular detour} at $\omega=0$ to capture all residues of the Green function. (The integrations from $-\infty$ to $-\mathrm{i}\infty$ and from $\mathrm{i}\infty$ to $\infty$ do not contribute.) The resulting self-energy can be written as a sum of two terms \begin{eqnarray} \label{contdef} \lefteqn{\langle\varphi^\sigma_{\mathbf{q}m}\left|\Sigma^\sigma(\omega)\right|\varphi^\sigma_{\mathbf{q}m}\rangle} \\ \nonumber &=& \frac{1}{N} \sum_\mathbf{k} \left [ \sum_{m'} \frac{1}{2\pi} \int_{-\infty}^{\infty}d\omega'\frac{T^{\sigma\sigma'}_{\mathbf{q}mm'}(\mathbf{k},-\mathrm{i}\omega')}{\omega+\mathrm{i}\omega'-\epsilon^{\sigma'}_{\mathbf{q-k}m'}} \right . \\ &+& \left . \left\{ \begin{array}{rr} {\displaystyle \sum_{\substack{m' \\ \omega<\epsilon^{\sigma'}_{\mathbf{q-k}m'}<\epsilon_\mathrm{F}}}} T^{\sigma\sigma'}_{\mathbf{q}mm'}(\mathbf{k},\omega-\epsilon^{\sigma'}_{\mathbf{q-k}m'}) & \textrm{ for }\omega<\epsilon_{\mathrm{F}}\\ -{\displaystyle \sum_{\substack{m' \\ \epsilon_\mathrm{F}<\epsilon^{\sigma'}_{\mathbf{q-k}m'}<\omega}}} T^{\sigma\sigma'}_{\mathbf{q}mm'}(\mathbf{k},\omega-\epsilon^{\sigma'}_{\mathbf{q-k}m'}) & \textrm{ for }\omega>\epsilon_{\mathrm{F}}\end{array}\right.\right ] \,, \nonumber \end{eqnarray} where $T^{\sigma\sigma'}_{\mathbf{q}mm'}(\mathbf{k},\omega)$ denotes the sum of the products $TUUUU$ over $n_1$, $n_2$, $n_3$, and $n_4$. The $\mathbf{k}$ summation should be understood as a $\mathbf{k}$ integration. It is important to perform this integration accurately. We have devised a new tetrahedron method for this purpose \cite{friedrich19}. All results presented in this paper have been calculated with the contour deformation technique using Eq.~(\ref{contdef}). } The evaluation of Eq.~(\ref{contdef}) can be expensive. Therefore, we exploit a few symmetries to accelerate the computation. The first is a symmetry in frequency space which the $T$ matrix inherits from the propagator $K$ \cite{friedrich14}, $T_{n_{1}n_{2},n_{3}n_{4}}^{\sigma\sigma^{\prime}}(\mathbf{k},-\mathrm{i}\omega) =T_{n_{3}n_{4},n_{1}n_{2}}^{\sigma\sigma^{\prime}*}(\mathbf{k},\mathrm{i}\omega)$. This leads to $T_{\mathbf{q}mm'}^{\sigma\sigma'}(\mathbf{k},-\mathrm{i}\omega)= T_{\mathbf{q}mm'}^{\sigma\sigma'*}(\mathbf{k},\mathrm{i}\omega)$. Hence, we can restrict the frequency mesh to the positive imaginary axis and utilize this symmetry for the frequency integration from $-\mathrm{i}\infty$ to $\mathrm{i}\infty$. \cf{Furthermore, it follows from this symmetry that the first term of Eq.~(\ref{contdef}) is real-valued. The second term is complex in general, which will become important in Sec.~\ref{sec:causality}.} We also exploit spatial symmetry in the evaluation of the $\mathbf{k}$ summations. Equation (\ref{eq::DiagonalSelfEnergyRealFrequency}) contains a sum over all $\mathbf{k}$ vectors. We restrict this summation to the extended irreducible Brillouin zone (corresponding to the irreducible zone that is created by the subset of symmetry operations that leave the $\mathbf{q}$ vector invariant \cite{friedrich10}, the so-called little group). The contribution of the symmetry-equivalent $\mathbf{k}$ points is obtained by a symmetrization procedure employing the symmetry transformation matrices of the Wannier set \cite{friedrich14}. In the same way, we can accelerate the computation of the two-particle propagator [Eq.~(\ref{K})], which involves a $\mathbf{k}$ summation, as well \cite{friedrich14}. As already mentioned above, we presently refrain from combining Eq.~(\ref{eq::ElMagSelfEnergy}) with the fully dynamical $GW$ self-energy. Such a study is deferred to a later work. Instead, we combine it with the LSDA solution whose exchange splitting has been corrected in such a way that the Goldstone condition (the spin-wave energy must vanish in the limit $\mathbf{k}\rightarrow 0$) is fulfilled. To be more precise, we introduce a parameter $\Delta_\mathrm{x}$, with which the spin-up and spin-down states are shifted with respect to each other, i.e., $\epsilon^{\uparrow/\downarrow}_{\mathbf{k}m}\rightarrow \epsilon^{\uparrow/\downarrow}_{\mathbf{k}m} \pm \Delta_\mathrm{x}/2$. (A positive $\Delta_\mathrm{x}$ decreases the exchange splitting.) The parameter $\Delta_\mathrm{x}$ is varied until the spin-wave excitation energy vanishes in the long-wavelength limit. The so-corrected single-particle system has been shown~\cite{mueller16} to be similar to a self-consistent solution obtained with the COHSEX self-energy~\cite{hedin65}, the static limit of $GW$. The corrected LSDA Green function can thus be understood as an approximation to the COHSEX Green function, which is then further renormalized with the $GT$ self-energy. We introduce another parameter to achieve partial self-consistency in the Green function. The application of the $GT$ self-energy does not only introduce lifetime broadening effects but also leads to an energetic shift of the electronic bands. In general, this entails that the Fermi energy has to be readjusted to ensure particle number conservation. Such a readjustment, however, has some undesired consequences. First, the spectral function vanishes at the old Fermi energy but not at the new one. Second and more seriously, spectral features, e.g., kinks in the quasiparticle bands, might end up at a wrong binding energy, in particular, if their energetic position is of the same order of magnitude as the shift of the Fermi energy. In an extreme case, the binding energy might even change sign. Clearly, in a fully self-consistent $GT$ calculation, the Fermi energy would converge eventually, and these problems would be solved. However, presently we cannot afford such a calculation. Therefore, we employ a correction that was originally proposed by Hedin \cite{hedin65} and discussed in detail in Refs.~\onlinecite{schindlmayr97} and \onlinecite{pollehn98}. In a one-shot calculation, there is a mismatch between the chemical potential of the non-interacting system, inherited by the self-energy, and the chemical potential of the renormalized Green function. In order to align the two potentials, one introduces an energy shift $\Delta_v$ in $\Sigma(\omega) \rightarrow \Sigma(\omega-\Delta_v)$, where $\Delta_v$ is chosen in such a way that the Fermi energy remains unchanged by the self-energy renormalization, giving the spectral function as \begin{equation} S^\sigma(\mathbf{k},\omega)=\frac{\mathrm{sgn}(\epsilon_\mathrm{F}-\omega)}{\pi} \sum_m\textrm{Im}\{\omega-\epsilon^\sigma_{\mathbf{k}m}-\Sigma^\sigma_{\mathbf{k}m}(\omega-\Delta_v)\}^{-1}. \label{Dyson0} \end{equation} Since the corrected energies approximate COHSEX quasiparticle energies and COHSEX does not share any diagrams with $GT$, no subtraction of the expectation value of the exchange-correlation potential as in one-shot $GW$ calculations is needed in this case. \cf{The spectral function in Eq.~(\ref{Dyson0}) is defined as the trace over the imaginary part of the renormalized Green function, the solution of the Dyson equation with the $GT$ self-energy. The sign is chosen such that the spectral function is non-negative for all $\omega$ and $\mathbf{k}$ (see Sec.~\ref{sec:causality}).} The correction is tantamount to shifting the exchange-correlation potential by the same amount $v_\mathrm{xc}^\sigma \rightarrow v_\mathrm{xc}^\sigma+\Delta_v$. In this way, the non-interacting reference system already contains information about the renormalized system through the shift $\Delta_v$, which can be regarded as some level of self-consistency. (Since the Fermi energy depends on the quasiparticle energies in the whole irreducible wedge of the Brillouin zone, the shift has to be calculated self-consistently until the Fermi energy and $\Delta_v$ converge.) This approach solves the aforementioned two problems in a natural way. The modification of $v^\sigma_\mathrm{xc}$ changes the single-particle energies accordingly so that, in summary, the latter are adjusted by $\epsilon^{\uparrow/\downarrow}_{\mathbf{k}m}\rightarrow \epsilon^{\uparrow/\downarrow}_{\mathbf{k}m} + \Delta_v \pm \Delta_\mathrm{x}/2$, which shifts the self-energy argument in the opposite direction, effectively undoing the shift from before, $\Sigma(\omega-\Delta_v)\rightarrow\Sigma(\omega)$. The $\Delta_v$ correction does not affect the Goldstone condition, which has been used to fix $\Delta_\mathrm{x}$, since only the differences of the single-particle states (of opposite spins) enter the evaluation of the spin excitations energies, and not their absolute values. We note that the two parameters, $\Delta_\mathrm{x}$ and $\Delta_v$, are determined from exact physical constraints, so they do not imply a deviation from the ``ab initio-ness'' of the method. With the aligned chemical potential, Eq.~(\ref{Dyson0}) can alternatively be written as \begin{equation} S^\sigma(\mathbf{k},\omega)=\frac{\mathrm{sgn}(\epsilon_\mathrm{F}-\omega)}{\pi} \sum_m\textrm{Im}\{\omega-\epsilon^\sigma_{\mathbf{k}m}-\Sigma^\sigma_{\mathbf{k}m}(\omega)+\Delta_v\}^{-1}, \label{Dyson} \end{equation} where the single-particle energies $\epsilon^\sigma_{\mathbf{k}m}$ are assumed to contain the two parameters $\Delta_\mathrm{x}$ and $\Delta_v$ as described above. (We note that the parameter $\Delta_v$ in Eq.~(\ref{Dyson}) seems to have no effect because the term $-\Delta_v$ de facto cancels the $\Delta_v$ in $\epsilon^\sigma_{\mathbf{k}m}$, but it should be remembered that $\Delta_v$ shifts the Fermi energy, too. We also note that it might be surprising that a spin-independent shift of the exchange-correlation potential and, thus, of the Kohn-Sham eigenvalues should have an effect at all, but one has to remember that quasiparticle energies correspond to total-energy differences and are therefore defined absolutely.) \cf{ \section{Violation of causality}\label{sec:causality} In this section, we present a detailed discussion about the violation of causality mentioned in the introduction. To this end, we have to resort to a general second-quantization formulation of the Green function and related response quantities. The Bloch $\mathbf{k}$ crystal momentum is suppressed for simplicity. The time-ordered Green function \begin{equation}\label{G} G^\sigma(12)=-\mathrm{i} \langle\Psi_0|{\cal T}[\hat{\psi}^\sigma(1)\hat{\psi}^{\sigma\dagger}(2)]|\Psi_0\rangle \end{equation} comprises both the electron and the hole propagator by virtue of the time-ordering operator $\cal T$: If $t_1>t_2$, the Heisenberg creation operator $\hat{\psi}^{\sigma \dagger}(2)$ acts first on the many-body ground state $|\Psi_0\rangle$ creating an electron at 2, which is later annihilated at 1. If otherwise $t_1<t_2$, then the operator $\cal T$ switches the order of the field operators and lets $\hat{\psi}^\sigma(1)$ act first, thus creating a hole at 1, which is later filled by an electron at 2. So, the Green function describes the propagation of an electron or a hole depending on the time order of the arguments $t_1$ and $t_2$. The Fourier transformation of Eq.~(\ref{G}) ($t_1-t_2\rightarrow\omega$) yields the Lehmann representation \footnote{To sketch the derivation, which can be found in many textbooks, e.g., Ref.~\onlinecite{mahan00}, we write Eq.~(\ref{G}) as the sum of two terms corresponding to the cases $t_1>t_2$ and $t_1<t_2$ with the help of the Heaviside function $\theta(t_1-t_2)=1$ for $t_1>t_2$ and 0 otherwise. \cf{(It should be noted that, in the fermionic case, the operator $\cal{T}$ is defined to introduce a factor $-1$ when it switches the field operators.)} Then, inserting the closure relation $\sum_m |\Psi_m^{\sigma\pm}\rangle\langle\Psi_m^{\sigma\pm}|=\hat{1}$ between the two field operators in each term, using $\hat{\psi}^\sigma(\mathbf{r}t)=e^{\mathrm{i}\hat{H}t}\hat{\psi}^\sigma_\mathrm{S}(\mathbf{r})e^{-\mathrm{i}\hat{H}t}$ and $\int\theta(t)e^{\mathrm{i}\omega t}dt=\mathrm{i}/(\omega+\mathrm{i}\eta)$ yields Eq.~(\ref{Lehmann}).} \begin{eqnarray} \nonumber G^{\sigma}(\mathbf{r}_1,\mathbf{r}_2;\omega) &=& \sum_m \frac{ \langle\Psi_0|\hat{\psi}_\mathrm{S}^\sigma(\mathbf{r}_1)|\Psi_{m}^{\sigma+}\rangle \langle\Psi_{m}^{\sigma+}|\hat{\psi}_\mathrm{S}^{\sigma\dagger}(\mathbf{r}_2)|\Psi_0\rangle } {\omega - ( E_{m}^{\sigma+} - E_0 ) + \textrm{i}\eta} \\ &+&\sum_m \frac{ \langle\Psi_{m}^{\sigma-}|\hat{\psi}_\mathrm{S}^\sigma(\mathbf{r}_1)|\Psi_0\rangle \langle\Psi_0|\hat{\psi}_\mathrm{S}^{\sigma\dagger}(\mathbf{r}_2)|\Psi_{m}^{\sigma-}\rangle } {\omega - ( E_0 - E_{m}^{\sigma-} ) - \textrm{i}\eta} \,, \label{Lehmann} \end{eqnarray} where $\eta$ is a positive infinitesimal and $\{|\Psi_m^{\sigma\pm}\rangle,E_m^{\sigma\pm}\}$ is a complete set of energy eigensolutions of the many-body system with one spin-$\sigma$ electron more ($+$) or less ($-$) than $|\Psi_0\rangle$. We now consider a diagonal element $G^\sigma_m(\omega)\coloneqq\langle\varphi_m^\sigma|G^\sigma(\omega)|\varphi_m^\sigma\rangle$ with $\varphi_m^\sigma(\mathbf{r})$ an eigenstate of the non-interacting reference system. Since the Schr\"odinger field operators $\hat{\psi}^\sigma_\mathrm{S}(\mathbf{r})$ and $\hat{\psi}^{\sigma\dagger}_\mathrm{S}(\mathbf{r})$ are adjoints of each other, the numerators, when projected onto $\varphi_m^\sigma(\mathbf{r})$, are real and non-negative. Furthermore, $E_0 - E_{m}^{\sigma-} \le \mu \le E_{m}^{\sigma+} - E_0$ with the chemical potential $\mu$. Then, it follows from $(\omega\pm\mathrm{i}\eta)^{-1}={\cal P}(1/\omega)\mp\mathrm{i}\pi\delta(\omega)$ that the imaginary part of $G^\sigma_m(\omega)$ must have a positive (negative) sign for $\omega<\mu$ ($\omega>\mu$). As $\eta$ enters Eq.~(\ref{Lehmann}) as a consequence of the time-ordering operator, we can trace the condition on the sign of $\mathrm{Im}G^\sigma_m(\omega)$ back to the correct time order of the field operators and thus to the causality. The mathematical formulas derived above bear similarity with those encountered in the treatment of the \emph{damped driven harmonic oscillator}. In this context, the energy differences would correspond to the resonance frequency of the free oscillator and $\eta$ could be interpreted as a damping coefficient. The latter appears linearly in the imaginary part of the free oscillator's frequency. So, a wrong sign in the imaginary part presupposes a wrong sign in $\eta$, which would thus have the effect of amplifying the oscillation rather than damping it. This solution could be described as one that goes backwards in time, thus violating causality. Through the Dyson equation, the imaginary part of $G^\sigma_m(\omega)$ can be related to the imaginary part of the self-energy \begin{equation} \label{GreenSigma} \mathrm{Im}G^\sigma_m(\omega)=\frac{\mathrm{Im}\Sigma^\sigma_m(\omega)}{[\omega-\epsilon_m^\sigma-\mathrm{Re}\Sigma^\sigma_m(\omega)]^2+[\mathrm{Im}\Sigma_m^\sigma(\omega)]^2}\,, \end{equation} where it has been used that Eq.~(\ref{Lehmann}) is trivially valid for the non-interacting $G$, too, in which case $E_{m}^{\sigma+} - E_0$ and $E_0 - E_{m}^{\sigma-}$ correspond to the single-particle energies $\epsilon_m^\sigma$ and the numerators are just products of single-particle wave functions $\varphi_m^\sigma(\mathbf{r}_1)\varphi_m^{\sigma*}(\mathbf{r}_2)$. Equation (\ref{Lehmann0}) shows such a Lehmann representation formulated with Bloch vectors. We have neglected the offdiagonal elements of $\Sigma^\sigma(\omega)$ for simplicity. According to Eq.~(\ref{GreenSigma}), the sign rule for $\mathrm{Im}G^\sigma_m(\omega)$ applies equally to $\mathrm{Im}\Sigma^\sigma_m(\omega)$, also see Eq.~(9.34) of \cite{FetterWalecka71}. In the last section, it was shown [Eq.~(\ref{contdef})] that the imaginary part of $\Sigma^\sigma_m(\omega)$ is completely determined by (a $\mathbf{k}$ integration of) the imaginary part of $T$-matrix elements, represented in terms of single-particle eigenfunctions. From Eq.~(\ref{contdef}), it follows that the condition on the sign of $\Sigma_m^\sigma(\omega)$ formulated above requires the imaginary part of the $T$ matrix elements to be non-negative for all $\omega$. The $T$ matrix, as we defined it above, contains all ladder diagrams starting from third order. In this sense, we can express it as \begin{equation} \label{T2} T^{\sigma\sigma^\prime} = T_{\ge 2}^{\sigma\sigma^\prime} - T_2^{\sigma\sigma^\prime}\,, \end{equation} where $T_2$ is the bare second-order ladder diagram and $T_{\ge 2}$ includes all ladder diagrams starting from second order. This reformulation looks trivial (and it is), but it will help us to understand the reason for the violation of causality. As a first step, we realize that all ladder diagrams are flanked from either side by an interaction line $W$, which can be factored out giving \begin{equation} \label{TinR} T^{\sigma\sigma^\prime}(12,34) = W(12) [ R^{(4)\sigma\sigma^\prime}(12,34) - K^{\sigma\sigma^\prime}(12,34) ] W(34)\,. \end{equation} Here, $K$ is the free electron-hole propagator of Eq.~(\ref{K}) and $R^{(4)}$ the interacting electron-hole propagator, which is related to the generalized four-point response function by $R=-2R^{(4)}$. Both quantities can be expressed in terms of a spin-spin correlation function \cite{friedrich18} $\mathrm{i}\langle\Psi_0|{\cal T}[ \hat{\psi}^{\sigma \dagger}(2) \hat{\psi}^{\sigma'}(1) \hat{\psi}^{\sigma'\dagger}(3) \hat{\psi}^ \sigma (4) ]|\Psi_0\rangle$. (Note the assumption $\sigma'=-\sigma$.) For the free (interacting) propagator $K$ ($R^{(4)}$), one would take $|\Psi_0\rangle$ to be the energy ground state of the non-interacting reference (real interacting) system. The spin-spin correlation function has a similar form as Eq.~(\ref{G}) and can be Fourier-transformed analogously giving \begin{eqnarray} \label{RLehmann} \lefteqn{R^{(4)\sigma\sigma'}(\mathbf{r}_1,\mathbf{r}_2,\mathbf{r}_3,\mathbf{r}_4;\omega)}\\ \nonumber &=& \sum_m \frac{ \langle\Psi_{m}| \hat{\psi}_\mathrm{S}^{\sigma\dagger} (\mathbf{r}_2) \hat{\psi}_\mathrm{S}^{\sigma'} (\mathbf{r}_1) |\Psi_0\rangle \langle\Psi_0| \hat{\psi}_\mathrm{S}^{\sigma'\dagger} (\mathbf{r}_3) \hat{\psi}_\mathrm{S}^{\sigma} (\mathbf{r}_4) |\Psi_{m}\rangle } {\omega - ( E_0 - E_{m} ) - \textrm{i}\eta} \\ &-& \sum_m \frac{ \langle\Psi_0| \hat{\psi}_\mathrm{S}^{\sigma\dagger} (\mathbf{r}_2) \hat{\psi}_\mathrm{S}^{\sigma'} (\mathbf{r}_1) |\Psi_{m}\rangle \langle\Psi_{m}| \hat{\psi}_\mathrm{S}^{\sigma'\dagger} (\mathbf{r}_3) \hat{\psi}_\mathrm{S}^{\sigma} (\mathbf{r}_4) |\Psi_0\rangle } {\omega - ( E_{m} - E_0 ) + \textrm{i}\eta} \,, \nonumber \end{eqnarray} where $\{|\Psi_m\rangle,E_m\}$ is a complete set of energy eigensolutions of the many-body system (with the same number of electrons as $|\Psi_0\rangle$). Inserting this expression and the corresponding one for $K$ (which could be simplified again using the single-particle eigenstates, which is, however, not necessary for the present discussion) into Eq.~(\ref{TinR}) yields the spectral representations of the two contributions $T_{\ge 2}$ and $T_2$. The $T$ matrix elements relevant for the self-energy are obtained by multiplying with two wave-function products, $\varphi_{m'}^{\sigma'*}(\mathbf{r}_2)\varphi_{m'}^{\sigma'}(\mathbf{r}_4)$ and $\varphi_m^{\sigma*}(\mathbf{r}_1)\varphi_m^\sigma(\mathbf{r}_3)$, and integrating over all space coordinates. The first product originates from the Green function in Eq.~(\ref{eq::ElMagSelfEnergy}) and the second from evaluating the diagonal element of the self-energy $\Sigma_m^\sigma(\omega)$, also see Eq.~(\ref{GTimp}). In analogy to the case of the Green function, this leads to numerators in the two terms $T_{\ge 2}$ and $T_2$ that are all real and non-negative. It is then straightforward to deduce from Eq.~(\ref{RLehmann}) that the imaginary parts of both $T_{\ge 2}$ and $T_2$ are non-negative in the whole frequency range. According to Eq.~(\ref{contdef}), the imaginary part of the self-energy contributions originating from $T_{\ge 2}$ and $T_2$ will both have the correct sign for $\omega>\mu$ and $\omega<\mu$ as a consequence. Here, the decisive point is that they have the correct sign \emph{individually}, but their difference [Eq.~(\ref{TinR})], might have a wrong sign in some frequency region and thus violate causality. This is not just a theoretical possibility but rather an expected behavior due to a redistribution of spectral weight in the renormalized pair propagator $R^{(4)}$ with respect to $K$, the free pair propagator. The imaginary part of the latter is formed by the spin-flip Stoner excitations, whereas the Bethe-Salpeter renormalization gives rise to the formation of collective excitations (spin waves) in $R^{(4)}$ at low energies. Spectral weight thus flows from higher to lower energies so that the difference $\mathrm{Im}(T_{\ge 2}-T_2)$ is expected to become negative at some energy. We indeed observe this violation of causality in our calculations but have chosen to accept it, as the sign change occurs at an energy of several eV, where the lifetime broadening has already become quite small and where diagrams neglected in the present work should become more relevant and recover the correct sign. \begin{figure} \includegraphics[width=0.49\columnwidth]{spec141414_contour_sign-up.png} \includegraphics[width=0.49\columnwidth]{spec141414_contour_sign-down.png} \caption{\label{fig::causal} Map of regions where the spectral function $S^\sigma(\mathbf{k},\omega)$ [(a) $\sigma=\uparrow$, (b) $\sigma=\downarrow$] of iron has positive ($+$) and negative ($-$) sign. The appearance of the negative sign implies a violation of causality, see text. The energy zero is set at the Fermi energy. } \end{figure} Figure \ref{fig::causal} shows a map of the sign of the spectral function $S^\sigma(\mathbf{k},\omega)$ as defined in Eq.~(\ref{Dyson}) for the two spin channels in the case of iron. The causality condition demands the spectral function to be non-negative everywhere, but we do see several regions where the sign is, in fact, negative. One notices a spin asymmetry: the wrong sign appears at negative energies in the spin-up channel, whereas it is seen at positive energies in the spin-down channel. As will be explained in the next section, this spin selectivity is due to the fact that a coupling to spin waves is possible only for majority holes (occupied states) and minority electrons (unoccupied states), which confirms our explanation that the causality violation originates from a shift of spectral weight from single-particle Stoner excitations to the low-energy spin waves, leading to the appearance of the wrong sign in the difference $\mathrm{Im}(T_{\ge 2}-T_2)$. Both terms individually have the correct sign. In fact, if we perform the calculation with $T_{\ge 2}$ alone (instead of the difference), the sign of the spectral function never becomes negative. However, such a calculation exhibits extreme and unphysical renormalization effects. The lifetime broadenings in the regions with negative sign must be considered incorrect. These regions are, however, relatively far away from the Fermi energy, below $-3.5$ eV and above $2.5$ eV for spin up and down, respectively, and the renormalizations and lifetime broadenings close to the Fermi energy, where the $GT$ self-energy diagrams dominate, should show the correct behavior. Clearly, if all infinitely many diagrams were included in the self-energy, the causality condition should be fulfilled. In this sense, the violation of causality is an effect of missing self-energy diagrams. For example, for larger absolute energies it is expected that the imaginary part of the $GW$ diagram, neglected in the present work, becomes more important. In fact, there is convincing argument that already the $GW$ self-energy term that is second order in $v$ restores the correct sign: This term is formally similar to the self-energy contribution derived from $T_2$ and $K$ [Eq.~(\ref{TinR})], except that the multiplication from both sides is with $v$ instead of $W$. Since $v(\mathbf{r},\mathbf{r}')>W(\mathbf{r},\mathbf{r}';\omega=0)$ for (nearly) all $\mathbf{r}$ and $\mathbf{r}'$, it seems natural to expect the corresponding self-energy contribution (also its imaginary part) to be larger than the one from $T_2$. The combination of $GT$ and $GW$ is planned as a future study. The question of whether $GW$ fully restores the correct sign must therefore await clarification until then. } \section{Calculations} \label{sec::Calculations} \begin{figure} \includegraphics[width=1.\columnwidth]{Fe.png} \includegraphics[width=1.\columnwidth]{Co.png} \includegraphics[width=1.\columnwidth]{Ni.png} \caption{\label{fig::bandstr}(Color online) Band structures of (a) Fe, (b) Co, and (c) Ni. The spectral functions [Eq.~(\ref{Dyson})] are shown in light blue to green. The LSDA band structure is shown as red lines for comparison. The energy zero is set at the Fermi energy. The lifetime broadening decreases in the order of increasing $d$ occupancy (Fe$\rightarrow$Co$\rightarrow$Ni). The spin and particle-hole asymmetries in the lifetime broadening are clearly seen. } \end{figure} Figure \ref{fig::bandstr} shows the $\mathbf{k}$- and energy-resolved spectral functions [Eq.~(\ref{Dyson})] in light blue to green together with the LSDA band structure as red lines for bcc Fe, fcc Co, and fcc Ni. The LSDA band structure is the uncorrected one (i.e, $\Delta_\mathbf{x}=\Delta_v=0$), while the non-interacting reference system for the $GT$ renormalization is taken from an LSDA solution with a parameter $\Delta_\mathrm{x}$ to enforce the Goldstone condition ($\Delta_\mathrm{x}=0.10$~eV, 0.36~eV, and 0.17~eV for the three materials, respectively) and a parameter $\Delta_v$ to align the chemical potentials of the non-interacting and interacting systems ($\Delta_v=$190~meV, 76~meV, and 25~meV). \cf{The parameters are discussed in Sec.~\ref{sec::Implementation}.} We employ a Wannier set of $s$, $p$, and $d$ orbital character constructed from the 18 lowest eigenstates in each spin channel. The Brillouin zone has been sampled with a 14$\times$14$\times$14 $\mathbf{k}$-point set. \cf{The $\mathbf{k}$ integrations of Eq.~(\ref{contdef}) are carried out using a newly developed tetrahedron method \cite{friedrich19}. For the first term of Eq.~(\ref{contdef}), we employ a frequency mesh along the imaginary axis containing 20 points from 0 to $2\mathrm{i}$~htr. A Pad\'{e} approximant, determined from these 20 points, yields $T_{\mathbf{q}mm'}^{\sigma\sigma'}(\mathbf{k},\mathrm{i}\omega)$ for the complete imaginary axis (interpolating between the points and extrapolating beyond $2\mathrm{i}$~htr) and enables an analytic integration along $\mathrm{i}\omega$. For the second term of Eq.~(\ref{contdef}), the matrix $T_{\mathbf{q}mm'}^{\sigma\sigma'}(\mathbf{k},\omega)$ is calculated for real frequencies $\omega$ in steps of 20~meV. The self-energy $\Sigma^\sigma_{\mathbf{q}m}(\omega)$ is evaluated on a mesh between $-6$~eV and $4$~eV (with respect to the Fermi energy) with a step size of 50~meV. To achieve very high resolution in energy, necessary to resolve sharp quasiparticle peaks (as found close to the Fermi energy), a spline interpolation is applied to obtain spectral functions in steps of 2~meV. The calculations were carried out on the supercomputer JURECA at the J\"ulich Supercomputing Centre \cite{JSC}. } Looking closely at the regions around the Fermi energy, one realizes that the lifetime broadenings disappear at energy zero and grow rapidly for lower and higher energies. Beyond a binding energy of around 4~eV (negative energies in the diagrams) in the spin-up channel, they decrease again (even to the extent that the sign changes to negative, see Sec.~\ref{sec:causality}.) The reason for this is the restriction of the self-energy to the $GT$ diagram, which is expected to give the low-energy scattering contribution because of the low energies of the magnon excitations contained in the $T$ matrix. At higher binding energies, other diagrams, which are neglected in the present study, should become increasingly important, for example, the $GW$ diagram. \begin{figure} \includegraphics[height=3cm]{SchematicProcesses.pdf} \caption{\label{schema}Schematic illustration of the scattering of majority (left) and minority holes (right) in a ferromagnet. The scattering phase space of the former is much larger than that of the latter.} \end{figure} The lifetime broadening is strongest in bcc Fe and weakens in the order of $d$ occupancy (Fe$\rightarrow$Co$\rightarrow$Ni). This is due to the decrease of the number of available minority conduction states, into which a majority electron can be excited, thus reducing the phase space of spin excitations. We also observe a spin asymmetry of the lifetime effects: the majority valence states are broadened more strongly than the minority ones. This can be attributed to the spin asymmetry of the density of states. The scattering diagram shown in Fig.~\ref{fig::GTSelfEnergy} may be described as an incoming majority hole at 1, which excites an electron-hole pair at 2 in the minority channel. The majority hole and minority electron then propagate through the system, successively scattering with each other until the minority electron-hole pair recombines at 4. The correlated propagation of the electron and the hole (having opposite spins) embodies the spin-flip excitations in the form of Stoner excitations and, due to their correlated motion, also of the collective spin-wave excitations. The probability for the pair formation and the pair propagation process depends on the number of available majority valence and minority conduction states, which is large in this case, see Fig.~\ref{schema}, left. (We note that the states also have similar orbital character.) On the other hand, an incoming minority hole would have to combine with majority conduction states. Naturally, the available phase space is much smaller in this case (Fig.~\ref{schema}, right), which explains the spin asymmetry in the lifetime broadening. The fact that minority bands are visible down to much larger binding energies than majority bands in ARPES is thus revealed to be a lifetime effect caused by electron-electron scattering. It should be noted that the qualitative behavior has already been observed in previous studies based on DMFT \cite{katsnelson99,grechnev07,sanchez-barriga09,sanchez-barriga12} and model calculations \cite{mills99}. Our calculated spectral functions for iron are remarkably similar to the ones published in Ref.~\onlinecite{sanchez-barriga12}. However, the fact that both the magnon excitation spectrum and the $GT$ self-energy are fully $\mathbf{k}$ dependent in our method gives rise to additional fine structure in the spectral functions, e.g., band anomalies, which have not been observed in theoretical studies so far. Apart from the spin asymmetry, there is evidence for a particle-hole asymmetry as well, i.e., a markedly different lifetime broadening above and below the Fermi energy; namely, in the majority channels of all three materials and also in the minority channel of Fe. Here, an explanation very similar to the one just given would predict the lifetime broadening of majority hole (minority electron) bands to be larger than the broadening of majority electron (minority hole) bands, in agreement with the spectra. In iron, and to a lesser degree in Co, the many-body renormalization is so strong that the quasiparticle character is virtually lost between $-2.5$ and $-0.5$~eV in the majority channel. A similar but less pronounced renormalization effect is also seen in the minority channel of iron at around 0.5~eV above the Fermi energy, see below. The $GT$ self-energy thus leads to a substantially more complex spectral function than the quasiparticle band structure of a $GW$ calculation, let alone the mean-field band structure of Kohn-Sham DFT. In the case of the former, one often only plots the energy-renormalized quasiparticle bands and drops the lifetime broadening altogether, which is legitimate because the main self-energy poles of $GW$ are typically energetically far away from the valence and low-lying conduction bands. The quasiparticle approximation is therefore a good one. To be more precise, the main $GW$ self-energy poles describe the coupling of the electrons and holes to plasmons (emission and absorption of plasmons), whose energies $\approx \epsilon_\mathrm{P}$ are typically in the order of tens of eV. The poles, thus, appear well separated from the valence (conduction) band energies $\epsilon^\sigma_{\mathbf{k}m}$ at around $\epsilon^\sigma_{\mathbf{k}m}\mp\epsilon_\mathrm{P}$ for valence and conduction states, respectively. As a result, one obtains a well-defined renormalized quasiparticle band structure and, at a much larger energy, the corresponding plasmon satellites. In the $GT$ self-energy, on the other hand, the plasmon energies $\epsilon_\mathrm{P}$ would have to be replaced by the much smaller magnon energies $\approx \epsilon_\mathrm{M}$. Furthermore, the spin of the Green function in the $GT$ self-energy is opposite to that of the self-energy itself [see Eq.~(\ref{eq::ElMagSelfEnergy})]. So, the $GT$ self-energy poles are to be expected at energies shifted from the \emph{opposite-spin} band energies, i.e., at around $\epsilon^{-\sigma}_{\mathbf{k}m}\mp\epsilon_\mathrm{M}$. In this sense, the magnon energies additionally compete with the exchange splitting. To sum up, the $GT$ self-energy poles will fall straight into the valence (and low-lying conduction) region, giving rise to a much richer photoelectron spectrum than, for example, in $GW$ calculations. The plot of the $\mathbf{k}$- and energy-resolved spectral function allows fine details of the quasiparticle band structure to be investigated. The resonant interaction with many-body states in the valence band region (these form the pole structure of the self-energy) can give rise to anomalies in the band dispersion and satellite bands, as we will discuss exemplarily for a few cases in the following. \cf{ \begin{figure} \includegraphics[width=1.\columnwidth]{anomaly.png} \includegraphics[width=0.5\columnwidth,angle=-90]{anomaly.eps} \caption{\label{fig::anomaly} (Color online) (a) Closeup of dispersion anomalies in a spin-down conduction band of iron. The LSDA bands are shown in red for comparison. (b) Same band renormalized only with the lowest-order diagram, which is of third order in $W$, see Fig.~\ref{fig::SelfEnergyExpansion}, and which cannot describe the scattering of electrons and spin waves. The anomaly disappears. The anomaly in (a) can therefore be attributed to electron--spin-wave scattering. (c) Spectral function (red line) along the white dashed line marked in (a). Real and imaginary parts of the self-energy are shown as blue and green lines, respectively. The spectral-function peaks are roughly at the intersections of the blue curve with a linear function, see text. } \end{figure} We first analyze a band anomaly in a spin-down conduction band of iron about 500~meV above the Fermi energy, see Fig.~\ref{fig::bandstr}(a), right panel. A closeup is shown in Fig.~\ref{fig::anomaly}(a). Whereas the respective Kohn-Sham band (red lines) is continuous and shows an approximately parabolic dispersion, the renormalized quasiparticle band seems to be cut in two: the band breaks off at around 250~meV and reappears at 700~meV with an incoherent continuum in-between. The energetic position of the anomaly, immediately above the Fermi energy, indicates that electronic scattering with low-energy spin waves is responsible for its formation. This is confirmed by Fig.~\ref{fig::anomaly}(b), which shows the band structure renormalized with only the first term of Eq.~(\ref{eq::TmatrixBSE}). This term, the third-order diagram of Fig.~\ref{fig::GTSelfEnergy}, can be regarded as not containing scattering contributions with collective spin-wave excitations, which, in turn, would require an infinite (or at least large) number of ladder diagrams to be summed to become a well defined many-body excitation. Fig.~\ref{fig::anomaly}(b) does not exhibit any band anomaly in the respective bands \footnote{The relatively strong energetic shift results from employing the same parameters $\Delta_\mathrm{x}$ and $\Delta_v$ as for the full $GT$ calculations, without readjusting them to the self-energy presently restricted to the third-order diagram.}. This proves that the scattering with spin waves is responsible for the appearance of the band anomaly. It also explains why DMFT cannot reproduce the anomaly because DMFT spatially confines the bosonic excitations in the impurity, which prevents the spin waves to become well-defined extended many-body excitations, despite the fact that infinitely many ladder diagrams are summed in principle. The spatial confinement in DMFT is equivalent to the missing $\mathbf{k}$ dependence of the bosonic excitation spectrum. In contrast, this $\mathbf{k}$ dependence (and also the one for the self-energy) is fully preserved in our theoretical approach. The electron-magnon interaction gives rise to a double-peak structure in the spectral function at the Bloch momentum of the anomaly [white dashed line in Fig.~\ref{fig::anomaly}(a)]. Figure \ref{fig::anomaly}(c) shows this spectral function as the red curve. The blue and green lines show the real and imaginary components of the self-energy matrix element $\Sigma^\sigma_{\mathbf{k}m}(\omega)=\langle\varphi^\sigma_{\mathbf{k}m}|\Sigma^\sigma(\omega)|\varphi^\sigma_{\mathbf{k}m}\rangle$, respectively, where $\varphi^\sigma_{\mathbf{k}m}$ is the respective Kohn-Sham state. The spectral function clearly exhibits two peaks. In general, peaks in the spectral function are expected where the denominator of Eq.~(\ref{Dyson}) becomes small, i.e., where the linear function $\omega-\epsilon^\sigma_{\mathbf{k}m}+\Delta_v$ becomes identical or, at least, close to the self-energy $\Sigma^\sigma_{\mathbf{k}m}(\omega)$. In fact, the two peaks are seen close to the intersections of the linear function (black dashed line) with the real part of the self-energy (blue line), whereas its imaginary part slightly shifts the peak positions and also suppresses a possible third peak related to the intersection in the middle. [If the self-energy is assumed to be given by a single pole $(\omega-\omega_0)^{-1}$ with a complex self-energy resonance at $\omega_0$, then equating with the linear function would yield a quadratic equation with only two possible solutions.] One might be tempted to identify one of the two peaks as the quasiparticle peak and the other as the magnon satellite, roughly corresponding to the two energies $\epsilon^\sigma_{\mathbf{k}m}$ and $\epsilon^{-\sigma}_{\mathbf{k}m}+\epsilon_{\mathrm{M}}$ discussed above. However, this is not quite correct because it would disregard the coupling between these two fundamental many-body states. Let us consider a simple $2\times 2$ Hamiltonian with the above two energies on the diagonal. The offdiagonal (coupling) matrix elements must be nonzero, otherwise there would not be any renormalization. As the quasiparticle band $\epsilon^\sigma_{\mathbf{k}m}$ disperses upwards in energy, it approaches and transverses the energy $\epsilon^{-\sigma}_{\mathbf{k}m}+\epsilon_{\mathrm{M}}$. The coupling and the strong lifetime effects make the band disappear. When the distance between the two energies has again grown large enough, the quasiparticle band reappears. The two peaks show up when the diagonal elements are roughly identical. The peak separation can then be identified as a measure for the electron-magnon coupling strength (offdiagonal elements). \begin{figure} (a)\includegraphics[width=0.5\columnwidth,angle=-90]{magnaron.eps} (b)\includegraphics[width=0.5\columnwidth,angle=-90]{magnaron1.eps} \caption{\label{fig::magnaron} (Color online) Same as Fig.~\ref{fig::anomaly}(c) for the two secondary bands [(a) upper one, (b) lower one] seen in the spin-up channel of nickel of Fig.~\ref{fig::bandstr}(c). The spectral function and self-energy are plotted for the $\Gamma$ point. } \end{figure} In the renormalized majority band structure of nickel, Fig.~\ref{fig::bandstr}(c), one observes two curious \emph{secondary} bands close to the center of the Brillouin zone, a weak one above the highest valence band and a somewhat stronger one between the highest and the second highest band. The latter one even seems to trace the dispersion of the quasiparticle bands below. Although the features look very different to the band anomaly discussed before, they have a similar origin. One might imagine the two energies $\epsilon^\sigma_{\mathbf{k}m}$ and $\epsilon^{-\sigma}_{\mathbf{k}m}-\epsilon_{\mathrm{M}}$ to be nearly constant with respect to the Bloch vector $\mathbf{k}$. The two peaks, the strong quasiparticle and the weak secondary one, then form two nearly parallel bands with little variation in strength. Figs.~\ref{fig::magnaron}(a) and (b) show the spectral function, together with the self-energy and the linear function $\omega-\epsilon^\sigma_{\mathbf{k}m}+\Delta_v$, at the $\Gamma$ point ($k=0$). In contrast to Fig.~\ref{fig::anomaly}(c), the two spectral-function peaks are of very different strength. The small secondary peak cannot be related to an intersection of the black dashed line with the blue solid one. It rather appears where the two functions have the closest approach. Due to the very different peak strengths, one can interpret the strong peak as the quasiparticle peak and the secondary one as a magnon satellite peak (the corresponding many-body states having small admixtures of each other). It might seem a curiosity that the satellite peak appears \emph{above} the quasiparticle peak instead below, as one is used to from $GW$ calculations. This can be attributed to the exchange splitting, which, as discussed above, competes with the magnon energy and can cause the satellite peak to emerge \emph{on the wrong side}, according to $\epsilon^{-\sigma}_{\mathbf{k}m}+\epsilon_{\mathrm{M}} =\epsilon^{\sigma}_{\mathbf{k}m}-\epsilon_{\mathrm{M}}+\Delta_{\mathbf{k}m}$ with the exchange splitting $\Delta_{\mathbf{k}m}$. } \iffalse \begin{figure} \includegraphics[height=4cm]{Fe_GH.png} \includegraphics[height=4cm]{Co_kink.png}\includegraphics[height=4cm]{Ni_kink.png} \caption{\label{fig::kinks1}Kinks in the dispersion of majority bands close to the Fermi energy caused by the electron-magnon scattering in (a) Fe, (b) Co, and (c) Ni. The LSDA bands are shown in red. } \end{figure} \deleted{ Figure \ref{fig::kinks1} shows selected majority valence bands of Fe, Co, and Ni close to the Fermi energy. While the LSDA bands (red lines) exhibit a very regular parabolic-like behavior, there are distinct signatures of many-body renormalization effects to be seen in the $GT$ bands. We note that these irregularities appear in some but not all $\mathbf{k}$ directions, which demonstrates the importance of the $\mathbf{k}$ dependence of the self-energy. We also see in Fig.~\ref{fig::kinks1}(a) that the effect is band-selective: only two out of three bands are affected, despite the fact that all bands have the same ($d$) orbital character. The kinks are seen at binding energies of about 180~meV and 130~meV for the two bands, respectively. These energies are in the order of magnitude of typical spin-wave energies, which leads one to suspect that these anomalies are caused by a coupling to extended spin-wave excitations. This conjecture is corroborated by a calculation in which the self-energy is restricted to the first diagram of Fig.~\ref{fig::GTSelfEnergy}, the diagram of third order in $W$. Since the spin-wave excitation requires the summation over a large number of diagrams (preferably infinitely many) to become well-defined, it is plausible that the third-order diagram alone cannot describe the collective spin-wave, but it does contain the Stoner excitations. Indeed, the resulting quasiparticle bands do not show any anomalies. We conclude that the kinks are caused by a scattering of the quasiparticle with collective spin-wave excitations. In Co and Ni [Figs.~\ref{fig::kinks1}(b) and (c)], similar kinks are observed, which are, however, somewhat less pronounced than in iron. Their energetic positions at 150~meV and 200~meV, respectively, are similar to the case of Fe. } \deleted{ In fact, there have been ARPES studies that reported kink structures in the band dispersion of iron \cite{schaefer04} and nickel \cite{hofmann09} at about 125~meV and 250~meV below the Fermi energy, respectively, thus in the same order of magnitude as in our theoretical calculation. However, in both cases, minority instead of majority bands were investigated, and we do not observe kink structures close to the Fermi energy in the minority channel. Cui \textit{et al.}~\cite{cui07}, on the other hand, published ARPES spectra of iron and claimed an anomaly in a majority band at around 270~meV to originate from electron-magnon interactions. However, in Ref.~\onlinecite{cui10} they distanced themselves from this claim and attributed the anomaly to an inaccuracy of the line-shape analysis in case of strong lifetime broadening. Indeed, our theoretical spectra do not exhibit a dispersion anomaly in the band they have measured. } \fi In iron and cobalt, we clearly see the onset of strong lifetime broadening when the quasiparticle bands suddenly lose intensity, and a ``waterfall'' structure forms. The ``waterfall'' structures seen here are very similar to a feature often observed in ARPES spectroscopy. This feature is interpreted as a footprint of a many-body resonance, at whose energy the spectral function is expected to show non-Fermi liquid behavior. The interpretation is confirmed by the present calculations. The ``waterfall'' structures are observed roughly at 250-600~meV in Fe [Fig.~\ref{fig::kinks2}(a)] and at 200-500~meV in Co [Fig.~\ref{fig::bandstr}(b)]. Nickel, on the other hand, exhibits the weakest lifetime broadening of all. While it is still strong, it does not lead to the disappearance of the quasiparticle character as in Fe and Co. The quasiparticle bands, although substantially broadened, remain well-defined over the whole energy range. Nevertheless, one can still make out a slight waterfall shape in the strongly dispersing band between $\Gamma$ and K. \cf{In one of the bands of Fe, one can discern a slight kink in the band dispersion at about 250~meV, just where the lifetime broadening sets in. Cui \textit{et al.}~\cite{cui07} published ARPES spectra of iron and claimed an anomaly in a majority band at around 270~meV to originate from electron-magnon interactions. However, in Ref.~\onlinecite{cui10} they distanced themselves from this claim and attributed the anomaly to an inaccuracy of the line-shape analysis in case of strong lifetime broadening. Indeed, our theoretical spectra do not exhibit a dispersion anomaly in the band they have measured.} \begin{figure} \includegraphics[width=1.\columnwidth]{Fe_kink.png} \caption{\label{fig::kinks2}(Color online) Renormalized band structure of iron in the (a) majority and (b) minority channel. High-energy anomaly in a minority band of iron. (a) and (b) $GT$-renormalized spectral function with $\Delta_v=190$~meV and (c) $\Delta_v=0$. The LSDA bands are shown as red lines. The red crosses are fitted peak positions from experiment (data reproduced from Ref.~\onlinecite{Mlynczak19}). } \end{figure} In the following, we investigate another band anomaly in a minority band of iron at the relatively large binding energy of 1.5~eV. Interestingly, very recently, an anomaly was found in ARPES experiments in the same minority band and roughly at the same energy and momentum as predicted by our theoretical calculation. We report about this comparison elsewhere (Ref. \onlinecite{Mlynczak19}) in detail. Figure~\ref{fig::kinks2 (b) and (c) present the theoretical (minority) spectral functions together with a comparison to the peak fits (red crosses) of the experimental momentum distribution curves. The two calculations (b) and (c) differ in the parameter $\Delta_v=190$~meV and $\Delta_v=0$. The form of the two plots is very similar, implying that the spectral function is largely unaffected by the parameter $\Delta_v$ at higher binding energies. In Ref.~\onlinecite{Mlynczak19}, we only show the spectrum for $\Delta_v=0$, which is why we have included this result here for comparison. After what we have discussed before, it might seem contradictory that a band anomaly should appear in a minority valence band. In fact, the scattering phase space is smaller for the minority valence states, but it is not zero. It does contain spin-flip Stoner excitations to which a propagating hole can couple. And, in the present case, these Stoner excitations form a particularly strong intensity in the spin excitation spectrum at around 0.7~eV \cite{Mlynczak19}. The many-body resonance is then formed by the Stoner pairs (spin $+\hbar$) together with majority hole $d$ states (spin $-\hbar/2$) that form a pronounced peak in the density of states at around 0.8~eV. This resonance is strong enough to give rise to the observed anomaly in the minority hole band (spin $+\hbar/2$) at a binding energy of 1.5~eV. We note at this point that the shape of the band anomaly can again be described as a ``waterfall'' structure. \begin{figure} \includegraphics[width=\columnwidth]{senergy.png} \caption{\label{fig::senergy}(Color online) Momentum and frequency dependence of the self-energy for iron. The imaginary component $|\mathrm{Im}\,\Sigma^\sigma_{\mathbf{q}m}(\omega)|$ is shown for the bands marked in color in Fig.~\ref{fig::bandk}, from bottom to top: green, blue, red, magenta, orange, black; left column spin up, right column spin down. } \end{figure} \begin{figure} \includegraphics[width=\columnwidth]{bandk.png} \caption{\label{fig::bandk}(Color online) Kohn-Sham LSDA band structure of iron. The colors mark the different bands for which the self-energy is plotted in Fig.~\ref{fig::senergy}.} \end{figure} \cf{ The present method fully retains the $\mathbf{k}$ dependence of both the bosonic (magnonic) excitation spectrum and the self-energy. This is in contrast to the DMFT approach, in which the spatial confinement to the impurity effectively obliterates the $\mathbf{k}$ dependence of the quantities. This leads to the fact that band anomalies that are caused by a coupling to extended spin waves are properly described in the present method but might be missed in DMFT, see, e.g., Fig.~\ref{fig::anomaly}(a). The $\mathbf{k}$ dependence of the spin excitation spectrum has been described already in Refs.~\cite{sasioglu10,friedrich14,friedrich18,mueller16} among other works. However, we also have the possibility to directly investigate the $\mathbf{k}$ dependence of the self-energy. To this end, we have plotted the self-energy matrix elements $\Sigma^\sigma_{\mathbf{k}m}(\omega)$ in Fig.~\ref{fig::senergy}. The respective bands are marked in color in Fig.~\ref{fig::bandk}. Overall, we observe a relatively weak dependence of the self-energy on the Bloch momentum, in particular for the localized $d$ bands. However, some bands do show a strong $\mathbf{k}$ dependence, especially the more strongly dispersing bands, such as the green band (Fig.~\ref{fig::bandk}). We also see the effect of hybridization: For example, the spin-up red and orange curves strongly hybridize close to the N point and exchange orbital character. This is reflected in the self-energies, which, towards the right, seem to exchange roles. Again, we would like to emphasize that a weak $\mathbf{k}$ dependence of the self-energy does not necessarily speak in favor of a purely local treatment of correlation effects. Equally important, or arguably more important, is the $\mathbf{k}$ dependence of the bosonic excitation spectrum. A purely local treatment would turn the quadratic spin-wave branch into an incoherent broad peak, which might wash out resonances of the electron-magnon scattering. } \section{Conclusions} \label{sec::Conclusions} In summary, we have described a Green-function technique to calculate the electron-magnon scattering from first principles. The $GT$ self-energy is constructed from the product of the single-particle Green function $G$ and the multiple scattering $T$ matrix, which can be viewed as an effective interaction that acts through the exchange of magnons. The $T$ matrix, thus, contains the full spin excitation spectrum, comprising collective spin-wave and single-particle Stoner excitations, and, diagrammatically, it describes the correlated motion of an electron-hole pair of opposite spins. When expanded in powers of the screened Coulomb interaction $W$, the $T$ matrix can be written as a summation over ladder diagrams of ever increasing order. We have shown that a formulation consistent with the Hedin equations requires the $GT$ self-energy to include the third-order ladder diagram as the lowest order. In this way, the self-energy approximation is free of double counting with the Hartree or exchange ($GW$) diagrams. \cf{ We showed that the so-constructed $GT$ self-energy can lead to a violation of causality, manifesting itself in an incorrect sign change of the spectral function. We elucidated the mathematical reason for this violation of causality: The particular subset of $GT$ diagrams can be written as the difference of two terms, respectively containing the renormalized and the bare electron-hole propagator. The renormalization redistributes the spectral weights in such a way that the difference can become negative in a certain energy region. However, since the sign change occurs far away from the Fermi energy, we accept this violation of causality and show that the resulting renormalized band structures yield a correct description of the electron-magnon scattering around the Fermi energy. It is expected that other diagrams neglected in the present work, in particular, the $GW$ self-energy, will act to restore the correct sign of the spectral function. However, this goes beyond the scope of the present study. } The $GT$ self-energy gives rise to a very strong lifetime broadening of the quasiparticle bands, to the extent that in a certain energy region the quasiparticle character is virtually lost in the majority valence bands of bcc Fe and fcc Co. The minority bands, however, are much less affected by the lifetime broadening, which we have attributed to the spin asymmetry of the density of states. For a similar reason, one also finds a particle-hole asymmetry. The spin asymmetry in the lifetime broadening explains the experimental fact that minority bands are seen to relatively large binding energies, while the majority bands disappear very soon below the Fermi energy. The strong renormalization effects also give rise to ``waterfall'' structures in the $\mathbf{k}$- and energy-resolved spectral functions, in particular for bcc Fe and fcc Co. The $T$ matrix contains the acoustic spin-wave excitations for all wave vectors, including the long-wavelength limit $\mathbf{k}\rightarrow 0$, where the excitation energy vanishes and the $T$ matrix becomes singular with a mathematical delta peak in its imaginary part. Since such a singularity is likely to give rise to numerical difficulties, we have investigated its contribution to the $GT$ self-energy. As a result, the contribution of the long-wavelength spin-wave excitation should be very small in the problematic limit $\mathbf{k}\rightarrow 0$. The contribution of the spin-wave excitations do not vanish for finite $k$, though. Our numerical calculations have shown that it is this contribution \cf{that can lead to the formation of dispersion anomalies in majority valence and minority conduction bands. There, the spectral function can exhibit double-peak structures, which may be interpreted as a quasiparticle and a magnon satellite peak. However, when the peaks are of equal strength, they are general mixtures of the two and the peak separation can be regarded as a measure of the electron-magnon coupling strength. } Our method can describe a band-dispersion anomaly in the minority channel of bcc Fe at a large binding energy of about 1.5~eV, which has very recently been observed in ARPES measurements \cite{Mlynczak19}. Whereas such a large energy is commonly not associated with a magnon excitation, we could elucidate \cite{Mlynczak19} that this band anomaly is caused by a many-body resonance formed by a superposition of Stoner excitations (an increased Stoner intensity is seen in the spin excitation spectrum at around $0.7$~eV) and majority hole states that form a peak in the density of states (at around $0.8$~eV). \cf{ While the overall shape of the renormalized band structures is similar to corresponding DMFT calculations, the latter have not been able to describe the band anomalies discussed in the present paper. We attribute this to the main approximation of the DMFT approach, the spatial confinement of the electronic correlation to a single impurity site. This neglects the $\mathbf{k}$ dependence of the self-energy, but, more importantly, it also does not allow for a $\mathbf{k}$ dependent bosonic (magnonic) excitation spectrum, effectively replacing the well-defined quadratic branch of extended spin waves by a broad feature of local excitations that cannot couple strongly to electrons. } In summary, we believe that the present work is an important step in the development of ab initio many-body methods for the treatment of magnetic materials, in which the magnon energies (collective and Stoner type), the exchange splitting, the band energies, and the electron-magnon coupling strength set different energy scales that compete with each other. Further studies in this direction might also elucidate the role of magnons in the potential formation of Cooper pairs in high-temperature superconductors. \begin{acknowledgments} The authors acknowledge valuable discussions with Ewa M{\l}y\'{n}czak, Lukas Plucinski, Ersoy \c{S}a\c{s}{\i}o\u{g}lu, Markus Betzinger, Manuel dos Santos Dias, Ferdi Aryasetiawan, and Hans Lustfeld. We thank Irene Aguilera for a critical reading of the manuscript. We gratefully acknowledge the computing time granted through JARA-HPC on the supercomputer JURECA at Forschungszentrum J\"ulich. \end{acknowledgments} \newpage \input{ElectronMagnon_proof.bbl} \end{document}

653cf0f1eb973876f5129c392f965ac832eee200

\section{Introduction} \label{sec:ch1intro} Radiative processes such as absorption, emission, and scattering are prevalent in astrophysical gases. Observations of such gases are made possible by the radiation that escapes and reaches the observer. This emission is often an integral part of a system as it is an energy loss (i.e. cooling) and can affect the evolution of the gas. Star forming regions, supernova remnants, planetary nebulae, and \HII regions are a few examples of systems where such cooling is important because the cooling timescales can be much shorter than the hydrodynamical timescales. Much of the gas in these systems is H and He, which cool inefficiently at low temperatures ($\leq$ 10\textsuperscript{3} K) because their lowest excitation levels occur at relatively high energies. In other words, the Boltzmann factor ($\exp{(\frac{-E}{kT})}$) is small at low temperatures when the gas is neutral and in ionization equilibrium. However, when the gas is shocked, the sudden increase in temperature can lead to cooling rates that are over two orders of magnitude higher than the equilibrium value. Atomic processes such as ionization and recombination also become important when the gas is not in ionization equilibrium. The combination of shock heating and strong cooling can lead to a wide range of temperatures and ionization fractions, which is why a careful treatment of non-equilibrium cooling is important when shock waves are present. One regime in which such processes play a key role is in the outflows from young stellar objects (YSOs). These outflows are observed as Herbig-Haro (HH) jets, and they contain many shocks and heterogeneous cooling regions. Both the shocks and cooling regions are readily observed in optical emission lines with instruments such as the Hubble Space Telescope (HST). The motivation of this work comes from time-resolved HST observations of HH objects \citep[see e.g.][]{Hartigan11}. In numerical simulations, non-equilibrium cooling is necessary to accurately predict the emission from an HH object. The ideal magnetohydrodynamic (MHD) equations with cooling coupled with ionization and recombination equations for various species can result in a better estimate of the temperature and ionization state of the gas. With such models, it is possible to produce synthetic emission maps, and in HH objects, emission is observed in H$\alpha$ (shock diagnostic) and [\SII] (post-shock cold flow diagnostic) \citep{Heathcote96}. As we describe in this paper, AstroBEAR calculates the non-equilibrium rate equations for each bright emission line in the cooling zone with a separate code, and these cooling rates combine to form a look-up table for the total cooling rate. That rate depends solely upon the ionization fraction of H, the electron density, and the electron temperature, parameters that AstroBEAR follows in each time step within the simulation. This scheme works well below 10\textsuperscript{4} K because the ionization states of all the strong coolants are either fixed or are tied to the ionization state of H through strong charge exchange reactions. At higher temperatures, the code takes forbidden line cooling from the Dalgarno-McCray cooling curves \citep{Dalgarno72}, but calculates non-equilibrium recombination cooling from H and He explicitly in each step. Our code's ability to accurately track the temperature and ionization state of the gas makes it possible to produce synthetic emission maps of various emission lines. Previous studies of outflows using AstroBEAR's new cooling method have already been done \citep{Hansen15a,Hansen17}, and synthetic emission maps of H$\alpha$ and [\SII] were presented in those works. Many other hydrodynamic codes calculate non-equilibrium cooling, and we list several of these codes here: YGUAZU-A \citep{Raga00}, FLASH \citep{Fryxell00}, PLUTO \citep{Tesileanu08}, ENZO \citep{Bryan14}, NIRVANA \citep{Ziegler16}, and RAMSES \citep{Teyssier02}. All of these codes have been implemented in 3-D and have support for magnetic fields. The implementation of non-equilibrium cooling in each of these codes is different, and each code may contain additional multi-physics such as radiation transfer, self-gravity, heat conduction, magnetic resistivity, etc. Thus, the choice of which code to use depends on the problem of interest and personal preference. Table~\ref{tab:codes} lists these codes and which ionic species they use to calculate cooling rates. The species listed in Table~\ref{tab:codes} are separated into two columns (``Non-equilibrium'' and ``Equilibrium''). In the ``Equilibrium'' column, these species contribute to the overall cooling but with ionization equilibrium assumed. In the ``Non-equilibrium'' column, species are either explicitly tracked or charge-exchange is used to obtain contributions to cooling without assuming ionization equilibrium (i.e., non-equilibrium). It is important to explicitly track H and He to correctly predict hydrodynamic quantities such as density and temperature, however, it is not always necessary to explicitly track the heavier elements (i.e., metals). To predict emission line fluxes accurately, species should be treated in the non-equilibrium regime. In AstroBEAR, the framework exists for future users of the code to implement explicit tracking of species (e.g., \OIII) if other emission lines are desired. \begin{deluxetable}{cc p{150pt}|p{130pt}} \tablenum{1} \tablecolumns{4} \tablewidth{0pt} \tablecaption{Capabilities and species included in cooling function for various hydrodynamic codes \label{tab:codes}} \tablehead{ & \colhead{Charge} & \multicolumn{2}{c}{\hspace{20pt}Species} \\ \colhead{Code} & \colhead{Exchange} & \colhead{Non-equilibrium} & \colhead{Equilibrium}} \startdata AstroBEAR & Y & \Hions, \Heions, \Nions, \Oions, \Sions & \Cionsvi, \Nionsviii, \Oionsviii, \Neionsxi, \Sionsxvii, \Mgions, \Siionsxiv, \Feions \\ YGUAZU-A & Y & \Hions, \Cions, \Nionsv, \Oionsv, \Neionsv, \Sionsv & \multicolumn{1}{c}{none} \\ FLASH & Y & \multicolumn{1}{c}{varies} & \multicolumn{1}{c}{varies} \\ PLUTO & Y & \Hions, \Heionsii, \Cionsv, \Nionsv, \Oionsv, \Neionsv, \Sionsvv & \MgII, \SiII, \FeII \\ ENZO & N & \Hions, \Heions, D, D\textsuperscript{+}, H$_2$, H$_2$\textsuperscript{+}, HD, HD\textsuperscript{+} & \multicolumn{1}{c}{none} \\ NIRVANA & N & \Hions, \Heions, H\textsuperscript{-}, H$_2$, H$_2$\textsuperscript{+} & \multicolumn{1}{c}{none} \\ RAMSES & N & \Hions, \Heions & \multicolumn{1}{c}{none} \enddata \end{deluxetable} Many of the codes treat non-equilibrium radiative cooling in a similar fashion. YGUAZU-A \citep{Raga00} uses the same cooling as its predecessor code CORAL \citep{Raga97}. YGUAZU-A/CORAL uses a network of 18 different rate equations for various ionic species to obtain a non-equilibrium cooling function. AstroBEAR also uses rate equations (for H, He, and S), but it uses table look-ups for the other species. Both codes are appropriate for approximate temperature ranges between 10\textsuperscript{3} and 10\textsuperscript{5} K. The codes track different species, so the preferred code also depends on which emission lines are desired for analysis. CORAL requires an assumption of axisymmetry, but YGUAZU-A can be run in 3-D. Magnetic fields may be important, especially in cooling regions behind shock fronts. For typical ISM conditions, the thermal and magnetic pressures are comparable, but at shock fronts the thermal pressure usually dominates. In cooling regions, the magnetic field is compressed along with the gas, thus the magnetic pressure can begin to dominate \citep{Hartigan15}. MHD codes like AstroBEAR are better-suited for simulating regions where fields may play an important role. FLASH is an open source code, and different users have incorporated different forms of cooling to suit their needs. For example, \citet{Gaspari11} used a metallicity dependent cooling function in FLASH which was appropriate for cooling in the intercluster medium (ICM) above 10\textsuperscript{4} K. With an assumed metallicity, the cooling rate was reduced to a function dependent only on temperature based on cooling functions from \citet{Sutherland93}. This treatment is different from AstroBEAR, but, in this case, FLASH is used in a completely different regime: the ICM. The current version of non-equilibrium cooling in AstroBEAR is more appropriate for higher density and lower temperature regions such as the interstellar medium (ISM). PLUTO has been used to study radiative shocks in YSO jets \citep{Tesileanu14}. Similar to the other aforementioned codes, PLUTO tracks several ionic species to calculate collisionally excited line radiation in the optically thin limit. Like AstroBEAR, it adds cooling contributions from \MgII, \SiII, and \FeII ~without explicitly tracking their number densities. PLUTO also uses look-up tables for more efficient computation. One can see from Table~\ref{tab:codes} that the PLUTO code has a relatively large network of ionic species, but we will show that explicitly tracking this many species is not necessary to produce accurate results in certain applications that require non-equilibrium cooling. Like FLASH, ENZO has been used to study the effects of cooling in the ICM \citep{Li15}. ENZO is different from the other codes in that it also includes cooling from deuterium, and molecular forms of hydrogen and deuterium; this makes it more suitable for low temperatures below 10\textsuperscript{3} K. The code can track all species of hydrogen, deuterium, and helium (including the aforementioned molecules). There is also an option to incorporate the use of look-up tables, which were calculated with the photoionization code CLOUDY \citep{Ferland98}, for metal cooling. The limitation is that the ENZO code assumes ionization equilibrium for the metals and does not explicitly track number densities like the other codes that track the heavier species. NIRVANA is another MHD code that has non-equilibrium cooling capabilities. The implementation is different from other codes in that it reads in a single text file which contains the desired reaction equations. Cooling is then predicted for each reaction with rate coefficients, which are derived from collision strengths found in various references. Reaction equations could be implemented to include metals and charge exchange, but the examples given in \citet{Ziegler16} only used the species listed in Table~\ref{tab:codes}. This implementation method makes it relatively simple to add new species (including molecules), but it becomes computationally expensive when incorporating many species due to the increased number of equations to be solved. This is a common issue with all hydrodynamic codes: tracking a large chemical network of ionic species becomes computationally expensive, especially when the code is three-dimensional and includes other multi-physical processes. For this reason, AstroBEAR, and codes similar to AstroBEAR, often use look-up tables to obtain cooling rates. Not all cooling tables work when the gas is out of ionization equilibrium. For example, mostly neutral gas at 30000 K cools much faster than ionized gas at the same temperature and density. For this reason, it is important that cooling routines follow the ionization fraction of H explicitly to give accurate results. RAMSES is yet another MHD code that, like AstroBEAR, uses AMR \citep{Teyssier02}. It is an N-body and hydrodynamics code designed to study structure formation at cosmological and galactic scales. The non-equilibrium cooling implementation is detailed in \citet{Rosdahl13} where the authors have named the extension of this code RAMSES-RT. Metal cooling is not included in this code in part because it is less important at these large scales. There are many other codes currently in use, and there are often different versions of a code which are used for different applications. The main point is that many of these codes treat non-equilibrium cooling in a similar fashion, but they are different in the specific processes and species that they include. The code user should pay careful attention to what is included in the cooling of a code to determine if it is appropriate for their application. Section~\ref{sec:ch1methods} contains our numerical methods including the ionization, recombination, non-equilibrium cooling, and production of synthetic emission maps in AstroBEAR. The best way to assess how well this implementation works is to see how closely it matches the predictions of temperature, density, and magnetic field throughout the post-shock cooling zones from a well-tested 1-D model that includes all the relevant physics. These comparison models were generated by the shock code built by Cox and Raymond (hereafter referred to as CRSC, \citet{Raymond79}), and they are shown in Section~\ref{sec:1Dshocks}. We also show some examples of synthetic emission maps from simulations of supersonic outflows in Section~\ref{sec:outflows}. Finally, in Section~\ref{sec:ch1conc} we summarize our paper and discuss ongoing and future work. \section{Methods} \label{sec:ch1methods} AstroBEAR is a highly parallelized adaptive mesh refinement (AMR) multi-physics code. The use of AMR is crucial in 2-D and 3-D to resolve regions of rapid cooling and ionization (behind shock fronts). See \cite{Cunningham09,Carroll12} for a detailed explanation of how AMR is implemented. More details of the code can also be found at https://astrobear.pas.rochester.edu/trac/. The code can solve the 3-D MHD equations with a variety of multi-physics and source terms, but for the present work, we will only consider the 3-D ideal MHD equations with non-equilibrium cooling: \begin{subequations}\label{group1.1} \begin{gather} \frac{\partial \rho}{\partial t} + \boldsymbol{\nabla} \cdot \rho \boldsymbol{v} = 0 ,\ \label{1.1a}\\[\jot] \frac{\partial \rho \boldsymbol{v}}{\partial t} + \boldsymbol{\nabla} \cdot (\rho \boldsymbol{v} \boldsymbol{v} + P\boldsymbol{I} - \boldsymbol{B}\boldsymbol{B}) = 0 ,\ \label{1.1b}\\[\jot] \frac{\partial E}{\partial t} + \boldsymbol{\nabla} \cdot ((E + P) \boldsymbol{v} - (\boldsymbol{v} \cdot \boldsymbol{B})\boldsymbol{B}) = -L ,\ \label{1.1c}\\[\jot] \frac{\partial \boldsymbol{B}}{\partial t} + \boldsymbol{\nabla} \cdot (\boldsymbol{v} \boldsymbol{B} - \boldsymbol{B} \boldsymbol{v}) = 0 ,\ \label{1.1d}\\[\jot] \frac{\partial n_i}{\partial t} + \boldsymbol{\nabla} \cdot n_i \boldsymbol{v} = \Gamma_i ,\ \label{1.1e} \end{gather} \end{subequations} where $\rho$ is the mass density, $\boldsymbol{v}$ is the velocity, $P$ is the total pressure (thermal + magnetic) defined as $P = P_{th} + \frac{1}{2} B^2$, $\boldsymbol{I}$ is the identity matrix, $\boldsymbol{B}$ is the magnetic field normalized by $\sqrt{4 \pi}$, and $E$ is the total energy such that $E = \frac{1}{\gamma - 1} P_{th} + \frac{1}{2}\rho v^2 + \frac{1}{2} B^2$ (with $\gamma = \frac{5}{3}$ for an ideal gas). $L$ is the cooling source term which will be a function of number density, temperature, and ionization. $n_i$ is the number density of species $i$, and $\Gamma_i$ is the sum of the ionization and recombination rates for species $i$. The equations above represent the conservation of mass \eqref{1.1a}, momentum \eqref{1.1b}, energy \eqref{1.1c}, and magnetic flux \eqref{1.1d}. Equation \eqref{1.1e} represents the evolution of the number densities of the different atomic species tracked within the code. In the subsections that follow, we will describe in detail the recombination and ionization $\Gamma_i$, the cooling source term $L$, and how the inclusion of various species $n_i$ are used to produce synthetic emission maps. \subsection{Ionization and Recombination} \label{subsec:ionrecomb} There currently are a total of 8 species in AstroBEAR whose number densities are tracked via equation \eqref{1.1e}: \HI, \HII, \HeI, \HeII, \HeIII, \SII, \SIII, and \SIV. The source term $\Gamma_i$ contains the ionization and recombination rates relevant to each species. Thus, $\Gamma_i$ can be expressed as 8 more equations: \begin{subequations}\label{group1.2} \begin{gather} \Gamma_{\HI} = -n_e n_{\HI} C_{\HI} + n_e n_{\HII} \alpha_{\HII} ,\ \label{1.2a}\\[\jot] \Gamma_{\HII} = n_e n_{\HI} C_{\HI} - n_e n_{\HII} \alpha_{\HII} ,\ \label{1.2b}\\[\jot] \Gamma_{\HeI} = -n_e n_{\HeI} C_{\HeI} + n_e n_{\HeII} \alpha_{\HeII} ,\ \label{1.2c}\\[\jot] \Gamma_{\HeII} = n_e n_{\HeI} C_{\HeI} - n_e n_{\HeII} \alpha_{\HeII} - n_e n_{\HeII} C_{\HeII} + n_e n_{\HeIII} \alpha_{\HeIII} ,\ \label{1.2d}\\[\jot] \Gamma_{\HeIII} = n_e n_{\HeII} C_{\HeII} - n_e n_{\HeIII} \alpha_{\HeIII} ,\ \label{1.2e}\\[\jot] \Gamma_{\SII} = -n_e n_{\SII} C_{\SII} + n_e n_{\SIII} \alpha_{\SIII} ,\ \label{1.2f}\\[\jot] \Gamma_{\SIII} = n_e*n_{\SII} C_{\SII} - n_e n_{\SIII} \alpha_{\SIII} - n_e n_{\SIII} C_{\SIII} + n_e n_{\SIV} \alpha_{\SIV} ,\ \label{1.2g}\\[\jot] \Gamma_{\SIV} = n_e n_{\SIII} C_{\SIII} - n_e n_{\SIV} \alpha_{\SIV} .\ \label{1.2h} \end{gather} \end{subequations} Here $n_e$ is the electron number density, $C$ is the collisional ionization rate in units of cm\textsuperscript{3}s\textsuperscript{-1} and $\alpha$ is the sum of the radiative recombination rate and dielectric recombination rate also in units of cm\textsuperscript{3}s\textsuperscript{-1} ($\alpha = \alpha_r + \alpha_d$). In these equations and everywhere in this paper, the subscript $i$ following $C$ or $\alpha$ refers to the initial species which interacts with an electron and becomes ionized or recombines. Following the prescription in \citet{Mazzotta98}, the collisional ionization rate for species $i$ can be written as \begin{equation} C_i = \frac{6.69 \times 10^{7}}{(kT)^{3/2}} \frac{\exp{(-x_i)}}{x_i} F(x_i,\sigma_i) , \ \label{1.3} \end{equation} where \begin{equation} x_i = \frac{I_i}{kT} .\ \label{1.4} \end{equation} Here $I_i$ is the ionization potential for species $i$ (e.g. 13.6 eV to ionize \HI into \HII), $T$ is temperature in K, $k$ is the Boltzmann constant in eV/K, and $F$ is a function which depends on $x_i$ and collisional cross section $\sigma_i$. The collisional cross sections which determine the value of $F$ were taken from \citet{Arnaud85}. We followed the fit used by \citet{Verner96} to determine radiative recombination rates. \begin{equation} \alpha_r = A \left[ \sqrt{\frac{T}{T_0}} \left( 1 + \sqrt{\frac{T}{T_0}} \right) ^{1-b} \left( 1 + \sqrt{\frac{T}{T_1}} \right) ^{1+b} \right] ^{-1} ,\ \label{1.5} \end{equation} where $A$, $T_0$, $T_1$, and $b$ are fitting parameters different for each species. Dielectric recombination is different from radiative recombination in that the ion first transitions to an excited state and then to its ground state. \citep{Mazzotta98} gives the dielectric recombination rate as \begin{equation} \alpha_d = \frac{c}{T^{3/2}} \exp{\left(-\frac{E}{T}\right)} ,\ \label{1.6} \end{equation} where $c$ and $E$ are the fitting parameters. The sum of $\alpha_r$ and $\alpha_d$ gives the total recombination rates $\alpha$ used in equations \eqref{group1.2}. It is also important to note how we keep track of the electron number density $n_e$. The gas is assumed to be quasi-neutral, thus the electron number density $n_e$ is equal to \begin{equation} n_e = \max(n_{\HII} + n_{\HeII} + 2 n_{\HeIII} , n_{min}) ,\ \label{1.7} \end{equation} where $n_{min}$ is a minimum density which we chose to be 0.01 times the total gas density. This minimum is necessary to ensure that there are always at least some electrons present to ionize the gas. The use of such a minimum is also physically realistic since the heavier metals (e.g., S or Fe) will always have some level of ionization due to the presence of far-ultraviolet (FUV) photons. This is also why we do not track neutral \SI; we assume that all sulfur is at least singly ionized (i.e., \SII). All of the collisional ionization rates $C$ and the recombination rates $\alpha$ multiplied by the appropriate number densities form the right hand sides of equations \eqref{group1.2}. AstroBEAR first solves the source-free version of equation \eqref{1.1e} (i.e. with $\Gamma = 0$) and then uses an operator split approach to handle the microphysical source terms. The source terms $\Gamma$ are integrated using a 5\textsuperscript{th} order accurate Cash-Karp Runge-Kutta method. This method uses the difference between the 4\textsuperscript{th} and 5\textsuperscript{th} order solutions to determine accuracy and adjust the time-step as appropriate. AstroBEAR uses the same method for other constant source terms (e.g., the cooling term $L$). \subsection{Non-equilibrium Cooling} \label{subsec:cooling} The cooling source term $L$ in equation \eqref{1.1c} is treated in the same way as $\Gamma$ in equation \eqref{1.1e}. As described in the previous subsection, the source term is integrated using a Cash-Karp Runge-Kutta method. $L$ is comprised of four main categories of coolants: hydrogen $L_H$, helium $L_{He}$, other heavy metals $L_Z$, and molecules $L_M$ as illustrated in equation \eqref{1.8}. \begin{equation} L = L_H + L_{He} + L_Z + L_M .\ \label{1.8} \end{equation} Each of these coolants can be further separated into more terms via the underlying processes responsible for the energy losses. $L_H$ contains energy losses due to \HI \ excitation, \HI \ ionization, and \HII \ recombination. Thus, $L_H$ can be written as \begin{equation} L_H = \HI_{ex} + n_e \left( n_{\HI} C_{\HI} I_{\HI} + n_{\HII} \alpha_{\HII} k T \right) .\ \label{1.9} \end{equation} Likewise, $L_{He}$ can be split into its excitation, ionization, and recombination terms: \begin{equation} L_{He} = \HeI_{ex} + \HeII_{ex} + n_e \left( n_{\HeI} C_{\HeI} I_{\HeI} + n_{\HeII} \alpha_{\HeII} k T + n_{\HeII} C_{\HeII} I_{\HeII} + n_{\HeIII} \alpha_{\HeIII} k T\right) .\ \label{1.10} \end{equation} In equations \eqref{1.9} and \eqref{1.10}, the terms labeled by subscripts `$ex$' represent cooling due to excitation. In previous versions of AstroBEAR, these were taken directly from the Dalgarno \& McCray (DM) cooling curve which assumes ionization equilibrium values for the number densities \citep{Dalgarno72}. Through the use of equation \eqref{1.1e}, we can now use non-equilibrium values for the number densities which results in more accurate energy losses. To convert the ionization and recombination rates, $C$ and $\alpha$ respectively, into energy loss rates we multiply by appropriate number densities and energies. The ionization terms are multiplied by the ionization potential $I_i$ for species $i$, and the recombination terms are multiplied by a factor of $2/3 kT$. This factor of 2/3 warrants additional explanation. Although the average kinetic energy of the electrons is $3/2 kT$ and the most probable kinetic energy is $kT$, the recombination cross sections of ions (including H) decline as electron energies increase. In general, values of the average energy lost per recombination to level $n$ depend both upon $n$ and the temperature, and vary from about $0.9 kT$ to $0.5 kT$ as $n$ changes from 2 to 6 for H at temperatures of interest. Hence we adopt 2/3 for the multiplicative constant in front of the $kT$ term for recombination losses. Typically, collisional cooling dominates by a factor of $\sim$ 10 over recombination, so the exact value of the constant has a minor effect on the overall energy budget in the simulations. Cooling from metal excitation is calculated from one of two different tables depending on the temperature. In between 2000 and 16500 K, we use a new cooling table which contains forbidden line cooling from \OI, \OII, \NI, \NII, \SII, \FeII, \SiII, \MgII, and \CII. The table uses strong charge exchange cross sections to lock the ratios of \NII/\NI \ and \OII/\OI \ to \HII/\HI, and it solves multilevel atom models to derive volume emission cooling terms. This 3-D table depends on the electron number density, ionization fraction, and temperature. Hence the tracking of various neutral and ionized species makes the use of this table possible and results in more accurate cooling due to metals. Cooling in this table (hereafter referred to as ``Z'' cooling) is strictly from metal-electron collisions. We also use a modified version of the Dalgarno \& McCray (DM) cooling curve \citep{Dalgarno72} where we subtract the contributions from H and He, leaving only the metal and molecular components. We altered the abundances to match the abundances used in the aforementioned 3-D metal cooling table which are more appropriate for the interstellar medium (see Table~\ref{tab:abundances}). The metal and molecular components of the DM curve are further altered by using non-equilibrium densities instead of the values one would get from ionization equilibrium. At every temperature, we use this modified DM curve for metal-hydrogen collisional cooling and molecular cooling, although the molecular component will only become important below 1000 K. Above 16500 K and below 2000 K, the metal-electron collision term of the modified DM curve is used instead of the value from the 3-D table. A smoothing function is applied at the boundaries of these two metal-electron cooling tables (2000 and 16500 K) to avoid discontinuous jumps in cooling rates. Figure~\ref{fig:coolingcurves} shows the following cooling curves: DM, an altered DM, and the new 3-D Z cooling table. \begin{deluxetable}{cc} \tablenum{2} \tablecolumns{2} \tablewidth{0pt} \tablecaption{Element abundances relative to hydrogen \label{tab:abundances}} \tablehead{\colhead{Element} & \colhead{Abundance}} \startdata H & 12 \\ C & 8.52 \\ N & 7.96 \\ O & 8.82 \\ Ne & 2.60 \\ Mg & 7.42 \\ Si & 7.52 \\ S & 7.20 \\ Fe & 7.60 \enddata \end{deluxetable} \begin{figure} \centering \includegraphics[width=0.75\linewidth]{f1.eps} \caption{Cooling curves used in AstroBEAR. The y-axis shows the cooling efficiency $\Lambda$ which leads to the energy loss term $L$ when multiplied by appropriate number densities ($n_e$ and total hydrogen number density $n_\text{H}$). The \emph{solid} line represents the original DM cooling curve used in previous versions of AstroBEAR, \emph{dashed} line is the DM curve but with the slightly altered abundances, and the \emph{gray circles} are all of the data points from the new 3-D metal cooling table. The upper boundary of the gray area represents the fully ionized limit, while the lower boundary represents the fully neutral limit. The new table allows for the possibility of having gas far from ionization equilibrium in the region where most of the forbidden lines originate and where magnetic pressure effects are greatest.} \label{fig:coolingcurves} \end{figure} \subsection{Calculating Synthetic Emission Lines} \label{subsec:emiss} In the following subsection, all emission equations give an emissivity in units of erg/cm\textsuperscript{3}/s. Tracking the hydrogen and helium species allows us to track their ionization fractions and thus the electron number density of the gas which is required for generating synthetic emission maps (equation \eqref{1.7}). For HH objects, [\SII] emission is strongest in cooling regions behind shocks \citep{Heathcote96}. In AstroBEAR, below $10^4 K$, all S is assumed to be \SII, and above $10^4 K$, the ionization and recombination rates are used to track the amount of \SIII \ and \SIV. This was employed to more accurately track the ionization state of S at higher temperatures and hence produce more accurate [SII] maps. Our total [\SII] emission is the sum of two lines with photon wavelengths of 673.1 and 671.6 nm. These correspond to electron transitions from level 2 to level 1 and level 3 to level 1, respectively. The total [\SII] emission can be written as \begin{equation} [\text{\SII}] = \sum_{j=2}^{3} n_j A_j h \nu_j ,\ \label{1.11} \end{equation} where $n_j$ is the number density of \SII \ in excited state $j$, $A_j$ is the Einstein A-coefficient for spontaneous emission from level $j$ to level 1, $h$ is Planck's constant, and $\nu_j$ is the frequency of the emitted photon such that $h\nu_j$ is in ergs. The values for $A_j$ are taken from \citet{Mendoza82}, and we solve the 5-level atom to determine the values for $n_j$. In solving for the level populations $n_j$ of this system, we used collision strengths from \citet{Keenan96}. Calculating the H$\alpha$ emission accurately is crucial to comparing these simulations with observations. H$\alpha$ will typically mark shock fronts within HH objects \citep{Heathcote96}. In AstroBEAR, the H$\alpha$ routine is dependent on electron number density, temperature, ionization fraction, and total hydrogen number density. The total H$\alpha$ emission is the sum of two terms: a collisional excitation term H$\alpha_c$ and a recombination term H$\alpha_r$. For the excitation term, we include excitations to levels 3, 4, and 5 using effective collision strengths $\upsilon$ from \citet{Anderson00} and \citet{Anderson02}. The excited atom then de-excites, and an H$\alpha$ photon is produced when the atom transitions from level 3 to level 2. Such a photon has energy $h\nu \simeq 3.0263 \times 10^{-12}$ ergs. The electrons in the higher levels (4 and 5) have to cascade down to level 3 and then level 2 to be included in the H$\alpha$ emission, thus their contributions are less significant than electrons already in level 3. The excitation term can be written as \begin{equation} \text{H}\alpha_c = n_e n_{\HI} h \nu \frac{8.63 \times 10^{-6}}{2\sqrt{T}} \sum_{j=3}^{5} \exp{\left(\frac{-E_j}{kT}\right)}\upsilon_j ,\ \label{1.12} \end{equation} where $E_j$ is the energy required to excite neutral H from level 1 to level $j$, and $\upsilon_j$ is the collision strength for such an excitation. The collision strength includes a factor which accounts for the fact that not all excitations lead to H$\alpha$ emission; some will produce H$\beta$ and H$\gamma$ emission as well. The recombination term is important when the gas is highly ionized at relatively low temperatures $\lesssim$ $10^4$ K. When ionized hydrogen is recombining, there is a nonzero probability that it will go through the H$\alpha$ transition. To calculate this term, the radiative recombination rate $\alpha_r$ is simply multiplied by the appropriate quantities as follows: \begin{equation} \text{H}\alpha_r = n_e n_{\HII} h \nu \alpha_r .\ \label{1.13} \end{equation} As in Subsection~\ref{subsec:ionrecomb}, recombination rates are taken from \citet{Verner96}, and they are valid above 3000 K. Both recombination and collisional excitation can be important depending on the conditions. In general, at low temperatures ($\lesssim$ $10^4$ K), H$\alpha$ emission is weak but is dominated by the recombination term because any hydrogen that is ionized will quickly recombine. The excitation term dominates at higher temperatures (up to 10\textsuperscript{5} K) as the gas is less likely to recombine and energies are favorable for excitation. At still higher temperatures, the amount of H$\alpha$ will decrease because the hydrogen becomes mostly ionized (no neutral H to excite). AstroBEAR does all of the emission calculations as post-processing after each outputted frame from a simulation. In order to generate a synthetic emission map, the gas is assumed to be optically thin, so the emission can be easily summed along a line of sight. The data can be rotated or inclined to produce a 2-D emission map from different observer perspectives. The emission maps in this paper show [\SII] in red and H$\alpha$ in green, and the ratio [\SII]/H$\alpha$ determines the color such that when this ratio approaches unity, the color will become yellow. \section{Testing AstroBEAR's Cooling Routines} \label{sec:1Dshocks} One of the first problems that one might study with an MHD code that uses equations~\eqref{group1.1} is a 1-D, stationary, magnetized, radiative shock. Exploring the behavior of such shocks with varying initial conditions helps us understand the various environments where these shocks exist such as within YSO jets. However, to make comparisons with observations, it is necessary for a code to accurately track temperature, densities, and other quantities in the post-shock region where the gas is cooling. We ran a number of models with AstroBEAR and compared the results with equivalent models run by CRSC. For our AstroBEAR models, we let the gas move in the x-direction with velocity $v_x$ and the magnetic field is in the y-direction $B_y$. This was chosen because any field component that is perpendicular to the shock front ($B_x$) will not be compressed by the shock. Given initial pre-shock conditions for density, velocity, temperature, field strength, and ionization fraction, the 1-D shock jump equations can be used to find the post-shock values. The 1-D MHD shock jump equations are as follows: \begin{subequations}\label{group1.14} \begin{gather} \left[ \rho v_x \right]^2_1= 0 ,\ \label{1.14a}\\[\jot] \left[ \rho v_x^{\ 2} + P_{th} + \frac{B_y^{\ 2}}{2} \right]^2_1 = 0 ,\ \label{1.14b}\\[\jot] \left[ (\frac{1}{2}\rho v_x^{\ 2} + \frac{\gamma}{\gamma - 1}P_{th} + B_y^{\ 2}) v_x \right]^2_1 = 0 ,\ \label{1.14c}\\[\jot] \left[ B_y v_x \right]^2_1 = 0 ,\ \label{1.14d}\\[\jot] \left[ n_i v_x \right]^2_1 = 0 ,\ \label{1.14e} \end{gather} \end{subequations} where the variables are defined in the same way as they were in Section~\ref{sec:ch1methods} for equations~\eqref{group1.1}. The bracket notation is used to define the difference between post-shock and pre-shock quantities ($\left[ A \right]^2_1 = A_2 - A_1$). For simplicity, the ionization fraction is assumed to remain constant across the shock jump. The cooling and ionization/recombination source terms must be accounted for in the post-shock region. We use a Runge-Kutta integration method to solve for the energy flux and species number density fluxes one computational cell at a time. From these values, we can derive a density, velocity, temperature, and magnetic field to populate all of the cells behind the shock. This can be done out to any distance, but we stop the calculation once the temperature drops to about 1000 K since this is where molecular cooling processes would start to become important. We will refer to the distance from the shock front to T = 4000 K as the cooling length $L_{cool}$. Once the domain is initialized with the pre-shock region, shock jump, and post-shock region, the code can begin taking hydrodynamic time-steps. Since the simulations are run in the frame of the shock, the entire radiative shock profile remains stationary and does not change. The CRSC \citep{Raymond79} treatment of a radiative shock is very different than that of AstroBEAR. It is not an MHD code; it instead follows a single parcel of gas as it travels through a post-shock region and radiates. CRSC, as compared to AstroBEAR, contains many more emission lines, explicitly tracks more ionized species, and accounts for more chemical processes such as photoionization. The CRSC also calculates non-local radiative transfer by following the optical depth of Lyman continuum photons. Such photons are mostly produced at the shock front and can reionize the gas in the post-shock cooling zone. This effect is mitigated in AstroBEAR by limiting the shock velocities to below 90 km/s. Modeling higher shock velocities would require a treatment of photoionization and radiative transfer in 3-D which can be computationally expensive. Higher shock velocities also require higher resolution which imposes a code limitation to which shock models we can reasonably compare. Despite these differences and limitations, radiative shock models of modest shock velocities ($<$ 90 km/s) from both codes are strikingly similar. Furthermore, the shock velocity limitation still allows for accurate modeling of HH objects. Although HH objects exhibit speeds of 200-400 km/s (determined by proper motions and radial velocities), the knots within such objects move into previously ejected material with relative velocities on the order of 10-100 km/s \citep{Hartigan11}. We tested the new AstroBEAR cooling routine by running dozens of models, and highlight four of them here which vary in pre-shock ionization fraction and shock velocity (see Table~\ref{tab:shockmodels}). Changing the ionization fraction is the best way to test the ionization/recombination routines in different regimes in AstroBEAR, and changing the shock velocity is the best way to test different temperature ranges. All four models have the same pre-shock values for total atomic number density and temperature of 100 cm\textsuperscript{-3} and 10000 K, respectively. Models A and B have shock velocities of 40 km/s and models C and D have shock velocities of 70 km/s. \begin{deluxetable}{cccc} \tablenum{1.2} \tablecolumns{4} \tablewidth{0pt} \tablecaption{Varied parameters in 1-D shock models \label{tab:shockmodels}} \tablehead{\colhead{Model} & \colhead{v [km/s]} & \colhead{B [$\mu$G]} & \colhead{X}} \startdata A & 40 & 94 & 0.01 \\ B & 40 & 94 & 0.40 \\ C & 70 & 77 & 0.01 \\ D & 70 & 77 & 0.40 \\ \enddata \end{deluxetable} Figure~\ref{fig:lowXlowM} shows the radiative shock profile for model A. This figure, and the three that follow, show the post-shock profiles for several MHD quantities and five different emission lines. We have plotted the temperature in K, density compression $n/n_o$, ionization fraction $n_e/n_o$, and inverse plasma beta $P_B/P$. As expected, the temperature drops and the density increases, hence the increase in thermal pressure is minimal. The magnetic pressure, however, increases along with the density due to magnetic flux freezing \citep{Hartigan15}, thus $P_B/P$ increases in the post-shock region. The emission lines that we have plotted here include H$\alpha$, [\OI] (630.0 nm), [\NII] (658.3 nm), and [\SII] (671.6 nm + 673.1 nm). H$\alpha$ always peaks at the shock front, as expected, and [\SII] peaks farther from the shock front. Within the post-shock cooling region, the [\SII] emission is much stronger than H$\alpha$. Oxygen and nitrogen tracers have not been implemented in AstroBEAR as was done for sulfur, so the [\OI] and [\NII] lines are considered to be upper limits. Even without additional tracers, the code predicts [\OI] and [\NII] emission reasonably well for some models as compared to the CRSC values. \begin{figure} \centering \includegraphics[width=0.7\linewidth]{f3.eps} \caption{Radiative shock profile of model A. The \emph{red} lines are from the AstroBEAR model, and the \emph{black} lines are from CRSC. The top plot shows the emission lines, and the bottom plot shows the temperature $T$, number density compression ratio $n/n_o$, inverse plasma beta $P_B/P$, and electron number density ratio $n_e/n_o$. The temperature values use the left vertical axis, and the ratios use the right vertical axis. Comparing $n/n_o$ to $n_e/n_o$ gives a sense for how ionization fraction changes with distance. The inverse plasma beta $P_B/P$ is greater than 1 which means that the magnetic field pressure dominates over thermal pressure forces.} \label{fig:lowXlowM} \end{figure} Model B (Figure~\ref{fig:highXlowM}) shows how a higher initial ionization fraction affects the shock jump and post-shock quantities. Higher ionization leads to a lower initial post-shock temperature which is counter-intuitive but a correct result of how we chose our models. One would expect higher ionization to lead to higher temperatures, which is true in most of the post-shock cooling zone and is evidenced by the emission lines \citep{Cox85}. The differences in the initial post-shock temperatures in our models can be explained via the mean atomic weight. The presence of more electrons decreases the mean atomic weight which in turn increases the sound speed. Since the shock velocity and pre-shock temperature are fixed, the shock for Model B is effectively weaker than Model A (lower Mach number), hence the temperature jump is less. For a more straightforward comparison, the initial post-shock temperatures are listed in Table~\ref{tab:Lcools}. In the post-shock region of Model B, the ionization fraction increases initially as the gas equilibrates to the high temperature, and then it begins to decrease as recombination begins to dominate. Also note that with this higher ionization, the metal emission lines emit over a longer extent in the post-shock cooling region. \begin{figure} \centering \includegraphics[width=0.7\linewidth]{f4.eps} \caption{Radiative shock profile of model B. The \emph{red} lines are from the AstroBEAR model, and the \emph{black} lines are from CRSC.} \label{fig:highXlowM} \end{figure} For Model C, we again have a low ionization fraction as in Model A, but we have increased the shock velocity to 70 km/s. The higher shock velocity means that this shock is stronger, thus the temperature and density jumps are higher. These initial post-shock temperatures are so high that the ionization processes are dominant, and the gas quickly reaches ionization fractions that are even higher than Model B. We again see that despite the higher shock velocity and temperatures, AstroBEAR agrees fairly well with the CRSC models in all MHD quantities. The emission line profiles are also in good agreement, with the H$\alpha$ and [\SII] lines being most accurate. \begin{figure} \centering \includegraphics[width=0.7\linewidth]{f5.eps} \caption{Radiative shock profile of model C. The \emph{red} lines are from the AstroBEAR model, and the \emph{black} lines are from CRSC.} \label{fig:lowXhighM} \end{figure} Figure~\ref{fig:highXhighM} shows Model D, the high ionization, high shock velocity model. As we saw before, the higher ionization actually decreases the temperature and density jump. This particular model is the most extreme case, and AstroBEAR still captures the correct trends. Higher shock velocity and higher ionization models would be more challenging for AstroBEAR as \OIII~ and \NIII~ start to become important in these regimes. In the future, tracers could be added to more accurately model velocities $\geq 90$ km/s and pre-shock ionization fractions $\geq 0.5$. \begin{figure} \centering \includegraphics[width=0.7\linewidth]{f6.eps} \caption{Radiative shock profile of model D. The \emph{red} lines are from the AstroBEAR model, and the \emph{black} lines are from CRSC.} \label{fig:highXhighM} \end{figure} An important result from running these models is obtaining a value for the cooling length $L_{cool}$ We typically define $L_{cool}$ as the distance from the shock front at which the gas has cooled to a certain temperature, and here we have defined the cooling length temperature to be 4000 K. Cooling lengths were calculated for all four models for both AstroBEAR and CRSC, and the results are reported in Table~\ref{tab:Lcools}. On average, the values for $L_{cool}$ from both codes vary by approximately 20\% relative error. AstroBEAR does appear to cool faster, but it still recovers the expected behaviors and trends of radiative shocks. \begin{deluxetable}{cc p{40pt}|cc} \tablenum{1.3} \tablecolumns{5} \tablewidth{0pt} \tablecaption{Cooling lengths and initial post-shock temperatures of the 1-D shock models \label{tab:Lcools}} \tablehead{& \multicolumn{2}{c}{Cooling length (AU)} & \multicolumn{2}{c}{Post-shock temperature (10\textsuperscript{4} K)} \\ \colhead{Model} & \colhead{AstroBEAR} & \colhead{CRSC} & \colhead{AstroBEAR} & \colhead{CRSC}} \startdata A & 8417 & 9793 & 2.54 & 2.50 \\ B & 732 & 1029 & 2.18 & 2.12 \\ C & 383 & 397 & 11.32 & 11.24 \\ D & 203 & 270 & 8.60 & 8.55 \\ \enddata \end{deluxetable} For any simulation that uses cooling, it is important to know the cooling lengths of the strongest radiative shocks within the grid because stronger shocks cool faster and will have a shorter $L_{cool}$. As discussed in \citet{Hansen15b}, the cooling length resolution is crucial to resolving instabilities caused directly or indirectly by strong cooling. \citet{Yirak10} found that a cooling length resolution of at least 10 cells per $L_{cool}$ was required to resolve such instabilities. For any AstroBEAR simulation that uses the new non-equilibrium cooling functions, a separate 1-D radiative shock simulation seeded with parameters from the full 2-D or 3-D simulation of interest can be a helpful guide. The 1-D models can give a value for $L_{cool}$ and thus a minimum for what the physical resolution should be if we want to resolve cooling processes within post-shock regions which will also lead to better values for the emission line calculations. In Figure~\ref{fig:DMcompare} we compare a shock profile from the new non-equilibrium cooling to the previously used DM cooling. This Figure only shows temperature down to 1000 K, and it is for model C. We see that with the new implementation, AstroBEAR cools somewhat faster, especially in the temperature range of the new 3-D metal cooling table. The new implementation is also more consistent with CRSC; the DM cooling model resulted in a cooling length that is approximately 10 times too long. DM cooling is still quite good at high temperatures, but it is clear that a non-equilibrium treatment is necessary for radiative shocks. One of the main reason non-equilibrium cooling is important for this application is that neutral H and He cool rapidly at high temperatures. \begin{figure} \centering \includegraphics[width=0.75\linewidth]{f7.eps} \caption{Improvement from previous DM cooling in AstroBEAR. The shock profile shown for comparison here is model C. The \emph{solid} line is the new non-equilibrium cooling in AstroBEAR, the \emph{dashed} is from the previously used DM cooling. Z cooling appears to be much stronger and leads to a much shorter cooling length for this shock.} \label{fig:DMcompare} \end{figure} \section{Applications to HH Objects} \label{sec:outflows} An important application of the new non-equilibrium cooling routines in AstroBEAR is simulations of HH objects. In previous studies, we have run simulations of pulsed jets \citep{Hansen15b} and interacting bow shocks \citep{Hansen17}. The example outflow and synthetic emission map we show here is similar to the simulations from the latter work. We initialize a 3-D grid with an overdense clump of radius 10 AU which moves supersonically, at M = 15, producing a radiative bow shock. The stationary ambient number density, temperature, and ionization fraction are 1000 cm\textsuperscript{-3}, 1250 K, and 0.01 respectively. These values are consistent with typical values in regions with YSO jets \citep[see][and references therein]{Frank14}. The clump is initialized with a density of 5 x 10\textsuperscript{5} cm\textsuperscript{-3}, temperature of 2000 K, and ionization fraction of 10\textsuperscript{-3}. The choice of initial conditions here is not crucial since the purpose of the simulations was to drive a radiative bow shock into the ambient. Supersonic clumps are a simple way to generate bow shocks in simulations, and these values do not affect the conclusions we can make about the emission from such bow shocks. We see the emission from a radiative bow shock in Figure~\ref{fig:emissmap} where H$\alpha$ is in green and [\SII] is in red. The emission map plots the ratio [\SII]/H$\alpha$ as a color such that it is yellow when the ratio approaches unity. We see that the green H$\alpha$ is strongest near the shock front, and the emission reddens as the gas cools in the post-shock region as expected. This is consistent with the 1-D radiative shock models from the previous Subsection. \begin{figure} \centering \includegraphics[width=0.5\linewidth]{f8.eps} \caption{Synthetic emission map from a radiative bow shock. The ratio [\SII]/H$\alpha$ is plotted such that green shows H$\alpha$, red shows [\SII], and yellow shows regions where this ratio approaches unity. The bow shock was created in simulation by launching a supersonic, over-dense clump.} \label{fig:emissmap} \end{figure} The use of non-equilibrium cooling and synthetic emission map production make it possible to compare simulation results with actual HST observations of HH objects. Figure~\ref{fig:HHcompare} shows an HST image of the bow shock of HH 1 from 2007. We again see bright H$\alpha$ at the shock front and [\SII] dominates in the post-shock cooling region. There are other similarities between Figure~\ref{fig:HHcompare} and Figure~\ref{fig:emissmap} such as the heterogeneous nature of the emission. These structures and other details of such outflows will are discussed in \citep{Hansen17}. \begin{figure} \centering \includegraphics[width=0.5\linewidth]{f9.eps} \caption{HST image of HH 1 from 2007. This bow shock shows H$\alpha$ and [\SII] features that are similar to those of the synthetic emission map of Figure~\ref{fig:emissmap}.} \label{fig:HHcompare} \end{figure} \section{Conclusions} \label{sec:ch1conc} We have presented the implementation of new non-equilibrium cooling routines into the MHD code AstroBEAR. These routines are robust and include ionization and recombination equations for various atomic species. The accuracy of the code was tested by comparing 1-D radiative shock models with CRSC. Despite large differences in the way each code handles the physics, the models were very similar with an average relative differences in cooling length of approximately 18\%. 1-D radiative shock models can also be used as a tool to determine the cooling length in other simulations. It is important to resolve $L_{cool}$ in order to accurately simulate cooling processes, including thermal instabilities, within post-shock regions. We also presented the implementation of post-processing routines in AstroBEAR built to produce synthetic emission maps of H$\alpha$ and [\SII]. The [\OI] and [\NII] lines are implemented in the same way as [\SII], and more lines may be added in the future. To produce other emission line fluxes, tracers for other species (e.g. \OIII) would have to be implemented, and the framework for implementing tracers already exists in the current version of AstroBEAR. With an accurate measure of temperature and ionization in post-shock regions, synthetic emission maps from AstroBEAR can be directly compared to HST images. We are currently using the new non-equilibrium cooling functionality in other research projects such as 3-D pulsed jets which will expand on the work done in \citet{Hansen15b}. There is much research that can be done with the current implementation, but further improvements to cooling in AstroBEAR can still be made. The cooling routines can be expanded to include more processes and track more species within the temperature regions relevant in this paper. Furthermore, much work is needed to implement accurate molecular cooling if simulations are to be conducted at temperatures below 1000 K. Molecular cooling is a challenging problem as it would need to include processes such as grain formation and photodissociation. \vspace{7 mm} \noindent\emph{Acknowledgements.} This work was supported by the Center for Integrated Research and Computing at the University of Rochester which provided computational resources. Financial support for this project was provided by the Laboratory for Laser Energetics at the University of Rochester; by Space Telescope Science Institute grants HST-AR-11251.01-A (2007), HST-AR-12128.01-A, and HST-AR013892.001-A; by the National Science Foundation under award AST-0807363; by the Department of Energy National Laser User Facility (NLUF) grants DE-NA0002037 and DE-NA0002722; by the Department of Energy under award DE-SC0001063. \bibliographystyle{apj}

0e646eaaa19b2a60d1d5fc6fe748a3b2467648ec

\section{Introduction} Recent advances in computer vision \cite{krizhevsky2012imagenet} and natural language processing \cite{sutskever2014sequence} can be attributed to the advent of deep learning and the presence of robust benchmarks to quantitatively measure progress, such as the ImageNet challenge \cite{russakovsky2015imagenet}. In the last few years, neural network-based function approximation has also proven successful in reinforcement learning, with AI agents now able to perform at superhuman levels in games like Go \cite{silver2016mastering} and the Atari \cite{mnih2015human} suite. Once again, research in Deep RL has been steered by the establishment of benchmarks like the Arcade Learning Environment \cite{bellemare2013arcade}, along with the OpenAI Gym interface \cite{brockman2016openai}, which has been widely adopted by the research community. These benchmarks are conspicuously missing 2 challenging characteristics: (a) abrupt context-dependent switching of strategy and (b) temporal sensitivity. For agents to operate in the real world, they need to be able to switch behaviors very abruptly, which necessitates (i) learning to identify critical points where behavior needs to change, and (ii) learning the different behaviors required in each context. Agents also need to have an understanding of time as an independent variable, along with the ability to adapt their behavior accordingly. While having no understanding of time as something that's always ticking might work for simulated or static real-world environments, it is not acceptable for real-world dynamic environments with moving entities and where decisions might have to be adaptively taken very quickly or very slowly, depending on the context. Since existing benchmarks do not focus on these properties, reinforcement learning research has not tackled these problems yet. In this paper, we introduce a challenging RL environment based on Space Fortress (SF) \cite{mane1989space}, an arcade-style game which was developed by psychologists in the 80s to study human skill acquisition, and is still used quite frequently \cite{towne2016understanding,destefano2016should}. The objective of the game is to fly a ship and destroy a fortress by firing missiles at it. The ship has to respect a minimum time difference between successive shots, while building up the fortress' vulnerability, and once the fortress becomes vulnerable, destroy it with a rapid double shot. As a RL testbed, Space Fortress possesses both the characteristics discussed above: context-dependent strategy change (change in required firing rate after the fortress becomes vulnerable) and time sensitivity (firing rate requirements independent of the agent's decision speed i.e., the frame rate). It also has a sparse reward structure, and, as we show, is not solved by any state-of-the-art RL algorithms such as Rainbow \cite{hessel2017rainbow}, Proximal Policy Optimization (PPO) \cite{schulman2017proximal} and Advantage Actor-Critic (A2C) \cite{mnih2016asynchronous}. While being an interesting and relevant challenge for reinforcement learning, the rich background on human skill acquisition research based on Space Fortress also makes it an attractive tool to study human-AI collaboration in a dynamic environment, compare skill acquisition techniques of humans vs artificial agents, and work on few-shot learning by leveraging lessons from cognitive architectures like ACT-R \cite{anderson2009can} which have previously learned the game with extremely high sample efficiency, albeit using handcrafted features and extensive domain knowledge. We make the following contributions. First, we present a new RL testbed that requires the agent to switch strategies abruptly based on context, and develop a conceptualization of time independent of its speed of decision making, and demonstrate empirically that performance on par with humans is beyond the capability of current state-of-the-art RL algorithms, even after relaxing the reward sparsity through shaping \cite{ng1999policy}. We identify the aspects of the game which can be varied to control both temporal and context sensitivity, allowing research on either \textit{in isolation}. Finally, we demonstrate that \textit{after introducing modifications} to ease identification of critical contexts, the PPO algorithm learns to play the game well enough to outperform humans, verifying that context insensitivity is the primary driver behind the poor performance of RL algorithms. We also present robust human benchmark results for Space Fortress, allowing future researchers to place new experimental results in context. We open-source\footnote{\url{https://github.com/agakshat/spacefortress}} the OpenAI Gym environment for Space Fortress as well as all the code used to run our experiments, to promote research in temporal and context-sensitive reinforcement learning algorithms. \section{Related Work} \label{sec:relatedwork} The Arcade Learning Environment (ALE) \cite{bellemare2013arcade} poses the challenge of building AI agents with competency across dozens of Atari 2600 games, like Space Invaders, Asteroids, Bowling and Enduro. Following the development of Deep Q Networks \cite{mnih2015human}, a lot of research in the RL community has focused on improving performance in one or more of the games with improvements like massive parallelization, sample efficiency \cite{wang2015dueling,schaul2015prioritized}, better exploration \cite{fortunato2017noisy,plappert2017parameter}, reward sparsity \cite{pathak2017curiosity,andrychowicz2017hindsight} and long-term strategies \cite{bacon2017option,kulkarni2016hierarchical}. In continuous control tasks on the MuJoCo testbed \cite{todorov2012mujoco}, on-policy actor critic methods \cite{schulman2017proximal,mnih2016asynchronous} have shown promise. \citet{bellemare2017distributional} estimated a probability distribution over the Q-value of a state (instead of just the mean of the Q-value), with greatly improved results. Rainbow \cite{hessel2017rainbow} combined a lot of orthogonal improvements in DQNs to achieve state of the art results. However, we show below that these algorithms fail to learn anything on Space Fortress. Games like Ms. Pacman and Seaquest in the ALE have previously required some context or temporal sensitivity, but these characteristics can't be controlled or varied, and form a minor part of the overall game. As a RL testbed, Space Fortress relies heavily on both context and temporal sensitivity, as we show in Section \ref{sec:experiments}, and both characteristics can be controlled directly to enable their study in isolation. There has also been a fair amount of prior work on reinforcement learning with sparse rewards. \citet{pathak2017curiosity} use curiosity as an intrinsic reward signal to efficiently direct exploration. State visitation counts have also been investigated for exploration \cite{bellemare2016unifying}, and \citet{osband2016deep} train multiple value functions and make use of bootstrapping and Thompson sampling for exploration. These works focus on learning with sparse rewards through better exploration of the state space, which does not help with Space Fortress where exploration is required in time and in latent contexts. \citet{zambrano2015continuous} trained agents to deal with actions that take a finite amount of time through neural reinforcement learning in grid worlds, which still did not require a conceptualization of time independent of the internal speed of decision making, hence differing from the proposed work. Finally, \citet{van2017higher,van2017towards} previously used A3C on a simple control task abstracted from Space Fortress, with no fortress destruction required. Crucially, this task removed the interesting characteristics of Space Fortress, namely contextual and temporal sensitivity, as well as reward sparsity. We release an implementation of the game as an OpenAI Gym environment to promote research, conduct an ablation study to ascertain the roles of context and temporal sensitivity and reward sparsity in poor performance, and then present results showing existing RL algorithms outperforming humans after we control for the above factors. \section{The Space Fortress RL Environment} \label{sec:sf} \begin{figure*}[t] \centering \begin{subfigure}{0.3\textwidth} \centering \includegraphics[width=\textwidth]{sfedit.png} \caption{} \end{subfigure} ~ \begin{subfigure}{0.3\textwidth} \centering \includegraphics[width=\textwidth]{w3.png} \caption{} \end{subfigure} ~ \begin{subfigure}{0.3\textwidth} \centering \includegraphics[width=\textwidth]{w2.png} \caption{} \end{subfigure} \caption{Game Screens in \textbf{Space Fortress}. The ship has to fly between the two hexagons, while the fortress can only change its orientation at a fixed position. The game score is displayed at the top, and the fortress' vulnerability is displayed as a bar which fills up on each shot. (a) The bar is empty, indicating that the fortress' vulnerability is 0, (b) The bar is now full, indicating that vulnerability is equal to 10 and a rapid double shot will now destroy the fortress and (c) The fortress has been destroyed. This is followed by a reset of the fortress and continuation of the game till end of episode (3 minutes game time)} \label{fig:sf} \end{figure*} We now describe the Space Fortress game, discuss its utility as a testbed for reinforcement learning, and present results from humans learning to play the game, intended as a baseline. The game environment can be seen in Fig. \ref{fig:sf}. \subsection{Game Description} \label{sec:gamedescription} The player/AI agent controls a ship, which has to fly around in a frictionless arena, firing missiles to destroy a fortress located centrally within the arena. Hitting the walls on either sides or being hit by shells fired by the fortress results in immediate ship death, which incurs a penalty on the agent. Destroying the fortress, however, requires a \textit{context-aware strategy}. Each missile that successfully hits the fortress increases its vulnerability $v$ by one. When $v<10$, the fortress is 'not vulnerable', and the ship must fire its missiles spaced \textit{more than 250ms} apart. Firing faster than this while $v<10$ leads to the fortress vulnerability getting reset back to zero. This is obviously undesirable and the agent must learn to shoot slowly. However, once $v = 10$, the fortress becomes vulnerable, and a rapid double fire (2 shots spaced \textit{less than 250ms} apart) is required to destroy the fortress. We refer to this 250ms time specification as the ``critical time interval''. It is important to note that once $v = 10$, shooting further at the fortress at a rate less than 4Hz will lead to no change in vulnerability. Hence the firing \textit{strategy completely reverses} at the point when vulnerability reaches 10, and the agent must learn to identify this critical point to perform well. Since the game is simply reset (without ending the episode) when the fortress is destroyed, it is crucial that the agent also recognize this second critical point of fortress destruction, and switch back its firing rate to continue playing well. This major dependence on contextual and temporal sensitivity is unique to Space Fortress among RL benchmarks. A single game lasts for 3 minutes. The game does not end in the event of either a fortress or ship destruction, and points are scored by destroying the fortress as many times as possible in those 3 minutes while avoiding getting shot down by the fortress or colliding with the arena. When a fortress is destroyed, its vulnerability resets to zero, and the game continues. When the ship is destroyed, it respawns at a random position and orientation, but the fortress' vulnerability is preserved. \subsection{Game Versions} Space Fortress requires the agent to master advanced controls in a frictionless environment, orienting and firing missiles at the fortress while avoiding shells and not colliding with the walls. Since current RL algorithms proved unable to solve the game in its entirety (see experiments in Section \ref{sec:defaultrewards}), we introduced another version of the game to reduce navigation complexity by having the ship automatically pointed at the fortress. Throughout the rest of the paper, the simpler version is called `Autoturn', while the original game is referred to as `Youturn'. \subsection{Human Evaluations} \label{sec:humanevaluations} The human player results were collected by the authors in the context of a study on human skill acquisition \cite{anderson_betts_bothell_hope_lebiere_2018}. 117 people were asked to play 20 games of Space Fortress, with 52 playing Autoturn and 65 playing Youturn. They were all \textit{given instructions about the rules of the game beforehand}, and told about the change in firing rate required when the fortress vulnerability reaches 10. Considering that humans would require some turns to learn to play the game, we report the following results in Table \ref{tab:humanresults}: (1) Best performance of any subject in any game, (2) Average performance of all subjects in the last 5 games, considering the first 15 as a learning phase, (3) Average performance of all subjects in the last 10 games, considering the first 10 as a learning phase, (4) Average performance of all subjects in the last 15 games, considering the first 5 as a learning phase and (5) Average performance of all subjects in all 20 games. The scores shown to the humans (and reported in Table \ref{tab:humanresults}) were as follows: +100 for fortress destruction, -100 for ship death and -2 for each missile shot to penalize excessive firing. \begin{table*} \centering \begin{tabular}{|c|c|c|c|c|c|c|c|} \hline \textbf{Game} & \textbf{N} & \textbf{Metric} & \textbf{Best} & \textbf{Last 5} & \textbf{Last 10} & \textbf{Last 15} & \textbf{All}\\ \hline \multirow{2}{*}{Autoturn} & \multirow{2}{*}{52} & Score & 3000 & 1989 & 1978 & 1940 & 1810 \\ & & FortressDeath & 40 & 30.311 & 30.044 & 29.591 & 28.181\\ \hline \multirow{2}{*}{Youturn} & \multirow{2}{*}{65} & Score & 2314 & 216 & 153 & 43 & -169 \\ & & FortressDeath & 32 & 14.36 & 13.704 & 12.882 & 11.4\\ \hline \end{tabular} \caption{Aggregated results for 102 humans playing Space Fortress. After being provided with instructions about the rules beforehand, each player played the game for 1 hour, or 20 games. Allowing for a few practice games, we report the average scores on the Last `$K$' games ($K \in \{5,10,15\}$), as well as the best individual score.} \label{tab:humanresults} \end{table*} \begin{table*} \centering \begin{tabular}{|c|c|c|>{\centering}m{1.3cm}|>{\centering}m{1.3cm}|m{1.2cm}|} \hline \textbf{S.No.} & \textbf{Algorithm } & \textbf{Game } & \textbf{Avg. Score } & \textbf{Best Score} & \textbf{Fortress Death} \\ \hline 1 & A2C & Autoturn & -2685 & -2242 & 0 \\ 2 & A2C & Youturn & -5859 & --5604 & 0 \\ \hline 3 & PPO & Autoturn & -2502 & -2178 & 0 \\ 4 & PPO & Youturn & -5269 & -4698 & 0 \\ \hline 3 & Rainbow & Autoturn & -8327 & -8264 & 0 \\ 4 & Rainbow & Youturn & -9378 & -9245 & 0 \\ \hline \end{tabular} \caption{Average game scores for RL agents, trained \textbf{with default (sparse) rewards}, for 45M steps} \label{tab:originalgameresults} \end{table*} \subsection{RL Setup} \label{sec:rlsetup} We now describe the exact game setup used for reinforcement learning on Space Fortress. \begin{itemize} \item Observations: The observations are in the form of pixel-level grayscale 84x84 size renderings of the game screen (similar to Fig. \ref{fig:sf}. Important information such as the time lapsed since the last shot is not a part of this observation, making the task partially observed. We provide the agent with a stack of the last 4 observations as input at each time step, allowing it to infer direction of movement of the ship and fortress using the difference between successive frames. \item Actions: The agent chooses from 5 actions: (i) No Operation, (ii) Fire (a missile), (iii) Thrust Forward (in the direction of current orientation), (iv) Thrust Right (rotate right without changing position) and (v) Thrust Left (rotate left without changing position). The game operates at a default frame rate of 30 FPS and there is no action repeat, which means an action is chosen every 33ms. Note that the Autoturn version only has 3 actions (since no turning is required). \item Rewards: In line with \citet{mnih2015human}, we found that learning was more stable when using clipped rewards. The fortress and ship destruction rewards were clipped to +1 and -1, respectively, and the missile penalty reduced to -0.05. Note that the results used for evaluation and reporting were not clipped, in order to follow the same scheme as described in Section \ref{sec:humanevaluations}. \end{itemize} \begin{table*} \centering \begin{tabular}{|c|c|c|>{\centering}m{1.3cm}|>{\centering}m{1.3cm}|m{1.2cm}|} \hline \textbf{S.No.} & \textbf{Algorithm } & \textbf{Game } & \textbf{Avg. Score } & \textbf{Best Score} & \textbf{Fortress Death} \\ \hline 1 & A2C & Autoturn & -4116 & -2100 & 0 \\ 2 & A2C & Youturn & -4781 & -3890 & 1.3 \\ \hline 3 & PPO & Autoturn & -1294 & -1108 & 1 \\ 4 & PPO & Youturn & -1435 & -1206 & 0.94 \\ \hline 3 & Rainbow & Autoturn & -6161 & -5960 & 0 \\ 4 & Rainbow & Youturn & -4894 & -4577 & 0 \\ \hline \end{tabular} \caption{Average game scores for RL agents, trained \textbf{with dense rewards}, for 45M steps.} \label{tab:denseresults} \end{table*} \begin{table*}[t] \centering \begin{tabular}{|c|c|c|c|>{\centering}m{1.3cm}|>{\centering}m{1.3cm}|m{1.2cm}|} \hline \textbf{S.No.} & \textbf{Algorithm } & \textbf{Architecture} & \textbf{Game } & \textbf{Avg. Score} & \textbf{Best Score} & \textbf{Fortress Death} \\ \hline 1 & A2C & SF-GRU & Autoturn & -1641 & -718 & 3 \\ 2 & A2C & SF-GRU & Youturn & -2444 & -1700 & 11 \\ \hline 3 & PPO & SF-FF & Autoturn & 2337 & 2818 & 41 \\ 4 & PPO & SF-FF & Youturn & 2235 & 2880 & 40 \\ \hline 5 & PPO & SF-GRU & Autoturn & 2510 & 2870 & 43 \\ 6 & PPO & SF-GRU & Youturn & 2356 & 2932 & 41 \\ \hline 7 & Rainbow & -- & Autoturn & -2973 & -2330 & 1.2\\ 8 & Rainbow & -- & Youturn & -4112 & -3934 & 0.0\\ \hline \end{tabular} \caption{Average game scores for RL agents, trained \textbf{after making context identification easier}, for 45M steps} \label{tab:easyresults} \end{table*} \section{Experiments and Results} \label{sec:experiments} In this section, we experimentally show that (a) no state-of-the-art reinforcement learning algorithm (Rainbow \cite{hessel2017rainbow}, A2C \cite{mnih2016asynchronous} and PPO \cite{schulman2017proximal}) can learn to play Space Fortress, (b) removing reward sparsity does not improve the performance and (c) making context identification easier through specific alterations in the reward structure allow PPO to achieve superhuman performance. We also discuss temporal sensitivity by examining effectiveness of transfer of learning across different settings of the game's critical time interval. \subsection{Network Architecture} For Rainbow, the Q-network architecture was identical to that in \citet{hessel2017rainbow}. For PPO and A2C, we experiment with two policy network architectures: \begin{itemize} \item \textbf{SF-GRU: } The agent's policy network takes the 1x84x84 environment observations as input, and outputs (a) a probability distribution over the actions, and (b) a value function estimate of the expected return. The input goes through two convolutional layers with 16, 32 filters of size 8,4 and stride 4,2 respectively, and ReLU activation. The output is flattened and passed through a linear layer with a ReLU non-linearity to get an output vector of size 256. This is then passed through a unidirectional Gated Recurrent Unit (GRU) cell \cite{cho2014learning} with a tanh non-linearity giving an output of size 256. Finally, this vector is passed as input to two linear layers that output the probability distribution over actions (using a softmax activation) and the value estimate of the expected return. \item \textbf{SF-FF: } Same as above, but with a fully connected layer of size 256 with ReLU non-linearity instead of the recurrent GRU cell. \end{itemize} For all experiments, we ran 16 processes collecting game experience in parallel, with discount factor $\gamma = 0.99$ and Generalized Advantage Estimation (GAE) \cite{schulman2015high} parameter $\lambda = 0.95$. PPO used value loss coefficient $c_{1} = 0.5$, entropy regularization coefficient $c_{2} = 0.05$ and learning rate $\num{1e-3}$, while A2C used $c_{1} = 0.5, c_{2} = 0.01$ and learning rate $\num{5e-4}$. Both A2C and PPO used $n = 1024$-step returns. These hyperparameters were found after extensive tuning. We also clipped the gradients of all the network parameters to 0.5, to prevent catastrophic updates from outlying samples of the expected gradient value. Since the PPO algorithm is more stable, we updated the policy 4 times every epoch - while A2C made only 1 update every epoch. \begin{figure*}[t] \centering \begin{subfigure}{0.48\textwidth} \centering \includegraphics[width=\textwidth]{autoturn_ppo.png} \caption{Autoturn} \label{fig:ppo_autoturn} \end{subfigure} ~ \begin{subfigure}{0.48\textwidth} \centering \includegraphics[width=\textwidth]{youturn_ppo.png} \caption{Youturn} \label{fig:ppo_youturn} \end{subfigure} \caption{Learning Curves for PPO on Space Fortress, for different reward structures and architectures. 'Average Human Score' refers to the average score over all 20 games, provided as a point of comparison. 'Default Rewards' discussed in Section \ref{sec:defaultrewards}, 'Dense Rewards' in Section \ref{sec:denserewards} and 'AECI' (After Easing Context Identification) in Section \ref{sec:easingcontextid}. Both SF-FF and SF-GRU architectures are able to achieve superhuman performance \textit{after making context identification easier}. The agent's performance is very poor with both default (sparse) rewards and dense rewards.} \label{fig:ppo_curves} \end{figure*} \begin{figure*} \centering \begin{subfigure}{0.32\textwidth} \centering \includegraphics[width=\textwidth]{temporal_transfer_125.png} \caption{125ms} \label{fig:temporal_125} \end{subfigure} ~ \begin{subfigure}{0.32\textwidth} \centering \includegraphics[width=\textwidth]{temporal_transfer_400.png} \caption{400ms} \label{fig:temporal_400} \end{subfigure} ~ \begin{subfigure}{0.32\textwidth} \centering \includegraphics[width=\textwidth]{temporal_transfer_600.png} \caption{600ms} \label{fig:temporal_600} \end{subfigure} \caption{Checking for positive transfer of learning while changing the critical time interval in Space Fortress. In each figure, we changed the time interval to a different value and verified whether transferring learned weights from the agent trained on 250ms as the critical interval led to any speedup in the learning process, and an improvement in the final performance.} \label{fig:temporal} \end{figure*} \subsection{With Default (Sparse) Rewards} \label{sec:defaultrewards} With the default sparse reward structure which rewards fortress destruction and penalizes ship destruction and missile firing, no algorithm is able to learn to destroy the fortress. A visual inspection of the game play revealed that the PPO and A2C agents (with both architectures) just learned to stop firing, since that leads to an immediate penalty. The Rainbow agent did not learn anything. Table \ref{tab:originalgameresults} presents the aggregated results for PPO, A2C and Rainbow on both versions of the game. The `Fortress Death' column in Table \ref{tab:originalgameresults} indicates the number of times the agent was able to destroy the fortress per game, on average. \subsection{With Dense Rewards} \label{sec:denserewards} Considering the results in Section \ref{sec:defaultrewards} and to understand how reward sparsity is impacting performance, we introduce an additional reward of +1 each time the fortress is hit by a missile, and a penalty of -1 if the fortress' vulnerability gets reset due to a faster firing speed than the context demands. This makes the reward density comparable to Atari games, on which PPO, A2C and Rainbow have all been shown to perform well. Their performance on Space Fortress with dense rewards can be seen in Table \ref{tab:denseresults}, where the scores for PPO and Rainbow have improved. From watching a video of the trained agent playing the game, we observed that the improvement stemmed from having learned to avoid ship death and to fire at the fortress, albeit without knowledge of the critical time interval and context-dependent strategy shifts, resulting in an inability to destroy the fortress with any consistency. Hence, Space Fortress clearly presents a challenge to the state of the art in reinforcement learning, and is a useful and relevant benchmark for further research. We now move on to studying the impact of context insensitivity of RL algorithms on the task performance (Section \ref{sec:easingcontextid}), and show that by making the identification of critical contexts easier with 2 simple modifications to the reward, PPO learns to play the game very well - outperforming humans comfortably. This clearly indicates that it is context insensitivity and the inability to identify critical points which is hampering performance, further making the case that the Space Fortress game is a useful benchmark for studying context sensitivity. \subsection{After Making Context Identification Easier} \label{sec:easingcontextid} As discussed in Section \ref{sec:gamedescription}, there are 2 critical points which the agent has to learn to identify and switch strategies at. The first is when the fortress becomes vulnerable, i.e. $v = 10$ and the agent has to switch from firing more than 250ms apart to a rapid double shot fired less than 250ms apart. The second is when the fortress is destroyed, and the agent has to switch back to its slow firing speed. To confirm our intuition that it is indeed the algorithms' inability to identify these critical points and accordingly adapt its firing strategy which lead to poor performance, we introduce 2 changes to the reward structure (with respect to the dense reward from Section \ref{sec:denserewards}) which make it trivial for the agent to identify the critical points where context changes: \begin{itemize} \item Instead of rewarding fortress hits (as in Section \ref{sec:denserewards}), we switch to rewarding fortress vulnerability change, by giving a reward of +1 for unit increase in vulnerability, and giving a penalty of -1 for decrease in vulnerability. This has the effect of rewarding fortress hits only until the fortress' vulnerability is building up to 10, at which point further hits are not rewarded. This clearly helps it identify the critical context at which the fortress becomes vulnerable. \item We give the agent a bonus reward of +2 for fortress destruction, to help it identify when the fortress is destroyed. \end{itemize} Table \ref{tab:easyresults} presents the results for agents trained after these 2 changes have been introduced to the reward structure to ease context identification. PPO with the recurrent architecture SF-GRU achieves the best performance in both score and number of fortress deaths, learning faster than SF-FF, and achieving a higher final score. The performance of both A2C and Rainbow also improves, although they are still unable to outperform humans. Fig. \ref{fig:ppo_curves} tracks the learning curves for PPO learning with all 3 reward settings (default, dense and after making context identification easier) in both game versions - Autoturn and Youturn. \subsection{Temporal Sensitivity} Having established that context insensitivity is the primary driver of poor performance of state of the art RL algorithms on Space Fortress, we now analyze the temporal sensitivity of the PPO algorithm. As described in Section \ref{sec:gamedescription}, Space Fortress has a dominant temporal aspect - missiles must hit the fortress at least 250ms apart when it is not vulnerable, and then the strategy must reverse to hit the fortress \textit{twice within 250ms} when it is vulnerable, in order to destroy it. In order to understand whether the RL algorithms had developed any understanding of time as an independent dimension, we modified the critical time interval from 250ms to other values, and checked for positive transfer of learning from the policy trained with 250ms as the critical time interval. We achieve transfer of learning by simply initializing the weights of the transferee with the learned weights of the transferer. Figure \ref{fig:temporal} compares the learning curves for an agent learning with PPO (using the SF-FF architecture) on Youturn, when the critical time interval is changed from 250ms to $\{125,400,600\}$ms. The blue line is for an agent learning from scratch, while the orange line is for an agent transferring learning from the PPO SF-FF agent trained on the 250ms interval. From Figure \ref{fig:temporal_400} and \ref{fig:temporal_600}, it can be seen that while the transfer of learning helps by initializing the weights in a favorable corner of the parameter space, the learning saturates very quickly and ends up with a final score much lower than achieved when the critical interval was 250ms. Modifying the critical time interval in Space Fortress is a useful technique to study temporal sensitivity of reinforcement learning algorithms. \section{Conclusion} This paper introduced Space Fortress as a new challenge for deep reinforcement learning research, with its time-sensitive game play, abrupt context-dependent shift of strategy and sparse rewards. We showed that state of the art RL algorithms (PPO, A2C and Rainbow) were unable to learn to play the game with neither the default sparse rewards nor the dense reward structure we defined. After making context identification easier through two minor tweaks in the reward structure, however, PPO was able to learn to play the game, outperforming humans comfortably. This ablation study allowed us to conclude that context insensitivity was the primary reason behind the poor performance of RL algorithms on Space Fortress, along with the inability to learn with sparse rewards. We then looked at whether PPO develops a concept of time as an independent variable - by checking for positive transfer of learning while changing the critical time interval of 250ms in Space Fortress. We found that while there was some positive transfer of learning, the agents saturated very quickly and did not achieve a good final score. By studying generalization and transfer across different settings of the critical time interval, Space Fortress can hence also be used as a benchmark to study temporal sensitivity of reinforcement learning algorithms. Learning to play Space Fortress without making any modifications to the reward structure will require reinforcement learning algorithms to be able to identify various latent contexts and adapt their strategies suitably. It will also require being able to learn with very sparse rewards. This is beyond the capability of current state of the art reinforcement learning algorithms, making Space Fortress a useful benchmark for research. \subsubsection*{Acknowledgments} This research was sponsored by AFOSR Grant FA9550-15-1-0442. The collection of human data and development of the OpenAI Gym interface for Space Fortress was supported by ONR grant N00014-15-1-2151. We would like to thank Shawn Betts and John Anderson for insightful discussions on the game of Space Fortress, and for the OpenAI Gym interface for Space Fortress which was used to run the experiments in this paper.

39fc2b1a9ba247a00bd594e02ed79a107981390b

\chapter{Composite boson bags in strong coupling QED \label{cha:qed}} At the beginning of this year yet another reformulation of strong coupling QCD with one flavor of staggered fermions was derived in \cite{Gattringer:2018mrg}. In this publication the author shows that the baryonic degrees of freedom in the QCD path integral are independent of the gauge fields. Moreover, he is able to separate the baryonic contributions from the remaining quark and diquark terms in the partition function. Finally he demonstrates that, in the strong coupling limit, the partition function completely factorizes into so-called \textit{baryons bags} (space-time regions in which the dynamics is given by free baryons) and a complementary domain (where the dynamical degree of freedoms are monomers and dimers of quarks and diquarks). As a natural continuation of this work, we aim to apply the baryon bag idea to strong coupling QCD with two flavors. In this appendix we present an intermediate project we studied as a toy model: strong coupling QED with two flavors of staggered fermions of opposite charge. In what follows we will show that in the case of QED the path integral factorizes into composite boson bags (CBB), where composite boson modes of two fermions propagate freely, and a complementary domain with monomers and dimers for the two flavors of fermions. For illustration one may think of the two flavors as "electrons" and "protons" and the composite bosons as "hydrogens". We will partly use this nomenclature in this appendix. Before coming to the concrete discussion of the model we would like to remark that the idea that fermionic degrees of freedom can be described as being contained inside dynamically determined space-time regions was firstly introduced in \cite{PhysRevD.82.025007} and developed into the so-called \textit{fermion bags} approach. This method constitutes a very powerful tool for treating the fermion sign problem for some fermionic lattice field theories \cite{PhysRevLett.108.140404,PhysRevD.85.091502, PhysRevD.86.021701, Chandrasekharan2013, PhysRevD.88.021701, PhysRevD.91.065035, PhysRevD.93.081701, Ayyar2016, PhysRevD.96.114502, PhysRevD.97.054501}. Among those works we outline \cite{Chandrasekharan2011}, in which the fermion bag approach was applied to strong coupling QED with one flavor of Wilson fermions. \section{Gauge integral at strong coupling} We consider strong coupling lattice QED with two flavors of staggered fermions with opposite charge. The corresponding action is given by \begin{align} S[\overline{e},e,\overline{p},p,U]\ &= \ \sum_{x} \bigg[ 2m_{e} \, \overline{e}_{x} e_{x} \, + \, \sum_{\nu} \gamma_{x,\nu} \Big( \overline{e}_{x} \, U_{x,\nu} \, e_{x + \hat{\nu}} \, - \, \overline{e}_{x + \hat{\nu}} \, U_{x,\nu}^{\, \star} \, e_{x} \Big) \bigg] \nonumber \\ &\, + \ \sum_{x} \bigg[ 2m_{p} \, \overline{p}_{x} p_{x} + \sum_{\nu} \gamma_{x,\nu} \Big( \overline{p}_{x} \, U_{x,\nu}^{\, \star} \, p_{x + \hat{\nu}} \, - \, \overline{p}_{x + \hat{\nu}} \, U_{x,\nu} \, p_{x} \Big) \bigg] \, . \label{eq:qed_action} \end{align} The fermionic degrees of freedom are the one-component Grassmann variables $e_{x}$, $\overline{e}_{x}$ for the electrons and $p_{x}$, $\overline{p}_{x}$ for the protons. They obviously have opposite charge, as is reflected in the expression of the action $S[\overline{e},e,\overline{p},p,U]$ where the role of the link variable $U_{x,\nu}$ and its complex conjugate $U_{x,\nu}^{\, \star}$ is interchanged for the forward and backward propagation of the electron and the proton. This also ensures overall electric neutrality of the two-flavor system as required by Gauss’ law. The gauge degrees of freedom are the link variables $U_{x,\nu} = e^{\,i\varphi_{x,\nu}} \in U(1)$. They live on the links $(x,\nu)$ of a four-dimensional lattice $\Lambda$ with volume $N_{s}^{3} \times N_{t}$ and they satisfy periodic boundary conditions in all directions. For the fermions, instead, we impose anti-periodic boundary conditions in the time ($\nu=4$) direction, and periodic boundary conditions for the spatial directions ($\nu=1,2,3$). In Eq.~(\ref{eq:qed_action}) $m_{e}$ and $m_{p}$ are the electron's and proton's bare masses respectively, while $\gamma_{x,\nu}$ are the usual staggered sign factors defined as \begin{equation} \label{eq:qed_staggered} \gamma_{x,1} = 1\, , \qquad \gamma_{x,2} = (-1)^{x_{1}}\, , \qquad \gamma_{x,3} = (-1)^{x_{1}+x_{2}}\, , \qquad \gamma_{x,4} = (-1)^{x_{1}+x_{2}+x_{3}} \, . \end{equation} The partition function of the system is obtained as \begin{align} Z \, &= \, \int \! D[\overline{e},e] D[\overline{p},p] \int \! D[U] \, e^{\,S[\overline{e},e,\overline{p},p,U]} \nonumber \\ &= \, \int \! D[\overline{e},e] D[\overline{p},p] \, \prod_{x} e^{\,2 m_{e}\, \overline{e}_{x} e_{x}} \, e^{\,2 m_{p}\, \overline{p}_{x} p_{x}} \int \! D[U] \, \prod_{x,\nu} L_{x,\nu}[\overline{e},e,\overline{p},p,U]\, , \label{eq:qed_partition} \end{align} where $\int \! D[\overline{e},e] D[\overline{p},p]$ are the product of Grassmann measures at all sites \begin{equation} \int \! D[\overline{e},e] D[\overline{p},p] \, = \, \prod_{x} \int dp_{x} d\overline{p}_{x} de_{x} d\overline{e}_{x} \, , \label{eq:qed_grassmannmeasure} \end{equation} while $\int \! D[U]$ is the product of U(1) Haar measures at all links \begin{equation} \int \! D[U] \, = \, \prod_{x,\nu} \, \int_{U(1)} dU_{x,\nu} \, = \, \prod_{x,\nu} \int_{0}^{2\pi} \dfrac{d\varphi_{x,\nu}}{2\pi} \, . \label{eq:qed_haarmeasure} \end{equation} In the second step of Eq.~(\ref{eq:qed_partition}) we used the fact that every term in the action (\ref{eq:qed_action}) is a commuting Grassmann bilinear to factorize the Boltzmann weight $e^{\,S[\overline{e},e,\overline{p},p,U]}$ into local factors. Moreover, we defined the link terms \begin{equation*} L_{x,\nu}[\overline{e},e,\overline{p},p,U] \, = \, e^{\,\gamma_{x,\nu}\, \overline{e}_{x}\, U_{x,\nu}\, e_{x + \hat{\nu}}} \, e^{-\gamma_{x,\nu}\, \overline{e}_{x + \hat{\nu}}\, U_{x,\nu}^{\star}\, e_{x}} \, e^{\,\gamma_{x,\nu} \, \overline{p}_{x} \, U_{x,\nu}^{\star} \, p_{x + \hat{\nu}}} \, e^{-\gamma_{x,\nu} \, \overline{p}_{x + \hat{\nu}} \, U_{x,\nu} \, p_{x}} \, . \end{equation*} In order to separate the hydrogen terms (composite boson terms) in the QED path integral we start by expanding the Boltzmann factors for the mass terms $e^{\,2 m_{e}\, \overline{e}_{x} e_{x}}$ and $e^{\,2 m_{p}\, \overline{p}_{x} p_{x}}$. In the expansions we exploit the nilpotency of the Grassmann variables thus terminating the power series after the linear terms. We also drop all space-time indices for brevity. \begin{align} e^{\,2 m_{e}\, \overline{e} e} \, e^{\,2 m_{p}\,\overline{p} p}& = \big[ 1 + 2 m_{e}\, \overline{e} e\big] \big[ 1 + 2 m_{p} \, \overline{p} p\big] = 1 + 2 m_{e} \, \overline{e} e + 2 m_{p} \, \overline{p} p + 4 m_{e} m_{p} \, \overline{e} e \, \overline{p} p \nonumber \\ & = \big[1 + 4 m_{e} m_{p} \, \overline{e} e \, \overline{p} p \big] \big[1 + 2 m_{e} \, \overline{e} e + 2 m_{p} \, \overline{p} p \big] \nonumber \\ &= \exp\big(2 M \, \overline{H} H\big) \sum_{s^{(e)} = 0}^{1} \big[2m_{e} \, \overline{e} e\big]^{s^{(e)}} \sum_{s^{(p)} = 0}^{1-s^{(e)}} \big[2m_{p} \, \overline{p} p\big]^{s^{(p)}} \, . \label{eq:qed_mass} \end{align} In the second line we used once more the nilpotency of the Grassmann variables to separate the hydrogen contribution to the mass terms, which we collected in the first factor in brackets in the second line. We then re-exponentiated this factor and used the identities $2M = 4 m_{e} m_{p}$ and $\overline{e} e \overline{p} p = \overline{e} \overline{p} p e = \overline{H} H$, where $H_{x}$ and $\overline{H}_{x}$ are the hydrogen fields defined as \begin{equation} \label{eq:qed_hydrogen} H_{x} \, \equiv \, p_{x} e_{x} \, , \qquad \qquad \overline{H}_{x} \, \equiv \, \overline{e}_{x} \overline{p}_{x}\, . \end{equation} In the last step of Eq.~(\ref{eq:qed_mass}) we also introduced the monomer variables $s_{x}^{(e)} \in \{0,1\}$ for the mass term of the electron fields and $s_{x}^{(p)} \in \{0,1-s_{x}^{(e)}\}$ for the proton's mass term. As we will see later, this rewriting will turn out to be useful for solving the Grassmann integrals in (\ref{eq:qed_partition}). Notice also that the monomer variables for the proton fields are constrained by the value of the monomer variables for the electron fields at the same site. This is because the mixed term, which would correspond to setting $s_{x}^{(e)} = s_{x}^{(p)} = 1$, was already factorized into the hydrogen contribution. A second important step to completely separate the hydrogens in the path integral of QED consists in solving the link integral in Eq.~(\ref{eq:qed_partition}), which has the following form (we suppress the link indices to shorten the notation, and we omit even powers of the staggered sign factor, since $\gamma^{2n} = +1$): \begin{align} \int \! dU \, &e^{\,\gamma \,\overline{e}_{x} U e_{y}} \, e^{-\gamma \, \overline{e}_{y} U^{\star} e_{x}} \, e^{\,\gamma \, \overline{p}_{x} U^{\star} p_{y}} \, e^{-\gamma \, \overline{p}_{y} U p_{x}} \nonumber \\ & = \int \! dU \, (1 + \gamma \, \overline{e}_{x} U e_{y}) \, (1 - \gamma \, \overline{e}_{y} U^{\star} e_{x}) \, (1 + \gamma \, \overline{p}_{x} U^{\star} p_{y}) \, (1 - \gamma \, \overline{p}_{y} U p_{x}) \nonumber \\[0.2em] & = 1 - \overline{e}_{x} e_{y} \overline{e}_{y} e_{x} - \overline{p}_{x} p_{y} \overline{p}_{y} p_{x} + \overline{e}_{x} e_{y} \overline{p}_{x} p_{y} + \overline{e}_{y} e_{x} \overline{p}_{y} p_{x} + \overline{e}_{x} e_{y} \overline{e}_{y} e_{x} \overline{p}_{x} p_{y} \overline{p}_{y} p_{x}\nonumber \\[0.5em] & = \big[1 + \overline{e}_{x} \overline{p}_{x} p_{y} e_{y} + \overline{e}_{y} \overline{p}_{y} p_{x} e_{x} + \overline{e}_{x} \overline{p}_{x} p_{y} e_{y} \overline{e}_{y} \overline{p}_{y} p_{x} e_{x} \big] \big[1 + \overline{e}_{x} e_{x} \overline{e}_{y} e_{y} + \overline{p}_{x} p_{x} \overline{p}_{y} p_{y} \big] \nonumber \\ & = e^{\overline{H}_{x} H_{y} + \overline{H}_{y} H_{x}} \sum_{d^{(e)} = 0}^{1} \big[\overline{e}_{x} e_{x} \overline{e}_{y} e_{y}\big]^{d^{(e)}} \sum_{d^{(p)} = 0}^{1-d^{(e)}} \big[\overline{p}_{x} p_{x} \overline{p}_{y} p_{y}\big]^{d^{(p)}} \, . \label{eq:qed_link} \end{align} In the first step we Taylor expanded each exponential. The only terms of the resulting product of binomials that survive the Haar integration are the ones in which the link variable $U$ and its complex conjugate $U^{\star}$ have the same exponent. In fact, the link integrals are now of the type \begin{equation} \int_{U(1)} \! dU \, (U)^{n} (U^{\star})^{m} \, = \, \int_{0}^{2\pi}\! \dfrac{d\varphi}{2 \pi} \, e^{i \varphi (n-m)} \, = \, \delta_{n,m} \, . \end{equation} In the third step of (\ref{eq:qed_link}) we have separated the propagation terms for the hydrogen from the electron and proton dimer contributions, using again the nilpotency of the Grassmann variables. Subsequently, we re-exponentiated the hydrogen contribution, that now has the form of a Boltzmann weight for the forward and backward hydrogen hops. Finally, we introduced the dimer variables $d_{x,\nu}^{\,(e)} \in \{0,1\}$ for the electron fields and $d_{x,\nu}^{\,(p)} \in \{0,1-d_{x,\nu}^{\,(e)}\}$ for the proton fields. We again constrain the proton dimer variable with the value of the electron dimer variable at the same link since the configuration $d_{x,\nu}^{\,(e)} = d_{x,\nu}^{\,(p)} = 1$ is already taken into account in the hydrogen contribution. Putting things together we obtain the following expression for the partition sum: \begin{equation} Z \, = \, \int \! D[\overline{e},e] D[\overline{p},p] \, e^{\,S_{H}[\overline{H},H]} \, W_{MD}[\overline{e},e,\overline{p},p] \, . \label{eq:qed_partition2} \end{equation} In this form the Boltzmann weight $e^{\,S_{H}[\overline{H},H]}$ factorizes the contributions of the hydrogens, which are now completely separated from the electron's and proton's monomer and dimer contributions. The hydrogen action \begin{equation} S_{H}[\overline{H},H] = \sum_{x} \Big[2 M \,\overline{H}_{x} H_{x} + \sum_{\nu} \big( \overline{H}_{x} H_{x+\hat{\nu}} + \overline{H}_{x+\hat{\nu}} H_{x}\big)\Big] \, \label{eq:qed_hydrogenaction} \end{equation} has a structure similar to a bosonic action, where backward and forward hopping terms have the same sign. Nevertheless, even though the fields $H_{x}$ and $\overline{H}_{x}$ are indeed bosonic, they still have to satisfy Pauli's exclusion principle, since their fundamental constituents are fermions (see (\ref{eq:qed_commutation}) -- (\ref{eq:qed_Hbarnilpotent}) below for their algebraic relations). $W_{MD}[\overline{e},e,\overline{p},p]$ collects the monomer and dimer contributions of the electrons and protons to the partition sum: \begin{align} W_{MD}[\overline{e},e,\overline{p},p] \, = \, &\sum_{\{s,d\}} \prod_{x} \, \big[2m_{e} \, \overline{e}_{x} e_{x}\big]^{s^{(e)}_{x}}\, \big[2m_{p} \, \overline{p}_{x} p_{x}\big]^{s^{(p)}_{x}} \nonumber \\ &\, \, \ \times \prod_{x,\nu}\, \big[\overline{e}_{x} e_{x} \overline{e}_{x+\hat{\nu}} e_{x+\hat{\nu}}\big]^{d_{x,\nu}^{\,(e)}} \, \big[\overline{p}_{x} p_{x} \overline{p}_{x+\hat{\nu}} p_{x+\hat{\nu}}\big]^{d_{x,\nu}^{\,(p)}} \, , \label{eq:qed_wmd} \end{align} where we used the notation \begin{equation} \sum_{\{s,d\}} = \left[\prod_{x} \sum_{s^{(e)}_{x} = 0}^{1} \sum_{s^{(p)}_{x} = 0}^{1-s^{(e)}_{x}}\right] \left[\prod_{x,\nu} \sum_{d^{\,(e)}_{x,\nu} = 0}^{1} \sum_{d^{\,(p)}_{x,\nu} = 0}^{1-d^{\,(e)}_{x,\nu}}\right] \label{eq:qed_sumsd} \end{equation} to denote the sum over the configurations of electron's and proton's monomers and dimers. \section{Factorization of the Grassmann integral} Having completely integrated out the gauge fields we now discuss the Grassmann integration, starting with the hydrogen contributions. Firstly, notice that the Grassmann integrals in Eq.~(\ref{eq:qed_partition2}) are non-vanishing only if each Grassmann variable $\overline{e}_{x}$ ,$e_{x}$ and $\overline{p}_{x}$, $p_{x}$ appears exactly once. When this condition is fulfilled we say that the Grassmann integral is \textit{saturated}. It is then clear that hydrogen variables saturate the Grassmann integrals on the sites they occupy. In fact, the expansion of the hydrogen Boltzmann weight $e^{\,S_{H}[\overline{H},H]}$ results in terms (hydrogen monomers, dimers and non-intersecting loops) that already contain both the flavors of staggered fermions we are considering (compare with the definition (\ref{eq:qed_hydrogen}) for the hydrogen fields $\overline{H}_{x}$ and $H_{x}$). Therefore, if we denote with $\mathcal{H}_{i}$ the collection of sites that are either occupied by an hydrogen monomer, are the endpoint of an hydrogen dimer, or are run through by an hydrogen loop, we then find that in the region defined by $\mathcal{H}_{i}$ all the Grassmann integrals are saturated by the hydrogen terms. We refer to these regions as \textit{hydrogen bags}. Additionally, we can define the union $\mathcal{H}$ of all bags and the complementary domain $\overline{\mathcal{H}}$ \begin{equation} \mathcal{H} = \cup_{i} \mathcal{H}_{i} \, , \qquad \overline{\mathcal{H}} = \Lambda/\mathcal{H}\, . \label{eq:qed_complementary} \end{equation} The complementary region $\overline{\mathcal{H}}$ contains the electron's and proton's monomer and dimer contributions described by the weight factor $W_{MD}[\overline{e},e,\overline{p},p]$. Now, recall that the hydrogens $H_{x}$ and $\overline{H}_{x}$ are nilpotent bosons. In other words, they satisfy the following properties: \begin{gather} \label{eq:qed_commutation} [H_{x},H_{y}] = 0\, , \qquad [H_{x},\overline{H}_{y}] = 0\, , \qquad [\overline{H}_{x},\overline{H}_{y}] = 0 \, ;\\ \label{eq:qed_Hnilpotent} H_{x} H_{x} = 0\, , \qquad H_{x} e_{x} = 0\, , \qquad H_{x} p_{x} = 0\, ; \\ \label{eq:qed_Hbarnilpotent} \overline{H}_{x} \overline{H}_{x} = 0\, , \qquad \overline{H}_{x} \overline{e}_{x} = 0\, , \qquad \overline{H}_{x} \overline{p}_{x} = 0\, . \end{gather} Eqs.~(\ref{eq:qed_Hnilpotent}) and (\ref{eq:qed_Hbarnilpotent}) imply that the two regions $\mathcal{H}$ and $\overline{\mathcal{H}}$ do not mix in the strong coupling limit. As a consequence, we can factorize the partition function in the following way \begin{equation} Z \, = \, \sum_{\{\mathcal{H}\}} \left[\prod_{i} Z_{\mathcal{H}_{i}}\right] \times Z_{\overline{\mathcal{H}}} \, . \label{eq:qed_partitionfactorized} \end{equation} The sum runs over all the possible ways of organizing the lattice into sets of hydrogen bags $ \mathcal{H} = \cup_{i} \mathcal{H}_{i}$. A single hydrogen bag causes a contribution of \begin{equation} Z_{\mathcal{H}_{i}} \, = \prod_{x \in \mathcal{H}_{i}} \int \! dH_{x} d\overline{H}_{x} \, \exp \Big(\sum_{x,y} \overline{H}_{x} D_{x,y}^{(i)} H_{y} \Big) \, , \label{eq:qed_partitionh} \end{equation} where we reorganized the canonical Grassmann integrations inside the region $\mathcal{H}_{i}$ as follows: \begin{equation} \prod_{x \in \mathcal{H}_{i}} \int dp_{x} d\overline{p}_{x} de_{x} d\overline{e}_{x} \, = \, \prod_{x \in \mathcal{H}_{i}} \int dp_{x} de_{x} d\overline{e}_{x} d\overline{p}_{x} \, = \, \prod_{x \in \mathcal{H}_{i}} \int dH_{x} d\overline{H}_{x} \, , \end{equation} with the hydrogen measures defined as \begin{equation} dH_{x} \, \equiv \, de_{x} dp_{x} \, , \qquad d\overline{H}_{x} \, \equiv \, d\overline{p}_{x} d\overline{e}_{x} \, . \label{eq:qed_hydrogenmeasure} \end{equation} In the exponent of the exponential in Eq.~(\ref{eq:qed_partitionh}) we introduced the operator \begin{equation} D_{x,y}^{(i)} = \theta_{x}^{(i)}2M \delta_{x,y} + \sum_{\nu} \theta_{x,\nu}^{(i)}\big[ \delta_{x+\hat{\nu},y} + \delta_{x,y+\hat{\nu}} \big] \, , \label{eq:qed_hydrogenoperator} \end{equation} with the site and link support functions on the hydrogen bag $\mathcal{H}_{i}$ given by \begin{equation} \theta_{x}^{(i)} = \left\{ \begin{array}{l} 1 \quad \text{if } x \in \mathcal{H}_{i} \\ 0 \quad \text{if } x \notin \mathcal{H}_{i} \\ \end{array}\right. \, , \qquad \theta_{x,\nu}^{(i)} = \left\{ \begin{array}{l} 1 \quad \text{if } (x,\nu) \in \mathcal{H}_{i} \\ 0 \quad \text{if } (x,\nu) \notin \mathcal{H}_{i} \\ \end{array}\right. \, . \label{eq:qed_supportfunctions} \end{equation} The partition function of the complementary region $\overline{\mathcal{H}}$ is given by \begin{align} Z_{\overline{\mathcal{H}}} \, = \, \sum_{\{s,d\, ||\, \overline{\mathcal{H}}\}} \, \prod_{x\in \overline{\mathcal{H}}} \, \int \! dp_{x} d\overline{p}_{x} de_{x} &d\overline{e}_{x} \prod_{x} \, \big[2m_{e} \, \overline{e}_{x} e_{x}\big]^{s^{(e)}_{x}}\, \big[2m_{p} \, \overline{p}_{x} p_{x}\big]^{s^{(p)}_{x}} \nonumber \\ &\, \ \times \prod_{x,\nu}\, \big[\overline{e}_{x} e_{x} \overline{e}_{x+\hat{\nu}} e_{x+\hat{\nu}}\big]^{d_{x,\nu}^{\,(e)}} \, \big[\overline{p}_{x} p_{x} \overline{p}_{x+\hat{\nu}} p_{x+\hat{\nu}}\big]^{d_{x,\nu}^{\,(p)}}. \label{eq:qed_partitionc} \end{align} $Z_{\overline{\mathcal{H}}}$ is a sum over all the configurations of electron's and proton's monomers and dimers that saturate the Grassmann integrals in the complementary region $\overline{\mathcal{H}}$: \begin{equation} \sum_{\{s,d\, ||\, \overline{\mathcal{H}}\}} \, = \, \left[\, \prod_{x \in \overline{\mathcal{H}}} \, \sum_{s^{(e)}_{x} = 0}^{1} \, \sum_{s^{(p)}_{x} = 0}^{1-s^{(e)}_{x}}\, \right] \left[\,\prod_{(x,\nu) \in \overline{\mathcal{H}}}\, \sum_{d^{\,(e)}_{x,\nu} = 0}^{1} \,\sum_{d^{\,(p)}_{x,\nu} = 0}^{1-d^{\,(e)}_{x,\nu}}\,\right] \, . \label{eq:qed_sumsdcomplementary} \end{equation} In Eq.~(\ref{eq:qed_partitionc}) all the fermion variables are already in the canonical order for the Grassmann integration. Monomers come with weight factors $2m_{e}$ for the electrons and $2m_{p}$ for the protons. Dimers account for a factor $+1$. The condition of saturation of the Grassmann integral can be implemented by means of the monomer-dimer constraint \begin{equation} C_{MD}[s,d] = \prod_{x \in \overline{\mathcal{H}}}\, \delta \Big( 1 - \big[s_{x}^{(e)} + \sum_{\nu}( d_{x,\nu}^{\, (e)} + d_{x - \hat{\nu},\nu}^{\, (e)} )\big]\Big) \,\delta \Big( 1 - \big[s_{x}^{(p)} + \sum_{\nu}( d_{x,\nu}^{\, (p)} + d_{x - \hat{\nu},\nu}^{\, (p)})\big]\Big) \, , \label{eq:qed_cmd} \end{equation} where we use the notation $\delta(n) \equiv \delta_{n,0}$ for the Kronecker deltas. Then, we can rewrite (\ref{eq:qed_partitionc}) in the following simple form \begin{equation} Z_{\overline{\mathcal{H}}} \, = \, \sum_{\{s,d\, ||\, \overline{\mathcal{H}}\}} \, \big(2m_{e} \big)^{\mathcal{N}(m_{e})} \big(2m_{p} \big)^{\mathcal{N}(m_{p})} \, C_{MD}[s,d]\, , \label{eq:qed_partitioncdual} \end{equation} where $\mathcal{N}(m_{e})$ and $\mathcal{N}(m_{p})$ denote the total number of electron and proton monomers respectively. The partition function $Z_{\overline{\mathcal{H}}}$ sums over all the possible configurations of the dual variables $\{s,d\}$. The allowed configurations must saturate the sites of the complementary domain $\overline{\mathcal{H}}$ by means of electron and proton monomers and dimers. This condition is enforced by the constraint $C_{MD}[s,d\,]$. Additionally, the configurations with $s_{x}^{(e)} = s_{x}^{(p)} = 1$ (hydrogen monomer), as well as with $d_{x,\nu}^{\,(e)} = d_{x,\nu}^{\,(p)} = 1$ (hydrogen dimer) are not allowed in the complementary region. This condition is already implemented in the definition of the sum (\ref{eq:qed_sumsdcomplementary}), where we restrict the value of the proton monomers and dimers to be $s_{x}^{(p)} \in \{0,1-s_{x}^{(e)}\}$ and $d_{x,\nu}^{\,(p)} \in \{0,1-d_{x,\nu}^{\,(e)}\}$. \section{CBB contributions as permanents} To obtain the composite boson bag contributions (\ref{eq:qed_partitionh}) to the partition function $Z$ we have to solve integrals of the form \begin{equation} I[D] \, = \int \! \prod_{l = 1}^{N} dH_{l} d\overline{H}_{l} \, \exp \Big(\sum_{i,j = 1}^{N} \overline{H}_{i} D_{i,j} H_{j} \Big) \, . \label{eq:qed_hydrogenintegral} \end{equation} If $H_{x}$ and $\overline{H}_{x}$ were Grassmann variables, the result of this integral could be trivially determined as $\det D$ (compare Sec.~\ref{sec:teo_grassmann}). However, we know that $H_{x}$ and $\overline{H}_{x}$ are commuting nilpotent variables, as they obey the properties (\ref{eq:qed_commutation}) -- (\ref{eq:qed_Hbarnilpotent}). Therefore we must explicitly compute the integral based on the algebra (\ref{eq:qed_commutation}) -- (\ref{eq:qed_Hbarnilpotent}). Before doing so, let us briefly discuss the rules for the integration over the hydrogen measures $dH_{x}$ and $d\overline{H}_{x}$ we defined in Eq.~(\ref{eq:qed_hydrogenmeasure}). They obey the same commutation relations as the hydrogen fields $H_{x}$ and $\overline{H}_{x}$ \begin{gather} \label{eq:qed_commutation2} [dH_{x},dH_{y}] = 0\, , \qquad [dH_{x},d\overline{H}_{y}] = 0\, , \qquad [d\overline{H}_{x},d\overline{H}_{y}] = 0 \, , \end{gather} and, together with the hydrogen fields, they obey the Grassmann integration rules \begin{gather} \label{eq:qed_integrationrules} \int dH_{x} \, = \, 0 \, , \qquad \int d\overline{H}_{x} \, = \, 0 \, , \qquad \int dH_{x} H_{x} \, = \, 1 \, , \qquad \int d\overline{H}_{x} \overline{H}_{x} \, = \, 1 \,. \end{gather} We can now proceed with the computation of the integral (\ref{eq:qed_hydrogenintegral}): \begin{equation} I[D] \, = \int \! \prod_{l = 1}^{N} dH_{l} d\overline{H}_{l} \, \prod_{i=1}^{N}\exp \Big(\overline{H}_{i} \sum_{j} D_{i,j} H_{j} \Big) \, = \, \int \! \prod_{l = 1}^{N} dH_{l} d\overline{H}_{l} \, \prod_{i=1}^{N} \Big(1 + \overline{H}_{i} \sum_{j} D_{i,j} H_{j} \Big) \,. \label{eq:qed_hydrogenintegral2} \end{equation} In the first step we used the fact that all the terms in the exponent of Eq.~(\ref{eq:qed_hydrogenintegral}) commute and brought the sum over the $i$ index down from the exponential as a product. Then we exploited the nilpotency of the hydrogen variables $\overline{H}_{i}$ and wrote the product of exponentials as a product of binomials. Now, from the integration rules (\ref{eq:qed_integrationrules}) follows that the only terms of the product of binomials that survive after the integration are the ones in which all the hydrogen variables appear exactly once: \begin{align} I[D] \, & = \, \int \! \prod_{l = 1}^{N} dH_{l} d\overline{H}_{l} \, \prod_{i=1}^{N} \overline{H}_{i} \sum_{j} D_{i,j} H_{j} \nonumber\\ & = \, \int \! \prod_{l = 1}^{N} dH_{l} d\overline{H}_{l} \ \overline{H}_{1} D_{1,j_{1}} H_{j_1} \,\overline{H}_{2} D_{2,j_{2}} H_{j_2} \dots \overline{H}_{N} D_{N,j_{N}} H_{j_N}\, , \label{eq:qed_hydrogenintegral3} \end{align} where a sum over repeated indices $j_{k}$, $k=1,\dots,N$ is understood. Then, using the commutation relations (\ref{eq:qed_commutation}) and (\ref{eq:qed_commutation2}) we obtain \begin{align} I[D] \, = \, \int \! \prod_{l = 1}^{N} dH_{l} d\overline{H}_{l} \ \overline{H}_{l} H_{l} \sum_{\sigma \in S_{N}} D_{1,\sigma(1)} D_{2,\sigma(2)} \dots D_{N,\sigma(N)} \, = \, \perm D \, , \label{eq:qed_hydrogenintegral4} \end{align} where the sum extends over all elements $\sigma$ of the symmetric group $S_{N}$, i.e., over all permutations of the numbers $1, 2, \dots, N$. In Eq.~(\ref{eq:qed_hydrogenintegral4}) we used the definition of the \textit{permanent} of a matrix $A$ \cite{minc_marcus_1984}: \begin{equation} \perm A \equiv \sum_{\sigma \in S_{N}} A_{1,\sigma(1)} A_{2,\sigma(2)} \dots A_{N,\sigma(N)} \, . \end{equation} Summarizing, we obtained that each hydrogen bag gives a contribution \begin{equation} Z_{\mathcal{H}_{i}} \, = \perm D^{(i)} \, . \label{eq:qed_partitionhdual} \end{equation} Putting together the results we obtained for the hydrogen bags and the complementary domain, the full partition function is given by \begin{equation} Z \, = \, \sum_{\{\mathcal{H}\}} \prod_{i} \perm D^{(i)} \times Z_{\overline{\mathcal{H}}} \, . \label{eq:qed_partitiondual} \end{equation} In the CBB formulation (hydrogen bags formulation) the partition function is a sum over all the possible configurations of the hydrogen bags. Each hydrogen bag $\mathcal{H}_{i}$ contributes with a weight factor $\perm D^{(i)}$, given by the permanent of the free boson operator (\ref{eq:qed_hydrogenoperator}). The remaining part of the lattice, i.e., the complementary domain $\overline{\mathcal{H}}$, has to be filled with monomers and dimers of electrons and protons. $Z_{\overline{\mathcal{H}}}$ sums over the allowed configurations of the monomers ($s_{x}^{(e)} \in \{0,1\}$ and $s_{x}^{(p)} \in \{0,1-s_{x}^{(e)}\}$) and dimers ($d_{x,\nu}^{\,(e)} \in \{0,1\}$ and $d_{x,\nu}^{\,(p)} \in \{0,1-d_{x,\nu}^{\,(e)}\}$) variables, and accounts a weight factor $2m_{e}$ ($2m_{p}$) for each electron (proton) monomer. We conclude this appendix with remarking that the permanent of a matrix with non-negative entries can be computed with a polynomial-time approximation algorithm \cite{Jerrum:2004:PAA:1008731.1008738} and thus makes the CBB formulation an interesting candidate for an efficient Monte Carlo simulation. \chapter{Introduction \label{cha:introduction}} Quantum Chromodynamics (QCD) is the theory that describes the strong interaction between quarks and gluons, the constituents of hadrons. It is a non-abelian gauge theory with symmetry group SU(3) and six flavors of quarks, two of which are almost massless. The resulting theory is extremely rich, showing properties such as asymptotic freedom, confinement of quarks, spontaneous chiral symmetry breaking and phase transitions. While asymptotic freedom allows the use of perturbation theory at high energies, the strong coupling at low energies requires the use of non-perturbative methods, such as lattice QCD which allows the theoretical study of QCD from first principle calculations. The phase diagram of QCD in the $T$ -- $\mu$ plane is thought to have a very rich structure, with a crossover at high temperature and zero chemical potential, which could eventually become a first order phase transition with increasing values of $\mu$, thus signaling the presence of a so-called critical point. More exotic phases, such as color superfluid/superconducting phases, are conjectured at higher values of chemical potential and small temperatures. As heavy ion collision experiments at LHC and RHIC probe the structure of the phase diagram of QCD at finite baryon chemical potential $\mu$, from the theory side the only region for which we have reliable results is the temperature axes, where numerical simulations predict a rapid crossover transition with the restoration of chiral symmetry at temperatures around 150 MeV. However, non-vanishing values of chemical potential cause the so-called sign problem which prevents the use of standard Monte-Carlo techniques on the lattice. Thus, the need of finding alternative approaches led to the development of a variety of methods over the years. Among those, the dual approach has shown to be a powerful tool in solving the sign problem of models with abelian gauge groups. Following that success, the aim of my PhD project was to extend the applicability of the dual approach to non-abelian lattice field theories. The effort we put in that direction resulted in the development of the \textit{abelian color cycle} (ACC) and \textit{abelian color flux} (ACF) dualization methods, which are at the center of the discussion of this thesis. While the ACC method is suitable for the dualization of gauge theories with non-abelian group, the ACF approach can be used for systems with matter. The key step of both these methods is the decomposition of the action into its minimal units. For gauge theories those minimal terms are complex numbers, which we refer to as abelian color cycles, while for fermions, the action is decomposed into Grassmann bilinears, which we called abelian color fluxes. As a result of these decompositions the actions are sums of commuting terms, and thus one can proceed with the dualization of the theory as in the abelian case, by factorizing and expanding the Boltzmann weight. The expansion indices, so-called dual variables, become the new degrees of freedom for the description of the system once the conventional fields are integrated out. In general, the integration over the gauge fields results into two contributions: a gauge weight factor and a gauge constraint. The weight factor organizes the partition function into a strong coupling series, where all terms are known in closed form. Moreover, for the specific cases of SU(2) and SU(3) pure gauge theory the weight factor also contain signs, which originate from the explicit signs in the parametrization of the group elements. The constraints, on the other hand, bring information about the original symmetry into the dual formulation, and in particular they allow to determine the configurations that contribute to the long range physics, which are worldsheets of the dual variables. Also the Grassmann integration results in a weight factor and a constraint. The fermion constraint implements Pauli's exclusion principle for the fermion dual variables, while the fermion weight factor collects the contributions of the admissible fermion configurations. We find that the long range physics for the fermion degrees of freedom in the dual representation is described by worldlines on the lattice which, however, come with signs. Obviously the appearance of gauge and fermion sign factors in our dual formulations implies that for a Monte Carlo simulation of the ACC and ACF dual representations a strategy for a partial resummation needs to be found. Nevertheless, our dual representations have some features which make them interesting per se, such as the possibility of identifying conserved charges with topological invariants, which thus become much easier to determine in the worldline formulation. In another application the ACF approach solves the complex action problem of the SU(2) principal chiral model in the conventional representation when chemical potentials are coupled to some of the conserved charges. For that system we were actually able to bring the dualization a step further by reformulating it in terms of variables which automatically solve the constraints. Both types of dual representations were successfully implemented in a Monte Carlo simulation. \newpage This thesis is organized as follows: \begin{description} \item[Chapter \ref{cha:theoretical}] collects some of the basic concepts of lattice field theories and explains the origin of the sign problem in lattice QCD. After reviewing some of the methods that have been developed in order to overcome the sign problem, we then focus on the dual approach and discuss the dualization of the U(1) Gauge -- Higgs model. \item[Chapter \ref{cha:pcm}] presents the SU(2) principal chiral model as first example for the application of the ACF approach. As was mentioned in the introduction, for this model the dual representation solves the complex action problem. At the end of the chapter we also discuss the possibility of completely resolving the constraints by bringing the ACF dualization a step further. \item[Chapter \ref{cha:su2}] gives the dual representation for SU(2) lattice gauge theory with staggered fermions. We start applying the ACC method to the pure gauge theory. We then show how the generalization of this approach to the theory with fermions leads us to the development of the ACF method. \item[Chapter \ref{cha:qcd}] shows how to obtain the worldline and worldsheet representation of lattice QCD using the ACC and ACF dualization methods. Also in this case we start with discussing the pure gauge theory, and then present the strong coupling limit before giving the result for full QCD. \item[Chapter \ref{cha:conclusions}] ends the thesis by summarizing the findings discussed in the previous chapters and outlining the main features of the dual approaches we developed. \end{description} \newpage From the projects described in this thesis so far the following papers and conference proceedings were published: \begin{itemize} \item C. Marchis and C. Gattringer, \textit{Dual representation of lattice QCD with worldlines and worldsheets of abelian color fluxes}, Phys. Rev. D \textbf{97} (2018) 034508, [arXiv:1712.07546 [hep-lat]] \item C. Gattringer, D. Göschl and C. Marchis, \textit{Worldlines and worldsheets for non-abelian lattice field theories: Abelian color fluxes and Abelian color cycles}, EPJ Web Conf. \textbf{175} (2018) 11007, [arXiv:1710.08745 [hep-lat]] \item C. Gattringer, D. Göschl and C. Marchis, \textit{Kramers-Wannier duality and worldline representation for the SU(2) principal chiral model}, Phys. Lett. B \textbf{778} (2018) 435, [arXiv:1709.04691 [hep-lat]] \item C. Gattringer and C. Marchis, \textit{Abelian color cycles: a new approach to strong coupling expansion and dual representations for non-abelian lattice gauge theory}, Nucl. Phys. B \textbf{916} (2017) 627, [arXiv:1609.00124 [hep-lat]] \item C. Gattringer and C. Marchis, \textit{Dualization of non-abelian lattice gauge theory with Abelian Color Cycles (ACC)}, PoS \textbf{LATTICE2016} (2016) 034, [arXiv:1611.01022 [hep-lat]] \end{itemize} All publications can be acquired freely as on-line e-prints at arXiv.org. \chapter{Dual representation for lattice QCD with one flavor of staggered fermions \label{cha:qcd}} As a natural extension of the methods we developed in Chapter \ref{cha:su2} for QC$_2$D, we here discuss the generalization of the abelian color cycle (ACC) and the abelian color flux (ACF) dual approaches to lattice QCD with one flavor of staggered fermions. The study of the phase diagram of QCD on the lattice has been tremendously slowed down by the so-called sign problem. This problem arises when a non-vanishing chemical potential is introduced in the action. As a result, the fermion determinant becomes complex and thus the applicability of standard Monte Carlo methods, which requires the interpretation of the Boltzmann weight as a probability, is prevented. As we learned in Sec.~\ref{sec:teo_u1} as well as in Chapter \ref{cha:pcm}, the sign problem is, in general, representation dependent. Unfortunately, up to this date no representation of lattice QCD has been found that solves the sign problem (and the one we present in this chapter makes no exception). Nevertheless, already more than 30 years ago, a monomer-dimer-loop representation of QCD at strong coupling was found that presented a sign problem much milder than the one of QCD itself \cite{Rossi:1984cv,Karsch:1988zx}, so that the standard technique of reweighting can be used for simulations \cite{deForcrand:2009dh}. In the strong coupling limit the inverse lattice coupling $\beta$ is set to 0. Thus, in this regime it is impossible to take the continuum limit and one expects large lattice artifacts. Nonetheless, strong coupling QCD shares some key features of QCD as, for examples, it confines colored objects and, for vanishing quark masses, it presents a chiral spontaneously broken phase in the vacuum and a restored chiral symmetric phase at high temperature or high chemical potential. These properties explain the abundance of literature about this subject. Older publications \cite{Adams:2003cca, Chandrasekharan:2003im, Chandrasekharan:2004kd, Chandrasekharan:2006tz, deForcrand:2009dh, Unger:2011in, deForcrand:2012vh, Kim:2016izx, deForcrand:2017fky} use the monomer-dimer-loop representation derived in \cite{Rossi:1984cv,Karsch:1988zx} to study the phase diagram of strong coupling QCD, using adaptation of the worm algorithm \cite{PhysRevLett.87.160601} or continuous time Monte Carlo as simulation techniques. In \cite{deForcrand:2014tha,deForcrand:2015daa} the first order correction in $\beta$ is included as a first step towards the continuum limit, while in \cite{Gagliardi:2017uag} a method to include corrections at higher orders of $\beta$ is proposed. Other diagrammatic representations for QCD and QCD-like lattice field theories can be found in \cite{Bruckmann:2017vri,Borisenko:2017gql,Vairinhos:2014uxa,Gattringer:2018mrg}. Differently from what was done in the past, the reformulation of lattice QCD we present in this chapter has the form of a strong coupling expansion in which all the terms are known in closed form. The long range physics, relevant for the continuum limit, has the structure of worldsheets for the gauge degrees of freedom, and worldlines for the matter fields. We also study in details the strong coupling limit. We find that in this regime baryons propagate as free fermion loops in a background of monomers and dimers. As a future challenge, we propose the use of fermion bags \cite{PhysRevD.82.025007,PhysRevD.85.091502,Chandrasekharan2013,Gattringer:2018mrg} as an updating strategy for our form of strong coupling QCD. \section{Abelian color cycle dualization of the SU(3) pure gauge theory \label{sec:qcd_acc}} We start the dualization of lattice QCD focusing on the pure gauge theory. The Wilson action for SU(3) reads \begin{equation} S_G[U] \; = \; -\dfrac{\beta}{3} \sum_{x,\nu < \rho} \R \Tr U_{x,\nu} \; U_{x+\hat{\nu},\rho} \, U_{x+\hat{\rho},\nu}^{\dagger} \, U_{x,\rho}^\dagger \; , \label{eq:qcd_sg} \end{equation} where $U_{x,\nu} \in$ SU(3) are the gauge degrees of freedom. They sit on the links $(x,\nu)$ of a four-dimensional lattice with periodic boundary conditions. The partition function $Z$ is obtained by integrating the Boltzmann weight $e^{-S_G[U]}$ over the product of Haar measures $\int \! D[U] = \prod_{x,\nu} \int_{\text{SU(3)}} dU_{x,\nu}$: \begin{equation} Z \, = \, \int \! D[U] \; e^{-S_G[U]} \, . \label{eq:qcd_partition} \end{equation} As we have seen in the previous chapter, the key step of the abelian color cycle (ACC) method consists in decomposing the action in its minimal terms. This is achieved by making explicit the real part, the trace and the matrix multiplications in Eq.~(\ref{eq:qcd_sg}). As a result, the action is rewritten as a sum of products of four link matrix elements $U_{x,\nu}^{ab}$ \begin{equation} S_G[U] \; = \; -\dfrac{\beta}{6} \sum_{x,\nu < \rho} \sum_{a,b,c,d=1}^{3} \Big[ U_{x,\nu}^{ab} U_{x+\hat{\nu},\rho}^{bc} U_{x+\hat{\rho},\nu}^{dc \, \star} U_{x,\rho}^{ad \, \star} + U_{x,\nu}^{ab \, \star} U_{x+\hat{\nu},\rho}^{bc \, \star} U_{x+\hat{\rho},\nu}^{dc} U_{x,\rho}^{ad} \Big]\, . \label{eq:qcd_accaction} \end{equation} The products $U_{x,\nu}^{ab} U_{x+\hat{\nu},\rho}^{bc} U_{x+\hat{\rho},\nu}^{dc \ \star} U_{x,\rho}^{ad \ \star}$ are what we call \textit{Abelian Color Cycles} (ACCs) \cite{Gattringer:2016lml}. These objects are complex numbers, and we interpret them as paths in color space closing around plaquettes, analogously to what we have done in Section \ref{sec:su2_acc} for the case of SU(2). The four color labels $a,b,c,d = 1,2,3$ denote the colors on each of the four corners of the plaquette $(x,\nu\rho)$. On the first site $x$, which is at the bottom left of the plaquette, the ACC runs trough color $a$, on site $x + \hat{\nu}$ trough color $b$, and so on, closing around the plaquette with mathematically positive orientation. Notice that, differently from what we found for the SU(2) gauge theory, also the complex conjugate ACCs $U_{x,\nu}^{ab \ \star} U_{x+\hat{\nu},\rho}^{bc \ \star} U_{x+\hat{\rho},\nu}^{dc} U_{x,\rho}^{ad}$ appear in the decomposition (\ref{eq:qcd_accaction}). This is a result of having to take the real part in the Wilson action (\ref{eq:qcd_sg}), an operation that was unnecessary in Section \ref{sec:su2_acc} because of the pseudo-reality of SU(2). Geometrically the complex conjugate ACCs will be interpreted as running through the plaquette with mathematically negative orientation. To give an example of the geometrical interpretation, in Fig.~\ref{fig:qcd_acc} we illustrate the $(1,2,3,3)$-ACC. It closes around the plaquette $(x,\nu\rho)$, which has three color layers represented in the figure with grey dashed lines. The four link matrix elements $U_{x,\nu}^{ab}$ constituting the ACC are represented with oriented arrows which connect color layers on neighboring sites, e.g., the matrix element $U_{x,\nu}^{12}$ links color $1$ on site $x$ to color $2$ on site $x + \hat{\nu}$. As we already mentioned, complex conjugation reverts the orientation of the arrows, so that, e.g., the link element $U_{x,\nu}^{13 \ \star}$ connects color $3$ on site $x + \hat{\rho}$ to color $1$ on site $x$, thus closing the cycle around the plaquette. Obviously, since for each of the four corners of a plaquette we have three different possible choices of color, then for SU(3) there are a total of $3^4=81$ different ACCs on every plaquette, each of which can have either positive or negative mathematical orientation. \begin{figure}[t] \begin{center} \vskip5mm \includegraphics[scale=0.6,clip]{chapters/plots/qcd_acc.pdf} \end{center} \caption{Graphical illustration of the (1,2,3,3)-ACC, which explicitly is given by $U_{x,\nu}^{12} U_{x+\hat{\nu},\rho}^{23} U_{x+\hat{\rho},\nu}^{33 \ \star} U_{x,\rho}^{13 \ \star}$. This ACC closes around the plaquette $(x,\nu\rho)$ running through the sequence $(1,2,3,3)$ of color indices at the corners of the plaquette. In the graphical representation the color degrees of freedom are shown as three distinct layers of the space-time lattice, labeled with 1,2 and 3 on the lhs.\ of the plot. Each of the matrix elements $U_{x,\nu}^{ab}$ constituting the ACC is represented by an arrow along the corresponding link $(x,\nu)$ connecting color $a$ with color $b$. For complex conjugate matrix elements the link is run through with negative orientation. \label{fig:qcd_acc}} \end{figure} The ACC decomposition (\ref{eq:qcd_accaction}) of the action (\ref{eq:qcd_sg}) allows us to proceed with the dualization of the theory as follows: \begin{align} Z \, &= \, \int \! D[U] \prod_{x,\nu<\rho} \prod_{a,b,c,d = 1}^{3} e^{\frac{\beta}{6} U_{x,\nu}^{ab} U_{x+\hat{\nu},\rho}^{bc} U_{x+\hat{\rho},\nu}^{dc \, \star} U_{x,\rho}^{ad \, \star}} \,e^{\frac{\beta}{6} U_{x,\nu}^{ab\, \star} U_{x+\hat{\nu},\rho}^{bc \, \star} U_{x+\hat{\rho},\nu}^{dc} U_{x,\rho}^{ad}} \nonumber \\ & = \, \int \! D[U] \prod_{x,\nu<\rho} \prod_{a,b,c,d = 1}^{3} \sum_{n_{x,\nu\rho}^{abcd} = 0}^{\infty} \sum_{\overline{n}_{x,\nu\rho}^{abcd} = 0}^{\infty} \dfrac{ \left( \beta/6 \right)^{n_{x,\nu\rho}^{abcd} + \overline{n}_{x,\nu\rho}^{abcd}}}{n_{x,\nu\rho}^{abcd}\, ! \overline{n}_{x,\nu\rho}^{abcd}\, !} \nonumber \\ & \hspace{2.5cm} \times \left( U_{x,\nu}^{ab} U_{x+\hat{\nu},\rho}^{bc} U_{x+\hat{\rho},\nu}^{dc \, \star} U_{x,\rho}^{ad \, \star} \right)^{n_{x,\nu\rho}^{abcd}} \left( U_{x,\nu}^{ab\, \star} U_{x+\hat{\nu},\rho}^{bc \, \star} U_{x+\hat{\rho},\nu}^{dc} U_{x,\rho}^{ad} \right)^{\overline{n}_{x,\nu\rho}^{abcd}} \, . \label{eq:qcd_partition2} \end{align} In the first step we wrote the Boltzmann weight as a product over all the plaquettes, and all the ACCs on that plaquette of the two local factors $e^{\frac{\beta}{6} U_{x,\nu}^{ab} U_{x+\hat{\nu},\rho}^{bc} U_{x+\hat{\rho},\nu}^{dc \ \star} U_{x,\rho}^{ad \ \star}}$ and $e^{\frac{\beta}{6} U_{x,\nu}^{ab \ \star} U_{x+\hat{\nu},\rho}^{bc \ \star} U_{x+\hat{\rho},\nu}^{dc} U_{x,\rho}^{ad}}$. These exponentials are the Boltzmann weights for the ACCs with positive and negative orientation respectively. In the second step of (\ref{eq:qcd_partition2}) we expanded each factor in a Taylor series, thus introducing two sets of expansion indices assigned to the plaquettes: $n_{x,\nu\rho}^{abcd} \in \mathbb{N}_{0}$ and $\overline{n}_{x,\nu\rho}^{abcd} \in \mathbb{N}_{0}$, where $a,b,c,d=1,2,3$ are the color indices. The variables $n_{x,\nu\rho}^{abcd}$ correspond to the units of flux around the plaquette $(x,\nu\rho)$ in positive orientation, with color indices $(a,b,c,d)$, while $\overline{n}_{x,\nu\rho}^{abcd}$ is used for flux with negative orientation. All the factors in the last line of (\ref{eq:qcd_partition2}) are complex numbers, such that we can freely commute them and reorganize the products to determine the integer valued powers for the link elements $U_{x,\nu}^{ab}$ and theirs complex conjugate $U_{x,\nu}^{ab\, \star}$. The partition sum becomes \begin{equation} \label{eq:qcd_partition3} Z \; = \; \sum_{\{n, \overline{n}\}} \left[ \prod_{x,\nu<\rho} \prod_{a,b,c,d} \dfrac{ \left( \beta/6 \right)^{n_{x,\nu\rho}^{abcd} + \overline{n}_{x,\nu\rho}^{abcd}}}{n_{x,\nu\rho}^{abcd}\, !\, \overline{n}_{x,\nu\rho}^{abcd}\, !} \right] \prod_{x,\nu} \int \! \! dU_{x,\nu} \; \prod_{a,b} \left( U_{x,\nu}^{ab} \right) ^{N_{x,\nu}^{ab}} \left( U_{x,\nu}^{ab \ \star} \right) ^{\overline{N}_{x,\nu}^{ab}} \; , \end{equation} where we introduced the short-hand notation \begin{equation*} \sum_{\{n, \overline{n}\}} = \prod_{x,\nu<\rho} \prod_{a,b,c,d = 1}^{3} \sum_{n_{x,\nu\rho}^{abcd} = 0}^{\infty} \sum_{\overline{n}_{x,\nu\rho}^{abcd} = 0}^{\infty} \, , \end{equation*} to express the sum over the configurations of the variables $n_{x,\nu\rho}^{abcd} \in \mathbb{N}_{0}$ and $\overline{n}_{x,\nu\rho}^{abcd} \in \mathbb{N}_{0}$. The powers $N_{x,\nu}^{ab} \in \mathbb{N}_{0}$ and $\overline{N}_{x,\nu}^{ab} \in \mathbb{N}_{0}$ for the matrix elements $U_{x,\nu}^{ab}$ and $U_{x,\nu}^{ab \, \star}$ are given by \begin{gather} \label{eq:qcd_N} N_{x,\nu}^{ab} \; = \; \sum_{\rho:\nu<\rho}n_{x,\nu\rho}^{abss} + \overline{n}_{x-\hat{\rho},\nu\rho}^{ssba} + \sum_{\sigma:\nu>\sigma}\overline{n}_{x,\sigma\nu}^{assb} + n_{x-\hat{\sigma},\sigma\nu}^{sabs}, \\ \label{eq:qcd_Nbar} \overline{N}_{x,\nu}^{ab} \; = \; \sum_{\rho:\nu<\rho}\overline{n}_{x,\nu\rho}^{abss} + n_{x-\hat{\rho},\nu\rho}^{ssba} + \sum_{\sigma:\nu>\sigma}n_{x,\sigma\nu}^{assb} + \overline{n}_{x-\hat{\sigma},\sigma\nu}^{sabs} \; . \end{gather} The label $s$ introduced in the last two equations stands for the independent sum over the color indices that are replaced by $s$, e.g., $n_{x,\nu\rho}^{assb} \equiv \sum_{c,d} n_{x,\nu\rho}^{acdb}$. To compute the Haar measure integrals in Eq.~(\ref{eq:qcd_partition3}) we use the following explicit parametrization for the SU(3) matrices \cite{Bronzan:1988wa} \begin{align} \label{eq:qcd_parametrization} &U_{x,\nu} =\\[1em] \nonumber &\left( \! \! \begin{array}{ccc} c_1 c_2 \, e^{i\phi_1} & s_1 \, e^{i\phi_3} & c_1 s_2 \, e^{i\phi_4}\\ s_2 s_3 \, e^{-i\phi_4 -i\phi_5} - s_1 c_2 c_3 \, e^{i\phi_1 +i\phi_2 -i\phi_3} & c_1 c_3 \, e^{i\phi_2} & - c_2 s_3 \, e^{-i\phi_1 -i\phi_5} - s_1 s_2 c_3 \, e^{i\phi_2 -i\phi_3 +i\phi_4}\\ - s_2 c_3 \, e^{-i\phi_2 -i\phi_4} - s_1 c_2 s_3 \, e^{i\phi_1 -i\phi_3 +i\phi_5} & c_1 s_3 \, e^{i\phi_5} & c_2 c_3 \, e^{-i\phi_1 -i\phi_2} - s_1 s_2 s_3 \, e^{-i\phi_3 +i\phi_4 +i\phi_5} \end{array} \!\!\right) , \end{align} where $c_{i} = \cos \theta_{x,\nu}^{(i)}$, $s_{i} = \sin \theta_{x,\nu}^{(i)}$, with $\theta_{x,\nu}^{(i)} \in [0, \pi/2]$, $i = 1,2,3$ and $\phi_{j} = \phi_{x,\nu}^{(j)}$, with $\phi_{x,\nu}^{(j)} \in [-\pi,\pi]$, $j=1,2,\dots,5$. The normalized Haar measure is \begin{equation} dU_{x,\nu} = \frac{1}{2\pi^5} \, d\theta_1 c_1^3 s_1 \, d\theta_2 c_2 s_2 \, d\theta_3 c_3 s_3 \, d\phi_1 \, d\phi_2 \, d\phi_3 \, d\phi_4 \, d\phi_5 \, . \label{eq:qcd_haarmeasure} \end{equation} In what follows it will prove convenient to perform the changes of variables: \begin{gather} \label{eq:qcd_cycleoccupationnumbers} n_{x,\nu\rho}^{abcd} - \overline{n}_{x,\nu\rho}^{abcd} = p_{x,\nu\rho}^{abcd} \, , \qquad p_{x,\nu\rho}^{abcd} \in \mathbb{Z}\, ; \\ \label{eq:qcd_auxiliary} n_{x,\nu\rho}^{abcd} + \overline{n}_{x,\nu\rho}^{abcd} = |p_{x,\nu\rho}^{abcd}| + 2 l_{x,\nu\rho}^{abcd} \, , \qquad l_{x,\nu\rho}^{abcd} \in \mathbb{N}_0 \, . \end{gather} Eqs.~(\ref{eq:qcd_cycleoccupationnumbers}) and (\ref{eq:qcd_auxiliary}) define the dual variables $p_{x,\nu\rho}^{abcd} \in \mathbb{Z}$ and $l_{x,\nu\rho}^{abcd} \in \mathbb{N}_0$, which will be the new dynamical degrees of freedom replacing the gauge fields once the conventional fields $U_{x,\nu}$ are integrated out. From the definition (\ref{eq:qcd_cycleoccupationnumbers}) and the interpretation of the $n_{x,\nu\rho}^{abcd}$ ($\overline{n}_{x,\nu\rho}^{abcd}$) as the activation numbers for the $(a,b,c,d)$-ACCs with positive (negative) orientation, it is clear that the new variables $p_{x,\nu\rho}^{abcd}$ activate $|p_{x,\nu\rho}^{abcd}|$ units of flux for the $(a,b,c,d)$-ACC on the plaquette $(x,\nu\rho)$, with the orientation of the flux given by the sign of the $p_{x,\nu\rho}^{abcd}$. We refer to the $p_{x,\nu\rho}^{abcd}$ as \textit{cycle occupation numbers}. As we will see later, the $p_{x,\nu\rho}^{abcd}$ will be subject to constraints, while the variables $l_{x,\nu\rho}^{abcd} \in \mathbb{N}_{0}$ will be unconstrained, thus we simply refer to them as \textit{auxiliary plaquette variables}. We also introduce the fluxes \begin{equation} \label{eq:qcd_Jfluxes} J_{x,\nu}^{ab} \; = \sum_{\rho:\nu<\rho} [\, p_{x,\nu\rho}^{abss} - p_{x-\hat{\rho},\nu\rho}^{ssba} \, ] - \!\! \sum_{\sigma:\nu>\sigma}[\, p_{x,\sigma\nu}^{assb} - p_{x-\hat{\sigma},\sigma\nu}^{sabs} \, ] \; , \end{equation} \begin{align} \nonumber S_{x,\nu}^{ab} &= \sum_{\rho:\nu<\rho} [ |p_{x,\nu\rho}^{abss}| + |p_{x-\hat{\rho},\nu\rho}^{ssba}| + 2(l_{x,\nu\rho}^{abss} + l_{x-\hat{\rho},\nu\rho}^{ssba}) ] \\ &\, + \sum_{\sigma:\nu>\sigma} [ |p_{x,\sigma\nu}^{assb}| + |p_{x-\hat{\sigma},\sigma\nu}^{sabs}| + 2(l_{x,\sigma\nu}^{assb} + l_{x-\hat{\sigma},\sigma\nu}^{sabs}) ] \; . \label{eq:qcd_Sfluxes} \end{align} Since the $J$-fluxes will enter the constraints, it is important to discuss their geometrical interpretation on the lattice. The $J_{x,\nu}^{ab}$ represents the total net amount of color flux on the link $(x,\nu)$ connecting color $a$ and $b$ of the neighboring sites $x$ and $x + \hat{\nu}$. This color flux is the sum of the contributions of all the ACCs attached to the link $(x,\nu)$ that have a path from color $a$ to color $b$ on that link. So, if we consider the plaquette $(x,\nu\rho)$, with $\nu < \rho$, we have 9 different ACCs that contribute to that flux, namely the ones corresponding to the cycle occupation numbers $p_{x,\nu\rho}^{abef}$, where $a$ and $b$ are the color indices which we fix at $x$ and $x + \hat{\nu}$. The colors $e$ and $f$, chosen independently from the set $\{1,2,3\}$, determine the ACC at the remaining two corners of the plaquette $(x,\nu\rho)$. We thus have $3^2 = 9$ possibilities. Since the flux of these ACCs has a positive orientation along the link $(x,\nu)$, the 9 ACCs contribute with a positive sign in the definition (\ref{eq:qcd_Jfluxes}) of the $J_{x,\nu}^{ab}$ fluxes. If we then consider another plaquette attached to the link $(x,\nu)$, e.g., the plaquette $(x,\sigma\nu)$ with $\sigma < \nu$, we find that also in this case there are 9 ACCs which contribute to the flux $J_{x,\nu}^{ab}$, corresponding to the 9 cycle occupation numbers $p_{x,\sigma\nu}^{aefb}$. Since these ACCs have negative flux along the link $(x,\nu)$, the $p_{x,\sigma\nu}^{aefb}$ contribute with a negative sign in the definition (\ref{eq:qcd_Jfluxes}) of the $J_{x,\nu}^{ab}$. For the remaining four plaquettes attached to the link $(x,\nu)$ an analogous discussion holds. As an example, in Fig.~\ref{fig:qcd_colorflux} we illustrate the sum of the contributions to the flux $J_{x,\nu}^{12}$ from the plaquette $(x,\sigma\nu)$, with $\sigma < \nu$. The flux from color 1 to 2 on the link $(x,\nu)$ is fixed and represented with an arrow oriented towards the positive $\nu$ direction. The 9 ACCs on the plaquette $(x,\sigma\nu)$ which contribute to this flux are summed over in the definition (\ref{eq:qcd_Jfluxes}) and we represent them with dashed lines in the figure. Since the ACCs on the $(x,\sigma\nu)$ plaquette have a negative orientation of the flux on the link we are considering, they contribute with a negative sign to the flux $J_{x,\nu}^{12}$. \begin{figure}[t] \begin{center} \vskip5mm \includegraphics[width=6.5cm]{chapters/plots/qcd_colorflux.pdf}% \end{center} \caption{Graphical illustration of the sum $- p_{x,\sigma\nu}^{1ss2}$ contributing to the $J_{x,\nu}^{12}$ flux. \label{fig:qcd_colorflux}} \end{figure} Having discussed the interpretation we give to the $J$-fluxes, we can now continue with the derivation of the dual representation for the partition function $Z$. Using the change of variables (\ref{eq:qcd_cycleoccupationnumbers}) and (\ref{eq:qcd_auxiliary}) we obtain the following expression: \begin{equation} \label{eq:qcd_partition4} Z \; = \; \sum_{\{p, l\}} \left[ \prod_{x,\nu<\rho} \prod_{a,b,c,d} \dfrac{ \left( \beta/6 \right)^{|p_{x,\nu\rho}^{abcd}| + 2l_{x,\nu\rho}^{abcd}}}{(|p_{x,\nu\rho}^{abcd}| + l_{x,\nu\rho}^{abcd})\, !\, l_{x,\nu\rho}^{abcd}\, !} \right] \prod_{x,\nu} \int \! \! dU_{x,\nu} \; \prod_{a,b} \left( U_{x,\nu}^{ab} \right) ^{N_{x,\nu}^{ab}} \left( U_{x,\nu}^{ab \ \star} \right) ^{\overline{N}_{x,\nu}^{ab}} \; . \end{equation} Hence, the partition function is now a sum over the configurations of the cycle occupation numbers $p_{x,\nu\rho}^{abcd} \in \mathbb{Z}$ and the auxiliary plaquette variables $l_{x,\nu\rho}^{abcd} \in \mathbb{N}_{0}$. In terms of these dual variables, $N_{x,\nu}^{ab}$ and $\overline{N}_{x,\nu}^{ab}$ can be written as \begin{equation} \label{eq:qcd_N2} N_{x,\nu}^{ab} = \frac{S_{x,\nu}^{ab} + J_{x,\nu}^{ab}}{2} \, , \qquad \qquad \overline{N}_{x,\nu}^{ab} = \frac{S_{x,\nu}^{ab} - J_{x,\nu}^{ab}}{2} \, . \end{equation} To compute the Haar measure integrals in Eq.~(\ref{eq:qcd_partition4}) we have to substitute the parametrization (\ref{eq:qcd_parametrization}) for the matrix elements $U_{x,\nu}^{ab}$ in (\ref{eq:qcd_partition3}). Notice, however, that some of the elements $U_{x,\nu}^{ab}$ of the matrix parametrization (\ref{eq:qcd_parametrization}) are not in the simple form $U_{x,\nu}^{ab} = r_{x,\nu}^{ab} \, e^{i \varphi_{x,\nu}^{ab}}$ (like in the case of SU(2)), but are sums $U_{x,\nu}^{ab} = \rho_{x,\nu}^{ab} \, e^{i \alpha_{x,\nu}^{ab}} + \omega_{x,\nu}^{ab} \, e^{i \beta_{x,\nu}^{ab}}$. Therefore an additional step is still required in order to perform analytically the Haar integration in (\ref{eq:qcd_partition4}). For those elements that are sums we make use of the binomial theorem $(x + y)^{N} = \sum_{m = 0}^{N} \binom{N}{m} x^{N-m} y^{m}$ and rewrite the integrand in (\ref{eq:qcd_partition4}) as \begin{align} \nonumber &\left( U_{x,\nu}^{ab} \right) ^{N_{x,\nu}^{ab}} \left( U_{x,\nu}^{ab \ \star} \right) ^{\overline{N}_{x,\nu}^{ab}} \, = \, \left( \rho_{x,\nu}^{ab}\, e^{i \alpha_{x,\nu}^{ab}} + \omega_{x,\nu}^{ab} \, e^{i \beta_{x,\nu}^{ab}} \right)^{N_{x,\nu}^{ab}} \left( \rho_{x,\nu}^{ab}\, e^{-i \alpha_{x,\nu}^{ab}} + \omega_{x,\nu}^{ab} \, e^{-i \beta_{x,\nu}^{ab}} \right)^{\overline{N}_{x,\nu}^{ab}} \\ & \hspace{1.5cm}= \, \sum_{m_{x,\nu}^{ab} = 0}^{N_{x,\nu}^{ab}} \sum_{\overline{m}_{x,\nu}^{ab} = 0}^{\overline{N}_{x,\nu}^{ab}} \binom{N^{ab}_{x,\nu}}{m^{ab}_{x,\nu}} \binom{\overline{N}^{ab}_{x,\nu}}{\overline{m}^{ab}_{x,\nu}} \left( \rho_{x,\nu}^{ab} \right)^{s_{x,\nu}^{ab}} \left( \omega_{x,\nu}^{ab} \right)^{S_{x,\nu}^{ab} - s_{x,\nu}^{ab}} e^{i \alpha_{x,\nu}^{ab} j_{x,\nu}^{ab}} \, e^{i \beta_{x,\nu}^{ab} \left(J_{x,\nu}^{ab} - j_{x,\nu}^{ab}\right)} \nonumber \\ &\text{with} \quad m_{x,\nu}^{ab} = 0,1\dots N_{x,\nu}^{ab} \, , \qquad \overline{m}_{x,\nu}^{ab} = 0,1 \dots \overline{N}_{x,\nu}^{ab} \, , \nonumber \\ &\text{and } \quad j_{x,\nu}^{ab} \equiv m_{x,\nu}^{ab} - \overline{m}_{x,\nu}^{ab} \, , \qquad \, \overline{s}_{x,\nu}^{ab} \equiv m_{x,\nu}^{ab} + \overline{m}_{x,\nu}^{ab} \, . \label{eq:qcd_binomial} \end{align} The new auxiliary variables $m_{x,\nu}^{ab}$ and $\overline{m}_{x,\nu}^{ab}$ that we use for the binomial decomposition (\ref{eq:qcd_binomial}) of the matrix elements with $(a,b) = (2,1),(2,3),(3,1)$ and $(3,3)$ live on the links of the lattice. To obtain the final result for the partition function we substitute the parametrization (\ref{eq:qcd_parametrization}) and the Haar measure (\ref{eq:qcd_haarmeasure}) in (\ref{eq:qcd_partition4}), and we use the binomial decomposition (\ref{eq:qcd_binomial}). For the partition function we obtain \begin{align} \nonumber &Z \; = \; 2^{4V} \sum_{\{p,l\}} \sum_{\{m,\overline{m}\}} \! \Bigg[ \prod_{x,\nu<\rho} \prod_{a,b,c,d} \dfrac{ \left( \beta/2 \right)^{|p_{x,\nu\rho}^{abcd}| + 2l_{x,\nu\rho}^{abcd}}}{\left(|p_{x,\nu\rho}^{abcd}| + l_{x,\nu\rho}^{abcd}\right)!l_{x,\nu\rho}^{abcd}!} \Bigg] \Bigg[ \prod_{x,\nu} (-1)^{S_{x,\nu}^{23} + S_{x,\nu}^{31} + s_{x,\nu}^{21} + s_{x,\nu}^{33}} \Bigg] \\ & \hspace{28mm} \times \; \Bigg[ \prod_{x,\nu} \prod_{a = 2,3} \prod_{b=1,3} \binom{N_{x,\nu}^{ab}}{m_{x,\nu}^{ab}} \binom{\overline{N}_{x,\nu}^{ab}}{\overline{m}_{x,\nu}^{ab}} \Bigg] \nonumber \\ & \hspace{23mm} \times \prod_{x,\nu} 2 \! \int_{0}^{\pi/2} \!\!\!\! d\theta_{x,\nu}^{(1)} \, (\cos\theta_{x,\nu}^{(1)})^{3 + S_{x,\nu}^{11} + S_{x,\nu}^{13} + S_{x,\nu}^{22} + S_{x,\nu}^{32}} \nonumber \\ & \hspace{53mm} (\sin\theta_{x,\nu}^{(1)})^{1 + S_{x,\nu}^{12} + s_{x,\nu}^{21} +s_{x,\nu}^{23} + s_{x,\nu}^{31} + s_{x,\nu}^{33}} \nonumber \\ & \hspace{30mm} \times \! 2 \! \int_{0}^{\pi/2} \!\!\!\! d\theta_{x,\nu}^{(2)} \, (\cos\theta_{x,\nu}^{(2)})^{1 + S_{x,\nu}^{11} + s_{x,\nu}^{21} + S_{x,\nu}^{23} - s_{x,\nu}^{23} + s_{x,\nu}^{31} + S_{x,\nu}^{33} - s_{x,\nu}^{33}} \nonumber \\ & \hspace{53mm} (\sin\theta_{x,\nu}^{(2)})^{1 + S_{x,\nu}^{13} + S_{x,\nu}^{21} - s_{x,\nu}^{21} + s_{x,\nu}^{23} + S_{x,\nu}^{31} - s_{x,\nu}^{31} + s_{x,\nu}^{33}} \nonumber \\ & \hspace{30mm} \times \! 2 \! \int_{0}^{\pi/2} \!\!\!\! d\theta_{x,\nu}^{(3)} \, (\cos\theta_{x,\nu}^{(3)})^{1 + s_{x,\nu}^{21} + S_{x,\nu}^{22} + s_{x,\nu}^{23} + S_{x,\nu}^{31} - s_{x,\nu}^{31} + S_{x,\nu}^{33} - s_{x,\nu}^{33}} \nonumber \\ & \hspace{53mm} (\sin\theta_{x,\nu}^{(3)})^{1 + S_{x,\nu}^{21} - s_{x,\nu}^{21} + S_{x,\nu}^{23} - s_{x,\nu}^{23} + s_{x,\nu}^{31} + S_{x,\nu}^{32} + s_{x,\nu}^{33}} \nonumber \\ & \hspace{31mm} \times \! \int_{0}^{2\pi} \! \dfrac{d\phi_{x,\nu}^{(1)}}{2\pi} \; e^{i\phi_{x,\nu}^{(1)}[J_{x,\nu}^{11}-J_{x,\nu}^{23}-J_{x,\nu}^{33}+j_{x,\nu}^{21}+j_{x,\nu}^{23}+j_{x,\nu}^{31}+j_{x,\nu}^{33}]} \nonumber \\ & \hspace{31mm} \times \! \int_{0}^{2\pi} \! \dfrac{d\phi_{x,\nu}^{(2)}}{2\pi} \; e^{i\phi_{x,\nu}^{(2)}[J_{x,\nu}^{22}-J_{x,\nu}^{31}-J_{x,\nu}^{33}+j_{x,\nu}^{21}+j_{x,\nu}^{23}+j_{x,\nu}^{31}+j_{x,\nu}^{33}]} \; \nonumber \\ & \hspace{31mm} \times \! \int_{0}^{2\pi} \! \dfrac{d\phi_{x,\nu}^{(3)}}{2\pi} \; e^{i\phi_{x,\nu}^{(3)}[J_{x,\nu}^{12}-j_{x,\nu}^{21}-j_{x,\nu}^{23}-j_{x,\nu}^{31}-j_{x,\nu}^{33}]} \nonumber \\ & \hspace{31mm} \times \! \int_{0}^{2\pi} \! \dfrac{d\phi_{x,\nu}^{(4)}}{2\pi} \; e^{i\phi_{x,\nu}^{(4)}[J_{x,\nu}^{13}-J_{x,\nu}^{21}-J_{x,\nu}^{31}+j_{x,\nu}^{21}+j_{x,\nu}^{23}+j_{x,\nu}^{31}+j_{x,\nu}^{33}]} \; \nonumber \\ & \hspace{31mm} \times \! \int_{0}^{2\pi} \! \dfrac{d\phi_{x,\nu}^{(5)}}{2\pi} \; e^{i\phi_{x,\nu}^{(5)}[J_{x,\nu}^{32}-J_{x,\nu}^{21}-J_{x,\nu}^{23}+j_{x,\nu}^{21}+j_{x,\nu}^{23}+j_{x,\nu}^{31}+j_{x,\nu}^{33}]} \; , \label{eq:qcd_partition5} \end{align} where we introduced the short hand notations \begin{equation*} \sum_{\{p\}} \, = \, \prod_{x,\nu<\rho} \, \prod_{a,b,c,d = 1}^{3} \, \sum_{p_{x,\nu\rho}^{abcd} = -\infty}^{\infty} \, , \qquad \sum_{\{l\}} \, = \, \prod_{x,\nu<\rho} \, \prod_{a,b,c,d = 1}^{3} \, \sum_{l_{x,\nu\rho}^{abcd} = 0}^{\infty} \, , \end{equation*} to express the sums over configurations of the cycle occupation numbers $p_{x,\nu\rho}^{abcd} \in \mathbb{Z}$ and the auxiliary plaquette variables $l_{x,\nu\rho}^{abcd} \in \mathbb{N}_{0}$, and \begin{equation} \sum_{\{m, \overline{m}\}} \, = \, \prod_{x,\nu} \, \prod_{a=2,3} \, \prod_{b=1,3} \, \sum_{m_{x,\nu}^{ab} = 0}^{N_{x,\nu}^{ab}} \, \sum_{\overline{m}_{x,\nu}^{ab} = 0}^{\overline{N}_{x,\nu}^{ab}} \, , \end{equation} for the sums over the configurations of the link-based auxiliary variables $m_{x,\nu}^{ab}$ and $\overline{m}_{x,\nu}^{ab}$ used in the binomial decomposition (\ref{eq:qcd_binomial}). The remarkable outcome of our approach is that, written as in (\ref{eq:qcd_partition5}), the Haar measure integrals can be solved in closed form. More precisely, the integrals over the $\theta_{x,\nu}^{(i)}$ angles $i=1,2,3$ give rise to beta functions \cite{Olver:2010:NHM:1830479} \begin{equation} \label{eq:qcd_betafunction} 2 \int_{0}^{\pi/2} \! d\theta (\cos \theta)^{n + 1} (\sin \theta)^{m + 1} = \B\left(\dfrac{n}{2} + 1\right| \left. \dfrac{m}{2} + 1\right) \, , \end{equation} whereas the integrals over the phases $\phi_{x,\nu}^{(j)}$ $j=1,2,\dots,5$ give rise to Kronecker deltas (we use the notation $\delta(n) \equiv \delta_{n,0}$), which impose constraints on the dual variables. Performing the gauge field integration we find \begin{equation} \label{eq:qcd_dualpartitiong} Z \, = \, \sum_{\{p\}} \, W_{G}[p] \, C_{G}[p] \, . \end{equation} The partition function is a sum over the configurations of the cycle occupation numbers $p_{x,\nu\rho}^{abcd}\in \mathbb{Z}$. The factor $C_{G}[p]$ collects the four constraints which arise from the integration of the four phases $\phi_{x,\nu}^{(j)}$, $j=1,2,4,5$: \begin{align} C_G[p] \; =& \; \prod_{x,\nu} \delta(J_{x,\nu}^{11} + J_{x,\nu}^{12} - J_{x,\nu}^{33} - J_{x,\nu}^{23}) \; \delta(J_{x,\nu}^{22} + J_{x,\nu}^{12} - J_{x,\nu}^{33} - J_{x,\nu}^{31}) \nonumber \\ & \hspace{3mm}\times \delta(J_{x,\nu}^{13} + J_{x,\nu}^{12} - J_{x,\nu}^{31} - J_{x,\nu}^{21}) \; \delta(J_{x,\nu}^{32} + J_{x,\nu}^{12} - J_{x,\nu}^{23} - J_{x,\nu}^{21}) \; . \label{eq:qcd_cg} \end{align} These Kronecker deltas enforce relations between different color components of the fluxes $J_{x,\nu}^{ab}$ (\ref{eq:qcd_Jfluxes}) at every link $(x,\nu)$, thus limiting the number of admissible configurations $\{p\}$ of the cycle occupation numbers $p_{x,\mu\rho}^{abcd}$. In Eq.~(\ref{eq:qcd_cg}) we have already taken into account another constraint, which origins from the $\phi_{x,\nu}^{(3)}$ integral in (\ref{eq:qcd_partition5}): \begin{equation} \label{eq:qcd_constrain} j_{x,\nu}^{21} + j_{x,\nu}^{23} + j_{x,\nu}^{31} + j_{x,\nu}^{33} \, = \, J_{x,\nu}^{12} \, . \end{equation} This constraint relates $J_{x,\nu}^{12}$ to the auxiliary fluxes $j_{x,\nu}^{ab} = m_{x,\nu}^{ab} - \overline{m}_{x,\nu}^{ab}$ for the variables $m_{x,\nu}^{ab}$ and $\overline{m}_{x,\nu}^{ab}$, introduced in (\ref{eq:qcd_binomial}) for the binomial decomposition of the $(2,1)$, $(2,3)$, $(3,1)$ and $(3,3)$ matrix elements of the parametrization (\ref{eq:qcd_parametrization}). To obtain (\ref{eq:qcd_cg}) we used (\ref{eq:qcd_constrain}) to substitute the sum $j_{x,\nu}^{21} + j_{x,\nu}^{23} + j_{x,\nu}^{31} + j_{x,\nu}^{33}$ with $J_{x,\nu}^{21}$ in the integrals over $\phi_{x,\nu}^{(j)}$, $j=1,2,4,5$ in (\ref{eq:qcd_partition5}). We included the resulting four Kronecker deltas in the expression (\ref{eq:qcd_cg}) for the gauge constraint $C_{G}[p]$, while (\ref{eq:qcd_constrain}) is incorporated in the weight $W_{G}[p]$. The weight factor $W_{G}[p]$ is itself a sum $\sum_{\{l,m,\overline{m}\}}$ over the auxiliary plaquette variables $l_{x,\nu\rho}^{abcd}\in \mathbb{N}_{0}$ and the auxiliary link variables $m_{x,\nu}^{ab} \in \{0,1 \dots N_{x,\nu}^{ab} \}$ and $\overline{m}_{x,\nu}^{ab} \in \{0,1 \dots \overline{N}_{x,\nu}^{ab} \}$: \small \begin{align} &W_{G} [p] \, = \, 2^{4V} \!\! \sum_{\{l,m,\overline{m}\}} \Bigg[\prod_{x,\nu} \delta(J_{x,\nu}^{12} - j_{x,\nu}^{21} - j_{x,\nu}^{23} - j_{x,\nu}^{31} - j_{x,\nu}^{33})\Bigg] \Bigg[\prod_{x,\nu} (-1)^{J_{x,\nu}^{12} + J_{x,\nu}^{23} + J_{x,\nu}^{31} - j_{x,\nu}^{23} - j_{x,\nu}^{31}} \Bigg] \nonumber \\ &\hspace{29mm} \times \Bigg[ \prod_{x,\nu} \prod_{a=2,3} \prod_{b=1,3} \binom{N_{x,\nu}^{ab}}{m_{x,\nu}^{ab}} \binom{\overline{N}_{x,\nu}^{ab}}{\overline{m}_{x,\nu}^{ab}}\Bigg] \Bigg[\prod_{x,\nu<\rho} \prod_{a,b,c,d} \dfrac{ \left( \beta/2 \right)^{|p_{x,\nu\rho}^{abcd}| + 2l_{x,\nu\rho}^{abcd}}}{\left(|p_{x,\nu\rho}^{abcd}| + l_{x,\nu\rho}^{abcd}\right)!l_{x,\nu\rho}^{abcd}!}\Bigg] \nonumber \\ &\!\times \!\Bigg[ \prod_{x,\nu} \B\left(\dfrac{S_{x,\nu}^{11} + S_{x,\nu}^{13} + S_{x,\nu}^{22} + S_{x,\nu}^{32}}{2} + 2\right.\left|\dfrac{S_{x,\nu}^{12} + s_{x,\nu}^{21} +s_{x,\nu}^{23} + s_{x,\nu}^{31} + s_{x,\nu}^{33}}{2} + 1\right) \nonumber \\ &\!\times\!\! \B\!\left(\!\!\dfrac{S_{x,\nu}^{11} \! + \! s^{21}_{x,\rho} \! + \! S_{x,\nu}^{23}\! -\! s_{x,\nu}^{23}\! + s_{x,\nu}^{31}\! + \!S_{x,\nu}^{33}\! - \!s_{x,\nu}^{33}}{2} \!+\! 1\right.\!\!\left|\dfrac{\!S_{x,\nu}^{13}\! +\! S_{x,\nu}^{21}\! - \!s_{x,\nu}^{21}\! +\! s_{x,\nu}^{23} \!+\! S_{x,\nu}^{31}\! -\! s_{x,\nu}^{31}\! +\! s_{x,\nu}^{33}}{2}\! + \!\!1\!\right) \nonumber \\ &\times\!\! \B\!\left(\!\!\dfrac{s_{x,\nu}^{21} \!+ \!S^{22}_{x,\rho}\! + \!s_{x,\nu}^{23} \!+\! S_{x,\nu}^{31}\! -\! s_{x,\nu}^{31}\! +\! S_{x,\nu}^{33}\! -\! s_{x,\nu}^{33}}{2}\! +\! 1\right.\!\!\left|\dfrac{\!S_{x,\nu}^{21} \!- \!s_{x,\nu}^{21} \!+ \!S_{x,\nu}^{23} \!- \!s_{x,\nu}^{23} \!+ \!s_{x,\nu}^{31} \!+ \!S_{x,\nu}^{32} \!+ \!s_{x,\nu}^{33}}{2}\! \!+\! 1\! \!\right) \!\!\!\Bigg]. \label{eq:qcd_wg} \end{align} \normalsize The configurations of the auxiliary link variables $m_{x,\nu}^{ab}$ and $\overline{m}_{x,\nu}^{ab}$ are constrained by the value of $J_{x,\nu}^{12}$ at every link $(x,\nu)$, as expressed by the Kronecker delta in the first line in (\ref{eq:qcd_wg}). The two factors in the second line in (\ref{eq:qcd_wg}) are the weights arising from the binomial decomposition and the Taylor expansion respectively. Finally, the beta functions arise from the $\theta_{x,\nu}^{(j)}$, $j=1,2,3$, integrals in (\ref{eq:qcd_partition5}). Notice that in (\ref{eq:qcd_wg}) there is an explicit sign factor. It origins from the minus signs in the parametrization (\ref{eq:qcd_parametrization}) of the SU(3) group elements. Let us now spend a few words discussing the gauge constraints in Eq.~(\ref{eq:qcd_cg}). Understanding the constraints is of crucial importance because, as we mentioned already several times in this thesis, they are the primary manifestation of the symmetry of the original theory in the dual representation. Explicitly the Kronecker deltas in (\ref{eq:qcd_cg}) require the following equalities to hold at every link $(x,\nu)$ of the lattice: \begin{gather} \label{eq:qcd_constraint1} J_{x,\nu}^{11} + J_{x,\nu}^{12} = J_{x,\nu}^{23} + J_{x,\nu}^{33} \; ,\\ \label{eq:qcd_constraint2} J_{x,\nu}^{22} + J_{x,\nu}^{12} = J_{x,\nu}^{33} + J_{x,\nu}^{31} \; ,\\ \label{eq:qcd_constraint4} J_{x,\nu}^{13} + J_{x,\nu}^{12} = J_{x,\nu}^{31} + J_{x,\nu}^{21} \; ,\\ \label{eq:qcd_constraint5} J_{x,\nu}^{23} + J_{x,\nu}^{21} = J_{x,\nu}^{32} + J_{x,\nu}^{12}\; . \end{gather} The requirements imposed by these constraints can be better understood by using linear combinations of (\ref{eq:qcd_constraint1})--(\ref{eq:qcd_constraint5}). So, for example, if we add $J_{x,\nu}^{11}$ on both sides of Eq.~(\ref{eq:qcd_constraint4}) we obtain \begin{equation*} J_{x,\nu}^{11} + J_{x,\nu}^{12} + J_{x,\nu}^{13} \, = \, J_{x,\nu}^{11} + J_{x,\nu}^{21} + J_{x,\nu}^{31} \, . \end{equation*} This relation is represented schematically in the top left plot of Fig~\ref{fig:qcd_constraints}. From the figure it is easy to see that the constraint requires the sum of all the fluxes coming out of color 1 at site $x$ on the link $(x,\nu)$ to be equal to the sum of all the color fluxes going into color 1 at site $x + \hat{\nu}$. Therefore, we can interpret this relation as a flux conservation constraint for the color layer 1. The analogous relation for color 2 can be obtained by adding $J_{x,\nu}^{22}$ on both sides of Eq.~(\ref{eq:qcd_constraint5}), while adding Eq.~(\ref{eq:qcd_constraint4}) to Eq.~(\ref{eq:qcd_constraint5}) and then $J_{x,\nu}^{33}$ on both sides of the resulting equation gives the relation for color 3. Thus we obtain the following three constraints \begin{gather} \label{eq:qcd_fluxconservation1} J_{x,\nu}^{11} + J_{x,\nu}^{12} + J_{x,\nu}^{13} \, = \, J_{x,\nu}^{11} + J_{x,\nu}^{21} + J_{x,\nu}^{31} \, , \\ \label{eq:qcd_fluxconservation2} J_{x,\nu}^{22} + J_{x,\nu}^{21} + J_{x,\nu}^{23} \, = \, J_{x,\nu}^{22} + J_{x,\nu}^{12} + J_{x,\nu}^{32} \, , \\ \label{eq:qcd_fluxconservation3} J_{x,\nu}^{33} + J_{x,\nu}^{32} + J_{x,\nu}^{31} \, = \, J_{x,\nu}^{33} + J_{x,\nu}^{23} + J_{x,\nu}^{13} \, , \end{gather} which we graphically illustrate in the plots in the first line of Fig.~\ref{fig:qcd_constraints}. Since (\ref{eq:qcd_fluxconservation1}) -- (\ref{eq:qcd_fluxconservation3}) impose the conservation of fluxes going through each of the 3 color layers at every link, we refer to them as \textit{color flux conservation} constraints. \begin{figure} \begin{center} \vskip5mm \includegraphics[width=15cm]{chapters/plots/qcd_constraints1.pdf}\\ \vskip10mm \includegraphics[width=8cm]{chapters/plots/qcd_constraints2.pdf}% \end{center} \caption{Schematic representation of the three \textit{color conservation constraints} (\ref{eq:qcd_fluxconservation1}) -- (\ref{eq:qcd_fluxconservation3}) (the three plots at the top) and the \textit{color exchange constraint} (\ref{eq:qcd_fluxexchange}) (bottom plot). They impose relations between the fluxes $J_{x,\nu}^{ab}$ and admissible configurations of the cycle occupation numbers $p_{x,\nu\rho}^{abcd}$ have to respect these constraints. \label{fig:qcd_constraints}} \end{figure} Another set of constraints that results from linearly combining Eqs.~(\ref{eq:qcd_constraint1})--(\ref{eq:qcd_constraint5}) is: \begin{equation} \label{eq:qcd_fluxexchange} J_{x,\nu}^{11} + J_{x,\nu}^{12} + J_{x,\nu}^{13} \, = \, J_{x,\nu}^{22} + J_{x,\nu}^{21} + J_{x,\nu}^{23} \, = \, J_{x,\nu}^{33} + J_{x,\nu}^{32} + J_{x,\nu}^{31} \, . \end{equation} The first equality in (\ref{eq:qcd_fluxexchange}) is obtained by adding Eq.~(\ref{eq:qcd_constraint1}) to Eq.~(\ref{eq:qcd_constraint4}) and using (\ref{eq:qcd_constraint2}) to substitute $J_{x,\nu}^{33} + J_{x,\nu}^{31}$ with $J_{x,\nu}^{22} + J_{x,\nu}^{21}$. The last is obtained by adding Eq.~(\ref{eq:qcd_constraint2}) to Eq.~(\ref{eq:qcd_constraint5}). The chain of equalities in (\ref{eq:qcd_fluxexchange}) requires the magnitude of the flux to be the same on each color layer at all links $(x,\nu)$. We refer to the set of constraints (\ref{eq:qcd_fluxexchange}) as \textit{color exchange} constraints, and we illustrate them in the plot at the bottom of Fig.~(\ref{fig:qcd_constraints}). We stress at this point that Eqs.~(\ref{eq:qcd_fluxconservation1}) -- (\ref{eq:qcd_fluxexchange}) are over-complete. Nonetheless, the gauge constraints $C_{G}[p]$ are easier to understand in the form (\ref{eq:qcd_fluxconservation1}) -- (\ref{eq:qcd_fluxexchange}). As a result of applying the ACC method to pure SU(3) gauge theory, we obtained a representation of the partition function where the dynamical degrees of freedom are integer valued. The cycle occupation numbers $p_{x,\mu\rho}^{abcd} \in \mathbb{Z}$, which are attached to plaquettes, have to satisfy the constraints encoded in the function $C_{G}[p]$. These constraints impose relations between the link color fluxes $J_{x,\nu}^{ab}$ generated by the cycle occupation numbers attached to the link $(x,\nu)$. We categorized the constraints into two sets: the color flux conservation constraints (\ref{eq:qcd_fluxconservation1}) -- (\ref{eq:qcd_fluxconservation3}), which force the flux going through a color to be the same at all sites, and the color flux exchange constraint (\ref{eq:qcd_fluxexchange}), which requires the fluxes going through the three colors to match. The admissible configurations, i.e., the configurations that obey the gauge constraint $C_{G}[p]$, come with the weight factor $W_{G}[p]$, whose explicit expression is given in (\ref{eq:qcd_wg}). $W_{G}[p]$ is itself a sum over the configurations of the auxiliary plaquette variables $l_{x,\nu\rho}^{abcd} \in \mathbb{N}_{0}$, and the auxiliary link variables $m_{x,\nu}^{ab} \in \{0,1\dots N_{x,\nu}^{ab}\}$ and $\overline{m}_{x,\nu}^{ab} \in \{0,1\dots \overline{N}_{x,\nu}^{ab}\}$, which come from the binomial decomposition of the matrix elements $(2,1)$, $(2,3)$, $(3,1)$ and $(3,3)$ of the parametrization (\ref{eq:qcd_parametrization}) of the SU(3) elements. The auxiliary plaquette variables $l_{x,\nu\rho}^{abcd} \in \mathbb{N}_{0}$ are unconstrained and act as a background field, while the link variables $m_{x,\nu}^{ab} \in \{0,1\dots N_{x,\nu}^{ab}\}$ and $\overline{m}_{x,\nu}^{ab} \in \{0,1\dots \overline{N}_{x,\nu}^{ab}\}$ are constrained by the value of the $J_{x,\nu}^{12}$ flux at every link. Admissible configurations of the cycle occupation numbers can be built in a simple way: one starts setting one cycle occupation number to one. This causes the constraints (\ref{eq:qcd_cg}) to be violated on all the four links that contour the plaquette. To obviate this problem one has two possibilities. The first one consists in activating at least two other cycle occupation numbers on the same plaquette, so that the constraints are fulfilled. However, the configurations obtained in this way are not relevant for the long range physics. Alternatively, one can activate a cycle occupation number on an adjacent plaquette, so that the constraints are satisfied on the common link. One must then proceed similarly for the other six links that contour the two plaquettes. For the resulting 2D surface the constraints will still be violated along the boundary, unless the surface is closed. Therefore, in our dual representation, the long range physics, which is relevant for the continuum limit, is described by closed \textit{worldsheets}. Before continuing with the discussion about fermions, we stress that in our dual representation (\ref{eq:qcd_dualpartitiong}) there is an explicit sign factor $(-1)^{\sum_{x,\mu}J_{x,\mu}^{12} + J_{x,\mu}^{23} + J_{x,\mu}^{31} - j_{x,\mu}^{23} - j_{x,\mu}^{31}}$. This sign factor origins from the explicit minus signs in the parametrization of the SU(3) matrices (\ref{eq:qcd_parametrization}). This implies that for a Monte Carlo simulation of the ACC dual form of the partition sum (\ref{eq:qcd_dualpartitiong}) a strategy for a partial resummation needs to be found. \section{Abelian color flux dualization of fermions: the strong coupling limit \label{sec:qcd_sc}} In this section we discuss the abelian color flux (ACF) dualization of fermions, focusing first on the strong coupling limit. In this limit $\beta = 0$, therefore the gauge action is absent and the continuum limit cannot be performed. Nevertheless, the strong coupling limit of lattice QCD shares some non-perturbative properties with QCD in the continuum, like confinement and the spontaneous breaking of chiral symmetry. For our purposes the worldline representation of strong coupling QCD is interesting because, even if the constraints have the same structure of those for the gauge degrees of freedom, they are simpler to interpret thanks to an additional constraint coming from the Pauli principle. The fermionic partition function is given by \begin{equation} Z_{F}[U] = \int \! D[\overline{\psi},\psi] \, e^{-S_{F}[U,\psi,\overline{\psi}]} \, , \label{eq:qcd_partitionf} \end{equation} where $\psi_{x}$ and $\overline{\psi}_{x}$ are 3-component Grassmann vectors \begin{equation} \psi_{x} = \left( \begin{array}{c} \psi_{x}^{1} \\ \psi_{x}^{2} \\ \psi_{x}^{3} \\ \end{array} \right) \ ,\qquad \overline{\psi}_{x} = \left( \overline{\psi}_{x}^{1} \ \ \overline{\psi}_{x}^{2} \ \ \overline{\psi}_{x}^{3} \right) \, , \label{eq:qcd_grassmann} \end{equation} that live on the sites of the four-dimensional lattice, with anti-periodic boundary conditions in Euclidean time, i.e., $\nu = 4$ and periodic boundary conditions in the spatial directions, i.e., $\nu =1,2,3$. The measure is a product over Grassmann measures $\int \! D[\overline{\psi},\psi] = \prod_{x}\prod_{a = 1}^{3} \int d\psi_{x}^{a} d\overline{\psi}^{a}_{x}$. The staggered fermion action is given by \begin{align} \nonumber &S_{F}[U,\psi,\overline{\psi}] = \sum_{x}\biggl[ m \overline{\psi}_{x} \psi_{x} + \sum_{\nu} \dfrac{\gamma_{x,\nu}}{2} \left( \overline{\psi}_x U_{x,\nu} \psi_{x + \hat{\nu}}\, e^{\mu \delta_{\nu,4}} -\, \overline{\psi}_{x + \hat{\nu}} U_{x,\nu}^{\dagger} \psi_{x}\, e^{-\mu \delta_{\nu,4}} \right) \biggr]\\ &= \sum_{x}\biggl[ m \sum_{a = 1}^{3} \overline{\psi}_{x}^{a} \psi_{x}^{a} + \sum_{\nu} \dfrac{\gamma_{x,\nu}}{2} \sum_{a, b = 1}^{3} \left( \overline{\psi}_{x}^{a} U_{x,\nu}^{ab} \psi_{x + \hat{\nu}}^{b}\, e^{\mu \delta_{\nu,4}} -\, \overline{\psi}_{x + \hat{\nu}}^{b} U_{x,\nu}^{ab \, \star} \psi_{x}^{a}\, e^{-\mu \delta_{\nu,4}} \right) \biggr] \, , \label{eq:qcd_actionf} \end{align} where $m$ is the fermion's mass, $\gamma_{x,\nu}$ are the staggered sign factors defined as \begin{equation*} \gamma_{x,1} = 1 \, , \quad \gamma_{x,2} = (-1)^{x_{1}} \, , \quad \gamma_{x,3} = (-1)^{x_{1} + x_{2}} \, , \quad \gamma_{x,4} = (-1)^{x_{1} + x_{2} + x_{3}} \, , \end{equation*} and $U_{x,\nu}$ are the gauge degrees of freedom, which live on the links $(x,\nu)$. To obtain the partition function in the strong coupling limit it is sufficient to integrate the fermionic partition function (\ref{eq:qcd_partitionf}): \begin{equation*} Z = \int \! D[U] \, Z_{F}[U] \, . \end{equation*} The measure $\int \! D[U]$ is the product of SU(3) Haar measures on the links of the lattice, $\int \! D[U] = \prod_{x,\nu} \int_{SU(3)} \! dU_{x,\nu}$ whose explicit expression is given in (\ref{eq:qcd_haarmeasure}). Notice that in Eq.~(\ref{eq:qcd_actionf}) we introduced a chemical potential $\mu$, which couples to the temporal hopping terms in the canonical way. In the first line of Eq.~(\ref{eq:qcd_actionf}) we used matrix-vector notation for gauge links and fermions, while in the second line the sums over color indices are made explicit, using the color labels $a,b = 1,2,3$. In this decomposition all terms of the action are single Grassmann bilinears, which therefore commute. Once again the rewriting of the action into its minimal units is what allows us to proceed further with the dualization. For the fermionic partition function we obtain \begin{align} &Z_{F}[U] = \nonumber \\ &= \int \! D[\overline{\psi},\psi] \prod_{x} \prod_{a = 1}^{3} e^{- m \overline{\psi}_{x}^{a} \psi_{x}^{a}} \prod_{x, \nu} \prod_{a, b = 1}^{3} e^{- \frac{\gamma_{x,\nu}}{2} \overline{\psi}_{x}^{a} U_{x,\nu}^{ab} \psi_{x + \hat{\nu}}^{b} e^{\mu \delta_{\nu,4}}} e^{\frac{\gamma_{x,\nu}}{2} \overline{\psi}_{x + \hat{\nu}}^{b} U_{x,\nu}^{ab \, \star} \psi_{x}^{a} e^{-\mu \delta_{\nu,4}}} \nonumber \\ &= \int \! D[\overline{\psi},\psi] \prod_{x} \prod_{a = 1}^{3} \sum_{s^a_{x} = 0}^{1} \left( m \overline{\psi}_{x}^{a} \psi_{x}^{a} \right)^{s_{x}^{a}} \prod_{x, \nu} \prod_{a, b = 1}^{3} \sum_{k_{x,\nu}^{ab} = 0}^{1} \left(- \frac{\gamma_{x,\nu}}{2}\, \overline{\psi}_{x}^{a} U_{x,\nu}^{ab} \psi_{x + \hat{\nu}}^{b}\, e^{\mu \delta_{\nu,4}} \right)^{k_{x,\nu}^{ab}} \nonumber\\ &\hspace*{6.1cm} \times \prod_{x, \nu} \prod_{a, b = 1}^{3} \sum_{\overline{k}_{x,\nu}^{ab} = 0}^{1} \left( \frac{\gamma_{x,\nu}}{2}\, \overline{\psi}_{x + \hat{\nu}}^{b} U_{x,\nu}^{ab \, \star} \psi_{x}^{a}\, e^{-\mu \delta_{\nu,4}} \right)^{\overline{k}_{x,\nu}^{ab}} \nonumber \\ \nonumber &= \left(\dfrac{1}{2}\right)^{\!\!3V} \! \! \! \sum_{\{s,k,\overline{k}\}} (2m)^{\sum_{x,a} s_{x}^{a}} \, e^{\mu \sum_{x,ab} [k_{x,\hat{4}}^{ab} - \overline{k}_{x,\hat{4}}^{ab}]} \prod_{x,\nu} \prod_{a, b} \left( U_{x,\nu}^{ab} \right)^{k_{x,\nu}^{ab}} \left( U_{x,\nu}^{ab \, \star} \right)^{\overline{k}_{x,\nu}^{ab}} \\ & \times \prod_{x,\nu} \prod_{a, b} (-1)^{k_{x,\nu}^{ab}} (\gamma_{x,\nu})^{k_{x,\nu}^{ab} + \overline{k}_{x,\nu}^{ab}} \int \! D[\overline{\psi},\psi] \prod_{x,a} (\overline{\psi}_{x}^{a} \psi_{x}^{a})^{s_{x}^{a}} \prod_{x, \nu} \prod_{a, b} ( \overline{\psi}_{x}^{a} \psi_{x + \hat{\nu}}^{b} )^{k_{x,\nu}^{ab}} ( \overline{\psi}_{x + \hat{\nu}}^{b} \psi_{x}^{a})^{\overline{k}_{x,\nu}^{ab}} \, . \label{eq:qcd_partitionf2} \end{align} In the first step we rewrote the exponential of sums as products of exponentials. In the second step we then Taylor-expanded each of these exponentials. Notice that the Taylor series terminate after the linear term. This is due to the nilpotency of the Grassmann variables that causes any higher order of the series to vanish. The three expansion indices that we introduced, one for every bilinear of the action, are the new dual variables for fermions: $s_{x}^{a} = 0, 1$ is the dual variable corresponding to the color component $a$ of the mass term on site $x$, $k_{x,\nu}^{ab} = 0, 1$ represents the forward hop from color $a$ to color $b$ on the link $(x,\nu)$, and $\overline{k}_{x,\nu}^{ab} = 0, 1$ is the respective backward hop on the same link. The graphical representation of the dual variables for fermions is shown in Fig. \ref{fig:qcd_grassmann}. The monomer variables $s_{x}^{a}$ are illustrated as circles around the color component $a$ of the site $x$, while the forward and backward hops $k_{x,\nu}^{ab}, \overline{k}_{x,\nu}^{ab}$ are shown as oriented arrows connecting color $a$ and $b$ on the link $(x,\nu)$. \begin{figure} \includegraphics[width=15cm]{chapters/plots/qcd_grassmann.pdf}% \caption{Graphical representation of the fermion's dual variables. In the first column we show the monomers $s_{x}^{a}$, while the arrows represent the dual variables $k_{x,\nu}^{ab}$ and $\overline{k}_{x,\nu}^{ab}$ for the forward and backward hopping respectively. With these link variables it is possible to build dimers and oriented loops that, together with monomers constitute the admissible configurations for fermions. \label{fig:qcd_grassmann}} \end{figure} In the last step of Eq.~(\ref{eq:qcd_partitionf2}) we reorganized the terms: we collected an overall factor $(1/2)^{3V}$ and brought all factors independent of the Grassmann variables in front of the Grassmann integrals. We also introduced the notation \begin{equation} \sum_{\{s,k,\overline{k}\}} = \Bigg[\prod_{x,a} \sum_{s_{x}^{a} = 0}^{1} \Bigg] \Bigg[\prod_{x,\nu} \prod_{a,b} \sum_{k_{x,\nu}^{ab} = 0}^{1} \sum_{\overline{k}_{x,\nu}^{ab} = 0}^{1} \Bigg] \, , \end{equation} to denote the sum over all the possible configurations of the fermion dual variables. The Grassmann integral in the last line of (\ref{eq:qcd_partitionf2}) is non-vanishing only if each Grassmann variable $\overline{\psi}_{x}^{a}\psi_{x}^{a}$ appears exactly once. This condition, which implements Pauli's exclusion principle, can be expressed by means of the following fermion constraint \begin{equation} \label{eq:qcd_cf} C_{F}[s,k,\overline{k}] \, = \, \prod_{x,a} \delta \bigg( 1 - s_{x}^{a} - \dfrac{1}{2} \sum_{\nu,b} \big[k_{x,\nu}^{ab} + \overline{k}_{x,\nu}^{ab} + k_{x - \hat{\nu},\nu}^{ba} + \overline{k}_{x - \hat{\nu},\nu}^{ba} \big] \bigg) \, . \end{equation} The configurations that satisfy the fermion constraint (\ref{eq:qcd_cf}) are the ones that completely saturate the four-dimensional lattice: the three color layers of every site must be either occupied by a monomer ($s_{x}^{a} = 1$), be the endpoint of a dimer ($k_{x,\nu}^{ab} = \overline{k}_{x,\nu}^{ab} = 1$), or run through by a loop (closed chains of $k_{x,\nu}^{ab} = 1$ and $\overline{k}_{x,\nu}^{ab} = 1$). Notice that Eq.~(\ref{eq:qcd_cf}) is the analogous of the triality constraint in \cite{Karsch:1988zx}. When $C_{F}[s,k,\overline{k}]$ is fulfilled, the Grassmann integral in the last line of Eq.~(\ref{eq:qcd_partitionf2}) gives either $+1$ or $-1$, depending on the values of the dual variables $s_{x}^{a}$, $k_{x,\nu}^{ab}$ and $\overline{k}_{x,\nu}^{ab}$. Forward and backward hops also contribute with signs, since they activate the staggered sign factor $(\gamma_{x,\nu})^{k_{x,\nu}^{ab} + \overline{k}_{x,\nu}^{ab}}$, as well as the factor $(-1)^{k_{x,\nu}^{ab}}$. Moreover, loops winding in the temporal direction also produce signs, caused by the anti-periodic boundary conditions for fermions. Because of the multitude of sign sources, we now discuss in detail the sign contributions coming from the monomers, the dimers and the loops. Monomers $\overline{\psi}_{x}^{a} \psi_{x}^{a}$ are activated by setting $s_{x}^{a} = 1$, which leads to the contribution of a factor $2m$. Furthermore, the Grassmann variables $\overline{\psi}_{x}^{a} \psi_{x}^{a}$ are already in the canonical order for the Grassmann integral and do not introduce any signs. Hence, we conclude that monomers are sign free objects in our representation. Dimers are built by setting $k_{x,\nu}^{ab} = \overline{k}_{x,\nu}^{ab} = 1$. This corresponds to activating the factor \begin{equation} \label{eq:qcd_dimer} \overline{\psi}_{x}^{a}\psi_{x + \hat{\nu}}^{b} \overline{\psi}_{x + \hat{\nu}}^{b}\psi_{x}^{a} \, = \, - \overline{\psi}_{x}^{a} \psi_{x}^{a} \overline{\psi}_{x + \hat{\nu}}^{b} \psi_{x + \hat{\nu}}^{b} \end{equation} in the Grassmann integral. The minus sign on the right hand-side of (\ref{eq:qcd_dimer}) results from the reordering of the Grassmann variables. Anyway it is compensated by the explicit minus sign arising from the activation of the forward hop $(-1)^{k_{x,\nu}^{ab}} = (-1)^{1} = -1$. Finally, the staggered sign factor for dimers is always positive, i.e., $(\gamma_{x,\nu})^{k_{x,\nu}^{ab} + \overline{k}_{x,\nu}^{ab}} = (\gamma_{x,\nu})^{2} = 1$, therefore also dimers do not contribute with negative signs. Loops are built by setting closed chains of $k_{x,\nu}^{ab}$ and $\overline{k}_{x,\nu}^{ab}$ variables to 1. Each loop $\mathcal{L}$ picks up an overall minus sign from the necessary interchanges to bring the Grassmann variables into the canonical order for the integration. Moreover, each forward hop of the loop will contribute with a minus sign. Hence, for trivially closing loops, if $|\mathcal{L}|$ is the length of the loop $\mathcal{L}$, the sign coming from the forward hops is $(-1)^{|\mathcal{L}|/2}$. On the other hand, the forward hops sign for loops that wind in one of the spatial directions $(\nu = 1,2,3)$ or in the temporal direction $(\nu = 4)$ depends on the size $N_{\nu}$ of the lattice in that direction. We choose to restrict our attention to lattices in which $N_{\nu}$ are multiples of 4, so that we do not need to distinguish different cases. Loops winding in the temporal direction pick up an additional minus sign because of the anti-periodic boundary conditions. We denote this sign as $(-1)^{W_{\mathcal{L}}}$, where $W_{\mathcal{L}}$ is the net number of windings of the loop $\mathcal{L}$ around the compactified time. Finally we have to work out the contributions from the staggered sign factors. Let us first consider a loop closing around a single plaquette $(x,\nu\rho)$ with $\nu < \rho$. For this simple case the staggered sign is given by \begin{equation} \label{eq:qcd_staggered} \gamma_{x,\nu} \gamma_{x + \hat{\nu},\rho} \gamma_{x + \hat{\rho},\nu} \gamma_{x,\rho} = (-1)^{\sum_{i=1}^{\nu-1} x_{i}} (-1)^{\sum_{i=1}^{\rho-1} x_{i} + 1} (-1)^{\sum_{i=1}^{\nu-1} x_{i}} (-1)^{\sum_{i=1}^{\rho-1}x_{i}}= -1 \,, \end{equation} where we used the compact definition of the staggered sign \begin{equation*} \gamma_{x,\nu} = (-1)^{\sum_{i=1}^{\nu-1} x_{i}}\, , \qquad x = (x_1, \ x_2, \ x_3, x_4) \, . \end{equation*} The result (\ref{eq:qcd_staggered}) holds independently from the position $x$, and from the plane $\nu,\rho$ on which the plaquette lies. We can then elongate the loop by one plaquette. The product of the six staggered signs corresponding to the six links of the loop can equivalently be computed by multiplying the staggered signs for the two plaquettes bounded by the loop, i.e., $(-1)(-1) = +1$. The equivalence holds because in the product the staggered sign on the link common to the two plaquettes gets squared. Obviously this procedure can be iterated to build loops of any shape, and the staggered sign factor can be expressed as $(-1)^{P_\mathcal{L}}$, where $P_\mathcal{L}$ is the number of plaquettes necessary to cover the surface bounded by the loop $\mathcal{L}$. Since we have three layers of colors, the admissible configurations may also contain loops that wind around the same contour up to three times (see Fig.~\ref{fig:qcd_loop} for a simple example of such loop). For these cases we need a multiple covered surface (e.g., a surface covered 3 times for the example in the bottom plot of Fig.~\ref{fig:qcd_loop}) and the total number $P_{\mathcal{L}}$ of plaquettes in the surface spanned by the loop is understood in the sense that it also takes into account multiple coverings. We finally remark that, even if the surface that has the loop $\mathcal{L}$ as its boundary is not unique, those surfaces that share the same boundary always differ by an even number of plaquettes. Hence, the result $(-1)^{P_\mathcal{L}}$ is valid independently from the surface chosen to compute $P_\mathcal{L}$. Summarizing we found that in our representation, while monomers and dimers are sign free configurations, loops come with non-trivial signs which can be expressed as: \begin{equation} \label{eq:qcd_signl} \sign(\mathcal{L}) = (-1)^{1 + |\mathcal{L}|/2 + P_\mathcal{L} + W_\mathcal{L}} \, , \end{equation} where $|\mathcal{L}|$ is the length of the loop $\mathcal{L}$, $P_\mathcal{L}$ is the number of plaquettes necessary to cover the surface bounded by the loop $\mathcal{L}$ and $W_\mathcal{L}$ is the net winding number in the compactified temporal direction. To obtain the full partition sum at strong coupling $Z = \int \! D[U] Z_{F}[U]$ we still have to integrate the fermionic partition function $Z_{F}[U]$ over the product of SU(3) Haar measures. Putting things together the partition function reads \begin{equation} Z = \sum_{\{s,k,\overline{k}\}} C_{F}[s,k,\overline{k}] \, W_{F}[s,k,\overline{k}] \int \! D[U] \prod_{x, \nu} \prod_{a, b} \left( U_{x,\nu}^{ab} \right)^{k_{x,\nu}^{ab}} \left( U_{x,\nu}^{ab \, \star} \right)^{\overline{k}_{x,\nu}^{ab}} \, . \label{eq:qcd_partitionsc} \end{equation} $C_{F}[s,k,\overline{k}]$ is the fermion constraint given in Eq.~(\ref{eq:qcd_cf}) and $W_{F}[s,k,\overline{k}]$ is the weight for the fermion configurations defined as: \begin{align} W_{F}[s,k,\overline{k}] &= \dfrac{1}{2^{3V}} \prod_{\mathcal{L}} \sign (\mathcal{L}) \prod_{x} \bigg[\prod_{a} (2m)^{s_{x}^{a}}\bigg] \bigg[\prod_{ab} e^{\mu[k_{x,4}^{ab} - \overline{k}_{x,4}^{ab}]}\bigg] \nonumber \\ &= \dfrac{1}{2^{3V}} \prod_{\mathcal{L}} \sign (\mathcal{L}) \, e^{\mu \beta W_{\mathcal{L}}} \bigg[\prod_{x,a} (2m)^{s_{x}^{a}}\bigg] \, . \label{eq:qcd_wf} \end{align} The product $\prod_{\mathcal{L}}$ runs over all loops $\mathcal{L}$ that are formed by the $k_{x,\nu}^{ab}$ and $\overline{k}_{x,\nu}^{ab}$ variables. In the second step we reorganized the factor that gives the $\mu$-dependence. The chemical potential $\mu$ multiplies the difference $k_{x,4}^{ab} - \overline{k}_{x,4}^{ab}$ between the forward and backward hops on the links in the temporal direction. That quantity is obviously vanishing for dimers, since they are activated by setting $k_{x,\nu}^{ab} = \overline{k}_{x,\nu}^{ab} = 1$. Also trivially closing loop do not contribute to the difference $k_{x,4}^{ab} - \overline{k}_{x,4}^{ab}$, since they have the same amount of forward and backward hops. Therefore, only loops that wind in the temporal direction couple with the chemical potential. For them \begin{equation*} \sum_{x\in \mathcal{L}} \sum_{a,b} [k_{x,4}^{ab} - \overline{k}_{x,4}^{ab}] = N_{t} W_{\mathcal{L}} = \beta W_{\mathcal{L}}\, , \end{equation*} where $N_{t} = \beta$ is the temporal extent of the lattice and $W_{\mathcal{L}}$ is the temporal winding number of the loop $\mathcal{L}$. The chemical potential then enhances loops winding around the temporal direction with positive orientation $(W_{\mathcal{L}} > 0)$, while it suppresses loops that have a negative orientation of the winding $(W_{\mathcal{L}} < 0)$. If we then compare the $\mu$-dependence in the last line of (\ref{eq:qcd_wf}) with the usual form $e^{\mu \beta \mathcal{N}}$ for the coupling of the chemical potential with the net-particle number $\mathcal{N}$, it is straightforward to identify the equality \begin{equation} \sum_{\mathcal{L}} W_{\mathcal{L}} \, = \, \mathcal{N} \, . \label{eq:qcd_wl} \end{equation} The geometrical interpretation (\ref{eq:qcd_wl}) of the net-particle number $\mathcal{N}$ as the total temporal net-winding number of all fermion loops $\sum_{\mathcal{L}} W_{\mathcal{L}}$ is one of the most beautiful features of our worldline formulation of QCD. A clear advantage coming from (\ref{eq:qcd_wl}) is that in the dual representation the net-particle number is easily determined as a topological quantity, i.e., the total net-winding number for all the loops of a configuration. On the contrary, in the conventional representation the net-particle number can be quite challenging to compute, since it is given by the discretized integral over the zero component of the conserved vector current. Moreover, the interpretation (\ref{eq:qcd_wl}) opens the door to the implementation of simulations of the canonical ensemble \cite{Orasch:2017niz,Giuliani:2017qeo}. To obtain the final result for the strong coupling partition sum $Z$, we still have to perform the integrals over the SU(3) Haar measures in (\ref{eq:qcd_partitionsc}). Similarly to what we have done in the last section, we insert the explicit expressions (\ref{eq:qcd_parametrization}) and (\ref{eq:qcd_haarmeasure}) for the matrix elements $U_{x,\nu}^{ab}$ and the path integral measure $D[U]$ in the integral in (\ref{eq:qcd_partitionsc}). For the matrix elements $(2,1)$, $(2,3)$, $(3,1)$ and $(3,3)$, which are sums of complex numbers, we again use the binomial decomposition (\ref{eq:qcd_binomial}). In this way we introduce the link variables $m_{x,\nu}^{ab} \in \{0, k_{x,\nu}^{ab}\}$ and $\overline{m}_{x,\nu}^{ab} \in \{0, \overline{k}_{x,\nu}^{ab}\}$. Since $k_{x,\nu}^{ab}, \overline{k}_{x,\nu}^{ab}, m_{x,\nu}^{ab},\overline{m}_{x,\nu}^{ab} \in \{0,1\}$, all binomial factors $\binom{k_{x,\nu}}{m_{x,\nu}}, \binom{\overline{k}_{x,\nu}}{\overline{m}_{x,\nu}}$ are equal to 1 and we can therefore drop them here. We can then solve the gauge integrals in closed form and for the partition function we obtain \begin{equation} Z \, = \, \sum_{\{s,k,\overline{k}\}} C_{F}[s,k,\overline{k}] \, W_{F}[s,k,\overline{k}] \, C_{G}[k,\overline{k}] \, W_{G}[k,\overline{k}] \, . \label{eq:qcd_partitionsc2} \end{equation} The gauge integration in (\ref{eq:qcd_partitionsc}) has generated a gauge constraint $C_{G}[k,\overline{k}]$ and a weight factor $W_{G}[k,\overline{k}]$. To represent the constraints and the weight factor in a transparent way, we introduce combinations of the dual variables $k_{x,\nu}^{ab}$, $\overline{k}_{x,\nu}^{ab}$ and the auxiliary variables $m_{x,\nu}^{ab}$, $\overline{m}_{x,\nu}^{ab}$ for $(a,b) = (2,1), (2,3), (3,1), (3,3)$ as follows: \begin{align} \label{eq:qcd_shortsc} &K_{x,\nu}^{ab} = k_{x,\nu}^{ab} - \overline{k}_{x,\nu}^{ab} \, , \qquad \ P_{x,\nu}^{ab} = k_{x,\nu}^{ab} + \overline{k}_{x,\nu}^{ab} \, , \\ &j_{x,\nu}^{ab} = m_{x,\nu}^{ab} - \overline{m}_{x,\nu}^{ab} \, , \qquad s_{x,\nu}^{ab} = m_{x,\nu}^{ab} + \overline{m}_{x,\nu}^{ab} \, . \nonumber \end{align} As for the pure gauge case the constraints are generated by the integration over the four phases $\phi_{x,\nu}^{(j)}$, $j=1,2,4,5$ of the parametrization (\ref{eq:qcd_parametrization}) of the SU(3) matrices. We collect the corresponding four Kronecker deltas in the gauge constraint \begin{align} C_G[k,\overline{k}] \; =& \; \prod_{x,\nu} \delta(K_{x,\nu}^{11} + K_{x,\nu}^{12} - K_{x,\nu}^{33} - K_{x,\nu}^{23}) \; \delta(K_{x,\nu}^{22} + K_{x,\nu}^{12} - K_{x,\nu}^{33} - K_{x,\nu}^{31}) \nonumber \\ & \hspace{3mm}\times \delta(K_{x,\nu}^{13} + K_{x,\nu}^{12} - K_{x,\nu}^{31} - K_{x,\nu}^{21}) \; \delta(K_{x,\nu}^{32} + K_{x,\nu}^{12} - K_{x,\nu}^{23} - K_{x,\nu}^{21}) \; . \label{eq:qcd_cgsc} \end{align} For writing Eq.~(\ref{eq:qcd_cgsc}) we actually made use of the additional constraint which results from the $\phi_{x,\nu}^{(3)}$ integration \begin{equation} \label{eq:qcd_scconstraint3} J_{x,\nu}^{12} = j_{x,\nu}^{21} + j_{x,\nu}^{23} + j_{x,\nu}^{31} + j_{x,\nu}^{33} \, , \end{equation} to substitute the sum of the auxiliary fluxes $j_{x,\nu}^{ab} = m_{x,\nu}^{ab} - \overline{m}_{x,\nu}^{ab}$ with the fermion color flux $K_{x,\nu}^{12} = k_{x,\nu}^{12} - \overline{k}_{x,\nu}^{12}$. Eq.~(\ref{eq:qcd_scconstraint3}) relates the value of the auxiliary variables $m_{x,\nu}^{ab}$ and $\overline{m}_{x,\nu}^{ab}$ to the $k_{x,\nu}^{12}, \overline{k}_{x,\nu}^{12}$ fermion color fluxes, thus limiting the amount of admissible configurations. Later we will discuss the implications of the constraint (\ref{eq:qcd_scconstraint3}) in more detail. The weight factor $W_{G}[k,\overline{k}]$ has the following form \small \begin{align} &W_{G} [k,\overline{k}] = 2^{4V} \! \! \sum_{\{m,\overline{m}\}} \! \! \Bigg[\prod_{x,\nu} \delta(K_{x,\nu}^{12} - j_{x,\nu}^{21} - j_{x,\nu}^{23} - j_{x,\nu}^{31} - j_{x,\nu}^{33})\Bigg] \Bigg[\prod_{x,\nu} (-1)^{K_{x,\nu}^{12} + K_{x,\nu}^{23} + K_{x,\nu}^{31} - j_{x,\nu}^{23} - j_{x,\nu}^{31}} \Bigg] \nonumber \\ &\times \Bigg[ \prod_{x,\nu} \B\left(\dfrac{P_{x,\nu}^{11} + P_{x,\nu}^{13} + P_{x,\nu}^{22} + P_{x,\nu}^{32}}{2} + 2\right.\left|\dfrac{P_{x,\nu}^{12} + s_{x,\nu}^{21} +s_{x,\nu}^{23} + s_{x,\nu}^{31} + s_{x,\nu}^{33}}{2} + 1\right) \nonumber \\ &\! \times\! \B\!\left(\!\!\dfrac{P_{x,\nu}^{11} \! + \! s^{21}_{x,\rho} \! + \! P_{x,\nu}^{23}\! -\! s_{x,\nu}^{23}\! + s_{x,\nu}^{31}\! + \!P_{x,\nu}^{33}\! - \!s_{x,\nu}^{33}}{2} \!+\! 1\right.\!\left|\dfrac{\!P_{x,\nu}^{13}\! +\! P_{x,\nu}^{21}\! - \!s_{x,\nu}^{21}\! +\! s_{x,\nu}^{23} \!+\! P_{x,\nu}^{31}\! -\! s_{x,\nu}^{31}\! +\! s_{x,\nu}^{33}}{2}\! + \!1\!\right) \nonumber \\ &\! \times\! \B\!\left(\!\!\dfrac{s_{x,\nu}^{21}\! \!+ \!P^{22}_{x,\rho}\! + \!s_{x,\nu}^{23}\! \!+\! P_{x,\nu}^{31}\! -\! s_{x,\nu}^{31}\! \! +\! P_{x,\nu}^{33}\! -\! s_{x,\nu}^{33}}{2}\! \! +\! 1\right.\!\left|\dfrac{\!P_{x,\nu}^{21} \!- \!s_{x,\nu}^{21}\! \!+ \!P_{x,\nu}^{23} \!- \!s_{x,\nu}^{23} \! \!+ \!s_{x,\nu}^{31} \! \!+ \!P_{x,\nu}^{32} \!+ \!s_{x,\nu}^{33}}{2}\! \! +\! 1\!\right)\! \! \Bigg]. \label{eq:qcd_wgsc} \end{align} \normalsize It sums over the configurations of the auxiliary variables $m_{x,\nu}^{ab}$, $\overline{m}_{x,\nu}^{ab}$, which we introduced for the binomial decomposition (\ref{eq:qcd_binomial}) of the matrix elements $(2,1)$, $(2,3)$, $(3,1)$ and $(3,3)$: \begin{equation} \sum_{\{m,\overline{m}\}} \, = \, \prod_{x,\nu} \prod_{a =2,3} \prod_{b=1,3} \sum_{m_{x,\nu}^{ab} = 0}^{k_{x,\nu}^{ab}} \sum_{\overline{m}_{x,\nu}^{ab} = 0}^{\overline{k}_{x,\nu}^{ab}} \, . \end{equation} The admissible configurations of the dual variables $m_{x,\nu}^{ab}$ and $\overline{m}_{x,\nu}^{ab}$ must satisfy the constraint (\ref{eq:qcd_scconstraint3}), which is enforced at every link $(x,\nu)$ of the lattice by the product of Kronecker deltas in (\ref{eq:qcd_wgsc}). The three beta functions $\B$ in Eq.~(\ref{eq:qcd_wgsc}) are the result of the three $\theta_{x,\nu}^{(i)}$ integrals, $i=1,2,3$ in (\ref{eq:qcd_partitionsc}). They have the same structure as the beta functions we found for the pure gauge case (compare with (\ref{eq:qcd_wg})). Summarizing the discussion about the dual representation of strong coupling QCD, we found that the partition function is a sum over the configurations of the dual variables for fermions: $s_{x}^{a} = 0,1$ for the monomers and $k_{x,\nu}^{ab}, \overline{k}_{x,\nu}^{ab} = 0,1$ for the forward and backward hops respectively. The configurations of the fermion dual variables must completely fill the lattice, as imposed by the fermion constraint $C_{F}[s,k,\overline{k}]$ given in (\ref{eq:qcd_cf}). Hence, the admissible configurations are those for which every color layer of every site of the lattice is either occupied my a monomer, is the endpoint of a dimer or is run through by a loop. The fermion configurations come with the weight factor $W_{F}[s,k,\overline{k}]$ in (\ref{eq:qcd_wf}): monomers contribute with a factor $2m$, while loops come with the sign function $\sign (\mathcal{L})$ given in (\ref{eq:qcd_signl}). Only loops winding in the compactified temporal direction couple with the chemical potential $\mu$, and in Eq.~(\ref{eq:qcd_wf}) we were able to find a nice geometrical interpretation for the net-particle number as the total net-winding number of a loop configuration. Dimers and loops are further restricted by the gauge constraint $C_{G}[k,\overline{k}]$ in (\ref{eq:qcd_cf}). In the next section we will discuss the structure of the admissible configurations determined by the gauge constraint. Moreover, dimers and loops give contributions to the gauge weight $W_{G}[k,\overline{k}]$. The sign factor $(-1)^{K_{x,\nu}^{12} + K_{x,\nu}^{23} + K_{x,\nu}^{31} - j_{x,\nu}^{23} - j_{x,\nu}^{31}}$ in (\ref{eq:qcd_wgsc}) has a simple interpretation in the strong coupling limit, as we will see later. Before coming to the detailed discussion of the structure of the strong coupling configurations, we would like to compare the strong coupling results for the gauge integration with the ones we obtained in the last section for the pure gauge case. The gauge constraint (\ref{eq:qcd_cgsc}) in the strong coupling limit imposes the same relations between color link fluxes as the gauge constraint (\ref{eq:qcd_cg}) in the pure gauge theory. This similarity in the structure of the constraints reflects the fact that both systems are SU(3) symmetric. However, the variables that generate the color link fluxes are different in the two cases. In the pure gauge case the $J_{x,\nu}^{ab}$ are generated by the cycle occupation numbers $p_{x,\nu\rho}^{abcd} \in \mathbb{Z}$. Since the $p_{x,\nu\rho}^{abcd}$ live on the plaquettes $(x,\nu\rho)$ of the lattice, each $J_{x,\nu}^{ab}$ receives contributions from the cycle occupation numbers on the 6 plaquettes that are attached to the link $(x,\nu)$ (compare with the definition (\ref{eq:qcd_Jfluxes})). On the other hand, the fermion fluxes $K_{x,\nu}^{ab}$ are generated from the fermion dual variables $k_{x,\nu}^{ab}$ and $\overline{k}_{x,\nu}^{ab}$ for the forward and backward hops on the link $(x,\nu)$. Another main difference is that while the $J$-fluxes take values in $\mathbb{Z}$, the fermion fluxes $K_{x,\nu}^{ab} \in \{-1, 0 , 1\}$. This will make the interpretation of the gauge constraint (\ref{eq:qcd_cgsc}) together with the constraint (\ref{eq:qcd_scconstraint3}) for the auxiliary variables $m_{x,\nu}^{ab}, \overline{m}_{x,\nu}^{ab}$ much easier than in the pure gauge case. Comparing $W_{G}[p]$ of the pure gauge theory in Eq.~(\ref{eq:qcd_wg}) with the weight factor $W_{G}[k,\overline{k}]$ for strong coupling QCD in (\ref{eq:qcd_wgsc}) we also notice a structural similarity. Both are sums over configurations of the auxiliary variables $m_{x,\nu}^{ab}$ and $\overline{m}_{x,\nu}^{ab}$ needed for the binomial decomposition. In both cases the sum of the auxiliary fluxes $j_{x,\nu}^{ab} = m_{x,\nu}^{ab} - \overline{m}_{x,\nu}^{ab}$ is constrained by the value of the $K_{x,\nu}^{12}$ ($J_{x,\nu}^{12}$ for the gauge case) color flux at every link. Furthermore, the same sign factors appear in the summands of both weight factors. Because of the Pauli principle, the fluxes in the strong coupling case are restricted to the values 0,1 and -1, such that all binomial coefficients are equal to 1, while in the pure gauge theory weight $W_{G}[p]$ in Eq.~(\ref{eq:qcd_wg}) the binomial coefficients can have non-trivial values. The pure gauge theory weight $W_{G}[p]$ additionally have plaquette based weight factors from the expansion of the gauge action which also depends on the auxiliary plaquette variables $l_{x,\nu\rho}^{abcd}$. Clearly these terms are absent in strong coupling QCD where we have no gauge action. Nonetheless the weight factors that come from the Haar measure integration are the same in both cases, key signal of the fact that not only the constraints carry the information about the SU(3) symmetry in our dual representation. \subsection{Strong coupling baryon loops} In this section we focus on the interpretation of the constraints (\ref{eq:qcd_cgsc}) and (\ref{eq:qcd_scconstraint3}), and we discuss the consequences that they have on the admissible configurations of QCD at strong coupling. The gauge constraint $C_{G}[k,\overline{k}]$ imposes relations between different components of the color fluxes $K_{x,\nu}^{ab} \equiv k_{x,\nu}^{ab} - \overline{k}_{x,\nu}^{ab}$ we defined in (\ref{eq:qcd_shortsc}). Explicitly those constraints are: \begin{gather} \label{eq:qcd_scconstraint1} K_{x,\nu}^{11} + K_{x,\nu}^{12} = K_{x,\nu}^{23} + K_{x,\nu}^{33} \, ,\\ \label{eq:qcd_scconstraint2} K_{x,\nu}^{22} + K_{x,\nu}^{12} = K_{x,\nu}^{33} + K_{x,\nu}^{31} \, ,\\ \label{eq:qcd_scconstraint4} K_{x,\nu}^{13} + K_{x,\nu}^{12} = K_{x,\nu}^{31} + K_{x,\nu}^{21} \, ,\\ \label{eq:qcd_scconstraint5} K_{x,\nu}^{32} + K_{x,\nu}^{12} = K_{x,\nu}^{23} + K_{x,\nu}^{21} \, . \end{gather} Comparing Eqs.~(\ref{eq:qcd_scconstraint1}) -- (\ref{eq:qcd_scconstraint5}) with the constraints (\ref{eq:qcd_constraint1}) -- (\ref{eq:qcd_constraint5}) in Sec.~\ref{sec:qcd_acc} for the pure gauge theory we notice the structural similarity between the two sets of equations. As a consequence also (\ref{eq:qcd_scconstraint1}) -- (\ref{eq:qcd_scconstraint5}) can be organized into the three \textit{color flux conservation} constraints (\ref{eq:qcd_fluxconservation1}) -- (\ref{eq:qcd_fluxconservation3}) (plots in the top row of Fig.~\ref{fig:qcd_constraints}) and the three \textit{color flux exchange} constraints (\ref{eq:qcd_fluxexchange}) (plots at the bottom of Fig.~\ref{fig:qcd_constraints}) as we did in Sec.~\ref{sec:qcd_acc}, where we substitute the color fluxes $J_{x,\nu}^{ab} \in \mathbb{Z}$ of the pure gauge theory with the fermion color fluxes $K_{x,\nu}^{ab} \in \{-1,0,1\}$. The fact that the $K_{x,\nu}^{ab}$ fluxes take values from such a small set, together with the further restrictions imposed by the constraints (\ref{eq:qcd_scconstraint1}) -- (\ref{eq:qcd_scconstraint5}), makes the determination of the admissible configurations at strong coupling much easier than in the pure gauge case. Moreover, the additional constraint \begin{equation} \label{eq:qcd_scconstraint3b} K_{x,\nu}^{12} = j_{x,\nu}^{21} + j_{x,\nu}^{23} + j_{x,\nu}^{31} + j_{x,\nu}^{33} \, , \end{equation} which relates the fermion flux $K_{x,\nu}^{12}$ to the auxiliary fluxes $j^{ab}_{x,\nu} \equiv m^{ab}_{x,\nu} - \overline{m}^{ab}_{x,\nu}$ of the variables $m^{ab}_{x,\nu} \in \{0,k^{ab}_{x,\nu}\}$ and $\overline{m}^{ab}_{x,\nu} \in \{0,\overline{k}^{ab}_{x,\nu}\}$ will be crucial for the determination of the gauge sign \begin{equation} \label{eq:qcd_gaugesign} (-1)^{K_{x,\nu}^{12} + K_{x,\nu}^{23} + K_{x,\nu}^{31} - j_{x,\nu}^{23} - j_{x,\nu}^{31} } \, . \end{equation} The admissible combinations of the strong coupling fluxes $K_{x,\nu}^{ab}$ at a single link come in two types: three lines of flux that run in the same direction (see Fig.~\ref{fig:qcd_strongloops}), or six lines of flux that form a closed loop on a single link (Fig.~\ref{fig:qcd_linkloops}). \begin{figure} \centering \includegraphics[width=7.5cm]{chapters/plots/qcd_strongloops.pdf} \caption{Baryon loop elements in the strong coupling limit. Only the six combinations shown here are admissible for the propagation of fluxes in the strong coupling limit. The elements with an odd number of color flux crossings come with an explicit minus sign. For the negative direction the same fluxes are admissible and have the same signs. The corresponding diagrams are obtained by reverting the arrows. \label{fig:qcd_strongloops}} \end{figure} \begin{figure} \centering \includegraphics[width=9.5cm]{chapters/plots/qcd_linkloops.pdf} \caption{Closed, non-propagating one-link loops at strong coupling. All of these loops come with a positive weight. Also the opposite orientation is possible, which is obtained by reverting all arrows. \label{fig:qcd_linkloops}} \end{figure} Obviously only the first type allows for long distance propagation and we refer to these strong coupling elements as \textit{strong coupling baryon fluxes}. The locally closing ones are referred to as \textit{one-link loops}. For the discussion of the complete list of strong coupling baryon fluxes we start with solutions of the constraint equations (\ref{eq:qcd_scconstraint1}) -- (\ref{eq:qcd_scconstraint5}) where we allow only the values $K_{x,\nu}^{ab}= 0,1$, i.e., we consider forward propagation. In addition to the gauge constraint $C_{G}[k,\overline{k}]$ also the fermion constraint $C_{F}[k,\overline{k}]$ has to be obeyed. This implies that, if we consider the link $(x,\nu)$, at site $x$ only a single line of flux can originate from each color $a$, and the three lines must end on different colors at site $x + \hat{\nu}$. For the forward propagation of the strong coupling baryon fluxes one finds exactly six solutions, which we represent in Fig.~\ref{fig:qcd_strongloops}. The six solutions for the backward propagation of the strong coupling baryon fluxes are obtained reverting the arrows in Fig.~\ref{fig:qcd_strongloops}, which corresponds to $K_{x,\nu}^{ab} \rightarrow - K_{x,\nu}^{ab}$. The gauge sign (\ref{eq:qcd_gaugesign}) for the strong coupling baryon fluxes in Fig.~\ref{fig:qcd_strongloops} is easily determined. We here discuss two examples to illustrate the method. The strong coupling baryon flux at the top left of Fig.~\ref{fig:qcd_strongloops} is built by setting $K_{x,\nu}^{11} = K_{x,\nu}^{22} = K_{x,\nu}^{33} = 1$. Therefore, for this example the fermion fluxes $K_{x,\nu}^{12}$, $K_{x,\nu}^{23}$ and $K_{x,\nu}^{31}$ that appear at the exponent of the sign factor (\ref{eq:qcd_gaugesign}) are vanishing. Moreover, since $K_{x,\nu}^{23} = K_{x,\nu}^{31} = 0$, also $j_{x,\nu}^{23} = j_{x,\nu}^{31} = 0$ (recall that $j_{x,\nu}^{23} = m_{x,\nu}^{ab} - \overline{m}_{x,\nu}^{ab}$ and that $m_{x,\nu}^{ab} = 0, k_{x,\nu}^{ab}$ and $\overline{m}_{x,\nu}^{ab} = 0, \overline{k}_{x,\nu}^{ab}$, with $(a,b) = (2,1), (2,3), (3,1), (3,3)$). Hence, the strong coupling baryon flux at the top left of Fig.~\ref{fig:qcd_strongloops} has positive gauge sign. Let us now consider the top center example of Fig.~\ref{fig:qcd_strongloops}. This strong coupling baryon flux has non-vanishing fluxes $K_{x,\nu}^{12} = K_{x,\nu}^{23} = K_{x,\nu}^{31} = 1$, which means that the sum $K_{x,\nu}^{12} + K_{x,\nu}^{23} + K_{x,\nu}^{31}$ at the exponent of the sign factor (\ref{eq:qcd_gaugesign}) is odd. However, the constraint (\ref{eq:qcd_scconstraint3b}) forces the sum of the auxiliary fluxes $j_{x,\nu}^{ab}$ to be equal to $K_{x,\nu}^{12} = 1$. Since $K_{x,\nu}^{21}$ and $K_{x,\nu}^{33}$ are vanishing, $j_{x,\nu}^{21} = j_{x,\nu}^{33} = 0$ and the constraint (\ref{eq:qcd_scconstraint3b}) is satisfied only if $j_{x,\nu}^{23} + j_{x,\nu}^{31} = 1$. As a result, the gauge sign for the top center example of Fig.~\ref{fig:qcd_strongloops} is also positive. The gauge signs for all the other strong coupling baryon fluxes in Fig.~\ref{fig:qcd_strongloops} can be determined in a similar way. In particular we find: \begin{equation} \label{eq:qcd_gaugesignsc} (-1)^{K_{x,\nu}^{12} + K_{x,\nu}^{23} + K_{x,\nu}^{31} - j_{x,\nu}^{23} - j_{x,\nu}^{31} } = (-1)^{\text{\# crossings of K-flux}}\, . \end{equation} In other words, at strong coupling the exponent of the gauge sign in (\ref{eq:qcd_gaugesign}) has the simple interpretation of the total number of $K$-flux crossings. In Fig.~\ref{fig:qcd_strongloops} we marked the strong coupling baryon fluxes where the sign (\ref{eq:qcd_gaugesignsc}) is negative with $-1$. The second class of solutions of (\ref{eq:qcd_scconstraint1}) -- (\ref{eq:qcd_scconstraint5}) are the one-link loops depicted in Fig.~\ref{fig:qcd_linkloops}. They are obtained by allowing all values $K_{x,\nu}^{ab} = -1,0,+1$ and enforcing the fermion constraints $C_{F}[s,k,\overline{k}]$, such that each node is run through by a loop. It is easy to see that the six solutions in Fig.~\ref{fig:qcd_linkloops} are obtained by combining one of the forward baryon fluxes from Fig.~\ref{fig:qcd_strongloops} with a matching backward baryon flux such that the fermion constraints are obeyed. One finds that only fluxes with the same sign in Fig.~\ref{fig:qcd_strongloops} can be combined among each other, such that the total sign from (\ref{eq:qcd_gaugesign}) is always +1. Recalling the definition (\ref{eq:qcd_signl}) of $\sign (\mathcal{L})$, we find that also the fermion loop sign is positive for the six one-link loops in Fig.~\ref{fig:qcd_linkloops}. In fact, for each of them there is an overall minus sign and then a factor $(-1)^{3}$ for the three forward hops. Thus one-link loops always come with a positive weight which is given by the products of beta functions in (\ref{eq:qcd_wgsc}). These weights can be summed, and all possible one-link loops may be combined into a dual element that plays a similar role as the monomers and dimers: they are all local fermionic monomials that can be used to saturate the fermion constraints on the sites and links that are not occupied by a strong coupling baryon loop. We now demonstrate that the interplay between the gauge sign (\ref{eq:qcd_gaugesignsc}) for the strong coupling baryon fluxes in Fig.~\ref{fig:qcd_strongloops} and the fermion loop sign makes the overall sign of baryon loops equivalent to the fermion loop sign (\ref{eq:qcd_signl}) itself. A general baryon loop is a closed non-intersecting path on the lattice made out of the strong coupling baryon fluxes represented in Fig.~\ref{fig:qcd_strongloops}. Let us start considering the simple strong coupling baryon loop shown in the top plot of Fig.~\ref{fig:qcd_loop}: a baryon loop made out of just parallel fluxes closing around a single plaquette. For this loop the gauge sign is evidently positive because there are no crossings of $K$-fluxes. Moreover, since the baryon loop is made out of three identical fermion loops $\mathcal{L}$ running parallel on the three different color layers, the fermion sign for the baryon loop can be determined as the third power of the sign function $\sign (\mathcal{L})$ of a single loop: \begin{equation} \big(\sign (\mathcal{L})\big)^3 = \sign (\mathcal{L}) \, . \label{eq:qcd_signb} \end{equation} Hence, we proved that for this simple example the baryon loops again obey the sign formula (\ref{eq:qcd_signl}) for staggered fermions. We then replace one of the parallel fluxes of the baryon loop with a strong coupling flux that has just one crossing (see the middle example in Fig.~\ref{fig:qcd_loop}). The resulting baryon loop have an overall negative gauge sign. However, the crossing also changes the connectivity properties of the loop: inserting one crossing either connects two fermion loops into one, as in the example in the middle of Fig.~\ref{fig:qcd_loop}, or splits a loop into two loops. Thus every crossing changes the number of loops by one, and since every loop comes with an overall minus sign, the crossing also changes the fermion sign. However, the overall sign of the baryon loop remains the same, because the change in the fermion sign is compensated by the change of the gauge sign. Consequently the sign of a baryon loop is always given by (\ref{eq:qcd_signb}), i.e., the signs of the strong coupling baryon loops are equivalent to the signs for loops of a single free staggered fermion. In our dual representation it is possible to bring the identification of the strong coupling baryon loops with the loops of a free staggered fermion a step further: when computing the link weight factor for the six strong coupling flux elements in Fig.~\ref{fig:qcd_strongloops}, one finds that they all come with the same weight of $1/12$. Thus for every link of the loop we can sum over all six possible fluxes and obtain a total link weight of $1/2$. The resulting weight for a baryon loop $\mathcal{L}_{B}$ is then given by: \begin{equation} \label{eq:qcd_wb} W(\mathcal{L}_{B}) = \sign (\mathcal{L}_{B}) \left(\dfrac{1}{2}\right)^{|\mathcal{L}_{B}|} \, , \end{equation} where $|\mathcal{L}_{B}|$ denotes the length of the loop $\mathcal{L}_{B}$. We conclude that the dual form of strong coupling QCD is a gas of free staggered fermion loops that come with the weight $W(\mathcal{L}_{B})$. These loops describe baryons and are embedded in a background of monomers, dimers and local link loops, such that the fermion constraints are obeyed at all sites of the lattice. We think that the interpretation of the strong coupling baryon loops as free staggered fermions could open the possibility of updating our form of strong coupling QCD with fermion bags \cite{PhysRevD.82.025007}. This idea has also led to a new reformulation of strong coupling QCD in terms of baryon bags, presented in \cite{Gattringer:2018mrg}. \begin{figure} \begin{center} \includegraphics[width=7cm]{chapters/plots/qcd_loop.pdf} \end{center} \caption{Examples of simple strong coupling baryon loops with different connectivity properties. \label{fig:qcd_loop}} \end{figure} \section{Worldlines and worldsheets representation of the full theory \label{sec:qcd_full}} We complete the presentation of the dual representation of QCD discussing the formulation in terms of worldsheets and worldlines of the full partition function. The partition sum of full QCD can be written as \begin{equation} \label{eq:qcd_partitionfull} Z = \int \! D[U] Z_{F} [U] e^{-S_{G}[U]} \, . \end{equation} In other words, we integrate the fermionic partition sum $Z_{F} [U]$ given in (\ref{eq:qcd_partitionf}) over the Haar measure $\int \! D[U]$, weighted with the gauge Boltzmann factor $e^{-S_{G}[U]}$, where $S_{G}[U]$ is the Wilson action (\ref{eq:qcd_sg}). We can therefore use the intermediate result (\ref{eq:qcd_partitionsc}) we obtained in the last section for the partition sum at strong coupling, and simply multiply the integrand by the Boltzmann weight $e^{-S_{G}[U]}$. What we are left with is the following gauge integral \begin{equation} \label{eq:qcd_gaugeintegral} \Bigg[\prod_{x,\nu} \int \! dU_{x,\nu} \Bigg] e^{-S_{G}[U]} \, \prod_{x, \nu} \prod_{a, b} \left( U_{x,\nu}^{ab} \right)^{k_{x,\nu}^{ab}} \left( U_{x,\nu}^{ab \, \star} \right)^{\overline{k}_{x,\nu}^{ab}} \, . \end{equation} We can now treat the Boltzmann weight $e^{-S_{G}[U]}$ as we did in Sec.~\ref{sec:qcd_acc}: we decompose the action into a sum over abelian color cycles as in Eq.~(\ref{eq:qcd_accaction}), and then we factorize and Taylor-expand the Boltzmann weight. As a result the integrand in (\ref{eq:qcd_gaugeintegral}) takes the following form: \begin{equation} \label{eq:qcd_gaugeintegral2} \Bigg[\prod_{x,\nu} \int \! dU_{x,\nu} \Bigg] \prod_{x, \nu} \prod_{a, b} \left( U_{x,\nu}^{ab} \right)^{N_{x,\nu}^{ab} + k_{x,\nu}^{ab}} \left( U_{x,\nu}^{ab \, \star} \right)^{\overline{N}_{x,\nu}^{ab} + \overline{k}_{x,\nu}^{ab}} \, . \end{equation} Eq.~(\ref{eq:qcd_gaugeintegral2}) can be solved in closed form following exactly the same steps we used in Sec.~\ref{sec:qcd_acc}, just replacing $N_{x,\nu}^{ab}$ with $N_{x,\nu}^{ab} + k_{x,\nu}^{ab}$ and $\overline{N}_{x,\nu}^{ab}$ with $\overline{N}_{x,\nu}^{ab} + \overline{k}_{x,\nu}^{ab}$. Again we rewrite \begin{equation} N_{x,\nu}^{ab} + k_{x,\nu}^{ab} = \dfrac{Q_{x,\nu}^{ab} + L_{x,\nu}^{ab}}{2} \, , \qquad \overline{N}_{x,\nu}^{ab} + \overline{k}_{x,\nu}^{ab} = \dfrac{Q_{x,\nu}^{ab} - L_{x,\nu}^{ab}}{2} \, , \end{equation} where \begin{equation} L_{x,\nu}^{ab} = J_{x,\nu}^{ab} + K_{x,\nu}^{ab} \, , \qquad Q_{x,\nu}^{ab} = S_{x,\nu}^{ab} + P_{x,\nu}^{ab} \, . \end{equation} $J_{x,\nu}^{ab}$ and $S_{x,\nu}^{ab}$ are the linear combinations of the cycle occupation numbers $p_{x,\nu\rho}^{abcd}$ given in (\ref{eq:qcd_Jfluxes}) and (\ref{eq:qcd_Sfluxes}), while $K_{x,\nu}^{ab}$ and $P_{x,\nu}^{ab}$ are the combinations of the fermion dual variables $k_{x,\nu}^{ab}$ and $\overline{k}_{x,\nu}^{ab}$ given in (\ref{eq:qcd_shortsc}). Putting things together we obtain the following dual form for the partition sum of full QCD: \begin{equation} \label{eq:qcd_partitionfull2} Z \; = \; \sum_{\{p,s,k,\overline{k}\}} C_{F}[s,k,\overline{k}] \; W_{F}[s,k,\overline{k}] \; C_{G}[p,k,\overline{k}] \; W_{G}[p,k,\overline{k}] \, . \end{equation} $C_{F}[s,k,\overline{k}]$ and $W_{F}[s,k,\overline{k}]$ are the fermion constraint (\ref{eq:qcd_cf}) and the fermion weight (\ref{eq:qcd_wf}) we discussed in the previous section. The gauge constraint $C_{G}[p,k,\overline{k}]$ and the gauge weight $W_{G}[p,k,\overline{k}]$ originate from the gauge integrals in (\ref{eq:qcd_gaugeintegral2}). They therefore have the same structure as the gauge constraints and weights we obtained in the pure gauge case and the strong coupling limit. In particular, \begin{align} C_G[p,k,\overline{k}] \; =& \; \prod_{x,\nu} \delta(L_{x,\nu}^{11} + L_{x,\nu}^{12} - L_{x,\nu}^{33} - L_{x,\nu}^{23}) \; \delta(L_{x,\nu}^{22} + L_{x,\nu}^{12} - L_{x,\nu}^{33} - L_{x,\nu}^{31}) \nonumber \\ & \hspace{3mm}\times \delta(L_{x,\nu}^{13} + L_{x,\nu}^{12} - L_{x,\nu}^{31} - L_{x,\nu}^{21}) \; \delta(L_{x,\nu}^{32} + L_{x,\nu}^{12} - L_{x,\nu}^{23} - L_{x,\nu}^{21}) \; , \label{eq:qcd_cgfull} \end{align} is again a product of four Kronecker deltas which impose relations between different color components of the link-based fluxes $L_{x,\nu}^{ab}$. In full QCD these fluxes take contributions both from the gauge degrees of freedom, represented by the cycle occupation numbers $p_{x,\nu\rho}^{abcd}$ attached to the link $(x,\nu)$, and the fermionic degrees of freedom, in the form of the hopping variables $k_{x,\nu}^{ab}$ and $\overline{k}_{x,\nu}^{ab}$. The weight factor \small \begin{align} &W_{G} [p,k,\overline{k}] = 2^{4V}\!\! \sum_{\{l,m,\overline{m}\}} \! \Bigg[\prod_{x,\nu} \delta(L_{x,\nu}^{12} - j_{x,\nu}^{21}\! - j_{x,\nu}^{23} \!- j_{x,\nu}^{31}\! - j_{x,\nu}^{33})\Bigg] \! \Bigg[\prod_{x,\nu} (-1)^{L_{x,\nu}^{12} + L_{x,\nu}^{23} + L_{x,\nu}^{31} - j_{x,\nu}^{23} - j_{x,\nu}^{31}} \Bigg] \nonumber \\ &\hspace{18mm} \times \Bigg[ \prod_{x,\nu} \prod_{a=2,3} \prod_{b=1,3} \binom{N_{x,\nu}^{ab} + k_{x,\nu}^{ab}}{m_{x,\nu}^{ab}} \binom{\overline{N}_{x,\nu}^{ab} + \overline{k}_{x,\nu}^{ab}}{\overline{m}_{x,\nu}^{ab}}\Bigg] \Bigg[\prod_{x,\nu<\rho} \prod_{a,b,c,d} \dfrac{ \left( \beta/2 \right)^{|p_{x,\nu\rho}^{abcd}| + 2l_{x,\nu\rho}^{abcd}}}{\left(|p_{x,\nu\rho}^{abcd}| + l_{x,\nu\rho}^{abcd}\right)!l_{x,\nu\rho}^{abcd}!}\Bigg] \nonumber \\ &\times \Bigg[ \prod_{x,\nu} \B\left(\dfrac{Q_{x,\nu}^{11} + Q_{x,\nu}^{13} + Q_{x,\nu}^{22} + Q_{x,\nu}^{32}}{2} + 2\right.\left|\dfrac{Q_{x,\nu}^{12} + s_{x,\nu}^{21} +s_{x,\nu}^{23} + s_{x,\nu}^{31} + s_{x,\nu}^{33}}{2} + 1\right) \nonumber \\ &\!\times\!\! \B\!\left(\!\!\dfrac{Q_{x,\nu}^{11} \! + \! s^{21}_{x,\rho}\! \! + \! Q_{x,\nu}^{23}\! -\! s_{x,\nu}^{23}\!\! + s_{x,\nu}^{31}\!\! + \!Q_{x,\nu}^{33}\! - \!s_{x,\nu}^{33}}{2} \!\!+\! 1\right.\!\left|\dfrac{\!Q_{x,\nu}^{13}\! +\! Q_{x,\nu}^{21}\! - \!s_{x,\nu}^{21}\!\! +\! s_{x,\nu}^{23}\! \!+\! Q_{x,\nu}^{31}\! -\! s_{x,\nu}^{31}\!\! +\! s_{x,\nu}^{33}}{2}\!\! + \!1\!\right) \nonumber \\ &\!\times\! \!\B\!\left(\!\!\dfrac{s_{x,\nu}^{21} \!\!+ \!Q^{22}_{x,\rho}\! + \!s_{x,\nu}^{23} \!\!+\! Q_{x,\nu}^{31}\! -\! s_{x,\nu}^{31}\!\! +\! Q_{x,\nu}^{33}\! -\! s_{x,\nu}^{33}}{2}\!\! +\! 1\right.\!\left|\dfrac{\!Q_{x,\nu}^{21} \!- \!s_{x,\nu}^{21} \!\!+ \!Q_{x,\nu}^{23} \!- \!s_{x,\nu}^{23} \!\!+ \!s_{x,\nu}^{31} \!\!+ \!Q_{x,\nu}^{32} \!+ \!s_{x,\nu}^{33}}{2}\!\! +\! 1\!\right)\!\!\Bigg], \label{eq:qcd_wgfull} \end{align} \normalsize contains the additional constraint \begin{equation} L_{x,\nu}^{12} = j_{x,\nu}^{21} + j_{x,\nu}^{23} + j_{x,\nu}^{31} + j_{x,\nu}^{33} \, , \label{eq:qcd_fullconstraint3} \end{equation} which relates the link flux $L_{x,\nu}^{12}$ with the fluxes $j_{x,\nu}^{ab} \equiv m_{x,\nu}^{ab} - \overline{m}_{x,\nu}^{ab}$ of the auxiliary variables $m_{x,\nu}^{ab}$ and $\overline{m}_{x,\nu}^{ab}$ we introduced for the binomial decomposition of the matrix elements $(2,1)$, $(2,3)$, $(3,1)$ and $(3,3)$. In full QCD $W_{G}[p,k,\overline{k}]$ sums over the configurations of the auxiliary plaquettes variables $l_{x,\nu\rho}^{abcd} \in \mathbb{N}_{0}$ as well as the link-based auxiliary variables $m_{x,\nu}^{ab} \in \{0,1\dots N_{x,\nu}^{ab} + k_{x,\nu}^{ab}\}$ and $\overline{m}_{x,\nu}^{ab} \in \{0,1\dots \overline{N}_{x,\nu}^{ab} + \overline{k}_{x,\nu}^{ab}\}$: \begin{equation} \sum_{\{l,m,\overline{m}\}} = \Bigg[\prod_{x,\nu<\rho} \prod_{a,b,c,d} \sum_{l_{x,\nu\rho}^{abcd} = 0}^{\infty}\Bigg] \Bigg[\prod_{x,\nu} \prod_{a=2,3} \prod_{b=1,3} \sum_{m_{x,\nu}^{ab} = 0}^{N_{x,\nu}^{ab} + k_{x,\nu}^{ab}} \sum_{\overline{m}_{x,\nu}^{ab} = 0}^{\overline{N}_{x,\nu}^{ab} + \overline{k}_{x,\nu}^{ab}}\Bigg] \, . \end{equation} The beta functions and the sign factor resulting from the Haar measure integration have the same dependence on the component of the link fluxes which, also for $W_{G}[p,k,\overline{k}]$ depend on both the cycle occupation numbers and the fermion dual variables. Moreover, $W_{G}[p,k,\overline{k}]$ collects the coefficients from the Taylor expansion of the local exponentials (\ref{eq:qcd_partition2}), as well as the binomial factors from the binomial decomposition (\ref{eq:qcd_binomial}). Summarizing, in the dual formulation the partition function is a sum over the dual variables for fermions $s_{x}^{a},k_{x,\nu}^{ab},\overline{k}_{x,\nu}^{ab} \in \{0,1\}$ and the cycle occupation numbers $p_{x,\nu\rho}^{abcd} \in \mathbb{Z}$ which represent the gauge degrees of freedom. The configurations of the fermion dual variables must satisfy the fermion constraint $C_{F}[s,k,\overline{k}]$ given in (\ref{eq:qcd_cf}), which enforces Pauli's exclusion principle. As a result, the admissible fermion configurations must saturate the lattice with monomers dimers and loops. Each fermion configuration comes with a weight given by $W_{F}[s,k,\overline{k}]$ in (\ref{eq:qcd_wf}): monomers contribute a factor of $2m$, loops come with a sign $\sign (\mathcal{L})$ given in (\ref{eq:qcd_signl}) and the chemical potential couples to the temporal winding number $W_{\mathcal{L}}$ of the loops. Dimers and loops, together with the gauge degrees of freedom, enter in the expressions of the gauge constraint $C_{G}[p,k,\overline{k}]$ and the gauge weight $W_{G}[p,k,\overline{k}]$ that we just discussed. Structurally these are the same constraints and weights as for pure gauge theory and strong coupling QCD, since they are generated by integrating the SU(3) link matrices. However, in full QCD they link the color flux contributions from both the gauge fields, via the cycle occupation numbers $p_{x,\nu\rho}^{abcd}$, and the fermion loops, via $k_{x,\nu}^{ab}$ and $\overline{k}_{x,\nu}^{ab}$. We conclude this section on full QCD with addressing two important aspects of the new representation: as in the case of pure SU(3) lattice gauge theory, our dual form of the partition sum (\ref{eq:qcd_partitionfull2}) has the structure of a strong coupling expansion, and again, our approach allows one to compute all coefficients of this expansion in closed form. Furthermore, it is obvious how to generalize the construction to several flavors: one simply uses multiple sets of dual fermion variables, which all couple in the same way to the gauge fields. Thus instead of the variables $k_{x,\nu}^{ab}$ and $\overline{k}_{x,\nu}^{ab}$ one has flavor sums over such variables, and the color fluxes at each link have contributions from all flavors. These flavor sums over the dual fermion variables enter the constraints and weights, which otherwise have the same form as presented in this section. \chapter{Summary \label{cha:conclusions}} In this thesis we developed two methods, the \textit{abelian color cycle} (ACC) and \textit{abelian color flux} (ACF) methods, with the aim of extending the applicability of the dual approach to non-abelian lattice field theories. We achieved this by making explicit the sums over the color indices in the action of the model we were considering. As a result, the action is decomposed into a sum over commuting terms, and thus the Boltzmann weight may be factorized into local factors. Then the dualization program can be carried out as in the abelian case, by expanding the single exponentials and integrating out the conventional fields. As a result the partition function is exactly rewritten in terms of new variables, the so-called dual variables, that are integer valued variables attached either to the links or the plaquettes of the lattice. The final partition function is a sum over the configurations of those dual variables, which come with weights and must satisfy constraints. The emergence of the constraints comes from the integration over the phases in the parametrization of the conventional fields. In general, those constraints impose flux conservation for the dual variables which imply that the long range physics is described by worldlines for the matter fields and worldsheets for the gauge degrees of freedom. When integrating fermion fields we see the emergence of additional fermionic constraints which implement Pauli's exclusion principle for the fermionic dual variables. As a first example for the application of the ACF approach, in Chapter \ref{cha:pcm} we presented the dualization of the SU(2) principal chiral model with coupled chemical potentials. In the dual formulation this model consists of two species of worldlines that are constrained to form closed loops, and two additional species of auxiliary variables which are unconstrained. As we outlined several times, the constraints implement the original symmetry of a theory in the dual representation. However, this is not the only way the symmetry is represented in the dual formulation. Also key are the types of dual variables, and their role in the final form of the partition function, as well as the form of the weight factors. In fact, in the principal chiral model the two constraints for the two species of flux variables are the same as the constraint for the flux variable we found in Sec.~\ref{sec:teo_u1} for the case of the U(1) Gauge -- Higgs model. Hence, the constraints alone would hint at a $U(1) \times U(1)$ symmetry. Then the auxiliary variables as well as the structure of the weight tie together the two species of worldlines and give the full SU(2) symmetry in the dual representation. Our worldline representation of the dual partition function of the SU(2) principal chiral model sums over the admissible configurations of the flux variables and the auxiliary variables which come with real and positive weights. Hence, in this particular case, the ACF method solves the sign problem that this model has in the conventional representation. Moreover, thanks to the simple structure of the constraints, in a second instant we were able to reformulate the model in terms of yet another set of dual variables which automatically fulfill the flux conservation constraints, thus obtaining a complete dualization of the partition function in the sense of Kramers and Wannier. In Chapters \ref{cha:su2} and \ref{cha:qcd} we derived our dual formulation for QC$_{2}$D and QCD respectively. In both cases we started with the ACC dualization of the pure gauge theories. The minimal terms the gauge actions are decomposed into are the so-called abelian color cycles, which we interpreted as paths in color space closing around plaquettes. In SU(2) there are a total of 16 different ACCs for each plaquette of the lattice, which can only have positive orientation. For SU(3) the number of the ACC increases to 81 for each plaquette, because we have 3 color layers instead of 2. Furthermore, they can have both positive and negative orientation. In the dualization those abelian color cycles are in one to one correspondence with a set of dual variables called cycle occupation numbers, which are positive integers for SU(2) and integer valued variables for SU(3). For SU(3) we additionally have plaquette auxiliary variables, which result from having to take the real part in the Wilson gauge action. Our dualization process factorizes the integrals over the link variables, thus allowing us to perform the Haar measure integrals analytically. The Haar measure integrations give rise to gauge constraints and gauge weight factors. The gauge constraints arise when integrating over the phases of the parametrization chosen for the conventional gauge fields, hence we find two constraints for the SU(2) cycle occupation numbers, and five constraints for the SU(3) cycle occupation numbers. Those constraints impose relations between different color components of the link fluxes generated by the cycle occupation numbers attached to that link. As a result we find that the admissible configurations which contribute to the long range physics are worldsheets of cycle occupation numbers. The weight factors collect the combinatorial factors arising from the integrals over the angles of the parametrization of the elements of the group, as well as the power series in the inverse gauge coupling which come from the Taylor expansion of the Boltzmann weight. Hence, our worldsheet representations for pure gauge theories correspond to a strong coupling series, where all terms are known in closed form. We stress at this point that the gauge weight factors contain explicit signs which originate from the negative signs in the parametrization of the group elements. To add the matter fields we consider staggered fermions. The minimal terms the fermionic actions are decomposed into are the so-called abelian color fluxes. The dual variables for fermions are then monomer variables, and forward and backward hopping variables which we illustrated in Fig.~\ref{fig:su2_grassmann} and Fig.~\ref{fig:qcd_grassmann}. The Grassmann integrals give rise to a fermion constraints which enforce Pauli's exclusion principle for the dual variables, and a fermion weight. The fermion constraints imply that the admissible fermion configurations are the ones that completely fill the lattice with monomers, dimers and loops. Loops come with a sign factor. Moreover, when considering non-vanishing values of the quark chemical potential, in the dual representation the chemical potential only couples with loops winding in the temporal direction. We thus can identify the net-particle number as the total net temporal winding number of the worldline configuration, which from a geometrical point of view is a very elegant representation. In fact, the identification of conserved quantities as topological invariant, not only makes the determination of such quantities easier in the dual representation, but also opens the possibility of performing simulations in the canonical ensemble. Finally, the Haar measure integration leads to the same constraints and weights discussed for the pure gauge case, with the difference that now the link fluxes receive contributions not only from the cycle occupation numbers, but also from the forward and backward hopping terms for the fermions. Obviously the appearance of gauge and fermion sign factors in our dual formulations implies that for a Monte Carlo simulation of the ACC and ACF dual representations a strategy for a partial resummation needs to be found. Nevertheless, for SU(2) we find that, when considering a coupled strong coupling hopping expansion, negative terms only appear at $\mathcal{O}\big(\beta^{4}\big)$ or $\mathcal{O}\big((\frac{1}{2m})^{4} \beta^{3} \big)$ and the leading orders are free of the sign problem. For SU(3) we discussed the strong coupling limit, and found that in our worldline representation strong coupling QCD is a gas of free staggered fermion loops, which describe baryons, that are embedded in a background of monomers, dimers and local link loops. We think that the interpretation of the strong coupling baryon loops as free staggered fermions could open the possibility of updating our form of strong coupling QCD with the fermion bags approach. Summarizing, we developed two methods that can be employed for the dualization of non-abelian lattice field theories. Unfortunately, the application of those methods to QCD leads to a worldlines and worldsheets representation that has a sign problem. However, both the ACC and the ACF dualization methods are very general and could be successful in overcoming the sign problem of other non-abelian field theories, as in the case of the SU(2) principal chiral model. \chapter{The SU(2) principal chiral model \label{cha:pcm}} In this chapter we present the application of the abelian color flux (ACF) approach to the SU(2) principal chiral model. The study of this model is motivated by its analogies with non-abelian field theories, like asymptotic freedom and dynamic mass generation. Also, this model can be used as a chiral effective model for low energy QCD with isospin chemical potential \cite{Rindlisbacher:2015xku}. However, here our interest lies on more theoretical aspects. We want to demonstrate the applicability of the abelian color flux (ACF) approach to general non-abelian theories, other than fermion theories with a gauge background. Furthermore, we want to address the question of how the symmetry of the conventional representation manifests itself in the dual worldline representation. In doing so we figured out that the ACF approach easily solves the complex action problem that this model has in the conventional representation when chemical potentials are coupled to some of the conserved currents. Furthermore, we were able to find yet another dual representation, as we will discuss in the last section of this chapter. \section{The continuum theory \label{sec:pcm_continuum}} The Euclidean continuum action of the SU(2) principal chiral model in $d$ dimensions is given by \begin{equation} \label{eq:pcm_continuumaction} S = \dfrac{J}{2} \int \! d^{d}x \, \Tr \left[ ( \partial_{\nu} U(x) )^{\dagger} (\partial_{\nu} U(x))\right] \, , \end{equation} where $J$ is the coupling and $U(x)\in$ SU(2) are the dynamical degrees of freedom. The labels $\nu = 1,2,\dots,d$ are space-time indices which are summed over in the expression for the action (\ref{eq:pcm_continuumaction}). The model enjoys the global symmetry $\text{SU(2)}_{L} \times \text{SU(2)}_{R}$ which corresponds to the left and right multiplications of $U(x)$ by arbitrary constant SU(2) matrices \begin{equation} \label{eq:pcm_symmetries} U(x) \, \rightarrow \, V_{L} \, U(x) \, V^{\dagger}_{R} \, , \qquad \qquad V_{L}, V_{R} \in \text{SU(2)} \, . \end{equation} Let us now consider one of the symmetries (\ref{eq:pcm_symmetries}) written in the infinitesimal form \begin{equation} \label{eq:pcm_infinitesimaltransformation} U(x) \, \rightarrow \, U^{\prime}(x) \, = \, U(x) + \alpha \, \Delta U(x) \, , \end{equation} where $\alpha$ is an infinitesimal parameter and $\Delta U(x)$ is the linearized deformation of the field $U(x)$ at order $\alpha$. Noether's theorem implies the existence of the conserved currents \begin{equation} \label{eq:pcm_conservedcurrents} j_{\nu} (x) = J \, \Tr \left[(\partial_{\nu} U^{\dagger} (x)) \Delta U(x)\right] \, , \qquad \text{with } \, \partial_{\nu} j_{\nu} (x) = 0 \, . \end{equation} The corresponding conserved charge then is \begin{equation} \label{eq:pcm_conservedcharge} Q = \int \!d^{d-1}x \, j_{d}(x) = J \int \!d^{d-1}x \Tr \left[(\partial_{d} U^{\dagger} (x)) \Delta U(x)\right] \, , \end{equation} where the integration runs over the $d-1$ dimensional space and $d$ denotes the temporal direction. In the Euclidean formalism this charge can be coupled to a chemical potential $\mu$ as follows \begin{equation} \label{eq:pcm_coupling} S_{\mu} = S + i \mu \beta Q = J \int d^{d} x \left\{ \frac{1}{2} \Tr \left[ ( \partial_{\nu} U(x) )^{\dagger} (\partial_{\nu} U(x))\right] + i \mu \Tr \left[\partial_{d} U^{\dagger} (x) \Delta U(x)\right] \right\} \, , \end{equation} where in the second step we used the fact that the size of the compactified time direction is equivalent to the inverse temperature: $\int d x_{4} = \beta$. For our purposes we consider two of the symmetries (\ref{eq:pcm_symmetries}), \begin{equation} \label{eq:pcm_symmetries12} U(x) \, \rightarrow \, e^{i \frac{\alpha_{1}}{2} \sigma_{3}} \, U(x) \, e^{i \frac{\alpha_{1}}{2} \sigma_{3}} \, , \qquad U(x) \, \rightarrow \, e^{i \frac{\alpha_{2}}{2} \sigma_{3}} \, U(x) \, e^{-i \frac{\alpha_{2}}{2} \sigma_{3}} \, , \end{equation} where $\alpha_{1} \in \mathbb{R}$ and $\alpha_{2} \in \mathbb{R}$ are independent parameters and $\sigma_{3}$ denotes the third Pauli matrix. Following the discussion above we can compute the Noether charges corresponding to these two symmetries and obtain \begin{align} \nonumber &Q_{1} = \dfrac{i J}{4} \int \!d^{d-1}x \Tr \left[(\partial_{d} U^{\dagger} (x)) [ \sigma_{3} U(x) + U(x) \sigma_{3} ] - U^{\dagger} (x) [ \sigma_{3} (\partial_{d} U(x)) + (\partial_{d} U(x)) \sigma_{3} ]\right] \, ,\\ &Q_{2} = \dfrac{i J}{4} \int \!d^{d-1}x \Tr \left[(\partial_{d} U^{\dagger} (x)) [ \sigma_{3} U(x) - U(x) \sigma_{3} ] - U^{\dagger} (x) [ \sigma_{3} (\partial_{d} U(x)) - (\partial_{d} U(x)) \sigma_{3} ]\right] \, . \label{eq:pcm_charges12} \end{align} In the next section we will obtain the lattice version of the principal chiral model with chemical potentials $\mu_{1}$ and $\mu_{2}$ coupled to the two charges. \section{Lattice discretization \label{sec:pcm_lattice}} The lattice discretization is carried out following the standard procedure, as described in Section \ref{sec:teo_lattice}: the continuous Euclidean space-time is replaced by a finite lattice, derivatives are discretized using Taylor expansion and the space-time integral is replaced by a sum over all the links of the lattice. The lattice version of the model is then defined by the action \begin{align} \nonumber S = - \dfrac{J}{2} \sum_{x,\nu} &\left(\Tr \left[e^{\, \delta_{\nu,d} \sigma_{3} \frac{\mu_{1} + \mu_{2}}{2}} \, U_{x} \, e^{\,\delta_{\nu,d}\sigma_{3}\frac{\mu_{1} - \mu_{2}}{2}} \, U_{x + \hat{\nu}}^{\dagger}\right] \right.\\ & + \left. \Tr \left[e^{-\delta_{\nu,d}\sigma_{3}\frac{\mu_{1} - \mu_{2}}{2}} \, U_{x}^{\dagger} \, e^{-\delta_{\nu,d}\sigma_{3}\frac{\mu_{1} + \mu_{2}}{2}} \, U_{x + \hat{\nu}} \right]\right) \, . \label{eq:pcm_action} \end{align} The matrices $U_{x} \in$ SU(2) live on the sites $x$ of a $d$-dimensional $N^{d-1} \times N_{t}$ lattice with periodic boundary conditions and lattice constant $a$ set to $1$. From the explicit expression (\ref{eq:pcm_action}) of the lattice action it is clear that the chemical potentials $\mu_{1}$ and $\mu_{2}$ give different weights to temporal ($\nu = d$) forward and backward nearest neighbor terms. In other words, they favor matter over anti-matter of the charge they couple to. Furthermore, it is important to notice that the introduction of the chemical potentials $\mu_{\lambda}$, $\lambda = 1,2$ also causes the action $S$ to become complex. Therefore in the conventional representation the model suffers from the complex action problem. As usual, the partition function is obtained integrating the Boltzmann weight $e^{-S}$ over the product of SU(2) Haar measures $\int \! D[U] = \prod_{x} \int_{SU(2)} dU_{x}$, giving rise to \begin{equation} \label{eq:pcm_partition} Z = \int \! D[U] \, e^{-S} \, . \end{equation} In what follows we will use the explicit parametrization of the SU(2) matrices \begin{equation} \label{eq:pcm_su2parametrization} U_{x} = \left[ \begin{array}{cc} \cos \theta_{x} \, e^{\,i \alpha_{x}} & \sin \theta_{x} \, e^{\, i \beta_{x}} \\ -\sin \theta_{x} \, e^{-i \beta_{x}} & \cos \theta_{x} \, e^{-i \alpha_{x}} \end{array}\right] \quad , \quad dU_{x} = 2 \sin \theta_{x} \cos \theta_{x} \, d\theta_{x} \dfrac{d\alpha_{x}}{2 \pi} \, \dfrac{d\beta_{x}}{2 \pi} \, , \end{equation} with $\theta_{x} \in [0, \pi/2]$, $\alpha_{x} \in [-\pi, \pi]$ and $\beta_{x} \in [-\pi, \pi]$. In the next section we will show how to obtain a worldline representation of the model described by the action (\ref{eq:pcm_action}) using the abelian color flux (ACF) method. Our dualization will result in an expression of the partition sum (\ref{eq:pcm_partition}) that only has real and positive terms. It will thus solve the complex action problem. \section{Worldline formulation \label{sec:pcm_worldline}} For the ACF dualization of the SU(2) principal chiral model with chemical potentials $\mu_{\lambda}$, $\lambda = 1,2$, we start by making all the traces and matrix products in the lattice action (\ref{eq:pcm_action}) explicit, i.e., we rewrite them as sums over the indices $a$, $b$ of the matrices $U_{x}^{ab}$: \begin{equation} \label{eq:pcm_action2} S = - J \sum_{x,\nu} \Big[ e^{\,\mu_{1} \delta_{\nu,d}} \, U_{x}^{11} U_{x + \hat{\nu}}^{11 \, \star} + e^{\,\mu_{2} \delta_{\nu,d}} \, U_{x}^{12} U_{x + \hat{\nu}}^{12 \, \star} + e^{-\mu_{2} \delta_{\nu,d}} \, U_{x}^{21} U_{x + \hat{\nu}}^{21 \, \star} + e^{-\mu_{1} \delta_{\nu,d}} \, U_{x}^{22} U_{x + \hat{\nu}}^{22 \, \star} \Big] \, . \end{equation} This expression of the action makes evident that non-zero values of the chemical potentials give rise to a complex action. In fact, from the parametrization (\ref{eq:pcm_su2parametrization}), the following identities between products of SU(2) matrix elements hold \begin{equation*} U_{x}^{22} U_{x + \hat{\nu}}^{22 \, \star} = (U_{x}^{11} U_{x + \hat{\nu}}^{11 \, \star})^{\star}\, , \qquad \qquad U_{x}^{21} U_{x + \hat{\nu}}^{21 \, \star} = (U_{x}^{12} U_{x + \hat{\nu}}^{12 \, \star})^{\star}\, . \end{equation*} Hence, when either of the chemical potentials $\mu_{1}$ or $\mu_{2}$ are finite, the imaginary parts of the terms $U_{x}^{11} U_{x + \hat{\nu}}^{11 \, \star}$ and $U_{x}^{12} U_{x + \hat{\nu}}^{12 \, \star}$ respectively do not cancel, consequently the action $S$ becomes complex. We will see at the end of our dualization program that the worldline formulation we will obtain completely solves the complex action problem. We continue the ACF dualization with introducing the matrices $M_{\nu}$ with entries \begin{equation} \label{eq:Mentries} M_{\nu}^{11} = e^{\mu_{1} \delta_{\nu,d}} \, , \qquad M_{\nu}^{22} = e^{-\mu_{1} \delta_{\nu,d}} \, , \qquad M_{\nu}^{12} = e^{\mu_{2} \delta_{\nu,d}} \, , \qquad M_{\nu}^{21} = e^{-\mu_{2} \delta_{\nu,d}} \, , \end{equation} and rewriting the partition sum as \begin{equation} Z = \int \! D[U] \exp \left( J \sum_{x,\nu} \sum_{a,b = 1}^{2} M_{\nu}^{ab} U_{x}^{ab} U_{x + \hat{\nu}}^{ab \, \star} \right) = \int \! D[U] \prod_{x,\nu} \prod_{a,b = 1}^{2} e^{\,J M_{\nu}^{ab} U_{x}^{ab} U_{x + \hat{\nu}}^{ab \, \star}} \, . \label{eq:pcm_partition2} \end{equation} In the second step of (\ref{eq:pcm_partition2}) the exponential of sums was factorized into a product of exponentials. Then, we expand each of the exponentials $e^{J M_{\nu}^{ab} U_{x}^{ab} U_{x + \hat{\nu}}^{ab \, \star}}$ in a power series, thus introducing summation variables $j_{x,\nu}^{ab}$, \begin{align} \nonumber Z& = \int \!\! D[U] \prod_{x,\nu} \prod_{a,b = 1}^{2} \sum_{j_{x,\nu}^{ab} = 0}^{\infty} \dfrac{(J M_{\nu}^{ab})^{j_{x,\nu}^{ab}}}{j_{x,\nu}^{ab}!} (U_{x}^{ab} U_{x + \hat{\nu}}^{ab \, \star})^{j_{x,\nu}^{ab}} \\ &= \sum_{\{ j \}} W_{J,\mu}[j] \int \!\! D[U] \prod_{x,\nu} \prod_{a,b} (U_{x}^{ab})^{j_{x,\nu}^{ab}} (U_{x}^{ab \, \star})^{j_{x - \hat{\nu},\nu}^{ab}} \, . \label{eq:pcm_partition3} \end{align} From the first line of this equation we can already observe that, since every matrix element $U_{x}^{ab}$ is multiplied by the same matrix element on the neighboring site $U_{x + \hat{\nu}}^{ab}$, there is no complex action problem left after the expansion. In the second step of (\ref{eq:pcm_partition3}) we have reordered the factors $U_{x}^{ab}$ and introduced the sum $\sum_{\{j\}}$ over all configurations of the summation variables $j_{x,\nu}^{ab}$, as well as the weight factors $W_{J,\mu}[j]$: \begin{align} \nonumber &\sum_{\{j\}} \equiv \prod_{x,\nu} \prod_{ab} \sum_{j_{x,\nu}^{ab} = 0}^{\infty} \, , \\ &W_{J,\mu}[j] \equiv \prod_{x,\nu} \prod_{ab} \dfrac{(J M_{\nu}^{ab})^{j_{x,\nu}^{ab}}}{j_{x,\nu}^{ab}!} = e^{\, \mu_{1} \sum_{x}[j_{x,d}^{11} - j_{x,d}^{22}]} \, e^{\, \mu_{2} \sum_{x}[j_{x,d}^{12} - j_{x,d}^{21}]} \prod_{x,\nu} \prod_{ab} \dfrac{J^{j_{x,\nu}^{ab}}}{j_{x,\nu}^{ab}!} \, . \label{eq:pcm_wjmu} \end{align} The integration over the product Haar measure $\int \! D[U] = \prod_{x} \int_{SU(2)} dU_{x}$ in (\ref{eq:pcm_partition3}) can be carried out analytically by inserting the explicit parametrization (\ref{eq:pcm_su2parametrization}) for the $U_{x}$ matrices and the Haar integration measure in Eq.~(\ref{eq:pcm_partition3}): \begin{align} \nonumber Z= &\sum_{\{ j \}} W_{J,\mu}[j] \prod_{x} 2 \int_{0}^{\frac{\pi}{2}} \! d\theta_{x} (\cos \theta_{x})^{1 + \sum_{\nu}[j_{x,\nu}^{11} + j_{x,\nu}^{22} + j_{x-\hat{\nu},\nu}^{11} + j_{x-\hat{\nu},\nu}^{22}]} \\ \nonumber &\times (\sin \theta_{x})^{1 + \sum_{\nu}[j_{x,\nu}^{12} + j_{x,\nu}^{21} + j_{x-\hat{\nu},\nu}^{12} + j_{x-\hat{\nu},\nu}^{21}]} \int_{-\pi}^{\pi} \! \dfrac{d\alpha_{x}}{2\pi} e^{i\alpha_{x} \sum_{\nu}[(j_{x,\nu}^{11} - j_{x,\nu}^{22}) - (j_{x-\hat{\nu},\nu}^{11} - j_{x-\hat{\nu},\nu}^{22})]}\\ & \times \int_{-\pi}^{\pi} \! \dfrac{d\beta_{x}}{2\pi} e^{i\beta_{x} \sum_{\nu}[(j_{x,\nu}^{12} - j_{x,\nu}^{21}) - (j_{x-\hat{\nu},\nu}^{12} - j_{x-\hat{\nu},\nu}^{21})]} \, . \label{eq:pcm_partition4} \end{align} The integrals over the angles $\alpha$ and $\beta$ give rise to Kronecker deltas for the integer valued combinations of $j_{x,\nu}^{ab}$ in the respective exponents. They enforce constraints for the $j$-fluxes at every site $x$. The integrals over the $\theta$ angles can be solved using the following representation of the beta function \cite{Olver:2010:NHM:1830479} \begin{equation} 2 \int_{0}^{\frac{\pi}{2}} \! d\theta \, (\cos \theta)^{1 + n} \,(\sin \theta)^{1 + m} \, = \, \B \bigg(\frac{n}{2} + 1 \bigg| \frac{m}{2} + 1 \bigg) = \dfrac{\frac{n}{2}! \frac{m}{2}!}{\big(\frac{n + m}{2} + 1 \big)!} \, , \label{eq:pcm_betafunction} \end{equation} where the last equality holds only if both $n$ and $m$ are even. It can be easily shown that this is the case in Eq.~(\ref{eq:pcm_partition4}). In fact, as we have already said, the $\alpha$ and $\beta$ integrals result in two constraints at every site $x$: \begin{align*} \sum_{\nu}[(j_{x,\nu}^{11} - j_{x,\nu}^{22}) - (j_{x-\hat{\nu},\nu}^{11} - j_{x-\hat{\nu},\nu}^{22})] = 0 \, , \quad \sum_{\nu}[(j_{x,\nu}^{12} - j_{x,\nu}^{21}) - (j_{x-\hat{\nu},\nu}^{12} - j_{x-\hat{\nu},\nu}^{21})] = 0 \, . \end{align*} The expressions on the left hand side of these two equations can be turned into the combinations of $j$-fluxes at the exponent of the cosine and the sine in Eq. (\ref{eq:pcm_partition4}) by adding on both sides of the equations even quantities. Hence, these exponents are even and the $\theta$ integral results in simple combinatorial factors, that are fractions of factorials, as given in (\ref{eq:pcm_betafunction}). In order to make the constraints more transparent it is favorable to introduce new \textit{flux variables} $k_{x,\nu}^{\lambda} \in \mathbb{Z}$, $\lambda = 1, 2$ and \textit{auxiliary variables} $m_{x,\nu}^{\lambda} \in \mathbb{N}_{0}$, $\lambda = 1, 2$, defined as linear combinations of the $j_{x,\nu}^{ab}$: \begin{align} \nonumber &k_{x,\nu}^{1} = j_{x,\nu}^{11} + j_{x,\nu}^{22} \, , \qquad k_{x,\nu}^{2} = j_{x,\nu}^{12} + j_{x,\nu}^{21} \, , \\ &m_{x,\nu}^{1} = \dfrac{j_{x,\nu}^{11} + j_{x,\nu}^{22} - |k_{x,\nu}^{1}|}{2} \, , \qquad m_{x,\nu}^{2} = \dfrac{j_{x,\nu}^{12} + j_{x,\nu}^{21} - |k_{x,\nu}^{2}|}{2} \, . \label{eq:pcm_dualvariables} \end{align} Using these the constraints turn into \begin{equation} \label{eq:pcm_constraints} \vec{\nabla} \vec{k}_{x}^{1} = 0 \quad \forall x \, , \qquad \vec{\nabla} \vec{k}_{x}^{2} = 0 \quad \forall x \, , \end{equation} where we introduced the discretized divergence \begin{equation} \label{eq:pcm_divergence} \vec{\nabla} \vec{k}_{x}^{\lambda} \equiv \sum_{\nu} [k_{x,\nu}^{\lambda} - k_{x- \hat{\nu},\nu}^{\lambda}] \, . \end{equation} We can now obtain the final form of the worldline representation of the partition sum just by inserting (\ref{eq:pcm_dualvariables}) into (\ref{eq:pcm_partition4}) after all integrals were solved \begin{equation} \label{eq:pcm_partitionworldline} Z = \sum_{\{k,m\}} \, W_{J} [k,m] \, W_{H} [k,m] \, W_{\mu}[k] \prod_{x} \, \prod_{\lambda=1}^{2} \, \delta \left(\vec{\nabla} \vec{k}_{x}^{\lambda}\right) \, . \end{equation} The partition function is a sum $\sum_{\{k,m\}}$ over all possible configurations of the dual variables $k_{x,\nu}^{\lambda} \in \mathbb{Z}$, $\lambda = 1, 2$ and $m_{x,\nu}^{\lambda} \in \mathbb{N}_{0}$, $\lambda = 1, 2$. The admissible configurations of the flux variables $k_{x,\nu}^{\lambda}$ are specified by the product of Kronecker deltas (we use the notation $\delta(n) \equiv \delta_{n,0}$) in (\ref{eq:pcm_partitionworldline}). These Kronecker deltas require the discretized divergence of both $k_{x,\nu}^{1}$ and $k_{x,\nu}^{2}$ to vanish at every site $x$. Therefore, the condition $\vec{\nabla} \vec{k}_{x}^{\lambda} = 0$ $\forall x$ implemented by the product of the Kronecker deltas in (\ref{eq:pcm_partitionworldline}) implies that at each site $x$ the total flux of $k_{x,\nu}^{\lambda}$ has to vanish. In other words, the fluxes of $k_{x,\nu}^{1}$ and of $k_{x,\nu}^{2}$ must form closed worldlines. The auxiliary variables $m_{x,\nu}^{\lambda}$ are unconstrained. Hence, in the worldline representation, the principal chiral model consists of two species of worldlines ($\lambda = 1$ and $\lambda = 2$) that are constrained to form closed loops, and two additional species of auxiliary variables which are unconstrained. An example of an admissible configuration of the $k$-flux variables is represented in Fig.~\ref{fig:pcm_config}. \begin{figure} \centering \includegraphics[scale=1.7]{chapters/plots/pcm.pdf} \caption{Example of an admissible configuration of the flux variables. $k_{x,\nu}^{1}$ is represented in red and $k_{x,\nu}^{2}$ in blue. The number of arrows denotes $|k_{x,\nu}^{\lambda}|$, i.e., the units of flux on the link $(x,\nu)$. The direction of the arrow reflects the sign of $k_{x,\nu}^{\lambda}$: it has positive (negative) orientation when $k_{x,\nu}^{\lambda}$ is positive (negative). \label{fig:pcm_config}} \end{figure} There we depicted $k_{x,\nu}^{1}$ in red and $k_{x,\nu}^{2}$ in blue. The number of arrows denotes $|k_{x,\nu}^{\lambda}|$, which has the interpretation of the units of flux on the link $(x,\nu)$, while the direction of the arrow depends on the sign of $k_{x,\nu}^{\lambda}$: it has positive (negative) orientation when $k_{x,\nu}^{\lambda}$ is positive (negative). Since the constraints are independent for the two species, the two kinds of worldlines must form closed loops or winding loops independently from each other. Nevertheless, as we will see later, they interact with each other, as well as with the auxiliary variables, through the weight factors $W$. In fact, the configurations of the $k_{x,\nu}^{\lambda}$ and $m_{x,\nu}^{\lambda}$ come with three different weights: $W_{J} [k, m]$, $W_{H} [k, m]$ and $W_{\mu} [k]$. $W_{J} [k, m]$ is the $J$-dependent combinatorial weight factor that arises from the Taylor expansion of the exponential factors in (\ref{eq:pcm_partition2}). It is given by \begin{equation} W_{J} [k, m] = \prod_{x,\nu} \prod_{\lambda = 1}^{2}\dfrac{J^{D_{x,\nu}^{\lambda}}}{(D_{x,\nu}^{\lambda} - m_{x,\nu}^{\lambda})!m_{x,\nu}^{\lambda}!} \, , \label{eq:pcm_wj} \end{equation} where we have introduced the abbreviation \begin{equation} D_{x,\nu}^{\lambda} \equiv |k_{x,\nu}^{\lambda}| + 2 m_{x,\nu}^{\lambda} \, . \label{eq:pcm_d} \end{equation} $W_{H} [k, m]$ collects the fractions of factorials (\ref{eq:pcm_betafunction}) resulting from the integration of the $\theta$ angle in the Haar measure integrals. Also in this case we use the abbreviation (\ref{eq:pcm_d}) and obtain \begin{equation} W_{H} [k, m] = \prod_{x} \dfrac{\prod_{\lambda = 1}^{2} \left(\frac{1}{2} \sum_{\nu} [D_{x,\nu}^{\lambda} + D_{x-\hat{\nu},\nu}^{\lambda}]\right)!}{\left(1 + \frac{1}{2} \sum_{\nu} \sum_{\lambda} [D_{x,\nu}^{\lambda} + D_{x-\hat{\nu},\nu}^{\lambda}]\right)!} \, . \label{eq:pcm_wh} \end{equation} Finally, $W_{\mu} [k]$ gives the $\mu$-dependence for the coupling with the chemical potentials \begin{equation} \label{eq:pcm_wmu} W_{\mu}[k] = \prod_{\lambda = 1}^{2} \prod_{x} e^{\ \mu_{\lambda} k_{x,d}^{\lambda}} = \prod_{\lambda = 1}^{2} e^{\ \mu_{\lambda} \sum_{x} k_{x,d}^{\lambda}} = e^{\ \mu_{1} \, \beta \, \omega_{1}[k]}\ e^{\ \mu_{2} \, \beta \, \omega_{2}[k]} \, . \end{equation} In the last step of (\ref{eq:pcm_wmu}) we used the identity $\sum_{x} k_{x,d}^{\lambda} = N_{t} \omega_{\lambda}[k]$ where $\omega_{\lambda}[k]$ is the temporal net winding number of the $k^{\lambda}$-flux and $N_{t}$ the extent of the lattice in time ($\nu = d$) direction. The identity holds because the admissible configurations of the $k_{x,\nu}^{\lambda}$ are closed worldlines. In (\ref{eq:pcm_wmu}) we also used the fact that $N_{t}$ is the inverse temperature in lattice units and replaced $N_{t}$ by the more conventional symbol $\beta$. From Eq.~(\ref{eq:pcm_wmu}) we can also deduce that the conserved charges we computed in Section \ref{sec:pcm_kw} in the worldline formulation have the topological interpretation of the integer valued net winding numbers of the respective worldlines. This feature is of great advantage, because it allows one to identify the conserved charges easily, and thus it also opens the possibility to perform canonical simulations (compare \cite{Orasch:2017niz}). Since all the weight factors (\ref{eq:pcm_wj}) -- (\ref{eq:pcm_wmu}) are real and positive, numerical simulations of this model in the worldline representation (\ref{eq:pcm_partitionworldline}) are indeed possible. In \cite{Gattringer:2017hhn} an exploratory numerical test was done with this representation. Here we just remark that, due to the fact that the $k$ variables are subject to constraints, simulation strategies have to involve techniques that implement updates that do not violate the constraints. Furthermore these techniques have to be set up such that the updates are still efficient. In \cite{Gattringer:2017hhn} we performed two types of simulations. One in which the $k$ variables are updated via a generalization of the worm algorithm \cite{PhysRevLett.87.160601} (taking into account the site terms using the variant described in \cite{Giuliani:2017mxu}), the other using local updates. The auxiliary $m$ variables were always updated with local Metropolis sweeps. We remark that our worldline representation is not unique and other alternative dual representations can be found \cite{Rindlisbacher:2015xku,Kovacs:1993ta,PhysRevD.49.6072}. Furthermore, the SU(2) principal chiral model is equivalent to the O(4) nonlinear sigma model. In this form dual formulations were presented in \cite{Bruckmann:2015sua,Wolff:2009kp}. As we outlined at the beginning of this chapter, one of our main goals was to identify how the original symmetry of the model translates in the dual representation. It is clear since quite some time now that the constraints arising from the integration of the conventional degrees of freedom carry information about the original symmetry of the system. Often for matter fields the constraints enforce site based flux conservation laws. In the specific case of the principal chiral model, constraints over the $k$-fluxes arise from the integration over the U(1) phases. The whole SU(2) description is recovered with the weight factors, that tie together the two worldlines $k^{1}$ and $k^{2}$, and the auxiliary variables $m^{\lambda}_{x,\nu}$, $\lambda=1,2$. \section{Kramers-Wannier dual representation \label{sec:pcm_kw}} Having discussed the ACF dualization of the principal chiral model with two chemical potentials, here we present a second reformulation of the same model. Starting from the worldline representation of the partition function given by Eqs.~(\ref{eq:pcm_partitionworldline})--(\ref{eq:pcm_wmu}), we show how it is possible to take the dualization a step further by rewriting the partition sum in terms of yet another set of dual variables. These variables live on what is called \textit{dual lattice}, and they automatically solve the constraints in Eq.~(\ref{eq:pcm_partitionworldline}). In this sense we can say that the system is fully dualized, \textit{a la} Kramers and Wannier \cite{Kramers:1941zz}. The first step for the Kramers-Wannier (KW) dualization program consists in finding new variables that can generate all the admissible configurations of the $k^{\lambda}$-fluxes and that, at the same time, resolve the flux conservation constraints enforced by the Kronecker deltas in Eq.~(\ref{eq:pcm_partitionworldline}). In our case these conditions can be fulfilled by introducing two kinds of new dual variables: plaquette fluxes and disorder loops. For a graphical illustration we refer to Fig.~\ref{fig:pcm_kw}. \begin{figure} \centering \includegraphics[scale=1.5]{chapters/plots/pcm_kw.pdf} \caption{Graphical representation of the plaquette fluxes $n_{x,\nu\rho}^{\lambda}$ (left), and the disorder loops $q_{\nu}^{\lambda}$ (right). In both plots $\sigma < \nu < \rho$. On the left we show the graphical illustration of the two sums in Eq.~(\ref{eq:pcm_hodge}). We depict the four plaquette fluxes that contribute positively to the $k_{x,\nu}^{\lambda}$ in 3 dimensions (there would be six contributions in 4 dimensions). On the right we show the disorder loops, which are forced to be on the coordinate axes by the support functions $\Theta_{x,\nu}^{(\nu)}$, and that implement closed winding loops due to the periodic boundary conditions. \label{fig:pcm_kw}} \end{figure} The plaquette variables are denoted by $n_{x,\nu\rho}^{\lambda} \in \mathbb{Z}$, $\lambda = 1,2$. They generate $|n_{x,\nu\rho}^{\lambda}|$ units of $k^{\lambda}$-flux around the plaquette $(x,\nu\rho)$, $\nu<\rho$, with mathematically positive orientation if $n_{x,\nu\rho}^{\lambda} > 0$ and negative orientation if $n_{x,\nu\rho}^{\lambda} < 0$. The disorder flux is written as $\Theta_{x,\rho}^{(\nu)} \, q_{\nu}^{\lambda}$, $q_{\nu}^{\lambda}\in\mathbb{Z}$, $\lambda = 1,2$ where \begin{equation} \Theta_{x,\rho}^{(\nu)} = \left\{ \begin{array}{l} 1 \quad \text{if } (x,\rho)\in \{ (\mathbf{0} + n \, \hat{\nu},\nu), n = 0, 1, \dots, N_{\nu}-1\} \\ 0 \quad \text{if } (x,\rho)\notin \{ (\mathbf{0} + n \, \hat{\nu},\nu), n = 0, 1, \dots, N_{\nu}-1\} \end{array} \right. \label{eq:pcm_supportfunction} \end{equation} is the support function for the coordinate axis in the $\nu$ direction. In the definition (\ref{eq:pcm_supportfunction}), $\mathbf{0}$ denotes the origin of lattice, defined as the corner with site $(0,0,\dots,0)$. $N_{\nu}$, $\nu = 1, \dots, d-1$ are the spatial extents of the lattice, and $N_{d}$ the temporal extent $N_{d} = N_{t}$. $q_{\nu}^{\lambda} \in \mathbb{Z}$ therefore introduces $|q_{\nu}^{\lambda}|$ units of $k^{\lambda}$-flux on the $\nu$ coordinate axis which is oriented in positive $\nu$-direction for $q_{\nu}^{\lambda} > 0$ and has negative orientation for $q_{\nu}^{\lambda} < 0$. Due to the periodic boundary conditions, the term $\Theta_{x,\nu}^{(\nu)} \, q_{\nu}^{\lambda}$ actually introduces a closed winding loop of $|q_{\nu}^{\lambda}|$ units of flux placed on the coordinate axes through the origin of our $d$-dimensional lattice. Since both the plaquette fluxes and the disorder variables implement only closed worldlines of $k^{\lambda}$-fluxes, they are unconstrained. The flux $k_{x,\nu}^{\lambda}$ receives contributions from all the plaquettes that contain the link $(x,\nu)$, as well as the disorder lines if $(x,\nu)$ sits on the $\nu$-coordinate axis: \begin{equation} \label{eq:pcm_hodge} k_{x,\nu}^{\lambda} = \sum_{\rho: \nu < \rho} \left[ n_{x,\nu\rho}^{\lambda} - n_{x - \hat{\rho},\nu\rho}^{\lambda}\right] - \sum_{\sigma: \nu > \sigma} \left[ n_{x,\sigma\nu}^{\lambda} - n_{x - \hat{\sigma},\sigma\nu}^{\lambda}\right] + \Theta_{x,\nu}^{(\nu)} q_{\nu}^{\lambda} \ . \end{equation} This transformation is a so-called Hodge decomposition and generates all possible configurations of $k_{x,\nu}^{\lambda}$, $\lambda = 1,2$ that obey the zero-divergence constraints. Substituting (\ref{eq:pcm_hodge}) in (\ref{eq:pcm_partitionworldline}) one obtains the following form of the partition function, \begin{equation} \label{eq:pcm_partition5} Z = \sum_{\{q,n,m\}} \, W_{J} [q,n,m] \, W_{H} [q,n,m] \, W_{\mu}[q] \, . \end{equation} The sum is now over all configurations of the variables $q_{\nu}^{\lambda}\in\mathbb{Z}$, $n_{x,\nu\rho}^{\lambda}\in\mathbb{Z}$ and the auxiliary variables $m_{x,\nu}^{\lambda}\in\mathbb{N}_{0}$, $\lambda=1,2$. The product of Kronecker deltas has disappeared, since the constraints are automatically satisfied by the configurations of the new variables $\{q,n\}$. The weight factors $W_{J} [q,n,m]$ and $W_{H} [q,n,m]$ are still given by (\ref{eq:pcm_wj}) and (\ref{eq:pcm_wh}) respectively, but now the combinations $D_{x,\nu}^{\lambda} \in \mathbb{N}_{0}$ read \begin{equation} D_{x,\nu}^{\lambda} \equiv \bigg|\sum_{\rho: \nu < \rho} \left[ n_{x,\nu\rho}^{\lambda} - n_{x - \hat{\rho},\nu\rho}^{\lambda}\right] - \sum_{\sigma: \nu > \sigma} \left[ n_{x,\sigma\nu}^{\lambda} - n_{x - \hat{\sigma},\sigma\nu}^{\lambda}\right] + \Theta_{x,\nu}^{(\nu)} q_{\nu}^{\lambda}\bigg| + 2 m_{x,\nu}^{\lambda} \, . \label{eq:pcm_d2} \end{equation} The $\mu$-dependent weight factor is \begin{equation} \label{eq:pcm_wmu2} W_{\mu}[q] = e^{\ \beta (\mu_{1} \, q_{d}^{1}\ + \mu_{2} \, q_{d}^{2} ) } \, . \end{equation} We obtained this expression by observing that, in terms of the new dual variables introduced in Eq.~(\ref{eq:pcm_hodge}), the net temporal winding number $\omega_{\lambda}[k]$ is indeed represented by the disorder variables $q_{d}^{\lambda}$ (compare with the definition of $q_{\nu}^{\lambda}$ we gave earlier). The last step to obtain the KW-dual of the principal chiral model consists in switching to the dual lattice. In general, the construction of the lattice dual to a $d$-dimensional hypercubical lattice can be done by shifting the lattice by half a lattice spacing in every direction. More mathematically we may introduce the notion of a \textit{simplex} of dimension $s$ as an $s$-dimensional element of the hypercubical lattice: a simplex of dimension 0 is a site of the lattice, a simplex of dimension 1 is a link on the lattice, a simplex of dimension 2 is a plaquette, a simplex of dimension 3 an elementary cube, and so on. Obviously, in a $d$-dimensional lattice there are simplices of dimension $0 \leq s \leq d$. From the way the dual lattice is constructed follows that each simplex of dimension $s$ of the original lattice is in one-to-one correspondence with one simplex of dimension $\tilde{s} = d - s$ of the dual lattice. This correspondence will become clearer in the next subsections, where we will discuss in detail the mapping of the worldline representation of the principal chiral model to the dual lattice for the special cases of two and four dimensions. The generalization to any other dimensions is straightforward. \subsection{Fully dualized form in two dimensions} Following the above discussion, in two dimensions the dual lattice is built by placing a new site in the center of each plaquette of the original lattice. We will denote the sites of the dual lattice, as well as the variables living on it, with a tilde, e.g., $\tilde{x}$, $\tilde{n}_{\tilde{x}}^{\lambda}$. For the two-dimensional case it is easy to identify the one-to-one correspondence we mentioned earlier: each site $x$ of the original lattice lies in the center of one plaquette $(\tilde{x},12)$ of the dual lattice; likewise, each site $(\tilde{x}+\hat{1}+\hat{2})$ of the dual lattice is placed at the center of one plaquette $(x,12)$ of the original lattice; finally, each link $(x,1)$ in the spatial direction of the original lattice crosses one link in the temporal direction $(\tilde{x}+\hat{1},2)$ of the dual lattice and vice versa (refer to Fig.~\ref{fig:pcm_dual} for a graphical illustration of this mapping). \begin{figure} \centering \includegraphics[scale=1]{chapters/plots/pcm_dual.pdf} \caption{In 2 dimensions the dual lattice is built by placing a new site in the center of each plaquette of the original lattice. Here we represent the dual lattice with dashed lines and the original lattice with solid lines. We graphically show the one-to-one correspondence discussed in the text: simplices of dimension $s$, $0 \leq s \leq 2$, of the original lattice (illustrated in blue) are mapped to simplices of dimension $\tilde{s} = 2 - s$ of the dual lattice (illustrated in red). \label{fig:pcm_dual}} \end{figure} The dual variables $m_{x,\nu}^{\lambda}$, $n_{x,\nu\rho}^{\lambda}$ and $q_{\nu}^{\lambda}$, $\lambda = 1,2$ are mapped to the dual lattice accordingly: \begin{equation} \label{eq:pcm_2dmaptodual} n_{x,12}^{\lambda} \rightarrow \tilde{n}_{\tilde{x} + \hat{1} + \hat{2}}^{\lambda} \, , \qquad m_{x,\nu}^{\lambda} \rightarrow \tilde{m}_{\tilde{x} + \hat{\nu}, \rho}^{\lambda} \, , \qquad \Theta_{x,\nu}^{(\nu)}\, q_{\nu}^{\lambda} \rightarrow \tilde{\Theta}_{\tilde{x} + \hat{\nu}, \rho }^{(\tilde{\nu})}\, \tilde{q}_{\tilde{\nu}}^{\lambda} \, , \quad \text{with } \rho\neq \nu \, . \end{equation} In (\ref{eq:pcm_2dmaptodual}) the support function $ \tilde{\Theta}_{\tilde{x}, \rho }^{(\tilde{\nu})}$ is nonzero only on the links $ (\tilde{x}, \rho )$ of the dual lattice that are dual to the $\nu$ coordinate axis. In the KW-dual formulation of the principal chiral model in 2 dimensions the dynamical degrees of freedom are the dual variables $\tilde{n}_{\tilde{x}}^{\lambda} \in \mathbb{Z}$ assigned to the sites of the dual lattice, the dual auxiliary variables $\tilde{m}_{\tilde{x}, \rho}^{\lambda} \in \mathbb{N}_{0}$ on the links of the dual lattice, and the dual disorder variables $\tilde{q}_{\tilde{\nu}}^{\lambda}\in\mathbb{Z}$ for the flux through the links $(\tilde{x},\rho)$ dual to the $\nu$ coordinate axes ($\rho \neq \nu$). At this point, to obtain the fully dualized form of the partition sum it is sufficient to apply the mapping (\ref{eq:pcm_2dmaptodual}) to the expression (\ref{eq:pcm_partition5}) of the partition function, \begin{equation} \label{eq:pcm_2dpartitionkw} Z = \sum_{\tilde{q}_{\tilde{2}}^{1}, \tilde{q}_{\tilde{2}}^{2} \in \mathbb{Z}} e^{\, \beta \, \left( \mu_{1} \tilde{q}_{\tilde{2}}^{1} + \mu_{2} \tilde{q}_{\tilde{2}}^{2}\right)} \sum_{\tilde{q}_{\tilde{1}}^{1}, \tilde{q}_{\tilde{1}}^{2} \in \mathbb{Z}} \, \sum_{\{\tilde{n}, \tilde{m}\}} W_{J} [\tilde{q}, \tilde{n}, \tilde{m}] W_{H} [\tilde{q}, \tilde{n}, \tilde{m}] \, . \end{equation} In (\ref{eq:pcm_2dpartitionkw}) we wrote explicitly the sum $\sum_{\{\tilde{q}\}}$ over the configurations of the disorder variables. In this way, we are able to factorize the $\mu$-dependence, and the partition sum has the form of a double fugacity expansion. Thus, the form (\ref{eq:pcm_2dpartitionkw}) of the partition function is suitable for canonical simulations of the system. The sum over the configurations of the dual site variables $\tilde{n}_{\tilde{x}}^{\lambda} \in \mathbb{Z}$ and the dual auxiliary variables $\tilde{m}_{\tilde{x}, \rho}^{\lambda} \in \mathbb{N}_{0}$ is expressed by $\sum_{\{\tilde{n}, \tilde{m}\}}$, and the weights of the configurations are \begin{equation} \label{eq:pcm_2dwj} W_{J} [\tilde{q}, \tilde{n}, \tilde{m}] = \prod_{\tilde{x}, \nu} \prod_{\lambda = 1}^{2}\dfrac{J^{\tilde{D}_{\tilde{x}, \nu}^{\lambda}}}{(\tilde{D}_{\tilde{x}, \nu}^{\lambda} - \tilde{m}_{\tilde{x}, \nu}^{\lambda})!\, \tilde{m}_{\tilde{x}, \nu}^{\lambda}!}, \end{equation} \begin{equation} \label{eq:pcm_2dwh} W_{H} [\tilde{n}, \tilde{q}, \tilde{m}] = \prod_{\tilde{p}} \dfrac{\prod_{\lambda} \left( \frac{1}{2} \sum_{(\tilde{x}, \nu) \in \partial \tilde{p}} \tilde{D}_{\tilde{x}, \nu}^{\lambda} \right) !}{\left(1 + \frac{1}{2} \sum_{\lambda} \sum_{(\tilde{x}, \nu) \in \partial \tilde{p}} \tilde{D}_{\tilde{x}, \nu}^{\lambda} \right) !} \, , \end{equation} where the combinations $\tilde{D}_{\tilde{x}, \nu}^{\lambda} \in \mathbb{N}_{0}$ assigned to the links of the dual lattice are \begin{equation} \tilde{D}_{\tilde{x}, 1}^{\lambda} = \left| \tilde{n}_{\tilde{x}}^{\lambda} - \tilde{n}_{\tilde{x} + \hat{1}}^{\lambda} + \tilde{\Theta}_{\tilde{x}, 1}^{(\tilde{2})} \, q_{\tilde{2}}^{\lambda}\, \right| + 2 m_{\tilde{x},1}^{\lambda} \, ,\quad \tilde{D}_{\tilde{x}, 2}^{\lambda} = \left| \tilde{n}_{\tilde{x} + \hat{2}}^{\lambda} - \tilde{n}_{\tilde{x}}^{\lambda} + \tilde{\Theta}_{\tilde{x}, 2}^{(\tilde{1})} \, q_{\tilde{1}}^{\lambda}\, \right| + 2 m_{\tilde{x},2}^{\lambda} \, . \end{equation} In Eq.~(\ref{eq:pcm_2dwh}) the product $\prod_{\tilde{p}}$ runs over all plaquettes $\tilde{p}$ of the dual lattice, and the sum $\sum_{(\tilde{x}, \nu) \in \partial \tilde{p}}$ is over all links $(\tilde{x},\nu)$ on the boundary $\partial \tilde{p}$ of $\tilde{p}$. Concluding, we found that in two dimensions the KW-dual of the principal chiral model with two chemical potentials is given by a sum over the configurations of the dual disorder variables $\tilde{q}_{\tilde{\nu}}^{\lambda} \in \mathbb{Z}$, the dual site variables $\tilde{n}_{\tilde{x}}^{\lambda} \in \mathbb{Z}$ and the dual auxiliary variables $\tilde{m}_{\tilde{x}, \nu}^{\lambda} \in \mathbb{N}_{0}$. All the degrees of freedom are unconstrained in the KW-formulation of the model. The only variables that couple to the chemical potentials are the disorder variables on the links dual to the temporal coordinate axis of the original lattice, i.e., $\tilde{q}_{\tilde{2}}^{\lambda} \in \mathbb{Z}$. Therefore, in this representation it is possible to completely factorize the $\mu$-dependence and write the partition sum as a double fugacity expansion. The canonical partition function then sums over the configurations of the disorder variables $\tilde{q}_{\tilde{1}}^{\lambda} \in \mathbb{Z}$, the site variables and the auxiliary variables, and it collects the weights $W_{J} [\tilde{q}, \tilde{n}, \tilde{m}]$ and $W_{H} [\tilde{q}, \tilde{n}, \tilde{m}]$ for those configurations. Notice that both $W_{J} [\tilde{q}, \tilde{n}, \tilde{m}]$ and $W_{H} [\tilde{q}, \tilde{n}, \tilde{m}]$ are real and positive, therefore also the KW-dual formulation (\ref{eq:pcm_2dpartitionkw}) is free of the sign problem, and it can be used for simulations at finite $\mu_{\lambda}$. \subsection{Fully dualized form in four dimensions} Having discussed the two-dimensional case, it is now straightforward to generalize the results to four dimensions. The dual lattice is constructed by placing the new sites in the center of each hypercube of the original lattice. Explicitly the map to the dual lattice reads \begin{equation*} \begin{array}{cccccc} \text{site }& x &\leftrightarrow & (\tilde{x}, 1234)& \text{hypercube } \, ,& \\ \text{link } &(x, \nu) &\leftrightarrow & (\tilde{x} + \hat{\nu}, \rho \sigma \tau) &\text{cube }& \rho < \sigma < \tau \, ,\\ \text{plaquette } &(x, \nu \rho) &\leftrightarrow & (\tilde{x} + \hat{\nu} + \hat{\rho}, \sigma \tau) &\text{plaquette } & \nu < \rho \text{ and } \sigma < \tau \, ,\\ \text{cube } &(x, \nu\rho\sigma) &\leftrightarrow & (\tilde{x} + \hat{\nu} + \hat{\rho} + \hat{\sigma}, \tau) &\text{link }& \nu < \rho < \sigma \, ,\\ \text{hypercube }& (x, 1234) &\leftrightarrow &\tilde{x} + \hat{1} + \hat{2} + \hat{3} + \hat{4}& \text{site } \, .& \end{array} \end{equation*} The variables are then mapped to the dual lattice as follows \begin{align} \nonumber &n_{x,\nu\rho}^{\lambda} \rightarrow \tilde{n}_{\tilde{x} + \hat{\nu} + \hat{\rho}, \sigma \tau}^{\lambda}\, , \qquad \text{with } \nu < \rho \text{ and } \sigma < \tau \, , \\ \nonumber &m_{x,\nu}^{\lambda} \rightarrow \tilde{m}_{\tilde{x} + \hat{\nu}, \rho \sigma \tau}^{\lambda}\, , \qquad \ \text{with } \rho < \sigma < \tau \, ,\\ &\Theta_{x,\nu}^{(\nu)}\, q_{\nu}^{\lambda} \rightarrow \tilde{\Theta}_{\tilde{x} +\hat{\nu}, \tilde{\nu} }^{(\tilde{\nu})}\, \tilde{q}_{\tilde{\nu}}^{\lambda} \, . \label{eq:pcm_4dmaptodual} \end{align} In the last line of (\ref{eq:pcm_4dmaptodual}) we use the notation $\tilde{\nu}$ to label the cubes of the dual lattice which are dual to links in direction $\nu$ of the original lattice, i.e., $(\tilde{x},\tilde{\nu}) \equiv (\tilde{x}, \sigma \tau \omega )$, $\sigma < \tau < \omega$. The support function $\tilde{\Theta}_{\tilde{x}, \tilde{\nu} }^{(\tilde{\nu})}$ on the dual lattice then is nonvanishing only on the cubes $\tilde{\nu}$ dual to the $\nu$ coordinate axis. In the KW-dual formulation we thus use the dual dynamical variables $\tilde{n}_{\tilde{x}, \sigma \tau}^{\lambda} \in \mathbb{Z}$ assigned to the plaquettes of the dual lattice, the dual auxiliary variables $\tilde{m}_{\tilde{x}, \tilde{\nu}}^{\lambda} \in \mathbb{N}_{0}$ on the cubes of the dual lattice, and the dual disorder variables $\tilde{q}_{\tilde{x}, \tilde{\nu}}^{\lambda} \in \mathbb{Z}$ for the flux through the dual cubes dual to the coordinate axes. The fully KW-dual form of the partition sum in four dimensions is \begin{equation} \label{eq:pcm_4dpartitionkw} Z = \sum_{\tilde{q}_{\tilde{4}}^{1}, \tilde{q}_{\tilde{4}}^{2} \in \mathbb{Z}} e^{\, \beta \, \left( \mu_{1} \tilde{q}_{\tilde{4}}^{1} + \mu_{2} \tilde{q}_{\tilde{4}}^{2}\right)} \left( \prod_{\nu = 1}^{3} \sum_{\tilde{q}_{\tilde{\nu}}^{1}, \tilde{q}_{\tilde{\nu}}^{2} \in \mathbb{Z}} \right) \, \sum_{\{\tilde{n}, \tilde{m}\}} W_{J} [\tilde{n}, \tilde{q}, \tilde{m}] W_{H} [\tilde{n}, \tilde{q}, \tilde{m}] \, , \end{equation} where the sum $\sum_{\{\tilde{n}, \tilde{m}\}}$ now runs over all configurations of the dual $\tilde{n}$- and $\tilde{m}$-variables on the dual lattice. In (\ref{eq:pcm_4dpartitionkw}) the sums over the dual disorder variables $\tilde{q}_{\tilde{\nu}}^{\lambda}$ were written explicitly up front. They are ordered such that the first double sum is over the temporal disorder variables $\tilde{q}_{\tilde{4}}^{\lambda}$ which carry the dependence on the chemical potentials $\mu_{\lambda}$. Thus, also in this case, the KW-dual partition sum (\ref{eq:pcm_4dpartitionkw}) is already organized in the form of a double fugacity expansion. The weight factor $W_{J} [\tilde{n}, \tilde{q}, \tilde{m}]$ from the Taylor expansion of the original Boltzmann factors reads \begin{equation} \label{eq:weightJPCMKW} W_{J} [\tilde{n}, \tilde{q}, \tilde{m}] = \prod_{\tilde{x}, \tilde{\nu}} \prod_{\lambda = 1}^{2}\dfrac{J^{\tilde{D}_{\tilde{x}, \tilde{\nu}}^{\lambda}}}{(\tilde{D}_{\tilde{x}, \tilde{\nu}}^{\lambda} - \tilde{m}_{\tilde{x}, \tilde{\nu}}^{\lambda})!\, \tilde{m}_{\tilde{x}, \tilde{\nu}}^{\lambda}!}, \end{equation} where the combinations $\tilde{D}_{\tilde{x}, \tilde{\nu}}^{\lambda} \in \mathbb{N}_{0}$ assigned to the dual cubes $(\tilde{x}, \tilde{\nu})$ are given by \begin{equation} \tilde{D}_{\tilde{x}, \tilde{\nu}}^{\lambda} = \left| \sum_{\rho: \nu < \rho} \left[ \tilde{n}_{\tilde{x} + \hat{\rho}, \sigma \tau}^{\lambda} - \tilde{n}_{\tilde{x}, \sigma \tau}^{\lambda} \right] - \sum_{\omega: \nu > \omega} \left[ \tilde{n}_{\tilde{x} + \hat{\omega}, \sigma \tau}^{\lambda} - \tilde{n}_{\tilde{x}, \sigma \tau}^{\lambda} \right] + \tilde{\Theta}_{\tilde{x}, \tilde{\nu}}^{(\tilde{\nu})} \, q_{\tilde{\nu}}^{\lambda}\, \right| + 2 m_{\tilde{x},\tilde{\nu}}^{\lambda} \, . \end{equation} The weight $W_{H} [\tilde{n}, \tilde{q}, \tilde{m}]$ that originates from the Haar measure integration and implements the SU(2) symmetry of the conventional representation in the KW-dual formulation is given by \begin{equation} \label{eq:pcm_4dwh} W_{H} [\tilde{n}, \tilde{q}, \tilde{m}] = \prod_{\tilde{h}} \dfrac{\prod_{\lambda} \left( \frac{1}{2} \sum_{(\tilde{x}, \tilde{\nu}) \in \partial \tilde{h}} \tilde{D}_{\tilde{x}, \tilde{\nu}}^{\lambda} \right) !}{\left(1 + \frac{1}{2} \sum_{\lambda} \sum_{(\tilde{x}, \tilde{\nu}) \in \partial \tilde{h}} \tilde{D}_{\tilde{x}, \tilde{\nu}}^{\lambda} \right) !} \, , \end{equation} where the product $\prod_{\tilde{h}}$ runs over all hypercubes $\tilde{h}$ of the dual lattice and the sum $\sum_{(\tilde{x}, \tilde{\nu}) \in \partial \tilde{h}}$ is over all dual cubes $(\tilde{x}, \tilde{\nu})$ in the boundary $\partial \tilde{h}$ of $\tilde{h}$. In the KW-dual form all constraints have disappeared and again all weight factors are real and positive, such that a simulation is possible at finite $\mu_{\lambda}$. The KW-dual form (\ref{eq:pcm_4dpartitionkw}) -- (\ref{eq:pcm_4dwh}) was tested in an exploratory numerical simulation in \cite{Gattringer:2017hhn}. A very good agreement was found between results for bulk variables from simulations in the conventional representation (for vanishing chemical potentials), in the worldline representation (both results from local updates and worm simulation results), as well as in the KW-dual form. The preliminary numerical findings presented there indicate that in some coupling regions the worm update is inefficient despite fine tuning the worm amplitude parameter. In these cases switching to a completely KW-dual form is a good choice for efficient simulations, and the absence of constraints in the KW-dual form might even allow for using strategies such as Swendsen–Wang type algorithms. \chapter{Theoretical background \label{cha:theoretical}} The concepts presented in this chapter are at the base of lattice field theory and are explained in much greater detail and clarity in many textbooks, such as \cite{Gattringer:2010zz,Rothe:1992nt,Montvay:1994cy,Peskin:1995ev}. Nonetheless, in Secs.~\ref{sec:teo_path} -- \ref{sec:teo_numerical} we summarize the main concepts of the lattice regularization of quantum field theories with the aim of making this thesis as self-contained as possible. Then, in Sec.~\ref{sec:teo_sign} we explain the technical problem which arises when introducing the quark chemical potential in the discretized QCD action. This is the so-called sign problem, which prevents the study of the QCD phase diagram with first principle lattice calculations and that since the earliest days of lattice field theory have been stimulating the development of a multitude of methods to overcome it. As we outlined in the introduction, this also stands at the very core of the motivation driving our research, and led us to the development of the abelian color cycle and abelian color flux methods as an extension of the applicability of the dual approach to non-abelian lattice field theories. Another important motivation for us is the understanding of general properties of the dual representations, such as the manifestation of the original symmetries in the dual reformulation, or the possibility of identifying conserved quantities as topological invariants. \section{Path integral quantization of field theories \label{sec:teo_path}} The quantization of fields on the lattice is done using the path integral approach. Since its introduction by Feynman \cite{Feynman}, the path integral method has become a very important tool for elementary particle physics. In this approach the action, rather than the Hamiltonian, is used as the fundamental quantity. The path integral approach reveals the close analogy between quantum field theories and statistical mechanics. The exploitation of this analogy is at the base of the Monte Carlo simulations performed on the lattice. In the path integral approach the quantum mechanical amplitude for a particle to travel from a point $x$ to a point $y$ within the time interval $t$ is expressed as an integral over all classical paths weighted by the exponential of $i$ times the classical action $S$ associated with that trajectory: \begin{equation} \label{eq:teo_path} \langle y | e^{i \hat{H} t}| x \rangle = \int D[x] \, e^{iS} \, , \end{equation} where $\hat{H}$ is the Hamiltonian operator of the quantum mechanical system and $\int D[x]$ denotes the functional integral over all paths. In (\ref{eq:teo_path}) quantum mechanical operators have been eliminated in favor of an infinite-dimensional integral. However, a rigorous definition of the path integral $\int D[x]$ is not possible unless a discretization is performed. Moreover, in the form (\ref{eq:teo_path}) the weight $e^{iS}$ is complex and strongly oscillating, which prevents the use of numerical methods to compute the propagation amplitude (\ref{eq:teo_path}). This problem can be bypassed by Wick rotating the time coordinate $t \rightarrow - i x_{4}$, where $x_{4}$ is the Euclidean time. Then, in Euclidean space-time (\ref{eq:teo_path}) reads \begin{equation} \label{eq:teo_pathe} \langle y | e^{- \hat{H} x_{4}}| x \rangle = \int D[x] \, e^{-S_{E}} \, , \end{equation} where $S_{E} = - i S$ is the Euclidean action. In the next section we will see how the Euclidean formalism for quantum field theories makes the analogy with statistical mechanics more transparent, and we will outline the main advantages of the lattice regularization. \section{Space-time discretization: lattice regularization of quantum field theories \label{sec:teo_lattice}} Consider a quantum field theory governed by the Euclidean action \begin{equation} \label{eq:teo_action} S_{E} = \int d^{4} x \, \mathcal{L} \big(\phi(x), \partial_{\mu} \phi(x) \big) \, , \end{equation} where $\mathcal{L}$ is a general expression of the Euclidean Lagrangian density as a function of the fields $\phi$ and their derivatives $\partial_{\mu} \phi$. In the continuum correlation functions can be computed as functional derivatives of the generating functional $Z[J]$, which is defined as \begin{equation} \label{eq:teo_generating} Z [J] = \int D[\phi]\, e^{- \int d^{4} x (\mathcal{L} - J \phi)} \, , \end{equation} where $\int D[\phi]$ is the infinite-dimensional functional integral over field configurations, and $J \phi$ is a source term. The lattice counterpart of the partition function $Z = Z[0]$ is \begin{equation} \label{eq:teo_partition} Z_{lat} = \int D[\phi]\, e^{- S_{lat} [\phi]} \, . \end{equation} Eq.~(\ref{eq:teo_partition}) is obtained by discretizing the space-time \begin{equation} \label{eq:teo_discretization} x \rightarrow a n \, , \quad n = (n_1,n_2,n_3,n_4) \quad \text{with } n_{i} = 0,1,\dots, N_i-1 \quad \text{for } i= 1,2,3,4 \, . \end{equation} In other words we introduce a four-dimensional lattice $\Lambda$, where the sites $n$ are separated by the lattice spacing $a$, which has the dimension of length. Then, the space-time integral in Eq.~(\ref{eq:teo_action}) becomes a sum over all lattice sites \begin{equation} \label{eq:teo_latticeintegral} \int d^{4} x \, \rightarrow \, a^{4} \sum_{n\in \Lambda} \, , \end{equation} and the value of the field at a space-time point $x$ is replaced by the value of the field at site $n$, i.e., \begin{equation} \label{eq:teo_latticefield} \phi (x) \, \rightarrow \, \phi (n) \, . \end{equation} Finally, the derivatives of the fields may be approximated for small lattice spacing $a$ using Taylor expansions as \begin{equation} \label{eq:teo_latticederivative} \partial_{\mu} \phi (x) \, \rightarrow \, \dfrac{1}{2a} \big(\phi (n + \hat{\mu}) - \phi (n - \hat{\mu})\big)\, . \end{equation} Performing the substitutions (\ref{eq:teo_latticeintegral}) -- (\ref{eq:teo_latticederivative}) in (\ref{eq:teo_action}) one obtains the discretized version $S_{lat}[\phi]$ of the Euclidean continuum action $S_E$. As a first advantage, the introduction of the lattice allows for a rigorous definition of the functional integral $\int D[\phi]$ in (\ref{eq:teo_partition}): \begin{equation} \label{eq:teo_measure} D[\phi] \equiv \prod_{n \in \Lambda} d \phi(n) \, , \end{equation} where $d \phi(n)$ is the integration measure for the fields at site $n$ which depends on the nature of the fields $\phi(n)$ one is considering. Another important advantage of the lattice discretization (\ref{eq:teo_discretization}) is the appearance of an UV cut-off for the discretized field theory. To show this consider, e.g., a function $f(x)$ in one dimension. The Fourier transform $\tilde{f}(k)$ in momentum space of the function $f(x)$ is given by \begin{equation} \label{eq:teo_fourier} \tilde{f}(k) \, = \, \int_{-\infty}^{\infty} dx f(x) \, e^{-i k x} \, . \end{equation} If one then restricts $x$ to be a multiple of lattice spacing $a$, i.e., $x = an$, (\ref{eq:teo_fourier}) becomes \begin{equation} \label{eq:teo_fourierdiscretized} \tilde{f}_{a} (k) \, = \, a \sum_{n \in \mathbb{Z}} f(an) \, e^{-i a k n} \, . \end{equation} It is easy to see that the discretized version (\ref{eq:teo_fourierdiscretized}) of the Fourier transform $\tilde{f}(k)$ is invariant under the substitution $k \to k + 2\pi/a$, i.e., it is a periodic function. Hence, in the Fourier representation of the function $f(an)$, \begin{equation} \label{eq:teo_fourierdecomposition} f(an) \, = \, \int_{-\pi/a}^{+\pi/a} \dfrac{dk}{2\pi} \, \tilde{f}_{a}(k) \, e^{ia k n} \, , \end{equation} the momentum integration is restricted to the so-called Brillouin zone $[-\pi/a, \pi/a]$. Thus we find that the introduction of a lattice $\Lambda$ automatically provides an UV regularization of the discretized field theory thanks to the arising of a momentum cut-off proportional to the inverse lattice spacing $a$. We would now like to address the long awaited discussion about the analogy between the path integral formulation of quantum field theories and statistical mechanics. The expression (\ref{eq:teo_partition}) of the discretized partition sum is reminiscent of the partition function of statistical mechanics. It has in fact the same structure of an integral over all possible configurations of an exponential statistical weight. This analogy plays such an important role in lattice field theories that the nomenclatures used in quantum field theory and statistical mechanics are often interchanged or used synonymously. So, for example, throughout this thesis we will call the Feynman path weight $e^{-S_{latt}}$ the Boltzmann weight for a given field configuration on the lattice. A powerful tool that comes with the identification of the lattice regularized quantum field theory with a statistical mechanical system is the Monte Carlo technique for simulation, which we will briefly discuss in Sec.~\ref{sec:teo_numerical}. Before coming to that, in the next two sections we will illustrate in more detail the lattice discretization of fermions and gauge fields for the specific case of lattice QCD. \section{Lattice discretization of fermions \label{sec:teo_fermions}} On the lattice the fermionic degrees of freedom are described by Grassmann numbers, which implement the correct Fermi statistic. We therefore start this section by discussing the main properties of Grassmann numbers and giving the rules for integration which we will need in the following chapters of this thesis. \subsection{Grassmann variables\label{sec:teo_grassmann}} The Grassmann numbers are numbers that anticommute with each other. So, if we consider the set of Grassmann numbers $\eta_{1}, \dots, \eta_{N}$, the following anticommutation relations hold \begin{equation} \label{eq:teo_anticommutation} \{\eta_{i}, \eta_{j}\} \, = \, \eta_{i} \eta_{j} + \eta_{j} \eta_{i} \, = \, 0 \, , \qquad i,j = 0,\dots , N \, . \end{equation} From Eq.~(\ref{eq:teo_anticommutation}) also follows that Grassmann variables are nilpotent, i.e., \begin{equation} \label{eq:teo_nilpotent} \eta_{i}^{2} \, = \, 0 \, . \end{equation} General functions of Grassmann numbers are therefore polynomials of the form \begin{equation} \label{eq:teo_function} f(\eta) = f_{0} + \sum_{i} f_{i} \eta_{i} + \sum_{i \neq j} f_{ij} \eta_{i} \eta_{j} + \dots + f_{12 \dots N} \eta_{1} \eta_{2} \dots \eta_{N} \, . \end{equation} Notice that Eq.~(\ref{eq:teo_nilpotent}) automatically implements Pauli's exclusion principle for fermions. To compute Grassmann integrals of the form \begin{equation} \label{eq:teo_integral} \int \prod_{i=1}^{N} d\eta_{i} \, f(\eta) \, , \end{equation} the two following rules are sufficient \begin{equation} \label{eq:teo_integrationrules} \int d\eta_{i} = 0 \, , \qquad \int d\eta_{i} \,\eta_{i} = 1 \, , \end{equation} where the integration measures $\{d\eta_{i}\}$ obey the same anticommutation relations as the Grassmann variables $\eta_{i}$: \begin{equation} \label{eq:teo_anticommutation2} \{d\eta_{i}, d\eta_{j}\} \, = \, \{d\eta_{i}, \eta_{j}\} \, = \, 0 \, , \qquad \forall i,j \, . \end{equation} To illustrate the application of the rules (\ref{eq:teo_integrationrules}), we compute the integral \begin{equation} I[D] \, = \, \int \! \prod_{l = 1}^{N} d\eta_{l} d\overline{\eta}_{l} \ e^{ \, \sum_{i,j = 1}^{N} \overline{\eta}_{i} D_{i,j} \eta_{j} } \, , \label{eq:teo_gaussian} \end{equation} where we denoted the $2N$ Grassmann variables as $\eta_{1}, \dots \eta_{N}, \overline{\eta}_{1}, \dots, \overline{\eta}_{N}$. All the terms in the exponent of the integrand in Eq.~(\ref{eq:teo_gaussian}) are quadratic in the Grassmann variables. Hence, they commute among each other, and we can rewrite the sum over the index $i$ in the exponent as a product of exponentials: \begin{equation} I[D] \, = \int \! \prod_{l = 1}^{N} d\eta_{l} d\overline{\eta}_{l} \, \prod_{i=1}^{N} e^{ \overline{\eta}_{i} \sum_{j} D_{i,j} \eta_{j} } \, = \, \int \! \prod_{l = 1}^{N} d\eta_{l} d\overline{\eta}_{l} \, \prod_{i=1}^{N} \Big(1 + \overline{\eta}_{i} \sum_{j} D_{i,j} \eta_{j} \Big) \,. \label{eq:teo_gaussian2} \end{equation} In the second step we exploited the nilpotency of the Grassmann variables $\overline{\eta}_{i}$ and wrote the product of exponentials as a product of binomials. From the integration rules (\ref{eq:teo_integrationrules}) follows that the only terms of the product of binomials that survive after the integration are the ones in which all the Grassmann variables appear exactly once: \begin{align} I[D] \, & = \, \int \! \prod_{l = 1}^{N} d\eta_{l} d\overline{\eta}_{l} \, \prod_{i=1}^{N} \overline{\eta}_{i} \sum_{j} D_{i,j} \eta_{j} \nonumber\\ & = \, \int \! \prod_{l = 1}^{N} d\eta_{l} d\overline{\eta}_{l} \ \overline{\eta}_{1} D_{1,j_{1}} \eta_{j_1} \,\overline{\eta}_{2} D_{2,j_{2}} \eta_{j_2} \dots \overline{\eta}_{N} D_{N,j_{N}} \eta_{j_N}\, , \label{eq:teo_gaussian3} \end{align} where a sum over repeated indices $j_{k}$, $k=1,\dots,N$ is understood. The product of Grassmann variables in (\ref{eq:teo_gaussian3}) is antisymmetric under the exchange of any pair of indices $j_{l}$ and $j_{l'}$. Moreover, only the terms where all the indices $j_{1},\dots,j_{N}$ are different survive. Hence, we can rewrite (\ref{eq:teo_gaussian3}) using the completely antisymmetric epsilon tensor $\epsilon_{j_{1}j_{2}\dots j_{N}}$ in $N$ dimensions: \begin{align} I[D] \, = \, \int \! \prod_{l = 1}^{N} d\eta_{l} d\overline{\eta}_{l} \ \overline{\eta}_{l} \eta_{l} \sum_{j_{1},\dots,j_{N}} \epsilon_{j_{1}j_{2}\dots j_{N}} D_{1,j_{1}} D_{2,j_{2}} \dots D_{N,j_{N}} \, = \, \det D \, , \label{eq:teo_gaussian4} \end{align} where we used the standard formula for the determinant of a matrix $D$. Eq.~(\ref{eq:teo_gaussian4}) and analogous formulas are important when dealing with fermionic systems, as we will see in the next sections. \subsection{Naive discretization \label{sec:teo_naive}} We now come to the discussion of the discretization of the fermion action of QCD. In Euclidean space-time the fermionic part of the QCD action is given by: \begin{equation} \label{eq:teo_sfcontinuum} S_{F}[\psi,\overline{\psi},A] \ = \ \int \!\! d^{4}x \ \overline{\psi}(x) \, [\gamma_{\mu} \,(\partial_{\mu} + i A_{\mu}(x)) + m] \, \psi(x) \, , \end{equation} where we used the Einstein summation convention, and matrix-vector notation for the color and Dirac indices. $\overline{\psi}(x)$ and $\psi(x)$ are Dirac 4-spinors, whose entries are Grassmann variables, so that Fermi statistics is enforced for the quarks. The gluons are described by the gauge fields $A_{\mu} (x) \in$ su(3). $\gamma_{\mu}$, $\mu = 1,2,3,4$ are the Dirac $\gamma$-matrices, which in Euclidean space satisfy the Euclidean anti-commutation relations $\{\gamma_{\mu}, \gamma_{\nu}\} = 2 \delta_{\mu \nu} \mathbf{1}$. The action (\ref{eq:teo_sfcontinuum}) is invariant under the gauge transformations: \begin{align} \label{eq:teo_gaugetransformation1} &\overline{\psi}(x) \rightarrow \overline{\psi}'(x) = \overline{\psi} (x) \Omega(x)^{\dagger} \, ,&&\\ \label{eq:teo_gaugetransformation2} &\psi(x) \rightarrow \psi'(x) = \Omega(x) \psi (x) \, ,&& \\ \label{eq:teo_gaugetransformation3} &A_{\mu}(x) \rightarrow A'_{\mu} (x) = \Omega (x) A_{\mu}(x) \Omega (x)^{\dagger} + i (\partial_{\mu}\Omega (x)) \Omega (x)^{\dagger} \, ,&& \end{align} where $\Omega(x) \in $ SU(3). To carry out the lattice discretization of the fermion action (\ref{eq:teo_sfcontinuum}) we first start with considering the free case, i.e., we set $A_{\mu}(x) = 0$: \begin{equation} S_{F}^{free}[\psi,\overline{\psi}] \ = \ \int \! d^{4}x \ \overline{\psi}(x) \, [\gamma_{\mu} \,\partial_{\mu} + m] \, \psi(x) \, . \end{equation} As we outlined in Sec.~\ref{sec:teo_lattice}, the main idea behind the lattice discretization of a field theory is the introduction of a four-dimensional space-time lattice $\Lambda$: \begin{equation} \label{eq:teo_lattice} \Lambda \, = \, \{n = (n_1, n_2, n_3, n_4) \, |\, n_1,n_2,n_3 = 1,\dots, N-1 \, ; \ n_4 = 1,\dots, N_T-1 \} \, . \end{equation} The four-vector $n\in \Lambda$ labels the sites of the lattice, which are separated by the lattice spacing $a$. Then, the transition from the continuum to the lattice is effected by making the following substitutions \begin{flalign} &\psi(x) \rightarrow \psi (n) \, , \nonumber \\[0.5em] &\int \! d^{4}x \rightarrow a^{4} \sum_{n \in \Lambda} \, , \\ &\partial_{\mu} \psi (x) \rightarrow \dfrac{1}{2 a} (\psi(n + \hat{\mu}) - \psi(n - \hat{\mu})) \, .\nonumber \end{flalign} For the free fermion action on the lattice $\Lambda$ we obtain: \begin{equation} \label{eq:teo_sffree} S_{F}^{free}[\psi,\overline{\psi}] \ = \ a^4 \sum_{n} \overline{\psi}(n) \left( \sum_{\mu = 1}^{4} \gamma_{\mu} \dfrac{\psi(n + \hat{\mu}) - \psi(n - \hat{\mu}) }{2 a} + m \psi(x) \right) \, . \end{equation} Obviously we want to preserve the gauge invariance under the local transformations (\ref{eq:teo_gaugetransformation1}) and (\ref{eq:teo_gaugetransformation2}), which on the lattice are implemented by choosing $\Omega(n) \in$ SU(3) for each site $n$ and changing the fermion fields according to: \begin{equation} \label{eq:teo_gaugelattice} \overline{\psi}(n) \rightarrow \overline{\psi}'(n) = \overline{\psi} (n) \Omega(n)^{\dagger} \, , \qquad \psi(n) \rightarrow \psi'(n) = \Omega(n) \psi (n) \, . \end{equation} While the mass term in (\ref{eq:teo_sffree}) is invariant under (\ref{eq:teo_gaugelattice}), the discretized derivative is not: \begin{equation*} \overline{\psi}'(n) \psi' (n + \hat{\mu}) = \overline{\psi} (n) \Omega(n)^{\dagger} \Omega (n + \hat{\mu}) \psi (n + \hat{\mu}) \, . \end{equation*} This problem is overcome by introducing in the expression of the discretized derivative in (\ref{eq:teo_sffree}) the link variables $U_{\mu} (n) \in$ SU(3), which transform as \begin{equation} \label{eq:teo_gaugelattice2} U_{\mu}(n) \rightarrow U'_{\mu}(n) = \Omega(n) U_{\mu}(n) \Omega(n + \hat{\mu})^{\dagger} \, . \end{equation} Then, terms of the form $\overline{\psi}(n) U_{\mu}(n) \psi (n + \hat{\mu})$ are invariant under the rotation $\Omega (n)$ of the color indices: \begin{align*} \overline{\psi}'(n) U'_{\mu}(n) \psi' (n + \hat{\mu}) &= \overline{\psi} (n) \Omega(n)^{\dagger} \Omega(n) U_{\mu}(n) \Omega(n + \hat{\mu})^{\dagger} \Omega (n + \hat{\mu}) \psi (n + \hat{\mu}) \\ &= \overline{\psi}(n) U_{\mu}(n) \psi (n + \hat{\mu}) \, . \end{align*} As we mentioned, the variables $U_{\mu}(n)$ are attached to the links $(n,\mu)$ of the lattice, i.e., they connect the neighboring sites $n$ and $n + \hat{\mu}$. We can then define the hermitian conjugate $U_{\mu} (n)^{\dagger}$ to be the corresponding link variable with negative orientation, namely \begin{equation*} U_{\mu} (n)^{\dagger} \, \equiv \, U_{-\mu} (n + \hat{\mu}) \, . \end{equation*} For a graphical representation of $U_{\mu}(n)$ and its hermitian conjugate refer to Fig.~\ref{fig:teo_link}. \begin{figure} \vspace{2.5mm} \centering \includegraphics[scale=1.7]{chapters/plots/teo_link.pdf} \vspace{2.5mm} \caption{Graphical representation of the link variable $U_{\mu}(n)$ and its hermitian conjugate $U_{\mu}(n)^{\dagger}$. \label{fig:teo_link}} \end{figure} Putting things together we obtain the following discretized version of the fermion action: \begin{equation} S_{F}[\psi,\overline{\psi},U] \ = \ a^4 \sum_{n} \overline{\psi}(n) \left( \sum_{\mu = 1}^{4} \gamma_{\mu} \dfrac{U_{\mu}(n) \psi(n + \hat{\mu}) - U^{\dagger}_{\mu}(n - \hat{\mu}) \psi(n - \hat{\mu}) }{2 a} + m \psi(n) \right) \, . \label{eq:teo_naivesf} \end{equation} Eq.~(\ref{eq:teo_naivesf}) is often referred to as the naive discretization of fermions because it actually gives rise to the so-called \textit{doubling problem}, which we will sketch shortly. The discretized fermion action (\ref{eq:teo_naivesf}) may be rewritten as \begin{equation} S_{F}[\psi,\overline{\psi},U] \ = \ a^4 \sum_{n,m} \overline{\psi}(n) D(n|m) \psi(m) \, . \label{eq:teo_naivesf2} \end{equation} where \begin{equation} \label{eq:teo_naive} D(n|m)_{_{\alpha\beta}^{a\,b}} \, = \, \sum_{\mu = 1}^{4} (\gamma_{\mu})_{\alpha \beta} \dfrac{U_{\mu}(n)_{ab} \delta_{n + \hat{\mu},m} - U_{-\mu}(n)_{ab} \delta_{n -\hat{\mu}, m}}{2 a} + m \delta_{ab} \delta_{\alpha \beta} \delta_{n,m} \, , \end{equation} is the naive Dirac operator, with $a,b=1,2,3$ color labels and $\alpha, \beta = 1,2,3,4$ Dirac indices. The inverse $D(n|m)^{-1}$ of the naive Dirac operator (\ref{eq:teo_naive}) is the quark propagator. The quark propagator in momentum space $\tilde{D}(p)^{-1}$ can be computed by inverting the Fourier transform of the discretized Dirac operator (\ref{eq:teo_naive}). In doing so one finds that the momentum space propagator $\tilde{D}(p)^{-1}$ for massless ($m=0$) free fermions has the correct continuum limit when sending $a \to 0$: the propagator has a single pole which corresponds to the massless single fermion described by the continuum Dirac operator. On the lattice, however, the propagator for free fermions has 15 additional unphysical poles, the so-called doublers. \subsection{Wilson fermions} The first solution to the doubling problem was given by Wilson: an extra term, the so-called \textit{Wilson term}, is added to the naive Dirac operator (\ref{eq:teo_naive}). The Wilson term gives a mass to the doublers which is proportional to the inverse lattice spacing $a$. Hence, when $a \to 0$ the doublers become heavy and decouple from the theory. Moreover, in the limit $a \to 0$ the Wilson term vanishes, thus giving the correct continuum limit. In compact notation the Wilson Dirac operator is given by \begin{equation} \label{eq:teo_wilsonfermions} D(n|m)_{_{\alpha\beta}^{a\,b}} \, = \, \left( m + \dfrac{4}{a}\right) \delta_{ab} \, \delta_{\alpha \beta} \, \delta_{n,m} - \dfrac{1}{2a}\sum_{\mu = \pm 1}^{\pm 4} (\mathbf{1} - \gamma_{\mu})_{\alpha \beta} \, U_{\mu}(n)_{ab} \, \delta_{n + \hat{\mu},m} \, , \end{equation} with \begin{equation*} \gamma_{-\mu} = - \gamma_{\mu} \, , \quad \mu = 1, 2, 3, 4 \, . \end{equation*} It is easy to show that the Dirac operator (\ref{eq:teo_wilsonfermions}) is $\gamma_{5}$-hermitian, i.e., it satisfies \begin{equation} \label{eq:teo_gamma5} \gamma_{5} D \gamma_{5} = D^{\dagger} \, . \end{equation} Eq.~(\ref{eq:teo_gamma5}) implies that the eigenvalues of the Dirac operator (\ref{eq:teo_wilsonfermions}) are either real or come in complex conjugate pairs. We will use this property in Sec.~\ref{sec:teo_numerical} to show the feasibility of Monte Carlo simulations of lattice QCD with fermions. As a remark we stress that the additional term $4/a$ in the expression (\ref{eq:teo_wilsonfermions}) explicitly breaks chiral symmetry. The doubling problem of the naive Dirac operator (\ref{eq:teo_naive}) was therefore traded with the impossibility of studying the effects of the spontaneous chiral symmetry breaking with the Dirac operator (\ref{eq:teo_wilsonfermions}). \subsection{Staggered fermions} In this section we introduce another discretization of the fermion action (\ref{eq:teo_sfcontinuum}), so-called staggered fermions first proposed by Kogut and Susskind in \cite{Kogut1975}, which we will use in Chapters \ref{cha:su2} and \ref{cha:qcd}. The staggered fermion action reads \begin{equation} S_{F}[\psi,\overline{\psi},U] \ = \ a^4 \sum_{n} \overline{\psi}(n) \left( \sum_{\mu = 1}^{4} \eta_{\mu}(n) \dfrac{U_{\mu}(n) \psi(n + \hat{\mu}) - U^{\dagger}_{\mu}(n - \hat{\mu}) \psi(n - \hat{\mu}) }{2 a} + m \psi(n) \right) , \label{eq:teo_staggered} \end{equation} where we have introduced the staggered sign functions $\eta_{\mu}(n)$, which are defined as \begin{equation*} \eta_{1}(n) = 1 \, , \quad \eta_{2}(n) = (-1)^{n_{1}} \, , \quad \eta_{3}(n) = (-1)^{n_{1} + n_{2}} \, , \quad \eta_{4}(n) = (-1)^{n_{1} + n_{2} + n_{3}} \, . \end{equation*} The staggered sign functions $\eta_{\mu}(n)$ play the role of the $\gamma$-matrices: the Dirac indices $\alpha, \beta = 1,2,3,4$ are absent in the staggered formulation, and instead the spinor degrees of freedom are distributed on several sites of the lattice. When $m = 0$, the action (\ref{eq:teo_staggered}) is invariant under the global transformation \begin{equation} \label{eq:teo_chiral} \psi(n) \rightarrow e^{i \alpha \eta_{5}(n)} \psi(n) \, , \quad \overline{\psi}(n) \rightarrow \overline{\psi}(n) \, e^{i \alpha \eta_{5}(n)}\, , \quad \text{with } \eta_{5}(n) = (-1)^{n_{1} + n_{2} + n_{3} + n_{4}} \, , \end{equation} which implements chiral symmetry for staggered fermions ($\eta_{5}(n)$ plays the role of $\gamma_{5}$). We remark that the form (\ref{eq:teo_staggered}) of the fermion action has a residual doubling problem, since it describes not only one flavor of fermions but four \section{Gauge theories on the lattice \label{sec:teo_gauge}} In Euclidean space-time the continuum gluon action is given by \begin{equation} \label{eq:teo_sgcontinuum} S_{G}[A] \, = \, \dfrac{1}{2g^{2}} \int \! d^{4}x \Tr [F_{\mu\nu}(x) F_{\mu\nu}(x)] \, , \end{equation} where $g$ is the gauge coupling, $F_{\mu\nu} (x)$ is the field strength tensor defined as \begin{equation} \label{eq:teo_fmunu} F_{\mu\nu} (x) = \partial_{\mu} A_{\nu}(x) - \partial_{\nu} A_{\mu}(x) + i [A_{\mu}(x), A_{\nu}(x)] \, , \end{equation} and $A_{\mu}(x) \in$ su(3) are the gauge fields. In Sec.~\ref{sec:teo_naive} we have introduced the link variables $U_{\mu}(n)$ to make the fermion action (\ref{eq:teo_sffree}) invariant under the gauge transformations (\ref{eq:teo_gaugelattice}). The variables $U_{\mu}(n)$ are elements of the group SU(3), and absorb the rotations in color space of the fermion variables $\psi (n)$ and $\overline{\psi} (n)$. They thus have the role of the gauge fields on the lattice, and their connection to the continuum gauge fields is given by \begin{equation} \label{eq:teo_gaugetransporter} U_{\mu}(n) = \mathcal{P} \exp \Big( i \int_{n}^{n + \hat{\mu}}dx \, A_{\mu} (x) \Big) \simeq \exp \big( i a \, A_{\mu} (n) \big) \, , \end{equation} i.e., they are the path ordered exponential integrals of the gauge fields $A_{\mu}$ along the link $(n,\mu)$ connecting the neighboring sites $n$ and $n + \hat{\mu}$. In (\ref{eq:teo_gaugetransporter}) we approximated the integral along the link $(n,\mu)$ by the length $a$ of the path and the value $A_{\mu}(n)$ of the field at the starting point. The gauge transformations of the variables $U_{\mu}(n)$ are given in (\ref{eq:teo_gaugelattice2}). We can then obtain the transformation rules for the ordered product of link variables along a path $\mathcal{L}$ as: \begin{align} \label{eq:teo_closedpath} \prod_{(n,\mu) \in \mathcal{L}} U'_{\mu}(n) &= U'_{\mu_0}(n_0) U'_{\mu_1}(n_1) \dots U'_{\mu_k}(n_k) \\ \nonumber &= \Omega(n_0) U_{\mu_0}(n_0) \Omega(n_1)^{\dagger} \Omega(n_1) U_{\mu_1}(n_1) \Omega(n_2)^{\dagger} \dots \Omega(n_{k}) U_{\mu_k}(n_k) \Omega(n_{k+1})^{\dagger} \, , \end{align} where $(n_{i}, \mu_i)\in \mathcal{L}$, $i = 0,1,\dots, k$ are the links connecting site $n_{i}$ to site $n_{i+1}$ on the path $\mathcal{L}$. Eq.~(\ref{eq:teo_closedpath}) shows that the transformation matrices $\Omega(n_{i})$ and $\Omega(n_{i})^{\dagger}$ cancel for $i = 1, \dots, k$, and the whole product transforms like a gauge transporter: \begin{equation} \prod_{(n,\mu) \in \mathcal{L}} U'_{\mu}(n) = \Omega(n_{0}) \Bigg[ \prod_{(n,\mu) \in \mathcal{L}} U_{\mu}(n) \Bigg] \Omega(n_{k+1})^{\dagger} \, . \end{equation} Therefore we find that the trace of the ordered product of link variables along a closed path on the lattice is gauge invariant. Wilson, when giving the first formulation of lattice gauge theories in \cite{Wilson74}, used the smallest gauge-invariant objects one can build on the lattice, i.e., the traces of the plaquette variable, which is the ordered product of link variables around a plaquette (see Fig.~\ref{fig:teo_plaquette} for illustration): \begin{equation} \label{eq:teo_plaquette} U_{\mu\nu}(n) = U_{\mu}(n) U_{\nu}(n + \hat{\mu}) U_{\mu} (n + \hat{\nu})^{\dagger} U_{\nu} (n)^{\dagger} \, . \end{equation} \begin{figure} \vspace{2.5mm} \centering \includegraphics[scale=1.7]{chapters/plots/teo_plaquette.pdf} \vspace{2.5mm} \caption{Graphical illustration of the plaquette variable $U_{\mu\nu}(n)$. \label{fig:teo_plaquette}} \end{figure} The Wilson gauge action is a sum over all plaquettes, counted with only one orientation: \begin{equation} \label{eq:teo_wilson} S_{G}[U] = \dfrac{\beta}{3} \sum_{n \in \Lambda}\sum_{\mu<\nu} \R \Tr \big[\mathbf{1} - U_{\mu\nu}(n) \big] \, . \end{equation} We now show that (\ref{eq:teo_wilson}) reproduces the continuum form (\ref{eq:teo_sgcontinuum}) of the gauge action when the continuum limit $a \to 0$ is taken. To do so, we expand the link variables in the form (\ref{eq:teo_gaugetransporter}) for small $a$. Then we use the Baker-Campbell-Hausdorff formula \begin{equation} \label{eq:teo_bchformula} \exp (A) \exp (B) \, = \, \exp \big( A + B + \dfrac{1}{2} [A,B] + \dots \big) \end{equation} iteratively to rewrite the plaquette variables in the action (\ref{eq:teo_wilson}) as follows \begin{align} \label{eq:teo_plaquettecontinuum} U_{\mu\nu}(n) \, = \, \exp \Big(&ia A_{\mu}(n) + i a A_{\nu}(n + \hat{\mu}) - \frac{a^{2}}{2}[A_{\mu}(n), A_{\nu}(n + \hat{\mu})] \nonumber \\ &- ia A_{\mu} (n + \hat{\nu}) -i a A_{\nu} (n) - \frac{a^{2}}{2} [A_{\mu}(n + \hat{\nu}), A_{\nu}(n)] \nonumber \\ &+ \frac{a^{2}}{2} [A_{\mu}(n),A_{\mu}(n + \hat{\nu})] + \frac{a^{2}}{2} [A_{\nu}(n + \hat{\mu}), A_{\nu}(n)] \nonumber \\ &+ \frac{a^{2}}{2} [A_{\mu}(n),A_{\nu}(n)] + \frac{a^{2}}{2} [A_{\nu}(n + \hat{\mu}), A_{\mu}(n + \hat{\nu})] + \mathcal{O}(a^{3})\Big) \, . \end{align} If one then substitutes in (\ref{eq:teo_plaquettecontinuum}) the Taylor expansion for the shifted fields \begin{equation*} A_{\mu} (n + \hat{\nu}) = A_{\mu}(n) + a \partial_{\nu} A_{\mu}(n) + \mathcal{O}(a^{2}) \, , \end{equation*} many of the terms in Eq.~(\ref{eq:teo_plaquettecontinuum}) cancel and one obtains \begin{equation} \label{eq:teo_plaquettecontinuum2} U_{\mu\nu}(n) \, = \, \exp \Big(ia^{2} \big(\partial_{\mu} A_{\nu}(n) - \partial_{\nu} A_{\mu}(n) + i [A_{\mu}(n),A_{\nu}(n)] \big) + \mathcal{O}(a^{3})\Big) \, . \end{equation} Now we can insert the form (\ref{eq:teo_plaquettecontinuum2}) in the Wilson action, and expand the exponential to obtain \begin{equation} \label{eq:teo_continuumlimit} S_{G}[U] = \dfrac{\beta}{3} \sum_{n \in \Lambda} \sum_{\mu<\nu} \R \Tr [\mathbf{1} - U_{\mu\nu}(n)] = \dfrac{\beta}{6} \sum_{n \in \Lambda} \sum_{\mu,\nu} \Tr [F_{\mu\nu}(n)^{2}] + \mathcal{O}(a^{2}) \, . \end{equation} Eq.~(\ref{eq:teo_continuumlimit}) states that the Wilson action (\ref{eq:teo_wilson}) approximates the continuum form (\ref{eq:teo_sgcontinuum}) in the limit $a \to 0$ up to $\mathcal{O}(a^{2})$. Comparing (\ref{eq:teo_continuumlimit}) with (\ref{eq:teo_sgcontinuum}) we also find that the parameter $\beta$, the so-called \textit{inverse coupling}, is related to the gauge coupling $g$ by \begin{equation} \label{eq:teo_beta} \beta = \dfrac{6}{g^{2}} \, . \end{equation} We remark that this definition of $\beta$ is specific for the SU(3) case, while the general expression for SU(N) groups is $\beta = 2N/g^{2}$. Another important remark regards the continuum limit. The one we discussed in this chapter is usually referred to as "naive continuum limit" and is used as a guiding principle in the construction of lattice theories. For example in Chapter \ref{cha:pcm} we will use the same steps as described in Sec.~\ref{sec:teo_naive} to discretize the SU(2) principal chiral model with coupled chemical potentials. However, when taking the limit $a \to 0$ the bare parameters of the action, like the gauge coupling $g$ and the quark mass $m$, show a dependence on the lattice spacing $a$. Obviously, when sending $a \to 0$ the so-called \textit{running} of the bare parameters has to reproduce the correct physics. So, a physical observable $P(g(a), m(a), \dots, a)$ has to assume its physical value $P_0$ when taking the continuum limit: \begin{equation} \label{eq:teo_physical} \lim_{a \to 0} P(g(a), m(a), \dots, a) = P_0 \, . \end{equation} This requirement can be formulated by means of a differential equation \cite{Callan70,Symanzik1970} \begin{equation} \label{eq:teo_callan} \dfrac{dP(g,m,\dots,a)}{d\ln a} = 0 \, . \end{equation} When using Eq.~(\ref{eq:teo_callan}) for pure gauge theories one finds that the running coupling $g(a)$ decreases when the lattice spacing $a$ decreases, which reproduces the behavior of asymptotic freedom. Then, for pure gauge theories the true continuum limit is performed by sending $\beta \to \infty$, and at the same time changing the number of lattice points $N$ and $N_{T}$ such that the physical extents $ L = a N $ and $T = a N_T \, $ of the lattice remain the same. \section{Numerical simulations \label{sec:teo_numerical}} The partition function (\ref{eq:teo_partition}) of QCD on the lattice is given by \begin{equation} \label{eq:teo_partitionqcd} Z \, = \, \int D[\psi,\overline{\psi}] D[U] \, e^{-S_{F}[\psi,\overline{\psi}, U ] - S_{G}[U]} \, , \end{equation} where $S_{F}[\psi,\overline{\psi}, U]$ and $S_{G}[U]$ are some lattice discretization of the fermion action (\ref{eq:teo_sfcontinuum}) and the gauge action (\ref{eq:teo_sgcontinuum}), e.g., (\ref{eq:teo_naivesf2}) and (\ref{eq:teo_wilson}) respectively. The integrals over the fermion field configurations are the product of Grassmann measures on the sites $n$ of the lattice \begin{equation} D[\psi,\overline{\psi}] \equiv \prod_{n \in \Lambda} \prod_{\alpha, a} d \psi(n)_{_{a}^{\alpha}} \, d \overline{\psi}(n)_{_{a}^{\alpha}} \, , \end{equation} while the integrals over the gauge field configurations are the product of Haar measures on the links $(n, \mu)$ of the lattice \begin{equation} \label{eq:teo_haarmeasure} D[U] \equiv \prod_{n \in \Lambda} \prod_{\mu = 1}^{4} d U_{\mu}(n) \, . \end{equation} Eq.~(\ref{eq:teo_partitionqcd}) can be rewritten as \begin{equation} \label{eq:teo_partitionqcd2} Z \, = \, \int D[U] \, e^{- S_{G}[U]}\, Z_{F}[U]\, , \end{equation} where \begin{equation} \label{eq:teo_partitionf} Z_{F} [U] \, = \, \int D[\psi,\overline{\psi}] \, e^{-S_{F}[\psi,\overline{\psi}, U ]} \, . \end{equation} When $S_{F}[\psi,\overline{\psi}, U]$ is written as in Eq.~(\ref{eq:teo_naivesf2}), (\ref{eq:teo_partitionf}) has the form of the Grassmann integral (\ref{eq:teo_gaussian}) we computed in Sec.~\ref{sec:teo_grassmann}, and thus \begin{equation} \label{eq:teo_fermiondeterminant} Z_{F} [U] \, = \, \det D(U) \, , \end{equation} which is referred to as \textit{fermion determinant}. The $\gamma_{5}$-hermicity (\ref{eq:teo_gamma5}) of the Wilson Dirac operator (\ref{eq:teo_wilsonfermions}) implies that the fermion determinant (\ref{eq:teo_fermiondeterminant}) is real, a property that is crucial for the Monte Carlo simulations we will discuss shortly. Given the form (\ref{eq:teo_partitionqcd2}) of the partition function, expectation values of operators that only depend on the gauge fields may be computed as: \begin{equation} \label{eq:teo_expectationU} \langle O \rangle \, = \, \dfrac{1}{Z} \int D[U] O(U) \det D(U) e^{-S_{G}[U]} \, . \end{equation} In general, in the path integral formulation of quantum field theories expectation values are given by: \begin{equation} \label{eq:teo_expectation} \langle O \rangle \, = \, \dfrac{1}{Z} \int D[C] O(C) e^{-S[C]} \, , \end{equation} where with the compact notation $C$ we denote the fermion and gauge field configurations. The expression (\ref{eq:teo_expectation}) usually cannot be computed analytically. However, Monte Carlo (MC) simulations are very powerful tools to approximate the integral over all the field configurations in (\ref{eq:teo_expectation}). The key idea behind MC techniques is the interpretation of the Boltzmann weight $e^{-S[C]}/Z$ as a probabilistic weight $P(C)$. It is now clear why the reality of the fermion determinant (\ref{eq:teo_fermiondeterminant}) is so important: only if $P(C)$ is real and positive it can be interpreted as a probability. Then, the MC simulation approximates the expectation value (\ref{eq:teo_expectation}) by the average of the observable evaluated on $M$ sample field configurations $C_{i}$ distributed with probability $P(C_i) = e^{-S[C_{i}]}/Z$: \begin{equation} \label{eq:teo_average} \langle O\rangle_{M} \approx \dfrac{1}{M} \sum_{i=1}^{M} O(C_i) \, . \end{equation} From probability theory we know that $\langle O\rangle_{M}$ of Eq.~(\ref{eq:teo_average}) reproduces the correct value $\langle O\rangle$ (\ref{eq:teo_expectation}) when $M \to \infty$, with a statistical error that scales as $1/\sqrt{M}$. The field configurations $C_i$ distributed with probability $P(C_{i})$ are obtained as a Markov chain: a memoryless stochastic process, i.e., a process for which the probability to get the next configuration $C'$ depends only on the present state $C$ and not on the other configurations $C_{i}$ generated in the Markov chain. This memoryless property goes under the name of Markov property and it requires the transition probability $T(C'|C)$ to depend only on the states $C'$ and $C$ and not on the label $i$. To obtain correct results the Markov chain has to satisfy two important properties. It has to be ergodic, i.e., it must be possible for the process to access all configurations, and the transition probability $T(C'|C)$ has to satisfy the balance equation: \begin{equation} \label{eq:teo_balance} \sum_{C} T(C'|C) P(C) = \sum_{C} T(C|C') P(C') \, . \end{equation} One way of implementing the Markov Chain in a simulation is the Metropolis algorithm \cite{Metropolis:1953am}, which uses a sufficient condition as a solution of the balance equation (\ref{eq:teo_balance}), called \textit{detailed balance}: \begin{equation} \label{eq:teo_detailedbalance} T(C'|C) P(C) = T(C|C') P(C') \, . \end{equation} The Metropolis algorithm consists of the following steps: \begin{itemize} \item[1:] A candidate configuration $C'$ is chosen according to an a priori selection probability $T_{0}(C'|C)$, with $C = C_{n-1}$. \item[2:] The configuration $C'$ is accepted as the new configuration $C_{n}$ of the Markov chain with the acceptance probability \begin{equation*} T_{A}(C'|C) = \min \Big(1, \dfrac{T_{0}(C'|C) e^{-S[C']}}{T_{0}(C'|C) e^{-S[C]}}\Big) \, . \end{equation*} \item[3:] Steps 1 and 2 are repeated. \end{itemize} The change from the field configuration $C_n$ to the field configuration $C_{n+1}$ is an \textit{update} of the MC simulation. In general one initializes the simulation with an arbitrary field configuration, and then the observables are calculated as in (\ref{eq:teo_average}) only after a sufficient number of equilibration and decorrelation updates. In the next section we will see that the introduction of a chemical potential $\mu$ coupled to the net-quark number $\mathcal{N}$ causes the fermion determinant $Z_{F}[U]$ to be complex, thus preventing the use of MC techniques. \section{Finite chemical potential and the sign problem \label{sec:teo_sign}} \subsection{Introducing the chemical potential on the lattice \label{sec:teo_chem}} In the continuum, the net-quark number $\mathcal{N}$ is given by the spatial volume integral of the temporal component $\overline{\psi} \gamma_{4} \psi$ of the Noether current $\overline{\psi} \gamma_{\mu} \psi$, i.e., it is a conserved charge. As such, it can be coupled to a \textit{quark chemical potential} $\mu$. The introduction of the quark chemical potential $\mu$ on the lattice is implemented by replacing the temporal hopping terms in (\ref{eq:teo_wilsonfermions}) with \begin{equation} \label{eq:teo_chemical} -\dfrac{1}{2a} \left( e^{a \mu} (\mathbf{1} - \gamma_{4})_{\alpha \beta} \, U_{4}(n)_{ab} \, \delta_{n + \hat{4},m} + e^{-a \mu} (\mathbf{1} + \gamma_{4})_{\alpha \beta} \, U_{4}(n - \hat{4})_{ab}^{\dagger} \, \delta_{n - \hat{4},m} \right) \, . \end{equation} In (\ref{eq:teo_chemical}) the forward propagation in the time direction is favored by a factor $e^{a \mu}$ while the backward propagation in time is suppressed by a factor $e^{-a \mu}$, thus giving the desired asymmetry between particles and anti-particles. However, the introduction of the chemical potential on the lattice causes a serious technical problem: for $\mu \neq 0$ the Dirac operator (\ref{eq:teo_wilsonfermions}) is no longer $\gamma_5$-hermitian. It instead satisfies the property \begin{equation} \label{eq:teo_nogamma5} D(-\mu^{\star})^{\dagger} \, = \, \gamma_{5} D(\mu) \gamma_{5} \, , \end{equation} which implies that the eigenvalues of the Dirac operator $D$ no longer come in complex conjugate pairs. This means that the fermion determinant $\det D$ is now complex, and thus it cannot be used as a probabilistic weight in a Monte Carlo simulation. This is known as the \textit{complex action problem} or \textit{sign problem}, and it has been preventing the study of the QCD phase diagram from first principle lattice calculations in the region of non-vanishing chemical potential. \subsection{Approaches to the sign problem \label{sec:teo_approaches}} Over the years a wide variety of methods have been developed in order to overcome the sign problem on the lattice. Here we discuss some of them briefly, and we refer the reader to the reviews at the annual lattice conferences, e.g., \cite{Chandrasekharan:2008gp, deForcrand:2010ys, Wolff:2010zu, Aarts:2013lcm, Gattringer:2014nxa, Sexty:2014dxa, Borsanyi:2015axp} for a more complete overview of the topic. \subsubsection{Reweighting \label{sec:teo_reweighting}} As we outlined in the previous section, the introduction of the chemical potential causes the fermion determinant $\det D$ to be complex. Using the notation of Sec.~\ref{sec:teo_numerical}, this means that $P(C)$ is now complex, i.e., $P(C) = |P(C)| e^{i \varphi (C)}$ and thus it cannot be interpreted directly as a probability. In reweighting, subsequent configurations $C'$ of the Markov chain are generated using the modulus $|P(C)|$. Then the expectation values (\ref{eq:teo_expectation}) are estimated as \begin{equation} \label{eq:teo_reweighting} \langle O \rangle \, = \, \dfrac{\int D[C] |P(C)| e^{i \varphi(C)} O(C)}{\int D[C] |P(C)| e^{i \varphi(C)} } \, = \, \dfrac{\langle e^{i \varphi(C)} O \rangle_{R} }{\langle e^{i \varphi(C)} \rangle_{R}} \, , \end{equation} where with the angular brackets with the subscript $R$ we denote the average (\ref{eq:teo_average}) computed over a set of configurations generated with the real probability distribution $|P(C)|$. When the average $\langle e^{i \varphi(C)} \rangle_{R}$ is vanishing, it indicates strong oscillations of the phases which cause huge cancellations in MC simulations and the sign problem is said to be severe. Moreover, $\langle e^{i \varphi(C)} \rangle_{R}$ can be computed as \begin{equation} \label{eq:teo_phase} \langle e^{i \varphi(C)} \rangle_{R} = \dfrac{Z}{Z_{R}} = e^{\ln Z - \ln Z_{R}} = e^{- \beta V \Delta f}\, , \end{equation} where $Z = \int D[C] P(C)$, $Z_{R} = \int D[C] |P(C)|$ and $\Delta f$ is the difference in free energy density of the two partition sums. Eq.~(\ref{eq:teo_phase}) means that, even when $\langle e^{i \varphi(C)} \rangle_{R}$ is finite, it decays exponentially with the volume $V$. Therefore, reweighting can only be applied to systems at small volumes and whose sign problem is mild. \subsubsection{Density of states \label{sec:teo_density}} The density of states is formally defined as \begin{equation} \label{eq:teo_density} \rho(E) \, = \, \int D[\phi] \delta \big( S[\phi] - E\big) \, . \end{equation} Using (\ref{eq:teo_density}) one can rewrite the partition sum as: \begin{equation*} Z \, = \, \int D[\phi] e^{-S[\phi]} = \int dE \rho (E) e ^{-E} \, , \end{equation*} and expectation values as \begin{equation*} \langle O \rangle \, = \, \dfrac{1}{Z} \int dE \, O(E) \, \rho (E) \, e ^{-E} \, . \end{equation*} Therefore, if $\rho(E)$ is determined numerically, $Z$ and $\langle O \rangle$ can be computed as numerical integrals of a single variable. There exist different methods for the numerical determination of the density, e.g., the LLR method \cite{Langfeld:2012ah}, and the FFA approach \cite{Gattringer:2015lra}. For a more detailed review over the density of states method in the context of the complex action problem see, e.g., \cite{Langfeld:2016kty,Gattringer:2016kco}. \subsubsection{Complex Langevin \label{sec:teo_langevin}} The Langevin method, as firstly formulated by Parisi and Wu in \cite{Parisi:1980ys}, treats Euclidean quantum field theory as the equilibrium limit of a statistical system coupled to a thermal reservoir. The system evolves in a fictitious time direction $t$ until it reaches equilibrium as $t\to \infty$. The evolution of the field is described by a stochastic differential equation, the Langevin equation, where the coupling with the heat reservoir is simulated by means of a stochastic noise field $\eta(t)$. In the equilibrium limit stochastic averages become identical to Euclidean expectation values \cite{DAMGAARD1987227}. The idea of applying this method to system for which MC methods are inapplicable came shortly after, \cite{PARISI1983393,Klauder:1983nn}, and complex Langevin was born. However, also this method faces technical problems, which mainly pertain convergence: sometimes it fails to reach convergence and sometimes it converges to a wrong limit. For a review of the status of complex Langevin see, e.g., \cite{Seiler:2017wvd} and references therein. \section{The dual approach \label{sec:teo_duality}} The dual approach consists in exactly rewriting the partition sum of a system in terms of new variables, so-called \textit{dual variables}. The advantages of reformulating field theories in terms of new degrees of freedom are known since long. For example, already almost 80 years ago Kramers and Wannier were able to determine the critical point of the two-dimensional Ising model using a duality transformation \cite{Kramers:1941zz}. What has been emerging since then is that different representations highlight different aspects of a system, and thus maybe better suited for different scopes (see, e.g., the review on duality \cite{Savit:1979ny}). In more recent years, in the framework of lattice field theory, the dual approach has been successful in solving the sign problem of a variety of abelian field theories (see, e.g., \cite{Gattringer:2014nxa}). As a general strategy, the Boltzmann weight is factorized into local exponentials which are then expanded into Taylor series. The expansion coefficients substitute the conventional degrees of freedom in the description of the system once the integration over the conventional fields is performed. The resulting partition function is a sum over the contributions of the admissible configurations of the expansion indices, i.e., the dual variables. Sometimes all those contributions are real and positive and, for those cases, the dual formulation of the partition sum is suitable for Monte-Carlo simulations and the sign problem is solved. As an explanatory example for the application of the dual approach to the sign problem of abelian theories we present the U(1) Gauge -- Higgs model \cite{Mercado:2013ola}. We also use this example to outline some important features of the dual approach, such as the emergence of constraints in the dual representation and their connection with the original symmetry of the system. \subsection{The U(1) Gauge -- Higgs model \label{sec:teo_u1}} On the lattice the action of the U(1) Gauge -- Higgs model can be written as the sum of the U(1) Wilson gauge action $S_{G}[U]$ and of the discretized action $S_{M}[U, \phi]$ for the Higgs field: \begin{equation} S[U, \phi] \, = \, S_{G}[U] \, + \, S_{M}[U, \phi] \, . \end{equation} The gauge action reads \begin{equation} \label{eq:teo_sg} S_{G}[U] \, = \, - \dfrac{\beta}{2} \sum_{x} \sum_{\nu < \rho} \big[U_{x,\nu\rho} + U_{x,\nu\rho}^{\star} \big] \, , \end{equation} where the plaquette variable \begin{equation} \label{eq:teo_plaquetteu1} U_{x,\nu\rho} \equiv U_{x,\nu} U_{x + \hat{\nu}, \rho} U_{x+ \hat{\rho}, \nu}^{\star} U_{x,\rho}^{\star}\, , \quad \end{equation} is the oriented product of the link variables $U_{x,\nu} \in U(1)$ around the plaquette $(x,\nu\rho)$. The action for the matter fields has the form \begin{equation} \label{eq:teo_sm} S_{M}[U,\phi] \, = \, \sum_{x} \Big[ \eta |\phi_{x}|^{2} + \lambda |\phi_{x}|^{4} - \sum_{\nu} \big( e^{\mu \delta_{\nu,4}} \phi_{x}^{\star} U_{x,\nu} \phi_{x+\hat{\nu}} + e^{-\mu \delta_{\nu,4}} \phi_{x+\hat{\nu}}^{\star} U_{x,\nu}^{\star} \phi_{x} \big)\Big] \, , \end{equation} where $\phi_{x} \in \mathbb{C}$ are the charged scalar Higgs fields attached to the sites $x$ of the lattice, the parameter $\eta$ denotes $8 + m^{2}$, where $m$ is the bare mass and $\lambda$ is the coupling of the quartic interaction. In (\ref{eq:teo_sm}) we also coupled a chemical potential $\mu$, which gives different weight to the forward and backward propagation of the $\phi_{x}$ fields, thus making the action $S_{M}[U,\phi]$ complex. Hence, the U(1) Gauge -- Higgs model suffers from the complex action problem in the conventional representation. The partition function of the model is given by \begin{equation} \label{eq:teo_partitionu1} Z \, = \, \int D[U] D[\phi] e^{-S_{G}[U] {-S_{M}[U,\phi]}} \, , \end{equation} where \begin{equation} \label{eq:teo_haar} \int D[U] \equiv \prod_{x,\nu} \int_{U(1)} dU_{x,\nu} \end{equation} is the integration over the U(1) Haar measure, while \begin{equation} \int D[\phi] \equiv \prod_{x} \int_{\mathbb{C}} \dfrac{d\phi_{x} }{2\pi} = \prod_{x} \int_{0}^{\infty} d r_{x} r_{x} \int_{0}^{2\pi} \dfrac{d\theta_{x}}{2\pi} \, , \qquad \text{with } \ \phi_{x} = r_{x} e^{i \theta_{x}} \, , \end{equation} is the integral over the complex plane for the matter fields. We start the dualization of the partition sum (\ref{eq:teo_partitionu1}) considering just the matter fields. The Boltzmann weight for the matter fields can be factorized as follows: \begin{equation} \label{eq:teo_esm} e^{-S_{M}[U,\phi]} = \Bigg[\prod_{x} e^{- \eta |\phi_{x}|^{2} - \lambda |\phi_{x}|^{4}} \Bigg] \Bigg[ \prod_{x,\nu} \exp \big( e^{\mu \delta_{\nu,4}} \phi_{x}^{\star} U_{x,\nu} \phi_{x+\hat{\nu}} \big) \exp \big( e^{-\mu \delta_{\nu,4}} \phi_{x+\hat{\nu}}^{\star} U_{x,\nu}^{\star} \phi_{x} \big) \Bigg] \, , \end{equation} i.e., we rewrite the exponential of sums as a product of exponentials. Then we expand the factors for the nearest neighbor terms: \begin{align} &\prod_{x,\nu} \exp \big( e^{\mu \delta_{\nu,4}} \phi_{x}^{\star} U_{x,\nu} \phi_{x+\hat{\nu}} \big) \exp \big( e^{-\mu \delta_{\nu,4}} \phi_{x+\hat{\nu}}^{\star} U_{x,\nu}^{\star} \phi_{x} \big)\nonumber \\ &\; = \prod_{x,\nu} \sum_{n_{x,\nu} = 0}^{\infty} \dfrac{\big( e^{\mu \delta_{\nu,4}} \phi_{x}^{\star} U_{x,\nu} \phi_{x+\hat{\nu}} \big)^{n_{x,\nu}} }{n_{x,\nu}! } \sum_{\overline{n}_{x,\nu} = 0}^{\infty} \dfrac{\big( e^{-\mu \delta_{\nu,4}} \phi_{x+\hat{\nu}}^{\star} U_{x,\nu}^{\star} \phi_{x} \big)^{\overline{n}_{x,\nu}}}{\overline{n}_{x,\nu}!} \nonumber \\ &\; = \sum_{\{n, \overline{n}\}} \Bigg[ \prod_{x, \nu} \dfrac{\big(U_{x,\nu} \big)^{n_{x,\nu}} \big(U_{x,\nu}^{\star} \big)^{\overline{n}_{x,\nu}}}{n_{x,\nu}! \overline{n}_{x,\nu}!} \Bigg] \Bigg[ \prod_{x} e^{\mu [n_{x,4} - \overline{n}_{x,4}]} \phi_{x}^{\star \sum_{\nu}[n_{x,\nu} + \overline{n}_{x- \hat{\nu},\nu}]} \phi_{x}^{ \sum_{\nu}[\overline{n}_{x,\nu} + n_{x- \hat{\nu},\nu}]}\Bigg] \, . \label{eq:teo_linkfactors} \end{align} In the first step each local exponential is expanded in a Taylor series. The expansion coefficients are the link variables $n_{x,\nu} \in \mathbb{N}_{0}$ for the forward hops and $\overline{n}_{x,\nu} \in \mathbb{N}_{0}$ for the backward hops. These integer valued variables will be the dynamical degrees of freedom once we will integrate out the conventional fields. In the last line of (\ref{eq:teo_linkfactors}) the terms are simply reorganized, and the notation $\sum_{\{n, \overline{n}\}}$ is introduced to denote the sum over all configurations of the expansion variables, i.e., \begin{equation*} \sum_{\{n, \overline{n}\}} \equiv \prod_{x,\nu} \sum_{n_{x,\nu} = 0}^{\infty} \sum_{\overline{n}_{x,\nu} = 0}^{\infty} \, . \end{equation*} We now insert the expansions (\ref{eq:teo_esm}) and (\ref{eq:teo_linkfactors}) in Eq.~(\ref{eq:teo_partitionu1}) and we write the complex fields as $\phi_{x} = r_{x} e^{i \theta_{x}}$ and the integration measure $\int D[\phi]$ in polar coordinates. For the partition sum we obtain: \begin{align} Z &= \sum_{\{n, \overline{n}\}} \Bigg[ \prod_{x, \nu} \dfrac{1}{n_{x,\nu}! \overline{n}_{x,\nu}!} \Bigg] \int D[U] e^{-S_{G}[U]} \Bigg[ \prod_{x,\nu} \big(U_{x,\nu} \big)^{n_{x,\nu} - \overline{n}_{x,\nu}} \Bigg] \nonumber \\ &\hspace{9mm} \times \Bigg[ \prod_{x} e^{\mu [n_{x,4} - \overline{n}_{x,4}]} \int_{0}^{\infty} d r_{x} r_{x}^{1 + \sum_{\nu}[(n_{x,\nu} + \overline{n}_{x,\nu})+ (n_{x- \hat{\nu},\nu} + \overline{n}_{x- \hat{\nu},\nu})]} e^{- \eta r_{x}^{2} - \lambda r_{x}^{4}} \Bigg] \nonumber \\ &\hspace{9mm} \times \Bigg[ \prod_{x} \int_{0}^{2\pi} \dfrac{d \theta_{x}}{2 \pi} e^{-i \theta_{x} \sum_{\nu}[(n_{x,\nu} - \overline{n}_{x,\nu})- (n_{x- \hat{\nu},\nu} - \overline{n}_{x- \hat{\nu},\nu})]} \Bigg] \, . \label{eq:teo_partition2} \end{align} The integration over the Higgs fields leads to two types of contributions: the integrals over the phases $\theta_{x}$ give rise to Kronecker deltas, which impose constraints on the variables $n_{x,\nu} \in \mathbb{N}_{0}$ and $\overline{n}_{x,\nu} \in \mathbb{N}_{0}$ at all sites $x$. The integrals over the radial coordinates $r_{x}$ together with the factorials in the first line of (\ref{eq:teo_partition2}) give rise to weights for the configurations $\{n,\overline{n}\}$ of the Higgs fields in the dual representation. To simplify the structure of the constraints it is useful to perform the change of variables: \begin{equation} \label{eq:teo_k} \begin{array}{ll} n_{x,\nu} - \overline{n}_{x,\nu} = k_{x,\nu} \, , \qquad &k_{x,\nu} \in \mathbb{Z} \, , \\ n_{x,\nu} + \overline{n}_{x,\nu} = |k_{x,\nu}| + 2 l_{x,\nu} \, , \qquad &l_{x,\nu} \in \mathbb{N}_{0} \, . \end{array} \end{equation} Then the partition function is given by \begin{equation} Z = \sum_{\{k, l\}} C_{M}[k] \, W_{M}[k,l] \int D[U] e^{-S_{G}[U]} \Bigg[ \prod_{x,\nu} \big(U_{x,\nu} \big)^{k_{x,\nu}} \Bigg] \, . \label{eq:teo_partition3} \end{equation} The dual variables for the Higgs fields are now the link based variables $k_{x,\nu} \in \mathbb{Z}$ and $l_{x,\nu} \in \mathbb{N}_{0}$ defined in (\ref{eq:teo_k}). The partition function sums over the configurations of those dual variables: \begin{equation*} \sum_{\{k, l\}} \equiv \prod_{x,\nu} \sum_{k_{x,\nu} = -\infty}^{+\infty} \sum_{l_{x,\nu} = 0}^{\infty} \, . \end{equation*} The configurations of the $k_{x,\nu}$ variables are constrained by the product of Kronecker deltas in the function $C_{M}[k]$: \begin{equation} \label{eq:teo_cm} C_{M}[k] = \prod_{x} \delta \Big( \vec{\nabla} \vec{k}_{x} \Big) \, , \end{equation} where $\vec{\nabla} \vec{k}_{x}$ is the discretized divergence of the link variables $k_{x,\nu}$, which explicitly is given by \begin{equation} \label{eq:teo_divergence} \vec{\nabla} \vec{k}_{x} \equiv \sum_{\nu} [k_{x,\nu} - k_{x- \hat{\nu},\nu}] \, . \end{equation} The constraint $C_{M}[k]$ in (\ref{eq:teo_cm}) requires the discretized divergence of the $k_{x,\nu}$ variables to vanish at every site $x$ of the lattice, i.e., it imposes the conservation of the $k$-fluxes. As a result the admissible configurations for the $k_{x,\nu}$ are closed \textit{worldlines}. We will see in Chapter \ref{cha:pcm} that analogous constraints arise from the abelian color flux dualization of the principal chiral model. In general constraints of the type (\ref{eq:teo_cm}) are flux conservation constraints that result from the integration of the phases in the parametrization of the conventional fields. On the other hand the $l_{x,\nu}$ are unconstrained variables that only contribute to the weight function $W_{M}[k,l]$: \begin{equation} \label{eq:teo_wm} W_{M}[k,l] = \prod_{x, \nu} \dfrac{1}{(|k_{x,\nu}|+ l_{x,\nu})! l_{x,\nu}!} \prod_{x} e^{\mu k_{x,4}} P \Big( \sum_{\nu} \big[ |k_{x,\nu}| + |k_{x - \hat{\nu},\nu}| + 2(l_{x,\nu} + l_{x - \hat{\nu},\nu}) \big]\Big) \, , \end{equation} where $P(n)$ are the elementary integrals \begin{equation} P(n) = \int_{0}^{\infty} d x \, x^{1 + n} \, e^{- \eta x^{2} - \lambda x^{4}} \, , \end{equation} which can easily be computed numerically. From Eq.~(\ref{eq:teo_wm}) we can also read off the $\mu$-dependence: the chemical potential $\mu$ multiplies the $k_{x,\nu}$ variables in the temporal direction $\nu = 4$. Since the $k_{x,\nu}$ variables are constrained to form closed loops by $C_{M}[k]$, only temporal winding loops couple to the chemical potential. For those loops $\sum_{x} k_{x,4}$ is equal to the extent of the lattice in the temporal direction $N_{T}$, i.e., the inverse temperature $\beta_{T}$, times the net temporal winding number of the loop $W_{T}$. If one then compares the resulting expression for the $\mu$ dependence $e^{\mu \beta_{T} W_{T}}$ with the usual form $e^{\mu \beta_{T} \mathcal{N}}$ for the coupling with the net-particle number $\mathcal{N}$, we can identify \begin{equation} W_{T} = \mathcal{N} \, . \end{equation} In other words, in the dual representation the net-particle number $\mathcal{N}$ is interpreted as the topological quantity $W_{T}$, which denotes the total net temporal winding of a given configuration. This is one of the most beautiful outcomes of the dualization procedure, as the net temporal winding number is much easier to determine than the net-particle number. We will see that a similar interpretation will arise also in the abelian color flux dualization of the fermion action of QCD. To complete the reformulation of the partition function $Z$ in (\ref{eq:teo_partition2}) we now focus on the dualization of the gauge fields. The gauge action $S_{G}[U]$ given in (\ref{eq:teo_sg}) is a sum over plaquettes. Hence, analogously to what we have done for the Higgs field, we may rewrite the Boltzmann weight $e^{- S_{G}[U]}$ as a product over plaquettes: \begin{align} e^{-S_{G}[U]} &= \prod_{x, \nu < \rho} e^{\frac{\beta}{2} \big(U_{x,\nu\rho} + U_{x,\nu\rho}^{\star} \big)} = \prod_{x, \nu < \rho} \sum_{p_{x,\nu\rho} = - \infty}^{+\infty} I_{p_{x,\nu\rho}}(\beta) \, U_{x,\nu\rho}^{\ p_{x,\nu\rho}} \nonumber \\ &= \sum_{\{p\}} \Bigg[\prod_{x, \nu < \rho} I_{p_{x,\nu\rho}}(\beta)\Bigg] \Bigg[\prod_{x, \nu } U_{x,\nu}^{ \ \sum_{\rho > \nu}[p_{x,\nu\rho} - p_{x - \hat{\rho},\nu\rho}] - \sum_{\sigma < \nu}[p_{x,\sigma\nu} - p_{x - \hat{\sigma},\sigma\nu}]} \Bigg] \, . \label{eq:teo_gauge} \end{align} In the second step we used the definition of the generating function of the modified Bessel functions \cite{Olver:2010:NHM:1830479}, \begin{equation*} e^{\frac{z}{2} (t + t^{-1})} = \sum_{m = - \infty}^{+\infty} t^{m} I_{m}(z) \, , \end{equation*} thus introducing the expansion variables $p_{x,\nu\rho} \in \mathbb{Z}$ attached to the plaquettes $(x,\nu\rho)$. We will refer to those variables as \textit{plaquette occupation numbers}. Then we inserted the explicit expression for the plaquettes (\ref{eq:teo_plaquetteu1}) in terms of the link variables and reorganized the product to find the powers of the link variables. We now insert the expression (\ref{eq:teo_gauge}) of the gauge Boltzmann weight inside Eq.~(\ref{eq:teo_partition3}). To obtain the final form of the dualized partition function (\ref{eq:teo_partition3}) we just need to solve the gauge integrals: \begin{align} C_{G}[p,k] &= \prod_{x, \nu } \int dU_{x,\nu} \, U_{x,\nu}^{ \ \sum_{\rho > \nu}[p_{x,\nu\rho} - p_{x - \hat{\rho},\nu\rho}] - \sum_{\sigma < \nu}[p_{x,\sigma\nu} - p_{x - \hat{\sigma},\sigma\nu}] + k_{x,\nu}} \nonumber \\ &= \prod_{x, \nu } \delta \Big( \sum_{\rho > \nu}[p_{x,\nu\rho} - p_{x - \hat{\rho},\nu\rho}] - \sum_{\sigma < \nu}[p_{x,\sigma\nu} - p_{x - \hat{\sigma},\sigma\nu}] + k_{x,\nu}\Big) \, . \label{eq:teo_cg} \end{align} Again the integration over the conventional degrees of freedom leads to constraints for the dual variables in the form of Kronecker deltas. In particular (\ref{eq:teo_cg}) requires the total net link flux to be vanishing at every link $(x,\nu)$. The flux on each link of the lattice receives contributions both from the $k_{x,\nu} \in \mathbb{Z}$ variables, which describe the Higgs fields in the dual representation, and the $p_{x,\nu\rho} \in \mathbb{Z}$ dual variables for the $U(1)$ gauge degrees of freedom. When the Higgs degrees of freedom are absent, i.e., in the pure gauge case, (\ref{eq:teo_cg}) implies that the only admissible gauge configurations are closed surfaces of cycle occupation numbers, also called \textit{worldsheets}. Putting things together we get the dual form of the partition function \begin{equation} \label{eq:teo_dualpartitionu1} Z = \sum_{\{p,k,l\}} C_{M}[k] \, W_{M}[k,l] \, C_{G} [p,k] \, W_{G}[p] \, . \end{equation} The sum $\sum_{\{p,k,l\}}$ runs over the configurations of the plaquette occupation numbers $p_{x,\nu\rho} \in \mathbb{Z}$, the dual variables for the Higgs $k_{x,\nu} \in \mathbb{Z}$ and the auxiliary variables $l_{x,\nu} \in \mathbb{N}_{0}$. The $k_{x,\nu}$ variables must satisfy the flux conservation constraint $C_{M}[k]$ given in (\ref{eq:teo_cm}), as well as the gauge constraint $C_{G}[p,k]$ in (\ref{eq:teo_cm}) together with the $p_{x,\nu\rho}$ variables. These two constraints allow us to interpret the dual variables for matter as worldlines and the dual variables for the gauge degrees of freedom as worldsheets. This kind of interpretation will recur also in the dualized theories we will present in this thesis. The weight factor $W_{M}[k,l]$ for the configurations of the matter flux variables was given in (\ref{eq:teo_wm}) while the weight for the gauge configurations is given by \begin{equation} \label{eq:teo_wg} W_{G}[p] = \prod_{x, \nu < \rho} I_{p_{x,\nu\rho}}(\beta) \, . \end{equation} Both $W_{M}[k,l]$ and $W_{G}[p]$ are real and positive factors. This means that the dual formulation (\ref{eq:teo_dualpartitionu1}) of the partition function is suitable for Monte Carlo simulations, and thus the sign problem is solved. Obviously, in order for the simulation to be efficient one must use simulation strategies that take into account the constraints (\ref{eq:teo_cm}) and (\ref{eq:teo_cg}). We refer the interested reader to the original paper \cite{Mercado:2013yta}, where the numerics are described in detail. \chapter{Dual representation for SU(2) lattice gauge theory with staggered fermions \label{cha:su2}} As we outlined in Sec.~\ref{sec:teo_sign}, QCD at finite quark density is not amenable to lattice Monte Carlo simulations because of the sign problem. One of the tactics that has been used to circumvent this problem is to consider non-abelian gauge theory with fermions, in which the color group is SU(2) rather than SU(3). The obvious advantage of studying this theory, known as \textit{two color QCD} or \textit{QC$_{2}$D}, is the absence of the complex action problem, due to the pseudo-reality of SU(2). Moreover, QC$_2$D is of particular interest since it shares most of the salient features of QCD, such as confinement and dynamical chiral symmetry breaking. Consequently, the phase structure of QC$_2$D has been extensively studied using lattice methods \cite{NAKAMURA1984391, DAGOTTO1986421, Kogut:2001na, Hands:2006ve, Braguta:2016cpw}. Our motivation for studying the SU(2) lattice gauge theory lies on a slightly different ground with respect to the studies we just mentioned, but it shares the same final goal of studying the phase diagram of lattice QCD from first principle calculations. Our aim was to develop a dual approach that would enable us to reformulate non-abelian gauge theories in terms of new dual variables, and SU(2) looked like the perfect playground to create and refine our techniques. Duality transformations are important tools for understanding quantum field theories, since often physical phenomena are better described by changing the representation. In particular, when using numerical simulations in the framework of lattice field theories, novel representations may give rise to new simulation strategies which allow one to explore parameter regions that were not accessible before (see, e.g., the reviews on worldline and dual representations in lattice field theories \cite{Chandrasekharan:2008gp,deForcrand:2010ys,Gattringer:2014nxa}). In the past the dualization of non-abelian lattice field theories has often involved the use of strong coupling expansion techniques based on the character expansion \cite{DROUFFE1978133,Savit:1979ny}. In the specific case of SU(2), exact dualizations were found, e.g., in 3 dimensions in terms of $6j$ symbols \cite{PhysRevLett.65.813,Anishetty:1992xa} or in the framework of spin foams \cite{Oeckl:2000hs,Cherrington:2007ax,Cherrington:2009ak,Cherrington:2009am}. In this chapter we present the \textit{abelian color cycle} dualization method, using four-dimensional SU(2) lattice gauge theory as a first example of application. We then show how the generalization of this approach to the theory with fermions leads us to the development of the \textit{abelian color flux} (ACF) method. Both these methods are versatile, as shown by the fact that we already applied the ACF approach to the SU(2) principal chiral model in the previous chapter. Nonetheless, the technicalities and the mathematical tools needed vary, depending on the theory one focuses on. \section{Abelian color cycle dualization of the pure SU(2) gauge theory \label{sec:su2_acc}} The Wilson action for SU(2) lattice gauge theory reads: \begin{equation} S_G[U] \; = \; -\dfrac{\beta}{2} \sum_{x,\mu < \nu} \Tr U_{x,\mu} \; U_{x+\hat{\mu},\nu} \, U_{x+\hat{\nu},\mu}^{\dagger} \, U_{x,\nu}^\dagger \, , \label{eq:su2_gaugeaction} \end{equation} where $U_{x,\mu} \in$ SU(2) are the dynamical degrees of freedom, which live on the links $(x,\mu)$ of a four-dimensional lattice, where we impose periodic boundary conditions. The partition sum $Z$ is then given by \begin{equation} \label{eq:su2_partition} Z \, = \, \int \! D[U] \, e^{-S_{G}[U]} \, = \, \prod_{x,\mu} \int_{SU(2)} \! dU_{x,\mu} \, e^{-S_{G}[U]}\, , \end{equation} where in the second step we have explicitly written the product over the links $(x,\mu)$ of the SU(2) Haar measures $\int \! D[U] = \prod_{x,\mu} \int_{SU(2)} dU_{x,\mu}$. As in the previous chapter, the explicit parametrization of the SU(2) matrices we adopt is \begin{equation} \label{eq:su2_parametrization} U_{x,\mu} = \left[ \begin{array}{cc} \cos \theta_{x,\mu} \, e^{\,i \alpha_{x,\mu}} & \sin \theta_{x,\mu} \, e^{\, i \beta_{x,\mu}} \\ -\sin \theta_{x,\mu} \, e^{-i \beta_{x,\mu}} & \cos \theta_{x,\mu} \, e^{-i \alpha_{x,\mu}} \end{array}\right] \, , \end{equation} with $\theta_{x,\mu} \in [0, \pi/2]$, $\alpha_{x,\mu}, \beta_{x,\mu} \in [-\pi, \pi]$ and Haar measure given by \begin{equation} \label{eq:su2_haar} dU_{x,\mu} = 2 \sin \theta_{x,\mu} \cos \theta_{x,\mu} \, d\theta_{x,\mu} \dfrac{d\alpha_{x,\mu}}{2 \pi} \, \dfrac{d\beta_{x,\mu}}{2 \pi} \, . \end{equation} The crucial step to dualize this theory consists in decomposing the action into its minimal terms, which we refer to as \textit{"Abelian Color Cycles"} (ACCs). This decomposition is carried out by writing explicitly both the trace and the matrix multiplications in the expression (\ref{eq:su2_gaugeaction}) of the gauge action: \begin{equation} \label{eq:su2_accaction} S_G[U] \; = \; -\dfrac{\beta}{2} \sum_{x,\mu < \nu} \sum_{a,b,c,d=1}^{2} U_{x,\mu}^{ab} U_{x+\hat{\mu},\nu}^{bc} U_{x+\hat{\nu},\mu}^{dc \ \star} U_{x,\nu}^{ad \ \star}. \end{equation} The products of the four link elements $U_{x,\mu}^{ab} U_{x+\hat{\mu},\nu}^{bc} U_{x+\hat{\nu},\mu}^{dc \, \star} U_{x,\nu}^{ad \, \star} \in \mathbb{C}$ are the ACCs. They are complex numbers and, as such, abelian. We will see later that this is a key property, because it enables the factorization and the expansion of the Boltzmann weight $e^{-S_{G}[U]}$, and thus it allows us to proceed with the dualization of the partition sum $Z$ as in the well-known case of abelian theories (in Sec.~\ref{sec:teo_u1} we discussed the example of dualizing U(1) pure gauge theory). Before doing so, let us spend a few words describing the geometrical interpretation we give to the ACCs on the lattice. First of all, notice that each abelian color cycle $U_{x,\mu}^{ab} U_{x+\hat{\mu},\nu}^{bc} U_{x+\hat{\nu},\mu}^{dc \, \star} U_{x,\nu}^{ad \, \star}$ carries two sets of indicies: the usual space-time position indices $(x,\mu\nu)$, and the color indices $a,b,c,d=1,2$. Clearly, $(x,\mu\nu)$ $\mu<\nu$ denotes the space-time position of the plaquette on which the ACC sits. The color indicies $a$, $b$, $c$ and $d$ come from the labeling of the four matrix entries that constitute the ACC (compare Eq.~(\ref{eq:su2_accaction})). Each of these four indices labels a corner of the plaquette $(x,\mu\nu)$, starting with $a$ at site $x$ and continuing in the mathematically positive direction for the other three color labels. If we now imagine the lattice to have two layers, each layer corresponding to a different color, then the four-tuple $(a,b,c,d)$ specifies the color path described by the ACC around the plaquette $(x,\mu\nu)$. For a better understanding of this interpretation, in Fig.~\ref{fig:su2_acc} we show the abelian color cycle $U_{x,\mu}^{21} U_{x+\hat{\mu},\nu}^{12} U_{x+\hat{\nu},\mu}^{22 \, \star} U_{x,\nu}^{22 \, \star}$ as an example. \begin{figure} \begin{center} \includegraphics[scale=0.6,clip]{chapters/plots/su2_acc.pdf} \end{center} \caption{Geometrical representation of an abelian color cycle (ACC) in the $\mu$--$\nu$ plane for the example $U_{x,\mu}^{21} U_{x+\hat{\mu},\nu}^{12} U_{x+\hat{\nu},\mu}^{22 \, \star} U_{x,\nu}^{22 \, \star}$. Each link element $U_{x,\mu}^{ab}$ is represented as an arrow connecting color $a$ at site $x$ to color $b$ at site $x+\hat{\mu}$ and we use two layers of the lattice to represent the two colors. The ACC shown here corresponds to the cycle occupation number $p_{x,\mu\nu}^{2122}\in\mathbb{C}$. \label{fig:su2_acc}} \end{figure} The two colors degrees of freedom correspond to the two layers of the lattice which, in the figure, we sketch in light grey as two copies of the plaquette we consider. The first link element of the $2122$ ACC is $U_{x,\mu}^{21}$. It is represented with a positively oriented arrow and it connects color $2$ on site $x$ to color $1$ on site $x+\hat{\mu}$. The next factor $U_{x+\hat{\mu},\nu}^{12}$ subsequently connects color $1$ on site $x+\hat{\mu}$ to color $2$ on site $x+\hat{\mu} + \hat{\nu}$, also with a positively oriented arrow. Then, since in our representation complex conjugation corresponds to a negative orientation of the arrow, the last two link elements $U_{x+\hat{\nu},\mu}^{22 \, \star} U_{x,\nu}^{22 \, \star}$ close the ACC around the plaquette. Summarizing, we interpret the ACCs $U_{x,\mu}^{ab} U_{x+\hat{\mu},\nu}^{bc} U_{x+\hat{\nu},\mu}^{dc \ \star} U_{x,\nu}^{ad \ \star}$ as paths in color space closing around plaquettes. The path in color is described by the labels $(a,b,c,d)$, one for every corner of the plaquette $(x,\mu\nu)$. This results in a total of $2^4 = 16$ possible ACCs for every plaquette of the lattice. We show all of these possibilities in Fig.~\ref{fig:su2_allcycles}. Notice that, in the case of SU(2), only ACCs with mathematically positive orientations are needed. This fact can be traced back to the pseudo-reality of SU(2), that guaranties the reality of the traces of SU(2) matrices. \begin{figure} \begin{center} \includegraphics[scale=1.10,clip]{chapters/plots/su2_cycles.pdf} \end{center} \caption{The 16 possible abelian color cycles which are attached to a given plaquette. In the dual representation their occupation is given by the corresponding cycle occupation number $p_{x,\mu\nu}^{abcd} \in \mathds{N}_0$.\label{fig:su2_allcycles}} \end{figure} We may now continue with the dualization of the partition sum (\ref{eq:su2_partition}): \begin{align} Z & = \int \! D[U] \, e^{-S_G[U]} \; = \; \int \! D[U] \prod_{x,\mu<\nu} \prod_{a,b,c,d = 1}^{2} e^{\frac{\beta}{2} U_{x,\mu}^{ab} U_{x+\hat{\mu},\nu}^{bc} U_{x+\hat{\nu},\mu}^{dc \ \star} U_{x,\nu}^{ad \ \star}} \nonumber \\ & = \int \! D[U] \prod_{x,\mu<\nu} \prod_{a,b,c,d = 1}^{2} \sum_{p_{x,\mu\nu}^{abcd} = 0}^{\infty} \dfrac{ \left( \beta/2 \right)^{p_{x,\mu\nu}^{abcd}}}{p_{x,\mu\nu}^{abcd}\, !} \left( U_{x,\mu}^{ab} U_{x+\hat{\mu},\nu}^{bc} U_{x+\hat{\nu},\mu}^{dc \ \star} U_{x,\nu}^{ad \ \star} \right)^{p_{x,\mu\nu}^{abcd}} \, . \label{eq:su2_partition2} \end{align} In the first line we rewrote the sums in the exponent as a product over all ACCs and all plaquettes of the lattice with the local exponentials $e^{\frac{\beta}{2} U_{x,\mu}^{ab} U_{x+\hat{\mu},\nu}^{bc} U_{x+\hat{\nu},\mu}^{dc \ \star} U_{x,\nu}^{ad \ \star}} $ as factors. We stress once more that this factorization is possible thanks to the ACC decomposition (\ref{eq:su2_accaction}), in which every term of the gauge action is a complex number and therefore commutes. In the second step we then expanded each local factor in a Taylor series, thus introducing the expansion coefficients $p_{x,\mu\nu}^{abcd} \in \mathbb{N}_{0}$, which will turn out to be our dual variables. Since each expansion coefficient $p_{x,\mu\nu}^{abcd} \in \mathbb{N}_{0}$ is related to one ACC $U_{x,\mu}^{ab} U_{x+\hat{\mu},\nu}^{bc} U_{x+\hat{\nu},\mu}^{dc \ \star} U_{x,\nu}^{ad \ \star}$, we refer to them as \textit{cycle occupation numbers}. We now rearrange the product in the last line of Eq.~(\ref{eq:su2_partition2}) so that all factors of $U_{x,\mu}^{ab}$ and $U_{x,\mu}^{ab \, \star}$ associated with a given link element are grouped together: \begin{equation} Z = \sum_{\{p\}} \left[ \prod_{x,\mu<\nu} \prod_{a,b,c,d} \dfrac{ \left(\beta/2 \right)^{p_{x,\mu\nu}^{abcd}}}{p_{x,\mu\nu}^{abcd}\, !} \right] \prod_{x,\mu} \int \! \! dU_{x,\mu} \; \prod_{a,b} \left( U_{x,\mu}^{ab} \right) ^{N_{x,\mu}^{ab}} \left( U_{x,\mu}^{ab \ \star} \right) ^{\overline{N}_{x,\mu}^{ab}} \, . \label{eq:su2_partition3} \end{equation} Written in this form the partition function is suitable for the Haar integration, as we shall discuss later. In (\ref{eq:su2_partition3}) we use the short-hand notation \begin{equation} \sum_{\{p\}} \; = \; \prod_{x,\mu<\nu} \prod_{a,b,c,d = 1}^{2} \sum_{p_{x,\mu\nu}^{abcd} = 0}^{\infty} \label{eq:su2_cyclesum} \end{equation} to denote the sum over all the configurations of the cycle occupation numbers $p_{x,\mu\nu}^{abcd} \in \mathbb{N}_{0}$. Moreover, we introduced $N_{x,\mu}^{ab}\in \mathbb{N}_{0}$ and $\overline{N}_{x,\mu}^{ab}\in \mathbb{N}_{0}$ for the exponents of the link variable elements $U_{x,\mu}^{ab}$ and $U_{x,\mu}^{ab \ \star}$ respectively. Explicitly they are given by the following linear combinations of cycle occupation numbers: \begin{equation} N_{x,\mu}^{ab} \; = \; \sum_{\nu:\mu<\nu}p_{x,\mu\nu}^{abss} + \sum_{\rho:\mu>\rho} p_{x-\hat{\rho},\rho\mu}^{sabs}\, , \qquad \overline{N}_{x,\mu}^{ab} \; = \; \sum_{\nu:\mu<\nu}p_{x-\hat{\nu},\mu\nu}^{ssba} + \sum_{\rho:\mu>\rho}p_{x,\rho\mu}^{assb} \, , \label{eq:su2_nnbar} \end{equation} where the label $s$ stands for the independent summation of the color indices replaced by it, e.g., $p_{x,\mu\nu}^{abss} = \sum_{c,d =1}^{2} p_{x,\mu\nu}^{abcd}$ or $p_{x,\mu\nu}^{sabs} = \sum_{c,d =1}^{2} p_{x,\mu\nu}^{cabd}$. At this point we substitute the parametrization (\ref{eq:su2_parametrization}) of the link elements and the expression (\ref{eq:su2_haar}) of the Haar measure in Eq.~(\ref{eq:su2_partition3}) and obtain: \begin{align} Z \; = &\; \sum_{\{p\}} \! \left[ \prod_{x,\mu<\nu} \prod_{a,b,c,d} \dfrac{ \left( \beta/2 \right)^{p_{x,\mu\nu}^{abcd}}}{p_{x,\mu\nu}^{abcd}!} \right] \nonumber \\ & \times \prod_{x,\mu} (-1)^{J_{x,\mu}^{21}} \; \, 2 \! \int_{0}^{\pi/2} \!\!\!\! d\theta_{x,\mu} \, (\cos\theta_{x,\mu})^{1 + S_{x,\mu}^{11} + S_{x,\mu}^{22}} \; (\sin\theta_{x,\mu})^{1 + S_{x,\mu}^{12} + S_{x,\mu}^{21}} \nonumber \\ & \hspace{20mm} \times \; \int_{0}^{2\pi} \! \dfrac{d\alpha_{x,\mu}}{2\pi} \; e^{i\alpha_{x,\mu}[J_{x,\mu}^{11}-J_{x,\mu}^{22}]} \; \int_{0}^{2\pi} \! \dfrac{d\beta_{x,\mu}}{2\pi} \; e^{i\beta_{x,\mu}[J_{x,\mu}^{12}-J_{x,\mu}^{21}]} \; . \label{eq:su2_partition4} \end{align} For a convenient notation we introduced the integer valued fluxes $J_{x,\mu}^{ab}\in \mathbb{Z}$ and $S_{x,\mu}^{ab}\in \mathbb{N}_{0}$, \begin{equation} \label{eq:su2_jfluxes} J_{x,\mu}^{ab} \; = \; N_{x,\mu}^{ab} - \overline{N}_{x,\mu}^{ab} \; = \; \sum_{\nu:\mu<\nu}[\, p_{x,\mu\nu}^{abss} - p_{x-\hat{\nu},\mu\nu}^{ssba} \, ] - \sum_{\rho:\mu>\rho}[\, p_{x,\rho\mu}^{assb} - p_{x-\hat{\rho},\rho\mu}^{sabs} \, ] \; , \end{equation} \begin{equation} \label{eq:su2_sfluxes} S_{x,\mu}^{ab} \; = \; N_{x,\mu}^{ab} + \overline{N}_{x,\mu}^{ab} \; = \; \sum_{\nu:\mu<\nu}[\, p_{x,\mu\nu}^{abss} + p_{x-\hat{\nu},\mu\nu}^{ssba}\, ] + \sum_{\rho:\mu>\rho}[\, p_{x,\rho\mu}^{assb} + p_{x-\hat{\rho},\rho\mu}^{sabs}\, ] \; . \end{equation} It is now easy to see that the link integrals in Eq.~(\ref{eq:su2_partition4}) can be solved in closed form. The integrals over the phases $\alpha_{x,\mu}$ and $\beta_{x,\mu}$ give rise to Kronecker deltas that impose constraints over the color link fluxes $J_{x,\mu}^{ab}$ we have just introduced. In particular, the constraints enforce the following equalities between components of the $J$-fluxes at all links $(x,\mu)$: \begin{equation} J_{x,\mu}^{11} \, - \, J_{x,\mu}^{22} \; = \; 0 \quad \forall x,\mu \qquad \text{and} \qquad J_{x,\mu}^{12} \, - \, J_{x,\mu}^{21} \; = \; 0 \quad \forall x,\mu \, . \label{eq:su2_constraints} \end{equation} These constraints can be easily understood after we discuss the geometrical interpretation of the currents $J_{x,\mu}^{ab}$. From their definition in Eq.~(\ref{eq:su2_jfluxes}), we know that the $J_{x,\mu}^{ab}$ sum over every cycle occupation number that contributes to the flux from color $a$ on site $x$ to color $b$ on site $x+\hat{\mu}$. In Fig.~\ref{fig:su2_j12} we show four of the plaquettes attached to the link $(x,\mu)$ and illustrate how they contribute to $J_{x,\mu}^{12}$ as an example. On the link $(x,\mu)$ the flux from color 1 to 2 is kept fixed and represented with solid arrows. For every plaquette attached to the link this flux gets contributions from four different cycle occupation numbers, which are summed over in the definition (\ref{eq:su2_jfluxes}), and illustrated with dotted lines in the figure. Hence, $J_{x,\mu}^{ab}$ is the total flux from color $a$ on site $x$ to color $b$ on site $x+\hat{\mu}$. \begin{figure} \begin{center} \includegraphics[scale=0.5,clip]{chapters/plots/su2_j12.pdf} \end{center} \caption{Graphical illustration of the contributions from the cycle occupation numbers to the $J$-flux using the example of the $J_{x,\mu}^{12}$ element. For a description of the figure see the text.\label{fig:su2_j12}} \end{figure} \begin{figure} \begin{center} \includegraphics[scale=0.8,clip]{chapters/plots/su2_constraints.pdf} \end{center} \caption{Geometrical illustration of the two constraints in Eq.~\protect(\ref{eq:su2_constraints}) for the fluxes $J_{x,\mu}^{ab}$ on all links $(x,\mu)$. The first constraint (lhs.~plot) requires the sum over all 1-1 fluxes to equal the sum over all 2-2 fluxes. The second constraint requires the sum over 1-2 fluxes to equal the sum over 2-1 fluxes.\label{fig:su2_constraints}} \end{figure} Now that we have given the interpretation of the $J_{x,\mu}^{ab}$ fluxes, it is straightforward to understand the meaning of the constraints in Eq.~(\ref{eq:su2_constraints}): for every link of the lattice the total fluxes on the two color layers have to be equal, and the total fluxes between the two layers have to match. We represent these conditions schematically in Fig.~\ref{fig:su2_constraints}. The constraints (\ref{eq:su2_constraints}) can also be used to simplify the result of the last link integral in Eq.~(\ref{eq:su2_partition4}). The integrals over the $\theta_{x,\mu}$ angles have the form of the well-known integral representation of the beta function $\B$ \begin{align} 2 \int_{0}^{\frac{\pi}{2}} \! d\theta \, (\cos \theta)^{1 + n} \,(\sin \theta)^{1 + m} \, &= \, \B \bigg(\frac{n}{2} + 1 \bigg| \frac{m}{2} + 1 \bigg) = \dfrac{\Gamma\big(\frac{n}{2} + 1 \big) \Gamma \big( \frac{m}{2} + 1 \big) }{\Gamma\big(\frac{n + m}{2} + 2 \big)} \nonumber \\ &= \, \dfrac{\frac{n}{2}! \frac{m}{2}!}{\big(\frac{n + m}{2} + 1 \big)!} \, . \label{eq:su2_betafunction} \end{align} In the second step we represent the beta function $\B$ in terms of gamma functions $\Gamma$, while the last simplified expression in terms of factorials holds only if $n$ and $m$ are even. In our case $n$ and $m$ are given by the sums \begin{equation} S_{x,\mu}^{11} \, + \, S_{x,\mu}^{22} \; = \; \text{even} \quad \forall x,\mu \qquad \text{and} \qquad S_{x,\mu}^{12} \, + \, S_{x,\mu}^{21} \; = \; \text{even} \quad \forall x,\mu \, \label{eq:su2_even} \end{equation} respectively. It is easy to prove the evenness of these combinations using the constraints (\ref{eq:su2_constraints}) and recalling the definitions (\ref{eq:su2_jfluxes}) and (\ref{eq:su2_sfluxes}) for the $J$- and $S$-fluxes: the expressions on the left hand sides of the equations in (\ref{eq:su2_even}) can be obtained from the expressions on the left hand sides of (\ref{eq:su2_constraints}) by adding even quantities on both sides of the equations. Therefore, the integrals over the $\theta_{x,\mu}$ angles in Eq.~(\ref{eq:su2_partition4}) result in the simple combinatorial factors \begin{equation} W_{H}[p] \; = \; \prod_{x,\mu} \dfrac{\left(\frac{S_{x,\mu}^{11} + S_{x,\mu}^{22}}{2}\right)! \left(\frac{S_{x,\mu}^{12} + S_{x,\mu}^{21}}{2}\right)!}{\left(\frac{S_{x,\mu}^{11} + S_{x,\mu}^{22} + S_{x,\mu}^{12} + S_{x,\mu}^{21}}{2} + 1\right)!} \; . \label{eq:su2_wh} \end{equation} We then introduce the $\beta$-dependent weight factor $W_{\beta}[p]$, which collects the powers and factorials from the expansion of the exponentials \begin{equation} \label{eq:su2_wbeta} W_{\beta}[p] \; = \; \prod_{x,\mu<\nu} \prod_{a,b,c,d} \dfrac{ \left( \frac{\beta}{2} \right)^{p_{x,\mu\nu}^{abcd}}}{p_{x,\mu\nu}^{abcd}!} \; . \end{equation} This weight organizes the partition function in the form of a power series of the inverse lattice coupling $\beta$, i.e., in a strong coupling expansion. Putting things together we obtain for the partition sum \begin{equation} Z \; = \; \sum_{\{p\}} W_{\beta}[p] \; W_H[p] \; (-1)^{\sum_{x,\mu}J_{x,\mu}^{21}} \; \prod_{x,\mu} \delta(J_{x,\mu}^{11}-J_{x,\mu}^{22}) \; \delta(J_{x,\mu}^{12}-J_{x,\mu}^{21}) \; . \label{eq:su2_dualgaugepartition} \end{equation} In its dual form (\ref{eq:su2_dualgaugepartition}) the partition function is a sum over configurations of cycle occupation numbers $p_{x,\mu\nu}^{abcd} \in \mathbb{N}_{0}$ attached to the plaquettes $(x, \mu < \nu)$ of the lattice. At each link $(x,\mu)$ the $p_{x,\mu\nu}^{abcd}$ have to obey constraints which are expressed in terms of the two Kronecker deltas (we use the notation $\delta(n) \equiv \delta_{n,0}$) in (\ref{eq:su2_dualgaugepartition}) which relate different components of the link fluxes $J_{x,\mu}^{ab}$, given in Eq.~(\ref{eq:su2_jfluxes}). Each configuration comes with two weight factors $W_{\beta}[p]$ and $W_{H}[p]$ that collect the coefficients of the Taylor expansion and the combinatorial factors from the Haar measure integrals respectively. Both these weight factors are real and positive (compare the explicit expressions (\ref{eq:su2_wh}) and (\ref{eq:su2_wbeta})). Nonetheless, the partition sum (\ref{eq:su2_dualgaugepartition}) also contains the explicit sign factor $(-1)^{\sum_{x,\mu}J_{x,\mu}^{21}}$. This minus sign origins from the minus sign in the (2,1) matrix element in the parametrization (\ref{eq:su2_parametrization}) of our SU(2) link variables. Using the constraint in Eq.~(\ref{eq:su2_dualgaugepartition}) we can give a simple interpretation of the sign factor: since by the constraints (\ref{eq:su2_constraints}) the $J_{x,\mu}^{21}$ flux equals the $J_{x,\mu}^{12}$ flux, the configurations that contribute to the partition function with a negative sign are the ones that have an odd number of flux crossings: \begin{equation*} (-1)^{\sum_{x,\mu} J_{x,\mu}^{12}} = (-1)^{\text{\# of flux crossings}} \, . \end{equation*} An example of a configuration with an odd number of crossing is given by \begin{equation*} p_{x,\mu\nu}^{1112} = 1 \, , \qquad p_{x,\mu\nu}^{1121} = 1 \, , \qquad p_{x,\mu\nu}^{2222} = 1 \, , \qquad p_{x,\mu\nu}^{2211} = 1 \, . \end{equation*} We illustrate this contribution in Fig.~\ref{fig:su2_unhappy}. \begin{figure} \begin{center} \includegraphics[scale=0.5,clip]{chapters/plots/su2_unhappy.pdf} \end{center} \caption{Example of a configuration that contributes to the partition function with a negative sign. We show the four cycle occupation numbers that are set to 1. Obviously all constraints are obeyed. The configuration has three flux crossings and thus a negative sign. \label{fig:su2_unhappy}} \end{figure} Obviously this configuration is admissible since it fulfils both the constraints in (\ref{eq:su2_constraints}) on each link of the plaquette but, since it has an odd number of crossings, it contributes to the partition function with a negative sign. This configuration is one of the lowest negative sign configurations which one can construct. In fact, the first negative configurations appear at order $\beta^{4}$. At this order negative configurations are local, i.e., they are built from four non-zero cycle occupation numbers on the same plaquette. A large class of admissible dual pure gauge configurations contributing to the partition sum with a positive weight are closed orientable and non-orientable surfaces of plaquettes where certain cycle occupation numbers are occupied. While the examples of negative configurations we found are local, the surfaces are non-local and contribute to the long distance properties of the theory. It is an interesting open question if all the configurations that contribute at long distance are positive. \section{Abelian color flux dualization of fermions \label{sec:su2_acf}} Having discussed the pure gauge part of the theory, we now focus on fermions. We consider one flavor of staggered fermions described by the action \begin{equation} S_F[\,\overline{\psi}, \psi,U] = \sum_x \Big[ m \, \overline{\psi}_x \psi_x + \sum_{\mu} \frac{\gamma_{x,\mu}}{2} \Big( \, \overline{\psi}_x U_{x,\mu} \psi_{x + \hat\mu} - \overline{\psi}_{x+\hat\mu} U_{x,\mu}^\dagger \psi_{x} \Big) \Big] \, , \label{eq:su2_actionf} \end{equation} where $m$ is the mass and $\gamma_{x,\mu}$ is the usual staggered sign factor defined as \begin{equation*} \gamma_{x,1} = 1 \, , \ \quad \gamma_{x,2} = (-1)^{x_{1}} \, , \ \quad \gamma_{x,3} = (-1)^{x_{1} + x_{2}} \, , \ \quad \gamma_{x,4} = (-1)^{x_{1} + x_{2} + x_{3}}\, . \end{equation*} $\psi_{x}$ and $\overline{\psi}_{x}$ are two-component Grassmann vectors in color space \begin{equation} \psi_{x} = \left( \begin{array}{c} \psi_{x}^{1} \\ \psi_{x}^{2} \\ \end{array} \right) \ ,\qquad \overline{\psi}_{x} = \left( \overline{\psi}_{x}^{1} \ , \ \overline{\psi}_{x}^{2} \right) \, , \label{eq:su2_grassmann} \end{equation} that live on the sites $x$ of our four-dimensional $N^{3} \times N_{t}$ lattice, with anti-periodic boundary conditions in Euclidean time ($\nu = 4$) for the fermions and periodic in the spatial directions ($\nu=1,2,3$). The fermionic partition function in a gauge background is given by \begin{equation} Z_F[U] \; = \; \int \!\! D[\, \overline{\psi}, \psi] \; e^{-S_F[\,\overline{\psi}, \psi,U]} \; , \label{eq:su2_partitionf} \end{equation} where the measure is a product over Grassmann measures, $D[\, \overline{\psi}, \psi] = \prod_{x,a} d \psi^a_x d \overline{\psi}^a_x$. The full partition sum $Z$ is then obtained as $Z = \int D[U] e^{-S_G[U]} Z_F[U]$. The key step in the \textit{abelian color flux} (ACF) dualization of the fermionic partition sum (\ref{eq:su2_partitionf}) is the decomposition of the staggered action (\ref{eq:su2_actionf}) into Grassmann bilinears. This is done by making the sums over the color indices explicit: \begin{equation} S_F[\,\overline{\psi}, \psi,U] = \sum_x \Big[ \, m \!\sum_{a} \overline{\psi}_x^{\,a} \psi_x^{a} + \sum_{\mu} \frac{\gamma_{x,\mu}}{2} \sum_{a,b} \Big( \, \overline{\psi}^{\,a}_x U_{x,\mu}^{\,ab} \psi_{x + \hat\mu}^{b} - \overline{\psi}_{x+\hat\mu}^{\,b} U_{x,\mu}^{\, ab \, \star} \psi_{x}^{a} \Big) \Big] \; . \label{eq:su2_acfactionf} \end{equation} As a result, every term in the decomposition (\ref{eq:su2_acfactionf}) of the staggered action commutes and, again, this is what enables us to proceed with the dualization. We find for the fermionic partition function \begin{align} Z_F[U] \, &= \int \!\!\! D[\, \overline{\psi}, \psi] \prod_x \prod_a e^{-m \overline{\psi}_x^{\,a} \psi_x^a} \; \prod_{x,\mu} \prod_{a,b} e^{- \,\frac{\gamma_{x,\mu}}{2} \overline{\psi}^{\,a}_x U_{x,\mu}^{\,ab} \psi_{x + \hat\mu}^b} \; e^{ \, \frac{\gamma_{x,\mu}}{2} \overline{\psi}_{x+\hat\mu}^{\,b} U_{x,\mu}^{\, ab\, \star} \psi_{x}^a} \nonumber \\ &= \int \!\!\!D[\, \overline{\psi}, \psi] \prod_x \prod_a \sum_{s_x^a = \, 0}^{1} \! ( -m \overline{\psi}_x^{\,a} \psi_x^a )^{s_x^a} \nonumber \\ & \qquad \qquad \times \prod_{x,\mu} \prod_{a,b} \sum_{k_{x,\mu}^{\;ab} = \, 0 }^{1} \!\!\! \Big( \!- \frac{\gamma_{x,\mu}}{2} \, \overline{\psi}^{\,a}_x U_{x,\mu}^{\,ab} \psi_{x + \hat\mu}^b \Big)^{k_{x,\mu}^{\; ab}} \! \sum_{ \overline{k}_{x,\mu}^{\;ab} = \, 0}^{1}\!\!\! \Big( \frac{\gamma_{x,\mu}}{2} \, \overline{\psi}_{x+\hat\mu}^{\,b} U_{x,\mu}^{\, ab\, \star} \psi_{x}^a \Big)^{\overline{k}_{x,\mu}^{\; ab}} \nonumber \\ &= \, \frac{1}{2^{2V}} \! \sum_{\{s,k,\overline{k}\}} \!\! (2m)^{\, \sum_{x,a} s_x^a} \; \prod_{x,\mu} \prod_{a,b} ( U_{x,\mu}^{\,ab} )^{k_{x,\mu}^{\; ab}} \, ( U_{x,\mu}^{\, ab\, \star} )^{\overline{k}_{x,\mu}^{\; ab}} \; (-1)^{k_{x,\mu}^{\; ab}} \; ( \, \gamma_{x,\mu}\,) ^{k_{x,\mu}^{\; ab} + \overline{k}_{x,\mu}^{\; ab} } \nonumber \\ & \qquad \qquad \times \!\! \int \!\! D[\, \overline{\psi}, \psi] \! \prod_x \prod_a (\overline{\psi}_x^{\,a} \psi_x^a )^{s_x^a} \prod_{x,\mu} \prod_{a,b} (\overline{\psi}^{\,a}_x \psi_{x + \hat\mu}^b)^{k_{x,\mu}^{\; ab}} (\overline{\psi}_{x+\hat\mu}^{\,b} \psi_{x}^a)^{\overline{k}_{x,\mu}^{\; ab}} \; . \label{eq:su2_partitionf2} \end{align} In the first step we have factorized the Boltzmann weight into the local color site factors $e^{-m \overline{\psi}_x^{\,a} \psi_x^a}$, the forward hops $e^{-\frac{\gamma_{x,\mu}}{2} \overline{\psi}_x^{\,a} U_{x,\mu}^{ab} \psi_{x+\hat{\mu}}^b}$ and the backward hops $e^{\frac{\gamma_{x,\mu}}{2} \overline{\psi}_{x+\hat{\mu}}^{\,b} U_{x,\mu}^{ab\star} \psi_{x}^a}$. Afterwards, we expanded each single exponential in a power series which, in the case of fermions, terminates after the linear term due to the nilpotency of the Grassmann variables. The expansion indices, $s_{x}^{a} = 0,1$, $k_{x,\mu}^{ab} = 0,1$ and $\overline{k}_{x,\mu}^{ab} = 0,1$ are the new dual variables for fermions. \begin{figure} \begin{center} \vspace*{0.2cm} \includegraphics[scale=0.45,clip]{chapters/plots/su2_grassmann.pdf} \end{center} \caption{Graphical representation of the dual variables for the fermions. The first two diagrams on the very left represent the monomers $s_x^a$, the arrows are for the dual variables $k_{x,\mu}^{\,ab}, \overline{k}_{x,\mu}^{\,ab}$. \label{fig:su2_grassmann}} \end{figure} Their graphical representation, shown in Fig.~\ref{fig:su2_grassmann}, is obtained as an extension of the interpretation we gave to the abelian color cycle in the previous section. Whenever one of the dual variables for the fermions is equal to 1, we say that it \textit{activates} the Grassmann bilinear it is associated to. So, for example, $s_{x}^{a}$ activates the color $a$ component of the mass term on site $x$. We refer to these objects as \textit{monomers}, and we represent them with a circle around the corresponding site and color, as done in the two diagrams on the very left of Fig.~\ref{fig:su2_grassmann}. $k_{x,\mu}^{ab}$ is used for the forward hop that connects color $a$ to color $b$ on the link $(x,\mu)$ and, analogously, $\overline{k}_{x,\mu}^{ab}$ represents the backward hop that connects color $b$ to color $a$ on the link $(x,\mu)$. They are illustrated as oriented arrows connecting color layers on neighboring sites. As we will see shortly, these flux variables can be used to build up either \textit{dimers}, by setting $k_{x,\mu}^{ab} = \overline{k}_{x,\mu}^{ab} = 1$, or \textit{loops}, by setting closed chains of $k$ and $\overline{k}$ to 1. In the last step of Eq.~(\ref{eq:su2_partitionf2}) we reorganized the terms by collecting all the factors that do not depend on the Grassmann variables in front of the Grassmann integral. The Grassmann integral will give as a result either $0$ or $\pm 1$, depending on the values of the fermion dual variables $s_{x}^{a},k_{x,\mu}^{ab},\overline{k}_{x,\mu}^{ab} = 0,1$. Obviously, for the Grassmann integral to be non-vanishing each Grassmann variable $\overline{\psi}_{x}^{a} \psi_{x}^{a}$ has to appear exactly once. This requirement can be enforced by means of a Kronecker delta at every site and color layer: \begin{equation} \label{eq:su2_cf} C_{F}[s,k,\overline{k}] = \prod_{x,a} \delta \Big( 1 - s_{x}^{a} - \dfrac{1}{2} \sum_{\mu,b} \big[k_{x,\mu}^{ab} + \overline{k}_{x,\mu}^{ab} + k_{x - \hat{\mu},\mu}^{ba} + \overline{k}_{x - \hat{\mu},\mu}^{ba} \big]\Big) \, . \end{equation} The way of satisfying this fermion constraint is to completely fill the two layers of our four-dimensional lattice with monomers, dimers and fermionic loops. For this reason, when the conditions imposed by the product of Kronecker deltas (\ref{eq:su2_cf}) are satisfied, we say that the Grassmann integrals are \textit{saturated}. We now continue the discussion of (\ref{eq:su2_partitionf2}) focusing on the sign factors. The sources of signs in the case of fermions are various, and therefore a detailed discussion about them is mandatory. Distinctly visible from Eq.~(\ref{eq:su2_partitionf2}) are the negative sign factors associated to the forward hops $(-1)^{k_{x,\mu}^{ab}}$, as well as the staggered sign factors $(\gamma_{x,\mu})^{k_{x,\mu}^{ab} + \overline{k}_{x,\mu}^{ab}}$. Other sources of signs are the anti-periodic boundary conditions along the time direction ($\mu=4$), and the Grassmann integral itself. As we mentioned earlier, the Grassmann integral is non-vanishing only if it is saturated, namely if the whole lattice is filled up with monomers, dimers and loops. It is clear from Eq.~(\ref{eq:su2_partitionf2}) that monomers, which are activated by setting $s_{x}^{a} = 1$, contribute just with a factor of $2m$. Moreover, the corresponding Grassmann variables $\overline{\psi}_{x}^{a} \psi_{x}^{a}$ are already in the canonical order for the Grassmann integration, which means that the corresponding Grassmann integral gives a factor of $+1$. We conclude that monomers are not a source of signs. Let us now consider a dimer, activated by setting $k_{x,\mu}^{ab} = \overline{k}_{x,\mu}^{ab} = 1$: \begin{equation} \overline{\psi}^{\,a}_x \, \psi_{x + \hat\mu}^b \, \overline{\psi}_{x+\hat\mu}^{\,b} \, \psi_{x}^a \; = \; - \; \overline{\psi}^{\,a}_x \, \psi_{x}^a \, \overline{\psi}_{x+\hat\mu}^{\,b} \, \psi_{x + \hat\mu}^b \; . \label{dimer} \end{equation} On the right hand side of the equality we reordered the Grassmann variables in the canonical way for the integration. By doing so we pick up a minus sign, which anyway is compensated by the minus sign coming from the forward hop ($(-1)^{k_{x,\mu}^{ab}} = (-1)^{1} = -1$). Moreover, the staggered sign factor is always positive for dimers ($(\gamma_{x,\mu})^{k_{x,\mu}^{ab} + \overline{k}_{x,\mu}^{ab}} = (-1)^{2} = +1$) and, ultimately, since dimers are backtracking loops of length 2, they always have vanishing net winding number. We conclude that dimers are not a source of signs either. The discussion about loops is a bit more complicated. Let us start with considering the product of the staggered sign factors along a plaquette: \begin{equation} \sigma_{x,\mu\nu} \; = \; \gamma_{x,\mu} \, \gamma_{x+\hat{\mu},\nu} \, \gamma_{x+\hat{\nu},\mu} \, \gamma_{x,\nu} \; = \; -\, 1 \,. \end{equation} This result always holds, independently of the plane $\mu,\nu$ on which the plaquette sits, the position $x$, or the orientation of the loop. If we then consider a loop closing around two adjacent plaquettes, the product of the staggered sing factors along that loop can be equivalently computed by multiplying the plaquette factors $\sigma_{x,\mu\nu}$ corresponding to the two plaquettes bordered by the loop, e.g.: \begin{align*} \sigma_{x,\mu\nu} \, \sigma_{x + \hat{\mu},\mu\nu}\; &\equiv \; \gamma_{x,\mu} \, \gamma_{x+\hat{\mu},\nu} \, \gamma_{x+\hat{\nu},\mu} \, \gamma_{x,\nu} \, \gamma_{x + \hat{\mu},\mu} \, \gamma_{x+2\hat{\mu},\nu} \, \gamma_{x+ \hat{\mu}+\hat{\nu},\mu} \, \gamma_{x + \hat{\mu},\nu} \; \\ &= \; \gamma_{x,\mu} \, \gamma_{x + \hat{\mu},\mu} \, \gamma_{x+2\hat{\mu},\nu} \, \gamma_{x+ \hat{\mu}+\hat{\nu},\mu} \, \gamma_{x+\hat{\nu},\mu} \, \gamma_{x,\nu} \,. \end{align*} The last equivalence holds because the staggered sign factor on the link common to the two adjacent plaquettes is squared, i.e., $(\gamma_{x + \hat{\mu},\nu})^{2} = 1$. This procedure can be iterated to form any kind of loops, and the resulting staggered sign factor can therefore be computed simply as: \begin{equation} \prod_{(x,\mu)\in \mathcal{L}} \gamma_{x,\mu} \; = \; (\, -1 \,)^{P_{\mathcal{L}}} \, , \label{eq:su2_signpl} \end{equation} where $P_{\mathcal{L}}$ stands for the number of plaquettes necessary to cover the surface bounded by the loop $\mathcal{L}$. Note that there is an ambiguity in this definition caused by the fact that, in more than two dimensions, the surface bounded by a loop is not unique. Nevertheless, different surfaces always differ by an even number of plaquettes, and the result in (\ref{eq:su2_signpl}) is the same also if different surfaces $P_{\mathcal{L}}$ are used. The loops pick up further signs from the sign factors $\prod_{x,\mu,a,b} (-1)^{k_{x,\mu}^{ab}}$ introduced by the forward hops of the loop. For trivially closing loops, where the number of forward hops is half the length of the loop, this sign factor can be expressed as $(-1)^{|\mathcal{L}|/2}$, where $|\mathcal{L}|$ is the length of the loop, i.e., the number of links that constitute the loop $\mathcal{L}$. Loops that wind around the compact time or space direction do not share the property of equal numbers of forward and backward hops and one would have to distinguish different cases. For simplicity we here assume that the temporal and spatial extents of the lattice are all multiples of 4 and it is easy to see that $(-1)^{|\mathcal{L}|/2}$ then always correctly takes into account the signs from forward hopping. Loops that wind in the temporal direction also pick up a minus sign due to the anti-periodic boundary conditions, and we express this sign as $(-1)^{\mathcal{W}_\mathcal{L}}$, where $\mathcal{W}_\mathcal{L}$ denotes the net temporal winding number of the loop $\mathcal{L}$. Finally, each loop also picks up a minus sign from the reordering of the Grassmann variables. Summarizing, we obtained that the sign of a loop $\mathcal{L}$ only depends on its geometry, and is given by \begin{equation} \mbox{sign}\,(\mathcal{L}) \; = \; (-1)^{ \, |\mathcal{L}|/2 \, + \, \mathcal{W}_\mathcal{L} \, + \, P_\mathcal{L} \, + \, 1 } \; , \label{eq:su2_loopsign} \end{equation} where $|\mathcal{L}|$ is the length of the loop, $\mathcal{W}_\mathcal{L}$ is the number of windings around the compact time direction, and $P_{\mathcal{L}}$ the number of plaquettes of a surface with $\mathcal{L}$ as its boundary. Putting things together we find the following expression for the fermionic partition sum, \begin{equation} Z_F[U] \, = \, \frac{1}{2^{2V}} \! \sum_{\{s,k,\overline{k}\}} C_{F}[s,k,\overline{k}\,] \; W_M[s] \; \prod_{\mathcal{L}} \, \mbox{sign} \,(\mathcal{L}) \prod_{x,\mu} \prod_{a,b} ( U_{x,\mu}^{\,ab} )^{k_{x,\mu}^{\; ab}} \, ( U_{x,\mu}^{\, ab\, \star} )^{\overline{k}_{x,\mu}^{\; ab}} \; , \label{eq:su2_partitionf3} \end{equation} where we collected the contributions of the monomers into the weight factor $W_M[s]$: \begin{equation} \label{eq:su2_wm} W_M[s] = \prod_{x,a}(2m)^{s_{x}^{a}} \, . \end{equation} The fermionic partition function (\ref{eq:su2_partitionf3}) is a sum over all the configurations of the fermion variables $s_{x}^{a}, k_{x,\mu}^{ab}, \overline{k}_{x,\mu}^{ab} \in \{0,1\}$. The admissible configurations must satisfy the fermion constraint $C_{F}[s,k,\overline{k}]$, which implement Pauli's exclusion principle for our dual representation. The only sources of sign for fermions are loops, which come with the sign function $\sign(\mathcal{L})$ given in (\ref{eq:su2_loopsign}). Monomers come with the weight (\ref{eq:su2_wm}), while dimers and loops activate factors $U_{x,\mu}^{ab}$ and $U_{x,\mu}^{ab \,\star}$. In the next section we discuss the integration of the gauge fields and obtain the final result for the partition function of the SU(2) lattice gauge theory with one flavor of staggered fermions. \section{The full theory \label{sec:su2_full}} The last step necessary in order to obtain the final dual form of the partition function $Z = \int \! D[U] \, e^{-S_{G}[U]} Z_{F}[U]$ consists in performing the link integrals \begin{align} &\int\! D[U] \, e^{-S_{G}[U]} \, \prod_{x,\mu} \prod_{a,b} ( U_{x,\mu}^{\,ab} )^{k_{x,\mu}^{\; ab}} \, ( U_{x,\mu}^{\, ab\, \star} )^{\overline{k}_{x,\mu}^{\; ab}} \nonumber \\ &= \, \prod_{x,\mu} \int \! dU_{x,\mu} \prod_{a,b} ( U_{x,\mu}^{\,ab} )^{N_{x,\mu}^{\; ab} + k_{x,\mu}^{\; ab}} \, ( U_{x,\mu}^{\, ab\, \star} )^{\overline{N}_{x,\mu}^{\; ab} + \overline{k}_{x,\mu}^{\; ab} } \, . \end{align} The solution to these integrals can be derived from the one obtained in Section \ref{sec:su2_acc}, substituting $N_{x,\mu}^{\; ab}$ with $N_{x,\mu}^{\; ab} + k_{x,\mu}^{\; ab}$ and $\overline{N}_{x,\mu}^{\; ab}$ with $\overline{N}_{x,\mu}^{\; ab} + \overline{k}_{x,\mu}^{\; ab}$. Thus we find that the integral over the $\theta_{x,\mu}$ angles results in the combinatorial factor \begin{equation} \label{su2_whfull} W_H [p,k,\overline{k}] \; = \; \prod_{x,\mu} \dfrac{\!\left(\!\frac{S_{x,\mu}^{11} +k_{x,\mu}^{\; 11}+\overline{k}_{x,\mu}^{\; 11}+ S_{x,\mu}^{22} +k_{x,\mu}^{\; 22}+\overline{k}_{x,\mu}^{\; 22}}{2}\!\right) ! \; \left(\frac{S_{x,\mu}^{12} +k_{x,\mu}^{\; 12}+\overline{k}_{x,\mu}^{\; 12} + S_{x,\mu}^{21}+k_{x,\mu}^{\; 21}+\overline{k}_{x,\mu}^{\; 21}}{2}\right) !}{\!\left(\frac{ S_{x,\mu}^{11} + k_{x,\mu}^{\; 11}+\overline{k}_{x,\mu}^{\; 11} + S_{x,\mu}^{22} + k_{x,\mu}^{\; 22}+\overline{k}_{x,\mu}^{\; 22} + S_{x,\mu}^{12} + k_{x,\mu}^{\; 12}+\overline{k}_{x,\mu}^{\; 12} + S_{x,\mu}^{21} + k_{x,\mu}^{\; 21}+\overline{k}_{x,\mu}^{\; 21} }{2} + 1\!\right) !} \; , \end{equation} while the integrals over the phases $\alpha_{x,\mu}$ and $\beta_{x,\mu}$ give rise to Kronecker deltas that enforce constraints over the color fluxes on each link $(x,\mu)$ of the lattice: \begin{gather} J_{x,\mu}^{11} \, + k_{x,\mu}^{11} \, - \overline{k}_{x,\mu}^{11} \; = \; J_{x,\mu}^{22} \, + k_{x,\mu}^{22} \, - \overline{k}_{x,\mu}^{22} \qquad \forall x,\mu \, , \label{eq:su2_fullconstraints1} \\ J_{x,\mu}^{12} \, + k_{x,\mu}^{12} \, - \overline{k}_{x,\mu}^{12} \; = \; J_{x,\mu}^{21} \, + k_{x,\mu}^{21} \, - \overline{k}_{x,\mu}^{21} \qquad \forall x,\mu \, . \label{eq:su2_fullconstraints2} \end{gather} In these constraints, which are illustrated schematically in Fig.~\ref{fig:su2_fullconstraints}, the color link fluxes receive contributions both from the gauge and the fermion fields. In our reformulation the gauge fields are represented by the cycle occupation numbers $p_{x,\mu\nu}^{abcd} \in \mathbb{N}_{0}$, which are variables assigned to the plaquettes. We write the contribution they give to the color link fluxes by means of the $J_{x,\mu}^{ab}$ fluxes (\ref{eq:su2_jfluxes}), which sum up all the cycle occupation numbers that are attached to the link $(x,\mu)$ and that have color path $(a,b)$ on that link. They are depicted in blue full lines in Fig.~\ref{fig:su2_fullconstraints}. \begin{figure}[t] \begin{center} \vskip5mm \includegraphics[scale=0.8,clip]{chapters/plots/su2_fullconstraints.pdf} \end{center} \caption{Geometrical illustration of the generalized constraints that combine the gauge fluxes $J_{x,\mu}^{ab}$ (full lines) and the matter fluxes $k_{x,\mu}^{\,ab} - \overline{k}_{x,\mu}^{\,ab}$ (dashed lines). Similar to the pure gauge case shown in Fig.~\ref{fig:su2_constraints} the sum over all 1-1 fluxes must equal the sum over all 2-2 fluxes and the sum over 1-2 fluxes must equal the sum over 2-1 fluxes. \label{fig:su2_fullconstraints}} \end{figure} The contribution of the fermions comes from the link variables $k_{x,\mu}^{ab} \in \{0,1\}$ and $\overline{k}_{x,\mu}^{ab} \in \{0,1\}$, and is shown with dashed purple lines in the figure. The final expression for the partition sum in our dual representation then is: \begin{align} Z = &\frac{1}{2^{2V}} \!\!\sum_{\{p,k,\overline{k},s \}} \!\! C_{F}[s,k,\overline{k}\,] \; W_M[s] \, W_\beta [p] \, W_H [p,k,\overline{k}]\; \prod_{x,\mu} (-1)^{\, J_{x,\mu}^{\,21} + k_{x,\mu}^{\,21} + \overline{k}_{x,\mu}^{\,21}} \; \prod_{\mathcal{L}} \, \sign \,(\mathcal{L}) \nonumber \\ \times & \prod_{x, \mu} \delta \!\left(\! J_{x,\mu}^{\,11} \!+\! k_{x,\mu}^{\,11} \!-\! \overline{k}_{x,\mu}^{\,11} - [J_{x,\mu}^{\,22} \!+\! k_{x,\mu}^{\,22} \!-\! \overline{k}_{x,\mu}^{\,22} ]\! \right) \delta \!\left(\! J_{x,\mu}^{\,12} \!+\! k_{x,\mu}^{\,12} \!-\! \overline{k}_{x,\mu}^{\,12} - [J_{x,\mu}^{\,21} \!+\! k_{x,\mu}^{\,21} \!-\! \overline{k}_{x,\mu}^{\,21} ]\! \right). \label{eq:su2_partitiondualfull} \end{align} It sums over all the possible configurations of the cycle occupation numbers $p_{x,\mu\nu}^{abcd} \in \mathbb{N}_{0}$, and the fermion dual variables $s_{x}^{a}, k_{x,\mu}^{ab}, \overline{k}_{x,\mu}^{ab}\in \{0,1\}$. The fermionic variables have to satisfy the constraint $C_{F}[s,k,\overline{k}]$ given in Eq.~(\ref{eq:su2_cf}), which ensures that the Grassmann integrals are saturated by requiring the two layers of our four-dimensional lattice to be filled with monomers, dimers and loops. Monomers come with the weight factor $W_{M}[s]$ (\ref{eq:su2_wm}), while dimers and loops, along with the cycle occupation numbers, give contributions to the combinatorial factor $W_{H}[p,k,\overline{k}]$. Loops, together with the cycle occupation numbers, are further constrained by the product of Kronecker deltas in Eq.~(\ref{eq:su2_partitiondualfull}) which impose equalities between different components of the total color flux at each link $(x,\mu)$. Notice that the dual representation we obtain suffers from the sign problem, which has two different origins: $(-1)^{\, J_{x,\mu}^{\,21} + k_{x,\mu}^{\,21} + \overline{k}_{x,\mu}^{\,21}}$ is the gauge sign, that can be traced back to the negative sign of the (2,1) entry in the parametrization (\ref{eq:su2_parametrization}) of the SU(2) matrices, and $\sign(\mathcal{L})$ is the fermion sign, which depends only on the geometry of the fermion loop $\mathcal{L}$. In the next and final section of this chapter we will show how the interplay of these two sources of sign makes the first terms of the joint hopping and strong coupling expansion positive. \section{The strong coupling limit \label{sec:su2_sc}} The aim of this section is to analyse the contributions to the partition function in its dual form (\ref{eq:su2_partitiondualfull}). We initially focus on the strong coupling limit \cite{PhysRevD.10.2445}, i.e., $\beta = 0$, and then add finite $\beta$ corrections. In the strong coupling limit all the cycle occupation numbers must be 0, as shown by the fact that $\lim_{\beta \to 0} W_{\beta}[p] = 0$ unless $p_{x,\mu\nu}^{abcd} = 0$ $\forall x,\mu,\nu,a,b,c,d$. Consequently, also the $J$- and $S$-fluxes vanish and, in particular, the constraints (\ref{eq:su2_fullconstraints1}) and (\ref{eq:su2_fullconstraints2}) reduce to: \begin{equation} k_{x,\mu}^{\,11} \!-\! \overline{k}_{x,\mu}^{\,11} \; = \; k_{x,\mu}^{\,22} \!-\! \overline{k}_{x,\mu}^{\,22} \quad \forall x, \mu \qquad \mbox{and} \qquad k_{x,\mu}^{\,12} \!-\! \overline{k}_{x,\mu}^{\,12} \; = \; k_{x,\mu}^{\,21} \!-\! \overline{k}_{x,\mu}^{\,21} \quad \forall x, \mu \; . \label{strongconstraints} \end{equation} From these constraints one infers that only special types of loops can be used to fill the lattice together with monomers and dimes. The four combinations of $k$-fluxes that give rise to admissible loop segments are (refer to Fig.\ref{fig:su2_strongloops} for a graphical illustration): \begin{equation*} k_{x,\mu}^{\,11} \, = \, k_{x,\mu}^{\,22} \, ,\quad \overline{k}_{x,\mu}^{\,11} \, = \, \overline{k}_{x,\mu}^{\,22} \, , \quad k_{x,\mu}^{\,12} \, = \, k_{x,\mu}^{\,21} \, ,\quad \overline{k}_{x,\mu}^{\,12} \, = \, \overline{k}_{x,\mu}^{\,21} \, . \end{equation*} \begin{figure}[t] \begin{center} \vskip5mm \includegraphics[scale=0.53,clip]{chapters/plots/su2_strongloops.pdf} \end{center} \caption{Loop elements in the strong coupling limit. Only the four double fluxes shown here are admissible for loops in the strong coupling limit. These strong coupling loops, together with monomers and the dimers shown in Fig.~\ref{fig:su2_dimers} have to completely fill the lattice. The elements with crossing color flux come with an explicit minus sign. \label{fig:su2_strongloops}} \end{figure} This means that admissible strong coupling loops are built from segments of two units of flux that run in the same direction. The fluxes can either run parallel in color space, or cross. The flux crossings come with negative sign, as we discussed in Sec.~\ref{sec:su2_acc}. Summarizing, at strong coupling the admissible configurations are the ones that completely fill the lattice with monomers, dimers (represented in Fig.~\ref{fig:su2_dimers}), and loops made of the strong coupling loop elements depicted in Fig.~\ref{fig:su2_strongloops}. \begin{figure}[t] \begin{center} \vskip5mm \includegraphics[scale=0.53,clip]{chapters/plots/su2_strongdimers.pdf} \end{center} \caption{Graphical representation of the possible dimers. Dimers saturate the Grassmann integrals for one of the two colors of two neighboring sites. Four different color combinations are possible. \label{fig:su2_dimers}} \end{figure} Examples of such loops are shown in Fig.~\ref{fig:su2_loops}. \begin{figure}[t] \begin{center} \vskip5mm \includegraphics[scale=0.6,clip]{chapters/plots/su2_loops.pdf} \end{center} \caption{Examples of fermion loops in the strong coupling limit. The double loop on the lhs.\ has a positive sign since here the signs of the individual loops in the color 1 and the color 2 layers are the same and thus this sign squares to $+1$. The loop on the rhs.\ differs from the lhs.\ by one link element with a color crossing which comes with an explicit minus sign. On the other hand the color crossing also connects the two loops from the lhs.\ into a single loop such that there is one less minus sign from the number of loops. In total the loop on the rhs.\ thus also has a positive sign. \label{fig:su2_loops}} \end{figure} The configuration on the lhs.\ of the plot has evidently a positive sign, since it is made of two identical loops $\mathcal{L}$ running parallel to each other on the two different color layers ($(\sign \mathcal{L})^2 = 1$). The configuration on the rhs.\ has a flux crossing, which gives rise to a factor of $-1$. On the other hand, the crossing connects the previously disconnected loops into a single loop, such that we have one minus sign less, while all other signs from the number of forward hops and from the staggered sign factors remain the same (compare with the definition of the loop sign factor (\ref{eq:su2_loopsign})). Thus the overall sign of the new loop with a single flux crossing is again $+1$. Any loop configuration at strong coupling can be built following this procedure. One starts with two units of flux running parallel along the path chosen for the loops (like in the plot on the lhs.\ of Fig.~\ref{fig:su2_loops}). Then, more general loop configurations can be obtained by interchanging loop elements of parallel flux (the two elements shown on the lhs.\ of Fig.~\ref{fig:su2_strongloops}) with flux crossings (the two elements on the rhs.\ of Fig.~\ref{fig:su2_strongloops}). Each interchange has the effect we just described: it adds a minus sign factor for the crossing and, at the same time, changes the total number of loops by $\pm 1$, thus leaving the overall sign unchanged. We conclude that loops in the strong coupling limit have positive weights. This result is remarkable because it implies that, at strong coupling, all the terms in the partition function are real and positive (recall that, for fermions, only loops come with signs). The first $\beta$ corrections can be derived easily and here, in order to illustrate such a calculation, we display the leading terms of a coupled expansion of the partition sum in $1/m$ and in $\beta$: \begin{align} Z \; = \; m^{2V} \bigg[& 1 \; + \; \left( \!\frac{1}{2m}\! \right)^{\!2} \!\! \times \frac{1}{2!} \times 16 V \nonumber \\ &+ \; \left( \! \frac{1}{2m}\! \right)^{\!4} \! \!\times \left( \! \frac{1}{2!} \! \right)^{\! 2} \!\! \times [ 128 V^2 - 264 V] \; + \; \left(\! \frac{1}{2m}\! \right)^4 \!\!\times \frac{2! + 1}{3!} \times 8 V \nonumber \\ & + \left(\! \frac{1}{2m}\! \right)^4 \!\!\times \frac{\beta}{2} \times \left( \! \frac{1}{2!} \! \right)^{\! 4} \!\! \times 192 V \; + \; \left(\! \frac{\beta}{2}\! \right)^2 \!\! \times \left( \! \frac{1}{2!} \! \right)^{\! 4} \!\! \times 48 V \nonumber \\ &+ \; {\cal O} \left(\!\left(\! \frac{1}{2m}\! \right)^{\!6\,}\!\right) + {\cal O} \left(\beta^4\right) \bigg] . \label{eq:su2_expansion} \end{align} The leading term of this expansion is the contribution of the configuration in which the lattice is saturated placing a monomer on all the $2V$ sites. The corresponding weight factor is $(2m)^{2V}$, which together with the overall factor $(1/2)^{2V}$ gives rise to the contribution $m^{2V}$. Notice that on the rhs.\ of Eq.~(\ref{eq:su2_expansion}) we wrote the factor $m^{2V}$ up front. Therefore, all further terms in the expansion are relative to the lattice completely filled with monomers. The next term corresponds to the configuration where we substitute two adjacent monomers with a dimer. We thus account for a suppressing factor of $(1/2m)^2$. The combinatorial factor $1/2!$ comes from the weight $W_{H}[p,k,\overline{k}]$, where we set $k_{x,\mu}^{ab} = \overline{k}_{x,\mu}^{ab} = 1$ on a single link $(x,\mu)$ for some color combination $a,b$. Lastly, the factor $16V$ is the degeneracy of the configuration we are considering, i.e., the number of ways to place a single dimer on our double layer lattice: on every link we have 4 ways to place a dimer (compare Fig.~\ref{fig:su2_dimers}) and, on a four-dimensional lattice, there are a total of $4V$ links. This results in $16V$ ways to place a single dimer. The following contribution, in the second line of Eq.~(\ref{eq:su2_expansion}), is the one where we place two dimers on different links. The suppressing factor this time is $(1/2m)^4$, since we interchanged four monomers with two dimers. The weight factor $W_{H}[p,k,\overline{k}] = (1/2!)^2$ is the square of the one for a single dimer. The number of ways to place two dimers on the lattice is the square of the degeneracy for a single dimer divided by two: $(16V)^2/2 = 128V^2$. To this number we have to subtract the $264V$ configurations in which the two dimers are either on the same link, or touch (not admissible). The second term in the second line of Eq.~(\ref{eq:su2_expansion}) accounts for the configurations where two dimers sit on the same link. There are only two ways of placing two dimers on one link: they must be either parallel ($k_{x,\mu}^{11} = \overline{k}_{x,\mu}^{11} = k_{x,\mu}^{22} = \overline{k}_{x,\mu}^{22} = 1$) or crossing ($k_{x,\mu}^{12} = \overline{k}_{x,\mu}^{12} = k_{x,\mu}^{21} = \overline{k}_{x,\mu}^{21} = 1$). Together with the number of links this gives $2\times 4V = 8V$ possibilities. $W_{H}[p,k,\overline{k}] = 2!/3!$ is the combinatorial weight for each of those configurations. The weight $1/3!$ accounts for self-crossing loops sitting on a single link. The subsequent term, i.e., the first term in the third line in Eq.~(\ref{eq:su2_expansion}), is the contribution of the configurations with a single fermion loop around a plaquette for which the gauge constraints (\ref{eq:su2_fullconstraints1}) and (\ref{eq:su2_fullconstraints2}) are satisfied with an ACC, thus coming with a factor $\beta/2$ for the non-trivial value of one cycle occupation number. The factor $(1/2m)^4$ is there because we replaced the monomers on the four corners of a plaquette with a fermion loop. For every plaquette there are $16$ different fermion loops we can place (compare with Fig.~\ref{fig:su2_allcycles}), each with two orientations. Together with the number of plaquettes, this gives $2 \times 16 \times 6V = 192V$. Then we must adequately place an ACC such that the gauge constraints are satisfied at every link of the plaquette. Since the ACCs only have positive orientation, there is just one way of doing so for every fermion loop: for positively oriented fermion loops (see the example on the lhs.\ in the top row of Fig.~\ref{fig:su2_saturateloops}) the ACC has the opposite color at each corner, while for negatively oriented fermion loops the ACC runs alongside the fermion loop but with opposite orientation (rhs.\ top of Fig.~\ref{fig:su2_saturateloops}). In both cases we have two units of flux on each link of the plaquette, such that we have the additional factor $W_H [p, k, \overline{k}] = (1/2!)^4$. Finally, the last term is for the configurations where the fermion constraint is fulfilled just with monomers, but where we place two ACCs on top of each other, such that they satisfy the gauge constraints. Since the ACCs only have positive orientation, the admissible configurations are the ones for which the matching ACCs have opposite color at each corner, thus we can form 8 matching pairs from the 16 ACCs. Counting the $6V$ plaquettes of the lattice, this results in $48V$ possibilities. The two non-trivial cycle occupation numbers come with a weight of $(\beta/2)^2$ and, since they generate two units of flux on the four links of the plaquette, $W_H [p, k, \overline{k}] = (1/2!)^4$ also for these configurations. Subsequent terms of the coupled strong coupling and hopping expansion can be computed following the same steps we discussed for the leading terms. The plot at the bottom of Fig.~\ref{fig:su2_saturateloops} shows an example of a configuration contributing at $\mathcal{O}\Big(\big(\frac{1}{2m}\big)^6 \beta^2\Big)$. We remark that we cross-checked some of the terms of the series (\ref{eq:su2_expansion}) with previous results: the term depending only on $\beta$ matches the corresponding term of conventional strong coupling expansion \cite{PhysRevD.10.2445}, while the leading $(2m)^{-2}$ and $(2m)^{-4}$ contributions of the $\beta$-independent terms were verified by comparison to the free case. We also stress that, in our dual formulation, the leading contributions of the expansion (\ref{eq:su2_expansion}) for the partition sum $Z$ are real and positive. Negative configurations only appear at $\mathcal{O}(\beta^4)$ (see Fig.~\ref{fig:su2_unhappy}) or at $\mathcal{O}\Big(\big(\frac{1}{2m}\big)^4 \beta^3\Big)$ when a fermion loop is included. \begin{figure}[t] \begin{center} \vskip5mm \includegraphics[scale=0.5,clip]{chapters/plots/su2_saturateloops.pdf} \end{center} \caption{Example for the saturation of a fermion loop with ACCs. \label{fig:su2_saturateloops}} \end{figure}

45eca41c965d312e357575d42f2eb27fcfab6415

\subsection{Score computations with LSH}\label{sec:scores} At a high level, the score $s_p$ of a curve $p \in S$ with query $q$ is given by the normalized number of collisions with $q$ under $L\geq 1$ hash functions from the LSH scheme $\mathcal G_\delta^k$ described below, where $\delta$ and $k$ are suitable parameters. {\bf LSH scheme $\mathcal G_\delta^k$.} Our starting point is the LSH scheme $\mathcal{\hat G}_\delta$ in~\cite{driemel_locality-sensitive_2017}, which maps each curve into a smaller curve with vertices from a random shifted grid \[ G_{\delta,t} = \brc{ (x_1,\dots,x_d) \in \Re^d ~|~ \forall~ i\in [d] ~\exists~ j \in {\rm I\!\hspace{-0.025em} N}:~ x_i = j \cdot \delta +t} \] where $\delta>0$ is the side of the grid and $t=(t_1,\ldots t_d)$ is a random variable uniformly distributed in $[0,\delta)^d$. For a curve $p$ with vertices $p_1,\dots,p_{m}$, the function $g_{\delta,t}(p)$ returns the curve obtained by: 1) replacing each vertex $p_i$ with its closest grid vertex in $G_{\delta,t}$; 2) removing consecutive duplicates in the new curve. The LSH family $\mathcal{\hat G}_\delta$ is defined as $\mathcal{\hat G}_\delta=\{g_{\delta,t}, \forall t\in [0,\delta)^d\}$. We also define $\mathcal{\hat G}^k_\delta$ as the LSH family obtained by concatenating $k\geq 1$ copies of hash functions uniformly and independently selected in $\mathcal{\hat G}_\delta$. We have that $\Pr_{g^k \in \mathcal {\hat G}^k_{\delta}}[g^k(q)=g^k(p)]= \Pr_{g\in \mathcal{\hat G}_{\delta}}[g(q)=g(p)]^k$: the lower collision probability of far curves allows to decrease false positives. FRESH\xspace requires the computation of a large number of hash values in $\mathcal{\hat G}^{k}_\delta$: indeed, $k \cdot L \cdot n$ hash values are computed at construction time and $k\cdot L$ hash values for each query. We speed up the hash computation with the tensoring approach. Tensoring was initially proposed in~\cite{andoni_efficient_2006} and then further studied in~\cite{christiani_fast_2017}; to the best of our knowledge, it has only been used in practice in~\cite{SundaramTSMIMD13}. The tensoring approach generates $L$ hash functions building on two collections of $\sqrt{L}$ hash functions, reducing the actual number of hash computations by a $\sqrt{L}$ factor. Specifically, let $\Lambda_1=\{g_1,\ldots, g_{L'}\}$ and $\Lambda_2=\{g'_1,\ldots, g'_{L'}\}$ be two groups of $L'=\sqrt{L}$ random hash functions from $\mathcal{\hat G}_{\delta}^{k/2}$. Then, it is possible to construct $L'\cdot L'=L$ LSH hash functions from $G_\delta^k$ by concatenating the pair $(g_i,g'_j)$ for all possible values of $i$ and $j$ in $\{1,\ldots L'\}$. This technique reduces the number of hash value computations for the initial data structure construction from $k \cdot L \cdot n$ to $k \cdot \sqrt{L}\cdot n$, and for the query procedure from $k\cdot L$ to $k \cdot \sqrt{L}$. Finally, as storing and searching signatures is quite inefficient, we map all signatures on integers with the multiply-shift hashing scheme $\mathcal H$ in~\cite{Dietzfelbinger97}. We denote with $\mathcal G_{\delta}^{k}$ the LSH hash family obtained by first using the tensoring approach to construct (a subset of) $\mathcal{\hat G}_{\delta}^{k}$, and then by applying the multiply-shift hashing $\mathcal H$ on the signature. We observe that the signature of a curve does not need to be generated and stored: while we scan a curve $p$ to compute its signature, the hash value $h(g(p))$ is built on the fly. {\bf Data structure.} The data structure of FRESH\xspace for efficiently computing the scores leverages on the traditional approach for solving range search with LSH. $L\geq 1$ hash functions $g_1,\ldots, g_L$ are randomly chosen from the above LSH family $\mathcal G_\delta^k$, for suitable values of $\delta$ and $k$; then for each $g_i$, a hash table $H_i$ is created for storing the $n$ input curves partitioned by $g_i$. For each query $q$, we compute the multiset $T_q = \cup_{i=1}^L H_i(g_i(q))$, where $H_i(g_i(q))$ denotes the set of curves colliding with $q$ under $g_i$. If $t\in T_q$ and its multiplicity in $T_q$ is $\hat s_t$, then its score $s_t$ is $\hat s_t/L$. Note that the hash tables do not need to store the complete curves but just their identifiers: thus, the space required by the data structure is $I+\BT{Ln}$ memory words, where $I$ is the number of words to store $S$. \subsection{Filtering false positives}\label{sec:filtering} All curves with non-zero score are not too far from the query: indeed, if the hash function uses a grid of side length $\delta$, then all colliding curves have maximum distance $\delta$. However, as in general $\delta > r$ (in our experiments $\delta = 4dr$, where $d$ is the point dimension), we may report some curves with distance in $(r,\delta]$. To improve the precision, a simple approach is to set a threshold $\Delta$ and verify all curves with scores less than $\Delta$. However, the limitations of this approach are: 1) it is not clear how to select the best $\Delta$ as it might be query dependent; 2) $\Delta$ does not directly allow to trade precision and running time. The approach used in FRESH\xspace is to verify a fraction $\tau$, with $0\leq \tau \leq 1$, of the curves in $O_q$ with smaller scores. The parameter $\tau$ can be used for trading performance (with $\tau=0$ no curve in $O_q$ is verified) with precision (with $\tau=1$, all curves in $O_q$ are verified which implies a 100\% precision). \subsection{Verifying the Fréchet distance}\label{sec:frechetalgo} Verifying that two curves $p$ and $q$ are within Fréchet distance $r$ is an expensive operation~\cite{Bringmann14}: to speed up this operation, we introduce the procedure \textsc{Verify} for checking if two curves $p$ and $q$ have continuous Fréchet distance less than or equal to $r$. \textsc{Verify} consists of two procedures, named \textsc{VerifySimpl} and \textsc{VerifyHeur}, that exploit strategies from~\cite{DriemelHW12,BaldusB2017,BuchinDDM17}: each procedure can successfully show that $d_F(p,q) \le r$ or $d_F(p,q) > r$, or it can fail and do not provide an answer. Procedure \textsc{VerifyHeur} exploits the heuristics \emph{Equal-time alignment}~\cite{BuchinDDM17}, \emph{Greedy algorithm}~\cite{BaldusB2017} and \emph{Negative filter}~\cite{BaldusB2017}, and it stops as soon as one of them succeeds. On the other hand, procedure \textsc{VerifySimpl} is a decision procedure based on the concept of simplification in \cite{DriemelHW12}: $p$ and $q$ are mapped on suitable smaller trajectories $p'$ and $q'$ through a transformation based on a parameter $\epsilon\geq 0$ ($\epsilon=0$ gives the original curves). Evaluating distance predicates on $p'$ and $q'$ allows to answer distance predicates on $p$ and $q$, by suitable setting the parameter $\epsilon$. Procedure \textsc{VerifyHeur} is the application of the following heuristics, stopping as soon as one of them succeeds. \begin{itemize} \item \emph{Equal-time alignment}~\cite{BuchinDDM17}. This heuristic performs a traversal of the two curves moving at the same speed on both, providing an upper bound to the Fréchet distance. If we define $\Phi_{x}(t) = t x$, this heuristic verifies \[ \max_{t \in [0, 1]} || p_{\Phi_{|p|}(t)} - q_{\Phi_{|q|}(t)} ||_2\le r \] which can be done in linear time. \item \emph{Greedy algorithm}~\cite{BaldusB2017}. It provides an upper bound on the continuous Fréchet distance by finding an alignment with a greedy approach. We construct the following traversal of $p$ and $q$: 1) $p_1$ and $q_1$ are matched; 2) after matching vertices $p_i$ and $q_j$, we match $p_{i'}$ and $q_{j'}$, for $(i',j')\in\{(i+ 1, j),(i, j + 1),(i+ 1, j + 1)\}$ minimizing $\|p'_{i'}-q_{j'}\|_2$. We ignore from these three options the ones that would make $i > |p|$ or $j > |q|$. If during the whole traversal we stayed at distance $\le r$, we can conclude that $p$ and $q$ are $r$-near. \item \emph{Negative filter}~\cite{BaldusB2017}. This heuristic seeks to prove that, for some vertices of $p$, there are no vertices of $q$ within distance $r$ they can be aligned to, providing a certificate that the two curves are at distance greater than $r$. For each vertex $p_j$ of $p$, we define $q^\leftarrow_{p_j}$ as the first vertex of $q$ that can be aligned with $p_j$. For this to be possible, such a vertex needs to be within distance $r$ from $p_j$, and needs to appear on $q$ after vertex $q^\leftarrow_{p_{j-1}}$, because of the definition of Fréchet distance. Since the first vertex of $p$ has to be aligned with the first vertex of $q$, we have that $q^\leftarrow_{p_1}=q_1$. Then, for $j \in [2, |p|]$ the heuristic proceeds in trying to define $q^\leftarrow_{p_j}$. If for some $j$ this is not possible, then $p$ and $q$ are farther than $r$. This heuristic is not symmetric, therefore we can apply it two times swapping arguments. \item \emph{Full verify}. If all of the above heuristics fail to verify the distance, we apply the exact algorithm in~\cite{AltGodau95} based on free space diagram. \end{itemize} To further speedup the verification of a pair of curves $p$ and $q$, we also adopt the decision procedure proposed in~\cite[Lemma 3.6]{DriemelHW12}, which we deem here \textsc{VerifySimpl}. This scheme is based on the concept of $\mu$-simplification (also presented in~\cite{DriemelHW12}), constructed as follows for a curve $p$ and $\mu > 0$. First mark $p_1$ and set it as the current vertex. Then, scan the curve from the current vertex until we reach the first $p_j$ such that $||p_j - p_1||_2 > \mu$: we mark $p_j$ and set it as the current vertex. The procedure is repeated until we reach the last vertex, which is marked as well. The marked vertices make up the simplified curve, which is denoted with $\simpl{p, \mu}$ and is computed in linear time. The decision scheme builds simplifications of $p$ and $q$, controlled by a parameter $\epsilon > 0$. Let $r'=r/(1+\epsilon/3)$. Define $\mu^-=r\epsilon/28$ and $\mu^+=r\epsilon/(28 \cdot(1+\epsilon/3))$, and let \[ r^- = r\cdot\left(1 + \epsilon/14\right) \quad \mbox{and} \quad r^+ = r\cdot\left(\frac{3(1+\epsilon/14)}{3 + \epsilon}\right) \] note that $r^+ < r^-$. First, we verify with \textsc{VerifyHeur} if \[d_F(\simpl{p, \mu^-}, \simpl{q, \mu^-}) > r^-\] If this is the case, the procedure reports that $d_F(p,q)> r$. Otherwise, we further verify with \textsc{VerifyHeur} if \[d_F(\simpl{p, \mu^+}, \simpl{q, \mu^+}) \le r^+\] If the answer is affirmative, we report $d_F(p,q) \le r$. It may be that neither of the two checks gives a positive answer. In this case, the procedure reports that it cannot give an answer. Procedure \textsc{Verify} is then the following: \begin{enumerate} \item In the first stage, we only consider the first ($p_1$ and $q_1$) and last vertices ($p_{|p|}$ and $q_{|q|}$) of $p$ and $q$. If $||p_1-q_1||_2>r$ or $||p_{|P|}- q_{|Q|}||_2>r$, then the two curves cannot be $r$-near by the definition of continuous traversal. We call this heuristic \textsc{Endpoints}. \item In the second stage, we look at the bounding boxes of the two curves. If the $\ell_1$ distance of corresponding corners of the bounding boxes is larger than $r$, then the two curves cannot be $r$-near~\cite{DutschV17}. We call this heuristic \textsc{BoundingBox}. \item In the third stage, we use \textsc{VerifySimpl} with decreasing values of $\epsilon$ (which will be fixed in the experimental analysis), corresponding to simplifications becoming less aggressive. For a given $\epsilon$, if \textsc{VerifySimpl} can give an answer, then we return it, otherwise we move to the next $\epsilon$. \item The fourth stage runs if none of the calls to \textsc{VerifySimpl} could return an answer: in this case we return the result of the invocation of \textsc{VerifyHeur} on the original curve. \end{enumerate} \subsection{Experimental setup}\label{sec:setup} \paragraph{\bf Hardware} We implement our algorithm in C++ with OpenMP, using the \texttt{gcc} compiler version 4.9.2. We run the experiments on a Debian GNU/Linux machine (kernel version 3.16.0) equipped with 24GB of RAM, and an Intel I7 Nehalem processor (clock frequency 3.07GHz). \paragraph{\bf Datasets} As benchmarks we use datasets from the UCR collection~\cite{UCRArchive}, which is comprised of 85 datasets of trajectories in one dimension. For brevity, we report on the 7 largest datasets of this collection. We also include in our benchmark a dataset of road trips in San Francisco that was used in the SIGSPATIAL 2017 challenge~\cite{werner2018acm}, along with the TDrive dataset~\cite{t-drive-dataset}. Both are datasets of trajectories in 2 dimensions. For each dataset, we perform a self-similarity join using a set of fixed Fréchet distance thresholds, by solving the $r$-range search problem for each curve of the dataset. The thresholds are set to the first and fifth percentiles of the pairwise distances for any given dataset, so that the output size is 1\% and 5\% of the number of possible pairs, respectively. Given the large number of possible pairs, these percentiles are computed on the pairwise distances of a sample of 1000 points of each dataset. Figure~\ref{fig:distance-distribution} gives the distribution of pairwise distances in the datasets we are considering. Each result is the average over at least 5 runs. \begin{figure*}[t] \includegraphics[width=\textwidth]{imgs/distances.pdf} \caption{Distribution of pairwise distances for all the datasets considered. The green line highlights the first percentile, the red one highlights the fifth percentile. \label{fig:distance-distribution}} \end{figure*} \begin{table} \newcommand{~$\star$}{~$\star$} \newcommand{~$\ddagger$}{~$\ddagger$} \begin{center} \small \begin{tabular}{llr} \toprule dataset & range & best time \\ \midrule Chlorine & 0.34 (first) & 24~$\star${} \\ & 0.52 (fifth) & 91~$\star${} \\ ECG5000 & 0.62 (first) & 29~$\star${} \\ & 0.92 (fifth) & 102~$\star${} \\ FordA & 1.07 (first) & 299~$\star${} \\ & 1.20 (fifth) & 1190~$\star${} \\ yoga & 0.14 (first) & 23~$\star${} \\ & 0.33 (fifth) & 87~$\star${} \\ SanFrancisco & 5213.21 (first) & 413~$\star${} \\ & 9205.43 (fifth) & 417~$\ddagger${} \\ StarLightCurves & 0.13 (first) & 548~$\star${} \\ & 0.21 (fifth) & 2949~$\star${} \\ TDrive & 0.17 (first) & 3913~$\star${} \\ & 0.23 (fifth) & 20372~$\star${} \\ TwoPatterns & 0.56 (first) & 76~$\star${} \\ & 0.68 (fifth) & 121~$\star${} \\ wafer & 0.14 (first) & 70~$\star${} \\ & 0.39 (fifth) & 134~$\star${} \\ \bottomrule \end{tabular} \end{center} \captionof{table}{Baseline times (in seconds) for the two different radii, which are defined, respectively, as the first and fifth percentile of all pairwise distances. Results marked with ~$\ddagger${} were obtained using the code by Baldus et al.~\cite{BaldusB2017}, the ones marked with ~$\star${} were obtained using the code by Dutsch et al.~\cite{DutschV17}. \label{tab:baseline}} \end{table} \paragraph{\bf Baseline} To establish a baseline, we ran the code provided by the three winners of the SIGSPATIAL 2017 challenge~\cite{BaldusB2017,BuchinDDM17,DutschV17}, compiled with all optimizations enabled and ran with the default parameters. Table~\ref{tab:baseline} reports these results. \begin{figure*}[t] \includegraphics[width=\textwidth]{imgs/precision-recall.pdf} \caption{Performance in terms of precision and recall of FRESH\xspace on all the datasets considered. The color of a point denotes the number of repetitions $L$, while the shape of a point represents the number of concatenations $k$.\label{fig:precision-recall}} \end{figure*} \subsection{Evaluating the LSH scheme} \label{sec:lshclassifier} \begin{figure*}[t] \begin{minipage}{.73\textwidth} \centering \includegraphics[width=\columnwidth]{imgs/mass-precision-correlation.pdf} \end{minipage} \begin{minipage}{.25\textwidth} \caption{Fraction of pairs below $4dr$ versus precision, for $k=2$, and $L=1024$. On dataset where such a fraction is high, the precision of the LSH scheme tends to be low. \label{fig:mass}} \end{minipage} \end{figure*} \begin{figure*}[t] \includegraphics[width=\textwidth]{imgs/wafer_positives.png} \caption{Curve 0 of the wafer dataset as a query (green) for $r=0.14$, $k=2$ and $L=1024$, in the context of relevant curves with respect to the LSH scheme: false positives (orange), false negatives (red) and true positives (blue). The spacing of the grid along the value axis is equals to $4r$, which is the size of the grid used by the LSH scheme for building signatures.\label{fig:wafer}} \end{figure*} We analyze how the LSH scheme affects the performance and quality of FRESH\xspace without the partial verification. In other words, each pair colliding in at least one of the $L$ repetitions (i.e., with a non-zero score) is reported as a positive match, without further verification. We test this setup using hash values obtained as the concatenation of $k=1,2,4$ hash functions and with $L=128,256,512,1024$ repetitions, setting the grid size to $\delta=4dr$. Figure~\ref{fig:precision-recall} reports, for each dataset and combination of parameters, the performance in the precision-recall space. The recall is the fraction of true positives reported by the algorithm over all the positives in the ground truth, whereas the precision is the fraction of true positives over the predicted positives (i.e., the sum of true positives and false positives). Both scores range from 0 to 1, with 1 being the best, hence in the plots of Figure~\ref{fig:precision-recall} we have that the closer the top right corner, the better the performance. Note that we use the precision instead of the false positive rate due to the large number of negatives in the ground truth, which makes very easy to attain a small false positive rate. In general, we have that increasing the number of repetitions $L$ improves the recall, lowering the precision, as expected. Symmetrically, increasing $k$ makes the LSH more selective, hence it increases the precision, at the expense of the recall. Note that on some datasets our LSH technique is more effective than on others. In general, using sufficiently many repetitions we can get good recall, while getting a good precision is harder, and may be very costly in terms of recall. We will address this problem in the next subsection. On the SanFrancisco and TDrive datasets we get perfect recall and low precision, almost irrespective of the configuration of parameters. This is due to the distance distribution of these datasets: by setting the query range to the first and fifth percentiles of distances, the algorithm constructs grids with a resolution so large that almost all curves collide with the queries. \begin{figure*}[t] \includegraphics[width=\textwidth]{imgs/scores-hist-2-1024.pdf} \caption{ The distribution of scores assigned to colliding pairs for $k=2$ and $L=1024$, with query radius equals to the first percentile of distances, shows that the majority of false positive pairs (fp, in orange) have lower scores than the true positive colliding pairs (tp, in blue), with some overlapping of the two distributions. The results for other configurations of parameters are similar. Note that in this plot each orange (resp. blue) bar is scaled with respect to the total number of false positives (resp. true positives) and not the total number of colliding pairs: this is to appreciate the overall distribution. \label{fig:score-distribution}} \end{figure*} Among the others, the wafer dataset deserves a particular attention. For the query range equals to the first percentile of the pairwise distances, Figure~\ref{fig:precision-recall} shows that the recall is just slightly above 0.5 at best. While a low precision can be fixed for all datasets, as we shall see in the next subsection, the recall on wafer seems resistant to increases of $L$. To understand why this happens, we can look at the behavior of a single query, as reported in Figure~\ref{fig:wafer}. Along with the one-dimensional query curve itself, we plot two curves that collide with the query under the LSH scheme, one false positive and one true positive, and a curve that did not collide but should have, i.e. a false negative. In terms of recall, the false negatives are the relevant curves to look at: having zero false negatives implies a perfect recall. Therefore, the poor performance on wafer is due to the fact that many curves are classified as being far from the query when they are actually close, which happens if the misclassified curve and the query do not collide in any of the $L$ repetitions. Looking at Figure~\ref{fig:wafer} we can see why this happens. The query (green curve), has a sudden jump downward around time 25, with no vertices in the segment connecting the extremes of the jump. The false negative curve (in red) has a similar jump around time 18. However, in this case, there is one vertex between the extremes of the jump. Under the LSH scheme described in Section~\ref{sec:algo}, two curves collide (and hence have a non-zero score) only if they have the same signature, which is computed by snapping vertices to a randomly shifted grid of resolution $4dr$, i.e. $4r$ for one-dimensional dataset such as wafer. The grid of Figure~\ref{fig:wafer} has a resolution $4r$ along the value axis. It is clear that, no matter the random shift of the grid, the point of the red curve in the middle of the jump will never snap to the same grid line as any point of the green curve in the analogous jump, because no such point exists. A simple solution to this problem is to add more vertices to the curves, by interpolation, in the jumps. This preprocessing does not change the Fréchet distance between any two curves. \subsection{Improving the precision by partial verification} \label{sec:exp:partial-verification} In this section we verify the trade-off between precision and running time proposed in Section~\ref{sec:filtering}. From the previous experiments we selected a configuration of parameters striking a good balance of recall and precision on most datasets: $k=2$ and $L=1024$. For $\tau\in\{0, 0.1, 0.2, 0.5, 1\}$ we run the algorithm evaluating the $\tau m$ pairs with lowest non-zero scores, where $m$ is the number of pairs with non-zero scores. When $\tau=0$, the algorithm runs in the same configuration used in the previous subsection, when $\tau=1$ the algorithm verifies all the colliding pairs. We apply 3 simplifications in the verification pipeline, using $\epsilon=10,1,0.1$, from coarsest to finest. First, we consider the distribution of scores before any verification happens, to assert that verifying the lowest-score pairs is actually sound (Figure~\ref{fig:score-distribution}). We have that the false positive pairs (colored in orange) have lower scores than the true positive colliding pairs (in blue), with some overlapping of the two distributions. Therefore, verifying pairs starting from the low-score ones seems like a sensible choice, since we are likely to get rid of many false positives, which we expect to improve the recall. Note that verifying some pairs does not remove true positives (neither it can introduce them), therefore the recall remains unchanged, irrespective to the fraction of pairs $\tau$ that we verify. \begin{figure*}[t] \includegraphics[width=\textwidth]{imgs/precision.pdf} \includegraphics[width=\textwidth]{imgs/speedup.pdf} \caption{ Precision and speedup per pair given for varying $\tau$, for $k=2$, $L=1024$. The black line on the speedup plots marks speedup 1, i.e. the performance of the best baseline algorithm. \label{fig:partial-evaluation}} \end{figure*} We now move to assess the influence of the fraction of verified pairs $\tau$ on the precision and the runtime performance (Figure~\ref{fig:partial-evaluation}). For measuring the latter, we focus on the \emph{speedup}, defined as the ratio between the time of the baseline and LSH based algorithm. As we expect, increasing $\tau$ increases the precision, with perfect precision when $\tau=1$, when all the pairs are verified and the algorithm reports no false positives. The speedup decreases with the increase of $\tau$: this is because we evaluate more and more pairs, which is a costly operation. We observe that on two-dimensional trajectories, the speedup that can be obtained is larger than on one-dimensional datasets, even at higher precision values. \begin{figure*}[t] \includegraphics[width=\textwidth]{imgs/breakdown.pdf} \caption{ Breakdown of the effect of the various heuristics used to decide whether a pair is a positive match or not. The hue of the colors increases with the cost of the heuristic, so \emph{full-negative} is more expensive to compute than \emph{endpoints-negative}. \label{fig:breakdown}} \end{figure*} Finally, we analyze the contribution to the decision process of the LSH and the various heuristics employed (Figure~\ref{fig:breakdown}). We concentrate on a single run, for each dataset, with $k=2$, $L=1024$ and the radius set to the first percentile of distances, evaluating all pairs with nonzero score. The parts shaded in gray denote pairs for which the algorithm was not able to reach a decision and needed to move to the next stage. Then, parts in shades of green (resp. red) denote pairs for which a positive (resp. negative) decision was reached using one of the heuristics. The pairs excluded by the LSH scheme are shaded in blue rather than red, to remark that even if they are rejected as negatives they may contain some false negatives: the larger the blue bar, the more effective the filtering power of the LSH scheme. Some datasets are more amenable to be processed with the LSH strategy, and this is in line with the precision results reported in Figure~\ref{fig:precision-recall}. Of the pairs surviving this first filtering, several can be discarded by looking at the endpoints, as shown by the \emph{endpoint-filtering} column in the plot. The simplifications have varying degrees of effectiveness, depending on the dataset: on some datasets coarser simplifications are effective, whereas on some others we have to use finer simplifications (i.e., with a smaller $\epsilon$). \subsection{Our results} The goal of this paper is to describe and experimentally evaluate FRESH\xspace, an approximate and randomized approach for $r$-range search under the continuous Fréchet distance. FRESH\xspace builds on the theoretical ideas in \cite{driemel_locality-sensitive_2017} and extends it by providing a solid and efficient framework for trading precision and performance. {\bf Algorithm design.} The core component of FRESH\xspace is a filter based on the LSH scheme for the discrete Fréchet distance in~\cite{driemel_locality-sensitive_2017}, which is boosted with multiply-shift hashing~\cite{Dietzfelbinger97} and tensoring~\cite{andoni_efficient_2006,christiani_fast_2017} for better performance. For a given input set $S$ with $n$ curves and a query curve $q$, the filter selects as candidate near neighbors all curves colliding with $q$ under at least one of $L$ hash functions randomly selected from the LSH scheme. This filters out a significant number of curves, without even reading them. All candidates are associated with a \emph{score}, representing the fraction of collisions under the $L$ hash functions. If FRESH\xspace is seen as a classifier for detecting near and far curves for a given query $q$, the score of a curve $p$ represents the probability that $p$ and $q$ are near. The second component of FRESH\xspace is a candidate pruning step for reducing false positives (i.e., far curves marked as near). The pruning consists in verifying that the fraction $0\leq \tau\leq 1$ of candidates with smaller scores have continuous Fréchet distance from the query not larger than $r$. As verifying the Fréchet distance is a costly operation, we propose a procedure exploiting a cascade of curve simplifications from~\cite{DriemelHW12} and verification heuristics from~\cite{BaldusB2017,BuchinDDM17}: each step can successfully show that the distance is larger or not than $r$, or it can fail and do not provide an answer; the procedure applies the aforementioned simplifications and heuristics until one of them succeeds. {\bf Performance/quality trade-off.} FRESH\xspace trades the quality of the results with the overall performance by suitably settings the aforementioned $L$ and $\tau$ parameters.\footnote{In addition to parameters $\tau$ and $L$, the FRESH\xspace algorithm has other second order parameters that are introduced in Section \ref{sec:algo}, which marginally affect performance and quality. However, from an application point of view, the trade-off is mainly captured by $L$ and $\tau$, and the remaining parameters can be left to the default value in the implementation.}: We measure the quality of the results in terms of: 1) \emph{recall}, that is the fraction of true positives reported by the algorithm over all the positives in the ground truth; 2) \emph{precision}, that is the fraction of true positives over the predicted positives (i.e. the sum of true positives and false positives). By increasing the number $L$ of hash functions used in FRESH\xspace, it is possible to increase the recall of our algorithm by increasing the query time (linear in $L$) and of the space requirements (equal to $L\cdot n+I$, where $I$ is the input size). Once the recall has been fixed, it is possible to improve the precision by increasing the $\tau$ parameter at the cost of a higher query time. The recall is not affected by this step and a perfect precision is reached by setting $\tau=1$. {\bf Practical and theoretical guarantees.} We have carried out an extensive experimental evaluation of the FRESH\xspace algorithm over several datasets. To evaluate FRESH\xspace, we use it as a primitive for solving a self-similarity join on each dataset $D$: specifically, for every curve in $D$, we perform an $r$-range search query over $D$. The experiments show that the scores computed under a query $q$ provide a good indicator of the distance from $q$, and thus filtering points according with scores is a sound approach. From a performance point of view, we compare FRESH\xspace with the exact solutions that won the ACM SIGSPATIAL 2017 challenge~\cite{BaldusB2017,BuchinDDM17,DutschV17}. When the recall is approximately 70-80\% and the precision is approximately 50\%, FRESH\xspace exhibits better running times with speedups above 5x for some inputs. Although the precision is low, the returned points are never too far from the query (up to a constant factor from $r$) by the property of the LSH scheme. With higher precision, the heuristics adopted in the exact solutions, in particular the bounding box approach in \cite{DutschV17}, are very effective with the 1-dimensional datasets (i.e., time series) considered in the experiments and highlight the limitations of FRESH\xspace in this setting. FRESH\xspace is also supported by the theoretical foundations of the LSH scheme in~\cite{driemel_locality-sensitive_2017}. The FRESH\xspace algorithm is described in Section \ref{sec:algo} and the experimental results in Section \ref{sec:exp}. The code of FRESH\xspace is available at \url{https://github.com/Cecca/FRESH}. We refer to the full version~\cite{fullversion} for a more detailed coverage of our results, including the theoretical analysis bounding the collision probability and further experiments. \subsection{Related works} {\bf Similarity search for curves.} Data structures for searching among curves under the Fréchet distance have been studied under different angles. One of the earlier theoretical works is~\cite{indyk2002} that proposes a nearest neighbor data structure for Fr\'echet distance. In 2011, \cite{bcg-ffq-11} revived the topic motivated by the availability of high-resolution trajectories of soccer players in the emerging area of sports analytics. A comprehensive study of the complexity of range searching under the Fr\'echet distance appeared in~\cite{AfshaniD18}, that also gives lower bounds on the space-query-time trade-off of range searching under the Fr\'echet distance. Recently, the annual data competition within the ACM SIGSPATIAL conference on geographic information science has drawn attention to the timeliness of this problem~\cite{werner2018acm}. The focus of the challenge was on exact solutions and hence none of the awarded submissions~\cite{BaldusB2017,BuchinDDM17,DutschV17} propose approximate solutions. An LSH for the discrete Fréchet distance is described in~\cite{driemel_locality-sensitive_2017}. A follow-up paper~\cite{emiris2018} provides better theoretical approximation bounds using a slightly different approach, but their results do not apply to the setting that we focus on in this paper. Sketches for the Hausdorff and discrete Fréchet distances are proposed in \cite{Astefanoaei18}, which gives an LSH scheme with similar properties of~\cite{driemel_locality-sensitive_2017}. {\bf Verifying the Fréchet distance.} In order to improve the precision of the proposed LSH scheme, we suggest to filter the query results by verifying the distances for selected curves. However, verifying the distance is a non-trivial and expensive operation. It is known that the (discrete or continuous) Fr\'echet distance between two fixed curves cannot be decided in strictly subquadratic time in the number of vertices of the curves, unless the Strong Exponential Time Hypothesis is false~\cite{Bringmann14}. The fastest algorithms for computing the continuous and discrete Fréchet distance are described in~\cite{buchin2014four} and~\cite{avraham2014}. Both algorithms take roughly quadratic time. However, \cite{DriemelHW12} shows that one can approximate the distance in near-linear time under certain realistic assumptions on the shape of the input curves. We use this algorithm to filter the query results, in order to improve the precision of our method. \section{Introduction}\label{sec:intro} \input{intro} \section{Preliminaries}\label{sec:prelim} \input{prelim} \section{FRESH\xspace algorithm}\label{sec:algo} \input{algo} \section{Experimental evaluation}\label{sec:exp} \input{experiments} \section{Conclusion}\label{concl} \input{concl} \bibliographystyle{plain} \balance

2ab7b54f849b0e8ee70bc680a3a2b467867e68e9

\section{Introduction} \subsection{The Chowla and Elliott conjectures} Define a \emph{$1$-bounded multiplicative function} to be a function $g: \N \to \D$ from the natural numbers $\N \coloneqq \{1,2,\dots\}$ to the unit disk $\Disk \coloneqq \{z \in \C: |z| \leq 1 \}$ satisfying $g(nm) = g(n) g(m)$ whenever $n,m$ are coprime. If in addition $g(nm)=g(n)g(m)$ for all $n,m \in \N$, we say that $g$ is \emph{completely multiplicative}. In addition, we adopt the convention that $g(n)=0$ when $n$ is zero or a negative integer. This paper is concerned with the structure of higher order correlations of such functions. To describe our results, we need some notation for a number of averages. \begin{definition}[Averaging notation] Let $f: A \to \C$ be a function defined on a non-empty finite set $A$. \begin{itemize} \item[(i)] (Unweighted averages) We define $$ \E_{n \in A} f(n) \coloneqq \frac{\sum_{n \in A} f(n)}{\sum_{n \in A} 1}.$$ \item[(ii)] (Logarithmic averages) If $A$ is a subset of the natural numbers $\N$, we define $$ \E_{n \in A}^{\log} f(n) \coloneqq \frac{\sum_{n \in A} \frac{f(n)}{n}}{\sum_{n \in A} \frac{1}{n}}.$$ \item[(iii)] (Doubly logarithmic averages) If $A$ is a subset of the natural numbers $\N$, we define $$ \E_{n \in A}^{\log\log} f(n) \coloneqq \frac{\sum_{n \in A} \frac{f(n)}{n \log(1+n)}}{\sum_{n \in A} \frac{1}{n \log(1+n)}}.$$ \end{itemize} Of course, the symbol $n$ can be replaced here by any other free variable. For any real number $X \geq 1$, we use $\E_{n \leq X} f(n)$ as a synonym for $\E_{n \in \N \cap [1,X]} f(n)$, and similarly for $\E_{n \leq X}^{\log} f(n)$ and $\E_{n \leq X}^{\log\log} f(n)$. If we use the symbol $p$ (or $p_1, p_2$, etc.) instead of $n$, we implicitly restrict $p$ to the set of primes ${\P} \coloneqq \{2,3,5,7,\dots\}$, thus for instance for $X \geq 2$, $\E_{p \leq X} f(p)$ is a synonym for $\E_{p \in {\P} \cap [2,X]} f(p)$, and similarly for $\E_{p \leq X}^{\log} f(p)$. \end{definition} \begin{remark} The use of $\log(1+n)$ in the $\E^{\log\log}$ notation instead of $\log n$ is only in order to avoid irrelevant divergences at $n=1$, and the shift by $1$ may otherwise be ignored. Because of the prime number theorem, prime averages such as $\E_{p \leq X} f(p)$ are often of ``comparable strength'' to logarithmic averages $\E_{n \leq X}^{\log} f(n)$, and similarly logarithmic prime averages such as $\E_{p \leq X}^{\log} f(p)$ are of comparable strength to $\E_{n \leq X}^{\log\log} f(n)$. See Lemma \ref{chug} for a more precise statement. \end{remark} Following Granville and Soundararajan \cite{gs}, given two $1$-bounded multiplicative functions $f,g: \N \to \Disk$, and $X \geq 1$, we define the \emph{pretentious distance} $\D(f,g;X)$ between $f$ and $g$ up to scale $X$ by the formula $$ \D( f, g; X) \coloneqq \left( \sum_{p \leq X} \frac{1 - \mathrm{Re}(f(p) \overline{g(p)})}{p} \right)^{1/2}.$$ It is conjectured that multiple correlations of $1$-bounded multiplicative functions should asymptotically vanish unless all of the functions involved ``pretend'' to be twisted Dirichlet characters in the sense of the pretentious distance. More precisely, the following conjecture is essentially due to Elliott. \begin{conjecture}[Elliott conjecture]\label{elliott-conj} Let $g_1,\dots,g_k: \N \to \Disk$ be $1$-bounded multiplicative functions for some $k \geq 1$. Assume that there exists $j\in \{1,\dots,k\}$ such that for every Dirichlet character $\chi$ one has \begin{equation}\label{suptx} \inf_{|t| \leq X} \D(g_j, n \mapsto \chi(n)n^{it}; X) \to \infty \end{equation} as $X \to \infty$. \begin{itemize} \item[(i)] (Unweighted Elliott conjecture) If $h_1,\dots,h_k \in \Z$ are distinct integers, then $$ \lim_{X\to \infty}\E_{n \leq X} g_1(n+h_1) \cdots g_k(n+h_k) = 0.$$ \item[(ii)] (Logarithmically averaged Elliott conjecture) If $h_1,\dots,h_k \in \Z$ are distinct integers, then $$ \lim_{X\to \infty}\E_{n \leq X}^{\log} g_1(n+h_1) \cdots g_k(n+h_k) = 0.$$ \end{itemize} \end{conjecture} Conjecture \ref{elliott-conj}(i) was first stated by Elliott in \cite{elliott}, \cite{elliott2}, with the condition \eqref{suptx} weakened to the assertion that $\D(g_j, n \mapsto \chi(n)n^{it}; X) \to \infty$ for each fixed $t$, with no uniformity in $t$ assumed. However, it was shown in \cite{mrt} that this version of the conjecture fails for a technical reason. By summation by parts, Conjecture \ref{elliott-conj}(i) implies Conjecture \ref{elliott-conj}(ii). At present, both forms of the Elliott conjecture are known for $k=1$ (thanks to Halász's theorem \cite{halasz}), while the $k=2$ case of the logarithmic Elliott conjecture was established in \cite{tao}. Specialising the above conjecture to the case of the Liouville function\footnote{For the definitions of the standard multiplicative functions used in this paper, see Subsection \ref{subsec: not}.} $\lambda$, we recover the following conjecture of Chowla \cite{chowla}, together with its logarithmically averaged form. \begin{conjecture}[Chowla conjecture]\label{chowla-conj} Let $k \geq 1$ be a natural number. \begin{itemize} \item[(i)] (Unweighted Chowla conjecture) If $h_1,\dots,h_k \in \Z$ are distinct integers, then $$ \lim_{X\to \infty}\E_{n \leq X} \lambda(n+h_1) \cdots \lambda(n+h_k) = 0.$$ \item[(ii)] (Logarithmically averaged Chowla conjecture) If $h_1,\dots,h_k \in \Z$ are distinct integers, then $$ \lim_{X\to \infty}\E_{n \leq X}^{\log} \lambda(n+h_1) \cdots \lambda(n+h_k) = 0.$$ \end{itemize} \end{conjecture} Note that for $k=1$, the unweighted Chowla conjecture is equivalent to the prime number theorem, while the logarithmically averaged 1-point Chowla conjecture has a short elementary proof. No further cases of the unweighted Chowla conjecture are currently known, but the logarithmically averaged Chowla conjecture has been established for $k=2$ in \cite{tao} and for all odd values of $k$ in \cite{tt} (with a second proof given in \cite{tt-odd}). The logarithmically averaged Chowla conjecture is also known to be equivalent to the logarithmically averaged form of a conjecture of Sarnak \cite{sarnak}; see \cite{tao-higher}. See also \cite{mrt} for a version of Elliott's conjecture where one averages over the shifts $h_i$. One can also formulate an analogous version of Chowla's conjecture for the M\"obius function, for which very similar\footnote{If one generalises the Chowla conjecture by using affine forms $a_in+h_i$ instead of shifts $n+h_i$, then a simple sieving argument can be used to show the equivalence of such generalised Chowla conjectures for the Liouville function and their counterparts for the M\"obius function; we leave the details to the interested reader.} results are known. In \cite{tt}, we obtained the following special case of the logarithmically averaged Elliott conjecture (Conjecture \ref{chowla-conj} (ii)). We say that a $1$-bounded multiplicative function $f: \N \to \Disk$ \emph{weakly pretends} to be another $1$-bounded multiplicative function $g: \N \to \Disk$ if $$ \lim_{X \to \infty} \frac{1}{\log\log X} \D(f,g;X)^2 = 0$$ or equivalently $$ \sum_{p\leq X} \frac{1-\mathrm{Re}(f(p)\overline{g(p)})}{p} = o(\log \log X).$$ \begin{theorem}[Special case of logarithmically averaged Elliott]\label{lae} \cite[Corollary 1.6]{tt} Let $k \geq 1$, and let $g_1,\dots,g_k: \N \to \Disk$ be $1$-bounded multiplicative functions such that the product $g_1 \cdots g_k$ does not weakly pretend to be any Dirichlet character $n \mapsto \chi(n)$. Then for any integers $h_1,\dots,h_k$, one has $$ \lim_{X\to \infty}\E_{n \leq X}^{\log} g_1(n+h_1) \cdots g_k(n+h_k) = 0.$$ \end{theorem} In particular this establishes the logarithmically averaged Chowla conjecture for odd values of $k$. This result was also recently used by Frantzikinakis and Host \cite{fh-furstenberg} to control the Furstenberg measure-preserving systems associated to $1$-bounded multiplicative functions, and to establish a version of the logarithmic Sarnak conjecture where the M\"obius function $\mu(n)$ is replaced by a $1$-bounded multiplicative function $g(n)$ and the topological dynamical system involved is assumed to be uniquely ergodic. Theorem \ref{lae} was deduced from a more general structural statement about the correlation sequence $a \mapsto \lim_{X\to \infty}\E_{n \leq X}^{\log} g_1(n+ ah_1) \cdots g_k(n+ah_k)$ for $1$-bounded multiplicative functions $g_1,\dots,g_k$, where one now permits the product $g_1\cdots g_k$ to weakly pretend to be a Dirichlet character. Here one runs into the technical difficulty that the asymptotic limits $\lim_{X\to \infty}\E_{n \leq X}^{\log}$ are not known \emph{a priori} to exist. To get around this difficulty, the device of \emph{generalised limit functionals}\footnote{Alternatively, one could employ ultrafilter limits, or pass to subsequences in which all limits of interest exist. The latter approach is for instance the one adopted in \cite{frantz-2}, \cite{fh-sarnak}, \cite{fh-furstenberg}.} was employed. By a generalised limit functional we mean a bounded linear functional $\lim^*_{X \to \infty}: \ell^\infty(\N) \to \C$ which agrees with the ordinary limit functional $\lim_{X \to \infty}$ on convergent sequences, maps non-negative sequences to non-negative numbers, and which obeys the bound $$ |\lim^*_{X \to \infty} f(X)| \leq \limsup_{X \to \infty} |f(n)|$$ for all bounded sequences $f$. As is well known, the existence of such generalised limits follows from the Hahn-Banach theorem. With these notations, we proved in \cite{tt} the following. \begin{theorem}[Structure of logarithmically averaged correlation sequences]\label{Str}\cite[Theorem 1.1]{tt} Let $k \geq 1$, and let $h_1,\dots,h_k$ be integers and $g_1,\dots, g_k:\N\to \Disk$ be $1$-bounded multiplicative functions. Let $\lim^*_{X \to \infty}$ be a generalised limit functional. Let $f: \Z \to \Disk$ denote the function \begin{align}\label{eq18} f(a) \coloneqq \lim^{*}_{X\to \infty}\E_{n\leq X}^{\log} g_1(n+ah_1) \cdots g_k(n+ah_k). \end{align} \begin{itemize} \item[(i)] If the product $g_1 \cdots g_k$ does not weakly pretend to be a Dirichlet character, then $f$ is identically zero. \item[(ii)] If instead the product $g_1 \cdots g_k$ weakly pretends to be a Dirichlet character $\chi$, then $f$ is the uniform limit of periodic functions $F_i$, each of which is $\chi$-\textnormal{isotypic} in the sense that $F_i(ab) = F_i(a) \chi(b)$ whenever $a$ is an integer and $b$ is an integer coprime to the periods of $F_i$ and $\chi$. \end{itemize} \end{theorem} Among other things, Theorem \ref{Str} yields Theorem \ref{lae} as a direct corollary. Theorem \ref{lae} in turn can be used to establish various results about the distribution of consecutive values of $1$-bounded multiplicative functions; to give just one example, in \cite[Corollary 7.2]{tt} it was used to show that every sign pattern in $\{-1,+1\}^3$ occurred with logarithmic density $1/8$ amongst the Liouville sign patterns $(\lambda(n), \lambda(n+1), \lambda(n+2))$. \subsection{From logarithmic averages to almost all ordinary averages} It would be desirable if many of the above results for logarithmically averaged correlations such as $\E_{n \leq X}^{\log} g_1(n+h_1) \cdots g_k(n+h_k)$ could be extended to their unweighted counterparts such as $\E_{n \leq X} g_1(n+h_1) \cdots g_k(n+h_k)$. However, such extensions cannot be automatic, since for instance the logarithmic averages $\E_{n \leq X}^{\log} n^{it}$ converge to $0$ for $t \neq 0$, but the unweighted averages $\E_{n \leq X} n^{it}$ diverge. Similarly, the statement $\E_{n \leq X}^{\log} \lambda(n)=o(1)$ has a short and simple elementary proof\footnote{One can for example prove this by writing $\E_{n \leq X}^{\log} \lambda(n)=-\mathbb{E}_{p\leq y}\E_{n \leq X}^{\log} \lambda(n)p1_{p\mid n}+o_{y\to \infty}(1)$, and then using the Turán-Kubilius inequality to get rid of the $p1_{p\mid n}$ factor; we leave the details to the interested reader.}, whereas the unweighted analogue $\E_{n \leq X} \lambda(n)=o(1)$ is equivalent to the prime number theorem and its proofs are more involved. Moreover, one can show\footnote{More generally, one can use partial summation to show that, for any bounded real-valued sequence $a:\mathbb{N}\to \mathbb{R}$, if $\lim_{X\to \infty}\mathbb{E}_{n\leq X}^{\log}a(n)=\alpha$, then there exists an increasing sequence $X_i$ such that $\lim_{i\to \infty}\mathbb{E}_{n\leq X_i}a(n)=\alpha$. In particular, if the logarithmic Elliott conjecture holds, then the ordinary Elliott conjecture also holds in the case of real-valued functions along some subsequence of scales (which may depend on the functions involved).} that if, for example, the correlation limit $\lim_{X\to \infty} \mathbb{E}_{n\leq X}\lambda(n)\lambda(n+1)$ exists, then it has to be equal to $0$, which means that proving the mere existence of the limit captures the difficulty in the two-point unweighted Chowla conjecture. Nevertheless, there are some partial results of this type in which control on logarithmic averages can be converted to control on unweighted averages for a subsequence of scales $X$. For instance, in \cite{gkl} it is shown using ergodic theory techniques that if the logarithmically averaged Chowla conjecture holds for all $k$, then there exists an increasing sequence of scales $X_i$ such that the Chowla conjecture for all $k$ holds for $X$ restricted to these scales. This was refined in a blog post \cite{tao-blog} of the first author, where it was shown by an application of the second moment method that if the logarithmically averaged Chowla conjecture held for some even order $2k$, then the Chowla conjecture for order $k$ would hold for all scales $X$ outside of an exceptional set ${\mathcal X}\subset \mathbb{N}$ of logarithmic density zero, by which we mean that $$ \lim_{X\to \infty}\E^{\log}_{n \leq X} 1_{\mathcal X}(n) = 0.$$ Unfortunately, as the only even number for which the logarithmically averaged Chowla conjecture is currently known to hold is $k=2$, this only recovers (for almost all scales) the $k=1$ case of the unweighted Chowla conjecture, which was already known from the prime number theorem. At present, the restriction to logarithmic averaging in many of the above results is needed largely because it supplies (via the ``entropy decrement argument'') a certain approximate dilation invariance, which roughly speaking asserts the approximate identity $$ g_1(p)\cdots g_k(p)\E_{n \leq X}^{\log} g_1(n+h_1) \cdots g_k(n+h_k) \approx \E_{n \leq X}^{\log} g_1(n+ph_1) \cdots g_k(n+ph_k)$$ for ``most'' primes $p$, and for extremely large values of $X$; see for instance \cite[Theorem 3.2]{fh-furstenberg} for a precise form of this statement, with a proof essentially provided in \cite[\S 3]{tt}. However, an inspection of the entropy decrement argument reveals that it also provides an analogous identity for unweighted averages, namely that \begin{equation}\label{nxp} g_1(p)\cdots g_k(p) \E_{n \leq X} g_1(n+h_1) \cdots g_k(n+h_k) \approx \E_{n \leq X/p} g_1(n+ph_1) \cdots g_k(n+ph_k) \end{equation} for ``most'' primes $p$, and ``most'' extremely large values of $X$; see Proposition \ref{dollop} for a precise statement. By using this form of the entropy decrement argument, we are able to obtain the following analogue of Theorem \ref{Str} for unweighted averages, which is the main technical result of our paper and is proven in Section \ref{main-sec}. \begin{theorem}[Structure of unweighted correlation sequences]\label{main} Let $k \geq 1$, and let $h_1,\dots,h_k$ be integers and $g_1,\dots, g_k:\N\to \Disk$ be $1$-bounded multiplicative functions. Let $\lim^*_{X \to \infty}$ be a generalised limit functional. For each real number $d>0$, let $f_d: \Z \to \Disk$ denote the function \begin{equation}\label{fda-def} f_d(a) \coloneqq \lim^*_{X \to \infty} \E_{n \leq X/d} g_1(n+ah_1) \cdots g_k(n+ah_k). \end{equation} \begin{itemize} \item[(i)] If the product $g_1 \cdots g_k$ does not weakly pretend to be any twisted Dirichlet character $n \mapsto \chi(n)n^{it}$, then $$\lim_{X\to \infty} \E^{\log\log}_{d \leq X} |f_d(a)| = 0$$ for all integers $a$. \item[(ii)] If instead the product $g_1 \cdots g_k$ weakly pretends to be a twisted Dirichlet character $n \mapsto \chi(n)n^{it}$, then there exists a function $f: \Z \to \Disk$ such that \begin{equation}\label{hd} \lim_{X\to \infty}\E^{\log\log}_{d \leq X} |f_d(a) - f(a) d^{-it}| = 0 \end{equation} for all integers $a$. Furthermore, $f$ is the uniform limit of $\chi$-isotypic\footnote{That is, we have $F_i(ab)=F_i(a)\chi(b)$ for any integers $a$ and $b$ with $b$ coprime to the periods of $F_i$ and $\chi$.} periodic functions $F_i$. \end{itemize} \end{theorem} We have defined $f_d$ for all real numbers $d>0$ for technical reasons, but we will primarily be interested in the behaviour of $f_d$ for natural numbers $d$; for instance, the averages $\lim_{X\to \infty}\E^{\log\log}_{d \leq X}$ appearing in the above theorem are restricted to this case. Roughly speaking, the logarithmic correlation sequence $f(a)$ appearing in Theorem \ref{Str} is analogous to the average $\lim_{X\to \infty}\E^{\log\log}_{d \leq X} f_d(a)$ of the sequences appearing here (ignoring for this discussion the question of whether the limits exist). These averages vanish when $t \neq 0$ in Theorem \ref{main}, and one basically recovers a form of Theorem \ref{Str}; but, as the simple example of averaging the single $1$-bounded multiplicative function $n \mapsto n^{it}$ already shows, in the $t \neq 0$ case it is possible for the $f_d(a)$ to be non-zero while the logarithmically averaged counterpart $f(a)$ vanishes. By combining Theorem \ref{main} with a simple application of the Hardy--Littlewood maximal inequality, we can obtain several new cases of the unweighted Elliott and Chowla conjectures at almost all scales, as follows. \begin{corollary}[Some cases of the unweighted Elliott conjecture at almost all scales]\label{elliott-1} Let $k \geq 1$, and let $g_1,\dots, g_k:\N\to \Disk$ be $1$-bounded multiplicative functions. Suppose that the product $g_1 \cdots g_k$ does not weakly pretend to be any twisted Dirichlet character $n \mapsto \chi(n) n^{it}$. \begin{itemize} \item[(i)] For any $h_1,\dots,h_k \in \Z$ and $\eps>0$, one has $$ |\E_{n \leq X} g_1(n+h_1) \cdots g_k(n+h_k)| \leq \eps $$ for all natural numbers $X$ outside of a set $\mathcal{X}_{\varepsilon}$ of logarithmic Banach density zero, in the sense that \begin{align}\label{eq6} \lim_{\omega \to \infty} \sup_{X \geq\omega} \E^{\log}_{X/\omega \leq n \leq X} 1_{\mathcal{X}_{\varepsilon}}(n) = 0. \end{align} \item[(ii)] There is a set ${\mathcal X}_0$ of logarithmic density zero, such that $$ \lim_{X \to \infty; X \not \in {\mathcal X}_0} \E_{n \leq X} g_1(n+h_1) \cdots g_k(n+h_k) = 0$$ for all $h_1,\dots,h_k \in \Z$. \end{itemize} \end{corollary} \begin{remark} We note that Corollary \ref{elliott-1} can be generalised to the case of dilated correlations $$ \E_{n \leq X} g_1(q_1n+h_1) \cdots g_k(q_kn+h_k), $$ where $q_1,\ldots, q_k\in \mathbb{N}$. To see this, one applies exactly the same trick related to Dirichlet character expansions as in \cite[Appendix A]{tt}. Similarly, Corollary \ref{elliott-2} below generalises to the dilated case. We leave the details to the interested reader. \end{remark} \begin{remark} We see by partial summation that if $f:\mathbb{N}\to \mathbb{C}$ is any bounded function such that for every $\varepsilon>0$ we have $\displaystyle |\lim_{X\to \infty; X\not \in \mathcal{X}_{\varepsilon}}\mathbb{E}_{n\leq X}f(n)|\leq \varepsilon$ for some set $\mathcal{X}_{\varepsilon}\subset \mathbb{N}$ of logarithmic Banach density $0$, then we also have the logarithmic correlation result $\displaystyle\limsup_{X\to \infty} |\mathbb{E}_{X/\omega(X)\leq n\leq X}^{\log}f(n)|\ll \varepsilon$ for any function $1\leq \omega(X)\leq X$ tending to infinity. Thus Corollary \ref{elliott-1} is a strengthening of our earlier result \cite[Corollary 1.6]{tt} on logarithmic correlation sequences. Similarly, Corollary \ref{elliott-2} below is a strengthening of \cite[Corollary 1.5]{tao}. \end{remark} \begin{remark} The logarithmic density (or logarithmic Banach density) appearing in Corollaries \ref{elliott-1} and \ref{elliott-2} is the right density to consider in this problem. Namely, if one could show that the set $\mathcal{X}_0$ has \emph{asymptotic} density $0$, then $[1,\infty)\setminus \mathcal{X}_0$ would intersect every interval $[x,(1+\varepsilon)x]$ for all large $x$, which would easily imply (together with \eqref{eq13} below) that the unweighted correlation converges to zero without any exceptional scales. \end{remark} \begin{remark} The twisted Dirichlet characters $\chi(n) n^{it}$ appear both in Conjecture \ref{elliott-conj} and in Theorems \ref{Str}, \ref{main}. However, there is an interesting distinction as to how they appear; in Conjecture \ref{elliott-conj}, $t$ is allowed to be quite large (as large as $X$) and $\chi(n) n^{it}$ is associated to just a single multiplicative function $g_j$, while in Theorems \ref{Str}, \ref{main}, the quantity $t$ is independent of $X$ and is now associated to the product $g_1 \dots g_k$. The dependence of $t$ on $X$ in Conjecture \ref{elliott-conj}(i) is necessary\footnote{In the case of the \emph{logarithmically averaged} Conjecture \ref{elliott-conj}(ii), in contrast, \eqref{suptx} might not be a necessary assumption, since the sequence of bad scales constructed in \cite[Theorem B.1]{mrt} is sparse and thus does not influence the behaviour of logarithmic averages. We thank the referee for this remark.}, as is shown in \cite{mrt}; roughly speaking, the individual $g_j$ can oscillate like $n^{it_j}$ for various large $t_j$ in such a fashion that these oscillations largely cancel and produce non-trivial correlations in the product $g_1(n+h_1) \dots g_k(n+h_k)$. Meanwhile, Theorem \ref{main} asserts in some sense that the shifted product $g_1(n+h_1) \dots g_k(n+h_k)$ oscillates ``similarly to'' the unshifted product $g_1(n) \dots g_k(n)$, so in particular if the latter began oscillating like $n^{it}$ for increasingly large values of $t$ then the former product should exhibit substantial cancellation. \end{remark} The proof of Corollary \ref{elliott-1} is found in Section \ref{corollaries-sec}. So far, all of our results have concerned correlations where the product of the multiplicative functions involved is non-pretentious. In the case of two-point correlations, however, we can prove Corollary \ref{elliott-1} under the mere assumption that one of the multiplicative functions involved is non-pretentious, thus upgrading the logarithmic two-point Elliott conjecture in \cite{tao} to an unweighted version at almost all scales. \begin{corollary}[The binary unweighted Elliott conjecture at almost all scales]\label{elliott-2} Let $g_1, g_2:\N\to \Disk$ be $1$-bounded multiplicative functions, such that there exists $j\in \{1,2\}$ for which \eqref{suptx} holds as $X \to \infty$ for every Dirichlet character $\chi$. \begin{itemize} \item[(i)] For any distinct $h_1,h_2 \in \Z$ and $\eps > 0$, one has $$ |\E_{n \leq X} g_1(n+h_1) g_2(n+h_2)| \leq \eps $$ for all natural numbers $X$ outside of a set $\mathcal{X}_{\varepsilon}$ of logarithmic Banach density zero (in the sense of \eqref{eq6}). \item[(ii)] There is a set ${\mathcal X}_0$ of logarithmic density zero such that $$ \lim_{X \to \infty; X \not \in {\mathcal X}_0} \E_{n \leq X} g_1(n+h_1) g_2(n+h_2) = 0$$ for all distinct $h_1,h_2 \in \Z$. \end{itemize} \end{corollary} When specialised to the case of the Liouville function, the previous corollaries produce the following almost-all result. \begin{corollary}[Some cases of the unweighted Chowla conjecture at almost all scales]\label{chow} There is an exceptional set ${\mathcal X}_0$ of logarithmic density zero, such that $$ \lim_{X \to \infty; X \not \in {\mathcal X}_0} \E_{n \leq X} \lambda(n+h_1) \cdots \lambda(n+h_k) = 0$$ for all natural numbers $k$ that are either odd or equal to $2$, and for any distinct integers $h_1,\dots,h_k$. The same result holds if one replaces one or more of the copies of the Liouville function $\lambda$ with the M\"obius function $\mu$. \end{corollary} We establish these results in Section \ref{corollaries-sec}. One can use these corollaries to extend some previous results involving the logarithmic density of sign patterns to now cover unweighted densities of sign patterns at almost all scales. For instance, by inserting Corollary \ref{chow} into the proof of \cite[Corollary 1.10(i)]{tt}, one obtains the following. \begin{corollary}[Liouville sign patterns of length three] There is an exceptional set ${\mathcal X}_0$ of logarithmic density zero, such that $$ \lim_{X \to \infty; X \not \in {\mathcal X}_0} \E_{n \leq X} 1_{(\lambda(n),\lambda(n+1),\lambda(n+2)) = (\epsilon_0, \epsilon_1, \epsilon_2)} = \frac{1}{8}$$ for all sign patterns $(\epsilon_0, \epsilon_1, \epsilon_2) \in \{-1,1\}^3$. \end{corollary} Similarly several other results in \cite{tt} and in \cite{tera-binary} can be generalised. For example, the result \cite[Theorem 1.16]{tera-binary} on the largest prime factors of consecutive integers can be upgraded to the following form. \begin{corollary}[The largest prime factors of consecutive integers at almost all scales] \label{cor1} Let $P^{+}(n)$ be the largest prime factor of $n$ with $P^{+}(1):=1$. Then there is an exceptional set $\mathcal{X}_0$ of logarithmic density $0$, such that \begin{align}\label{eq10} \lim_{X\to \infty; X\not \in \mathcal{X}_0} \mathbb{E}_{n\leq X}1_{P^{+}(n)<P^{+}(n+1)}=\frac{1}{2}. \end{align} \end{corollary} The same equality with ordinary limit in place of the almost-all limit is an old conjecture formulated in the correspondence of Erd\H{o}s and Turán \cite[pp. 100--101]{sos}, \cite{erdos}. We remark on the proof of Corollary \ref{cor1} in Remark \ref{rem1}. In \cite[Theorem 1.6]{tera-binary} it was proved that \eqref{eq10} holds for the logarithmic average $\mathbb{E}_{n\leq X}^{\log}$ (without any exceptional scales). It would of course be desirable if we could upgrade ``almost all scales'' to ``all scales'' in the above results. We do not know how to do so in general, however there is one exceptional (though conjecturally non-existent) case in which this is possible, namely if there are unusually few sign patterns in the multiplicative functions of interest. We illustrate this principle with the following example. \begin{theorem}[Few sign patterns implies binary Chowla conjecture]\label{sign-thm} Suppose that for every $\eps>0$, there exist arbitrarily large natural numbers $K$ such that the set $\{ (\lambda(n+1), \dots, \lambda(n+K)): n \in \N \} \subset \{-1,+1\}^K$ of sign patterns of length $K$ has cardinality less than $\exp( \eps \frac{K}{\log K} )$. Then, for any natural number $h$, one has \begin{align*} \lim_{X\to \infty}\E_{n \leq X} \lambda(n) \lambda(n+h) = 0. \end{align*} \end{theorem} \begin{remark} The best known lower bounds for the number $s(K)$ of sign patterns of length $K$ for the Liouville function are very far from $\exp(\varepsilon \frac{K}{\log K})$. It was shown by Matom\"aki, Radziwi\l{}\l{} and Tao \cite{mrt-sign} that $s(K)\geq K+5$, and Frantzikinakis and Host showed in \cite{fh-sarnak} that $s(K)/K\to \infty$ as $K\to \infty$, but the rate of growth is inexplicit in that result. This was very recently improved to $s(K)\gg K^2$ by McNamara \cite{mcnamara}. If one assumes the Chowla conjecture (in either the unweighted or logarithmically averaged forms), it is not difficult to conclude that in fact $s(K) =2^K$ for all $K$. \end{remark} We prove this result in Section \ref{sign-sec}. Roughly speaking, the reason for this improvement is that the entropy decrement argument that is crucially used in the previous arguments becomes significantly stronger under the hypothesis of few sign patterns. A similar result holds for the odd order cases of the Chowla conjecture if one assumes the sign pattern control for \emph{all} large $K$ (rather than for a sequence of arbitrarily large $K$) by adapting the arguments in \cite{tt-odd}, but we do not do so here. It is also possible to strengthen this theorem in a number of further ways (for instance, restricting attention to sign patterns that occur with positive upper density, or to extend to other $1$-bounded multiplicative functions than the Liouville function), but we again do not do so so here. One should view Theorem \ref{sign-thm} as stating that if there is ''too much structure'' in the Liouville sequence (in the sense that it has a small number of sign patterns), then the binary Chowla conjecture holds. This is somewhat reminiscent of various statements in analytic number theory that rely on the assumption of a Siegel zero; for example, Heath-Brown \cite{hb} proved that if there are Siegel zeros, then the twin prime conjecture (which is connected to the two-point Chowla conjecture) holds. Nevertheless, the proof of Theorem \ref{sign-thm} does not resemble that in \cite{hb}. \subsection{Isotopy formulae} The conclusion of Theorem \ref{main}(ii) asserts, roughly speaking, that $f_d(a)$ ``behaves like'' a multiple of $\chi(a) d^{-it}$ in a certain asymptotic sense. The following corollary of that theorem makes this intuition a bit more precise. \begin{theorem}[Isotopy formulae] \label{isotopy} Let $k\geq 1$, let $h_1,\ldots, h_k$ be integers and $g_1,\ldots, g_k:\mathbb{N}\to \mathbb{D}$ be $1$-bounded multiplicative functions. Suppose that the product $g_1\cdots g_k$ weakly pretends to be a twisted Dirichlet character $n\mapsto \chi(n)n^{it}$. \begin{itemize} \item[(i)] (Archimedean isotopy) There exists an exceptional set $\mathcal{X}_0$ of logarithmic density zero, such that \begin{align*} \lim_{X\to \infty; X\not \in \mathcal{X}_0}\left(\mathbb{E}_{n\leq X}g_1(n+h_1)\cdots g_k(n+h_k)-q^{it}\mathbb{E}_{n\leq X/q}g_1(n+h_1)\cdots g_k(n+h_k)\right)=0 \end{align*} for all rational numbers $q>0$. \item[(ii)] (Non-archimedean isotopy) There exists an exceptional set $\mathcal{X}_0$ of logarithmic density zero, such that \begin{align*} \lim_{X\to \infty; X\not \in \mathcal{X}_0}\left(\mathbb{E}_{n\leq X}g_1(n-ah_1)\cdots g_k(n-ah_k)-\chi(-1)\mathbb{E}_{n\leq X}g_1(n+ah_1)\cdots g_k(n+ah_k)\right)=0 \end{align*} for all integers $a$. \end{itemize} \end{theorem} \begin{remark} This generalises \cite[Theorem 1.2(iii)]{tt}, which implies $f(-a)=\chi(-1)f(a)$ with $f(a)$ a generalised limit of a logarithmic correlation defined in \eqref{eq18} (indeed, Theorem \ref{isotopy}(ii) implies by partial summation that $f(-a)=\chi(-1)f(a)$ in the notation of \eqref{eq18}). In \cite{tt}, we only considered logarithmically averaged correlations, and for such averages Theorem \ref{isotopy}(i) does not make sense, as logarithmic averages are automatically slowly varying. However, for unweighted averages Theorem \ref{isotopy}(i) gives nontrivial information about the behaviour of the correlation at nearby scales. \end{remark} We give the proof of Theorem \ref{isotopy} in Section \ref{isotopy-sec}. We show in that section that, perhaps surprisingly, the non-archimedean isotopy formula (Theorem \ref{isotopy}(ii)) allows us to evaluate the correlations of some multiplicative functions whose product \emph{does pretend} to be a Dirichlet character. Among other things, we use the isotopy formula to prove a version of the even order logarithmic Chowla conjectures where we twist one of the copies of the Liouville function by a carefully chosen Dirichlet character and the shifts of $\lambda$ are consecutive. \begin{corollary}[Even order correlations of a twisted Liouville function] \label{even-chowla} Let $k\geq 4$ be an even integer, and let $\chi$ be an odd Dirichlet character of period $k-1$ (there are $\frac{\varphi(k-1)}{2}$ such characters). Then there exists an exceptional set $\mathcal{X}_0$ of logarithmic density $0$, such that \begin{align}\label{eq1} \lim_{X\to \infty; X\not \in \mathcal{X}_0}\mathbb{E}_{n\leq X}\chi(n)\lambda(n)\lambda(n+a)\cdots \lambda(n+(k-1)a)=0 \end{align} for all integers $a$. \end{corollary} By partial summation, we see from \eqref{eq1} that we have the logarithmic correlation result \begin{align*} \lim_{X \to \infty}\mathbb{E}_{n \leq X}^{\log} \chi(n)\lambda(n)\lambda(n+1)\cdots \lambda(n+k-1)=0, \end{align*} which is already new. We stated Corollary \ref{even-chowla} only for even $k$, but of course the result also holds for odd $k$ by Corollary \ref{elliott-1}. The assumption that $\chi$ is an odd character is crucial above, as will be seen in Section \ref{isotopy-sec}; the isotopy formulae are not able to say anything about the untwisted even order correlations of the Liouville function. We likewise show in Section \ref{isotopy-sec} that the archimedean isotopy formula (Theorem \ref{isotopy}(i)) gives a rather satisfactory description of the \emph{limit points} of the correlations \begin{align}\label{eq19} \mathbb{E}_{n\leq X}g_1(n+h_1)\cdots g_k(n+h_k), \end{align} where the product $g_1\cdots g_k$ weakly pretends to be a twisted Dirichlet character $n\mapsto \chi(n)n^{it}$ with $t\neq 0$. Indeed, our Theorem \ref{limits} shows that once one continuously excludes the scales at which the correlation \eqref{eq19} is close to zero, the argument of the quantity in \eqref{eq19} is in a sense uniformly distributed on the unit circle. This uniform distribution is indeed expected when $g_j$ are pretentious; for example, one has $\mathbb{E}_{n\leq X}n^{it}=\frac{X^{it}}{1+it}+o(1)$, which uniformly distributes on the circle of radius $1/|1+it|$ with respect to logarithmic density. \subsection{Proof ideas} We now briefly describe (in informal terms) the proof strategy for Theorem \ref{main}, which follows the ideas in \cite{tt}, but now contains some ``Archimedean'' arguments (relating to the Archimedean characters $n \mapsto n^{it}$) in addition to the ``non-Archimedean'' arguments in \cite{tt} (that related to the Dirichlet characters $n \mapsto \chi(n)$). The new features compared to \cite{tt} include extensive use of the fact that the correlations $f_d(a)$ are ``slowly varying'' in terms of $d$ (this is made precise in formula \eqref{fd1d2}), and the use of this to derive ``approximate quasimorphism properties'' for certain quantities related to these correlations (these are detailed below). We then prove that the approximate quasimorphisms are very close to actual quasimorphisms (which in our case are Dirichlet characters or Archimedean characters), which eventually leads to the desired conclusions. As already noted, one key ingredient is (a rigorous form of) the approximate identity \eqref{nxp} that arises from the entropy decrement argument. In terms of the correlation functions $f_d(a)$, this identity takes the (heuristic) form $$ f_{dp}(a) G(p) \approx f_{d}(ap)$$ for any integers $a,d$ and ``most'' $p$, where $G \coloneqq g_1 \cdots g_k$; see Proposition \ref{dollop} for a precise statement. Compared to \cite{tt}, the main new difficulty is the dependence of $f_d$ on the $d$ parameter. Assuming for simplicity that $G$ has modulus $1$ (which is the most difficult case), we thus have $$ f_{dp}(a) \approx f_{d}(ap) \overline{G(p)}$$ for any integers $a,d$ and ``most'' $p$. Iterating this leads to \begin{equation}\label{p1p2} f_{p_1 p_2}(a) \approx f_1(ap_1p_2) \overline{G(p_1)} \overline{G(p_2)} \end{equation} for ``most'' primes $p_1,p_2$ (more precisely, the difference between the two sides of the equation is $o(1)$ when suitably averaged over $p_1,p_2$; see Corollary \ref{mango}). On the other hand, results from ergodic theory (such as \cite{leibman}, \cite{le}) give control on the function $f_1(a)$, describing it (up to negligible errors) as a nilsequence, which can then be decomposed further into a periodic piece $f_{1,0}$ and an ``irrational'' component. The irrational component was already shown in \cite{tt} to give a negligible contribution to the equation \eqref{p1p2} after performing some averaging in $p_1,p_2$, thanks to certain bilinear estimates for nilsequences. As such, one can effectively replace $f_1$ here by the periodic component $f_{1,0}$ (see \eqref{2m1-0} for a precise statement). We thus reach the relation \begin{align*} f_{p_1 p_2}(a) \approx f_{1,0}(ap_1p_2) \overline{G(p_1)} \overline{G(p_2)} \end{align*} for ``most'' $p_1,p_2$. Let $q$ be the period of $f_{1,0}$. If we pick two large primes $p_1\equiv c\pmod {q}$ and $p_1'\equiv bc\pmod{q}$ for arbitrary $b,c\in (\mathbb{Z}/q\mathbb{Z})^{\times}$ with $p_1\approx p_1'$ (using the the prime number theorem), we get \begin{align*} f_{1,0}(acp_2) \overline{G(p_1)}\approx f_{1,0}(abcp_2) \overline{G(p_1')}, \end{align*} for ``most'' $p_1,p_1',p_2$, since the averages $f_d(a)$ are slowly varying as a function of $d$ (see equation \eqref{fd1d2} for the precise meaning of this). Choosing $p_2\equiv 1 \pmod{q}$, we see that the quotient $f_{1,0}(ac)/f_{1,0}(abc)$ is independent of $a$ (since $p_1,p_1'$ were independent of $a$). Substituting then $a=a_1$ and $a=a_2$ to the quotient, we get the approximate identity \begin{align}\label{eq7} f_{1,0}(a_1c)f_{1,0}(a_2bc)\approx f_{1,0}(a_1bc)f_{1,0}(a_2c); \end{align} see Proposition \ref{prop1} for a precise version of this, where we need to average over $c$ to make the argument rigorous. We may assume that $f(a_0)\neq 0$ for some $a_0$, as otherwise there is nothing to prove, and this leads to $f_{1,0}(a_0)\neq 0$. Taking $a_1\equiv a_0c^{-1} \pmod{q}$, $a_2\equiv a_0\pmod{q}$ in \eqref{eq7}, we are led to \begin{align*} f_{1,0}(a_0)f_{1,0}(a_0bc)\approx f_{1,0}(a_0b)f_{1,0}(a_0c) \end{align*} Thus, the function $\psi(x)=f_{1,0}(a_0x)/f_{1,0}(a_0)$ satisfies the approximate quasimorphism equation $$ \psi(b_1 b_2) \approx \psi(b_1) \psi(b_2)$$ for $b_1,b_2 \in (\Z/q\Z)^\times$ ranging in the invertible residue classes in $\Z/q\Z$ and some unknown function $\psi: (\Z/q\Z)^\times \to \C$ (to make the above deductions rigorous, we need to take as $\psi(x)$ an averaged version of $x\mapsto f_{1,0}(a_0x)/f_{1,0}(a_0)$). Moreover, the function $\psi(x)$ takes values comparable to $1$. Of course, Dirichlet characters obey the quasimorphism equation exactly; and we can use standard ``cocycle straightening'' arguments to show conversely that any solution to the quasimorphism equation must be very close to a Dirichlet character $\chi$ (see Lemma \ref{dirs} for a precise statement). This will be used to show that $f_{1,0}$ and $f_d$ are essentially $\chi$-isotypic. Once this isotopy property is established, one can then return to \eqref{p1p2} and analyse the dependence of various components of \eqref{p1p2} on the Archimedean magnitudes of $p_1,p_2$ rather than their residues mod $q$. One can eventually transform this equation again to the quasimorphism equation, but this time on the multiplicative group $\R^+$ rather than $(\Z/q\Z)^\times$ (also, the functions $\psi$ will be ``log-Lipschitz'' in a certain sense). Now it is the Archimedean characters $n \mapsto n^{it}$ that are the model solutions of this equation, and we will again be able to show that all other solutions to this equation are close to an Archimedean character (see Lemma \ref{dira} for a precise statement). Once one has extracted both the Dirichlet character $\chi$ and the Archimedean character $n \mapsto n^{it}$ in this fashion, the rest of Theorem \ref{main} can be established by some routine calculations. \subsection{Acknowledgments} We thank the anonymous referees for careful reading of the paper and for useful comments. TT was supported by a Simons Investigator grant, the James and Carol Collins Chair, the Mathematical Analysis \& Application Research Fund Endowment, and by NSF grant DMS-1266164. JT was supported by UTUGS Graduate School and project number 293876 of the Academy of Finland. He thanks UCLA for excellent working conditions during a visit there in April 2018, during which a large proportion of this work was completed. \subsection{Notation}\label{subsec: not} We use the usual asymptotic notation $X \ll Y$, $Y \gg X$, or $X = O(Y)$ to denote the bound $|X| \leq CY$ for some constant $C$. If $C$ needs to depend on parameters, we will denote this by subscripts, thus for instance $X \ll_k Y$ denotes the estimate $|X| \leq C_k Y$ for some $C_k$ depending on $k$. We also write $o_{n \to \infty}(Y)$ for a quantity bounded in magnitude by $c(n)Y$ for some $c(n)$ that goes to zero as $n \to \infty$ (holding all other parameters fixed). For any set $\mathcal{X}\subset \mathbb{N}$ with infinite complement, we define the limit operator $\lim_{X\to \infty; X\not \in \mathcal{X}} f(X)$ as $\lim_{n\to \infty} f(x_n)$, where $x_1,x_2,\ldots$ are the elements of the complement $\mathbb{N}\setminus \mathcal{X}$ in strictly increasing order. We use $a\ (q)$ to denote the residue class of $a$ modulo $q$. If $E$ is a set, we write $1_E$ for its indicator function, thus $1_E(n)=1$ when $n \in E$ and $1_E(n)=0$ otherwise. We use the following standard multiplicative functions throughout the paper: \begin{itemize} \item The \emph{Liouville function} $\lambda$, which is the $1$-bounded completely multiplicative function with $\lambda(p)=-1$ for all primes $p$; \item The \emph{M\"obius function} $\mu$, which is equal to $\lambda$ at square-free numbers and $0$ elsewhere. \item \emph{Dirichlet characters} $\chi$, which are $1$-bounded completely multiplicative functions of some period $q$, with $\chi(n)$ non-zero precisely when $n$ is coprime to $q$; and \item \emph{Archimedean characters} $n \mapsto n^{it}$, where $t$ is a real number. \item \emph{twisted Dirichlet characters} $n \mapsto \chi(n)n^{it}$, which are the product of a Dirichlet character and an Archimedean character. \end{itemize} In the arguments that follow, asymptotic averages of various types feature frequently, so we introduce some abbreviations for them. \begin{definition}[Asymptotic averaging notation] If $f: \N \to \C$ is a function, we define the asymptotic average $$ \E_{n \in \N} f(n) \coloneqq \lim_{X \to \infty} \E_{n \leq X} f(n) $$ provided that the limit exists. We adopt the convention that assertions such as $\E_{n \in \N} f(n) = \alpha$ are automatically false if the limit involved does not exist. Similarly define $\E_{n \in \N}^{\log} f(n)$ and $\E_{n \in \N}^{\log\log} f(n)$. If $f: {\P} \to \C$ is a function, we similarly define $$ \E_{p \in \P} f(p) \coloneqq \lim_{X \to \infty} \E_{p \leq X} f(p) $$ and $$ \E_{p \in \P}^{\log} f(p) \coloneqq \lim_{X \to \infty} \E_{p \leq X}^{\log} f(p).$$ Moreover, given a generalised limit functional $\lim^*_{X \to \infty}$, we define the corresponding asymptotic limits $\E^*_{n \in \N}$, $\E^{\log,*}_{n \in \N}$, $\E^{\log\log,*}_{n \in \N}$, $\E^*_{p \in {\P}}$, $\E^{\log,*}_{p \in {\P}}$ by replacing the ordinary limit functional by the generalised limit, thus for instance $$ \E^{\log,*}_{n \in \N} f(n) \coloneqq \lim^*_{X \to \infty} \E^{\log}_{n \leq X} f(n).$$ If an ordinary asymptotic limit such as $\E^{\log}_{n \in \N} f(n)$ exists, then $\E^{\log,*}_{n \in \N} f(n)$ will attain the same value; but the latter limit exists for all bounded sequences $f$, whereas the ordinary limit need not exist. In later parts of the paper we will also need an additional generalised limit $\lim^{**}_{X \to \infty}$, and one can then define generalised asymptotic averages such as $\E^{\log,**}_{n \in \N} f(n)$ accordingly. \end{definition} \begin{remark}\label{avg-com} If $f$ is a bounded sequence and $\alpha$ is a complex number, a standard summation by parts exercise shows that the statement $\E_{n \in \N} f(n) = \alpha$ implies $\E_{n \in \N}^{\log} f(n) = \alpha$, which in turn implies $\E_{n \in \N}^{\log\log} f(n) = \alpha$, and similarly $\E_{p \in \P} f(p) = \alpha$ implies $\E_{p \in {\P}}^{\log} f(p) = \alpha$; however, the converse implications can be highly non-trivial or even false. For instance, as mentioned earlier, it is not difficult to show that $\E_{n \in \N}^{\log} n^{it} = 0$ for any $t\neq 0$, but the limit $\E_{n \in \N} n^{it}$ does not exist. (On the other hand, from the prime number theorem and partial summation one has $\E_{p \in \P} p^{it} = 0$.) In the same spirit, if $A$ is the set of integers whose decimal expansion has leading digit $1$, then one easily computes ``Benford's law'' $\E_{n\in \N}^{\log}1_{A}(n)=(\log 2)/(\log 10)$, whereas $\E_{n\in \N} 1_{A}(n)$ fails to exist. \end{remark} \section{Proof of main theorem}\label{main-sec} In this section we establish Theorem \ref{main}. We first establish a version of the Furstenberg correspondence principle. \begin{proposition}[Furstenberg correspondence principle]\label{corr} Let the notation and hypotheses be as in Theorem \ref{main}. Then for any real number $d > 0$, there exist random functions $\mathbf{g}^{(d)}_1,\dots,\mathbf{g}^{(d)}_k: \Z \to \Disk$ and a random profinite integer\footnote{The \emph{profinite integers} $\hat \Z$ are the inverse limit of the cyclic groups $\Z/q\Z$, with the weakest topology that makes the reduction maps $n \mapsto n\ (q)$ continuous. This is a compact abelian group and therefore it has a well-defined probability Haar measure.} $\mathbf{n}^{(d)} \in \hat \Z$, all defined on a common probability space $\Omega^{(d)}$, such that $$ \E^{(d)} F( ((\mathbf{g}^{(d)}_i(h))_{1 \leq i \leq k, -N \leq h \leq N}, \mathbf{n}^{(d)}\ (q))) = \lim^{*}_{X \to \infty} \E_{n \leq X/d} F((g_i(n+h))_{1 \leq i \leq k, -N \leq h \leq N}, n\ (q))$$ for any natural numbers $N,q$ and any continuous function $F: \Disk^{k(2N+1)}\times \mathbb{Z}/q\mathbb{Z} \to \C$, where $\E^{(d)}$ denotes the expectation on the probability space $\Omega^{(d)}$. Furthermore, the random variables $\mathbf{g}^{(d)}_1,\dots,\mathbf{g}^{(d)}_k: \Z \to \Disk$ and $\mathbf{n}^{(d)} \in \hat \Z$ are a stationary process, by which we mean that for any natural number $N$, the joint distribution of $(\mathbf{g}^{(d)}_i(n+h))_{1 \leq i \leq k, -N \leq h \leq N}$ and $\mathbf{n}^{(d)}+n$ does not depend on $n$ as $n$ ranges over the integers. \end{proposition} \begin{proof} Up to some minor notational changes, this is essentially \cite[Proposition 3.1]{tt}, applied once for each value of $d$. The only difference is that the logarithmic averaging $\E^{\log}_{x_m/w_m \leq n \leq x_m}$ there has been replaced by the non-logarithmic averaging $\E_{n \leq X/d}$. However, an inspection of the arguments reveal that the proof of the proposition is essentially unaffected by this change. \end{proof} Let $G: \N \to \Disk$ denote the multiplicative function $G \coloneqq g_1 \cdots g_k$. We now adapt the entropy decrement arguments from \cite[\S 3]{tt} to establish the approximate relation \begin{equation}\label{fdap} f_d(ap) \approx f_{dp}(a) G(p) \end{equation} for integers $a$, real numbers $d>0$, and ``most'' primes $p$. Fix $a,d$, and let $p$ be a prime. From \eqref{fda-def} we have $$ f_{dp}(a) G(p) = \lim^{*}_{x \to \infty} \E_{n \leq x/dp} g_1(p) g_1(n+ah_1) \cdots g_k(p) g_k(n+ah_k). $$ From multiplicativity, we can write $g_j(p) g_j(n+ah_j)$ as $g_j(pn + aph_j)$ unless $n = -ah_j\ (p)$. The latter case contributes $O\left(\frac{1}{p}\right)$ to the above limit (where we allow implied constants to depend on $k$), thus $$ f_{dp}(a) G(p) = \lim^{*}_{x \to \infty} \E_{n \leq x/dp} g_1(pn+aph_1) \cdots g_k(pn+aph_k) + O\left( \frac{1}{p} \right).$$ If we now make $pn$ rather than $n$ the variable of summation, we conclude that $$ f_{dp}(a) G(p) = \lim^{*}_{x \to \infty} \E_{n \leq x/d} g_1(n+aph_1) \cdots g_k(n+aph_k) p 1_{p|n} + O\left( \frac{1}{p} \right).$$ Comparing this with \eqref{fda-def}, we conclude that $$ f_{dp}(a) G(p) - f_d(ap) = \lim^{*}_{x \to \infty} \E_{n \leq x/d} g_1(n+aph_1) \cdots g_k(n+aph_k) (p 1_{p|n}-1) + O\left( \frac{1}{p} \right)$$ and hence by Proposition \ref{corr} \begin{equation}\label{fdag} f_{dp}(a) G(p) - f_d(ap) = \E^{(d)} \mathbf{g}^{(d)}_1(aph_1) \cdots \mathbf{g}^{(d)}_k(aph_k) (p 1_{p|\mathbf{n}^{(d)}}-1) + O\left( \frac{1}{p} \right). \end{equation} On the other hand, by repeating the proof of \cite[Theorem 3.6]{tt} verbatim (see also \cite[Remark 3.7]{tt}), we have the following general estimate: \begin{proposition}[Entropy decrement argument]\label{eda} Let $\mathbf{g}_1,\dots,\mathbf{g}_k: \Z \to \Disk$ be random functions and $\mathbf{n} \in \hat \Z$ be a stationary process, let $a,h_1,\dots,h_k$ be integers, and let $0 < \eps < \frac{1}{2}$ be real. Then one has $$ \E_{2^m \leq p < 2^{m+1}} |\mathbb{E}\mathbf{g}_1(aph_1) \cdots \mathbf{g}_k(aph_k) (p 1_{p|\mathbf{n}}-1) | \leq \eps$$ for all natural numbers $m$ outside of an exceptional set ${\mathcal M}$ obeying the bound \begin{equation}\label{eps} \sum_{m \in {\mathcal M}} \frac{1}{m} \ll_{a, h_1,\dots,h_k} \eps^{-4} \log \frac{1}{\eps}. \end{equation} \end{proposition} Note that the bound \eqref{eps} is uniform in the random functions $\mathbf{g}_1,\dots,\mathbf{g}_k$ (although the set $\mathcal{M}$ may depend on these functions). Summing the result over different dyadic scales gives us the following version of \eqref{fdap}. \begin{proposition}[Approximate isotopy]\label{dollop} Let the notation and hypotheses be as in Theorem \ref{main}. Let $a$ be an integer, and let $\eps > 0$ be real. Then for sufficiently large $P$, we have $$ \sup_{d >0} \E^{\log}_{p \leq P} |f_{dp}(a) G(p) - f_d(ap)| \leq \eps$$ where the supremum is over positive reals. \end{proposition} A key technical point for our application is that while $P$ may depend on $a,\eps$, it can be taken to be uniform in $d$. \begin{proof} Let $a,\eps,P$ be as in the proposition, and let $d >0$. We may assume that $\varepsilon>0$ is small. By the prime number theorem, we have $$ \E^{\log}_{p \leq P} |f_{dp}(a) G(p) - f_d(ap)| \ll \E^{\log}_{m \leq (\log P)/(\log 2)} \E_{2^m \leq p < 2^{m+1}} |f_{dp}(a) G(p) - f_d(ap)|.$$ By \eqref{fdag} and Proposition \ref{eda}, we have $$ \E_{2^m \leq p < 2^{m+1}} |f_{dp}(a) G(p) - f_d(ap)| \leq \eps^2$$ for all $m$ outside of an exceptional set ${\mathcal M}_{a,\eps,d}$ obeying the bound $$ \sum_{m \in {\mathcal M}_{a,\eps,d}} \frac{1}{m} \ll_{a, h_1,\dots,h_k} \eps^{-8} \log \frac{1}{\eps}.$$ In the exceptional set ${\mathcal M}_{a,\eps,d}$, we use the trivial bound $$ \E_{2^m \leq p < 2^{m+1}} |f_{dp}(a) G(p) - f_d(ap)| \ll 1$$ to conclude that $$ \E^{\log}_{p \leq P} |f_{dp}(a) G(p) - f_d(ap)| \ll \eps^2 + O_{a,h_1,\dots,h_k}\left( \frac{\eps^{-8} \log \frac{1}{\eps}}{\log\log P} \right),$$ and the claim follows by choosing $P$ large in terms of $a,\varepsilon, h_1,\ldots, h_k$. \end{proof} As in \cite{tt}, we iterate this approximate formula to obtain \begin{corollary}\label{mango} For any integer $a$ one has $$ \limsup_{P_1 \to \infty} \limsup_{P_2 \to \infty} \E_{p_1 \leq P_1}^{\log} \E_{p_2 \leq P_2}^{\log} |f_{p_1 p_2}(a) G(p_1) G(p_2) - f_1(ap_1 p_2)| = 0$$ \end{corollary} \begin{proof} Let $a$ be an integer, let $\eps>0$ be real, let $P_1$ be sufficiently large depending on $a,\eps$, and let $P_2$ be sufficiently large depending on $a,\eps,P_1$. From Proposition \ref{dollop} one has $$ \E^{\log}_{p_1 \leq P_1} |f_{p_1 p_2}(a) G(p_1) - f_{p_2}(ap_1)| \ll \eps$$ for all primes $p_2$, and hence $$ \E_{p_1 \leq P_1}^{\log} \E_{p_2 \leq P_2}^{\log} |f_{p_1 p_2}(a) G(p_1) G(p_2) - f_{p_2}(ap_1) G(p_2)| \ll \eps.$$ On the other hand, from a second application of Proposition \ref{dollop} one has $$ \E^{\log}_{p_2 \leq P_2} |f_{p_2}(ap_1) G(p_2) - f_{1}(ap_1 p_2)| \ll \eps$$ for all $p_1 \leq P_1$, and hence $$ \E_{p_1 \leq P_1}^{\log} \E_{p_2 \leq P_2}^{\log} |f_{p_2}(ap_1) G(p_2) - f_1(ap_1 p_2)| \ll \eps.$$ From the triangle inequality we thus have $$ \E_{p_1 \leq P_1}^{\log} \E_{p_2 \leq P_2}^{\log} |f_{p_1 p_2}(a) G(p_1) G(p_2) - f_1(ap_1 p_2)| \ll \eps$$ under the stated hypotheses on $\eps,P_1,P_2$. Taking limit superior in $P_2$ and then in $P_1$, we conclude that $$ \limsup_{P_1 \to \infty} \limsup_{P_2 \to \infty} \E_{p_1 \leq P_1}^{\log} \E_{p_2 \leq P_2}^{\log} |f_{p_1 p_2}(a) G(p_1) G(p_2) - f_1(ap_1 p_2)| \ll \eps$$ for any $\eps>0$, and the claim follows. \end{proof} Next, we have the following structural description of $f_1$. \begin{proposition}\label{pe} Let $f_1$ be as in Theorem \ref{main}. For any $\eps>0$, one can write $$ f_{1} = f_{1,0} + g$$ where $f_{1,0} = f_{1,0}^{(\eps)}$ is periodic, and the error $g = g^{(\eps)}$ obeys the bilinear estimate \begin{equation}\label{unweighted} \E_{p_1 \leq x} \E_{p_2 \leq y} \alpha_{p_1} \beta_{p_2} g(a p_1 p_2) \ll \eps \end{equation} as well as the logarithmic counterpart $$ \E^{\log}_{p_1 \leq x} \E^{\log}_{p_2 \leq y} \alpha_{p_1} \beta_{p_2} g(a p_1 p_2) \ll \eps $$ whenever $a$ is a non-zero integer, $x$ is sufficiently large depending on $a,\eps$; $y$ is sufficiently large depending on $x,a,\eps$; and $\alpha_{p_1}, \beta_{p_2} = O(1)$ are bounded sequences. \end{proposition} \begin{proof} We freely use the notation from \cite[Sections 4--5]{tt}. By summation by parts it suffices to obtain a decomposition obeying \eqref{unweighted}. By repeating the proof of \cite[Corollary 4.6]{tt} verbatim\footnote{In \cite[Corollary 4.6]{tt}, $a$ was required to be a natural number rather than a non-zero integer, however one can easily adapt the arguments to the case of negative $a$ with only minor modifications (in particular, one has to modify the definition of $\mathbf{X}_m$ slightly to allow $l$ to be negative).}, we can write $$ f_1 = f_{1,1} + f_{1,2}$$ where $f_{1,1}$ is a nilsequence of some finite degree $D$, and $f_{1,2}$ obeys the asymptotic $$ \lim_{x \to \infty} \E_{p \leq x} |f_{1,2}(ap)| = 0$$ for any non-zero integer $a$. We can now neglect the $f_{1,2}$ term as it can be absorbed into the $g$ error. Next, applying \cite[Proposition 5.6]{tt}, we can decompose $$ f_{1,1} = f_{1,0} + \sum_{i=1}^D \sum_{j=1}^{J_i} c_{i,j} \chiup_{i,j}$$ for some periodic function $f_{1,0}$, some non-negative integers $J_1,\dots,J_D$, some irrational nilcharacters $\chiup_{i,j}$ of degree $i$, and some linear functionals $c_{i,j}$. Using \cite[Lemma 5.8]{tt} (noting that if $\chiup$ is an irrational nilcharacter, then so is $\chiup(a \cdot)$) we see that each of the terms $c_{i,j} \chiup_{i,j}$ can be absorbed into the error term $g$. The claim then follows from the triangle inequality. \end{proof} Finally, we record a simple log-Lipschitz estimate \begin{equation}\label{fd1d2} |f_{d_1}(a) - f_{d_2}(a)| \leq 2|\log d_1 - \log d_2| \end{equation} for any integer $a$ and any real $d_1,d_2 > 0$; this follows by using \eqref{fda-def} and the triangle inequality to estimate $|f_{d_1}(a)-f_{d_2}(a)|\leq \frac{2|d_1-d_2|}{\max\{d_1,d_2\}}$ and then the mean value theorem to $x\mapsto \log x$. We return to the proof of Theorem \ref{main}. If we have $$ \limsup_{X \to \infty} \E^{\log\log}_{d \leq X} |f_d(a)| = 0$$ for all $a$, then the claim follows by setting $f=0$, so we may assume without loss of generality that there exists an integer $a_0$ such that $$ \limsup_{X \to \infty} \E^{\log\log}_{d \leq X} |f_d(a_0)| > 0.$$ Thus, by the Hahn-Banach theorem, we may find a generalised limit $\lim^{**}_{X \to \infty}$ (which may or may not be equal to the previous generalised limit $\lim_{X \to \infty}^*$) such that $$ \lim^{**}_{X \to \infty} \E^{\log\log}_{d \leq X} |f_d(a_0)| > 0,$$ and thus using the generalised limit asymptotic notation associated to $\lim^{**}_{X \to \infty}$ (see Subsection \ref{subsec: not}), we have \begin{equation}\label{fda} \E^{\log\log,**}_{d \in \N} |f_d(a_0)| \gg 1. \end{equation} For future reference we record the following convenient lemma relating the averaging operator $\E^{\log\log,**}_{d \in \N}$ with $\E^{\log,**}_{p \in {\P}}$: \begin{lemma}[Comparing averages over integers and primes]\label{chug} Let $f: \N \to \C$ be a function which is bounded log-Lipschitz in the sense that there is a constant $C$ such that $|f(d)| \leq C$ and $|f(d)-f(d')| \leq C |\log d - \log d'|$ for all $d,d' \in \N$. Then for any natural number $a$, one has $$\limsup_{X\to \infty}|\E^{\log\log}_{d \leq X} f(d)-\E^{\log}_{p \leq X} f(ap)|=0,$$ so in particular $$ \E^{\log\log,**}_{d \in \N} f(d) = \E^{\log,**}_{p \in {\P}} f(ap).$$ \end{lemma} \begin{proof} We allow implied constants to depend on $C,a$. Let $\eps>0$, and assume $X$ is sufficiently large depending on $C,\eps$. Then from the prime number theorem and the bounded log-Lipschitz property we have \begin{align*} \E^{\log}_{p \leq X} f(ap) &= \frac{1}{\log\log X} \sum_{p \leq X} \frac{f(ap)}{p} + O(\eps) \\ &= \frac{1}{\log\log X} \sum_{d \leq X} \frac{1}{\eps d} \sum_{d \leq p \leq (1+\eps) d} \frac{f(ap)}{p} + O(\eps) \\ &= \frac{1}{\log\log X} \sum_{d \leq X} \frac{1}{\eps d} \sum_{d \leq p \leq (1+\eps) d} \frac{f(ad)}{d} + O(\eps) \\ &= \frac{1}{\log\log X} \sum_{d \leq X} \frac{f(ad)}{d \log(2+d)} + O(\eps). \end{align*} Again by the bounded log-Lipschitz property, we have \begin{align*} f(ad)=\frac{1}{a}\sum_{ad\leq d'< a(d+1)}f(d')+O(1/d), \end{align*} and inserting this into the preceding computation, we get \begin{align*} \E^{\log}_{p \leq X} f(ap)&=\frac{1}{\log\log X} \sum_{d' \leq aX} f(d') \cdot\frac{1}{a}\sum_{d'/a-1< d\leq d'/a}\frac{1}{d\log (2+d)}+O(\varepsilon)\\ &=\frac{1}{\log\log X} \sum_{d' \leq X} \frac{f(d')}{d'\log(2+d')}+O(\varepsilon). \end{align*} Taking the absolute value of the difference of the two sides of this equation, applying $\limsup_{X \to \infty}$ and then sending $\eps \to 0$, we obtain the claim. \end{proof} Now, let $\eps>0$ be a sufficiently small parameter. If one had $$ \sum_p \frac{1-|g_j(p)|}{p} = \infty$$ for some $1 \leq j \leq k$, then by Wirsing's theorem \cite{wirsing} as in \cite[\S 6]{tt} one would have $f_d(a)=0$ for all $a,d$. Thus we may assume that $$ \sum_p \frac{1-|g_j(p)|}{p} < \infty$$ for all $j$, which implies in particular that one has \begin{equation}\label{gp-inv} 1-\varepsilon\leq |G(p)|\leq 1 \end{equation} for all but finitely many $p$. For any integer $a$, we see from Corollary \ref{mango} that $$ \limsup_{P_1 \to \infty} \limsup_{P_2 \to \infty} \E_{p_1 \leq P_1}^{\log} \E_{p_2 \leq P_2}^{\log} |f_{p_1 p_2}(a) G(p_1) G(p_2) - f_1(ap_1 p_2)| \ll \eps.$$ By \eqref{gp-inv} we then have $$ \limsup_{P_1 \to \infty} \limsup_{P_2 \to \infty} \E_{p_1 \leq P_1}^{\log} \E_{p_2 \leq P_2}^{\log} |f_{p_1 p_2}(a) - \overline{G(p_1)} \overline{G(p_2)} f(ap_1 p_2)| \ll \eps$$ Applying Proposition \ref{pe}, we conclude that \begin{equation}\label{2m1-0} \limsup_{P_1 \to \infty} \limsup_{P_2 \to \infty} \E_{p_1 \leq P_1}^{\log} \E_{p_2 \leq P_2}^{\log} |f_{p_1 p_2}(a) - \overline{G(p_1)} \overline{G(p_2)} f_{1,0}(ap_1 p_2)| \ll \eps. \end{equation} In particular we have \begin{equation}\label{2m1} \E_{p_1 \in {\P}}^{\log,**} \E_{p_2 \in {\P}}^{\log,**} |f_{p_1 p_2}(a) - \overline{G(p_1)} \overline{G(p_2)} f_{1,0}(ap_1 p_2)| \ll \eps. \end{equation} Heuristically, \eqref{2m1} asserts the approximation \begin{equation}\label{2m1-approx} f_{p_1 p_2}(a) \approx \overline{G(p_1)} \overline{G(p_2)} f_{1,0}(ap_1 p_2) \end{equation} for ``most'' $a,p_1,p_2$. This turns out to be a remarkably powerful approximate equation, giving a lot of control on the functions $G$, $f_d$, and $f_{1,0}$. Roughly speaking, we will be able to show that the only way to solve \eqref{2m1-approx} (in a manner compatible with \eqref{fda} and \eqref{fd1d2}) is if $G(p) \approx \chi(p) p^{it}$, $f_d(a) \approx f(a) d^{-it}$, and $f_{1,0} \approx f$ for some $\chi$-isotypic $q$-periodic function $f$. Conversely, it is easy to see that if $G, f_d, f_{1,0}$ are of the above form, then they obey \eqref{2m1-approx}. We first use \eqref{2m1} to control $f_{1,0}$. Let $q$ denote the period of $f_{1,0}$ (which depends on $\varepsilon$); by abuse of notation, we view $f_{1,0}$ as a function on $\Z/q\Z$ as well as on $\Z$. We then have \begin{proposition}[Initial control on $f_{1,0}$]\label{prop1} Let $a_0$ be as in \eqref{fda}. We have \begin{equation}\label{f10-large} \E_{c \in (\Z/q\Z)^\times} |f_{1,0}( a_0 c)| \gg 1. \end{equation} Furthermore, for any integers $a_1,a_2$ and any natural number $b$ coprime to $q$, we have \begin{equation}\label{fito} \E_{c \in (\Z/q\Z)^\times} |f_{1,0}(a_1 c)f_{1,0}(a_2bc) - f_{1,0}(a_1 bc) f_{1,0}(a_2 c)| \ll \eps. \end{equation} \end{proposition} \begin{proof} By Lemma \ref{chug}, \eqref{fda} and \eqref{fd1d2}, we see that $$\E_{p_2 \in {\P}}^{\log,**} |f_{p_1 p_2}(a_0)| = \E_{d \in \N}^{\log\log,**} |f_{d}(a_0)| \gg 1$$ for any $p_1$, and hence $$ \E_{p_1 \in {\P}}^{\log,**} \E_{p_2 \in {\P}}^{\log,**} |f_{p_1 p_2}(a_0)| \gg 1.$$ On the other hand, from \eqref{2m1} we have \begin{equation}\label{m1} \E_{p_1 \in {\P}}^{\log,**} \E_{p_2 \in {\P}}^{\log,**} |f_{p_1 p_2}(a_0) - \overline{G(p_1)} \overline{G(p_2)} f_{1,0}(a_0 p_1 p_2)| \ll \eps. \end{equation} From the triangle inequality, we have $$ |f_{p_1 p_2}(a_0)| \ll |f_{1,0}(a_0 p_1 p_2)| + |f_{p_1 p_2}(a_0) - \overline{G(p_1)} \overline{G(p_2)} f_{1,0}(a_0 p_1 p_2)|+\varepsilon,$$ and hence (since $\eps$ is assumed small) $$ \E_{p_1 \in {\P}}^{\log,**} \E_{p_2 \in {\P}}^{\log,**} |f_{1,0}(a_0 p_1 p_2)| \gg 1.$$ By the periodicity of $f_{1,0}$ and the prime number theorem in arithmetic progressions, we conclude \eqref{f10-large}. Next, let $a_1,a_2,b$ be as in the proposition. Applying \eqref{2m1} twice, we see that \begin{equation}\label{mas-1} \E_{p_1 \in {\P}}^{\log,**} \E_{p_2 \in {\P}}^{\log,**} |f_{p_1 p_2}(a_1) - \overline{G(p_1)} \overline{G(p_2)} f_{1,0}(a_1 p_1 p_2)| \ll \eps \end{equation} and \begin{equation}\label{mas-2} \E_{p_1 \in {\P}}^{\log,**} \E_{p_2 \in {\P}}^{\log,**} |f_{p_1 p_2}(a_2) - \overline{G(p_1)} \overline{G(p_2)} f_{1,0}(a_2 p_1 p_2)| \ll \eps. \end{equation} We now eliminate the functions $f_{p_1p_2}$ and $G$ from these estimates. As in the proof of Lemma \ref{chug}, we can use the prime number theorem in arithmetic progressions to rearrange the left-hand side of \eqref{mas-1} as $$ \E_{p_1 \in {\P}}^{\log,**} \E_{c \in (\Z/q\Z)^\times} \E_{d \in \N}^{\log\log,**} \E_{d \leq p_2 < (1+\eps) d; p_2 = c\ (q)} |f_{p_1 p_2}(a_1) - \overline{G(p_1)} \overline{G(p_2)} f_{1,0}(a_1 p_1 p_2)| + O(\eps)$$ and hence after a change of variables $c \mapsto bc$ (and renaming $p_2$ as $p'_2$) $$ \E_{p_1 \in {\P}}^{\log,**} \E_{c \in (\Z/q\Z)^\times} \E_{d \in \N}^{\log\log,**} \E_{d \leq p'_2 < (1+\eps) d; p'_2 = bc\ (q)} |f_{p_1 p'_2}(a_1) - \overline{G(p_1)} \overline{G(p'_2)} f_{1,0}(a_1 p_1 p'_2)| \ll \eps.$$ From \eqref{fd1d2}, we have $f_{p_1 p_2}(a_1), f_{p_1 p'_2}(a_1) = f_{p_1 d}(a_1) + O(\eps)$; from the periodicity of $f_{1,0}$ we also have $f_{1,0}(a_1 p_1 p_2) = f_{1,0}(a_1 c p_1)$ and $f_{1,0}(a_1 p_1 p'_2) = f_{1,0}(a_1bc p_1)$. We conclude that $$ \E_{p_1 \in {\P}}^{\log,**} \E_{c \in (\Z/q\Z)^\times} \E_{d \in \N}^{\log\log,**} \E_{d \leq p_2 < (1+\eps) d; p_2 = c\ (q)} |f_{p_1 d}(a_1) - \overline{G(p_1)} \overline{G(p_2)} f_{1,0}(a_1 cp_1)| \ll \eps$$ and $$ \E_{p_1 \in {\P}}^{\log,**} \E_{c \in (\Z/q\Z)^\times} \E_{d \in \N}^{\log\log,**} \E_{d \leq p'_2 < (1+\eps) d; p'_2 = bc\ (q)} |f_{p_1 d}(a_1) - \overline{G(p_1)} \overline{G(p'_2)} f_{1,0}(a_1bcp_1)| \ll \eps$$ and hence by the triangle inequality and \eqref{gp-inv} we have \begin{align*} &\E_{p_1 \in {\P}}^{\log,**} \E_{c \in (\Z/q\Z)^\times} \E_{d \in \N}^{\log\log,**} \E_{d \leq p_2 < (1+\eps) d; p_2 = c\ (q)} \\ &\quad \E_{d \leq p'_2 < (1+\eps) d; p'_2 = bc\ (q)}|\overline{G(p_2)} f_{1,0}(a_1 cp_1) - \overline{G(p'_2)} f_{1,0}(a_1bcp_1)| \ll \eps. \end{align*} We have thus eliminated $f_{p_1 p_2}$ and one factor of $G$; we still seek to eliminate the other factor of $G$. To do this, we replace $a_1$ by $a_2$ in the above analysis to obtain \begin{align*} &\E_{p_1 \in {\P}}^{\log,**} \E_{c \in (\Z/q\Z)^\times} \E_{d \in \N}^{\log\log,**} \E_{d \leq p_2 < (1+\eps) d; p_2 = c\ (q)} \\ &\quad \E_{d \leq p'_2 < (1+\eps) d; p'_2 = bc\ (q)}|\overline{G(p_2)} f_{1,0}(a_2 cp_1) - \overline{G(p'_2)} f_{1,0}(a_2bcp_1)| \ll \eps. \end{align*} At this point, let us note that $|f_{1,0}(a)|\ll 1$ for $a\in \mathbb{Z}$. To see this, we use Corollary \ref{mango} to conclude that \begin{align*} \limsup_{P_1 \to \infty} \limsup_{P_2 \to \infty} \E_{p_1 \leq P_1; p_1\equiv 1\ (q)}^{\log} \E_{p_2 \leq P_2; p_2\equiv 1\ (q)}^{\log} |f_{p_1 p_2}(a) - \overline{G(p_1)} \overline{G(p_2)} f_{1,0}(a)| =0. \end{align*} Then from the triangle inequality, \eqref{gp-inv}, and the trivial bound $|f_{p_1p_2}(a)|\ll 1$ we reach the conclusion $|f_{1,0}(a)|\ll 1$. Next observe the identity \begin{align*} &\overline{G(p_2)} (f_{1,0}(a_1 cp_1)f_{1,0}(a_2bcp_1) - f_{1,0}(a_1bcp_1) f_{1,0}(a_2 cp_1))\\ &= f_{1,0}(a_2 bcp_1) (\overline{G(p_2)} f_{1,0}(a_1 cp_1) - \overline{G(p'_2)} f_{1,0}(a_1bcp_1))\\ &- f_{1,0}(a_1 bcp_1) (\overline{G(p_2)} f_{1,0}(a_2 cp_1) - \overline{G(p'_2)} f_{1,0}(a_2bcp_1)); \end{align*} we thus have from the triangle inequality, the boundedness of $|f_{1,0}(a)|$, and \eqref{gp-inv} that \begin{align*} &|f_{1,0}(a_1 cp_1)f_{1,0}(a_2bcp_1) - f_{1,0}(a_1bcp_1) f_{1,0}(a_2 cp_1)| \\ &\ll |\overline{G(p_2)} f_{1,0}(a_1 cp_1) - \overline{G(p'_2)} f_{1,0}(a_1bcp_1)| + |\overline{G(p_2)} f_{1,0}(a_2 cp_1) - \overline{G(p'_2)} f_{1,0}(a_2bcp_1)| \end{align*} for all but finitely many $p_1,p_2$, and thus by further application of the triangle inequality \begin{align*} & \E_{p_1 \in {\P}}^{\log,**} \E_{c \in (\Z/q\Z)^\times} \E_{d \in \N}^{\log\log,**}\E_{d \leq p_2 < (1+\eps) d; p_2 = c\ (q)} \\ &\quad \E_{d \leq p'_2 < (1+\eps) d; p'_2 = bc\ (q)} |f_{1,0}(a_1 cp_1)f_{1,0}(a_2bcp_1) - f_{1,0}(a_1bcp_1) f_{1,0}(a_2 cp_1)| \ll \eps. \end{align*} As the expression being averaged does not depend on $d,p_2,p'_2$, this bound simplifies to $$ \E_{p_1 \in {\P}}^{\log,**} \E_{c \in (\Z/q\Z)^\times} |f_{1,0}(a_1 cp_1)f_{1,0}(a_2bcp_1) - f_{1,0}(a_1bcp_1) f_{1,0}(a_2 cp_1)| \ll \eps$$ and by the prime number theorem in arithmetic progressions and the periodicity of $f_{1,0}$, this simplifies further (cf. Lemma \ref{chug}) to give the desired bound \eqref{fito}. \end{proof} Let $a$ be an integer, and let $b$ be coprime to $q$. Applying \eqref{fito} with $a_1 = a$ and $a_2 = a_0 c'$ for $c'$ coprime to $q$, and averaging, we conclude that $$ \E_{c' \in (\Z/q\Z)^\times} \E_{c \in (\Z/q\Z)^\times} |f_{1,0}(a c)f_{1,0}(a_0bc c') - f_{1,0}(a bc) f_{1,0}(a_0 c c')| \ll \eps$$ and hence $$ \E_{c' \in (\Z/q\Z)^\times} \E_{c \in (\Z/q\Z)^\times} \left|f_{1,0}(a c)f_{1,0}(a_0bc c') \overline{f_{1,0}(a_0cc')}- f_{1,0}(a bc) |f_{1,0}(a_0 c c')|^2 \right| \ll \eps.$$ By the triangle inequality, this implies that $$ \E_{c \in (\Z/q\Z)^\times} \left|f_{1,0}(a c) \E_{c' \in (\Z/q\Z)^\times} f_{1,0}(a_0bc c') \overline{f_{1,0}(a_0cc')} - f_{1,0}(a bc) \E_{c' \in (\Z/q\Z)^\times} |f_{1,0}(a_0 c c')|^2 \right| \ll \eps.$$ Making the change of variables $c'' = cc'$, this is $$ \E_{c \in (\Z/q\Z)^\times} \left|f_{1,0}(a c) \E_{c'' \in (\Z/q\Z)^\times} f_{1,0}(a_0bc'') \overline{f_{1,0}(a_0c'')} - f_{1,0}(a bc) \E_{c'' \in (\Z/q\Z)^\times} |f_{1,0}(a_0c'')|^2 \right| \ll \eps.$$ If we define the function $\psi: (\Z/q\Z)^\times \to \C$ by $$ \psi(b) \coloneqq \frac{\E_{c'' \in (\Z/q\Z)^\times} f_{1,0}(a_0bc'') \overline{f_{1,0}(a_0c'')}}{\E_{c'' \in (\Z/q\Z)^\times} |f_{1,0}(a_0c'')|^2}$$ then by \eqref{f10-large} and Cauchy-Schwarz, we have $\psi(b) = O(1)$ for all $b \in (\Z/q\Z)^\times$, and \begin{equation}\label{ecc} \E_{c \in (\Z/q\Z)^\times} |f_{1,0}(a c) \psi(b) - f_{1,0}(a bc)| \ll \eps \end{equation} for all $a \in \Z/q\Z$ and $b \in (\Z/q\Z)^\times$. By definition, $\psi(1)=1$. Next, we establish an approximate multiplicativity property of $\psi$, known as the \emph{quasimorphism} property \cite{quasi} in the literature. If $b_1, b_2 \in (\Z/q\Z)^\times$, then from three applications of \eqref{ecc} one has \begin{align*} \E_{c \in (\Z/q\Z)^\times} |f_{1,0}(a_0 c) \psi(b_1) - f_{1,0}(a_0 b_1c)| &\ll \eps \\ \E_{c \in (\Z/q\Z)^\times} |f_{1,0}(a_0 b_1 c) \psi(b_2) - f_{1,0}(a_0 b_1 b_2 c)| &\ll \eps \\ \E_{c \in (\Z/q\Z)^\times} |f_{1,0}(a_0 c) \psi(b_1 b_2) - f_{1,0}(a_0 b_1 b_2 c)| &\ll \eps. \end{align*} Applying the triangle inequality (after multiplying the first inequality by $|\psi(b_2)|$, we conclude that $$ \E_{c \in (\Z/q\Z)^\times} |f_{1,0}(a_0 c) (\psi(b_1 b_2) - \psi(b_1) \psi(b_2))| \ll \eps$$ and hence by \eqref{f10-large} we have the \emph{quasimorphism equation} $$ \psi(b_1 b_2) = \psi(b_1) \psi(b_2) + O(\eps).$$ We now apply a stability theorem to replace this quasimorphism on $(\Z/q\Z)^\times$ by a homomorphism (i.e., a Dirichlet character). \begin{lemma}[Stability of Dirichlet characters]\label{dirs} Let $\varepsilon>0$, and let $\psi: (\Z/q\Z)^\times \to \C$ be a function obeying the bound $\psi(b) = O(1)$ for all $b \in \Z/q\Z$, the identity $\psi(1) = 1$, and the quasimorphism equation $\psi(b_1 b_2) = \psi(b_1) \psi(b_2) + O(\eps)$ for all $b_1,b_2 \in (\Z/q\Z)^\times$. Then there exists a Dirichlet character $\chi: (\Z/q\Z)^\times \to \mathbb{S}^{1}$ of period $q$ such that $\psi(b) = \chi(b) + O(\eps)$ for all $b \in (\Z/q\Z)^\times$. \end{lemma} This lemma is a special case of Kazhdan\footnote{We thank Assaf Naor for this reference. Ben Green also pointed out to us the closely related fact that the bounded cohomology of amenable groups is trivial; see for instance \cite[Theorem 3.7]{frigerio}.} \cite{kazhdan}, and also follows from \cite[Proposition 5.3]{bgs} (which cites \cite{bfl} for a more general result), but for the convenience of the reader we give a self-contained proof here. \begin{proof} We can assume that $\eps$ is smaller than any given positive absolute constant, as the claim is trivial otherwise. Since $1 = \psi(1) = \psi(b) \psi(b^{-1}) + O(\eps)$ and $\psi(b^{-1}) = O(1)$, we see that $1 \ll |\psi(b)| \ll 1$ for all $b \in (\Z/q\Z)^\times$. We introduce the cocycle $\rho: (\Z/q\Z)^\times \times (\Z/q\Z)^\times \to \C$ by defining $\rho(b_1,b_2)$ for $b_1,b_2 \in (\Z/q\Z)^\times$ to be the unique complex number of size $O(\eps)$ such that \begin{equation}\label{psib} \psi(b_1 b_2) = \psi(b_1) \psi(b_2) \exp( \rho(b_1, b_2) ); \end{equation} this is well-defined for $\eps$ small enough. For $b_1,b_2,b_3 \in (\Z/q\Z)^\times$, we have $$ \psi(b_1 b_2 b_3) = \psi(b_1 b_2) \psi(b_3) \exp( \rho(b_1b_2, b_3) ) = \psi(b_1) \psi(b_2) \psi(b_3) \exp( \rho(b_1, b_2) + \rho(b_1 b_2, b_3) )$$ and $$ \psi(b_1 b_2 b_3) = \psi(b_1) \psi(b_2 b_3) \exp( \rho(b_2, b_3) ) = \psi(b_1) \psi(b_2) \psi(b_3) \exp( \rho(b_1, b_2 b_3) + \rho(b_2, b_3) )$$ which on taking logarithms yields (for $\eps$ small enough) the \emph{cocycle equation} $$ \rho(b_1, b_2) + \rho(b_1 b_2, b_3) = \rho(b_1, b_2 b_3) + \rho(b_2, b_3).$$ Averaging in $b_3$, we conclude the \emph{coboundary equation} $$ \rho(b_1, b_2) + \phi(b_1 b_2) = \phi(b_1) + \phi(b_2)$$ where $\phi(b) \coloneqq \E_{b_3 \in (\Z/q\Z)^\times} \rho(b,b_3)$. If we then define the function $\chi: \Z/q\Z \to \C$ by $$ \chi(b) \coloneqq \psi(b) \exp( \phi(b) ),$$ then $\psi(b) = \chi(b) + O(\eps)$ for all $b \in (\Z/q\Z)^\times$, and from \eqref{psib} we have $$ \chi(b_1 b_2) = \chi(b_1) \chi(b_2)$$ for all $b_1,b_2 \in (\Z/q\Z)^\times$, thus $\chi: (\Z/q\Z)^\times \to \C$ is a homomorphism and therefore a Dirichlet character of period $q$. The claim follows. \end{proof} Let $\chi$ be the Dirichlet character of period $q$ provided by the above lemma, then from \eqref{ecc} and the triangle inequality we have the \emph{approximate isotopy equation} $$ \E_{c \in (\Z/q\Z)^\times} |f_{1,0}(a c) \chi(b) - f_{1,0}(a bc)| \ll \eps $$ for all $a \in \Z/q\Z$ and $b \in (\Z/q\Z)^\times$. We can rearrange this as $$ \E_{c \in (\Z/q\Z)^\times} |f_{1,0}(a c) - \overline{\chi(b)} f_{1,0}(a bc)| \ll \eps $$ and average in $b$ to conclude that \begin{equation}\label{tfa} \E_{c \in (\Z/q\Z)^\times} |f_{1,0}(a c) - \tilde f(ac)| \ll \eps \end{equation} for all $a$, where $\tilde f: \Z/q\Z \to \C$ is the function $$ \tilde f(a) \coloneqq \E_{b \in (\Z/q\Z)^\times} \overline{\chi(b)} f_{1,0}(a b).$$ Observe that $\tilde f$ is $\chi$-isotypic in the sense that $$ \tilde f(ab) = \chi(b) \tilde f(a)$$ whenever $a \in \Z/q\Z$ and $b \in (\Z/q\Z)^\times$. From \eqref{tfa} and \eqref{f10-large}, one has $$ \E_{c \in (\Z/q\Z)^\times} |\tilde f(a_0 c)| \gg 1$$ and hence by the $\chi$-isotypy of $\tilde f$ \begin{equation}\label{tfa0} |\tilde f(a_0)| \gg 1. \end{equation} Now we work to control $f_d$. Let $a$ be an integer. From \eqref{tfa} and the prime number theorem in arithmetic progressions, we have $$ \limsup_{P_1 \to \infty} \limsup_{P_2 \to \infty} \E_{p_1 \leq P_1}^{\log} \E_{p_2 \leq P_2}^{\log} |f_{1,0}(a p_1 p_2) - \tilde f(a p_1 p_2)| \ll \eps $$ From this, \eqref{gp-inv}, \eqref{2m1-0}, and the triangle inequality, we conclude that $$ \limsup_{P_1 \to \infty} \limsup_{P_2 \to \infty} \E_{p_1 \leq P_1}^{\log} \E_{p_2 \leq P_2}^{\log} |f_{p_1 p_2}(a) - \overline{G(p_1)}\, \overline{G(p_2)} \tilde f(a p_1 p_2)| \ll \eps. $$ Using the $\chi$-isotopy of $\tilde f$, we can write this as \begin{equation}\label{pip-0} \limsup_{P_1 \to \infty} \limsup_{P_2 \to \infty} \E_{p_1 \leq P_1}^{\log} \E_{p_2 \leq P_2}^{\log} |f_{p_1 p_2}(a) - \overline{G}\chi(p_1) \overline{G}\chi(p_2) \tilde f(a)| \ll \eps. \end{equation} This has the following useful consequence. \begin{lemma}[Isotopy]\label{isotop} Let the notation be as above. Let $a$ be an integer and let $b$ be an integer coprime to $q$. Then we have $$ \limsup_{X\to \infty}\E_{d \leq X}^{\log\log} |f_d(ab) - \chi(b) f_d(a)| \ll \eps.$$ \end{lemma} \begin{proof} It suffices to prove the claim with an arbitrary generalised limit $\lim^{*}_{X\to \infty}$ in place of $\limsup_{X\to \infty}$. From \eqref{pip-0} we have $$ \E_{p_1 \in {\P}}^{\log,*} \E_{p_2 \in {\P}}^{\log,*} |f_{p_1 p_2}(a) - \overline{G}\chi(p_1) \overline{G}\chi(p_2) \tilde f(a)| \ll \eps$$ and $$ \E_{p_1 \in {\P}}^{\log,*} \E_{p_2 \in {\P}}^{\log,*} |f_{p_1 p_2}(ab) - \overline{G}\chi(p_1) \overline{G}\chi(p_2) \tilde f(ab)| \ll \eps.$$ As $\tilde f$ is isotypic, $\tilde f(ab) = \chi(b) \tilde f(a)$. From the triangle inequality and \eqref{gp-inv}, we conclude that $$ \E_{p_1 \in {\P}}^{\log,*} \E_{p_2 \in {\P}}^{\log,*} |f_{p_1 p_2}(ab) - \chi(b) f_{p_1 p_2}(a)| \ll \eps.$$ On the other hand, since $d\mapsto |f_{d}(ab)-\chi(b)f_{d}(a)|$ is bounded log-Lipschitz by \eqref{fd1d2}, by Lemma \ref{chug} for any $p_1$ we have $$ \E_{p_2 \in {\P}}^{\log,*} |f_{p_1 p_2}(ab) - \chi(b) f_{p_1 p_2}(a)|=\E_{d\in \N}^{\log\log, *}|f_{d}(ab)-\chi(b)f_{d}(a)|, $$ and now the claim now follows by taking the average $\E_{p_1\in \P}^{\log,*}$ on both sides. \end{proof} Now we derive another consequence of \eqref{pip-0}. Let $x>0$ be a positive real, and let $a$ be an integer. From \eqref{pip-0} we have $$ \E_{p'_1 \in {\P}}^{\log, **} \E_{p_2 \in {\P}}^{\log,**} |f_{p'_1 p_2}(a) - \overline{G}\chi(p'_1) \overline{G}\chi(p_2) \tilde f(a)| \ll \eps. $$ By the prime number theorem, this can also be written as $$ \E_{p_1 \in {\P}}^{\log,**} \E_{x p_1 \leq p'_1 \leq (1+\eps) xp_1} \E_{p_2 \in {\P}}^{\log,**} |f_{p'_1 p_2}(a) - \overline{G}\chi(p'_1) \overline{G}\chi(p_2) \tilde f(a)| \ll \eps. $$ From \eqref{gp-inv} we have $$1-\varepsilon \leq |\overline{G}\chi(p_1)|, |\overline{G}\chi(p'_1)| , |\overline{G}\chi(p_2)|\leq 1$$ for all but finitely many $p_1,p_1',p_2$, so that \begin{align*} &|f_{p'_1 p_2}(a) - \overline{G}\chi(p'_1) G \overline{\chi}(p_1) f_{p_1 p_2}(a)|\\ &\ll |f_{p_1 p_2}(a) - \overline{G}\chi(p_1) \overline{G}\chi(p_2) \tilde f(a)| + |f_{p'_1 p_2}(a) - \overline{G}\chi(p'_1) \overline{G}\chi(p_2) \tilde f(a)|+O(\varepsilon). \end{align*} Thus by the triangle inequality we have $$ \E_{p_1 \in {\P}}^{\log,**} \E_{x p_1 \leq p'_1 \leq (1+\eps) xp_1} \E_{p_2 \in {\P}}^{\log,**} |f_{p'_1 p_2}(a) - \overline{G}\chi(p'_1) G \overline{\chi}(p_1) f_{p_1 p_2}(a)| \ll \eps. $$ From \eqref{fd1d2} we have $f_{p'_1 p_2}(a) = f_{x p_1 p_2}(a) + O(\eps)$ (recall that $f_d$ is defined for any real $d>0$), thus $$ \E_{p_1 \in {\P}}^{\log,**} \E_{x p_1 \leq p'_1 \leq (1+\eps) xp_1} \E_{p_2 \in {\P}}^{\log,**} |f_{x p_1 p_2}(a) - \overline{G}\chi(p'_1) G \overline{\chi}(p_1) f_{p_1 p_2}(a)| \ll \eps $$ and thus by the triangle inequality $$ \E_{p_1 \in {\P}}^{\log,**} \E_{p_2 \in {\P}}^{\log,**} |f_{x p_1 p_2}(a) - \alpha_{p_1}(x) f_{p_1 p_2}(a)| \ll \eps, $$ where $$ \alpha_{p_1}(x) \coloneqq \E_{x p_1 \leq p'_1 < (1+\eps) x p_1} \overline{G}\chi(p'_1) G \overline{\chi}(p_1).$$ By Lemma \ref{chug}, this implies that \begin{equation}\label{edl-0} \E_{p_1 \in {\P}}^{\log,**} \E_{d \in \N}^{\log\log,**} |f_{x d}(a) - \alpha_{p_1}(x) f_{d}(a)| \ll \eps \end{equation} which by the triangle inequality implies that \begin{equation}\label{edl} \E_{d \in \N}^{\log\log,**} |f_{x d}(a) - \alpha(x) f_{d}(a)| \ll \eps \end{equation} where $$ \alpha(x) \coloneqq \E_{p_1 \in {\P}}^{\log,**} \alpha_{p_1}(x).$$ By construction, we have $\alpha(x) = O(1)$ for all $x$. Setting $a=a_0$ in \eqref{edl} and using \eqref{fd1d2} to write $f_{xd}(a)=f_{d}(a)+O(\varepsilon)$ for $|x-1|\leq \varepsilon$, we deduce from \eqref{fda} that $\alpha(x)=1+O(\varepsilon)$ for $|x-1|\leq \varepsilon$. Next, for $x,y > 0$, we have the estimates \begin{align*} \E_{d \in \mathbb{N}}^{\log\log, **} |f_{x d}(a_0) - \alpha(x) f_{d}(a_0)| &\ll \eps\\ \E_{d \in \mathbb{N}}^{\log\log, **} |f_{x y d}(a_0) - \alpha(y) f_{xd}(a_0)| &\ll \eps\\ \E_{d \in \mathbb{N}}^{\log\log, **} |f_{x y d}(a_0) - \alpha(xy) f_{d}(a_0)| &\ll \eps, \end{align*} which by the triangle inequality and \eqref{fda} implies the quasimorphism equation $$ \alpha(xy) = \alpha(x) \alpha(y) + O(\eps).$$ We now require the Archimedean analogue of Lemma \ref{dirs} (which is also a special case of the results of \cite{kazhdan}). \begin{lemma}[Stability of Archimedean characters]\label{dira} Let $\alpha: (0,+\infty) \to \C$ be any function obeying the bound $\alpha(x) = O(1)$ for all $x > 0$, such that $\alpha(x) = 1 + O(\eps)$ when $|x-1| \leq \eps$, and $\alpha(xy) = \alpha(x) \alpha(y) + O(\eps)$ for all $x,y>0$. Then there exists a real number $t$ such that $\alpha(x) = x^{-it} + O(\eps)$ for all $x>0$. \end{lemma} \begin{proof} As before, we can assume $\eps$ is smaller than any given positive constant, as the claim is trivial otherwise. Since $\alpha(1) = 1+O(\eps)$ and $\alpha(1) = \alpha(x) \alpha(1/x) + O(\eps)$, we have the bounds $1 \ll |\alpha(x)| \ll 1$ for all $x$. By construction, we also have $\alpha(xy) = \alpha(x) + O(\eps)$ whenever $1 \leq y \leq 1+\eps$. By replacing $\alpha$ with the discretised version \begin{align*} \alpha_1(x):=\begin{cases}\alpha(\varepsilon^2\lfloor \frac{x}{\varepsilon^2}\rfloor),\quad x\geq \varepsilon,\\ \alpha(\frac{1}{n}),\quad x\in (\frac{1}{n+1},\frac{1}{n}],\,\, 0<x<\varepsilon,\end{cases} \end{align*} we may assume that $\alpha$ is Lebesgue measurable. The function $\alpha_1$ continues to enjoy the same properties as $\alpha$, since $\alpha_1(x)=\alpha(x)+O(\varepsilon)$ for all $x>0$. To simplify notation, we denote $\alpha_1$ by $\alpha$ in what follows. We introduce the cocycle $\rho: (0,+\infty) \times (0,+\infty) \to \C$ by defining $\rho(x_1,x_2)$ for $x_1,x_2 > 0$ to be the unique complex number of size $O(\eps)$ such that \begin{align}\label{eq3} \alpha(x_1 x_2) = \alpha(x_1) \alpha(x_2) \exp( \rho(x_1, x_2) ); \end{align} this is well-defined and measurable for $\eps$ small enough. Arguing exactly as in the proof of Lemma \ref{dirs}, we obtain the cocycle equation $$ \rho(x_1, x_2) + \rho(x_1 x_2, x_3) = \rho(x_1, x_2 x_3) + \rho(x_2, x_3).$$ Taking an asymptotic logarithmic average in $x_3$, we conclude the coboundary equation \begin{equation}\label{eq21} \rho(x_1, x_2) + \phi(x_1 x_2) = \phi(x_1) + \phi(x_2) \end{equation} where $$ \phi(x) \coloneqq \plim_{M \to \infty} \frac{1}{\log M} \int_{1}^{M} \rho(x, x_3) \frac{dx_3}{x_3}.$$ If we then define the function $\tilde \alpha: (0,+\infty) \to \C$ by $$ \tilde \alpha(x) \coloneqq \alpha(x) \exp( \phi(x) )$$ then $\tilde \alpha(x) = \alpha(x) + O(\eps)$ for all $x>0$, and from \eqref{eq3} and \eqref{eq21} we have $$ \tilde \alpha(xy) = \tilde \alpha(x) \tilde \alpha(y)$$ for all $x,y > 0$, thus $\tilde \alpha: (0,+\infty) \to \C$ is a homomorphism. Also, by construction one has $\tilde \alpha(x) = O(1)$ for all $x$, so $\tilde \alpha$ in fact takes values in the unit circle $\mathbb{S}^{1}$. We have $\tilde \alpha(x) = 1+O(\eps)$ when $|x-1| \leq \eps$, and we will use this additional information to show that $\tilde \alpha(x)=x^{it}$ for some real $t$ and all $x>0$. If $|x-1| \leq \eps/n$ for some natural number $n$, then $\tilde \alpha(x)^n, \tilde \alpha(x) = 1 + O(\eps)$, which implies that $\tilde \alpha(x) = 1 + O(\eps/n)$. This implies that $\tilde \alpha(x) = 1 + O(|x-1|)$, and so $\tilde \alpha$ is continuous at $1$ and hence continuous on all of $(0,+\infty)$. Next, if $x_0 \coloneqq 1 + \eps$ then we have $\tilde \alpha(x_0) = x_0^{it}$ for some $t=O(1)$; taking roots we conclude that $\tilde \alpha(x_0^{1/n}) = (x_0^{1/n})^{it}$ for all natural numbers $n$, and hence $\tilde \alpha(x_0^{m/n}) = (x_0^{m/n})^{it}$ for all natural numbers $n$ and integers $m$. By continuity we conclude that $\tilde \alpha(x) = x^{it}$ for all $x \in (0,+\infty)$, as required. \end{proof} From the above lemma, we conclude that there is a real number $t$ with the property that for every integer $a$ and real $x>0$, one has \begin{equation} \label{eq25} \E_{d \in \N}^{\log\log,**} |f_{x d}(a) - x^{-it} f_{d}(a)| \ll \eps.\end{equation} In particular, for every prime $p_1$, one has \begin{equation*} \E_{d \in \N}^{\log\log,**} |f_{p_1 d}(a_0) - p_1^{-it} f_{d}(a_0)| \ll \eps,\end{equation*} and thus \begin{equation}\label{eq22} \E_{p_1 \in {\P}}^{\log,**} \E_{d \in \N}^{\log\log,**} |f_{p_1 d}(a_0) - p_1^{-it} f_{d}(a_0)| \ll \eps\end{equation} On the other hand, from Proposition \ref{dollop} one has that if $P_1$ is sufficiently large depending on $a_0,\eps$, then $$ \sup_{d>0} \E^{\log}_{p_1 \leq P_1} |f_{p_1 d}(a_0) G(p_1) - f_d(a_0p_1)| \ll \eps.$$ Hence on averaging in $d$ and taking limits in the $d$ average and then in the $p_1$ average, we conclude that \begin{equation}\label{eq23} \limsup_{P_1 \to \infty} \E^{\log}_{p_1 \leq P_1} \E_{d \in \N}^{\log\log,**} |f_{p_1 d}(a_0) G(p_1) - f_d(a_0p_1)| \ll \eps.\end{equation} Meanwhile, from Lemma \ref{isotop} we have $$ \E_{d \in \N}^{\log\log,**} |f_d(a_0p_1) - \chi(p_1) f_d(a_0)| \ll \eps$$ for all sufficiently large $p_1$, and thus \begin{equation} \label{eq24} \limsup_{P_1 \to \infty} \E^{\log}_{p_1 \leq P_1} \E_{d \in \N}^{\log\log,**} |f_d(a_0p_1) - \chi(p_1) f_d(a_0)| \ll \eps.\end{equation} Applying the triangle inequality to \eqref{eq22}, \eqref{eq23}, \eqref{eq24}, we obtain $$ \limsup_{P_1 \to \infty} \E^{\log}_{p_1 \leq P_1} \E_{d \in \N}^{\log\log,**} |G(p_1) - \chi(p_1) p_1^{it}| |f_d(a_0)| \ll \eps $$ and hence by \eqref{fda} we have $$ \limsup_{P_1 \to \infty} \E^{\log}_{p_1 \leq P_1} |G(p_1) - \chi(p_1) p_1^{it}| \ll \eps.$$ To summarise the above analysis, we have shown that for every $\eps>0$ there exists a Dirichlet character $\chi = \chi_\eps$ and a real number $t = t_\eps$ such that $$ \limsup_{P_1 \to \infty} \E^{\log}_{p_1 \leq P_1} |G(p_1) - \chi_\eps(p_1) p_1^{it_\eps}| \ll \eps.$$ \emph{A priori}, the character $\chi_\eps$ and the real number $t_\eps$ depend on $\eps$. But if $\eps,\eps'>0$ are sufficiently small, we have from the triangle inequality that $$ \limsup_{P_1 \to \infty} \E^{\log}_{p_1 \leq P_1} |\chi_{\eps'}(p_1) p_1^{it_{\eps'}} - \chi_\eps(p_1) p_1^{it_\eps}| \ll \eps + \eps'.$$ But from the prime number theorem in arithmetic progressions and partial summation, we see that the left-hand side is $\gg 1$ unless $t_\eps = t_{\eps'}$ and the Dirichlet characters are \emph{co-trained} in the sense that they are both induced from the same primitive character $\chi$. We conclude that there exists a primitive character $\chi$ independent of $\eps$, and a real number $t_0$ independent of $\eps$, such that $t_\eps = t_0$ and $\chi_\eps$ is induced from $\chi$ for $\eps$ sufficiently small. In particular, as $\chi_{\varepsilon}(p_1)$ and $\chi(p_1)$ agree for all but $O_{\varepsilon}(1)$ primes $p_1$, we have for each $\eps>0$ that $$ \limsup_{P_1 \to \infty} \E^{\log}_{p_1 \leq P_1} |G(p_1) - \chi(p_1) p_1^{it_0}| \ll \eps $$ and thus \begin{equation}\label{lop} \E^{\log}_{p_1 \in {\P}} |G(p_1) - \chi(p_1) p_1^{it_0}| = 0. \end{equation} Thus $G$ weakly pretends to be the twisted Dirichlet character $n \mapsto n^{it_0} \chi(n)$. This (vacuously) establishes part (i) of Theorem \ref{main}. Now let $\eps>0$ be small, and let $a$ be an integer. From \eqref{pip-0} (and the fact that $\chi_\eps$ is induced from $\chi$), and making the dependence of $\tilde f_\eps$ on $\eps$ explicit, we have $$ \E_{p_1 \in {\P}}^{\log,**} \E_{p_2 \in {\P}}^{\log,**} |f_{p_1 p_2}(a) - \overline{G}\chi(p_1) \overline{G}\chi(p_2) \tilde f_\eps(a)| \ll \eps $$ and hence by \eqref{lop} and the triangle inequality $$ \E_{p_1 \in {\P}}^{\log,**} \E_{p_2 \in {\P}}^{\log,**} |f_{p_1 p_2}(a) - p_1^{-it_0} p_2^{-it_0} \tilde f_\eps(a)| \ll \eps $$ or equivalently $$ \E_{p_1 \in {\P}}^{\log,**} \E_{p_2 \in {\P}}^{\log,**} |(p_1p_2)^{it_0} f_{p_1 p_2}(a) - \tilde f_\eps(a)| \ll \eps. $$ Applying \eqref{fd1d2}, Lemma \ref{chug} and \eqref{eq25} (where we can in fact take $\varepsilon\to 0$, since the deduction succeeding this formula shows that $t=t_0$ is independent of $\varepsilon$), we have $$\E_{p_2 \in {\P}}^{\log,**} |(p_1p_2)^{it_0} f_{p_1 p_2}(a) - \tilde f_\eps(a)| = \E_{d \in \N}^{\log\log,**} |(p_1d)^{it_0} f_{p_1 d}(a) - \tilde f_\eps(a)| = \E_{d \in \N}^{\log\log,**} |d^{it_0} f_d(a) - \tilde f_\eps(a)| $$ for any $p_1$, and hence $$ \E_{d \in \N}^{\log\log,**} |d^{it_0} f_{d}(a) - \tilde f_\eps(a)| \ll \eps. $$ We thus see from the triangle inequality that $$ |\tilde f_\eps(a) - \tilde f_{\eps'}(a)| \ll \eps + \eps'$$ and so $\tilde f_\eps$ converges uniformly to a limit $f$ with \begin{equation}\label{ttfa} |\tilde f_\eps(a) - f(a)| \ll \eps \end{equation} and thus by the triangle inequality, we have $$ \E_{d \in \N}^{\log\log,**} |d^{it_0} f_{d}(a) - f(a)| \ll \eps $$ whenever $\eps>0$, which gives \begin{equation}\label{dlog} \E_{d \in \N}^{\log\log,**} |d^{it_0} f_{d}(a) - f(a)| = 0. \end{equation} From \eqref{fda} we see in particular that $f(a_0) \neq 0$. By construction, each $\tilde f_\eps$ is $\chi$-isotypic in the sense that $\tilde f_\eps(ab) = \chi(b) \tilde f_\eps(a)$ whenever $a,b$ are integers with $b$ coprime to the periods of both $\chi$ and $\tilde f_\eps$. Hence, what remains to be shown is that \eqref{dlog} holds also when taking the average with respect to the ordinary limit. Now let $\eps>0$ be arbitrary. Inserting \eqref{lop} into \eqref{pip-0}, we see that $$ \limsup_{P_1 \to \infty} \limsup_{P_2 \to \infty} \E_{p_1 \leq P_1}^{\log} \E_{p_2 \leq P_2}^{\log} |f_{p_1 p_2}(a) - (p_1 p_2)^{-it_0} \tilde f_\eps(a)| \ll \eps $$ and hence by \eqref{ttfa} and sending $\eps \to 0$ we get \begin{equation*} \limsup_{P_1 \to \infty} \limsup_{P_2 \to \infty} \E_{p_1 \leq P_1}^{\log} \E_{p_2 \leq P_2}^{\log} |f_{p_1 p_2}(a) - (p_1 p_2)^{-it_0} f(a)| = 0. \end{equation*} For any $\varepsilon>0$ and any $P_1$ large enough in terms of $\varepsilon$, we apply Lemma \ref{chug}, Proposition \ref{dollop}, formula \eqref{lop} and Lemma \ref{isotop} to write \begin{align*} &\limsup_{P_2 \to \infty} \E_{p_1 \leq P_1}^{\log} \E_{p_2 \leq P_2}^{\log} |f_{p_1 p_2}(a) - (p_1 p_2)^{-it_0} f(a)|\\ =& \limsup_{P_2 \to \infty} \E_{p_1 \leq P_1}^{\log} \E_{d \leq P_2}^{\log\log} |f_{p_1 d}(a) - (p_1 d)^{-it_0} f(a)| \\ =& \limsup_{P_2 \to \infty} \E_{p_1 \leq P_1}^{\log} \E_{d \leq P_2}^{\log\log} |\overline{G(p_1)}f_d(ap_1) - (p_1 d)^{-it_0} f(a)|+O(\varepsilon)\\ =&\limsup_{P_2 \to \infty} \E_{p_1 \leq P_1}^{\log} \E_{d \leq P_2}^{\log\log} |p_1^{-it_0}\overline{\chi}(p_1)f_d(ap_1) - (p_1 d)^{-it_0} f(a)|+O(\varepsilon)\\ =& \limsup_{P_2 \to \infty} \E_{d \leq P_2}^{\log\log} |f_{d}(a) - d^{-it_0} f(a)|+O(\varepsilon), \end{align*} and hence, sending $\varepsilon\to 0$, we obtain $$ \E_{d \in \N}^{\log\log} |f_d(a) - d^{-it_0} f(a)| = 0.$$ This establishes part (ii) of Theorem \ref{main} (recalling as before that as $G$ weakly pretends to be a twisted Dirichlet character $n \mapsto \chi(n) n^{it}$, it can only weakly pretend to be another twisted Dirichlet character $n \mapsto \chi'(n) n^{it'}$ if $t=t'$ and $\chi, \chi'$ are co-trained). \section{Proofs of corollaries}\label{corollaries-sec} In this section we use Theorem \ref{main} to prove Corollaries \ref{elliott-1}, \ref{elliott-2}, \ref{chow}. We begin with Corollary \ref{elliott-1}. \begin{proof}[Proof of Corollary \ref{elliott-1}] Suppose the claim failed, then we can find $k, g_1,\dots,g_k$ as in that corollary, as well as $h_1,\dots,h_k \in \Z$ and $\eps>0$, such that the set $$ {\mathcal X} \coloneqq \{ X \in \N: |\E_{n \leq X} g_1(n+h_1) \cdots g_k(n+h_k)| > \eps \}$$ does not have logarithmic Banach density zero. In particular, one can find sequences $X_i \geq \omega_i \to \infty$ and $0 < \delta < 1/2$ such that \begin{align}\label{eq4} \E^{\log}_{X_i/\omega_i \leq X \leq X_i} 1_{{\mathcal X}}(x) \geq \delta \end{align} for all $i$. Intuitively, if the exceptional set $\mathcal{X}$ was big in the sense of \eqref{eq4}, there would have to be a lot of ``points of density'' of $\mathcal{X}$ (in a sense to be specified later). To make this rigorous, we introduce for each $i$ the function $a_i: \R \to [0,1]$ given by $$ a_i(s) \coloneqq \sum_{X_i/\omega_i \leq X \leq X_i: X \not \in {\mathcal X}} 1_{\log(X-1) < s \leq \log X}.$$ Note that $a_i(s)$ is the indicator function of the event that there exists an integer $X\not \in \mathcal{X}$ with $X\in [e^{s},e^{s}+1)$ and $X_i/\omega_i\leq X\leq X_i$. The function $a_i$ is a piecewise constant function supported on an interval of length $(1 + o_{i \to \infty}(1)) \log \omega_i$ and has integral \begin{align*} \int_\R a_i(s)\ ds &= \sum_{X_i/\omega_i \leq X \leq X_i: X \not \in {\mathcal X}} \log \frac{X}{X-1} \\ &= \left(\sum_{X_i/\omega_i \leq X \leq X_i: X \not \in {\mathcal X}} \frac{1}{X}\right) + O(1) \\ &= \log \omega_i + O(1) - \sum_{X_i/\omega_i \leq X \leq X_i: X \in {\mathcal X}} \frac{1}{X} \\ &\leq (1-\delta + o_{i \to \infty}(1)) \log \omega_i. \end{align*} We introduce the one-sided Hardy--Littlewood maximal function $$ Ma_i(s) \coloneqq \sup_{r>0} \frac{1}{r} \int_{s-r}^s a_i(s')\ ds'.$$ It is a well-known consequence of the rising sun lemma \cite{riesz} that one has the Hardy--Littlewood maximal inequality $$ m( \{ s \in \R: Ma_i(s) \geq \lambda \} ) \leq \frac{1}{\lambda} \int_\R a_i(s)\ ds$$ for any $\lambda>0$, where $m$ denotes Lebesgue measure. Applying this with $\lambda \coloneqq (1-\delta)^{1/2}$, we conclude that $$ m(\{ s \in \R: Ma_i(s) \geq (1-\delta)^{1/2} \} ) \leq ((1-\delta)^{1/2} + o_{i \to \infty}(1)) \log \omega_i.$$ In particular, one can find a real number $s_i$ with \begin{equation}\label{txi} \log X_i - ((1-\delta)^{1/2} + o_{i \to \infty}(1)) \log \omega_i \leq s_i \leq \log X_i \end{equation} such that $$ Ma_i(s_i) < (1-\delta)^{1/2} $$ which implies that \begin{equation}\label{ar} \int_{s_i-r}^{s_i} a_i(t)\ dt \leq (1-\delta)^{1/2} r \end{equation} for all $r>0$. Informally, the estimate \eqref{ar} asserts that the natural number $\lfloor \exp(s_i) \rfloor$ is a ``multiplicative point of density'' for the exceptional set ${\mathcal X}$. By passing to subsequences, and using a diagonalisation argument, we may assume that the limits \begin{align}\label{eq5} f_d(a) \coloneqq \lim_{i \to \infty} \E_{n \leq \lfloor \exp(s_i)\rfloor/d} g_1(n+ah_1) \cdots g_k(n+ah_k), \end{align} exist for every natural number $d$ and integer $a$. In particular, the limit of the right-hand side of \eqref{eq5} is the same along any generalised limit $\lim^{**}$. If we now apply Theorem \ref{main}(i) to a generalised limit of the form $$ \lim^*_{X \to \infty} f(X) \coloneqq \lim^{**}_{i \to \infty} f( \lfloor \exp(s_i) \rfloor ),$$ where $\lim^{**}$ is any generalised limit, we conclude that $$ \E^{\log\log}_{d \in \N} |f_d(1)| = 0.$$ Thus, if we let $\mu>0$ denote a small constant (depending on $\delta,\eps$) to be chosen later, and $D$ is sufficiently large depending on $\mu$, we have $$ \E^{\log\log}_{d \leq D} |f_d(1)| \leq \mu.$$ Thus by the triangle inequality $$ \limsup_{i \to \infty} \E^{\log\log}_{d \leq D} |\E_{n \leq \lfloor \exp(s_i)\rfloor/d} g_1(n+h_1) \cdots g_k(n+h_k)| \leq \mu, $$ and hence for all sufficiently large $i$ (depending on $\delta,\eps,\mu,D$) we find $$ \E_{d\leq D}^{\log \log} |\E_{n \leq \lfloor \exp(s_i)\rfloor/d} g_1(n+h_1) \cdots g_k(n+h_k)| \leq 2\mu. $$ This implies $$ \sum_{\log D \leq d \leq D} \frac{1}{d \log d} |\E_{n \leq \lfloor \exp(s_i)\rfloor/d} g_1(n+h_1) \cdots g_k(n+h_k)| \ll \mu \log\log D, $$ say. In particular, by Markov's inequality one has \begin{equation}\label{df} |\E_{n \leq \lfloor \exp(s_i)\rfloor/d} g_1(n+h_1) \cdots g_k(n+h_k)| \leq \frac{\eps}{2} \end{equation} for all $\log D \leq d \leq D$ outside of an exceptional set ${\mathcal D}_i$ with \begin{equation}\label{di} \sum_{d\in {\mathcal D}_i} \frac{1}{d \log d} \ll \frac{\mu}{\eps} \log\log D. \end{equation} If $\log D \leq d \leq D$ lies outside of ${\mathcal D}_i$, then one has $$ |\E_{n \leq X} g_1(n+h_1) \cdots g_k(n+h_k)| < \eps$$ for all $X$ between $\frac{\exp(s_i)}{d+1}-1$ and $\frac{\exp(s_i)}{d}+1$. In particular, all such $X$ lie outside of ${\mathcal X}$. Using \eqref{txi} (which places $\exp(s_i)/d$ below $X_i$ and well above $X_i/\omega_i$), we conclude that $$ a_i(t) = 1$$ on the interval $[s_i - \log(d+1), s_i - \log(d)]$. In particular, $$ \int_d^{d+1} a_i( s_i - \log u)\ du = 1.$$ For $d \in {\mathcal D}_i$ we use the trivial bound $$ \int_d^{d+1} a_i( s_i - \log u)\ du \geq 0.$$ From \eqref{di} we conclude that \begin{equation}\label{eq27} \sum_{\log D \leq d \leq D} \frac{1}{d \log d} \int_d^{d+1} a_i( s_i - \log u)\ du \geq (1 - O(\frac{\mu}{\eps})) \log\log D. \end{equation} The left-hand side, up to errors that can be absorbed into the $O( \frac{\mu}{\eps}) \log\log D$ term, can be rewritten as $$ \int_{\log D}^D a_i( s_i - \log u) \frac{du}{u\log u}$$ which by the change of variables $s = s_i - \log u$ becomes $$ \int_{s_i - \log D}^{s_i - \log\log D} a_i(s) \frac{ds}{s_i-s}.$$ However, from Fubini's theorem and \eqref{ar} we have \begin{align*} \int_{s_i - \log D}^{s_i - \log\log D} a_i(s) \frac{ds}{s_i-s} &= \int_{s_i - \log D}^{s_i - \log\log D} a_i(s) (\int_{s_i - \log D}^s \frac{dt}{(s_i-t)^2} + \frac{1}{\log D})\ ds \\ &= \int_{s_i-\log D}^{s_i - \log\log D} (\int_t^{s_i-\log\log D} a_i(s)\ ds) \frac{dt}{(s_i-t)^2} + \frac{1}{\log D} \int_{s_i-\log D}^{s_i-\log\log D} a_i(s)\ ds \\ &\leq \int_{s_i-\log D}^{s_i - \log\log D} (\int_t^{s_i} a_i(s)\ ds) \frac{dt}{(s_i-t)^2} + \frac{1}{\log D} \int_{s_i-\log D}^{s_i} a_i(s)\ ds \\ &\leq \int_{s_i-\log D}^{s_i - \log\log D} (1 - \delta)^{1/2} (s_i-t) \frac{dt}{(s_i-t)^2} + \frac{1}{\log D} (1-\delta)^{1/2} \log D \\ &= (1-\delta)^{1/2} (\log\log D - \log\log\log D + 1) \end{align*} and the right-hand side is equal to $(1-\delta)^{1/2} \log\log D$ up to errors that can be absorbed into the $O( \frac{\mu}{\eps}) \log\log D$ term. For $\mu$ small enough, this gives a contradiction when compared with \eqref{eq27}, proving Corollary \ref{elliott-1}(i). We are left with proving part (ii) of Corollary \ref{elliott-1}. Since sets of logarithmic Banach density zero automatically have logarithmic density zero, we already know from Corollary \ref{elliott-1}(i) that for each tuple $(h_1,\dots,h_k)$ of integers and every $m \geq 1$, there is a set ${\mathcal X}_{h_1,\dots,h_k,m}$ of logarithmic density zero such that $$ | \E_{n \leq X} g_1(n+h_1) \cdots g_k(n+h_k) |\leq \frac{1}{m}$$ for all $X$ outside of ${\mathcal X}_{h_1,\dots,h_k,m}$. Since the number of tuples $(h_1,\dots,h_k,m)$ is countable, a standard diagonalisation construction then gives a further set ${\mathcal X}_0$, still of logarithmic density zero, such that for each $h_1,\dots,h_k,m$, all but finitely many of the elements of ${\mathcal X}_{h_1,\dots,h_k,m}$ are contained in ${\mathcal X}_0$. For instance, one could remove finitely many elements from ${\mathcal X}_{h_1,\dots,h_k,m}$ to create a subset ${\mathcal X}'_{h_1,\dots,h_k,m}$ with the property that $$ \E_{X \leq Y}^{\log} 1_{{\mathcal X}'_{h_1,\dots,h_k,m}}(X) \leq 2^{-h_1-\cdots-h_k-m}$$ for all $Y \geq 1$, and then take ${\mathcal X}_0$ to be the union of all the ${\mathcal X}'_{h_1,\dots,h_k,m}$, which thus differs from a finite union of these sets by a set of arbitrarily small logarithmic density (and finite unions of the sets ${\mathcal X}'_{h_1,\dots,h_k,m}$ have logarithmic density $0$). By construction one then has $$ \limsup_{X \to \infty; X \not \in {\mathcal X}_0} | \E_{n \leq X} g_1(n+h_1) \cdots g_k(n+h_k) |\leq \frac{1}{m}$$ for all $h_1,\dots,h_k,m$, and the claim follows. \end{proof} \begin{remark}\label{18-rem} An inspection of the above argument shows that one could have replaced the sequence $n \mapsto g_1(n+h_1) \dots g_1(n+h_k)$ by any other bounded sequence $n \mapsto F(n)$ for which the analogue of Theorem \ref{main}(i) holds, or more precisely that $$ \E^{\log\log}_{d \in \N} |\lim^*_{X \to \infty} \E_{n \leq X/d} F(n)| = 0$$ for any generalised limit $\lim^*_{X \to \infty}$. \end{remark} Next we prove Corollary \ref{elliott-2}. \begin{proof}[Proof of Corollary \ref{elliott-2}] By Corollary \ref{elliott-1}, we are done unless $g_1 g_2$ weakly pretends to be a twisted Dirichlet character $n \mapsto \chi(n)n^{it}$, so suppose that this is indeed the case for some $\chi$ and $t$. Then for any generalised limit $\lim^*_{X \to \infty}$, the corresponding correlations $f_d(a)$ defined by \eqref{fda-def} obey the property \eqref{hd} for some function $f: \Z \to \Disk$. If this function $f$ was vanishing at $a=1$ for every choice of the generalised limit, then one could repeat the proof of Corollary \ref{elliott-1} to obtain the claim (cf. Remark \ref{18-rem}). Thus suppose instead that we can find a generalised limit $\lim^*_{X \to \infty}$ such that $f(1) \neq 0$ for the function $f$ provided by Theorem \ref{main}(ii). By \eqref{hd} and the triangle inequality, this implies that $$ \E^{\log\log}_{d \in \N} f_d(1) d^{it} = f(1) \neq 0.$$ In particular, for $D$ sufficiently large, one has $$ |\E_{d \leq D}^{\log\log} f_d(1) d^{it} | \gg 1$$ and hence by summation by parts we have $$ |\E_{d \leq D}^{\log} f_d(1) d^{it} | \gg 1$$ for a sequence of arbitrarily large $D$. If $D$ obeys the above estimate, then by \eqref{fda-def} we have $$ |\lim^*_{X \to \infty} \E_{d \leq D}^{\log} d^{it} \E_{n \leq X/d} g_1(n+h_1) g_2(n+h_2)| \gg 1$$ and thus there exist arbitrarily large $X$ such that $$ |\E_{d \leq D}^{\log} d^{it} \E_{n \leq X/d} g_1(n+h_1) g_2(n+h_2)| \gg 1.$$ This implies that $$ |\E_{d \leq D}^{\log} d^{it} \E_{cX/d \leq n \leq X/d} g_1(n+h_1) g_2(n+h_2)| \gg 1$$ for some small constant $c>0$ (not depending on $D$ and $X$). This yields $$ \bigg|\sum_{\log D\leq d \leq D} d^{it} \sum_{cX/d \leq n \leq X/d} g_1(n+h_1) g_2(n+h_2)\bigg| \gg X \log D$$ The left-hand side can be rearranged (discarding negligible errors, assuming $D$ is large enough) as $$ \bigg|\sum_{cX/D \leq n \leq X/\log D} \bigg(\sum_{cX/n \leq d \leq X/n} d^{it}\bigg) g_1(n+h_1) g_2(n+h_2)\bigg| \gg X \log D.$$ By summation by parts, for $cX/D\leq n \leq X/\log D$ we have \begin{align*} \sum_{cX/n \leq d \leq X/n} d^{it} = \alpha n^{-it} \frac{X}{n} + o_{D \to \infty}(1),\quad \alpha=\frac{X^{it}-(cX)^{it}\cdot c}{1+it}, \end{align*} where in particular the quantity $\alpha$ is bounded and is independent of $n$. For $D$ large enough, we conclude that $$ \bigg|\sum_{cX/D \leq n \leq X/\log D} \frac{n^{-it}}{n} g_1(n+h_1) g_2(n+h_2)\bigg| \gg \log D.$$ and hence $$ |\E_{X/D \leq n \leq X}^{\log} n^{-it} g_1(n+h_1) g_2(n+h_2)| \gg 1.$$ Approximating $n^{-it}$ by $(n+h_1)^{-it}$, we conclude that there exist arbitrarily large $D$ such that $$ |\E_{X/D \leq n \leq X}^{\log} (n+h_1)^{-it} g_1(n+h_1) g_2(n+h_2)| \gg 1$$ for arbitrarily large $X$. But this contradicts the $k=2$ case of the logarithmically averaged Elliott conjecture \cite[Corollary 1.5]{tao} applied to the functions $n \mapsto n^{-it} g_1(n)$ and $n \mapsto g_2(n)$ (note that the hypothesis \eqref{suptx} for $g_1$ implies the same hypothesis for $n \mapsto n^{-it} g_1(n)$). This completes the proof of part (i) of Corollary \ref{elliott-2}. Part (ii) of Corollary \ref{elliott-2} is then deduced from Corollary \ref{elliott-2}(i) using precisely the same diagonalisation argument that was used to deduce Corollary \ref{elliott-1}(ii) from Corollary \ref{elliott-1}(i). \end{proof} \begin{remark} The above argument shows more generally that if the logarithmically averaged Elliott conjecture\footnote{One needs the variant where we sum over $X/\omega(X) \leq n \leq X$ rather than $n \leq X$.} (resp. Chowla conjecture) is proven for a given value of $k$, then the unweighted form of the Elliott conjecture (resp. Chowla conjecture) for that value of $k$ holds at almost all scales. (Note in the case of the Chowla conjecture that the parameter $t$ will vanish, since $\lambda^k=1$ for even $k$ and $\lambda^k=\lambda$ does not pretend to be any twisted Dirichlet character for odd $k$.) \end{remark} \begin{remark}\label{rem1} With small modifications, we can adapt the above proofs to prove Corollary \ref{cor1}. Firstly, by approximating the indicator function $1_{P^{+}(n)<P^{+}(n+1)}$ as in \cite[Section 4]{tera-binary} by a linear combination of indicator functions of the form $1_{P^{+}(n)<n^{\alpha}, P^{+}(n+1)<n^{\beta}}$, we can reduce the proof to showing \begin{align}\label{eq9} \lim_{X\to \infty; X\not \in \mathcal{X}_0}\mathbb{E}_{n\leq X}1_{P^{+}(n)<n^{\alpha}}1_{P^{+}(n+1)<n^{\beta}} =\rho(1/\alpha)\rho(1/\beta), \end{align} where $\rho(\cdot)$ is the Dickmann function and $\alpha, \beta \in (0,1)$ are any rational numbers. Since the set of rationals is countable, by a diagonal argument (as in the proof of Corollary \ref{elliott-2}(ii)) it suffices to prove \eqref{eq9} with $\alpha, \beta$ fixed. One starts by proving a version of the structural theorem (Theorem \ref{main}) in the case of the functions $g_1(n)=1_{P^{+}(n)<n^{\alpha}}$, $g_2(n)=1_{P^{+}(n)<n^{\beta}}$; these are not quite multiplicative functions, but they can be approximated as $1_{P^{+}(n)<n^{\alpha}}=1_{P^{+}(n)<X^{\alpha}}+O(1_{P^{+}(n)\in [(X/\log X)^{\alpha},X^{\alpha}]})$ for $n\in [X/\log X,X]$. The $O(\cdot)$ term has negligible contribution in the entropy decrement argument by standard estimates on smooth numbers, so the proof of Proposition \ref{dollop} goes through for the generalised limits associated to the correlations of $g_1$ and $g_2$ with $G\equiv 1$ (so certainly \eqref{gp-inv} holds). We did not use the specific properties of $g_1,g_2$ anywhere else in the proof of Theorem \ref{main}, so that proof goes through, giving \begin{equation}\label{eq28}\mathbb{E}_{d\in \mathbb{N}}|\lim^{*}_{X\to \infty}\mathbb{E}_{n\leq X/d}g_1(n)g_2(n+1)-c^{*}|=0\end{equation} for all generalised limits $\lim^{*}$ and some constant $c^{*}$ depending on $\lim^{*}$. From \cite[Proof of Corollary 1.19]{tera-binary}, we have a logarithmic version of \eqref{eq9}, so following the proof of Corollary \ref{elliott-2} verbatim, we see that $c^{*}=\rho(1/\alpha)\rho(1/\beta)$. Then from Remark \ref{18-rem} we deduce \eqref{eq9}. We leave the details to the interested reader. \end{remark} \begin{proof}[Proof of Corollary \ref{chow}] We observe from Corollary \ref{elliott-1}(i) (for odd $k$) or Corollary \ref{elliott-2}(ii) (for $k=2$) that for any distinct integers $h_1,\dots,h_k$ and $\eps>0$, one has $$ |\E_{n \leq X} \lambda(n+h_1) \cdots \lambda(n+h_k)| \leq \eps $$ for all $X$ outside of a set $\mathcal{X}_{k,\varepsilon}$ of logarithmic Banach density zero, and hence also of logarithmic density zero. The claim then follows by the same diagonalisation argument used to prove Corollary \ref{elliott-1}(ii) and Corollary \ref{elliott-2}(ii). \end{proof} \section{Consequences of the isotopy formulae}\label{isotopy-sec} Before proving the isotopy formula in the form of Theorem \ref{isotopy}, let us state a variant of it that involves the quantities $f_d(a)$ present in Theorem \ref{main}. In what follows, a sequence $b_n$ of integers is said to be \emph{asymptotically rough} if for any given prime $p$, one has $p \nmid b_n$ for all sufficiently large $n$. For instance, any increasing sequence of primes is asymptotically rough, as is the sequence $-1,-1,-1,\dots$. \begin{lemma}\label{le1} Let the notation and hypotheses be as in Theorem \ref{main}. Let $n \mapsto \chi(n)n^{it}$ be a twisted Dirichlet character that weakly pretends to be $g_1 \cdots g_k$, if one exists; otherwise, choose $\chi$ and $t$ arbitrarily. Let $a$ be an arbitrary integer. \begin{itemize} \item[(i)] (Archimedean isotopy) For any natural number $h$, one has $$ \lim_{X\to \infty}\E^{\log\log}_{d \leq X} |f_{hd}(a) - h^{-it} f_d(a)| = 0.$$ \item[(ii)] (Non-archimedean isotopy) For any asymptotically rough sequence $b_n$ of natural numbers, one has $$ \lim_{n\to \infty}\lim_{X\to \infty}\E^{\log\log}_{d \leq X} |f_{d}(ab_n) - \chi(b_n) f_d(a)|=0.$$ In particular, since the sequence $b_n = -1$ is asymptotically rough, one has \begin{equation}\label{lln} \lim_{X\to \infty}\E^{\log\log}_{d \leq X} |f_{d}(-a) - \chi(-1) f_d(a)| = 0. \end{equation} \end{itemize} \end{lemma} A variant of Lemma \ref{le1}(ii) (for logarithmic averaging, and with $b_n$ specialised to the primes in an arithmetic progression $1\ (q)$ for $q$ a period of $\chi$) was obtained in \cite[Corollary 3.7]{fh-furstenberg}. \begin{proof} We may assume without loss of generality that $g_1 \cdots g_k$ weakly pretends to be $n \mapsto \chi(n)n^{it}$, as the claims follow from Theorem \ref{main}(i) otherwise. Extracting out the contribution to \eqref{hd} from multiples of $h$, we see that $$ \E^{\log\log}_{d \in \N} |f_{hd}(a) - f(a) (hd)^{-it}| = 0,$$ and also by \eqref{hd} we have \begin{align*} \E^{\log\log}_{d \in \N} |f_{d}(a) - f(a) d^{-it}| = 0. \end{align*} Now the claim follows from the triangle inequality. To prove\footnote{Note that Lemma \ref{isotop} does not directly imply Claim (ii), since the Dirichlet character present in that lemma depends on the error $\varepsilon$.} Claim (ii), we observe from \eqref{hd} that \begin{align*} \E^{\log\log}_{d \in \N} |f_{d}(ab_n) - f(ab_n)d^{-it}| =0 \end{align*} for all $n$, and \begin{align*} \E^{\log\log}_{d \in \N} |f_{d}(a) - f(a)d^{-it}| =0. \end{align*} Putting together the above two equalities we have \begin{align}\label{eq11} \E^{\log\log}_{d \in \N} |f_{d}(ab_n) - \chi(b_n)f_d(a)| =|f(ab_n)-\chi(b_n)f(a)| \end{align} By Theorem \ref{main}, $f$ is the uniform limit of $\chi$-isotypic periodic functions $F_i$. For each such $F_i$, we have $F_i(ab_n) = \chi(b_n) F_i(a)$ for all sufficiently large $n$, since the sequence $b_n$ is asymptotically rough. Thus also $f(ab_n)=\chi(b_n)f(a)+o_{n\to \infty}(1).$ Combining this with \eqref{eq11}, the claim follows. \end{proof} We then use Lemma \ref{le1} to deduce the isotopy formulae (Theorem \ref{isotopy}). \begin{proof}[Proof of Theorem \ref{isotopy}] We start with the proof of (i). By a diagonalisation argument, similarly as in the proof of Corollary \ref{elliott-1}(ii), it suffices to show that for any fixed rational $q>0$ there exists a set $\mathcal{X}_{0,q}$ of logarithmic density $0$ such that the claim holds with $\mathcal{X}_{0,q}$ in place of $\mathcal{X}_0$. Next, we argue that it suffices to consider the case $q\in \mathbb{N}$. Suppose that the case $q\in \mathbb{N}$ has been established, and let $q=a/b$ with $a,b\in \mathbb{N}$. Then if $\mathcal{X}_{0,q}:=(1/b)\mathcal{X}_{0,a}\cup (1/b)\mathcal{X}_{0,b}$ (which is still a set of logarithmic density zero), we have \begin{align*} &\lim_{X\to \infty; X\not \in \mathcal{X}_{0,q}}\left(\mathbb{E}_{n\leq X}g_1(n+h_1)\cdots g_k(n+h_k)-(a/b)^{it}\mathbb{E}_{n\leq bX/a}g_1(n+h_1)\cdots g_k(n+h_k)\right)\\ &=\lim_{X\to \infty; X\not \in \mathcal{X}_{0,q}}\left(b^{-it}\mathbb{E}_{n\leq bX}g_1(n+h_1)\cdots g_k(n+h_k)-(a/b)^{it}\mathbb{E}_{n\leq bX/a}g_1(n+h_1)\cdots g_k(n+h_k)\right)=0. \end{align*} Hence we may assume from now on that $q\in \mathbb{N}$. Observe that the statement of Lemma \ref{le1}(i) with $a=1$ can be written in the form \begin{align*} \mathbb{E}_{d\in \mathbb{N}}^{\log \log}|\lim^{*}_{X\to \infty}\mathbb{E}_{n\leq X/d}\big(g_1(n+h_1)\cdots g_k(n+h_k)-q^{-it}\mathbb{E}_{b\in \mathbb{Z}/q\mathbb{Z}}g_1(qn+b+h_1)\cdots g_k(qn+b+h_k)\big)|=0 \end{align*} for every generalised limit $\lim^{*}_{X\to \infty}$. By following the proof of Corollary \ref{elliott-1}(i) verbatim (see also Remark \ref{18-rem}), this leads to \begin{align}\label{eq12} \lim_{X\to \infty;X\not \in \mathcal{X}_{0,q}} \mathbb{E}_{n\leq X}\big(g_1(n+h_1)\cdots g_k(n+h_k)-q^{-it}\mathbb{E}_{b\in \mathbb{Z}/q\mathbb{Z}}g_1(qn+b+h_1)\cdots g_k(qn+b+h_k)\big)=0 \end{align} for some set $\mathcal{X}_{0,q}$ of logarithmic density zero. But rewriting \eqref{eq12}, it becomes the identity asserted in Theorem \ref{isotopy}(i). We turn to the proof of part (ii), which is similar. Again by a diagonalisation argument, it suffices to prove the statement for fixed $a$ rather than all $a$. From Lemma \ref{le1}(ii) we have \begin{align*} \mathbb{E}_{d\in \mathbb{N}}^{\log \log}|\lim^{*}_{X\to \infty}\mathbb{E}_{n\leq X/d}(g_1(n-ah_1)\cdots g_k(n-ah_k)-\chi(-1)g_1(n+ah_1)\cdots g_k(n+ah_k))|=0 \end{align*} for every generalised limit $\lim^{*}_{X\to \infty}$. Just as in the proof of part (i) of the Theorem, by the proof of Corollary \ref{elliott-1}(i) (cf. Remark \ref{18-rem}) we get \begin{align*} \lim_{X\to \infty;X\not \in \mathcal{X}_{0,a}} \mathbb{E}_{n\leq X}(g_1(n-ah_1)\cdots g_k(n-ah_k)-\chi(-1)g_1(n+ah_1)\cdots g_k(n+ah_k))|=0 \end{align*} for some set $\mathcal{X}_{0,a}$ of logarithmic density zero, and this is what we wished to prove. \end{proof} Morally speaking, the Archimedean isotopy formula implies that the argument of the correlation \eqref{eq19} becomes equidistributed at large scales whenever $t\neq 0$. Unfortunately we cannot quite establish this claim as stated, because of the discontinuous nature of the complex argument function. However, if we insert a continuous mollifier to remove this discontinuity, we can obtain equidistribution. More precisely, we have the following result. \begin{theorem}[Equidistribution of argument away from zero] \label{limits} Let $k\geq 1$, let $h_1,\ldots, h_k$ be integers and $g_1,\ldots, g_k:\mathbb{N}\to \mathbb{D}$ be $1$-bounded multiplicative functions. Suppose that the product $g_1\cdots g_k$ weakly pretends to be a twisted Dirichlet character $n\mapsto \chi(n)n^{it}$, where $t\neq 0$. Let us denote \begin{align*} S(X):=\mathbb{E}_{n\leq X}g_1(n+h_1)\cdots g_k(n+h_k). \end{align*} Let $\psi: \C \to \C$ be a continuous function that vanishes in a neighbourhood of the origin, and let $$ \overline{\psi}(z) \coloneqq \frac{1}{2\pi} \int_0^{2\pi} \psi(e^{i\theta} z)\ d\theta$$ be $\psi$ averaged over rotations around the origin. Then we have $$ \E^{\log}_{X \in \N} \psi(S(X)) - \overline{\psi}(S(X)) = 0.$$ \end{theorem} \begin{proof} Since $S$ is bounded, we may assume that $\psi$ is compactly supported. By replacing $\psi$ by $\psi - \overline{\psi}$ we may assume that $\overline{\psi}=0$. Approximating $\psi$ uniformly by partial Fourier series (e.g., using Fej\'er summation) in the angular variable, and using linearity, we may assume that $\psi$ takes the form $\psi(re^{i\theta}) = \Psi(r) e^{ik\theta}$ for some non-zero integer $k$ and some continuous compactly supported function $\Psi$ that vanishes in a neighbourhood of the origin (cf. the standard proof of the Weyl equidistribution criterion \cite{weyl}). In particular we have the isotopy formula \begin{equation}\label{psi-iso} \psi(\omega z) = \omega^k \psi(z) \end{equation} for all $z \in \C$ and $\omega \in S^1$. Let $q > 1$ be an integer to be chosen later. From Theorem \ref{isotopy}(i), outside of an exceptional set $\mathcal{X}_0$ of logarithmic density zero, we have $$ \lim_{X \to \infty; X \not \in {\mathcal X}_0} S(X) - q^{it} S(X/q) = 0.$$ From \eqref{psi-iso} and the uniform continuity of $\psi$, this implies that $$ \lim_{X \to \infty; X \not \in {\mathcal X}_0} \psi(S(X)) - q^{ikt} \psi(S(X/q)) = 0.$$ Taking logarithmic averages, we conclude that $$ \E^{\log}_{X \in \N} \psi(S(X)) - q^{ikt} \psi(S(X/q)) = 0.$$ On the other hand, in analogy to \eqref{fd1d2}, we have the log-Lipschitz bound \begin{align}\label{eq13} |S(x)-S(y)|\ll |\log x-\log y|. \end{align} We can use this and the uniform continuity of $\psi$ to estimate, for $X_0$ large enough, \begin{align*} \mathbb{E}_{X\leq X_0}^{\log}\psi(S(X/q))&=\mathbb{E}_{X\leq X_0/q}^{\log}\mathbb{E}_{0\leq b<q} \psi(S(X+b/q))+o(1)\\ &=\mathbb{E}_{X\leq X_0/q}^{\log} \psi(S(X))+o(1)\\ &=\mathbb{E}_{X\leq X_0}^{\log} \psi(S(X))+o(1). \end{align*} Hence $$ \E^{\log}_{X \in \N} \psi(S(X)) - \psi(S(X/q)) = 0.$$ By the triangle inequality, we conclude that $$ (1 - q^{ikt}) \E^{\log}_{X \in \N} \psi(S(X)) = 0.$$ Since $t \neq 0$, we can select $q$ so that $q^{ikt} \neq 1$ for all $k\in \mathbb{N}$. The claim follows. \end{proof} Suppose that $\Psi: [0,+\infty) \to [0,+\infty)$ is a non-negative continuous function vanishing near the origin, and let $I \subset \R/2\pi \Z$ be an arc in the unit circle $\R/2\pi \Z$. Applying Theorem \ref{limits} to upper and lower approximants to the discontinuous function $z \mapsto \Psi(|z|) 1_I(\mathrm{arg}(z))$, and taking limits, we conclude that $$ \E^{\log}_{X \in \N} (1_I(\mathrm{arg}(S(X))) - \frac{|I|}{2\pi}) \Psi(|S(X)|) = 0$$ where $|I|$ denotes the length of $I$. Informally, this asserts that the argument $\mathrm{arg}(S(X))$ is uniformly distributed in the unit circle, so long as one inserts a continuous weight of the form $\Psi(|S(X)|)$. It would be more aesthetically pleasing if we could replace this weight with a discontinuous cutoff such as $1_{|S(X)| \geq \eps}$, but we were unable to exclude the possibility that $|S(X)|$ lingers very close to $\eps$ for very many scales $X$, with the event that $|S(X)| \geq \eps$ being coupled in some arbitrary fashion to the argument of $S(X)$, leading to essentially no control on the argument of $S(X)$ restricted to the event $|S(X)| \geq \eps$. On the other hand, if one was able to show that $S(X)$ did not concentrate at the origin in the sense that $$ \limsup_{X_0 \to \infty} \E^{\log}_{X \leq X_0} 1_{|S(X)| \leq \eps} \to 0$$ as $\eps \to 0$, then the above arguments do show that $$ \E^{\log}_{X \in \N} 1_I(\mathrm{arg}(S(X))) = \frac{|I|}{2\pi}$$ for all intervals $I$, so that $\mathrm{arg}(S(X))$ is indeed asymptotically equidistributed on the unit circle. Alternatively, by selecting the cutoff $\eps$ using the pigeonhole principle to ensure that $|S(X)|$ does not linger too often in a neighbourhood of $\eps$, one can prove statements such as the following: If $\delta > 0$, then for all sufficiently large $X_0$ outside of a set of logarithmic density zero, one can find $0 < \eps \leq \delta$ with the approximate equidistribution property $$ \E^{\log}_{X \leq X_0} \left(1_I(\mathrm{arg}(S(X))) - \frac{|I|}{2\pi}\right) 1_{|S(X)| \geq \eps} \leq \delta$$ for all intervals $I \subset \R/2\pi \Z$. We leave the proof of this assertion to the interested reader. Now we investigate the consequences of the non-archimedean isotopy formula (Theorem \ref{isotopy}(ii)). Many of these consequences tell us that the correlation \eqref{eq19} tends to $0$ along almost all scales also in some cases that are not covered by Corollary \ref{elliott-1}(i). \begin{definition} We say that a tuple $(g_1,\ldots, g_k)$ of functions is \emph{reflection symmetric} if $g_i=g_{k+1-i}$ for all $1\leq i\leq \frac{k+1}{2}$. Similarly, we say that a tuple $(h_1,\ldots, h_k)$ of integers is \emph{progression-like} if $h_1+h_{k}=h_i+h_{k+1-i}$ for all $1\leq i\leq \frac{k+1}{2}$. In particular, all arithmetic progressions are progression-like. \end{definition} \begin{theorem} \label{even} Let $k\geq 1$ and let $h_1,\ldots, h_k$ be integers. Suppose that $\chi$ is an odd Dirichlet character (i.e., $\chi(-1)=-1$) with $\chi(n+h_1+h_k)=\chi(n)$ for all $n$. Let $g_1,\ldots, g_k:\mathbb{N}\to \mathbb{D}$ be multiplicative functions such that the product $g_1\cdots g_k$ weakly pretends to be a Dirichlet character $\psi$ with $\psi$ even. Suppose additionally that the tuple $(g_1,\ldots, g_k)$ is reflection symmetric and that the tuple $(h_1,\ldots, h_k)$ is progression-like. Then there exists an exceptional set $\mathcal{X}_0$ of logarithmic density $0$, such that \begin{align*} \lim_{X\to \infty; X\not \in \mathcal{X}_0}\mathbb{E}_{n\leq X} \chi(n)g_1(n+h_1)g_2(n+h_2)\cdots g_k(n+h_k)=0. \end{align*} \end{theorem} \begin{proof} Note that the function $g_1\cdots g_k\chi$ weakly pretends to be $\psi \chi$, which is an odd character. Hence by Theorem \ref{isotopy}(ii) there exists some set $\mathcal{X}_0$ of logarithmic density $0$, such that for $X\not \in \mathcal{X}_0$ we have \begin{align*} &\mathbb{E}_{n\leq X} \chi(n)g(n)g_1(n+h_1)\cdots g_k(n+h_k)\\ &=-\mathbb{E}_{n\leq X} \chi(n)g_1(n-h_1)g(n-h_2)\cdots g(n-h_k)+o(1). \end{align*} By translation invariance, the periodicity assumption $\chi(n+h_1+h_k)=\chi(n)$, and the progression-likeness of $(h_1,\ldots, h_k)$, the latter expression equals \begin{align*} &-\mathbb{E}_{n\leq X} \chi(n+h_1+h_k)g_1(n+h_k)g_2(n+h_1+h_k-h_2)\cdots g_k(n+h_1)+o(1)\\ &=-\mathbb{E}_{n\leq X}\chi(n)g_1(n+h_k)g_2(n+h_{k-1})\cdots g_k(n+h_1)+o(1). \end{align*} Since the tuple $(g_1,\ldots, g_k)$ is reflection symmetric, this equals the the original correlation with a minus sign, proving the statement. \end{proof} Corollary \ref{even-chowla} is an immediate consequence of Theorem \ref{even}. \begin{proof}[Proof of Corollary \ref{even-chowla}] Taking $g_1=\ldots=g_k=\lambda$ and $(h_1,\ldots, h_k)=(0,a,\ldots, (k-1)a)$ in Theorem \ref{even}, we readily obtain the claim. \end{proof} In other words, the shifted products of the Liouville function can be shown to be orthogonal to some suitable Dirichlet characters also when there is an even number of shifts. As already mentioned, also the weaker, logarithmic version of Corollary \ref{even-chowla} is new. The next theorem is in the same spirit as Theorem \ref{even}, but with somewhat different conditions. \begin{theorem} \label{odd} Let $k\geq 1$ be odd, and let $g_1,\ldots, g_k:\mathbb{N}\to \mathbb{D}$ be multiplicative functions such that the product $g_1\cdots g_k$ weakly pretends to be a Dirichlet character $\chi$ with $\chi$ odd. Suppose also that the tuple $(g_1,\ldots, g_k)$ is reflection symmetric and that $(h_1,\ldots, h_k)$ is a progression-like tuple of integers. Then there exists an exceptional set $\mathcal{X}_0$ of logarithmic density $0$, such that \begin{align*} \lim_{X\to \infty; X\not \in \mathcal{X}_0}\mathbb{E}_{n\leq X} g_1(n+h_1)g_2(n+h_2)\cdots g_k(n+h_k)=0. \end{align*} \end{theorem} \begin{proof} As with Theorem \ref{even}, this follows directly from the isotopy formula (Theorem \ref{isotopy}) and translation invariance. \end{proof} This theorem can for example be applied to the variants \begin{align*} \lambda_q(n):=e\left(\frac{2\pi i\Omega(n)}{q}\right) \end{align*} of the Liouville function that take values in the $q$th roots of unity. Here $\Omega(n)$ is the number of prime factors of $n$ with multiplicities. We obtain the following. \begin{corollary}\label{cor-1} Let $k\geq 1$ be odd, $q\in \mathbb{N}$, and let $\chi$ be an odd Dirichlet character. Then there exists an exceptional set $\mathcal{X}_0$ of logarithmic density $0$, such that \begin{align*} \lim_{X\to \infty; X\not \in \mathcal{X}_0}\mathbb{E}_{n\leq X} \lambda_q(n)\chi(n)\lambda_q(n+a)\chi(n+a)\cdots \lambda_q(n+(k-1)a)\chi(n+(k-1)a)=0. \end{align*} \end{corollary} \begin{proof} We apply Theorem \ref{odd} with $g_j(n)=g(n):=\chi(n)\lambda_q(n)$ and $(h_1,\ldots, h_k)=(0,\ldots, (k-1)a)$. Then if $q\nmid k$, the function $g^k\chi^k$ does not weakly pretend to be any twisted Dirichlet character, since $g^k$ does not do so. In this case, we may appeal to Corollary \ref{elliott-1}(i) to obtain the claim. Suppose then that $q\mid k$. Then $g^k$ weakly pretends to be $\chi^k$, which is an odd character, so Theorem \ref{odd} is applicable. \end{proof} \begin{example} Let $\chi_3$ be the odd Dirichlet character of modulus $3$ and $\chi_8$ any odd Dirichlet character of modulus $8$. Then from Corollary \ref{cor-1} and partial summation, for any sequences $1\leq \omega_m\leq x_m$ of reals tending to infinity we have \begin{align*} \lim_{m\to \infty}\mathbb{E}^{\log}_{x_m/\omega_m\leq n\leq x_m}\chi_3(n)\lambda_3(n)\lambda_3(n+3)\lambda_3(n+6)=0 \end{align*} and \begin{align*} \lim_{m\to \infty}\mathbb{E}^{\log}_{x_m/\omega_m\leq n\leq x_m}\lambda_3(n)\lambda_3(n+2)\lambda_3(n+4)\chi_8(n+6)=0. \end{align*} This is seen by applying the corollary to the functions $g_j(n)=\lambda_3(n)\chi_3(n)$ with $a=3$ and $g_j(n)=\lambda_3(n)\chi_8(n)$ with $a=2$ and using $n(n+2)(n+4)\equiv n+6 \pmod{8}$ for $n$ odd. \end{example} We then turn to bounding more general correlations of multiplicative functions where the shifts involved no longer form a progression-like tuple. In the case of triple correlations, we obtain savings that are explicit but nevertheless far from the desired $o(1)$ bound. \begin{theorem}[Savings in logarithmic three-point Elliott conjecture] Let $g:\mathbb{N}\to \mathbb{D}$ be a multiplicative function, and let $h_1,h_2,h_3$ be distinct integers. Suppose that $g$ is non-pretentious in the sense that \begin{align*} \liminf_{X\to \infty}\inf_{|t|\leq X}\mathbb{D}(g,n\mapsto\chi(n)n^{it},x)=\infty \end{align*} for every Dirichlet character $\chi$. Then for any sequences $1\leq \omega_m\leq x_m$ tending to infinity we have \begin{align}\label{eqtn6} \limsup_{m\to \infty}|\mathbb{E}^{\log}_{x_m/\omega_m\leq n\leq x_m} g(n+h_1)g(n+h_2)g(n+h_3)|\leq \frac{1}{\sqrt{2}}. \end{align} \end{theorem} \begin{remark} This looks superficially similar to \cite[Lemma 5.3]{klurman-mangerel} (and also to \cite[Proposition 7.1]{tt}, which achieves the better upper bound of $\frac{1}{2}$ for real-valued multiplicative functions). However, importantly, the shifts $h_i$ are allowed to be arbitrary here, while in the aforementioned results they had to form an arithmetic progression for the method to work. \end{remark} \begin{proof} If $h_1,h_2,h_3$ is an arithmetic progression, we may apply \cite[Lemma 5.3]{klurman-mangerel}, so we may henceforth suppose that $h_1,h_2,h_3$ is not an arithmetic progression. If the function $g^3$ does not weakly pretend to be any Dirichlet character, we get the bound $0$ for the $\limsup$ by \cite[Theorem 1.2(ii)]{tt}. Suppose then that $g^3$ weakly pretends to be some character $\chi$. By the isotopy formula (Theorem \ref{isotopy}), partial summation and translation invariance, we have \begin{align}\label{eqtn5} \begin{aligned} &\mathbb{E}^{\log}_{x_m/\omega_m\leq n\leq x_m}g(n+h_1)g(n+h_2)g(n+h_3)\\ &=\chi(-1)\mathbb{E}^{\log}_{x_m/\omega_m\leq n\leq x_m}g(n+h_1)g(n+h_1+h_3-h_2)g(n+h_3)+o_{m\to \infty}(1). \end{aligned} \end{align} In particular, the first part of \eqref{eqtn5} is the average of both parts of the equation. Hence, the average on the left-hand side of \eqref{eqtn6} is up to $o_{m\to \infty}(1)$ bounded by \begin{align*} &\frac{1}{2}|\mathbb{E}^{\log}_{x_m/\omega_m\leq n\leq x_m}g(n+h_1)g(n+h_3)(g(n+h_2)+\chi(-1)g(n+h_1+h_3-h_2))|\\ &\leq \frac{1}{2}\mathbb{E}^{\log}_{x_m/\omega_m\leq n\leq x_m} |g(n+h_2)+\chi(-1)g(n+h_1+h_3-h_2))|. \end{align*} By the Cauchy-Schwarz inequality, this is bounded by \begin{align*} &\frac{1}{2}(\mathbb{E}^{\log}_{x_m/\omega_m\leq n\leq x_m} |g(n+h_2)+\chi(-1)g(n+h_1+h_3-h_2))|^2)^{\frac{1}{2}}\\ &\leq\frac{1}{2}(\mathbb{E}^{\log}_{x_m/\omega_m\leq n\leq x_m}(2+2\chi(-1)\textnormal{Re}(g(n+h_2)\overline{g}(n+h_1+h_3-h_2))))^{\frac{1}{2}}. \end{align*} Since $h_2\neq h_1+h_3-h_2$ by assumption, we can apply \cite[Corollary 1.5]{tao} to see that the term involving real parts contributes $o_{m\to \infty}(1)$. Then we indeed get a bound of $\frac{1}{\sqrt{2}}+o_{m\to \infty}(1)$ for the correlation. \end{proof} \begin{remark} For specific multiplicative functions one can do slightly better by not applying Cauchy-Schwarz. For example, in the case $g(n)=\lambda_3(n)$ one gets a bound of $\frac{2}{3}$ for the correlation by using the fact (following from \cite[Corollary 1.5]{tao}) that $(\lambda_3(n),\lambda_3(n+h))$ takes for fixed $h\neq 0$ each of the possible $9$ values with logarithmic density $1/9$. \end{remark} \section{The case of few sign patterns}\label{sign-sec} In this section we prove Theorem \ref{sign-thm}. Assume the hypotheses of that theorem. Let $h$ be a natural number. By the Hahn-Banach theorem, it suffices to show that $$ \E^*_{n \in \N} \lambda(n) \lambda(n+h) = 0$$ for every generalised limit $\lim^*_{X \to \infty}$. Accordingly, let us fix such a limit. As usual, we introduce the correlation sequences \begin{equation}\label{fda-def-again} f_d(a) \coloneqq \lim^*_{X \to \infty} \E_{n \leq X/d} \lambda(n) \lambda(n+ah) \end{equation} for every real $d>0$. Our task is now to show that $$ f_1(1) = 0.$$ Proposition \ref{dollop} (noting that $G(p) = 1$ in our case) establishes the approximate isotopy formula $$ \sup_{d >0} \E^{\log}_{p \leq P} |f_{dp}(a) - f_d(ap)| \leq \eps$$ whenever $\eps>0$ and $P$ is sufficiently large depending on $\eps$. But because of our hypothesis of few sign patterns, we can obtain a stronger result in which the logarithmic weighting on the averages is removed. \begin{proposition}[Improved approximate isotopy formula] Let $f_d(a)$ be as in \eqref{fda-def-again}, let $\eps>0$, and let $a$ be a natural number. Assume the hypotheses of Theorem \ref{sign-thm}. Then there exist arbitrarily large $m$ such that $$ \sup_{d > 0} \E_{2^m \leq p < 2^{m+1}} |f_{dp}(a) - f_d(ap)| \leq \eps.$$ \end{proposition} This formula also applies for negative $a$, but in this argument we only require the case of positive $a$ (in fact, for the binary correlations considered here, we only need the case $a=1$). \begin{proof} This will be a modification of the arguments in \cite[\S 3]{tt}, and we freely use the notation from that paper. Let $d>0$ be real, let $a$ be an integer, and let $m$ be a large integer to be chosen later. We allow implied constants to depend on $h,a$, but they will remain uniform in $d,m,\eps$. From \eqref{fdag} we have the formula $$ f_{dp}(a) - f_d(ap) = \E^{(d)} \mathbf{g}^{(d)}(0) \mathbf{g}^{(d)}(aph) (p 1_{p|\mathbf{n}^{(d)}}-1) + O(\eps) $$ for all $2^m \leq p < 2^{m+1}$, if $m$ is sufficiently large depending on $\eps$, and where $\mathbf{g}^{(d)} = \mathbf{g}^{(d)}_1 =\mathbf{g}^{(d)}_2$ and $\mathbf{n}^{(d)}$ are the random variables provided by Proposition \ref{corr} (with $g_1=g_2=\lambda$). We can thus write the expression $$ \E_{2^m \leq p < 2^{m+1}} |f_{dp}(a) - f_d(ap)| $$ as $$ \E^{(d)} \E_{2^m \leq p < 2^{m+1}} c_p \mathbf{g}^{(d)}(0) \mathbf{g}^{(d)}(aph) (p 1_{p|\mathbf{n}^{(d)}}-1) + O(\eps)$$ for some sequence of complex numbers $c_p$ with $|c_p| \leq 1$. By stationarity we can also write this expression as $$ \E^{(d)} \E_{1 \leq l \leq 2^m} \E_{2^m \leq p < 2^{m+1}} c_p \mathbf{g}^{(d)}(l) \mathbf{g}^{(d)}(l+aph) (p 1_{\mathbf{n}^{(d)} = -l\ (p)}-1) + O(\eps)$$ and thus $$ \E_{2^m \leq p < 2^{m+1}} |f_{dp}(a) - f_d(ap)| = \E^{(d)} F( \mathbf{X}^{(d)}, \mathbf{Y}^{(d)} )$$ where $\mathbf{X}^{(d)} = \mathbf{X}^{(d)}_m \in \{-1,+1\}^{(2ah+1) 2^m}$, $\mathbf{Y}^{(d)} = \mathbf{Y}^{(d)}_m \in \prod_{2^m \leq p < 2^{m+1}} \Z/p\Z$ are the random variables \begin{align*} \mathbf{X}^{(d)} &\coloneqq \left( \mathbf{g}^{(d)}(l) \right)_{1 \leq l \leq (2ah+1) 2^m} \\ \mathbf{Y}^{(d)} &\coloneqq ( \mathbf{n}^{(d)}\ (p))_{2^m \leq p < 2^{m+1}} \end{align*} and $F: \{-1,+1\}^{(2ah+1) 2^m} \times \prod_{2^m \leq p < 2^{m+1}} \Z/p\Z \to \C$ is the function \begin{align*} & F( (g_l)_{1 \leq l \leq (2ah+1)2^m}, (n_p)_{2^m \leq p < 2^{m+1}})\\ &\quad \coloneqq \E_{1 \leq l \leq 2^m} \E_{2^m \leq p < 2^{m+1}} c_p g_l g_{l+aph} (p 1_{n_p=-l\ (p)} - 1). \end{align*} Repeating the arguments in \cite[\S 3]{tt} verbatim (but without the additional conditioning on the $\mathbf{Y}_{<m}$ random variable), we conclude that $$ \E^{(d)} |F( \mathbf{X}^{(d)}, \mathbf{Y}^{(d)} )| \leq \eps $$ unless we have the mutual information bound $$ \mathbf{I}( \mathbf{X}^{(d)} : \mathbf{Y}^{(d)} ) > \eps^5 \frac{2^m}{m}.$$ At this point we deviate from the arguments in \cite[\S 3]{tt} by using the trivial bound $$ \mathbf{I}( \mathbf{X}^{(d)} : \mathbf{Y}^{(d)} ) \leq \mathbf{H}( \mathbf{X}^{(d)} )$$ to conclude that we will have the desired bound $$ \E_{2^m \leq p < 2^{m+1}} |f_{dp}(a) - f_d(ap)| \leq \eps $$ whenever $\mathbf{X}^{(d)}$ obeys the entropy bound $$ \mathbf{H}( \mathbf{X}^{(d)} ) \leq \eps^5 \frac{2^m}{m}.$$ By Jensen's inequality, this bound will hold if $\mathbf{X}^d$ attains at most $\exp( \eps^5 \frac{2^m}{m} )$ values with positive probability. Using the correspondence principle (Proposition \ref{corr}), this claim in turn is equivalent to the number of possible sign patterns $(\lambda(n+l))_{1 \leq l \leq (2ah+1) 2^m}$ not exceeding $\exp( \eps^5 \frac{2^m}{m} )$; note that this assertion does not depend on $d$, so we in fact obtain the uniform bound $$ \sup_{d>0} \E_{2^m \leq p < 2^{m+1}} |f_{dp}(a) - f_d(ap)| \leq \eps $$ in this case. But by the hypothesis of Theorem \ref{sign-thm}, this assertion holds for arbitrarily large values of $m$. \end{proof} Now we establish Theorem \ref{sign-thm}. By the above proposition, for any $\eps>0$, there exist arbitrarily large $m$ such that $$ f_1(1) = \E_{2^m \leq p < 2^{m+1}} f_{P}(p) + O( \eps ),$$ where $P:=2^m$. By \eqref{fda-def-again}, it suffices to show that $$ \limsup_{X \to \infty} |\E_{P \leq p<2P} \E_{n \leq X/P} \lambda(n) \lambda(n+ph)| \ll \eps$$ for sufficiently large $P$. But this follows from the results in \cite[\S 3]{tao}, specifically Lemmas 3.6, 3.7 and equation (2.9) of that paper\footnote{When applying these results, note that (in the notation of \cite[Lemma 3.6]{tao}) the length of the sum over $j\in [1,H]$ and the range $\mathcal{P}_H$ of primes $p$ do not need to be controlled by the same parameter $H$.} (see also Remark 3.8 for a simplification in the case of the Liouville function). We remark that the equation \cite[(2.8)]{tao} relies crucially on the Matom\"aki-Radziwi{\l}{\l} theorem \cite{mr} (as applied in \cite{mrt}). \begin{remark} A similar argument also gives the odd order cases of the Chowla conjecture if one strengthens the hypothesis of Theorem \ref{sign-thm} to hold for \emph{all} sufficiently large $K$, rather than for arbitrarily large $K$, by using the arguments in \cite[\S 3]{tt-odd} (but with the exceptional sets ${\mathcal M}_1$ in those arguments now being empty, and using unweighted averaging in $n$ rather than logarithmic averaging). We leave the details to the interested reader. \end{remark}

a674d388c420af3c63faac7cde8945886fcf051a

\section{Introduction} A full understanding of the dynamics in liquids is still a pending challenge in chemical physics research\cite{Berthier11,Ediger12}. A better understanding of the dynamics has implications not only for small organic molecular liquids as will be presented in this paper, but also for other materials such as polymers, proteins and metals\cite{Arbe12,Khodadadi15,Luo17}. The dynamics in liquids spans many orders of timescales from the equilibrium liquid above the melting temperature, where relaxation times are shorter than nanoseconds, to the viscous slowing down in the supercooled state as the glass transition is approached and the $\alpha$-relaxation reaches \SI{100}{\s}. Slowing down of the liquid dynamics and the glass transition are traditionally observed by cooling a liquid, but they are also seen upon compression. New insight into liquid dynamics has been gained in the past decades by introducing pressure as an additional variable to temperature \cite{Roland05}. T{\"o}lle et al. \cite{Tolle01} observed isochronal superposition of neutron spectra on picosecond timescales. Isochronal superposition is the observation of an invariance of the spectral shape and linewidth along lines in the temperature-pressure phase diagram of constant relaxation time\cite{Frick03_IS}. The observation of isochronal superposition implies that the spectral shape is determined by the relaxation time or that they are controlled by the same underlying mechanism\cite{Ngai05}. Density scaling is another experimental observation from pressure experiments, based on the ideas from the soft repulsive potential\cite{Tolle01,Dreyfus}. It was shown for several systems that the viscosity or the relaxation time as a function of temperature and pressure could be brought to collapse when expressed in terms of the parameter $\Gamma=\rho^\gamma/T$ (Refs.~\onlinecite{Casalini04_DS,Alba-Simionesco04,Roland05}), where $\gamma$ is a material specific constant. Density scaling has since been shown to apply to numerous different systems and implies that in these systems the relaxation time is governed by the parameter $\Gamma$. A complete understanding of the dynamics in liquids must encompass pressure behaviour, and therefore also the observations of density scaling and isochronal superposition from pressure experiments. One proposed approach is to assume the existence of isomorphs as suggested by isomorph theory \cite{Gnan09,Dyre14}. The basic idea is that two thermodynamic state points are isomorphic if they have the same potential energy landscape upon scaling. This gives rise to curves in the phase diagram along which dynamics on all timescales presented in reduced dimensionless units are invariant. The dynamics in this view is on all timescales determined by the same control parameter, $\Gamma$, which in the simplest case is given by $\Gamma=\rho^\gamma/T$. Density scaling and isochronal superposition are therefore natural consequences of isomorphs. A conjecture of isomorph theory is that it only works for $R$-simple systems, i.e. systems that in molecular dynamics simulations show a strong correlation between the potential energy and the virial \cite{Bailey08I,Pedersen10,Ingebrigtsen12a,Dyre06}. From an experimental point of view, the conjecture is that the idea of $R$-simple systems is manifested in systems without directional bonding or competing interactions. Moreover, isomorph theory only describes inter-molecular interactions and to be certain to exclude intra-molecular modes we consider experimental candidates for simple liquids to be those with no large secondary relaxations. From dielectric spectroscopy, it has previously been shown that isochronal superposition works better for van der Waals liquids than for hydrogen-bonding liquids \cite{Roed13}, although several other studies also with dielectrics\cite{Hensel04,Grzybowska06,Adrjanowicz16,Puosi16} and molecular dynamics simulations \cite{Romanini17} have shown that density scaling and isochronal superposition apply to hydrogen-bonding systems as well. In a recent paper, we showed that the picosecond dynamics along the glass transition as a function of temperature and pressure, i.e. fast relaxations and vibrations that are completely separated from the $\alpha$-relaxation, measured with neutron spectroscopy were invariant for two simple van der Waals liquids\cite{hwh18}. For these simple liquids, we observed isochronal superposition of the dynamics separated by 14 orders of magnitude, which was interpreted as a genuine signature of these liquids having isomorphs. The implication of this finding is that for simple liquids with isomorphs, the phase diagram can be reduced to just one dimension with the control parameter $\Gamma$, in contrast to an experimentalist's normal temperature-pressure phase diagram. A hydrogen-bonding liquid with directional bonding was chosen to test a more complex system in contrast to the simple van der Waals liquids. We showed how the relation between picosecond dynamics and the glass transition broke down for the hydrogen-bonding complex liquid, in agreement with the predictions from isomorph theory. Most previous experimental studies of glass-forming liquids, also the ones at short timescales as probed for example by neutron spectroscopy, take place in the viscous or supercooled liquid, thus, in the proximity of the glass transition. In this paper, we focus on the dynamics in the liquid state with fast $\alpha$-relaxation times. We expand the test of density scaling and isochronal superposition to experimental timescales in the equilibrium liquid when the $\alpha$-relaxation is on the same timescale as neutron spectroscopy. Timescales that also coincides with those accessible with computer simulations. We apply for the first time systematic inelastic fixed window pressure studies from the backscattering instrument IN16B at the Institut Laue-Langevin and use those to test density scaling. Most of the data presented in this paper have been measured in a newly developed pressure cell for performing simultaneous dielectric and neutron spectroscopy as a function of temperature and pressure \cite{Sanz18}, thereby expanding the accessible neutron scattering to include longer timescales, while assuring that the data are collected under exactly the same experimental conditions. In this paper, we present scaling of the $\alpha$-relaxation on pico- to millisecond timescales for three different systems, two simple van der Waals liquids and a more complex hydrogen-bonding sample, studied with neutron and dielectric spectroscopy. We begin by introducing experimental details in the next section. Data are presented in the Sec.~III containing three subsections, one for each sample, followed by the discussion in Sec.~IV. \section{Experimental details} The three samples presented in this paper are two van der Waals liquids, isopropyl benzene (cumene) and 5-polyphenyl ether (PPE), and the hydrogen-bonding liquid dipropylene glycol (DPG) (Fig.~\ref{fig:exp:molecules}). Cumene and DPG were purchased from Sigma Aldrich and PPE from Santolubes, and all three samples were used as acquired. The two van der Waals liquids are considered simple in the sense that they do not possess pronounced secondary relaxations or intra-molecular motion visible on either subnanosecond timescales in the neutron signal or in the dielectric signal \cite{hwh17}, and have previously been shown to have behaviour in agreement with predictions from isomorph theory \cite{hwh18,Xiao15,Roed13}. The hydrogen-bonding liquid DPG represents a complex system with directional bonding, a clear excess wing visible in the dielectric signal \cite{Casalini04} and methyl-group rotation observable on nanosecond timescale with neutron spectroscopy, and has a high pressure response for a hydrogen-bonding liquid. The molecules are sketched in Fig.~\ref{fig:exp:molecules} and details about the samples can be found in Table~\ref{table:exp:details}. \begin{figure}[htpb!] \centering \includegraphics[width=0.25\columnwidth]{cumene_text} \hspace{1 em} \includegraphics[width=0.6\columnwidth]{DPG_text} \\ \includegraphics[width=0.9\columnwidth]{PPE_text} \caption{Sketches of the molecular structure of the three samples studied in this chapter. DPG and PPE are mixtures of isomers.}\label{fig:exp:molecules} \end{figure} \begin{table} \centering \begin{tabular}{l l l l l l} & $T_g(P_\mathrm{amb})$ \hspace{2 em} & $m$ \hspace{2 em} & $\gamma$ \hspace{2 em} & d$T_g$/d$P$ \hspace{1 em} & Refs.\\ & \hspace{1 em} [K] & & & [K/MPa] & \\ \hline cumene & \hspace{1 em} 126 & 70 & 4.8 & 0.085 & \cite{Ransom17,hwh18}\\ PPE & \hspace{1 em} 245 & 80 & 5.5 & 0.18\phantom{0} & \cite{Hecksher13,Gundermann13_thesis,hwh18} \\ DPG & \hspace{1 em} 195 & 65 & 1.5 & 0.075 & \cite{Grzybowska06,Grzybowski06_gamma,hwh18} \\ \end{tabular} \caption{Details on the three samples: cumene (isopropyl benzene), PPE (5-polyphenyl ether) and DPG (dipropylene glycol). The fragility, $m$, is a measure of the super-Arrhenius behaviour introduced by Angell \cite{Angell91}: $m=\frac{\mathrm{d} \log_{10}\tau(T)}{\mathrm{d}(T_g/T)}\biggr\rvert_{T_g}$.}\label{table:exp:details} \end{table} The data presented in this paper have been measured at the Institut Laue-Langevin (ILL) at the backscattering instrument IN16B and IN13 and the time-of-flight instrument IN5 \cite{Exp-6-05-961,Exp-LTP-6-7}. The timescale of the neutron experiments is determined by the energy resolution and the energy transfer of the neutron instruments; the coarser the energy resolution, the faster the timescale. We measured cumene and DPG on IN16B, accessing nanosecond dynamics with the standard Si(111) setting with $\lambda=\SI{6.271}{\angstrom}$ in the $Q$-range from 0.4 to \SI{1.8}{\per\angstrom} with an energy resolution of $\sim\SI{0.75}{\micro\eV}$ corresponding to nanosecond timescales in the energy range $\pm\SI{30}{\micro\eV}$. The elastic intensity of cumene was also measured at the backscattering instrument IN13 with an energy resolution of \SI{8}{\micro\eV} in the $Q$-range $0.2-\SI{4.9}{\per\angstrom}$, i.e. an order of magnitude faster than IN16B. Cumene and PPE were both measured on IN5 with a wavelength of \SI{5}{\angstrom} with an energy resolution of \SI{0.1}{\milli\eV} corresponding to a timescale of $\sim\SI{10}{\pico\s}$, i.e. two orders of magnitude faster than IN16B, in the $Q$-range from $1.2-\SI{1.9}{\per\angstrom}$ binned in steps of $\SI{0.1}{\per\angstrom}$ in the energy range $\pm\SI{2}{\milli\eV}$ . Full spectra of cumene were also measured at IN5 with wavelength \SI{8}{\angstrom} with energy resolution \SI{0.015}{\milli\eV} accessing a timescale of roughly \SI{0.1}{\nano\s} in the $Q$-range from $0.5-\SI{1.2}{\per\angstrom}$ in energy range $\pm\SI{1}{\milli\eV}$. All data reduction, subtraction of background, normalisation to monitor and vanadium, corrections for cell geometry and instrument efficiency, have been done in LAMP\cite{LAMP}. Except for the IN16B data on cumene, all data were measured in a newly developed pressure cell for doing simultaneous dielectric and neutron spectroscopy under high pressure \cite{Sanz18}. The temperature range for the high-pressure cell is $2-\SI{320}{\K}$ and is controlled by the standard orange cryostat available at the ILL. The pressure range is $0.1-\SI{500}{\mega\pascal}$. Both the temperature and pressure range are determined by the aluminium alloy of the high-pressure cell, which is used because of its low background in the neutron signal. This new sample cell is a unique tool for combining techniques, thereby accessing dynamics on different timescales with high precision, making sure that data are collected under exactly the same sample conditions. We utilise fast switching between the elastic and inelastic fixed window scans unique for the backscattering instrument IN16B (Ref.~\onlinecite{Frick12}). These scans offer a fast overview of how a parameter such as temperature or pressure in the full $Q$-range affects the dynamics. The IFWS data presented in this paper were measured with an energy offset at \SI{2}{\micro\eV} and a maximum is observed when the spectral weight at the offset energy is highest as a relaxational process passes through the instrument window, for example caused by an increase of the quasielastic broadening. The nature of the dynamics, whether it is a local process or of translational character, can be determined from the $Q$-dependence. That is, if a relaxation shows change in the $Q$-dependence as a function of temperature or pressure it is a delocalised translational motion, whereas a local process will be independent of $Q$ as a function of pressure or temperature \cite{Bee88}. All spectra are shown summed over $Q$ for better statistics since we generally observe the same trend for all values of $Q$, unless otherwise stated. More data and details on samples and data are available in Ref.~\onlinecite{hwh_thesis}. Equations of state to calculate the density at various state points for cumene is from Ref.~\onlinecite{Ransom17}, for PPE from Ref.~\onlinecite{Gundermann13_thesis}, and for DPG from Ref.~\onlinecite{Grzybowski11}. The scaling exponents used for testing density scaling for the three samples were found in previous publications: Ransom et al. \cite{Ransom17} measured the glass transition temperature of cumene in a large pressure range, up to $\sim\SI{4}{\giga\pascal}$, and found $\gamma=4.8$, and used this to density scale light scattering data\cite{Li95} and viscosity data\cite{Barlow66,Ling68}. The scaling exponent for PPE was found from dielectric data to be $\gamma=5.5$ (Ref.~\onlinecite{Gundermann13_thesis}) and used in scaling of dielectric spectra\cite{Xiao15}. The two van der Waals liquids complement each other in terms of experimental accessibility: cumene has a low glass transition temperature which means that the $\alpha$-relaxation reaches subnanoseconds, probed by our scattering experiments, within the $(T,P)$-range accessible by our sample cell, and furthermore, cumene crystallizes easily in the range between the melting point and the glass transition. PPE, on the other hand, does not crystallize and we can therefore easily study dynamics in the viscous dynamic range, whereas the $\alpha$-relaxation is too slow within the $(T,P)$-limits of our equipment compared to the neutron timescales. The hydrogen-bonding liquid DPG was found to have a scaling exponent of $\gamma=1.5$ (Ref.~\onlinecite{Grzybowski06_gamma}), a quite low value of $\gamma$ compared to the two van der Waals liquids, but typical for hydrogen-bonding liquids \cite{Casalini04_DS,Roland05,Niss06}. This scaling exponent allows for scaling of the $\alpha$-relaxation in the range from microseconds to milliseconds \cite{Grzybowska06}, while it does not provide a satisfactory collapse for dynamics close to the glass transition. As mentioned earlier, according to isomorph theory, the relevant quantities to investigate are in reduced units, which are per definition dimensionless (presented with a tilde). Examples of length and time units, $l_0$ and $t_0$, respectively, used in this paper are: \begin{equation} \begin{aligned} l_0 & = \rho^{-1/3} \\ t_0 & = \rho^{-1/3}(m_p/k_\mathrm{B}T)^{1/2}, \end{aligned} \end{equation} where $m_p$ is the average particle mass and $\rho$ is the number density. For length, wave vector and frequency, respectively, the reduced dimensionless units denoted with a tilde are then given by, \begin{equation}\label{eq:exp:reducedunits} \begin{aligned} \tilde{\mathbf{r}} & = \mathbf{r}/l_0 = \mathbf{r} \rho^{1/3} \\ \tilde{{Q}} & = {Q}~l_0 = Q\rho^{-1/3} \\ \tilde{\omega} & = \omega~t_0 = \omega~\rho^{-1/3}(m_p/k_\mathrm{B}T)^{1/2}. \end{aligned} \end{equation} The frequency unit $\omega$ multiplied by $\hbar$ corresponds to the energy transfer. We assume the average particle mass is constant and set $m_p=k_\mathrm{B}=1$. Effectively, the reduced unit for energy transfer thus becomes \begin{equation}\label{eq:exp:energy_red} \tilde{\omega} = \omega\rho^{-1/3}T^{-1/2}, \end{equation} where $\rho$ is now the volumetric mass density. In the dynamic range investigated in this paper, the change in density is in the percent range, thus, the inverse of the cubic root of the density will only have a small effect on the scaling of the energy transfer. Plotting data in reduced energy units will therefore mainly be affected by temperature changes. The scaling has a visible effect only for higher energy transfer, and we will therefore ignore the reduced units for the low frequency IFWS in the next section. Expressing $Q$ in reduced units will result in changes of around 1\%, which will be within the uncertainty of the data and is therefore also neglected. \section{Results} In this section, we present high-pressure neutron scattering data from the two van der Waals liquids, cumene and PPE, and the hydrogen-bonding liquid, DPG, presented in three subsections according to sample. \subsection{van der Waals liquid -- cumene} First, we present the fixed window scans from IN16B on cumene in Fig.~\ref{fig:res:cumene_fws}. In the top panel, the elastic and inelastic fixed energy window scan (EFWS and IFWS) intensity are shown as a function of pressure for three isotherms. From the EFWS ($\Delta E=\SI{0}{\micro\eV}$), the intensity is observed to increase on compression as the mobility decreases moving towards the glass transition. In the IFWS ($\Delta E=\SI{2}{\micro\eV}$), a maximum is observed in the intensity as a function of pressure as the $\alpha$-relaxation moves through the instrument window upon compression. Density scaling predicts that the control parameter of the dynamics is $\Gamma=\rho^\gamma/T$ and we therefore plot the three isotherms from the FWS on IN16B as a function of $\rho^\gamma/T$ in the bottom panel of Fig.~\ref{fig:res:cumene_fws}, using the scaling exponent $\gamma=4.8$ determined by Ransom et al. \cite{Ransom17}. We observe how the data collapse to just one curve for both the EFWS and the IFWS. The same is observed for the elastic intensity from IN13, also a backscattering instrument, with an energy resolution that corresponds to a timescale one order of magnitude faster than IN16B. We are able to express the nanosecond dynamics along three isotherms in the equilibrium liquid with the $\alpha$-relaxation on nanosecond timescale as a function of the control parameter $\Gamma=\rho^\gamma/T$ using a $\gamma$ value found in the supercooled liquid at the glass transition. The scaling exponent, $\gamma$, is according to isomorph theory allowed to change with temperature and density and to depend on the dynamic range, and it has, of course, itself an uncertainty. With this type of experiment and uncertainty, we are not able to determine whether $\gamma$ is state-point dependent. In fact, the data can be made to collapse using values of $\gamma$ in the range from $4.2-4.8$. This type of analysis does not offer a precise way of determining $\gamma$. However, within the experimental uncertainty that we have for this kind of experiment, e.g. knowing the exact temperature, we note how well the value found by Ransom et al. \cite{Ransom17} in the highly viscous liquid using completely different experimental techniques works here in the equilibrium liquid and therefore use $\gamma=4.8$ throughout this paper for cumene. \begin{figure}[htpb!] \centering \includegraphics[width=0.49\columnwidth]{fws_cumene_efws} \includegraphics[width=0.49\columnwidth]{fws_cumene_ifws} \\ \includegraphics[width=0.49\columnwidth]{fws_cumene_efws_ds} \includegraphics[width=0.49\columnwidth]{fws_cumene_ifws_ds}\\ \caption{EFWS (left) and IFWS (right) from IN16B on cumene summed over $Q$. Intensity of EFWS and IFWS plotted for three isotherms (\SI{200}{\K}, \SI{222}{\K}, \SI{231}{\K}) as a function of pressure (top) and as a function of $\Gamma=\rho^\gamma/T$ with $\gamma=4.8$ (bottom).}\label{fig:res:cumene_fws} \end{figure} \begin{figure}[htpb!] \centering \includegraphics[width=0.49\columnwidth]{cumene_in13} \includegraphics[width=0.49\columnwidth]{cumene_in13_ds} \caption{Elastic intensity of cumene from IN13 summed over $Q$. Left: three isotherms (\SI{200}{\K}, \SI{222}{\K}, \SI{231}{\K}) as a function of pressure. Right: Density scaling of left plotted as a function of $\Gamma=\rho^\gamma/T$, $\gamma=4.8$.}\label{fig:res:cumene_in13} \end{figure} For the interpretation of the fixed window scans, we assume that by knowing the intensity of the elastic peak ($\Delta E=\SI{0}{\micro\eV}$) and the intensity at some offset energy, in this case, $\Delta E=\SI{2}{\micro\eV}$, we have enough information to interpret how the overall spectral shape and linewidth or quasielastic broadening changes with some variable, in this case $T$ and $P$. Therefore, an indirect assumption of isochronal superposition goes into this interpretation, i.e. we assume that the linewidth of the spectral shape changes in a similar way for temperature and pressure. This assumption can be tested by determining an isochrone from the EFWS and IFWS of the three isotherms from IN16B by identifying the same place in intensity on the curves for the three isotherms and investigating the full spectra. In Fig.~\ref{fig:res:cumene_spectra}, we present full spectra from IN16B measured along the isochrone determined from the FWS and an isotherm for comparison. The studied state points are shown in the $(T,P)$-phase diagram. We observe an invariance of the spectral shape and quasielastic broadening along the isochrone. For the isotherm, we observe less elastic intensity and more relaxation in the instrument window on decreasing pressure as the $\alpha$-relaxation time becomes faster on decreased pressure. This means that just one spectrum at a given $(T,P)$ gives the spectral shape and the FWS, which give the temperature and pressure dependence of the width, we are able to predict the spectral shape and linewidth along all state points that are isochronal to that spectrum. \begin{figure}[htpb!] \centering \begin{minipage}{0.3\columnwidth} \vspace{-7 em} {\noindent cumene \\ \vspace{0.5em} IN16B \\ \vspace{0.5em} $\sim\SI{e-9}{\s}$} \end{minipage} \includegraphics[width=0.32\columnwidth]{phasediagram_cu1} \\ \includegraphics[width=0.49\columnwidth]{cumene_in16b_isochrone_red} \includegraphics[width=0.49\columnwidth]{cumene_in16b_isotherm1_red} \caption{Spectra of cumene from IN16B summed over $Q$. The $(T,P)$-phase diagram shows the studied state points for the isochrone as determined from fixed window scan on IN16B and an isotherm at \SI{231}{\K}. The blue dashed line in the spectra corresponds to an energy offset of \SI{2}{\micro\eV} used for the inelastic fixed window scans.}\label{fig:res:cumene_spectra} \end{figure} \begin{figure}[htpb!] \centering \includegraphics[width=0.49\columnwidth]{fft_cu_isochrone_red} \includegraphics[width=0.49\columnwidth]{fft_cu_isotherm2_red} \caption{Fourier transform of combined cumene spectra from IN16B and IN5 with $\lambda=\SI{5}{\angstrom}$ and \SI{8}{\angstrom} for $Q=\SI{1.2}{\per\angstrom}$. Full line is a stretched exponential fit to data from IN16B and IN5 with $\lambda=\SI{5}{\angstrom}$ and \SI{8}{\angstrom}.}\label{fig:res:cumene_fft} \end{figure} At the same state points as presented in Fig.~\ref{fig:res:cumene_spectra} for IN16B, spectra were measured on IN5 at two different resolutions, where we observed the same trend. In Fig.~\ref{fig:res:cumene_fft}, we show the Fourier transformed combined spectra of cumene from IN16B and IN5 with $\lambda=5$ and \SI{8}{\angstrom} at $Q=\SI{1.2}{\per\angstrom}$. The Fourier transformed data have been corrected for resolution by merely dividing by the low-temperature spectrum of cumene for each spectrometer and wavelength. The total dynamic range spans close to four decades, and we observe a good degree of invariance for the intermediate scattering funtion, $I(Q,t)$, along the isochrone determined from the FWS on IN16B that are shown on the same scale as the isotherm for comparison. For the isotherm, we observe as expected the relaxation time decreasing as pressure is released. \begin{figure}[htpb!] \centering \includegraphics[width=0.49\columnwidth]{fft_time_cu_isochrone} \includegraphics[width=0.49\columnwidth]{fft_time_cu_isotherm2} \caption{The relaxation time $\tau_w$ as a function of $Q$ for cumene obtained from a stretched exponential fit to the Fourier transformed combined spectra from IN5 and IN16B along an isochrone and an isotherm. An example of the fit is shown as black lines in Fig.~\ref{fig:res:cumene_fft}. The dashed lines correspond to a $Q^{-2}$ dependence of the relaxation time.}\label{fig:res:cumene_fft_q} \end{figure} Relaxation processes in disordered systems are typically fitted with a stretched exponential as a phenomenological fitting function. We fitted the intermediate scattering functions with a stretched exponential as a function of $Q$: \begin{equation} I(Q,t)=Ae^{-\left(\frac{t}{\tau_w(Q)}\right)^\beta}, \end{equation} where $A$ is the amplitude, $\tau_w$ is the $Q$-dependent relaxation time and $\beta$ is the stretching parameter. The stretching parameter was fixed at $\beta=0.5$ for all spectra as this gave the best fits, which indicates that the shape of all the spectra, independent of temperature and pressure, is the same, which means that isochronal superposition of the $\alpha$-relaxation becomes an evident observation. The $Q$-dependence of the relaxation time is shown in Fig.~\ref{fig:res:cumene_fft_q} for the isochrone determined from the FWS on IN16B and for comparison, the $Q$-dependence of an isotherm. The relaxation time is observed to be fairly invariant along the isochrone, while the relaxation time is observed to slow down with increased pressure along the isotherm. For both the isochrone and the isotherm, the relaxation time is observed to become faster with increasing $Q$, showing a $Q^{-2}$-dependence in a large part of the accessible $Q$-range. This behaviour suggests a Gaussian behaviour which is observed to deviate for high $Q$, i.e. short distances, where the relaxation time is observed to flatten out away from the Gaussian behaviour. At low $Q$, there might be a slight tendency for the relaxation time to bend off into a $Q^{-2/\beta}$ behaviour as suggested in Ref.~\onlinecite{Colmenero93}, although this is not clear from these data sets. The $Q$-dependence observed for cumene suggests translational motion typical for structural relaxation. \begin{figure}[htpb!] \centering \includegraphics[width=0.49\columnwidth]{cu_mct_pred_ds_beta} \includegraphics[width=0.49\columnwidth]{cu_mct_pred_beta} \caption{Comparison of the simple model in Eq.~\ref{eq:res:ifws_model} (full lines) based on the relaxation times obtained from the full cumene spectra, and data on three isotherms from IFWS on IN16B. The density scaled IFWS data are shown to the left and as a function of pressure to the right.}\label{fig:res:cu_mct} \end{figure} Using the mode-coupling parameters reported in Ref.~\onlinecite{Li95}, and an expression for the temparature dependence of the relaxation time reported in Ref.~\onlinecite{Gotze87}, then substituting $\Gamma=\rho^\gamma/T$ for the temperature dependence and $\Gamma_c=\rho_c^\gamma/T_c$ for the critical temperature, we end up with an expression for the relaxation time: \begin{equation}\label{eq:res:tau_mct} \tau=\frac{\tau_0}{((\Gamma-\Gamma_c)/\Gamma_c)^n}, \end{equation} where $n$ is the MCT critical exponent. Only the prefactor $\tau_0$ is fitted to data in the equilibrium liquid above the critical temperature, which is reported to be $\sim\SI{160}{\K}$ at ambient pressure. We use this expression of the $\alpha$-relaxation time and then suggest a simple model for our neutron data with a single relaxation process moving through the instrument window at a fixed energy, $\SI{2}{\micro\eV}$, similar to the IFWS from IN16B. If the relaxation process was an exponential process, this would correspond to a Lorentzian, we therefore add a phenomenological stretching to account for the stretched exponential behaviour. The simple model is then given by \begin{equation}\label{eq:res:ifws_model} I\propto\frac{\tau^\beta}{1+(\omega^2\tau^2)^\beta}. \end{equation} In Fig.~\ref{fig:res:cu_mct}, we have used this model with the expression for the relaxation time as a function of $\rho^\gamma/T$ given in Eq.~\ref{eq:res:tau_mct} to recreate the IFWS shown as a function of $\rho^\gamma/T$ and finally, as a function of pressure. This model suggests only one distribution of relaxation times and is in qualitative agreement with our data, recreating the position of the maximum of the IFWS from the relaxation time obtained from fits to the full spectra and Eq.~\ref{eq:res:tau_mct}, thereby supporting our interpretation of the FWS. \subsection{van der Waals liquid -- PPE} The second van der Waals liquid is PPE, which only barely shows relaxation in the IN16B instrument window in the temperature and pressure range accessible with this setup. Instead, PPE data are reported from IN5 at \SI{5}{\angstrom} in the combined dielectric and neutron sample cell, i.e. accessing dynamics with neutrons on picosecond timescale. Two pairs of isochronal state points were found with high precision from the dielectric loss peak in the simultaneous dielectric and neutron measurements. The two pairs of isochronal state points are shown in Fig.~\ref{fig:res:ppe_diel} with relaxation times of roughly \SI{e-2}{\s} and \SI{e-6}{\s} for isochrone (A) and (B), respectively. The dielectric spectra are shown both as a function of frequency on an absolute frequency scale and in reduced frequency units. We observe that within the experimental uncertainty there is no influence on the relative position of the state points, i.e. isochronal state points on an absolute scale are also isochronal in reduced units. This observation supports the idea of identifying possible isomorphs by isochrones, an approach that was also taken studying the ultraviscous liquid in Ref.~\onlinecite{hwh18}. \begin{figure}[htpb!] \centering \includegraphics[width=0.49\columnwidth]{ppe_diel} \includegraphics[width=0.49\columnwidth]{ppe_diel_reduced} \caption{Two pairs of isochronal state points determined from the dielectric loss peak on PPE measured on IN5, $\lambda=\SI{5}{\angstrom}$. Relaxation time for isochrone A is $\tau_\alpha\approx\SI{e-2}{\s}$ and for B $\tau_\alpha\approx\SI{e-6}{\s}$. Dielectric loss plotted as a function of frequency on an absolute scale (left) and as a function of the reduced frequency unit (right).}\label{fig:res:ppe_diel} \end{figure} The IN5 neutron spectra are shown in Fig.~\ref{fig:res:ppe_spectra} for the two pairs of isochronal state points (A) and (B). The relaxation times of PPE covered here on isochrones and isotherms are located between the relaxation time found for cumene in the previous subsection and the glass transition, a dynamic range not accessible in cumene because of crystallization. For the isotherm, we observe the dynamics slowing down as the liquid is compressed, i.e. less quasielastic broadening. For the isochronal state points, we observe an invariance of the spectral shape and the quasielastic broadening. This collapse is far from trivial; what we see on picosecond timescales is interpreted as the tail of the $\alpha$-relaxation stretching over many orders of magnitude, with vibrations and fast relaxation are here merged with the $\alpha$-relaxation\cite{hwh17}, suggesting that the whole relaxation curve is invariant for isochronal state points also in the liquid state \cite{hwh18}. \begin{figure}[htpb!] \centering \begin{minipage}{0.3\columnwidth} \vspace{-7 em} {\noindent PPE \\ \vspace{0.5 em} \noindent IN5, $\lambda=\SI{5}{\angstrom}$ \\ \vspace{0.5 em} \noindent $\sim\SI{e-11}{\s}$} \end{minipage} \includegraphics[width=0.32\columnwidth]{phasediagram_ppe1} \\ \includegraphics[width=0.49\columnwidth]{ppe_isochrone_red_zoom} \includegraphics[width=0.49\columnwidth]{ppe_isotherm2_red} \caption{Spectra on PPE from IN5 at $\lambda=\SI{5}{\angstrom}$ summed over $Q$. Top: The phase diagram shows the studied state points as a function of temperature and pressure and indicates the two pair of isochronal state points. Bottom: Two pairs of isochronal state points found from dielectrics (left) and an isotherm at \SI{315}{\K} for comparison (right).}\label{fig:res:ppe_spectra} \end{figure} \subsection{Hydrogen-bonding liquid -- DPG} The third set of data that we present is measured on the hydrogen-bonding liquid DPG at IN16B. We present fixed window scans, full spectra and simultaneously acquired dielectric data, securing the precision of the measurements. In Fig.~\ref{fig:res:dpg_fws}, we show the measured EFWS and IFWS intensity at $\Delta E=\SI{0}{\micro\eV}$ and $\Delta E=\SI{2}{\micro\eV}$, respectively, as a function of pressure and as a function of $\rho^\gamma/T$, using $\gamma=1.5$, for five isotherms, and a similar plot with the dielectric loss signal at a fixed frequency of \SI{e5}{\Hz}. We observe an increase in the EFWS intensity upon compression as the mobility is decreased. From the IFWS, we observe that the high-temperature isotherms go through a maximum, similar to what was observed for cumene (Fig.~\ref{fig:res:cumene_fws}), i.e. the $\alpha$-relaxation passes through the neutron instrument window. Equivalently, for the low temperature isotherms, the $\alpha$-relaxation is observed to pass through the fixed frequency of the dielectrics. As the $\alpha$-relaxation moves out of the dielectric window, we observe DC-conductivity, also evident from the full frquency range in Fig.~\ref{fig:res:dpg_spectra}. We observe how the respective intensities can be made to collapse to a master curve for the dynamics on neutron nanosecond timescale when plotted as a function of $\rho^\gamma/T$ , and to some degree for the dielectric signal measured at roughly microsecond timescale. \begin{figure}[htpb!] \centering \includegraphics[width=0.49\columnwidth]{fws_dpg_efws} \includegraphics[width=0.49\columnwidth]{fws_dpg_efws_ds} \\ \includegraphics[width=0.49\columnwidth]{fws_dpg_ifws} \includegraphics[width=0.49\columnwidth]{fws_dpg_ifws_ds} \\ \includegraphics[width=0.49\columnwidth]{ds_dpg_100kHz} \includegraphics[width=0.49\columnwidth]{ds_dpg_100kHz_ds} \caption{FWS from IN16B on DPG summed over $Q$. Intensity of EFWS (top) and IFWS (middle) and fixed frequency dielectric loss (bottom) plotted for five isotherms (\SI{240}{\K}, \SI{270}{\K}, \SI{280}{\K}, \SI{300}{\K}, \SI{315}{\K}) as a function of pressure (left) and as a function of $\Gamma=\rho^\gamma/T$ (right).}\label{fig:res:dpg_fws} \end{figure} \begin{figure}[htpb!] \centering \begin{minipage}{0.3\columnwidth} \vspace{-7 em} {\noindent DPG \\ \vspace{0.5 em} \noindent IN16B \\ \vspace{0.5 em} \noindent $\sim\SI{e-9}{\s}$} \end{minipage} \includegraphics[width=0.32\columnwidth]{phasediagram_dpg1} \\ \includegraphics[width=0.49\columnwidth]{dpg_diel} \\ \includegraphics[width=0.49\columnwidth]{dpg_in16b_isochrone_red} \includegraphics[width=0.49\columnwidth]{dpg_in16b_isobar_red} \caption{Top: $(T,P)$-phase diagram with the studied state points of DPG on IN16B. Centre: dielectric loss as a function of frequency. Bottom: Spectra on DPG summed over $Q$ showing an isochrone (left) determined from FWS on IN16B and the dielectrics and an isobar for comparison (right).}\label{fig:res:dpg_spectra} \end{figure} Just as for cumene, an isochrone was determined from the EFWS and IFWS intensity, but in this case also from the DC-conductivity and the wing of the relaxation observed in the dielectrics. A temperature-pressure phase diagram illustrating the studied state points is shown in Fig.~\ref{fig:res:dpg_spectra} along with the full spectra from IN16B and the dielectric signal at those state points. The isochrone determined from the fixed window neutron signal is observed also to be invariant in the dielectrics where mainly the DC-conductivity is observed and just the tail of the $\alpha$-relaxation at high frequencies. The full spectral shape and quasielastic broadening of the IN16B data is observed to be invariant for the isochrone. An isobar is shown for comparison, where we observe the dynamics speeding up as the temperature is increased, causing more broadening and less elastic intensity. \begin{figure}[htpb!] \centering \includegraphics[width=0.49\columnwidth]{fws_dpg_Tscans} \\ \includegraphics[width=0.49\columnwidth]{fws_dpg_dsE.pdf} \includegraphics[width=0.49\columnwidth]{fws_dpg_dsI.pdf} \caption{Top: EFWS and IFWS from IN16B on DPG along two isobars at ambient pressure and \SI{300}{\mega\pascal}. Bottom: Density scaling of the EFWS (left) and IFWS (right) that are shown above and in Fig.~\ref{fig:res:dpg_fws}, i.e. for both the isobars and isotherms.}\label{fig:res:dpg_TP} \end{figure} \begin{figure}[htpb!] \centering \includegraphics[width=0.49\columnwidth]{fws_dpg_q_300I} \includegraphics[width=0.49\columnwidth]{fws_dpg_q_240I} \caption{$Q$-dependence of the IFWS of two isotherms from IN16B on DPG at \SI{300}{\K} (left) and \SI{240}{\K} (right).}\label{fig:res:dpg_Q} \end{figure} The spectra from IN16B shown in Fig.~\ref{fig:res:dpg_spectra} are all from state points where the $\alpha$-relaxation is near the timescale of the backscattering instrument window of IN16B, i.e. nanoseconds. In Fig.~\ref{fig:res:dpg_TP} (top), we present fixed window scans from two isobars, at ambient pressure and at \SI{300}{\mega\pascal}, carried out from the high temperature region, where the $\alpha$-relaxation is in the instrument window, and all the way into the glassy state. On cooling at IN16B, DPG is observed to have a distinct signal with a maximum close to the glass transition, $T_g(P_\mathrm{atm})=\SI{195}{\K}$. This signal is well separated from the $\alpha$-relaxation on IN16B which is roughly \SI{100}{\K} higher at ambient pressure. While the $\alpha$-relaxation with a maximum at $\sim\SI{300}{\K}$ is observed to move upon increased pressure, the second bumb at lower temperatures does not respond to pressure, also visible in the pressure scans in Fig.~\ref{fig:res:dpg_fws}, which indicates that this contribution is from intra-molecular dynamics. If we compare the $Q$-dependence for the isotherm at \SI{240}{\K} to that at \SI{300}{\K} (Fig.~\ref{fig:res:dpg_Q}), we see that for the high temperature isotherm, there is a clear shift on compression in the relaxation process towards higher $Q$, indicating that this is of translational character, whereas there is no $Q$-dependence observed for the isotherm at \SI{240}{\K}, suggesting a local process with a higher intensity at higher $Q$. This signal most likely stems from the methyl-group rotation of DPG and is manifested in the IFWS as the second broad bump at lower temperatures than the $\alpha$-relaxation with a maximum around \SI{200}{\K} (Fig.~\ref{fig:res:dpg_TP}). By combining the two isobars with the five isotherms in Fig.~\ref{fig:res:dpg_fws}, we observe that when the EFWS and IFWS are plotted as a function of $\rho^\gamma/T$, density scaling breaks down for the methyl-group rotation dynamics and the dynamics can no longer be expressed solely by $\Gamma=\rho^\gamma/T$. \section{Discussion} In the previous section, we showed how dynamics of the $\alpha$-relaxation taking place on timescales accessible with neutron and dielectric spectroscopy can be made to collapse into one curve when performing density scaling in the equilibrium liquid. The dynamics involved covers timescales from picoseconds to milliseconds and suggests that the timescale of the $\alpha$-relaxation is determined by the control parameter $\Gamma=\rho^\gamma/T$. We have shown isochronal superposition over several decades in time accessed by different spectrometers for the simple van der Waals liquid PPE, i.e. an invariance in the quasielastic broadening on picosecond timescale along isochrones determined from dielectrics. Also for DPG, we observed that the DC-conductivity visible from dielectrics and the full spectral shape and linewidth were invariant along the isochrone determined from the FWS on IN16B and the dielectric spectra. For cumene, that was measured on two different spectrometers and combined to cover almost four orders of magnitude in the time domain, allowed for a more careful analysis of the relaxaton. The spectral shape of the Fourier-transformed cumene data was fitted with the same stretching exponent $\beta=0.5$ for all temperatures and pressure, which suggests time-temperature-pressure superposition, i.e. that the shape of the $\alpha$-relaxation is always the same in the equilibrium liquid far from the glass transition. This observation implies an extra level of simplicity that seems to apply to some liquids that makes the observation of isochronal superposition an evident observation. Most density scaling studies have been performed in the viscous liquid close to the glass transition. In this study, we move away from the slow-flowing liquid and into the liquid state. The $\alpha$-relaxation times for cumene and DPG were in this study mainly in the instrument window of IN16B, i.e. on nanosecond timescale. Density scaling for the $\alpha$-relaxation dynamics was shown for the two liquids with very different values of $\gamma$, both reported elsewhere and found in different dynamic range measured with different techniques. While the scaling exponent $\gamma$ is allowed to vary, we observe that for cumene $\gamma=4.8$ (Ref.~\onlinecite{Ransom17}) works well within the experimental uncertainty that we have to work with. The $\gamma$-value of cumene is fairly high, meaning that density plays a much larger role for the dynamics, as compared to the hydrogen-bonding liquid DPG, where the scaling exponent is just $\gamma=1.5$ (Ref.~\onlinecite{Grzybowski06_gamma}), meaning that the dynamics therefore mainly is affected by temperature. In agreement with previous observations \cite{Adrjanowicz16,Puosi16,Romanini17}, we observe density scaling to work well also for the tested hydrogen-bonding liquid when it comes to the $\alpha$-relaxation. But we observed density scaling to break down for the intra-molecular dynamics of the methyl-group rotation that was identified from the lack of pressure response and the $Q$-dependent signal. Previously, we have observed isochronal superposition to break down for the hydrogen-bonding liquid DPG along the glass transition when fast relaxation and vibrations were completely separated from the $\alpha$-relaxation \cite{hwh18}. The fact that we observe density scaling to break down in this study for the intra-molecular motion in DPG is in agreement with observations from molecular dynamics simulations \cite{Veldhorst15,Olsen16}. Spring-like bonds were introduced as intra-molecular motion and what was referred to as pseudo-isomorphs were recovered when only the intermolecular dynamics were treated, i.e. the scaling only worked for intermolecular motion. A partly deuterated DPG sample that would mask the methyl-group rotation in the neutron signal would therefore be an obvious next approach to test whether density scaling in this way can be restored for the full dynamic range. \subsection{Acknowledgements} This work was supported by the Danish Council for Independent Research (Sapere Aude: Starting Grant). We acknowledge the workshop at IMFUFA and everyone involved in the LTP-6-7 at the ILL for technical support: instrument responsibles Judith Peters on IN13 and Jacques Olivier on IN5; Eddy Leli\'{e}vre-Berna, cryogenics and high-pressure from the SANE group. \bibliographystyle{apsrev4-1}

2a8aa260ce93b4a75b1e15926c089e272f755826

\section{Introduction} In this paper we study the Generalised Nonvanishing Conjecture in every dimension, and derive some surprising nonvanishing results for pluricanonical bundles which were not predicted by the Minimal Model Program. \begin{GenNonvanCon} Let $(X,\Delta)$ be a klt pair such that $K_X+\Delta$ is pseudoeffective. Let $L$ be a nef $\mathbb{Q}$-divisor on $X$. Then for every $t\geq0$ the numerical class of the divisor $K_X+\Delta+tL$ belongs to the effective cone. \end{GenNonvanCon} If the numerical class of a divisor belongs to the effective cone, then we call the divisor \emph{num-effective}, see Section \ref{sec:prelim}. The Generalised Nonvanishing Conjecture as well as the Generalised Abundance Conjecture were introduced in our paper \cite{LP18a} as generalisations of several central conjectures in algebraic geometry. This conjecture contains as special cases the Nonvanishing Conjecture -- often viewed as part of the Abundance Conjecture -- and the nonvanishing part of the Semiampleness Conjecture (for nef line bundle on varieties with numerically trivial canonical class). \subsection{Numerical versus linear equivalence} Constructing many non-tri\-vi\-al effective or basepoint free divisors on a projective variety is one of the main goals of the Minimal Model Program. Assume, for instance, that $(X,\Delta)$ is a projective klt pair such that $K_X+\Delta$ is semiample. It is a basic question to determine which $\mathbb{Q}$-divisors $D$ in the numerical class of $K_X+\Delta$ are effective (that is, $\kappa(X,D)\geq0$) or semiample. If the numerical and $\mathbb{Q}$-linear equivalence on $X$ coincide, i.e.\ if $H^1(X,\mathcal{O}_X) = 0 $, then \emph{all} such $D$ are trivially effective, respectively semiample, and one would expect that this is the only case where such behaviour happens. Our first main result is that, surprisingly, such behaviour happens \emph{almost always}. More precisely, whether each such $D$ is effective, respectively semiample, depends not (only) on the vanishing of the cohomology group $H^1(X,\mathcal{O}_X)$, but on the behaviour of the Euler-Poincar\'e characteristic of the structure sheaf of $X$. We obtain the following unconditional results in dimensions at most $3$, which were previously not even conjectured. \begin{thmA}\label{cor:F} Let $(X,\Delta)$ be a projective klt pair of dimension at most $3$ such that $K_X+\Delta$ is pseudoeffective. Assume that $\chi(X,\mathcal{O}_X)\neq0$. \begin{enumerate} \item[(i)] Then for every $\mathbb{Q}$-divisor $G$ with $K_X+\Delta\equiv G$ we have $\kappa(X,G)\geq0$. \item[(ii)] If $K_X+\Delta$ is semiample, then every $\mathbb{Q}$-divisor $G$ with $K_X+\Delta\equiv G$ is semiample. \end{enumerate} \end{thmA} Theorem \ref{cor:F} is a special case of Corollaries \ref{cor:surfaces} and \ref{cor:threefolds} below. Another instance where our results hold unconditionally \emph{in every dimension} is the case where the numerical dimension is $1$; this is the first highly nontrivial case beyond those when the numerical dimension is $0$ or maximal. \begin{thmA} \label{thm:G} Let $(X,\Delta)$ be a projective terminal pair of dimension $n$ such that $K_X + \Delta$ is nef and let $L$ be a nef $\mathbb{Q}$-Cartier divisor on $X$. Assume that $\nu(X,K_X+\Delta+L) = 1$ and $\chi(X,\mathcal{O}_X) \neq 0$. Then $K_X+\Delta+tL$ is num-effective for all $t\geq0$. \end{thmA} This is proved in Section \ref{sec:num1}. As an illustration, consider a terminal Calabi-Yau pair $(X,\Delta)$, i.e.\ a terminal pair such that $K_X + \Delta \equiv 0$. Let $L$ be a nef divisor on $X$ of numerical dimension $1$. If $\chi(X,\mathcal{O}_X) \neq 0$, then $L$ is num-effective. \subsection{Metrics with algebraic singularities} Theorem \ref{cor:F} above is a special instance of more general results which work in all dimensions but require inductive hypotheses. Let $(X,\Delta)$ be a klt pair such that $K_X+\Delta$ is nef and $X$ is not uniruled, and let $L$ be a nef $\mathbb{Q}$-divisor on $X$. Our main technical result, Theorem \ref{thm:nonvanishingForms} below, gives a general criterion for the existence of sections of some multiple of $K_X+\Delta+L$. The criterion generalises the main results of \cite{LP18}: the proof uses the birational stability of the cotangent bundle from \cite{CP15} (this is where the non-uniruledness of $X$ is necessary) together with the very carefully chosen MMP techniques in Theorem \ref{thm:MMPtwist} (this is where the nefness of $K_X+\Delta$ is used). We then apply Theorem \ref{thm:nonvanishingForms} to line bundles possessing metrics with \emph{generalised algebraic singularities}; the importance of this class of metrics is explained in Section \ref{sec:prelim}. We expect these methods to be crucial in the resolution of the Generalised Nonvanishing Conjecture. In Section \ref{sec:algsing2} we prove our central inductive statement. \begin{thmA}\label{thm:C} Assume the termination of flips in dimensions at most $n-1$, and the Abundance Conjecture in dimensions at most $n-1$. Let $(X,\Delta)$ be a projective klt pair of dimension $n$ such that $K_X+\Delta$ is pseudoeffective, and let $L$ be a nef $\mathbb{Q}$-divisor on $X$. Suppose that $K_X+\Delta$ and $K_X+\Delta+L$ have singular metrics with generalised algebraic singularities and semipositive curvature currents, and that $\chi(X,\mathcal{O}_X)\neq0$. \begin{enumerate} \item[(i)] Then for every $\mathbb{Q}$-divisor $L'$ with $L\equiv L'$ there exists a positive rational number $t_0$ such that $$\kappa(X,K_X+\Delta+tL')\geq0\quad\text{for every }0\leq t\leq t_0.$$ \item[(ii)] Assume additionally the semiampleness part of the Abundance Conjecture in dimension $n$. Then $$K_X+\Delta+tL\quad\text{is num-effective for every }t\geq0.$$ \end{enumerate} \end{thmA} With notation from Theorem \ref{thm:C}, we note that by \cite[Theorem A]{LP18a}, all divisors of the form $K_X+\Delta+L$ should possess singular metrics with generalised algebraic singularities and semipositive curvature currents. \medskip We mention a special case, where the conclusion of Theorem \ref{thm:C} is much stronger: when $K_X+\Delta$ and $K_X+\Delta+L$ additionally have smooth metrics with semipositive curvature, or, more generally, singular metrics with semipositive curvature currents and vanishing Lelong numbers, then $K_X+\Delta+L$ is actually semiample. This is Theorem \ref{thm:semipositive}, and the proof uses crucially the main result of \cite{GM17}. As a corollary of Theorem \ref{thm:C}, we obtain the following: \begin{corA}\label{cor:D} Let $(X,\Delta)$ be a projective klt pair of dimension $4$ such that $K_X+\Delta$ is pseudoeffective and let $L$ be a nef $\mathbb{Q}$-divisor on $X$. Suppose that $K_X+\Delta$ and $K_X+\Delta+L$ have singular metrics with generalised algebraic singularities and semipositive curvature currents, and that $\chi(X,\mathcal{O}_X)\neq0$. Then there exists a positive rational number $t_0$ such that $$\kappa(X,K_X+\Delta+tL)\geq0\quad\text{for every }0\leq t\leq t_0.$$ \end{corA} Theorem \ref{thm:C} has an unexpected consequence for the Minimal Model Program in terms of the existence of sections of divisors numerically equivalent to adjoint divisors, generalising Theorem \ref{cor:F} to every dimension. The following result is a special case of Corollary \ref{cor:numequiv} and Theorem \ref{thm:semipositive1}. \begin{thmA}\label{thm:E} Assume the termination of flips in dimensions at most $n-1$, and the existence of good models in dimensions at most $n$. Let $(X,\Delta)$ be a projective klt pair of dimension $n$ such that $K_X+\Delta$ is pseudoeffective. Assume that $\chi(X,\mathcal{O}_X)\neq0$. \begin{enumerate} \item[(i)] Then for every $\mathbb{Q}$-divisor $G$ with $K_X+\Delta\equiv G$ we have $\kappa(X,G)\geq0$. \item[(ii)] If $K_X+\Delta$ is semiample, then every $\mathbb{Q}$-divisor $G$ with $K_X+\Delta\equiv G$ is semiample. \end{enumerate} \end{thmA} \subsection{The Weak Generalised Nonvanishing Conjecture} As an important part of this paper on which the previous results depend, we show that the assumptions of the Generalised Nonvanishing Conjecture can be weakened considerably. In fact, we reduce the Generalised Nonvanishing Conjecture to the following. \begin{WeakGenNonvanCon} Let $(X,\Delta)$ be a $\mathbb{Q}$-fac\-to\-ri\-al projective klt pair of dimension $n$ such that $X$ is not uniruled and $K_X + \Delta $ is nef. Let $L$ be a nef $\mathbb{Q}$-divisor on $X$ such that $n(X,K_X + \Delta + L) = n$. Then for every $t\geq0$ the numerical class of the divisor $K_X+\Delta+tL$ belongs to the effective cone. \end{WeakGenNonvanCon} Here, for any $\mathbb{Q}$-Cartier divisor $D$ on $X$, we denote by $n(X,D)$ the nef dimension of $D$, see Section \ref{sec:prelim}. The condition $n(X,D) = \dim X $ is equivalent to saying that $X$ cannot be covered by curves $C$ such that $D \cdot C = 0$. Conjecturally, when $n(X,K_X + \Delta + L) = \dim X$ as above, then $K_X+\Delta+L$ is big. This conjecture is weaker than the Generalised Nonvanishing Conjecture in three different ways: the variety $X$ is not uniruled; the pair $(X,\Delta)$ is a minimal model; and the nef dimension of $K_X+\Delta+L$ is maximal; see Section \ref{sec:prelim} for details. \medskip In Section \ref{sec:reductions} we prove: \begin{thmA}\label{thm:mainreduction} Assume the termination of klt flips in dimensions at most $n-1$, the Abundance Conjecture for klt pairs in dimensions at most $n-1$, the existence of minimal models of klt pairs in dimension $n$, and the Weak Generalised Nonvanishing Conjecture in dimensions at most $n$. Let $(X,\Delta)$ be an $n$-dimensional klt pair such that $K_X+\Delta$ is pseudoeffective and let $L$ be a nef $\mathbb{Q}$-divisor on $X$. \begin{enumerate} \item[(i)] Then there exists a positive rational number $t_0$ such that $K_X+\Delta+tL$ is num-effective for every $0\leq t\leq t_0$. \item[(ii)] Assume additionally the semiampleness part of the Abundance Conjecture in dimension $n$. Then $K_X+\Delta+tL$ is num-effective for every $t\geq0$, that is, the Generalised Nonvanishing Conjecture holds in dimension $n$. \end{enumerate} \end{thmA} The heart of the argument is in Theorem \ref{thm:Support}, which essentially deals with the situation when $X$ is not uniruled, together with Theorem \ref{thm:B}, which deals with the situation when the nef dimension $n(X,K_X+\Delta+L)$ is not maximal. \medskip A more general version of the Generalised Nonvanishing Conjecture was proposed by Han and Liu \cite{HL18}. They do not require the divisor $K_X+\Delta$ to be pseudoeffective and allow log canonical singularities, and they confirm the conjecture in dimension $2$. \section{Preliminaries}\label{sec:prelim} In this section we collect notation and technical lemmas that we use later. \subsection{Num-effectivity, models, MMP} Unless otherwise stated, a $\mathbb{Q}$-di\-vi\-sor on a projective variety is assumed to be a Weil $\mathbb{Q}$-divisor. We write $D \geq 0$ for an effective $\mathbb{Q}$-divisor $D$ on a normal variety $X$. A $\mathbb{Q}$-divisor $L$ on a projective variety $X$ is \emph{num-effective} if the numerical class of $L$ belongs to the effective cone of $X$. A $\mathbb{Q}$-divisor $L$ on a projective variety $X$ is \emph{num-semiample} if there exists a semiample $\mathbb{Q}$-divisor $L'$ on $X$ such that $L\equiv L'$. We need the following easy lemma. \begin{lem}\label{lem:easylemma} Let $D$ and $D'$ be effective $\mathbb{Q}$-Cartier $\mathbb{Q}$-divisors on a normal projective variety $X$ and let $L$ be a $\mathbb{Q}$-Cartier $\mathbb{Q}$-divisor on $X$. Assume that $\Supp D=\Supp D'$. Then: \begin{enumerate} \item[(a)] $D+tL$ is big for all $t>0$ if and only if $D'+tL$ is big for all $t>0$, \item[(b)] there exists $t_0>0$ such that $D+tL$ is num-effective for all $0\leq t\leq t_0$ if and only if there exists $t_0'>0$ such that $D'+tL$ is num-effective for all $0\leq t\leq t_0'$. \end{enumerate} \end{lem} \begin{proof} Assume that $D+tL$ is big for all $t>0$. Pick a positive integer $m$ such that $mD'\geq D$. Then for each $t>0$, the divisor $$m(D'+tL)=(D+mtL)+(mD'-D)$$ is big. This shows (a). Now assume that $D+tL$ is num-effective for all $0\leq t\leq t_0$, for some positive $t_0$. Pick a positive integer $m$ such that $mD'\geq D$. Then for every $0\leq t\leq t_0/m$ there exists an effective divisor $F$ such that $D+mt_0 L\equiv F$, hence $$m(D'+tL)=D+mtL+(mD'-D)\equiv F+(mD'-D)\geq0,$$ which gives (b). \end{proof} A \emph{fibration} is a projective surjective morphism with connected fibres between two normal varieties. If $f\colon X\to Y$ is a surjective morphism of normal projective varieties and if $D$ is a $\mathbb{Q}$-divisor on $X$, then $D$ is \emph{$f$-exceptional} if $\codim_Y f(\Supp D) \geq 2$. \begin{dfn} Let $X$ and $Y$ be normal projective varieties, and let $D$ be a $\mathbb{Q}$-Cartier $\mathbb{Q}$-divisor on $X$. Let $\varphi\colon X\dashrightarrow Y$ be a birational contraction, and assume that $\varphi_*D$ is $\mathbb{Q}$-Cartier. Then $\varphi$ is \emph{$D$-non-positive} (respectively \emph{$D$-negative}) if there exists a smooth resolution of indeterminacies $(p,q)\colon W\to X\times Y$ of $\varphi$ such that $$ p^*D\sim_\mathbb{Q} q^*\varphi_*D+E,$$ where $E\geq0$ is a $q$-exceptional $\mathbb{Q}$-divisor (respectively, $E\geq0$ is a $q$-exceptional $\mathbb{Q}$-divisor and $\Supp E$ contains the proper transform of every $\varphi$-exceptional divisor). If additionally $\varphi_*D$ is semiample, the map $\varphi$ is a \emph{good model} for $D$. \end{dfn} \begin{lem}\label{lem:MMPnumeff} Let $\varphi\colon X\dashrightarrow Y$ be a birational contraction between normal projective varieties, where $X$ has rational singularities. Let $D$ be a $\mathbb{Q}$-Cartier $\mathbb{Q}$-divisor on $X$ such that $\varphi_*D$ is $\mathbb{Q}$-Cartier on $Y$ and such that the map $\varphi$ is $D$-non-positive. If $\varphi_*D$ is num-effective, then $D$ is num-effective. \end{lem} \begin{proof} Let $(p,q)\colon W\to X\times Y$ be a smooth resolution of indeterminacies of $\varphi$. Since $\varphi$ is $D$-non-positive, there exists an effective $q$-exceptional divisor $E$ on $W$ such that $$p^*D\sim_\mathbb{Q} q^*\varphi_*D+E.$$ If there exists an effective $\mathbb{Q}$-divisor $G$ on $Y$ such that $\varphi_*D\equiv G$, then $$p^*D\equiv q^*G+E\geq0,$$ hence $D$ is num-effective by \cite[Lemma 2.14]{LP18a}. \end{proof} A \emph{pair} $(X,\Delta)$ consists of a normal variety $X$ and a $\mathbb{Q}$-divisor $\Delta\geq0$ such that $K_X+\Delta$ is $\mathbb{Q}$-Cartier. The standard reference for the foundational definitions and results on the singularities of pairs and the Minimal Model Program is \cite{KM98}, and we use these freely in this paper. We recall additionally that flips for klt pairs exist by \cite[Corollary 1.4.1]{BCHM}. We also use throughout the paper that for every projective klt pair $(X,\Delta)$, a small $\mathbb{Q}$-factorialisation of $(X,\Delta)$ exists, see \cite[Corollary 1.37]{Kol13}; and that a terminalisation of $(X,\Delta)$ exists, see \cite[paragraph after Corollary 1.4.3]{BCHM}. A small $\mathbb{Q}$-factorialisation of $X$ is an isomorphism over the $\mathbb{Q}$-factorial locus of $X$, see \cite[1.40]{Deb01}. The following result is well-known. \begin{lem}\label{lem:numlintrivial} Let $(X,\Delta)$ be a projective klt pair and let $f\colon X\to Y$ be a fibration to a projective variety $Y$. If $K_X+\Delta\equiv_f0$, then $K_X+\Delta\sim_{f,\mathbb{Q}}0$. \end{lem} \begin{proof} Let $F$ be a general fibre of $f$. Then $(K_X+\Delta)|_F\sim_\Q0$ by \cite[Corollary V.4.9]{Nak04}, and thus the pair $(X,\Delta)$ has a good model over $Y$ by \cite[Theorem 2.12]{HX13}. Therefore, $K_X+\Delta$ is $f$-semiample, and there exists a fibration $\pi\colon X\to Z$ over $Y$ and a $\mathbb{Q}$-divisor $A$ on $Z$ which is ample over $Y$ such that $K_X+\Delta\sim_\mathbb{Q}\pi^*A$. As $K_X+\Delta\equiv_f0$, the map $\pi$ must be an isomorphism, and the lemma follows. \end{proof} \subsection{Numerical dimension}\label{subsec:numdim} We recall the definition of the numerical dimension from \cite{Nak04,Kaw85}. \begin{dfn}\label{dfn:kappa} Let $X$ be a normal projective variety, let $D$ be a pseudoeffective $\mathbb{Q}$-Cartier $\mathbb{Q}$-divisor on $X$ and let $A$ be any ample divisor on $X$. Then the {\em numerical dimension\/} of $D$ is $$\nu(X,D)=\sup\big\{k\in\mathbb{N}\mid \limsup_{m\rightarrow\infty}h^0\big(X, \mathcal O_X(\lfloor mD\rfloor+A)\big)/m^k >0\big\}.$$ When the divisor $D$ is nef, then equivalently $$\nu(X,D)=\sup\{k\in\mathbb{N}\mid D^k\not\equiv0\}.$$ If $D$ is not pseudoeffective, we set $\nu(X,D)=-\infty$. \end{dfn} The Kodaira dimension and the numerical dimension of a $\mathbb{Q}$-divisor $D$ are preserved under any $D$-non-positive birational map, see for instance \cite[\S2.2]{LP18a}. If $X$ is a normal projective variety, and if $D$ and $D'$ are pseudoeffective $\mathbb{Q}$-divisors on $X$ such that $D'-D$ is pseudoeffective, then $\nu(X,D')\geq\nu(X,D)$ by the proof of \cite[Proposition V.2.7]{Nak04}. If $D$ is a pseudoeffective $\mathbb{R}$-Cartier $\mathbb{R}$-divisor on a smooth projective variety $X$, and if $\Gamma$ is a prime divisor on $X$, we denote by $\sigma_\Gamma(D)$ Nakayama's multiplicity of $D$ along $\Gamma$, see\ \cite[Chapter III]{Nak04} for the definition and the basic properties. For the proof of the following lemma we refer to \cite[Lemma 2.3]{LP18a}. \begin{lem}\label{lem:1} Let $X$, $X'$, $Y$ and $Y'$ be normal varieties, and assume that we have a commutative diagram \[ \xymatrix{ X' \ar[d]_{\pi'} \ar[r]^{f'} & Y' \ar[d]^{\pi}\\ X \ar[r]_{f} & Y, } \] where $\pi$ and $\pi'$ are projective birational. Let $D$ be a $\mathbb{Q}$-Cartier $\mathbb{Q}$-divisor on $X$ and let $E\geq0$ be a $\pi'$-exceptional $\mathbb{Q}$-Cartier $\mathbb{Q}$-divisor on $X'$. If $F$ and $F'$ are general fibres of $f$ and $f'$, respectively, such that $F'=\pi'^{-1}(F)$, then $$\nu(F,D|_F)=\nu\big(F',(\pi'^*D+E)|_{F'}\big).$$ \end{lem} \subsection{Nef reduction} In this paper we use crucially the nef reduction map. The following is the main result of \cite{BCE+}. \begin{thm}\label{thm:nefreduction} Let $L$ be a nef divisor on a normal projective variety $X$. Then there exists an almost holomorphic dominant rational map $f\colon X\dashrightarrow Y$ with connected fibres to a normal projective variety $Y$, called the \emph{nef reduction of $L$}, such that: \begin{enumerate} \item[(i)] $L$ is numerically trivial on all compact fibres $F$ of $f$ with $\dim F=\dim X-\dim Y$, \item[(ii)] for every very general point $x\in X$ and every irreducible curve $C$ on $X$ passing through $x$ and not contracted by $f$, we have $L\cdot C>0$. \end{enumerate} The map $f$ is unique up to birational equivalence of $Y$ and the \emph{nef dimension} $n(X,L)$ of $L$ is defined as the dimension of $Y$: $$ n(X,L) = \dim Y. $$ \end{thm} It is immediate from the definition that, with notation above, we have $n(X,L)=\dim X$ if and only if $L\cdot C>0$ for every irreducible curve $C$ on $X$ passing through a very general point on $X$. In the following lemmas we study how the nef dimension behaves under morphisms and birational contractions. \begin{lem}\label{lem:nefredsum} Let $X$ be a normal projective variety of dimension $n$. Let $F$ and $G$ be pseudoeffective divisors on $X$ such that $(F+G)\cdot C_X>0$ for every curve $C_X$ on $X$ passing through a very general point of $X$. Then for every positive real number $t$ and every curve $C$ on $X$ passing through a very general point of $X$ we have $(F+tG)\cdot C>0$. \end{lem} \begin{proof} The proof is the same as that of \cite[Lemma 2.11]{LP18a}. \end{proof} \begin{lem}\label{lem:generalintersections} Let $f\colon X\to Y$ be a birational morphism between normal projective varieties of dimension $n$. Let $D$ be a nef $\mathbb{Q}$-divisor on $Y$ and let $G=f^*D+E$, where $E\geq0$ is $f$-exceptional. If $G\cdot C_X>0$ for every curve $C_X$ on $X$ passing through a very general point on $X$, then $n(Y,D)=n$. \end{lem} \begin{proof} Denote $V:=f(\Exc f)$. Assume that $G\cdot C_X>0$ for every curve $C_X$ on $X$ passing through a point in a very general subset $U_X\subseteq X$, and assume that $n(Y,D)<n$. Let $\varphi\colon Y\dashrightarrow Z$ be the nef reduction of $D$. Then $\dim Y<n$ and recall that $\varphi$ is almost holomorphic. The set $U_Y=Y\setminus f(X\setminus U_X)$ is a very general subset of $Y$ such that $f^{-1}(U_Y)\subseteq U_X$. If $F$ is a very general fibre of $\varphi$, then $U_Y\cap F$ is a very general subset of $F$, and $\codim_F(F\cap V)\geq2$. Therefore, a very general complete intersection curve $C_Y$ in $F$ avoids the set $F\cap V$ and intersects $U_Y\cap F$. Then for $C:=f^{-1}(C_Y)$ we have $G\cdot C=0$ and $C\cap U_X\neq\emptyset$, a contradiction. \end{proof} \begin{lem}\label{lem:nefdim1} Let $f\colon X\to Y$ be a birational morphism between normal projective varieties of dimension $n$. Let $D$ be a $\mathbb{Q}$-divisor on $Y$ and let $G=f^*D+E$, where $E$ is a pseudoeffective divisor on $X$. Assume that $D\cdot C_Y>0$ for every curve $C_Y$ on $Y$ passing through a very general point on $Y$. Then $G\cdot C_X>0$ for every curve $C_X$ on $X$ passing through a very general point on $X$. \end{lem} \begin{proof} Assume that $D\cdot C_Y>0$ for every curve $C_Y$ on $Y$ passing through a point in a very general subset $U_Y\subseteq Y$. We may assume that $U_Y$ is contained in the locus on $Y$ over which $f$ is an isomorphism. Since $E$ is pseudoeffective, there exists a very general subset $U_X\subseteq X$ such that $E\cdot C_X\geq0$ for every curve $C_X$ on $X$ passing through a point in $U_X$. Then for every curve $C$ on $X$ passing through a point in $U_X\cap f^{-1}(U_Y)$ we have $G\cdot C\geq f^*D\cdot C=D\cdot f(C)>0$. \end{proof} \begin{cor}\label{cor:nefdim} Let $\varphi\colon X\dashrightarrow Y$ be a birational contraction between normal projective varieties of dimension $n$. Let $D$ be a $\mathbb{Q}$-Cartier $\mathbb{Q}$-divisor on $X$ such that $\varphi_*D$ is $\mathbb{Q}$-Cartier and nef, and assume that $\varphi$ is $D$-non-positive. If $D\cdot C>0$ for every curve $C$ on $X$ passing through a very general point on $X$, then $n(Y,\varphi_*D)=n$. \end{cor} \begin{proof} Let $(p,q)\colon W\to X\times Y$ be a resolution of indeterminacies of $f$. Since $\varphi$ is $D$-non-positive, there exists an effective $q$-exceptional divisor $E$ on $W$ such that $$p^*D\sim_\mathbb{Q} q^*\varphi_*D+E.$$ By Lemma \ref{lem:nefdim1} we have $p^*D\cdot C_W>0$ for every curve $C_W$ on $W$ passing through a very general point on $W$. But then the conclusion follows from Lemma \ref{lem:generalintersections}. \end{proof} \subsection{Metrics on $\mathbb{Q}$-divisors} We briefly recall the properties of metrics on line bundles on projective varieties. For a more thorough treatment, see \cite{DPS01,De12,Bou04}. In the sequel, if $F$ is a prime divisor on a smooth projective variety $X$, then $[F]$ denotes the current of integration along $F$. If $h$ is a singular metric on a some line bundle, and if $\varphi$ is a corresponding local plurisubharmonic function, then $\mathcal I(h)=\mathcal I(\varphi)$ denote the multiplier ideal associated to $h$, respectively $\varphi$. Let $X$ be a smooth projective variety and let $D$ be a Cartier divisor on $X$. Then $D$ is pseudoeffective if and only if $\mathcal{O}_X(D)$ has a singular metric $h$ whose associated curvature current $\Theta_h(D)$ is semipositive (as a current); we also write $\Theta_h(D)\geq0$.\footnote{Often in the literature such currents are called positive.} Equivalently, $D$ is pseudoeffective if and only if its Chern class $c_1(D)$ contains a closed semipositive current. Recall that by Demailly's regularisation technique \cite[\S15.B]{De12}, every pseudoeffective line bundle on a smooth projective variety can be equipped with a sequence of singular metrics with algebraic singularities whose curvature currents converge to a semipositive current; for the definition of metrics with algebraic singularities, see for instance \cite{LP18}. Thus, line bundles possessing singular metrics with algebraic singularities and semipositive curvature current form an important subclass of pseudoeffective line bundles; for instance, all hermitian semipositive line bundles belong to this class, as well as all num-effective line bundles. In this paper, we consider a more general class of metrics with generalised algebraic singularities. This class of singularities behaves very well under the operations of the Minimal Model Program. \begin{dfn} \label{dfn:metrics} Let $X$ be a normal complex projective variety and $D$ a Cartier divisor on $X$. We say that $D$ has a metric with \emph{generalised algebraic singularities} and semipositive curvature current, if there exists a resolution of singularities $\pi\colon Y \to X$ such that the line bundle $\pi^*\mathcal{O}_X(D)$ has a singular metric $h$ whose curvature current is semipositive, and the Siu decomposition (see, for instance, \cite[\S2.2]{Bou04}) of $\Theta_h(\pi^*D)$ has the form $$\Theta_h(\pi^*D)=\Theta+\sum\lambda_j [D_j],$$ where $\Theta\geq0$ is a current whose all Lelong numbers are zero, and $\sum\lambda_j [D_j]$ is a $\mathbb{Q}$-divisor on $Y$. We then say that the metric $h$ \emph{descends} to $Y$. If $D$ is a $\mathbb{Q}$-Cartier divisor, we say that $D$ has a metric with generalised algebraic singularities and semipositive curvature current, if there exists a positive integer $m$ such that $mD$ is Cartier and $mD$ has a metric with generalised algebraic singularities and semipositive curvature current. \end{dfn} Since the pullback of a $(1,1)$-current whose all Lelong numbers are zero by a proper morphism is again a $(1,1)$-current whose all Lelong numbers are zero by \cite[Corollary 4]{Fav99}, the pullback of any $(1,1)$-current with generalised algebraic singularities is again a $(1,1)$-current with generalised algebraic singularities. The class of metrics (or currents) with generalised algebraic singularities contains several other important classes of metrics: metrics with algebraic singularities, metrics whose all Lelong numbers are zero (in particular, all hermitian semipositive metrics are in this class), as well as currents of integration along effective divisors. \begin{rem}\label{rem:WZD} In the notation from Definition \ref{dfn:metrics}, the current $\Theta$ is nef by \cite[Corollary 6.4]{Dem92}. Therefore, $D$ has a weak Zariski decomposition in the sense of \cite[Definition 1.3]{Bir12b}. \end{rem} The following lemma allows to calculate the multiplier ideal of a metric with generalised algebraic singularities. \begin{lem} Let $X$ be a complex manifold and let $\varphi$ and $\psi$ be two plurisubharmonic functions on $X$. Let $x$ be a point in $X$ and assume that the Lelong number of $\varphi$ at $x$ is zero. Then $$\mathcal I(\varphi+\psi)_x=\mathcal I(\psi)_x.$$ In particular, let $\mathcal L$ be a line bundle on a projective manifold $X$ and let $h$ be a singular metric on $\mathcal L$ with semipositive curvature current whose Siu decomposition is $$\Theta_h(\mathcal L)=\Theta+\sum\lambda_i[D_i],$$ where all Lelong numbers of $\Theta$ are zero and $\sum D_i$ is a simple normal crossings divisor. Then $$\textstyle\mathcal I(h^{\otimes m})=\mathcal{O}_Y\big({-}\sum\lfloor m\lambda_j\rfloor D_j\big)\quad\text{for every positive integer }m.$$ \end{lem} \begin{proof} Since $\varphi $ is locally bounded from above, we have $$\mathcal I(\varphi+\psi)_x\subseteq\mathcal I(\psi)_x,$$ and it suffices to show the reverse inclusion; note that by \cite[Theorem 2.6(ii)]{DEL00} one has even the stronger inclusion $$ \mathcal I(\varphi+\psi)_x\subseteq \mathcal I(\varphi)_x \cdot \mathcal I(\psi)_x.$$ By \cite{GZ15} there exists a positive rational number $p>1$ such that $\mathcal I(p\psi)=\mathcal I(\psi)$, and let $q>1$ be a rational number such that $\frac1p+\frac1q=1$. Fix $f\in\mathcal I(p\psi)_x$, and we may assume that $f\in\mathcal I(p\psi)(U)$ for an open set $x\in U\subseteq X$. Since $q\varphi$ has the Lelong number zero at $x$, we have $\mathcal I(q\varphi)_x=\mathcal{O}_{X,x}$ by Skoda's lemma \cite[Lemma 5.6]{De12}, hence we may assume that $|f|^2e^{-2q\varphi}$ is integrable on $U$. Therefore, H\"older's inequality gives $$\int_U|f|^2e^{-2(\varphi+\psi)}\leq\left(\int_U|f|^2e^{-2q\varphi}\right)^\frac{1}{q}\left(\int_U|f|^2e^{-2p\psi}\right)^\frac{1}{p}<\infty,$$ hence $f\in\mathcal I(\varphi+\psi)(U)$, as desired. \end{proof} The following results connect metrics with generalised algebraic singularities with birational geometry. \begin{lem}\label{lem:Siu} Let $f\colon X\to Y$ be a surjective morphism from a smooth projective variety $X$ to a normal projective variety $Y$. Let $D$ be a pseudoeffective $\mathbb{Q}$-Cartier divisor on $Y$ and let $G=f^*D+\sum e_iE_i$, where $E_i$ are $f$-exceptional prime divisors on $X$ and $e_i\geq0$ are rational numbers. Let $T$ be a closed semipositive current in $c_1(G)$. Then $T-\sum e_i[E_i]$ is a semipositive current. \end{lem} \begin{proof} Let $T_{\min}$ be a closed semipositive current with minimal singularities in $c_1(G)$, see for instance \cite[\S2.8]{Bou04}. For each $i$, let $\lambda_{\min,i}$ and $\lambda_i$ denote the Lelong numbers of $T_{\min}$, respectively $T$, along $E_i$. Then we have $$\sigma_{E_i}(G)\leq \lambda_{\min,i}$$ by \cite[Proposition 3.6(i)]{Bou04}, and $\lambda_{\min,i}\leq \lambda_i$ by the definition of the current with minimal singularities. On the other hand, we have $$e_i\leq \sigma_{E_i}(G)$$ by \cite[Lemma 2.16]{GL13}. The lemma follows by combining all this with the Siu decomposition of $T$. \end{proof} \begin{cor}\label{cor:algsinMMP} Let $X$ be a normal projective variety and let $D$ be a $\mathbb{Q}$-Cartier $\mathbb{Q}$-divisor on $X$. Let $f\colon X\dashrightarrow Y$ be a birational contraction such that $f_*D$ is $\mathbb{Q}$-Cartier, and assume that $f$ is $D$-non-positive. If $D$ has a singular metric with generalised algebraic singularities and semipositive curvature current, then $f_*D$ has a singular metric with generalised algebraic singularities and semipositive curvature current. \end{cor} \begin{proof} Let $(p,q)\colon W\to X\times Y$ be a resolution of indeterminacies of $f$ such that $W$ is smooth. Then there exist $q$-exceptional prime divisors $E_i$ on $W$ and rational numbers $e_i\geq0$ such that \begin{equation}\label{eq:compare} p^*D\sim_\mathbb{Q} q^*f_*D+\sum e_iE_i. \end{equation} We may assume that the given metric with generalised algebraic singularities on $D$ descends to $W$. Then the associated curvature current $T$ has the Siu decomposition of the form $$T=\Theta+\sum\lambda_i [D_i],$$ where $\Theta\geq0$ has all Lelong numbers zero, $\sum D_i$ is a simple normal crossings divisor and $\lambda_i$ are positive rational numbers for all $i$. By \eqref{eq:compare} and by Lemma \ref{lem:Siu} we have $\sum\lambda_i D_i\geq \sum e_iE_i$, thus $$\Theta+\sum \lambda_i [D_i]-\sum e_i[E_i]$$ is a semipositive curvature current in $c_1(q^*f_*D)$. This proves the result. \end{proof} We use below crucially the following hard Lefschetz theorem from \cite[Theorem 0.1]{DPS01}. \begin{thm}\label{thm:DPS} Let $X$ be a compact K\"ahler manifold of dimension $n$ with a K\"ahler form $\omega$. Let $\mathcal L$ be a pseudoeffective line bundle on $X$ with a singular hermitian metric $h$ such that $\Theta_h(\mathcal L) \geq 0$. Then for every nonnegative integer $q$ the morphism \[ \xymatrix{ H^0\big(X,\Omega^{n-q}_X\otimes\mathcal L\otimes\mathcal I(h)\big) \ar[r]^{\omega^q\wedge\bullet} & H^q\big(X, \Omega^n_X\otimes \mathcal L\otimes\mathcal I(h)\big) } \] is surjective. \end{thm} \section{Reduction to the Weak~Generalised~Nonvanishing~Conjecture}\label{sec:reductions} In this section we prove Theorem \ref{thm:mainreduction}. In other words, in order to prove the Generalised Nonvanishing Conjecture, we may assume that $X$ is not uniruled, that $K_X +\Delta$ is nef, and that the nef dimension of $K_X+\Delta+L$ is maximal. We achieve this in several steps. An important step is the proof of the following result announced in the introduction. \begin{thm} \label{thm:B} Assume the termination of flips in dimensions at most $n-1$, the Abundance Conjecture in dimensions at most $n-1$, and the Generalised Nonvanishing Conjecture in dimensions at most $n-1$. Let $(X,\Delta)$ be a projective klt pair of dimension $n$ with $K_X+\Delta$ pseudoeffective, and let $L$ be a nef Cartier divisor on $X$ such that $K_X+\Delta+L$ is nef and $n(X,K_X+\Delta+L)<n$. Then $K_X+\Delta+tL$ is num-effective for every $t\geq0$. \end{thm} The proof of Theorem \ref{thm:B} is similar to the proof of \cite[Theorem 5.3]{LP18a}, and the proof of Theorem \ref{thm:Support} follows the strategy of \cite{DL15}. \medskip We start with the following lemma. \begin{lem}\label{lem:termination} Assume the Weak Generalised Nonvanishing Conjecture in dimension $n$, the existence of minimal models of klt pairs in dimension $n$, the Abundance Conjecture in dimensions at most $n-1$, and the termination of flips in dimension $n-1$. Assume Theorem \ref{thm:B} in dimension $n$. Let $(X,\Delta)$ be a klt pair of dimension $n$ such that $K_X+\Delta$ is pseudoeffective. Then $\kappa(X,K_X+\Delta)\geq0$ and every sequence of flips of $(X,\Delta)$ terminates. \end{lem} \begin{proof} By \cite[Corollary 1.2]{HMX14} it suffices to show that $\kappa(X,K_X+\Delta)\geq0$. By our assumptions, there exists a minimal model of $(X,\Delta)$. Hence, replacing $(X,\Delta)$ by this minimal model, we may assume that $K_X+\Delta$ is nef. By Theorem \ref{thm:B} and by \cite[Theorem 0.1]{CKP12} we have $\kappa(X,K_X+\Delta)\geq0$ if $n(X,K_X+\Delta)<n$. Therefore, we may assume that $n(X,K_X+\Delta)=n$. If $X$ is not uniruled, then we conclude by the Weak Nonvanishing Conjecture. If $X$ is uniruled, by passing to a terminalisation of $(X,\Delta)$ we may assume that $(X,\Delta)$ is terminal. Then $K_X$ is not pseudoeffective, and we conclude that $\kappa(X,K_X+\Delta)\geq0$ by \cite[Theorem 3.3]{DL15}. \end{proof} \subsection{Achieving nefness of $K_X+\Delta+L$}\label{subsec:nefness} The first step is to achieve the nefness of $K_X+\Delta+L$. \begin{pro}\label{pro:contocon} Let $(X,\Delta)$ be a projective klt pair of dimension $n$ such that $K_X+\Delta$ is pseudoeffective. Assume that any sequence of flips of the pair $(X,\Delta)$ terminates. Let $L$ be a nef Cartier divisor on $X$ and let $m>2n$ be a positive integer. Then there exists an $L$-trivial $(K_X+\Delta)$-MMP $\varphi\colon X\dashrightarrow Y$ such that $K_Y+\varphi_*\Delta+m\varphi_*L$ is nef. \end{pro} \begin{proof} Apart from the different assumptions, this is \cite[Proposition 4.1]{LP18a}. The only thing to note is that the proof of that result uses the termination of flips only for the given pair $(X,\Delta)$, and not for all klt pairs. \end{proof} Therefore, we obtain: \begin{pro} \label{pro:reduction} Assume the Weak Nonvanishing Conjecture in dimension $n$, the existence of minimal models of klt pairs in dimension $n$, the Abundance Conjecture (for klt pairs) in dimensions at most $n-1$, and the termination of flips in dimension $n-1$. Assume Theorem \ref{thm:B} in dimension $n$. In order to prove Theorem \ref{thm:mainreduction}, it suffices to assume that $L$ is Cartier and $K_X + \Delta + L$ is nef. \end{pro} \begin{proof} First note that any sequence of flips of the pair $(X,\Delta)$ terminates by Lemma \ref{lem:termination}. Then by Proposition \ref{pro:contocon} there exists an $L$-trivial $(K_X+\Delta)$-MMP $\varphi\colon X\dashrightarrow Y$ such that $K_Y+\varphi_*\Delta+m\varphi_*L$ is nef for some fixed $m\gg0$. Denote $\Delta_Y:=\varphi_*\Delta$ and $L_Y:=m\varphi_*L$. Then $K_Y+\Delta_Y+L_Y$ is nef and by Lemma \ref{lem:MMPnumeff} it suffices to show that $K_Y+\Delta_Y+tL_Y$ is num-effective for every $t\geq0$. \end{proof} \subsection{Non-maximal nef dimension}\label{subsec:nonmaxnefdim} We first deal with the case of non-maximal nef dimension -- we prove Theorem \ref{thm:B}. More precisely, we have the following inductive step. \begin{thm}\label{thm:induction} Assume the termination of flips in dimensions at most $n-1$, the Abundance Conjecture in dimensions at most $n-1$, and the Weak Generalised Nonvanishing Conjecture in dimensions at most $n-1$. Then Theorem \ref{thm:mainreduction} in dimension $n-1$ implies Theorem \ref{thm:B} in dimension $n$. \end{thm} \begin{proof} Let $(X,\Delta)$ be a projective klt pair of dimension $n$ with $K_X+\Delta$ pseudoeffective, and let $L$ be a nef Cartier divisor on $X$ such that $K_X+\Delta+L$ is nef and $n(X,K_X+\Delta+L)<n$. We need show that $K_X+\Delta+tL$ is num-effective for every $t\geq0$. \medskip \emph{Step 1.} In this step we show that we may assume the following: \medskip \emph{Assumption 1.} There exist a morphism $\varphi\colon X\to Y$ to a smooth projective variety $Y$ with $\dim Y<\dim X$ and a nef $\mathbb{Q}$-divisor $L_Y$ on $Y$ such that $L\equiv \varphi^*L_Y$ and $\nu\big(X,(K_X+\Delta)|_F\big)=0$ for a general fibre $F$ of $\varphi$. {\it However, we may not any more assume that $K_X+\Delta+L$ is nef.} \medskip To this end, first note that by passing to a small $\mathbb{Q}$-factorialisation and by \cite[Lemma 2.14]{LP18a}, we may assume first that $X$ is $\mathbb{Q}$-factorial. Fix a positive integer $m>2n$. Then $$n(X,K_X+\Delta+mL)<n$$ by Lemma \ref{lem:nefredsum}. Let $$\varphi\colon X\dashrightarrow Y$$ be the nef reduction of $K_X+\Delta+mL$. Then $\dim Y = n(X,K_X+\Delta+mL) <n$, and $K_X+\Delta+mL$ is numerically trivial on a general fibre of $\varphi$. By \cite[Lemma 2.8]{LP18a} this implies that both $K_X+\Delta$ and $L$ are numerically trivial on a general fibre of $\varphi$. Recall that $\varphi$ is almost holomorphic. Let $(\widehat\pi,\widehat\varphi)\colon\widehat{X}\to X\times Y$ be a smooth resolution of indeterminacies of the map $\varphi$. Write $$K_{\widehat{X}}+\widehat{\Delta}\sim_\mathbb{Q}\widehat{\pi}^*(K_X+\Delta)+\widehat{E},$$ where $\widehat{\Delta}$ and $\widehat{E}$ are effective $\mathbb{Q}$-divisors without common components. Then if $\widehat F$ is a general fibre of $\widehat \varphi$, then $\nu\big(\widehat F,(K_{\widehat{X}}+\widehat{\Delta})|_{\widehat{F}}\big)=0$ by Lemma \ref{lem:1}, and $\widehat{L}:=\widehat{\pi}^*L$ is numerically trivial on $\widehat{F}$. By \cite[Lemma 3.1]{LP18a} there exist a birational morphism $\pi\colon Y'\to Y$ from a smooth projective variety $Y'$, a nef $\mathbb{Q}$-divisor $L_{Y'}$ on $Y'$, a smooth projective variety $X'$ and a commutative diagram \[ \xymatrix{ X' \ar[d]_{\varphi'} \ar[r]^{\pi'} & \widehat X \ar[d]^{\widehat\varphi}\\ Y' \ar[r]_{\pi} & Y } \] such that, if we set $L':=\pi'^*\widehat L$, then $$L'\equiv \varphi'^*L_{Y'}.$$ We may write $$K_{X'}+\Delta'\sim_\mathbb{Q}\pi'^*\big(K_{\widehat{X}}+\widehat\Delta\big)+E',$$ where $\Delta'$ and $E'$ are effective $\mathbb{Q}$-divisors on $X'$ without common components. Then if $F'$ is a general fibre of $\varphi'$, then $\nu\big(F',(K_{X'}+\Delta')|_{F'}\big)=0$ by Lemma \ref{lem:1}. Since $$K_{X'}+\Delta'+tL'\sim_\mathbb{Q}\pi'^*\widehat{\pi}^*(K_X+\Delta+tL)+\pi'^*\widehat{E}+E'$$ for $t\geq0$, and since $X$ is $\mathbb{Q}$-factorial, it suffices to show that $K_{X'}+\Delta'+tL'$ is num-effective for any $t\geq0$. Therefore, by replacing $X$ by $X'$, $\Delta$ by $\Delta'$, $L$ by $L'$, and $\varphi$ by $\varphi'$, we achieve Assumption 1. \medskip \emph{Step 2.} We now run any relative $(K_X+\Delta)$-MMP with scaling of an ample divisor over $Y$, and note that this MMP is $L$-trivial by Assumption 1. Since $\nu\big(X,(K_X+\Delta)|_F\big)=0$ for a general fibre $F$ of $\varphi$ by Assumption 1, the pair $(F,\Delta|_F)$ has a good model by \cite[Corollaire 3.4]{Dru11} and \cite[Corollary V.4.9]{Nak04}. Therefore, by \cite[Theorem 2.12]{HX13}, this MMP terminates with a relative good model $X_{\min}$ of $(X,\Delta)$ over $Y$. Denote by $$\theta\colon X\dashrightarrow X_{\min}$$ the corresponding birational morphism and by $$\varphi_{\min}\colon X_{\min}\to Y$$ the induced morphism to $Y$. Set $\Delta_{\min}:=\theta_*\Delta$ and $L_{\min}:=\theta_*L$. Let $$\tau\colon X_{\min}\to T$$ be the map to the relative canonical model $T$ over $Y$, with induced morphism $\varphi_T\colon T\to Y$, yielding the commutative diagram \[ \xymatrix{ X \ar[dr]_{\varphi} \ar@{-->}[rr]^{\theta} && X_{\min} \ar@/^0.5pc/[d]^{\tau} \ar[dl]^{\varphi_{\min}}\\ & Y\ar@{<-}[r]_{\varphi_T} & T. } \] Notice that $\dim T<n$ by Assumption 1. Moreover, there exists a $\varphi_T$-ample $\mathbb{Q}$-divisor $A$ on $T$ such that $K_{X_{\min}}+\Delta_{\min}\sim_\mathbb{Q}\tau^*A$. By \cite[Theorem 0.2]{Amb05a} there exists an effective $\mathbb{Q}$-divisor $\Delta_T$ on $T$ such that the pair $(T,\Delta_T)$ is klt and such that $$K_{X_{\min}}+\Delta_{\min}\sim_\mathbb{Q}\tau^*(K_T+\Delta_T).$$ Therefore, since $\theta$ is $L$-trivial as noted above, for any $t\geq0$ we have $$K_{X_{\min}}+\Delta_{\min}+tL_{\min}\sim_\mathbb{Q}\tau^*(K_T+\Delta_T+t\varphi_T^*L_Y).$$ By the assumptions of the theorem and since we assume Theorem \ref{thm:mainreduction} in dimension $n-1$, the Generalised Nonvanishing Conjecture holds in dimensions at most $n-1$. Therefore, for each $t\geq0$ the divisor $K_T+\Delta_T+t\varphi_T^*L_Y$ is num-effective, and therefore, the previous equation shows that for every $t\geq0$ the divisor $K_{X_{\min}}+\Delta_{\min}+tL_{\min}$ is num-effective. But then each $K_X+\Delta+tL$ is num-effective by Lemma \ref{lem:MMPnumeff}. \end{proof} \subsection{An auxiliary result} \begin{thm}\label{thm:Support} Assume the termination of flips in dimensions at most $n-1$, and the Weak Generalised Nonvanishing Conjecture in dimension $n$. Let $(X,\Delta)$ be an $n$-dimensional projective $\mathbb{Q}$-factorial terminal pair such that $K_X+\Delta$ is pseudoeffective, and let $L$ be a nef $\mathbb{Q}$-divisor on $X$. Assume that there exists an effective $\mathbb{Q}$-divisor $D$ such that $$K_X + \Delta \sim_\mathbb{Q} D \geq 0 \quad \text{and} \quad \Supp\Delta\subseteq\Supp D.$$ Then there exists a rational $t_0>0$ such that $K_X+\Delta+tL$ is num-effective for every $0<t\leq t_0$. \end{thm} \begin{proof} We follow the arguments of \cite{DL15}. \medskip \emph{Step 1.} We need to show that there exists a rational $t_0>0$ such that \begin{equation}\label{eq:304} D+tL\text{ is num-effective for every }0<t\leq t_0. \end{equation} Let $f\colon X'\to X$ be a log resolution of the pair $(X,D)$. Then we may write \begin{equation}\label{eq:305} K_{X'}+\Delta_1\sim_\mathbb{Q} f^*(K_X+\Delta)+F', \end{equation} where $\Delta_1$ and $F'$ are effective $\mathbb{Q}$-divisors with no common components, and $\Delta_1=f^{-1}_*\Delta$ since $(X,\Delta)$ is a terminal pair. For a rational number $0<\varepsilon\ll1$, denote $\Delta_2:=\Delta_1+\varepsilon(f^*D+F')$ and $D_2:=(1+\varepsilon)(f^*D+F')$. Then the assumptions of the theorem and \eqref{eq:305} give $$K_{X'}+\Delta_2\sim_\mathbb{Q} D_2\geq0\quad \text{with}\quad \Supp\Delta_2=\Supp D_2,$$ where the pair $(X',\Delta_2)$ is terminal and log smooth. Denote $L':=f^*L$. Then $$\textstyle D_2+tL'\sim_\mathbb{Q} (1+\varepsilon)f^*\big(D+\frac{t}{1+\varepsilon}L\big)+(1+\varepsilon)F'.$$ Therefore, since $F'$ is effective and $f$-exceptional, by \eqref{eq:304} it suffices to show that there exists a rational $t_2>0$ such that \begin{equation}\label{eq:306} D_2+tL'\text{ is num-effective for every }0<t\leq t_2. \end{equation} Set $\Delta':=\lceil\Delta_2\rceil$ and $D':=D_2+\lceil\Delta_2\rceil-\Delta_2$. Then $$K_{X'}+\Delta'\sim_\mathbb{Q} D'\geq0\quad \text{with}\quad \Supp\Delta'=\Supp D',$$ where the pair $(X',\Delta')$ is log smooth and $\Delta'$ is a reduced divisor. By \eqref{eq:306} and by Lemma \ref{lem:easylemma}, it suffices to show that there exists a rational $t_0'$ such that \begin{equation}\label{eq:307} D'+tL'\text{ is num-effective for every }0<t\leq t_0'. \end{equation} \medskip \emph{Step 2.} Let $m$ be the smallest positive integer such that $$m(K_{X'}+\Delta')\sim mD',$$ and denote $G'=mD'$. Let $\pi\colon X''\to X'$ be the normalisation of the corresponding $m$-fold cyclic covering ramified along $G'$. The variety $X''$ is irreducible by \cite[Lemma 3.15(a)]{EV92} since $m$ is minimal. Then there exists an effective Cartier divisor $D_1''$ on $X''$ such that \begin{equation}\label{eq:2} \pi^*G'=mD_1''\quad\text{and}\quad\pi^*(K_{X'}+\Delta')\sim D_1'', \end{equation} and let $\Delta_1''=(D_1'')_\textrm{red}$. By the Hurwitz formula, we have \begin{equation}\label{eq:2a} K_{X''}+\Delta_1''=\pi^*(K_{X'}+\Delta'). \end{equation} Pick a rational number $0<\delta\ll1$ such that $$\Delta'':=\Delta_1''-\delta\pi^*\Delta'\geq0,\ D'':=D_1''-\delta\pi^*\Delta'\geq0\text{ and } \Supp D''=\Supp D_1''.$$ Then $$K_{X''}+\Delta''=\pi^*\big(K_{X'}+(1-\delta)\Delta'\big)\quad\text{and}\quad K_{X''}+\Delta''\sim_\mathbb{Q} D'',$$ and therefore, the pair $(X'',\Delta'')$ is klt by \cite[Proposition 5.20]{KM98}. Denote $L'':=\pi^*L'$. By \eqref{eq:307} and by \cite[Lemma 2.15]{LP18a} it suffices to show that there exists a rational $t_1''>0$ such that $$D_1''+tL''\text{ is num-effective for every }0<t\leq t_1''.$$ Since $\Supp D''=\Supp D_1''$, by Lemma \ref{lem:easylemma} it suffices to show that there exists a rational $t_0''>0$ such that \begin{equation}\label{eq:308} D''+tL''\text{ is num-effective for every }0<t\leq t_0''. \end{equation} \medskip \emph{Step 3.} We claim that $X''$ is not uniruled and that $\kappa(X'',K_{X''})\geq0$. Assuming the claim, we show first that it implies the theorem. The assumptions of the theorem and \cite[Corollary 1.2]{HMX14} then imply that any sequence of flips of the pair $(X'',\Delta'')$ terminates. Therefore, by Proposition \ref{pro:contocon} there exists an $L''$-trivial $(K_{X''}+\Delta'')$-MMP $\varphi\colon X''\dashrightarrow Y$ such that $K_Y+\varphi_*\Delta''+m\varphi_*L''$ is nef for some fixed $m\gg0$. Denote $\Delta_Y:=\varphi_*\Delta''$ and $L_Y:=\varphi_*L''$. By \eqref{eq:308} and by Lemma \ref{lem:MMPnumeff} it suffices to show that there exists a rational $t_0^Y>0$ such that \begin{equation}\label{eq:3c} K_Y+\Delta_Y+tL_Y\text{ is num-effective for every }0<t\leq t_0^Y. \end{equation} As above, any sequence of flips of the pair $(Y,\Delta_Y)$ terminates. As in Step 1 and the beginning of Step 2 of the proof of \cite[Lemma 5.2]{LP18a}, there is a birational contraction $$\theta\colon (Y,\Delta_Y)\dashrightarrow (Y_{\min},\Delta_{\min})$$ which is a composition of a sequence of operations of a $(K_Y+\Delta_Y)$-MMP, there is a positive rational number $\lambda$ and a divisor $L_{\min}:=\theta_*(mL_Y)$ on $Y_{\min}$ such that: \begin{enumerate} \item[(a)] $K_{Y_{\min}}+\Delta_{\min}$ is nef and $s(K_{Y_{\min}}+\Delta_{\min})+L_{\min}$ is nef for all $s\geq\lambda$, \item[(b)] the map $\theta$ is $\big(s(K_Y+\Delta_Y)+mL)$-negative for $s>\lambda$. \end{enumerate} Denote $M:=\lambda(K_{Y_{\min}}+\Delta_{\min})+L_{\min}$; thus $M$ is nef by (a). By the Weak Generalised Nonvanishing Conjecture, the divisor $$K_{Y_{\min}}+\Delta_{\min}+tM=t\big((\lambda+1/t)(K_{Y_{\min}}+\Delta_{\min})+L_{\min}\big)$$ is num-effective for every $t>0$. Then by (b) and by Lemma \ref{lem:MMPnumeff}, the divisor $$K_Y+\Delta_Y+tL_Y\quad\text{is num-effective}$$ for all $0<t\leq m/\lambda$. This proves (i) and finishes the proof of the theorem. \medskip \emph{Step 4.} It remains to prove the claim from Step 3. Let $g\colon W\to X''$ be a log resolution of the pair $(X'',\Delta_1'')$. By \eqref{eq:2} and \eqref{eq:2a} we may write \begin{equation}\label{eq:312} K_W+\Delta_W\sim_\mathbb{Q} g^*(K_{X''}+\Delta_1'')+E_W\sim g^*D_1''+E_W, \end{equation} where $\Delta_W$ and $E_W$ are effective $\mathbb{Q}$-divisors with no common components. We claim that $\Delta_W$ and $E_W$ are \emph{integral} divisors. Let $E''$ be a geometric valuation over $X''$. Then by \cite[Proposition 2.14]{DL15}, there exists a geometric valuation $E'$ over $X'$ and an integer $1\leq r\leq m$ such that $$a(E'',X'',\Delta_1'')+1=r\big(a(E',X',\Delta')+1\big).$$ Since $(X',\Delta')$ is log smooth and $\Delta'$ is reduced, we have $a(E',X',\Delta')\in\mathbb{Z}$, hence $a(E'',X'',\Delta_1'')\in\mathbb{Z}$, which proves the claim. \medskip Therefore, the divisor $$G_W:=g^*D_1''+E_W-\Delta_W$$ is Cartier, and by \eqref{eq:312} we have $$K_W\sim_\mathbb{Q} G_W.$$ Then it suffices to show that \begin{equation}\label{eq:309} G_W\geq0, \quad\text{and in particular,}\quad \kappa(W,K_W)\geq0. \end{equation} Indeed, this implies that $W$ is not uniruled. Since $K_{X''}\sim_\mathbb{Q} g_*K_W$, the claim from Step 3 follows. In order to show \eqref{eq:309}, fix a component $S''$ of $\Delta_W$. Then $a(S'',X'',\Delta_1'')=-1$. By \cite[Proposition 2.14]{DL15}, there exists a geometric valuation $S'$ over $X'$ and an integer $1\leq r\leq m$ such that $\pi\big(c_{X''}(S'')\big)=c_{X'}(S')$ and $$a(S'',X'',\Delta_1'')+1=r\big(a(S',X',\Delta')+1\big).$$ Here, $c_{X''}(S'')$ denotes the center of $S''$ on $X''$, and analogously for $c_{X'}(S')$. This implies $a(S',X',\Delta')=-1$, thus $c_{X'}(S')\subseteq\Supp\Delta'$ because $(X',\Delta')$ is log smooth. From here we obtain $c_{X''}(S'')\subseteq\pi^{-1}(\Supp\Delta')=\Supp D_1''$, and in particular, $S''\subseteq\Supp g^*D_1''$. Therefore, $$\mult_{S''}g^*D_1''\geq1=\mult_{S''}\Delta_W.$$ Now \eqref{eq:309} follows by the definition of $G_W$. This finishes the proof. \end{proof} \subsection{Non-pseudoeffective canonical class} \begin{lem}\label{lem:Knotpsef} Assume the termination of flips in dimensions at most $n-1$, the Abundance Conjecture in dimensions at most $n-1$, the existence of minimal models in dimension $n$, and the Weak Generalised Nonvanishing Conjecture in dimension $n$. Assume Theorem \ref{thm:B} in dimension $n$. Let $(X,\Delta)$ be an $n$-dimensional projective $\mathbb{Q}$-factorial terminal pair such that $K_X+\Delta$ is pseudoeffective, and let $L$ be a nef $\mathbb{Q}$-divisor on $X$. Assume that $K_X$ is not pseudoeffective. Then $\kappa(X,K_X+\Delta)\geq0$ and there exists a rational $t_0>0$ such that $K_X+\Delta+tL$ is num-effective for every $0< t\leq t_0$. \end{lem} \begin{proof} Define $$\tau:=\min\{t\in\mathbb{R}\mid K_X+t\Delta\text{ is pseudoeffective}\}.$$ Observe that $0<\tau\leq1$ by our assumptions. Then by \cite[Theorem 3.3]{DL15} there exists a $\mathbb{Q}$-divisor $D_\tau\geq0$ such that $K_X+\tau\Delta\sim_\mathbb{Q} D_\tau$. This yields $$K_X + \Delta \sim_\mathbb{Q} D \geq 0 , \quad \text{where} \quad D =D _\tau + (1-\tau )\Delta.$$ Pick a rational number $0<\mu\ll1$ such that the pair $\big(X,(1+\mu)\Delta\big)$ is terminal, and denote $\Delta':=(1+\mu)\Delta$. Observe that $$K_X+\Delta'\sim_\mathbb{Q} D+\mu\Delta\quad\text{and}\quad \Supp\Delta'\subseteq\Supp(D+\mu\Delta).$$ Then by Theorem \ref{thm:Support} there exist a rational number $\lambda>0$ and an effective $\mathbb{Q}$-divisor $G$ such that \begin{equation}\label{eq:121a} K_X+(1+\mu)\Delta+\lambda L\equiv G. \end{equation} Pick a rational number $0<\delta\ll1$ such that the pair $\big(X,(1-\delta)\Delta+\frac{\delta}{\mu}G\big)$ is terminal, and denote $\Delta_X:=(1-\delta)\Delta+\frac{\delta}{\mu}G$. Then by \eqref{eq:121a} we have $$\textstyle K_X+\Delta_X\equiv \big(1+\frac{\delta}{\mu}\big)\big(K_X+\Delta+\frac{\delta\lambda}{\delta+\mu}L\big).$$ In particular, $K_X+\Delta_X$ is pseudoeffective. Since then $\kappa(X,K_X+\Delta_X)\geq0$ by Lemma \ref{lem:termination}, the result follows. \end{proof} \subsection{Proofs of Theorems \ref{thm:mainreduction} and \ref{thm:B}}\label{subsec:mainreduction} We finally have all the ingredients to prove Theorems \ref{thm:mainreduction} and \ref{thm:B}. \begin{thm}\label{thm:induction2} Assume the termination of flips in dimensions at most $n-1$, the Abundance Conjecture in dimensions at most $n-1$, the existence of minimal models in dimension $n$, and the Weak Generalised Nonvanishing Conjecture in dimension $n$. Then Theorem \ref{thm:B} in dimension $n$ implies Theorem \ref{thm:mainreduction} in dimension $n$. \end{thm} \begin{proof} Let $(X,\Delta)$ be an $n$-dimensional projective klt pair such that $K_X+\Delta$ is pseudoeffective, and let $L$ be a nef $\mathbb{Q}$-divisor on $X$. By Proposition \ref{pro:reduction} and by Lemma \ref{lem:MMPnumeff} we may assume that $K_X+\Delta+L$ is nef. \medskip \emph{Step 1.} If $n(X,K_X+\Delta+L)<n$, then we conclude by Theorem \ref{thm:B}. \medskip \emph{Step 2.} Therefore, from now on we may assume that $$n(X,K_X+\Delta+L)=n.$$ We first prove that in this case part (i) of Theorem \ref{thm:mainreduction} implies part (ii) of Theorem \ref{thm:mainreduction}. Indeed, then there exists an effective $\mathbb{Q}$-divisor $N$ such that $$K_X+\Delta+t_0L\equiv N.$$ Pick a small positive rational number $\varepsilon$ such that $(X,\Delta+\varepsilon N)$ is klt. We have $$\kappa(X,K_X+\Delta+\varepsilon N)\geq0$$ by (i), hence by \cite[Corollary 1.2]{HMX14} there exists a $(K_X+\Delta+\varepsilon N)$-MMP $$\varphi\colon X\dashrightarrow Y$$ which terminates. The divisor $K_Y+\varphi_*\Delta+\varepsilon \varphi_*N$ is semiample, since we are assuming the semiampleness part of the Abundance Conjecture in dimension $n$. Since $n(X,K_X+\Delta+L)=n$, we have $$(K_X+\Delta+\varepsilon N)\cdot C>0$$ for every curve $C$ on $X$ passing through a very general point on $X$ by Lemma \ref{lem:nefredsum}, and therefore $$n(Y,K_Y+\varphi_*\Delta+\varepsilon\varphi_*N)=n$$ by Corollary \ref{cor:nefdim}. But then $K_Y+\varphi_*\Delta+\varepsilon\varphi_*N$ is big, hence $K_X+\Delta+\varepsilon N$ is also big. As $$K_X+\Delta+\varepsilon N\equiv(1+\varepsilon)(K_X+\Delta)+\varepsilon t_0 L,$$ and $K_X+\Delta$ and $L$ are both pseudoeffective, it is easy to see that $K_X+\Delta+tL$ is big for all $t>0$. \medskip \emph{Step 3.} It remains to prove (i). In this step we show that we may additionally assume the following: \medskip \emph{Assumption 1.} The pair $(X,\Delta)$ is $\mathbb{Q}$-factorial and terminal, and $K_X+\Delta$ is nef. \medskip To this end, let $\pi\colon \widetilde X\to X$ be a small $\mathbb{Q}$-factorialisation. As $K_{\widetilde X}+\pi_*^{-1}\Delta\sim_\mathbb{Q}\pi^*(K_X+\Delta)$, by replacing $X$ by $\widetilde X$, $\Delta$ by $\pi_*^{-1}\Delta$ and $L$ by $\pi^*L$, we may assume that $X$ is $\mathbb{Q}$-factorial by \cite[Lemma 2.14]{LP18a}. By Lemma \ref{lem:termination}, we have $$\kappa(X,K_X+\Delta)\geq0$$ and any sequence of flips of the pair $(X,\Delta)$ terminates. Then as in Step 1 and the beginning of Step 2 of the proof of \cite[Lemma 5.2]{LP18a}, there is a birational contraction $$\theta\colon (X,\Delta)\dashrightarrow (X_{\min},\Delta_{\min})$$ which is a composition of a sequence of operations of a $(K_X+\Delta)$-MMP, a positive rational number $\lambda$ and a divisor $L_{\min}:=\theta_*L$ on $X_{\min}$ such that: \begin{enumerate} \item[(a)] $K_{X_{\min}}+\Delta_{\min}$ is nef and $s(K_{X_{\min}}+\Delta_{\min})+L_{\min}$ is nef for all $s\geq\lambda$, \item[(b)] the map $\theta$ is $\big(s(K_X+\Delta)+L)$-negative for $s>\lambda$, \item[(c)] $(K_{X_{\min}}+\Delta_{\min}+tL_{\min})\cdot C>0$ for every curve $C$ on $X_{\min}$ passing through a very general point on $X_{\min}$ and for every $t>0$. \end{enumerate} Denote $$M:=\lambda(K_{X_{\min}}+\Delta_{\min})+L_{\min}.$$ Then $M$ is nef by (a) and $$n(X_{\min},K_{X_{\min}}+\Delta_{\min}+tM)=n\quad\text{for all }t>0$$ by (c) and by Corollary \ref{cor:nefdim}. Assume that there exists a rational $t_0^{\min}>0$ such that $K_{X_{\min}}+\Delta_{\min}+tM$ is num-effective for every $0<t\leq t_0^{\min}$. Since $$\textstyle K_{X_{\min}}+\Delta_{\min}+tM=(t\lambda+1)\big(K_{X_{\min}}+\Delta_{\min}+\frac{t}{t\lambda+1} L_{\min}\big),$$ the divisor $K_{X_{\min}}+\Delta_{\min}+t L_{\min}$ is num-effective for every $0<t\leq \frac{t_0^{\min}}{t_0^{\min}\lambda+1}$. But then by (b) and by Lemma \ref{lem:MMPnumeff}, the divisor $K_X+\Delta+t L$ is num-effective for every $0<t\leq \frac{t_0^{\min}}{t_0^{\min}\lambda+1}$. Therefore, by replacing $X$ by $X_{\min}$, $\Delta$ by $\Delta_{\min}$ and $L$ by $M$, we achieve that the variety $X$ is $\mathbb{Q}$-factorial, and the divisor $K_X+\Delta$ is nef. Now, let $$\xi\colon (X',\Delta')\to (X,\Delta)$$ be the composition of a terminalisation and a small $\mathbb{Q}$-factorialisation. Since $K_{X'}+\Delta'\sim_\mathbb{Q}\xi^*(K_X+\Delta)$, it follows that, by replacing $X$ by $X'$, $\Delta$ by $\Delta'$ and $L$ by $\xi^*L$, we may additionally assume that $(X,\Delta)$ is a terminal pair, hence we achieve Assumption 1. \medskip \emph{Step 4.} Now, if $K_X$ is pseudoeffective, then $X$ is not uniruled and Theorem \ref{thm:mainreduction}(i) follows from the Weak Generalised Nonvanishing Conjecture. If $K_X$ is not pseudoeffective, then Theorem \ref{thm:mainreduction}(i) follows from Lemma \ref{lem:Knotpsef}. This completes the proof. \end{proof} Theorems \ref{thm:mainreduction} and \ref{thm:B} follow immediately from Theorems \ref{thm:induction} and \ref{thm:induction2}. \section{A criterion for Nonvanishing}\label{sec:MMP} In this section we prove a criterion -- Theorem \ref{thm:nonvanishingForms} -- for effectivity of divisors of the form $K_X+\Delta+L$, where $(X,\Delta)$ is a klt pair such that $K_X+\Delta$ is nef and $X$ is not uniruled, and if $L$ is a nef $\mathbb{Q}$-divisor on $X$. It is crucial for the remainder of the paper. Theorem \ref{thm:nonvanishingForms} contains \cite[Theorem 4.3 and Theorem 8.1]{LP18} as special cases. \begin{thm} \label{thm:nonvanishingForms} Assume the existence of good models for klt pairs in dimensions at most $n-1$. Let $(X,\Delta)$ be a projective terminal pair of dimension $n$ such that $K_X$ pseudoeffective and $K_X+\Delta$ is nef. Let $L$ be a nef Cartier divisor on $X$ and let $t$ be a positive integer such that $t(K_X+\Delta)$ is Cartier. Denote $M=t(K_X+\Delta)+L$, and let $\pi\colon Y\to X$ be a resolution of $X$. Assume that for some positive integer $p$ we have $$ H^0\big(Y,(\Omega^1_Y)^{\otimes p} \otimes \mathcal{O}_Y(m\pi^*M)\big) \neq 0$$ for infinitely many integers $m$. Then $\kappa (X,M) \geq 0$. \end{thm} \begin{proof} Let $\rho\colon X'\to X$ be a small $\mathbb{Q}$-factorialisation of $X$. Then the pair $(X',\rho^{-1}_*\Delta)$ is terminal and $K_{X'}+\rho^{-1}_*\Delta\sim_\mathbb{Q}\rho^*(K_X+\Delta)$. We may assume that $\pi$ factors through $\rho$. By replacing $(X,\Delta)$ by $(X',\rho^{-1}_*\Delta)$ and $L$ by $\rho^*L$, we may thus assume that $X$ is $\mathbb{Q}$-factorial. \medskip If $M\equiv0$, then the assumptions imply that $K_X\equiv0$ (hence $K_X\sim_\Q0$ by \cite[Theorem 8.2]{Kaw85b}), $\Delta=0$ and $L\equiv0$. Then $\kappa(X,M)=\kappa(X,L)\geq0$ by the first part of the proof of \cite[Theorem 8.1]{LP18}. \medskip Therefore, from now on we may assume that $M\not\equiv0$. We apply \cite[Lemma 4.1]{LP18} with $\mathcal E := (\Omega^1_Y)^{\otimes p} $ and $\mathcal L := \pi^*\mathcal{O}_X(M)$. Then there exist a positive integer $r$, a saturated line bundle $\mathcal M$ in $\bigwedge^r\mathcal E$, an infinite set $\mathcal S\subseteq\mathbb{N}$ and integral divisors $N_m\geq0$ for $m\in\mathcal S$ such that \begin{equation}\label{eq:infmany} \mathcal{O}_Y(N_m) \simeq \mathcal M\otimes \mathcal L^{\otimes m}. \end{equation} Since $X$ is terminal and $K_X$ is pseudoeffective, the divisor $K_Y$ is also pseudoeffective, hence \cite[Proposition 4.2]{LP18} implies that there exist a positive integer $\ell$ and a pseudoeffective divisor $F$ such that \begin{equation}\label{eq:relation} N_m+ F \sim m\pi^*M+\ell K_Y. \end{equation} Noting that $\pi_*N_m$ is effective and that $\pi_*F$ is pseudoeffective, by pushing forward the relation \eqref{eq:relation} to $X$ we get $$\pi_*N_m+\pi_*F\sim_\mathbb{Q} mM +\ell K_X,$$ and hence $$\pi_*N_m+(\pi_*F+\ell\Delta)\sim_\mathbb{Q} mM +\ell (K_X+\Delta).$$ Now we conclude by Theorem \ref{thm:MMPtwist}. \end{proof} \begin{thm} \label{thm:MMPtwist} Assume the existence of good models for klt pairs in dimensions at most $n-1$. Let $(X,\Delta)$ be a $\mathbb{Q}$-factorial projective klt pair of dimension $n$ such that $K_X+\Delta$ is nef, and let $L$ be a nef Cartier divisor on $X$. Let $t$ be a positive integer such that $t(K_X+\Delta)$ is Cartier, and denote $M=t(K_X+\Delta)+L$. Assume that there exist a pseudoeffective $\mathbb{Q}$-divisor $F$ on $X$, a positive integer $\ell$ and an infinite subset $\mathcal S\subseteq\mathbb{N}$ such that \begin{equation}\label{eq:rel2a} N_m+F\sim_\mathbb{Q} mM+\ell(K_X+\Delta) \end{equation} for all $m\in\mathcal S$, where $N_m\geq0$ are integral Weil divisors. Then $$\kappa(X,M)=\max\{\kappa(X,N_m)\mid m\in\mathcal S\}\geq0.$$ \end{thm} \begin{proof} Steps 1--4 follow closely the proof of \cite[Theorem 3.1]{LP18}, and we include the details for the benefit of the reader. \medskip Note first that \eqref{eq:rel2a} implies \begin{equation}\label{eq:rel2b} N_p-N_q\sim_\mathbb{Q} (p-q)M\quad\text{for all }p,q\in\mathcal S. \end{equation} It suffices to show that \begin{equation}\label{eq:kappalargem} \kappa(X,M)\geq\kappa(X,N_m)\quad\text{for all large }m\in\mathcal S. \end{equation} Indeed, then $\kappa(X,M)\geq0$, hence \eqref{eq:rel2b} gives $\kappa(X,M)\leq\kappa(X,N_m)$ for all large $m\in\mathcal S$ and $\kappa(X,N_q)\leq \kappa(X,N_p)$ for $p,q\in\mathcal S$ with $q<p$, which implies the theorem. \medskip \emph{Step 1.} We claim that for every $m\in\mathcal S$ and every rational number $\lambda>0$ we have \begin{equation}\label{eq:kodaira} \kappa(X,M+\lambda N_m)\geq\kappa(X,N_m), \end{equation} and in particular, \begin{equation}\label{eq:kodaira3} \kappa(X,M+\lambda N_m)\geq0 \text{ for every }m\in\mathcal S \text{ and every rational }\lambda>0. \end{equation} Indeed, fix $m\in\mathcal S$ and a rational number $\lambda>0$. Pick $m'\in\mathcal S$ so that $\frac{1}{m'-m}<\lambda$. Then by \eqref{eq:rel2b} we have \begin{align}\label{eq:33a} M+\lambda N_m& \textstyle =\frac{1}{m'-m}\big((m'-m)M+N_m\big)+\big(\lambda-\frac{1}{m'-m}\big)N_m\\ &\textstyle \sim_\mathbb{Q} \frac{1}{m'-m} N_{m'}+\big(\lambda-\frac{1}{m'-m}\big)N_m,\notag \end{align} which proves \eqref{eq:kodaira}. \medskip There are now three cases to consider. \medskip \emph{Step 2.} First assume that $$M+\lambda N_p \text{ is big for some }p\in\mathcal S \text{ and some rational number }\lambda>0.$$ Since $$\textstyle M \sim_\mathbb{Q} \frac{1}{1+\lambda p+\lambda \ell/t} \big(M+\lambda N_p+\lambda F+\frac{\lambda \ell}{t} L\big)$$ by \eqref{eq:rel2a}, this implies that $M$ is big, which shows \eqref{eq:kappalargem}. \medskip \emph{Step 3.} Now assume that $$\kappa(X,M+\lambda N_p)=0 \text{ for some }p\in\mathcal S \text{ and some rational number }\lambda>0.$$ Fix $q\in\mathcal S$ such that $0<\frac{1}{q-p}<\lambda$. Then as in \eqref{eq:33a} we have $$\textstyle M+\lambda N_p\sim_\mathbb{Q} \frac{1}{q-p} N_q+\big(\lambda-\frac{1}{q-p}\big)N_p,$$ hence $\kappa(X,N_q)\leq\kappa(X,M+\lambda N_p)=0$, and therefore $\kappa(X,N_q)=0$. Let $r\in\mathcal S$ be such that $r>q$. Then by \eqref{eq:rel2b} we have $$M\sim_\mathbb{Q}\frac{1}{q-p}(N_q-N_p)\quad\text{and}\quad M\sim_\mathbb{Q} \frac{1}{r-p}(N_r-N_p),$$ so that $$(r-p)N_q\sim_\mathbb{Q}(q-p)N_r+(r-q)N_p\geq0.$$ Since $\kappa(X,N_q)=0$, this implies $$(r-p)N_q=(q-p)N_r+(r-q)N_p, $$ and hence \begin{equation}\label{eq:kappaSupp} \Supp N_r\subseteq\Supp N_q \quad\text{and}\quad \kappa(X,N_r)=0. \end{equation} Therefore, for $r>q$, all divisors $N_r$ are supported on the reduced Weil divisor $\Supp N_q$. By \cite[Lemma 2.1]{LP18} there are positive integers $p'\neq q'$ larger than $q$ in $\mathcal S$ such that $N_{p'}\leq N_{q'}$, and thus by \eqref{eq:rel2b}, $$(q'-p')M\sim_\mathbb{Q} N_{q'}-N_{p'}\geq0,$$ hence $\kappa(X,M)\geq 0$. This together with \eqref{eq:kappaSupp} shows \eqref{eq:kappalargem}. \medskip \emph{Step 4.} Finally, by \eqref{eq:kodaira3} and by Steps 2 and 3 for the rest of the proof we may assume that \begin{equation}\label{eq:kodaira1} 0<\kappa(X,M+\lambda N_p)<n \quad\text{for every }p\in\mathcal S\text{ and every }\lambda>0. \end{equation} In this step we show that \begin{equation}\label{eq:positivekodaira} \kappa(X,K_X+\Delta)\geq0. \end{equation} Indeed, fix $p\in\mathcal S$ and denote $P=M+N_p$. Then $\kappa(X,P)\in\{1,\dots,n-1\}$ by \eqref{eq:kodaira1}, hence by \cite[Lemma 2.3]{LP18} there exists a resolution $\pi\colon Y\to X$ and a fibration $f\colon Y\to Z$: \[ \xymatrix{ Y \ar[d]_{\pi} \ar[r]^{f} & Z\\ X & } \] such that $\dim Z=\kappa(X,P)$, and for a very general fibre $F$ of $f$ and for every $\pi$-exceptional $\mathbb{Q}$-divisor $G$ on $Y$ we have \begin{equation}\label{eq:exceptional} \kappa\big(F,(\pi^*P+G)|_F\big)=0. \end{equation} There exist effective $\mathbb{Q}$-divisors $\Delta_Y$ and $E$ without common components such that $$K_Y+\Delta_Y\sim_\mathbb{Q}\pi^*(K_X+\Delta)+E,$$ and it is enough to show that $\kappa(Y,K_Y+\Delta_Y)\geq0$. By \eqref{eq:exceptional} we have $$\kappa\big(F,\big(t(K_Y+\Delta_Y)+\pi^*L+\pi^*N_p\big)|_F\big)=\kappa\big(F,(\pi^*P+tE)|_F\big)=0.$$ In particular, $(K_Y+\Delta_Y)|_F$ is not big on $F$ since otherwise $\big(t(K_Y+\Delta_Y)+\pi^*L+\pi^*N_p\big)|_F$ would also be big on $F$. But then $\kappa(Y,K_Y+\Delta_Y)\geq0$ by \cite[Proposition 3.2]{LP18}, which proves the claim \eqref{eq:positivekodaira}. \medskip \emph{Step 5.} From \eqref{eq:kodaira1} and \eqref{eq:positivekodaira} it immediately follows that $$0\leq\kappa(X,K_X+\Delta+\lambda N_p)<n\quad \text{ for every }p\in\mathcal S\text{ and every }\lambda>0.$$ Assume that there exist two elements $p<q$ in $\mathcal S$ and positive rational numbers $\lambda_p$ and $\lambda_q$ such that $$\kappa(X,K_X+\Delta+\lambda_p N_p)=\kappa(X,K_X+\Delta+\lambda_q N_q)=0.$$ Then \eqref{eq:positivekodaira} implies $\kappa(X,N_p)=\kappa(X,N_q)=0$. Thus, as in Step 3 we have $\kappa(X,M)\geq0$ and $\kappa(X,N_r)=0$ for all $r\in\mathcal S$ with $r\geq q$, which yields \eqref{eq:kappalargem}. \medskip \emph{Step 6.} Therefore, by Step 5 we may assume that \begin{equation}\label{eq:kodaira2} 0<\kappa(X,K_X+\Delta+\lambda N_p)<n \quad\text{for every }p\in\mathcal S\text{ and every }\lambda>0. \end{equation} Fix integers $w>k$ in $\mathcal S$ and a rational number $0<\varepsilon\ll1$ such that \begin{enumerate} \item[(a)] the pair $(X,\Delta+\varepsilon N_k)$ is klt, \item[(b)] $\varepsilon(w-k)>2n$. \end{enumerate} Then in Steps 6--8 we show that for every $s\in\mathcal S$ with $s>w$ we have \begin{equation}\label{eq:finalclaim} \kappa(X,M)\geq\kappa(X,N_s), \end{equation} which then implies \eqref{eq:kappalargem} and proves the theorem. \medskip Fix now any $s\in\mathcal S$ with $s>w$, and fix a rational number $0<\delta<\varepsilon$ such that: \begin{enumerate} \item[(c)] the pairs $(X,\Delta+\delta N_w)$ and $(X,\Delta+\delta N_s)$ are klt. \end{enumerate} Since we are assuming the existence of good models for klt pairs in lower dimensions, by \eqref{eq:kodaira2} and by \cite[Theorem 2.5]{LP18} applied to a birational model of the Iitaka fibration of $K_X+\Delta+\delta N_w$, there exists a $(K_X+\Delta+\delta N_w)$-MMP $\theta\colon X\dashrightarrow X_{\min}$ which terminates with a good model for $(X,\Delta+\delta N_w)$. \medskip We claim that: \begin{enumerate} \item[(i)] $\theta$ is $N_w$-negative, \item[(ii)] $\theta$ is $(K_X+\Delta)$-trivial as well as $L$-trivial; hence, the proper transforms of $t(K_X+\Delta)$ and $L$ at every step of this MMP are nef Cartier divisors by \cite[Theorem 3.7(4)]{KM98}, \item[(iii)] $\theta$ is $(K_X+\Delta+\delta N_s)$-negative. \end{enumerate} Indeed, it is enough to show the claim for the first step of the MMP, as the rest is analogous; one important point is that relations \eqref{eq:rel2a} and \eqref{eq:rel2b} also continue to hold for the proper transforms of the divisors involved. Let $c_R\colon X\to Z$ be the contraction of a $(K_X+\Delta+\delta N_w)$-negative extremal ray $R$ in this MMP. Since $K_X+\Delta$ is nef, we immediately get that $R$ is $N_w$-negative, which gives (i). For (ii), since $K_X+\Delta$ and $L$ are both nef, it suffices to show that $c_R$ is $M$-trivial. By the boundedness of extremal rays \cite[Theorem 1]{Kaw91}, there exists a rational curve $C$ contracted by $c_R$ such that $$(K_X+\Delta+\varepsilon N_k)\cdot C\geq {-}2n.$$ If $c_R$ were not $M$-trivial, then $M\cdot C\geq1$ as $M$ is nef and Cartier. But then \eqref{eq:rel2b} and the condition (b) above yield $$(K_X+\Delta+\varepsilon N_w)\cdot C=(K_X+\Delta+\varepsilon N_k)\cdot C+\varepsilon(w-k)M\cdot C>0.$$ On the other hand, by (i) we have $$(K_X+\Delta+\varepsilon N_w)\cdot C=(K_X+\Delta+\delta N_w)\cdot C+(\varepsilon-\delta)N_w\cdot C<0,$$ a contradiction which shows (ii). Finally, by \eqref{eq:rel2b} and by (ii) we have $$(K_X+\Delta+\delta N_s)\cdot R=(K_X+\Delta+\delta N_w)\cdot R+\delta(s-w)M\cdot R<0,$$ which gives (iii). \medskip \emph{Step 7.} Now, by (ii) we have $$\kappa(X,M)=\kappa(X_{\min},\theta_*M),$$ the divisor $\theta_*F$ is pseudoeffective, and $$\kappa(X_{\min},\theta_*N_m)\geq\kappa(X,N_m)\quad\text{for }m\in\mathcal S$$ by \cite[Lemma 2.8]{LP18}. Furthermore, $K_{X_{\min}}+\theta_*\Delta+\delta \theta_*N_w$ is semiample, and by \eqref{eq:rel2b} we have $$K_{X_{\min}}+\theta_*\Delta+\delta \theta_*N_s\sim_\mathbb{Q} K_{X_{\min}}+\theta_*\Delta+\delta \theta_*N_w+\delta(s-w)\theta_*M,$$ hence $K_{X_{\min}}+\theta_*\Delta+\delta \theta_*N_s$ is likewise nef by (ii). By (iii) and by \eqref{eq:kodaira2} we have $$\kappa(X_{\min},K_{X_{\min}}+\theta_*\Delta+\delta\theta_*N_s)=\kappa(X,K_X+\Delta+\delta N_s)\in\{1,\dots,n-1\},$$ hence $K_{X_{\min}}+\theta_*\Delta+\delta\theta_*N_s$ is semiample by \cite[Theorem 2.5]{LP18} applied to a birational model of the Iitaka fibration of $K_{X_{\min}}+\theta_*\Delta+\delta\theta_*N_s$. Therefore, in order to show \eqref{eq:kappalargem}, by replacing $X$ by $X_{\min}$, $\Delta$ by $\theta_*\Delta$, $L$ by $\theta_*L$, $F$ by $\theta_*F$ and $N_m$ by $\theta_*N_m$ for every $m\in\mathcal S$, we may additionally assume: \medskip \emph{Assumption 1.} The divisors $K_X+\Delta+\delta N_w$ and $K_X+\Delta+\delta N_s$ are semiample, and $\kappa(X,K_X+\Delta+\delta N_s)\in\{1,\dots,n-1\}$. \medskip \emph{Step 8.} Let $\varphi_w\colon X\to S_w$ and $\varphi_s\colon X\to S_s$ be the Iitaka fibrations associated to $K_X+\Delta+\delta N_w$ and $K_X+\Delta+\delta N_s$, respectively. Then there exist ample $\mathbb{Q}$-divisors $A_w$ on $S_w$ and $A_s$ on $S_s$ such that \begin{equation}\label{eq:pullbacks} K_X+\Delta+\delta N_w\sim_\mathbb{Q}\varphi_w^*A_w\quad\text{and}\quad K_X+\Delta+\delta N_s\sim_\mathbb{Q}\varphi_s^*A_s. \end{equation} If $\xi$ is a curve on $X$ contracted by $\varphi_s$, then by \eqref{eq:rel2b} we have \begin{align*} 0&=(K_X+\Delta+\delta N_s)\cdot \xi=(K_X+\Delta+\delta N_w)\cdot \xi+\delta(s-w)M\cdot \xi\\ & =(K_X+\Delta+\delta N_w)\cdot \xi+\delta(s-w)(K_X+\Delta)\cdot \xi+\delta(s-w)L\cdot \xi, \end{align*} hence $$(K_X+\Delta+\delta N_w)\cdot \xi=(K_X+\Delta)\cdot \xi=L\cdot \xi=0$$ since $K_X+\Delta+\delta N_w$, $K_X+\Delta$ and $L$ are nef. This implies two things: first, that $K_X+\Delta$ is $\varphi_s$-numerically trivial, and hence by Lemma \ref{lem:numlintrivial} there exists a $\mathbb{Q}$-Cartier $\mathbb{Q}$-divisor $B_s$ on $S_s$ such that \begin{equation}\label{eq:pullbackklt} K_X+\Delta\sim_\mathbb{Q}\varphi_s^*B_s. \end{equation} Second, every curve contracted by $\varphi_s$ is also contracted by $\varphi_w$. Hence, by the Rigidity lemma \cite[Lemma 1.15]{Deb01} there exists a morphism $\psi\colon S_s\to S_w$ such that $\varphi_w=\psi\circ\varphi_s$. \[ \xymatrix{ & X \ar[ld]_{\varphi_s} \ar[dr]^{\varphi_w} & \\ S_s \ar[rr]^{\psi} & & S_w } \] Therefore, denoting $$\textstyle L_s:=\frac{1}{\delta(s-w)} (A_s-\psi^*A_w)-tB_s$$ and $$\textstyle F_s:=(\ell+\frac{1}{\delta}) B_s+\frac{1}{\delta(s-w)}(w A_s-s\psi^* A_w),$$ by \eqref{eq:rel2a}, \eqref{eq:rel2b}, \eqref{eq:pullbacks} and \eqref{eq:pullbackklt} we have \begin{equation}\label{eq:twopullbacks} L\sim_\mathbb{Q} \varphi_s^*L_s\quad\text{and}\quad F\sim_\mathbb{Q} \varphi_s^*F_s. \end{equation} This implies that $L_s$ is nef and $F_s$ is pseudoeffective. Now by \eqref{eq:rel2a}, \eqref{eq:pullbacks} and \eqref{eq:twopullbacks} we have \begin{align*} (1+\delta\ell+\delta st)M & =t(K_X+\Delta+\delta N_s)+\delta t F+(1+\delta\ell)L\\ & \sim_\mathbb{Q}\varphi_s^*\big(tA_s+\delta t F_s+(1+\delta\ell)L_s\big). \end{align*} Finally, since $tA_s+\delta t F_s+(1+\delta\ell)L_s$ is a big divisor on $S_s$, we obtain \begin{align*} \kappa(X,M)&=\kappa\big(S_s,tA_s+\delta t F_s+(1+\delta\ell)L_s\big)=\dim S_s\\ &=\kappa(X,K_X+\Delta+\delta N_s)\geq\kappa(X,N_s), \end{align*} where the last inequality follows by \eqref{eq:positivekodaira}. This shows \eqref{eq:finalclaim} and finishes the proof. \end{proof} \section{Metrics with generalised algebraic singularities, I}\label{sec:Algsing} The goal of this section is to apply Theorem \ref{thm:nonvanishingForms} in the case when one knows that the nef divisors involved possess metrics with generalised algebraic singularities. This will have surprising consequences for the relationship between effectivity and num-effectivity. Additionally, we assume that $K_X$ is pseudoeffective and $K_X + \Delta$ is nef. The general case, where $K_X$ is not necessarily pseudoeffective and $K_X + \Delta$ is pseudoeffective rather then nef, will be treated in the following section. We start with the following result. Part (i) contains \cite[Corollary 4.5 and Corollary 8.5]{LP18} as special cases. \begin{thm}\label{thm:nv0} Assume the existence of good models for klt pairs in dimensions at most $n-1$. Let $(X,\Delta)$ be a projective terminal pair of dimension $n$ such that $K_X$ is pseudoeffective and $K_X+\Delta$ is nef, and let $L$ be a nef $\mathbb{Q}$-divisor on $X$. Assume that $\chi(X,\mathcal{O}_X)\neq0$. \begin{enumerate} \item[(i)] Suppose that the divisor $K_X+\Delta+L$ has a singular metric with generalised algebraic singularities and semipositive curvature current. Then for every $\mathbb{Q}$-divisor $L'$ with $L\equiv L'$ we have $\kappa(X,K_X+\Delta+L') \geq 0$. \item[(ii)] In particular, if $\kappa(X,K_X+\Delta+L)\geq0$, then for every $\mathbb{Q}$-divisor $D$ with $K_X+\Delta+L\equiv D$ we have $\kappa(X,D)\geq0$. \end{enumerate} \end{thm} \begin{proof} Part (ii) follows immediately from part (i). The assumptions imply that $K_X+\Delta+L'$ has a singular metric with generalised algebraic singularities and semipositive curvature current. Therefore, we may assume that $L=L'$. Let $\rho\colon X'\to X$ be a $\mathbb{Q}$-factorialisation of $X$. Then $\rho$ is an isomorphism in codimension $1$, hence $$K_{X'}+\rho_*^{-1}\Delta=\rho^*(K_X+\Delta)$$ and the pair $(X',\rho_*^{-1}\Delta)$ is terminal. By replacing $(X,\Delta)$ by $(X',\rho_*^{-1}\Delta)$, we may thus assume that $X$ is $\mathbb{Q}$-factorial. Choose a positive integer $\ell$ such that $M:=\ell(K_X+\Delta+L)$ is Cartier, and so that there exist a resolution $\pi\colon Y\to X$ and a metric $h$ with generalised algebraic singularities on $\pi^*\mathcal{O}_X(M)$ as in Section \ref{sec:prelim}. Then the corresponding curvature current $T$ of $h$ has the Siu decomposition $$T = \Theta+\sum_{j=1}^r \lambda_j D_j,$$ where the semipositive current $\Theta$ has all Lelong numbers zero, the numbers $\lambda_j$ are positive and rational and the divisor $\sum D_j$ has simple normal crossings. We have \begin{equation}\label{eq:metric1} \textstyle\mathcal I(h^{\otimes m})=\mathcal{O}_Y\big({-}\sum_{j=1}^r\lfloor m\lambda_j\rfloor D_j\big). \end{equation} Arguing by contradiction, we assume that $\kappa(X,M)=-\infty$. Then by Theorem \ref{thm:nonvanishingForms}, for all $p\geq 0$ and for all $m\gg0$ we have $$ H^0\big(Y,\Omega^p_Y \otimes \pi^*\mathcal{O}_X(mM)\big)=0,$$ and thus $$H^0\big(Y,\Omega^p_Y\otimes \pi^*\mathcal{O}_X(mM)\otimes\mathcal I(h^{\otimes m})\big) = 0.$$ Theorem \ref{thm:DPS} implies that for all $p\geq 0$ and for all $m\gg0$: $$H^p\big(Y,\mathcal{O}_Y(K_Y+m\pi^*M)\otimes\mathcal I(h^{\otimes m})\big) = 0,$$ which together with \eqref{eq:metric1} and Serre duality yields \begin{equation}\label{eq:euler} \textstyle\chi\big(Y,\mathcal{O}_Y\big(\sum_{j=1}^r\lfloor m\lambda_j\rfloor D_j-m\pi^*M\big)\big) = 0 \end{equation} for all $m\gg0$. Let $q$ be a positive integer such that $q\lambda_j\in\mathbb{N}$ for all $j$, and $D=\sum q\lambda_j D_j-q\pi^*M$. Then \eqref{eq:euler} implies $$\chi\big(Y,\mathcal{O}_Y(mD)\big) = 0\quad \text{for all }m\gg0.$$ But then $\chi(Y,\mathcal{O}_Y) = 0$. Since $X$ has rational singularities, this implies $\chi(X,\mathcal{O}_X) = 0$, a contradiction which finishes the proof. \end{proof} When the metrics have all Lelong numbers zero, then the conclusion is even stronger: \begin{thm}\label{thm:semipositive} Assume the existence of good models for klt pairs in dimensions at most $n-1$. Let $(X,\Delta)$ be a $\mathbb{Q}$-factorial projective terminal pair of dimension $n$ such that $K_X$ is pseudoeffective and $K_X+\Delta$ is nef, and let $L$ be a nef $\mathbb{Q}$-Cartier divisor on $X$. \begin{enumerate} \item[(i)] Assume that $K_X+\Delta+L$ has a singular metric with semipositive curvature current and vanishing Lelong numbers. If there exists a positive integer $m$ such that $m(K_X+\Delta+L)$ is Cartier and $$\chi\big(X,\mathcal{O}_X\big(m(K_X+\Delta+L)\big)\big)\neq0,$$ then for every $\mathbb{Q}$-divisor $L'$ with $L\equiv L'$ we have $\kappa(X,K_X+\Delta+L')\geq0$. \item[(ii)] If $K_X+\Delta$ and $K_X+\Delta+L$ have singular metrics with semipositive curvature currents and vanishing Lelong numbers (in particular, if $K_X+\Delta$ and $K_X+\Delta+L$ are semiample), and if $\chi(X,\mathcal{O}_X)\neq0$, then for every $\mathbb{Q}$-divisor $L'$ with $L\equiv L'$ and for every rational number $t\geq0$, the divisor $K_X+\Delta+tL'$ is semiample. \item[(iii)] If $K_X+\Delta$ has a singular metric $h$ with semipositive curvature current and vanishing Lelong numbers and if there exists a positive integer $\ell$ such that $\ell(K_X+\Delta)$ is Cartier and $$\chi\big(X,\mathcal{O}_X\big(\ell(K_X+\Delta)\big)\big)\neq0,$$ then every $\mathbb{Q}$-divisor $G$ with $K_X+\Delta\equiv G$ is semiample. \end{enumerate} \end{thm} \begin{proof} In (i) and (ii), the assumptions imply that $K_X+\Delta+L'$ has a singular metric vanishing Lelong numbers and semipositive curvature current. Therefore, we may assume that $L=L'$. Denote $M:=K_X+\Delta+L$. We first show (i). We follow closely the proof of Theorem \ref{thm:nv0}. We may assume that $X$ is $\mathbb{Q}$-factorial. Arguing by contradiction, assume that $\kappa(X,M) = - \infty$. By assumption, there exist a positive integer $\ell$ and a desingularisation $\pi\colon Y \to X$ such that $\ell M$ is Cartier and $\pi^*\mathcal{O}_X(\ell M)$ has a singular metric $h$ with semipositive curvature current and vanishing Lelong numbers. This implies that $$ \mathcal I(h^{\otimes m}) = \mathcal{O}_Y \quad\text{for all positive integers $m$.}$$ Then by Theorem \ref{thm:nonvanishingForms}, for all $p\geq 0$ and for all $m\gg0$ we have $$ H^0\big(Y,\Omega^p_Y \otimes \pi^*\mathcal{O}_X(m\ell M)\big)=0,$$ which together with Theorem \ref{thm:DPS} and Serre duality implies $$H^p\big(Y,\pi^*\mathcal{O}_X(-m\ell M)\big) = 0\quad\text{for all }p\geq0\text{ and }m\gg0,$$ hence $\chi\big(Y,\pi^*\mathcal{O}_X(m\ell M)\big) = 0$ for all $m\gg0$, hence for all $m$. Since $X$ has rational singularities, we deduce $$ \chi\big(X,\mathcal{O}_X(m\ell M)\big) = 0\quad\text{for all $m$},$$ a contradiction. \medskip Next we show (ii). By (i) and since $\chi(X,\mathcal{O}_X)\neq0$, we have $$\kappa(X,K_X+\Delta)\geq0\quad\text{and}\quad \kappa(X,M)\geq0.$$ Then there exists an effective $\mathbb{Q}$-divisor $G$ such that $M\sim_\mathbb{Q} G$. Fix a rational number $0<\varepsilon\ll1$ such that the pair $(X,\Delta+\varepsilon G)$ is klt, and note that $K_X+\Delta+\varepsilon G$ has a singular metric $h$ with semipositive curvature current and vanishing Lelong numbers, since both $K_X+\Delta$ and $G$ have such metrics. By \cite[Theorem 5.1]{GM17}, the divisors $K_X+\Delta$ and $K_X+\Delta+\varepsilon G$ are semiample. Let $\varphi_0\colon X\to S_0$ and $\varphi_\varepsilon\colon X\to S_\varepsilon$ be the Iitaka fibrations of $K_X+\Delta$ and $K_X+\Delta+\varepsilon G$, respectively. Let $C$ be a curve contracted by $\varphi_\varepsilon$. Then $(K_X+\Delta+\varepsilon G)\cdot C=0$ implies $$(K_X+\Delta)\cdot C=G\cdot C=0$$ as both $K_X+\Delta$ and $G$ are nef, hence by the Rigidity lemma \cite[Lemma 1.15]{Deb01} there exists a morphism $\xi\colon S_\varepsilon\to S_0$ such that $\varphi_0=\xi\circ\varphi_\varepsilon$. \[ \xymatrix{ X \ar[d]_{\varphi_0} \ar[r]^{\varphi_\varepsilon} & S_\varepsilon \ar[dl]^\xi \\ S_0 & } \] There exist ample $\mathbb{Q}$-divisors $A_\varepsilon$ on $S_\varepsilon$ and $A_0$ on $S_0$ such that $$K_X+\Delta+\varepsilon G\sim_\mathbb{Q}\varphi_\varepsilon^*A_\varepsilon\quad\text{and}\quad K_X+\Delta\sim_\mathbb{Q}\varphi_0^*A_0,$$ and denote $L_\varepsilon=\frac{1}{\varepsilon}\big(A_\varepsilon-(1+\varepsilon)\xi^*A_0\big)$. Then $$L\sim_\mathbb{Q}\varphi_\varepsilon^*L_\varepsilon,$$ and hence $L_\varepsilon$ is nef. Since for every $t\geq1$ we have $$\textstyle K_X+\Delta+tL\sim_\mathbb{Q} \frac{1}{1+\varepsilon}\varphi_\varepsilon^*\big(A_\varepsilon+\big(t+(t-1)\varepsilon\big)L_\varepsilon\big)$$ and $A_\varepsilon+\big(t+(t-1)\varepsilon\big)L_\varepsilon$ is ample on $S_\varepsilon$, the divisor $K_X+\Delta+tL$ is semiample for every $t\geq1$. As $K_X+\Delta$ is semiample, it follows that $K_X+\Delta+tL$ is semiample for every $t\geq0$. Finally, for (iii), there exists a $\mathbb{Q}$-divisor $L\equiv0$ such that $G=K_X+\Delta+L$. By (i) we have $\kappa(X,K_X+\Delta)\geq0$ and $\kappa(X,K_X+\Delta+L)\geq0$, and then the proof is the same as that of (ii). \end{proof} Theorem \ref{thm:nv0}(ii) and Theorem \ref{thm:semipositive}(iii) show that the Generalised Nonvanishing Conjecture, and in particular the techniques of this paper, have unexpected consequences even for the usual Nonvanishing Conjecture. We will partly generalise Theorem \ref{thm:semipositive} in Theorem \ref{thm:semipositive1} below. In small dimensions, Theorem \ref{thm:semipositive} holds almost unconditionally by our results in \cite{LP18a}. We will present particularly interesting cases in Section \ref{sec:algsing2}. \section{Metrics with generalised algebraic singularities, II}\label{sec:algsing2} In this section we generalise the results from the previous section without assuming that we deal with minimal varieties. We use the same strategy as in Section \ref{sec:reductions}. The only issue is that we have to check that the reduction results preserve the assumptions on the singularities and the Euler-Poincar\'e characteristic. We thus have to analyse in detail the results of Section \ref{sec:reductions}. We start with the following lemma. \begin{lem}\label{lem:algsing} Assume the termination of flips in dimensions at most $n-1$, and the Abundance Conjecture for klt pairs in dimensions at most $n-1$. Let $(X,\Delta)$ be a projective klt pair of dimension $n$ such that $K_X+\Delta$ is pseudoeffective. Suppose that $K_X+\Delta$ has a singular metric with generalised algebraic singularities and semipositive curvature current. If $\chi(X,\mathcal{O}_X)\neq0$, then $\kappa(X,K_X+\Delta)\geq0$ and every sequence of flips of $(X,\Delta)$ terminates. \end{lem} \begin{proof} By \cite[Corollary 1.2]{HMX14} it suffices to show that $\kappa(X,K_X+\Delta)\geq0$. By Remark \ref{rem:WZD}, the pair $(X,\Delta)$ has a weak Zariski decomposition, hence it has a minimal model by \cite[Theorem 1.5]{Bir12b}; note that we only need the termination of klt flips in dimensions at most $n-1$, since the termination of klt flips implies the termination of dlt flips by the special termination \cite{Fuj07a}. Let $(X_m,\Delta_m)$ be a minimal model of $(X,\Delta)$. Then $K_{X_m}+\Delta_m$ has a singular metric with generalised algebraic singularities and semipositive curvature current by Corollary \ref{cor:algsinMMP}. Since klt pairs have rational singularities, the Euler-Poincar\'e characteristic of the structure sheaf is preserved during any MMP of klt pairs, hence $$\chi(X_m,\mathcal{O}_{X_m})=\chi(X,\mathcal{O}_X)\neq0.$$ By passing to a terminalisation, we may assume that $(X_m,\Delta_m)$ is a terminal pair. If $X$ is uniruled, then $\kappa(X_m,K_{X_m}+\Delta_m)\geq0$ by \cite[Theorem 3.3]{DL15}. If $X$ is not uniruled, then $K_X$ is pseudoeffective by \cite[Corollary 0.3]{BDPP}, hence $\kappa(X,K_X+\Delta)=\kappa(X_m,K_{X_m}+\Delta_m)\geq0$ by Theorem \ref{thm:nv0}. \end{proof} \begin{thm}\label{thm:Supportalgsings} Assume the termination of flips in dimensions at most $n-1$, and the Abundance Conjecture for klt pairs in dimensions at most $n-1$. Let $(X,\Delta)$ be an $n$-dimensional $\mathbb{Q}$-factorial terminal pair such that $K_X+\Delta$ is pseudoeffective, and let $L$ be a nef $\mathbb{Q}$-divisor on $X$. Suppose that $K_X+\Delta+L$ has a singular metric with generalised algebraic singularities and semipositive curvature current and that $\chi(X,\mathcal{O}_X)\neq0$. Assume that there exists an effective $\mathbb{Q}$-divisor $D$ such that $$K_X + \Delta \sim_\mathbb{Q} D \geq 0 \quad \text{and} \quad \Supp\Delta\subseteq\Supp D.$$ Then for every $\mathbb{Q}$-divisor $L_0$ with $L\equiv L_0$ there exists a rational $t_0>0$ such that we have $\kappa(X,K_X+\Delta+tL_0)\geq0$ for every $0<t\leq t_0$. \end{thm} \begin{proof} The assumptions imply that $K_X+\Delta+L_0$ has a singular metric with generalised algebraic singularities and semipositive curvature current. Therefore, we may assume that $L=L_0$. We follow closely the proof of Theorem \ref{thm:Support}. \medskip \emph{Step 1.} Let $f\colon X'\to X$ be a log resolution of the pair $(X,D)$, and we write $$K_{X'}+\Delta_1\sim_\mathbb{Q} f^*(K_X+\Delta)+F',$$ where $\Delta_1$ and $F'$ are effective $\mathbb{Q}$-divisors with no common components. Denote $D_1:=f^*D+F'$ and $L':=f^*L$, so that $K_{X'}+\Delta_1\sim_\mathbb{Q} D_1$, and the divisors $$K_{X'}+\Delta_1\quad\text{and}\quad K_{X'}+\Delta_1+L'$$ have singular metrics with generalised algebraic singularities and semipositive curvature currents. We may also assume that these metrics descend to $X'$, see Section \ref{sec:prelim}. Then Step 1 of the proof of Theorem \ref{thm:Support} shows that there exists a $\mathbb{Q}$-divisor $H\geq0$ with $\Supp H=\Supp D_1$ such that, if we denote $\Delta':=\Delta_1+H$ and $D':=D_1+H$, then the pair $(X',\Delta')$ is log smooth with $\Delta'$ reduced, and $$K_{X'}+\Delta'\sim_\mathbb{Q} D'\geq0\quad \text{with}\quad \Supp\Delta'=\Supp D'.$$ In particular, there exists $0<\eta_0\ll1$ such that for all $0\leq\eta\leq\eta_0$ the divisors $$K_{X'}+(1-\eta)\Delta'\quad \text{and}\quad K_{X'}+(1-\eta)\Delta'+L'$$ have singular metrics with generalised algebraic singularities and semipositive curvature currents which descend to $X'$. As in Step 1 of the proof of Theorem \ref{thm:Support}, it suffices to show that there exists a rational $t_0'$ such that \begin{equation}\label{eq:claimclaim1} \kappa(X',D'+tL')\geq0\quad\text{for every }0<t\leq t_0'. \end{equation} \medskip \emph{Step 2.} As in Step 2 of the proof of Theorem \ref{thm:Support} we construct a cyclic covering $\pi\colon X''\to X'$, a reduced divisor $\Delta_1''$ on $X''$ such that $$K_{X''}+\Delta_1''=\pi^*(K_{X'}+\Delta'),$$ and an effective Cartier divisor $D_1'':=\pi^*D'$ on $X''$ such that $$K_{X''}+\Delta_1''\sim D_1''\geq0\quad \text{with}\quad \Supp\Delta_1''=\Supp D_1''.$$ Pick a rational number $0<\delta\leq\eta_0$ such that $$\Delta'':=\Delta_1''-\delta\pi^*\Delta'\geq0,\ D'':=D_1''-\delta\pi^*\Delta'\geq0\text{ and } \Supp D''=\Supp D_1''.$$ Then \begin{equation}\label{eq:331} K_{X''}+\Delta''=\pi^*\big(K_{X'}+(1-\delta)\Delta'\big)\quad\text{and}\quad K_{X''}+\Delta''\sim_\mathbb{Q} D'', \end{equation} and therefore, the pair $(X'',\Delta'')$ is klt by \cite[Proposition 5.20]{KM98}. Denote $L'':=\pi^*L'$. Then by Step 1 the divisors $$K_{X''}+\Delta''\quad\text{and}\quad K_{X''}+\Delta''+L''$$ have singular metrics with generalised algebraic singularities and semipositive curvature currents. Moreover, $X''$ is not uniruled and $\kappa(X'',K_{X''})\geq0$ as in Step 4 of the proof of Theorem \ref{thm:Support}. \medskip \emph{Step 3.} Since we are assuming the termination of flips in dimension $n-1$, \cite[Corollary 1.2]{HMX14} implies that any sequence of flips of the pair $(X'',\Delta'')$ terminates. Therefore, by Proposition \ref{pro:contocon} there exists an $L''$-trivial $(K_{X''}+\Delta'')$-MMP $\varphi\colon X''\dashrightarrow Y$ such that $K_Y+\varphi_*\Delta''+m\varphi_*L''$ is nef for some fixed $m\gg0$. Denote $\Delta_Y:=\varphi_*\Delta''$ and $L_Y:=\varphi_*L''$. As in the previous paragraph, any sequence of flips of the pair $(Y,\Delta_Y)$ terminates. As in Step 1 and the beginning of Step 2 of the proof of \cite[Lemma 5.2]{LP18a}, there is a birational contraction $$\theta\colon (Y,\Delta_Y)\dashrightarrow (Y_{\min},\Delta_{\min})$$ which is a composition of a sequence of operations of a $(K_Y+\Delta_Y)$-MMP, a positive rational number $\lambda$ and a divisor $L_{\min}:=\theta_*(mL_Y)$ on $Y_{\min}$ such that: \begin{enumerate} \item[(a)] $K_{Y_{\min}}+\Delta_{\min}$ is nef and $s(K_{Y_{\min}}+\Delta_{\min})+L_{\min}$ is nef for all $s\geq\lambda$, \item[(b)] the map $\theta$ is $\big(s(K_Y+\Delta_Y)+mL_Y)$-negative for $s>\lambda$. \end{enumerate} Set $\mu:=\max\{\lambda,m\}$. Then by Step 1 the divisor $$M:=(\mu+1)(K_{X'}+(1-\delta)\Delta')+mL'$$ has a singular metric with generalised algebraic singularities and semipositive curvature current which descends to $X'$. \medskip \emph{Step 4.} By assumption, we have the Siu decomposition $$M \equiv \Theta + \sum \lambda_i [D_i],$$ where $\Theta\geq0$ has vanishing Lelong numbers, $\lambda_i$ are positive rational numbers and $\sum D_i$ is a simple normal crossings divisor. Let $(p,q)\colon T\to X''\times Y_{\min}$ be a resolution of indeterminacies of the birational map $\theta\circ\varphi\colon X''\dashrightarrow Y_{\min}$. Set $g := \pi\circ p\colon T\to X'$. \[ \xymatrix{ T \ar[r]_p \ar@/^1.0pc/[rrr]^q \ar[dr]_g & X'' \ar[d]^{\pi} \ar@{-->}[rr]_{\theta\circ\varphi} && Y_{\min} \\ & X' && } \] Then $$ M_T := g^*M \equiv g^*\Theta + \sum \lambda_i g^*[D_i].$$ We may assume that $\sum \lambda_i g^*D_i$ has simple normal crossings support. \medskip \emph{Step 5.} Denote $$M'':=\pi^*M\quad\text{and}\quad M_{\min}:=(\theta\circ\varphi)_*M''=q_*M_T.$$ Then by \eqref{eq:331} we have $$M''=(\mu+1)(K_{X''}+\Delta'')+mL'',$$ and hence $$M_{\min}=(\mu+1)(K_{Y_{\min}}+\Delta_{\min})+L_{\min}.$$ The divisor $M_{\min}$ is nef by (a). \medskip We claim that $\kappa(Y_{\min},M_{\min})\geq0$. The claim immediately implies the theorem: indeed, then $\kappa(X',M')=\kappa(X'',M'')\geq0$ by (b), which gives \eqref{eq:claimclaim1}, as desired. \medskip \emph{Step 6.} It remains to prove the claim. Assume by contradiction, that $\kappa(Y_{\min},M_{\min})=-\infty$. Since $M_T=p^*M''$, by (b) there exists an effective $q$-exceptional $\mathbb{Q}$-divisor $E$ on $T$ such that $$M_T\sim_\mathbb{Q} q^*M_{\min}+E.$$ The proof of Lemma \ref{lem:Siu} shows that $\sum \lambda_i g^*D_i\geq E$, so that $$\textstyle q^*M_{\min}\equiv g^*\Theta + \big(\sum \lambda_i g^*D_i-E\big).$$ Set $$\textstyle N := (\sum \lambda_i g^*D_i-E) - q^*M_{\min}.$$ Then as in the proof of Theorem \ref{thm:nv0}, for all $p \geq 0$ and all $\ell$ sufficiently divisible we have \begin{equation}\label{eq:N} H^p\big(T, \mathcal{O}_T(\ell N)\big) = 0. \end{equation} Since $$\textstyle N \sim_\mathbb{Q} \sum \lambda_i g^*D_i-M_T=p^*\pi^*\big(\sum \lambda_i D_i-M\big)$$ and since $X''$ has rational singularities by Step 2, Equation \eqref{eq:N} implies $$\textstyle H^p\big(X'', \mathcal{O}_{X''}\big(\ell\pi^*\big(\sum \lambda_iD_i - M\big)\big)\big) = 0$$ for all $p \geq 0$ and all $\ell$ sufficiently divisible. As $\pi$ is finite, \cite[Lemma 4.1.14]{Laz04} gives $$\textstyle H^p\big(X', \mathcal{O}_{X'}\big(\ell\big(\sum \lambda_i D_i - M\big)\big)\big) = 0$$ for all $p \geq 0$ and all $\ell$ sufficiently divisible. As in the proof of Theorem \ref{thm:nv0}, this implies $$\chi(X',\mathcal{O}_{X'})=0,$$ hence $\chi(X,\mathcal{O}_X)=0$ since $X$ has rational singularities. This is a contradiction which proves the claim and finishes the proof of the theorem. \end{proof} \begin{thm}\label{thm:tau<1algsings} Assume the termination of flips in dimensions at most $n-1$, and the Abundance Conjecture in dimensions at most $n-1$. Let $(X,\Delta)$ be an $n$-dimensional $\mathbb{Q}$-factorial terminal pair such that $K_X+\Delta$ is pseudoeffective, let $L$ be a nef $\mathbb{Q}$-divisor on $X$, and assume that there exists a minimal model of $(X,\Delta)$. Suppose that $K_X+\Delta+L$ has a singular metric with generalised algebraic singularities and semipositive curvature current and that $\chi(X,\mathcal{O}_X)\neq0$. Assume that $K_X$ is not pseudoeffective. Then for every $\mathbb{Q}$-divisor $L_0$ with $L\equiv L_0$ there exists a rational $t_0>0$ such that we have $\kappa(X,K_X+\Delta+tL_0)\geq0$ for every $0<t\leq t_0$. \end{thm} \begin{proof} The assumptions imply that $K_X+\Delta+L_0$ has a singular metric with generalised algebraic singularities and semipositive curvature current. Therefore, we may assume that $L=L_0$. The proof is analogous to that of Lemma \ref{lem:Knotpsef}, by invoking Theorem \ref{thm:Supportalgsings} instead of Theorem \ref{thm:Support}, and by invoking Lemma \ref{lem:algsing} instead of Lemma \ref{lem:termination}. \end{proof} Finally, we have all the ingredients to prove Theorem \ref{thm:C}; Corollary \ref{cor:D} is an immediate consequence. \begin{proof}[Proof of Theorem \ref{thm:C}] The assumptions imply that $K_X+\Delta+L'$ has a singular metric with generalised algebraic singularities and semipositive curvature current. Therefore, we may assume that $L=L'$. \medskip \emph{Step 1.} Note that the assumptions imply that for each $0\leq t\leq1$, the divisor $K_X+\Delta+tL$ has a singular metric with generalised algebraic singularities and semipositive curvature current. By Lemma \ref{lem:algsing} we have $\kappa(X,K_X+\Delta)\geq0$ and every sequence of flips of $(X,\Delta)$ terminates. By Proposition \ref{pro:contocon}, by Lemma \ref{lem:MMPnumeff} and by Corollary \ref{cor:algsinMMP} we may assume that $K_X+\Delta+L$ is nef. We follow the proof of Theorem \ref{thm:induction2} closely. However, note that the conclusion in part (i) is stronger than that of Theorem \ref{thm:mainreduction}(i). We are thus somewhat more careful how we conduct the proof. \medskip \emph{Step 2.} It suffices to show (i). Indeed, if $n(X,K_X+\Delta+L)<n$, then in this case of part (ii) we conclude by Theorem \ref{thm:B}. Thus, we may (for part (ii)) assume that $n(X,K_X+\Delta+L)=n$. Then this is done analogously as in Step 2 of the proof of Theorem \ref{thm:induction2}. \medskip \emph{Step 3.} Therefore, from now on we prove (i). Then analogously as in Step 3 of the proof of Theorem \ref{thm:induction2}, by using that metrics with generalised algebraic singularities and semipositive curvature currents are preserved by an MMP by Corollary \ref{cor:algsinMMP}, we show that we may additionally assume the following: \medskip \emph{Assumption 1.} The pair $(X,\Delta)$ is $\mathbb{Q}$-factorial and terminal, and $K_X+\Delta$ is nef. \medskip We only note that we do not prove (but we also do not need) item (c) in Step 3 of the proof of Theorem \ref{thm:induction2}. Now, if $K_X$ is pseudoeffective, then (i) follows from Theorem \ref{thm:nv0}. If $K_X$ is not pseudoeffective, then (i) follows from Theorem \ref{thm:tau<1algsings}. This completes the proof. \end{proof} When the variety is not uniruled, or if we add an additional assumption, we get more: \begin{thm} Assume the termination of flips in dimensions at most $n-1$, and the Abundance Conjecture in dimensions at most $n-1$. Let $(X,\Delta)$ be a projective klt pair of dimension $n$ such that $K_X+\Delta$ is pseudoeffective, and let $L$ be a nef $\mathbb{Q}$-divisor on $X$. Suppose that $K_X+\Delta$ and $K_X+\Delta+L$ have singular metrics with generalised algebraic singularities and semipositive curvature currents, and that $\chi(X,\mathcal{O}_X)\neq0$. Assume either that $X$ is not uniruled, or assume the semiampleness part of the Abundance Conjecture in dimension $n$. Then there exists a positive rational number $t_0$ such that for every $\mathbb{Q}$-divisor $L'$ with $L\equiv L'$ we have $$\kappa(X,K_X+\Delta+tL')\geq0\quad\text{for every }0\leq t\leq t_0.$$ \end{thm} \begin{proof} When $X$ is not uniruled, the proof is the same as the proof of Theorem \ref{thm:C}, by noticing that we do not need Theorem \ref{thm:tau<1algsings}. \medskip Now, for the rest of the proof we do not assume that $X$ is not uniruled, but we assume the semiampleness part of the Abundance Conjecture in dimension $n$. Fix a $\mathbb{Q}$-divisor $L'$ with $L\equiv L'$. We will show that there exists a rational number $t_0>0$, independent of $L'$, such that $\kappa(X,K_X+\Delta+tL')\geq0$ for all $0\leq t\leq t_0$. By Theorem \ref{thm:C} there exists a positive rational number $r_0'$ such that $$\kappa(X,K_X+\Delta+tL')\geq0\quad\text{for all }0\leq t\leq r_0'.$$ Since we are assuming the termination of flips in dimension $n-1$, by \cite[Corollary 1.2]{HMX14} every sequence of flips of $(X,\Delta)$ terminates. Therefore, by Proposition \ref{pro:contocon} there exists an $L$-trivial (hence $L'$-trivial) $(K_X+\Delta)$-MMP $\varphi\colon X\dashrightarrow Y$ such that $K_Y+\varphi_*\Delta+m\varphi_*L$ is nef for some fixed $m\gg0$. Denote $\Delta_Y:=\varphi_*\Delta$, $L_Y:=\varphi_*L$ and $L_Y':=\varphi_*L'$. As in the previous paragraph, any sequence of flips of the pair $(Y,\Delta_Y)$ terminates. As in Step 1 and the beginning of Step 2 of the proof of \cite[Lemma 5.2]{LP18a}, there is a birational contraction $$\theta\colon (Y,\Delta_Y)\dashrightarrow (Y_{\min},\Delta_{\min})$$ which is a composition of a sequence of operations of a $(K_Y+\Delta_Y)$-MMP, a positive rational number $\lambda$ and a divisor $L_{\min}:=\theta_*L_Y$ on $Y_{\min}$ such that: \begin{enumerate} \item[(a)] $K_{Y_{\min}}+\Delta_{\min}$ is nef and $s(K_{Y_{\min}}+\Delta_{\min})+mL_{\min}$ is nef for all $s\geq\lambda$, \item[(b)] the map $\theta$ is $\big(s(K_Y+\Delta_Y)+mL_Y)$-negative for $s>\lambda$. \end{enumerate} Denote $L_{\min}':=\theta_*L_Y'$. Then clearly: \begin{enumerate} \item[(a$_0$)] $s(K_{Y_{\min}}+\Delta_{\min})+mL_{\min}'$ is nef for all $s\geq\lambda$, \item[(b$_0$)] the map $\theta$ is $\big(s(K_Y+\Delta_Y)+mL_Y')$-negative for $s>\lambda$. \end{enumerate} Denote $M:=\lambda(K_{Y_{\min}}+\Delta_{\min})+mL_{\min}'$. Then $M$ is nef by (a$_0$). Since $\kappa(Y_{\min},K_{Y_{\min}}+\Delta_{\min}+t L_{\min}')\geq0$ for $0\leq t\leq r_0'$, there exist a rational number $0<\delta\leq \min\{r_0',m/\lambda\}$ and an effective divisor $G'$ such that $$K_{Y_{\min}}+\Delta_{\min}+\delta L_{\min}'\sim_\mathbb{Q} G'.$$ Then $G'$ is nef by (a$_0$), and set $G:=\frac{m}{m-\delta\lambda}G'$. We have $$\textstyle K_{Y_{\min}}+\Delta_{\min}+\frac{\delta}{m-\delta\lambda}M=\frac{m}{m-\delta\lambda}(K_{Y_{\min}}+\Delta_{\min}+\delta L_{\min}')\sim_\mathbb{Q} G.$$ Fix a rational number $0<\varepsilon\ll1$ such that the pair $(Y_{\min},\Delta_{\min}+\varepsilon G)$ is klt, and note that $K_{Y_{\min}}+\Delta_{\min}+\varepsilon G$ is nef since $K_{Y_{\min}}+\Delta_{\min}$ and $G$ are nef. Since $\kappa(Y_{\min},K_{Y_{\min}}+\Delta_{\min}+\varepsilon G)\geq0$ and since we are assuming semiampleness part of the Abundance Conjecture in dimension $n$, the divisors $K_{Y_{\min}}+\Delta_{\min}$ and $K_{Y_{\min}}+\Delta_{\min}+\varepsilon G$ are semiample. Then we conclude that $K_{Y_{\min}}+\Delta_{\min}+tM$ is semiample for all $t\geq0$ as in the proof of Theorem \ref{thm:semipositive}(ii). In particular, the divisor $(\lambda+1)(K_{Y_{\min}}+\Delta_{\min})+mL_{\min}'$ is semiample. But by (b$_0$) this implies $$\textstyle \kappa\big(X,K_X+\Delta+\frac{m}{\lambda+1}L'\big)=\kappa\big(Y,K_Y+\Delta_Y+\frac{m}{\lambda+1}L_Y'\big)\geq0.$$ This finishes the proof. \end{proof} We deduce the following surprising fact from Theorem \ref{thm:C} when $L=0$. \begin{cor}\label{cor:numequiv} Assume the termination of flips in dimensions at most $n-1$, the Abundance Conjecture in dimensions at most $n-1$, and the semiampleness part of the Abundance Conjecture in dimension $n$. Let $(X,\Delta)$ be a projective klt pair of dimension $n$ such that $K_X+\Delta$ is pseudoeffective. Suppose that $K_X+\Delta$ has a singular metric with generalised algebraic singularities and semipositive curvature current. If $\chi(X,\mathcal{O}_X)\neq0$, then for every $\mathbb{Q}$-Cartier divisor $L$ with $L\equiv 0$ and for every rational number $t\geq0$ we have $\kappa(X,K_X+\Delta+tL)\geq0$. \end{cor} \begin{proof} Fix a $\mathbb{Q}$-divisor $L$ with $L\equiv 0$. Then by Theorem \ref{thm:C} there exists a positive rational number $t_0$ such that $\kappa(X,K_X+\Delta+tL)\geq0$ for all $0\leq t\leq t_0$. Since we are assuming the termination of flips in dimension $n-1$, by \cite[Corollary 1.2]{HMX14} every sequence of flips of $(X,\Delta)$ terminates. Let $\varphi\colon X\dashrightarrow Y$ be any $(K_X+\Delta)$-MMP which terminates with a minimal model $(Y,\Delta_Y)$, where $\Delta_Y:=\varphi_*\Delta$. Denote $L_Y:=\varphi_*L$, and note that this MMP is $L$-trivial. Then there exists an effective divisor $G$ such that $K_Y+\Delta_Y+t_0L_Y\sim_\mathbb{Q} G$. Fix a rational number $0<\varepsilon\ll1$ such that the pair $(Y,\Delta_Y+\varepsilon G)$ is klt, and note that $K_Y+\Delta_Y+\varepsilon G$ is nef since both $K_Y+\Delta_Y$ and $G$ are nef. Since $\kappa(Y,K_Y+\Delta_Y+\varepsilon G)\geq0$ and since we are assuming semiampleness part of the Abundance Conjecture in dimension $n$, the divisors $K_Y+\Delta_Y$ and $K_Y+\Delta_Y+\varepsilon G$ are semiample. Then we conclude that $K_Y+\Delta_Y+tL_Y$ is semiample for all $t\geq0$ as in the proof of Theorem \ref{thm:semipositive}(ii), which implies the result. \end{proof} We also have the following partial generalisation of Theorem \ref{thm:semipositive}, where we do not assume that the variety is not uniruled: \begin{thm}\label{thm:semipositive1} Assume the termination of flips in dimensions at most $n-1$, and the Abundance Conjecture in dimensions at most $n-1$. Let $(X,\Delta)$ be a $\mathbb{Q}$-factorial projective klt pair of dimension $n$, and let $L$ be a nef $\mathbb{Q}$-Cartier divisor on $X$. Assume that $K_X+\Delta$ and $K_X+\Delta+L$ have singular metrics with semipositive curvature currents whose Lelong numbers are all zero. If $\chi(X,\mathcal{O}_X)\neq0$, then for every $\mathbb{Q}$-divisor $L'$ with $L\equiv L'$ and for every rational number $t\geq0$, the divisor $K_X+\Delta+tL'$ is semiample. \end{thm} \begin{proof} Fix a $\mathbb{Q}$-divisor $L'$ with $L\equiv L'$. Then by Theorem \ref{thm:C} there exists a positive rational number $t_0$ such that $\kappa(X,K_X+\Delta+tL')\geq0$ for all $0\leq t\leq t_0$. We may assume that $t_0\leq1$. Then there exists an effective divisor $G$ such that $K_X+\Delta+t_0L\sim_\mathbb{Q} G$. Fix a rational number $0<\varepsilon\ll1$ such that the pair $(X,\Delta+\varepsilon G)$ is klt, and note that $K_Y+\Delta_Y+\varepsilon G$ has a singular metric with semipositive curvature current whose Lelong numbers are all zero since both $K_X+\Delta$ and $G$ do. Since $\kappa(X,K_X+\Delta+\varepsilon G)\geq0$, by \cite[Theorem 5.1]{GM17} the divisors $K_X+\Delta$ and $K_X+\Delta+\varepsilon G$ are semiample. Then we conclude that $K_X+\Delta+tL'$ is semiample for all $t\geq0$ as in the proof of Theorem \ref{thm:semipositive}(ii). \end{proof} We summarise the situation for surfaces and threefolds in the following two results. \begin{cor}\label{cor:surfaces} Let $(X,\Delta)$ be a projective klt surface pair such that $K_X+\Delta$ is pseudoeffective, and let $L$ be a nef $\mathbb{Q}$-Cartier divisor on $X$. Assume that $\chi(X,\mathcal{O}_X)\neq0$. \begin{enumerate} \item[(i)] Then for every $\mathbb{Q}$-divisor $D$ with $K_X+\Delta\equiv D$ we have $\kappa(X,D)\geq0$. \item[(ii)] If $K_X+\Delta+L$ is nef, then every $\mathbb{Q}$-divisor $D$ with $K_X+\Delta+L\equiv D$ is semiample. \end{enumerate} \end{cor} \begin{proof} Since $\dim X=2$, the assumptions of Corollary \ref{cor:numequiv} are satisfied, which gives (i). Furthermore, the assumptions of Theorem \ref{thm:semipositive1} are satisfied by \cite[Corollary C]{LP18a}, which shows (ii). \end{proof} \begin{cor}\label{cor:threefolds} Let $(X,\Delta)$ be a projective klt threefold pair such that $K_X+\Delta$ is pseudoeffective, and let $L$ be a nef $\mathbb{Q}$-Cartier divisor on $X$. Assume that $\chi(X,\mathcal{O}_X)\neq0$. \begin{enumerate} \item[(i)] Then for every $\mathbb{Q}$-divisor $D$ with $K_X+\Delta\equiv D$ we have $\kappa(X,D)\geq0$. \item[(ii)] If $K_X+\Delta$ is nef, then every $\mathbb{Q}$-divisor $D$ with $K_X+\Delta\equiv D$ is semiample. \item[(iii)] If $\nu(X,K_X+\Delta)>0$ and $K_X+\Delta$ is nef, then every $\mathbb{Q}$-divisor $D$ with $K_X+\Delta+L\equiv D$ is semiample. \end{enumerate} \end{cor} \begin{proof} Since $\dim X=3$, the assumptions of Corollary \ref{cor:numequiv} are satisfied, which gives (i). For (ii), the divisor $K_X + \Delta$ is semiample by the Abundance Conjecture on threefolds, hence Theorem \ref{thm:semipositive1} applies. Finally, for (iii) the assumptions of Theorem \ref{thm:semipositive1} are satisfied by \cite[Corollary D]{LP18a}. \end{proof} \section{Numerical dimension 1} \label{sec:num1} In case of numerical dimension one, the results of Section 5 hold unconditionally. More precisely, the following theorem and Theorem \ref{thm:G} generalise \cite[Theorem 6.3 and Theorem 6.7]{LP18}. \begin{thm} \label{thm:nu=1} Let $(X,\Delta)$ be a projective terminal pair of dimension $n$ such that $K_X + \Delta$ is nef. Let $L$ be a nef Cartier divisor on $X$ and let $t$ be a positive integer such that $t(K_X+\Delta)$ is Cartier. Denote $M=t(K_X+\Delta)+L$. Assume that $\nu(X,M) = 1$ and let $\pi\colon Y\to X$ be a resolution of $X$. Assume that for some positive integer $p$ we have $$ H^0\big(Y,(\Omega^1_Y)^{\otimes p} \otimes \mathcal{O}_Y(m\pi^*M)\big) \neq 0$$ for infinitely many integers $m$. Then $M$ is num-effective. \end{thm} \begin{proof} By repeating verbatim the proof of Theorem \ref{thm:nonvanishingForms}, we obtain that there exist a resolution $\pi\colon Y\to X$, a positive integer $\ell$, infinitely many positive integers $m$, effective Weil divisors $N_m$ on $Y$ and a pseudoeffective divisor $F$ on $Y$ such that \begin{equation}\label{eq:repeat} \pi_*N_m+(\pi_*F+\ell\Delta)\sim_\mathbb{Q} mM +\ell (K_X+\Delta). \end{equation} Assume that there exist positive integers $m_1<m_2$ such that $N_{m_1}=N_{m_2}=0$. Then \eqref{eq:repeat} gives $$(m_2-m_1)M\sim_\Q0,$$ and the result follows. Therefore, there exists a positive integer $m_0$ such that $N_{m_0}\neq0$. By adding $\frac{\ell}{t}L$ to the relation \eqref{eq:repeat} for $m=m_0$, we obtain $$\textstyle \pi_*N_{m_0}+\big(\pi_*F+\ell\Delta+\frac{\ell}{t}L\big)\sim_\mathbb{Q} \big(m_0+\frac{\ell}{t}\big) M.$$ We then conclude by \cite[Theorem 6.1]{LP18}. \end{proof} \begin{proof}[Proof of Theorem \ref{thm:G}] Fix a rational number $t\geq0$ and set $M:=K_X+\Delta+tL$. We will show that $M$ is num-effective. Assume first that $t>0$. Then we first claim that $\nu(X,M)=1$. Indeed, if $t\leq1$, then $$1=\nu\big(X,t(K_X+\Delta+L)\big)\leq\nu(X,M)\leq\nu(X,K_X+\Delta+L)=1$$ as $(1-t)(K_X+\Delta)$ and $(1-t)L$ is pseudoeffective, and similarly if $t\geq1$. If there exist a resolution $\pi\colon Y\to X$, a positive integer $m$ such that $mM$ is Cartier, and a singular metric $h$ on $\pi^*\mathcal{O}_X(mM)$ with semipositive curvature current, such that $\mathcal I(h)\neq\mathcal{O}_Y$, then the result follows from \cite[Theorem 6.5]{LP18}. Otherwise, pick a resolution $\pi\colon Y\to X$, a positive integer $m$ such that $mM$ is Cartier, and a singular metric $h$ on $\pi^*\mathcal{O}_X(mM)$ with semipositive curvature current. Then the proof is the same as that of Theorem \ref{thm:nv0}, by invoking Theorem \ref{thm:nu=1} instead of Theorem \ref{thm:nonvanishingForms}. Finally, if $t=0$, then $\nu(X,M)\leq\nu(X,K_X+\Delta+L)=1$ since $L$ is pseudoeffective. If $\nu(X,M)=0$, then $M\equiv 0$. Otherwise, $\nu(X,M)=1$, and then $M$ is num-effective by the argument above, by setting $L=0$. \end{proof} \section{Maps to abelian varieties}\label{sec:abelian} In this section we apply our methods to varieties admitting a morphism to an abelian variety. \begin{thm}\label{thm:overabelian} Assume the Generalised Nonvanishing Conjecture in dimensions at most $n-1$, the Abundance Conjecture in dimensions at most $n-1$ and the termination of flips in dimensions at most $n-1$. Let $(X,\Delta)$ be a projective klt pair of dimension $n$ such that $K_X+\Delta$ is pseudoeffective, and let $L$ be a nef $\mathbb{Q}$-divisor on $X$. Assume that there exists a nonconstant morphism $\alpha\colon X\to A$ to an abelian variety $A$. Then $K_X+\Delta+tL$ is num-effective for all $t\geq0$. \end{thm} \begin{proof} By replacing $L$ by some multiple, we may assume that $L$ is Cartier. Consider the Stein factorisation of $\alpha$: $$X \, {\buildrel {\beta} \over {\longrightarrow} }\, Y \, {\buildrel {\gamma} \over {\longrightarrow} } \, A.$$ Let $F$ be a general fibre of $\beta$. Then $\kappa\big(F,(K_X+\Delta)|_F\big)\geq0$ by the Generalised Nonvanishing Conjecture in dimension $\dim F$ and by \cite[Theorem 0.1]{CKP12}, hence $\kappa(X,K_X+\Delta)\geq0$ by \cite[Lemma 3.5]{Hu18}. Therefore, it suffices to prove that $K_X+\Delta+tL$ is num-effective for any $t>2n$. Fix any $m>2n$. By \cite[Corollary 1.2]{HMX14}, any sequence of flips of the pair $(X,\Delta)$ terminates. Therefore, by Proposition \ref{pro:contocon} there exists an $L$-trivial $(K_X+\Delta)$-MMP $\varphi\colon X\dashrightarrow Y$ such that $K_Y+\varphi_*\Delta+m\varphi_*L$ is nef. We claim that this MMP is, in fact, a relative MMP over $Y$ (hence over $A$). Indeed, it suffices to show this on the first step of the MMP, as the rest is analogous. Let $c_R\colon X\to Z$ be the contraction of a $(K_X+\Delta)$-negative extremal ray, and let $H$ be an ample divisor on $Y$. Then by the Cone theorem \cite[Theorem 3.7]{KM98}, $R$ is spanned by the class of some rational curve $C$ on $X$. Since abelian varieties contain no rational curves, neither does $Y$, and therefore $C$ has to be contracted by $\beta$. In particular, $\beta^*H\cdot C=0$. If $C'$ is any other curve contracted by $c_R$, then $C'$ is numerically proportional to $C$. Hence, $\beta^*H\cdot C'=0$ and so $C'$ is contracted by $\beta$. By the Rigidity lemma \cite[Lemma 1.15]{Deb01}, the morphism $\beta$ factors through $c_R$, which proves the claim. By Lemma \ref{lem:MMPnumeff} it suffices to show that $K_Y+\varphi_*\Delta+m\varphi_*L$ is num-effective. Therefore, by replacing $X$ by $Y$, $\Delta$ by $\Delta_Y$ and $L$ by $mL_Y$, we may assume additionally that $K_X+\Delta+L$ is nef. We need then to show that $K_X+\Delta+L$ is num-effective. If $n(X,K_X+\Delta+L)<n$, then $K_X+\Delta+L$ is num-effective by Theorem \ref{thm:B}. If $n(X,K_X+\Delta+L)=n$, then for a very general fibre $F$ of $\beta$ we have $$n\big(F,(K_X+\Delta+L)|_F\big)=\dim F$$ by \cite[Lemma 2.10]{LP18a}. Then $(K_X+\Delta+L)|_F$ is big by \cite[Lemma 5.2]{LP18a}, and hence $K_X+\Delta+L$ is $\beta$-big. Then $K_X+\Delta+L$ is num-effective by \cite[Theorem 4.1]{BirChen15}. \end{proof} \bibliographystyle{amsalpha}

4c8deafbd42984382f8e84b40f4e11870f635cd4

\section{Introduction} For over 80 years, since the first observations by F.~Zwicky, Dark Matter (DM) constitutes a basic pillar of the cosmological model of our Universe and although there are compelling evidence in all astrophysical scales, the particle content of DM remains elusive. From astrophysical observations we know that DM should be ``cold", non-baryonic, weakly interacting with a density of $\Omega h^{2}=0.1186\pm 0.002$~\cite{pdg}. A plethora of particles has been proposed over the years that match these criteria, none predicted by the Standard Model. This generic class of DM candidates is known as Weakly Interacting Massive Particles (WIMPs). Initially, WIMPs in the 10 GeV/c$^{2}$ -- 1 TeV/c${}^{2}$ mass range were favored by supersymmetric models. But the non-findings of passive experiments and at the LHC motivate searches in masses below 10 GeV/c${}^{2}$, where many new theoretical approaches such as the asymmetric dark model and dark sector predict DM candidates. Direct DM detection in this mass range is beyond the capabilities of the conventional experiments based on dual phase TPCs and solid state detectors which use heavy elements such as Xe, Ge, Ar or Si as targets due to two main reasons. First, is the kinematic issue i.e. it is less likely to produce energetic recoils with light WIMP scattering on heavy targets because of the minimum WIMP relative velocities required that are scarce in our Galaxy. The second issue is that heavy recoils have a large ionization deficit that is described by the ionization quenching factor~\cite{Santos}. For example a Xe recoil with kinetic energy equal to 500 eV will induce in a medium of the same type will deposit $\sim50$ eV through ionization. As a result, it would require a detector with a energy threshold at the single ionization electron to be detected. Both the kinematic and ionization issues are much less prominent for lighter targets, making them more appropriate for the detection of light DM candidates. In this work, we present the concept of the NEWS-G experiment which by using an innovative gaseous detector, the Spherical Proportional Counter (SPC) filled with Ne, He and H gas mixtures as targets, aiming to search for light WIMPs with unprecedented sensitivity. \section{The Spherical Proportional Counter} The SPC is a spherical gaseous detector~\cite{spc1}, presented in Fig.~\ref{fig:spc}, consisting of a grounded spherical shell which acts as the cathode and a small ball -the anode- placed at the center of the sphere, supported by a grounded metallic rod, to which the high voltage (HV) is applied and from which the signal is read-out (the sensor). The electric field, which is given by $$E(r) = \frac{r_{A}r_{C}}{r_{C}-r_{A}} \frac{V_{0}}{r^{2}}$$ where $r_{A}$ is the anode radius, $r_{C}$ the cathode radius, $V_{0}$ the voltage applied on the anode and $r$ the distance from the center of the detector. The dependence on the inverse of the detector radius squared results in the natural division of the detector into the drift volume where under the influence of the low electric field the electrons drift towards the anode until a few mm from the anode surface, i.e. the amplification region, where the charge multiplication takes place. The simplicity of the design permits the construction of a detector of this type solely by radiopure materials. A large detector can be constructed using tens of kgs of copper with the exception of parts of the sensor weighing less than one gram. \begin{figure} \centering \begin{subfigure}[b]{0.45\textwidth} \includegraphics[width=\textwidth]{spc_draw.jpeg} \caption{} \label{fig:spc1} \end{subfigure} % \begin{subfigure}[b]{0.35\textwidth} \includegraphics[width=\textwidth]{sensor2.png} \caption{} \label{fig:spc2} \end{subfigure} \caption[]{(a) SPC design and principle of operation and (b) illustration of the basic read out sensor \cite{recoils}. } \label{fig:spc} \end{figure} The spherical geometry provides several advantages against other detector geometries, such as the parallel plate detectors or the cylindrical counters when it comes to building large detectors ($\sim$ 1 m${}^3$). The sphere has the lowest surface to volume ratio and it is the vessel that can easier sustain high pressures. In addition, it is the detector geometry that has the lowest capacitance compared to parallel plate or cylindrical detectors. The reason for this is that its capacitance, which is approximated by $C \simeq 4\pi \epsilon_{0} r_{A}$ when $r_{C}\gg r_{A}$, depends only on the radius of the anode and not on the size of the vessel, making it easily scalable. Also the high gain achieved with anode balls with order mm diameter make the detector construction more robust. All these advantages allow the development of large detectors with single electron detection threshold and increased signal to background ratio. Another important advantage is that it provides the possibility to discriminate between events with long tracks and point like events and also provides information about the interaction radius. This is due to the difference in the collection time of the primary electrons in these events which is depicted in the rise time of the pulses~\cite{spc1,recoils}. The symmetry of the spherical shape along with the possibility to reject events originating from a given radius provide a valuable handle to reject background which usually comes from the surface in rare event search experiments. \section{The NEWS-G/LSM detector - First WIMP search run} An SPC is already installed at the Laboratoire Souterrain de Modane (LSM) inside the Frejus tunnel, one of the deepest laboratories in the world, under a rock thickness of 4800 m water equivalent for protection against cosmic radiation. The detector, named SEDINE, consists of a 60 cm diameter sphere made of ultra pure (NOSV) copper (Fig.~\ref{fig:sedine1}) and a 6.3 mm diameter spherical sensor made of silicon located at the center of the vessel (Fig.~\ref{fig:sedine2}). It is additionally protected from external radiation by a multi-layered cubic shielding composed of, from the inside to the outside, 8 cm of copper, 15 cm of lead and 30 cm of polyethylene (Fig.~\ref{fig:sedine3}). NEWS-G used SEDINE to perform its first search. The detector was filled with a mixture of neon (99.3\% in pressure) and methane (0.7\%) at a total pressure of 3.1 bar, corresponding to 280 g of target mass. The sensor was biased to 2520 V and a CANBERRA Model 2006 charge sensitive preamplifier was used to read-out the signal. The detector was operated under these conditions, in sealed mode for 42.7 days without interruption. \begin{figure} \centering \begin{subfigure}[b]{0.25\textwidth} \includegraphics[width=\textwidth]{sedine.png} \caption{} \label{fig:sedine1} \end{subfigure} % \begin{subfigure}[b]{0.15\textwidth} \includegraphics[width=\textwidth]{sensor_sedine.png} \caption{} \label{fig:sedine2} \end{subfigure} % \begin{subfigure}[b]{0.36\textwidth} \includegraphics[width=\textwidth]{sedine-sielding.png} \caption{} \label{fig:sedine3} \end{subfigure} % \caption[]{(a) Picture of SEDINE, the 60 cm diameter prototype installed at LSM, (b) the sensor installed in SEDINE with a 6.3 mm in diameter Si ball and a 380 \textmu m diameter insulated HV wire routed through a grounded copper rod and (c) the cubic shielding of SEDINE comprised of layer of copper, lead and polyethylene.} \label{fig:spc} \end{figure} \subsection{Calibration} The detector was calibrated using an ${}^{37}$Ar gaseous X-ray source and a ${}^{241}$Am--${}^{9}$Be neutron source. The ${}^{37}$Ar gas was added to the mixture of Ne:CH${}_{4}$ at the end of the run, providing a large sample of mono-energetic events at 2.82 keV and 270 eV from X-rays induced by electron capture in the K- and L-shells, respectively. The ${}^{241}$Am--${}^{9}$Be neutron source was used to induce nuclear recoil events, homogeneously distributed in the detector volume down to the critical low energy range of 150 eVee - 250 eVee where our sensitivity to sub-GeV/c${}^{2}$ WIMPs is expected. This calibration is of great importance as it allows the detector response measurement in rise time for point-like energy depositions in the detector volume. \begin{figure} \centering \begin{subfigure}[b]{0.4\textwidth} \includegraphics[width=\textwidth]{ar-calib.png} \caption{} \label{fig:calib1} \end{subfigure} % \begin{subfigure}[b]{0.4\textwidth} \includegraphics[width=\textwidth]{ambe-calib.png} \caption{} \label{fig:calib2} \end{subfigure} \caption[]{(a) Agreement between of ${}^{37}$Ar calibration data and simulation. The measured energy spectrum (red markers) exhibits two peaks at 2.82 keV and 270 eV from the electron capture in the K- and L-shell, respectively. The energy spectrum derived from the simulation (dark blue histogram) is scaled to the number of events recorded during the calibration run for comparison. (b) Comparison between the rise time distribution of events in the energy range 150 eVee - 250 eVee for events from neutron calibration data (red markers) and for simulated events with energy deposits homogeneously distributed in the volume (dark blue histogram) \cite{news}.} \label{fig:calib} \end{figure} \section{Analysis of the LSM run and first NEWS-G results} \subsection{Event simulation method - Background modeling} For the determination of our sensitivity to light DM we simulated the detector response to ionizing radiation. In this simulation method we took into account all the basic processes that take place during the detector operation such as (a) the primary ionization for electronic recoils of energy $E_{R}$ with mean energy to produce an electron-ion pair $W_{\gamma}(E_{R})$ and for nuclear recoils with $W_{n}(E_{R}) = W_{\gamma}(E_{R})/Q(E_{R})$ where $Q(E_{R})$ is the SRIM derived quenching factor, (b) the electric field given by finite-element software after modeling the full detector geometry, (c) the drift and diffusion of primary charges which was parametrized using Magboltz coefficients, and (d) the avalanche process which was simulated using Garfield++, accounting for Penning transfers and fitting the results with a Polya distribution \footnote{$ P\Bigg( \frac{n}{\langle n \rangle} \Bigg) = \frac{(1+\theta)^{(1+\theta)}}{\Gamma(1+\theta)}\Bigg( \frac{n}{\langle n \rangle} \Bigg)^{\theta}exp\Bigg[ -(1+\theta) \frac{n}{\langle n \rangle} \Bigg]$, where $\langle n \rangle$ is the mean gain and $\theta$ the shape parameter}. The final pulse for an event was constructed by adding the contributions of all PEs that reached the sensor. To these pulses we also added a baseline randomly chosen from empty pre-traces of events recorded during the physics-run to account for realistic noise. The simulated pulses were processed with the same algorithms as real pulses. The simulation method was validated by comparing the simulation to the results of the ${}^{37}$Ar calibration with the measurements which we present at (Fig.~\ref{fig:calib1}). We can see the good agreement but also the reproduction of the non-Gaussianity of the 2.82 keV line caused by the field anisotropy in the avalanche region. In Fig.~\ref{fig:calib2} we present also the comparison of the rise time distribution of the neutron induced nuclear recoils during the ${}^{241}$Am--${}^{9}$Be calibration and the simulated nuclear recoils. The good agreement between the two distributions validates the Magboltz derived drift parameters used in the simulation to calculate the rise time of the pulses. Using this method we modeled our background which is categorized into surface and volume events. The surface events are uniformly distributed in the vessel and originate from radon daughter decays. Volume events originate from high energy \textgamma-rays from ${}^{208}$Tl and ${}^{40}$K present in the rock, and from the decay chains of ${}^{238}$U and ${}^{232}$Th contained in the copper shell and the shielding itself. The results of these simulations have shown that the pulse rise time provides useful statistical discrimination against surface events down to our analysis threshold of 150~eVee. \subsection{Data analysis} A conservative analysis threshold was set at 150 eVee, above the trigger threshold of $\sim$ 36 eVee, to ensure a $\sim$ 100 \% trigger efficiency. Further on, a quality cut was applied to reject events within a 4 second window after triggering to remove non-physical events thus introducing a dead time resulting to 20.1\% exposure loss. After the quality cut we defined a preliminary Region Of Interest (ROI) in rise time 10 \textmu s - 32 \textmu s and energy 150 eV - 4000 eV. Side band regions were used to determine the expected backgrounds in the preliminary ROI. The event rate measured in the 4000 eVee - 6000 eVee energy range was used to extrapolate the expected Compton background down to lower energy assuming a flat recoil energy spectrum. The side band region was used together with our simulation to extrapolate the expected number and distribution of surface events leaking in the preliminary ROI. For a total exposure of 34.1 live-days$\times$0.283 kg = 9.7 kg$\cdot$days, 1620 events were recorded in the preliminary ROI. The ROI was tuned by using a Boosted Decision Tree (BDT); a machine learning algorithm which was trained with $10^{5}$ simulated background events and signal events for 8 different WIMP masses (from 0.5 to 16 GeV/c$^{2}$). For each WIMP mass, the events where classified as background-like or signal-like using the BDT score (ranging between -1 and 1 respectively) resulting in a WIMP-mass-dependent fine-tuned ROI in the rise time vs. energy plane. Details about the principle of the analysis and the simulations can be found in the recent NEWS-G~\cite{news}. \subsection{Results} For each WIMP mass and considering as candidates all the events observed in the corresponding fine-tuned ROI, a 90 \% Confidence Level (C.L.) upper limit on the spin-independent WIMP-nucleon scattering cross section was derived using Poisson statistics. The recoil energy spectrum used to derive the sensitivity to light DM is based on standard assumptions of the DM-halo model\footnote{dark matter density $\rho_{DM} = 0.3\;{\rm GeV}/c^{2}/{\rm cm}^{3}$, galactic escape velocity of $v_{\rm esc} = 544\;{\rm km/s}$, asymptotic circular velocity of $v_{0} = 220\;{\rm km/s}$}. The resulting exclusion limit is presented at Fig.~\ref{fig:news} as a solid red line, setting new constraints on the spin-independent WIMP-nucleon scattering cross-section below 0.6 GeV/c$^{2}$ and excludes at 90\% C.L. a cross-section of $4.4\times10^{-37}$ cm${}^{2}$ for a 0.5 GeV/c$^{2}$ light DM candidate mass. \begin{figure} \centering \includegraphics[width=0.6\textwidth]{result.png} \caption[]{Constraints in the Spin-Independent WIMP-nucleon cross section versus WIMP mass plane. The result from this analysis is shown in solid red~\cite{news}.} \label{fig:news} \end{figure} \section{Future prospects - Conclusions} With the first results of NEWS-G we demonstrated the high potential of Spherical Proportional Counters being used for the search of low-mass DM. A new physics run is under way by operating SEDINE filled with He based gas mixture and a new physics run with H rich gas mixtures is envisioned. These runs will allow for more optimized momentum transfers for low-mass particles in the GeV/c$^{2}$ mass range, and increase our sensitivity to sub-GeV/c$^{2}$ light DM. The next phase of the experiment will build upon the experience acquired from the operation of the SEDINE prototype at the LSM. It will consist of a 140 cm diameter sphere, made of extremely low activity copper (in the range of a few \textmu Bq/kg of U and Th impurities) that is going to begin operation in SNOLab by summer 2019. The design for the shielding of the new sphere is more compact and advanced. It consists of shell of 22 cm of low activity lead lined with 3 cm of archaeological lead, inside a 40 cm thick polyethylene shield. The new setup will include dedicated handling to avoid radon entering the detector at any time. With such improvements we expect a significant reduction of the backgrounds levels, relative to the above results, and allow sensitivity down to cross sections of $\mathcal{O}$ $({10}^{-41}$ cm$^{2}$). The use of H and He targets will allow us to reach WIMP mass sensitivity down to 0.1 GeV/c$^{2}$. Finally, NEWS-G is developing novel detector instruments to facilitate the efficient operation of large detectors, such as the NEWS-G/SNO detector that may be influenced by the especially low electric field in the far radii of the detector volume. For this reason, new multi-ball sensors (ACHINOS~\cite{achinos}) are being developed (Fig.~\ref{fig:achinos1}) that are composed of multiple anode balls equidistantly placed on a virtual spherical surface and all biased at the same potential. Their collective influence results in an increased field magnitude at large radii while maintaining the ability to reach high gain operation provided by anode diameter in the order of a mm. Developments such as this open the way for even larger detectors and operation in higher pressure. \begin{figure} \centering \begin{subfigure}[b]{0.18\textwidth} \includegraphics[width=\textwidth]{achinos2.jpg} \caption{} \label{fig:achinos1} \end{subfigure} % \begin{subfigure}[b]{0.35\textwidth} \includegraphics[width=\textwidth]{achinos.png} \caption{} \label{fig:achinos1} \end{subfigure} \caption[]{(a) An ACHINOS prototype with 11 balls of 2 mm in diameter constructed using 3D printed materials (b) the electric field magnitude versus the radius in the case of a single 2-mm in diameter anode ball and in the case of the 11-ball ACHINOS sensor distributed on a 36-mm diameter sphere both placed in the center of a 300-mm diameter detector. The electric field in the far radii is $\sim9$ times higher for a detector equipped with an ACHINOS sensor than a single ball sensor placed at the same bias~\cite{achinos}.} \label{fig:achinos} \end{figure} \section{Acknowledgments} This work is funded by the French National Research Agency (ANR-15-CE31-0008). \newpage \section*{References}

649f2c6b669a78f75c4a484e3953e49d5861a852

\section{Introduction}\label{Sec:Intro} Plasmonics has experienced a drastic paradigm shift during the past decade, with rapid advances in nanofabrication and characterisation allowing metallic nanostructures to enter the quantum regime~\cite{tame_natphys9,zhu_natcom7,bozhevolnyi_nanophot6,Bozhevolnyi,tserkezis_ijmpb31_0}, defined by components of just a few nm in size~\cite{palpant_prb57,berciaud_nl5,scholl_nat483,raza_natcom6} and separations of only a few \AA~\cite{duan_nl12,kern_nl12}. In such situations, quantum effects like nonlocal screening~\cite{abajo_jpcc112,raza_jpcm27}, surface-enhanced Landau damping~\cite{shahbazyan_prb94,khurgin_acsphoton4}, electron spill-out~\cite{teperik_prl110,toscano_natcom6} and tunnelling~\cite{savage_nat491,scholl_nl13,tan_sci343,lerch_ijmpb31}, which cannot be captured by the common local response approximation (LRA) of classical electrodynamics, quickly become important and dominate the optical response. To deal with this issue, and overcome the limitations of the common free-electron-gas description in theoretical studies, a variety of quantum-informed models has been proposed. Benefiting from the enormous experience acquired over the decades through studies of metals in solid-state physics, models based on screening and the resulting nonlocal dielectric function~\cite{Ashcroft}, the $d$-parameter formalism of Feibelman~\cite{feibelman_progsurfsci12}, or (time-dependent) density functional theory (TD-DFT)~\cite{lang_prb1,Marques,varas_nanophot5,zhang_ijmpb31} have been developed. For instance, the hydrodynamic Drude model (HDM), already introduced in the 1930s~\cite{bloch_zphys81}, keeps reemerging in various forms due to its success in reproducing screening effects in noble metals~\cite{ruppin_prl31,fuchs_prb35,leung_prb42,mcmahon_jpcc114,raza_prb84,david_acsnano8,trugler_ijmpb31}, and approaches to extend its applicability are being proposed~\cite{toscano_natcom6,luo_prl111,ciraci_prb95}. One such extension, that also accounts in an efficient way for Landau damping, is provided by the generalised nonlocal optical response (GNOR) theory~\cite{mortensen_natcom5}. Other approaches to tackle Landau damping are typically based on a modified, quantum-mechanically obtained damping rate~\cite{shahbazyan_prb94,khurgin_acsphoton4}. On the other hand, the quantum-corrected model (QCM)~\cite{esteban_natcom3} has proven efficient in treating plasmonic dimers with gap distances in the quantum-tunnelling regime~\cite{savage_nat491,scholl_nl13} when the sizes involved render full TD-DFT calculations prohibitive. Finally, $d$-parameter approaches~\cite{yan_prl115,christensen_prl118} have attracted much interest recently, as they introduce the induced charges and currents of a quantum-mechanical calculation into an otherwise classical computation, accounting in principle for all aspects of quantum predictions relevant to plasmonics. To move through this diversity of complementing -- or sometimes contradicting -- approaches, a robust test bed for assessing their validity and pertinence is required. Here we show that a promising such template can be found in chiral metallic nanoparticle (NP) helices~\cite{fan_jpcc115,kuzyk_nat483}, whose optical response is dominated by a double-peak absorption resonance in the visible. By calculating absorption spectra of such helices for left- and right-circularly polarised (LCP and RCP) incident light, we show that different models produce fundamentally different circular dichroism (CD) spectra, characterised by a sign change that can be traced even when the two individual absorption peaks are hard to resolve. In particular, we compare HDM, GNOR and QCM to the standard LRA, and explain how gradually decreasing the distance between NPs in a helical arrangement allows to monitor the transition from one model's predominance to the other's, through distinct, experimentally resolvable spectral features that are significantly different in each case, in accordance with the particular physics governing each model. \section{Theoretical methods}\label{Sec:models} A fundamental factor to the rapid growth of plasmonics has been that, in the vast majority of situations, the theoretical description is extremely efficient within classical electrodynamics, with the metal adequately described by a \emph{local}, usually scalar dielectric function $\varepsilon$. The main assumption of LRA is that, in the linear regime, the displacement vector $\mathbf{D}$ in a metallic NP is simply proportional to the externally applied electric field $\mathbf{E}$~\cite{Bohren}, \begin{equation}\label{Eq:localD} \mathbf{D} (\mathbf{r}, \omega) = \varepsilon_{0} \varepsilon (\omega) \mathbf{E} (\mathbf{r}, \omega)~, \end{equation} where $\varepsilon_{0}$ is the vacuum permittivity, and $\varepsilon$ depends only on the angular frequency $\omega$ (dispersive medium), but not on the position $\mathbf{r}$. For simple free-electron metals $\varepsilon$ follows the Drude model~\cite{Ashcroft}, \begin{equation}\label{Eq:Drude} \varepsilon = \varepsilon_{\infty} - \frac{\omega_{\mathrm{p}}^{2}}{({\omega^{2} + \mathrm {i} \omega \gamma)}}~, \end{equation} where $\omega_{\mathrm{p}}$ is the plasmon frequency, $\gamma$ is the damping rate, and $\varepsilon_{\infty}$ the contribution of bound electrons. In noble metals, the contribution from interband transitions is typically included in $\varepsilon_{\infty}$, either through fitting Eq.~(\ref{Eq:Drude}) to measured data~\cite{johnson_prb6}, or by subtracting the Drude part from the experimental dielectric function~\cite{abajo_jpcc112}. With these assumptions, one just needs to solve the wave equation for the electric field, \begin{equation}\label{Eq:WaveLRA} \nabla \times \nabla \times \mathbf{E} (\mathbf{r}, \omega) = \left(\frac{\omega}{c}\right)^{2} \varepsilon (\omega) \mathbf{E} (\mathbf{r}, \omega)~, \end{equation} subject to the appropriate boundary conditions. In LRA, a hard wall boundary condition is usually adopted, implying that the induced charges are exactly localised at the metal-dielectric interface, explicitly excluding the possibility of spill-out, while the electron gas at the NP interior is assumed incompressible. The assumption of locality starts to become problematic when NP sizes, or their separations in plasmonic aggregates, are comparable to the electron mean free path. Going one step backwards, a more fundamental (still in the linear regime) constitutive relation is \begin{equation}\label{Eq:nonlocalD} \mathbf{D} (\mathbf{r}, \omega) = \varepsilon_{0} \int \mathrm{d} \mathbf{r}' \varepsilon (\mathbf{r}, \mathbf{r}', \omega) \mathbf{E} (\mathbf{r}', \omega)~, \end{equation} where the displacement at position $\mathbf{r}$ might depend on the electric field at positions $\mathbf{r}'$. In a homogeneous medium the spatial dependence becomes $\mathbf{r}-\mathbf{r}'$, simplifying the Fourier-transformed Eq.~(\ref{Eq:nonlocalD}) to $\mathbf{D} (\mathbf{k}, \omega) = \varepsilon_{0} \varepsilon(\mathbf{k}, \omega) \mathbf{E} (\mathbf{k}, \omega)$, where $\mathbf{k}$ is the wavevector. To proceed one needs an explicit form for $\varepsilon(\mathbf{k}, \omega)$, such as those obtained in the theory of screening~\cite{Ashcroft}. An efficient description is provided by HDM, where the equation of motion (at time $t$) for an electron moving with velocity $\mathbf{v} (\mathbf{r}, t)$ in a metal with electron density $n (\mathbf{r}, t)$ is given by \begin{equation}\label{Eq:HydroMotion} \left[\frac{\partial}{\partial t} + \mathbf{v} \cdot \nabla \right] \mathbf{v} = - \frac{e}{m} \left(\mathbf{E} + \mathbf{v} \times \mathbf{B} \right) - \frac{1}{m} \nabla \frac{\delta G[n]}{\delta n} - \gamma \mathbf{v}~. \end{equation} In the above, $e$ is the (positive) electron charge and $m$ its mass; $\mathbf{B}$ is the magnetic field, so that the first term on the right-hand side corresponds to the Lorentz force, while the last term describes damping. Finally, the functional $G$ takes into account the internal kinetic energy of the electron gas, together with the exchange and correlation contributions. In its simplest form, the Thomas--Fermi approximation, only the kinetic energy is accounted for, and the second term becomes $(\beta^{2}/n) \nabla n$, where, in the high-frequency limit, $\beta^{2} = 3 v_{\mathrm{F}}^{2}/5$ ($v_{\mathrm{F}}$ being the Fermi velocity of the metal). This hydrodynamic pressure term accounts for the finite compressibility and the fact that electrons are fermions obeying the exclusion principle~\cite{Griffiths}. Eq.~(\ref{Eq:HydroMotion}), combined with the continuity equation, $\frac{\partial}{\partial t} n = - \nabla \cdot (n \mathbf{v})$, can be solved assuming a small deviation from the equilibrium electron density $n_{0}$ to produce the system of coupled electromagnetic (EM) equations~\cite{raza_jpcm27} \begin{eqnarray}\label{Eq:CoupledHydro} && \nabla \times \nabla \times \mathbf{E} (\mathbf{r}, \omega) = \left(\frac{\omega}{c}\right)^{2} \varepsilon_{\infty} \mathbf{E} (\mathbf{r}, \omega) + \mathrm{i} \omega \mu_{0} \mathbf{J} (\mathbf{r}, \omega)~, \nonumber\\ && \frac{\beta^{2}}{\omega \left(\omega + \mathrm{i} \gamma \right)} \nabla \left[\nabla \cdot \mathbf{J} (\mathbf{r}, \omega) \right] + \mathbf{J} (\mathbf{r}, \omega) = \sigma \mathbf{E} (\mathbf{r}, \omega)~, \end{eqnarray} where $\mu_{0}$ is the vacuum permeability, $\mathbf{J} (\mathbf{r}, \omega) = -e n_{0} \mathbf{v} (\mathbf{r}, \omega)$ is the induced current density, and $\sigma = \varepsilon_{0} \mathrm{i} \omega_{\mathrm{p}}^{2}/[\omega (\omega + \mathrm{i} \gamma)]$ is the Drude conductivity. While exact analytic solutions can be found for spherical and cylindrical NPs~\cite{ruppin_prl31,raza_prb84}, for arbitrary geometries one solves Eqs.~(\ref{Eq:CoupledHydro}) numerically. Here we use a commercial finite-element method (FEM) solver (Comsol Multiphysics 5.1)~\cite{toscano_oex20}, assuming the additional boundary condition $\mathbf{J} \cdot \hat{\mathbf{n}} = 0$ (where $\hat{\mathbf{n}}$ is the unit vector normal to the metal-dielectric interface), which implies that the electron density vanishes abruptly at the interface (no spill-out). This condition can be relaxed self-consistently, by adding exchange-correlation terms to the functional $G[n]$ of Eq.~(\ref{Eq:HydroMotion})~\cite{toscano_natcom6}. With HDM describing electron convection, one might ask how the optical response is affected by diffusive currents. In the presence of diffusion the continuity equation becomes \begin{equation}\label{Eq:ContinuityDiffusion} -\mathrm{i} \omega e n(\mathbf{r}, \omega) = D \nabla^{2}\left[e n(\mathbf{r}, \omega)\right] + \nabla \cdot \left[-e n_{0} \mathbf{v}(\mathbf{r}, \omega)\right]~, \end{equation} where $D$ is the diffusion constant. Combining this with the diffusive version of Fick's law, \begin{equation}\label{Eq:Fick} \mathbf{J}(\mathbf{r}, \omega) = -e n_{0} \mathbf{v}(\mathbf{r}, \omega) + e D \nabla n(\mathbf{r}, \omega)~, \end{equation} simply modifies the second of Eqs.~(\ref{Eq:CoupledHydro}) by adding the term $-\mathrm{i} D/\omega$ to the prefactor of $\nabla \left[\nabla \cdot \mathbf{J} (\mathbf{r}, \omega) \right]$; the solving methodologies developed for HDM apply thus immediately. Using the Boltzman equation of motion as the starting point, it can be shown that diffusion is the main \textit{low-frequency} contribution in the bulk, and becomes negligible at optical frequencies. Consequently, if a diffusion term is to be included, it must describe processes at the surface of the metal. Indeed, comparison with TD-DFT and modified-damping models shows that GNOR accounts efficiently for surface-enhanced Landau damping~\cite{tserkezis_ijmpb31}. For NP separations of just few {\AA}, a prominent feature left out of the hydrodynamic treatment (by assuming that $\mathbf{J} \cdot \hat{\mathbf{n}} = 0$ at the metal surface) is the probability of direct electron tunnelling between NPs~\cite{savage_nat491,scholl_nl13}. This is where QCM~\cite{esteban_natcom3,esteban_faraday178} comes into play, to account for the infeasibility of fully quantum mechanical calculations for realistic NP sizes, where millions of electrons are involved. The model starts by calculating (typically with TD-DFT) the electron tunnelling probability in a nanogap (e.g. two flat interfaces), $T(\Omega, l)$, as a function of energy $\Omega$ and position in the gap $l$. This in turn is translated into a distance-dependent dc conductivity through \cite{esteban_natcom3} \begin{equation}\label{Eq:QCMconductivity} \sigma_{0}(l) = \frac{l}{2\pi^{2}} \int_{0}^{\Omega_{\mathrm{F}}} \mathrm{d}\Omega \; T(\Omega, l)~, \end{equation} where $\Omega_{\mathrm{F}}$ is the Fermi energy of the metal. A tunnelling damping rate $\gamma_{\mathrm{g}}$ is then obtained from \begin{equation}\label{Eq:QCMGamma} \gamma_{\mathrm{g}} = \frac{\omega_{\mathrm{g}}^{2}}{4\pi \sigma_{0}(l)}~, \end{equation} where the gap plasma frequency $\omega_{\mathrm{g}}$ is assumed equal to that of the bulk metal, $\omega_{\mathrm{p}}$. These two parameters, $\omega_{\mathrm{g}}$ and $\gamma_{\mathrm{g}}$, are finally introduced into a classical EM calculation, to model the Drude permittivity of a bridge connecting the metallic components [see schematics in Fig.~\ref{fig5}(a)]. This approach has been employed to mimic the role of tunnelling-mediated dissipation in different geometries, such as NP aggregates~\cite{hohenester_prb91} or nanomatryoshkas of alternating metal and dielectric layers~\cite{zapata_oex23}. \section{Optical response of NP helices}\label{Sec:LRA} The quantum-informed models described in the previous section will be employed to explore the optical response of metallic NP helices. The helices considered in the largest part of the paper consist of 9 silver spheres with radius $R = 5$\;nm, described by the experimental dielectric function of Johnson and Christy~\cite{johnson_prb6}, in air. The spheres revolve around the $z$ axis by $\pi/2$ angle steps, so that 9 NPs produce two full revolutions, as shown schematically in Fig.~\ref{fig1}(b). While it was not modelled in our simulations, a supporting pillar of diameter $d$ and height 40\;nm is implied and shown in the schematics: such a pillar is usually produced with DNA-origami nanofabrication techniques~\cite{kuzyk_acsphoton5}. Our choice of NP size serves therefore a dual purpose: it is small enough to ensure that quantum effects will be relevant even for non-interacting NPs, while being the typical size supported by DNA-origami pillars~\cite{zhou_acr50}. The centre of each NP is vertically shifted by $R$, while $d$ can vary from 14 to 12.5\;nm, producing surface-to-surface NP distances of 3.5 to 0.3\;nm. The system is illuminated by circularly polarised light propagating either along or normally to the helix axis ($z$ axis). \begin{figure}[h] \centerline{\includegraphics*[width=1\columnwidth]{figure1.eps}} \caption{(a) Absorption cross section ($\sigma_{\mathrm{abs}}$, normalised to the geometrical cross section of a single NP), calculated within LRA for a linearly ($z$)- or circularly-polarised incident plane wave propagating along the $x$ axis, for a helix of 9 silver NPs ($R = 5$\;nm) revolving around a supporting pillar of diameter $d = 12.5$\;nm, as shown schematically in (b) (left-hand schematic). Black, red, and blue lines correspond to $z$-polarised, LCP, and RCP light, respectively. (c) Same as in (a), for propagation along the helix ($z$) axis [right-hand schematics in (b)]. (d) Zoom in the 440--480\;nm spectral window of (c), for silver NPs described by a dielectric function whose imaginary part is artificially reduced by 90\%. The spectra for linearly $x$- and $y$-polarised light are shown by a thick black and grey line respectively, while LCP (RCP) spectra are represented by a thin red (blue) line.} \label{fig1} \end{figure} In Fig.~\ref{fig1}(a) we plot the calculated absorption cross section, $\sigma_{\mathrm{abs}}$, normalised to the geometrical cross section of a single NP, as a function of wavelength $\lambda$, for a helix revolving around a pillar with $d = 12.5$\;nm, which reduces the interparticle gap to just 0.3\;nm. For the strong interaction resulting from such narrow gaps, the long-wavelength resonances in the spectra can be understood in view of the embedded-chain model as collective chain modes in the direction of the electric field of the incident plane wave~\cite{taylor_jpcc120}, while short-wavelength resonances are associated with single-NP and higher-order hybrid modes. In this respect, when the incident plane wave propagates normally to the helix axis (here along the $x$ axis), the main difference between a linearly, $z$-polarised wave with electric field $\mathbf{E} = E_{0}\;\widehat{\mathbf{z}}$ ($E_{0}$ being the amplitude of the electric field) and an L(R)CP wave [taken as $\mathbf{E} = (E_{0}/\sqrt{2}) (\widehat{\mathbf{z}} \pm \mathrm{i} \widehat{\mathbf{y}})$ here] regards the excitation of short, kinked chains along the $y$ axis in the latter case, which manifests itself as an intense absorption peak around 450\;nm in the spectra. Naturally, since the system lacks mirror symmetry, LCP and RCP illumination produces different intensities of the absorption peaks, and thus a strong CD signal, in agreement with the results of Ref.~\cite{fan_jpcc115} (emphasis will be placed on CD spectra in the next section, where the different quantum-informed models are compared). A far more interesting optical response is obtained for propagation along the helix axis, as can be seen in Fig.~\ref{fig1}(c). In this case, instead of a single long-wavelength peak as in Fig.~\ref{fig1}(a), a double peak appears in the spectral window of 440--480\;nm, centred at 452 and 464\;nm, with different intensity for each branch under LCP and RCP illumination [$\mathbf{E} = (E_{0}/\sqrt{2}) (\widehat{\mathbf{x}} \pm \mathrm{i} \widehat{\mathbf{y}})$ for L(R)CP light]. An additional, higher-order hybrid chain mode is excited at 448\;nm under $x$ polarisation, but as it always appears just as a shoulder at the high-energy end of the fundamental chain resonance~\cite{tserkezis_ppsc31}, its presence does not significantly affect the doublet of interest here. To understand the origin of this doublet, we plot in Fig.~\ref{fig1}(d) the spectra for linear polarisation along the $x$ or $y$ axis (thick black and grey line, respectively), with absorptive losses artificially reduced by 90\%, by manually modifying the imaginary part of the dielectric function of silver. It is clear that the 452\;nm mode is only excited by $y$ polarisation, while the 464\;nm only by $x$ polarisation. These resonances can therefore be understood as chain modes of embedded chains growing along the $x$ or $y$ axis, depending on the polarisation, appearing at different wavelengths due to the finite size of the helix. In an infinite helix, the number of embedded chains along both axes is the same, and the two peaks coincide. This can be verified by calculations for shorter (5 NPs, one revolution) and longer (13 NPs, three revolutions) helices, with the spectral split gradually closing, from 21 to 8\;nm, as shown in Fig.~\ref{fig2}(a). This interpretation of the modes sustained by such helices is further supported by the near-field profiles of Fig.~\ref{fig2}(b), where the highest field intensities are obtained at the gaps along the corresponding effective chain for each incident linear ($x$ or $y$) polarisation. \begin{figure}[h] \centerline{\includegraphics*[width=0.9\columnwidth]{figure2.eps}} \caption{(a) Normalised absorption spectra (vertically shifted for clarity) for the Ag NP helices schematically shown on the right. From bottom to top, LCP (red) and RCP (blue) absorption spectra are calculated for helices of 5 (one full revolution, solid lines), 9 (two revolutions, dashed lines), and 13 NPs (three revolutions, dotted lines). (b) Near-field enhancement ($|\mathbf{E}|/E_{0}$) plots for the 9-NP helices of Fig.~\ref{fig1}(c), at $\lambda = 464$\;nm. Left- and right-hand panels correspond to $x$- and $y$-linearly polarised incident light, respectively. In the top row the field is plotted in the $x-z~(y=0)$ plane, and in the bottom row in the $y-z~(x=0)$ plane, as shown in the schematics in the middle panel.}\label{fig2} \end{figure} \begin{table}[ht] \centering \caption{Character table of the $C_{2}$ point group.}\label{Table1} \begin{tabular}{|c|c c|} \hline ~ & $\quad \mathcal{E} \quad$ & $\quad \mathcal{C}_{2x} \quad$ \\ \hline $Q_{1}$ & 1 & 1 \\ $Q_{2}$ & 1 & -1 \\ \hline \end{tabular} \end{table} To further analyse the spectral doublet, we turn to group theory. Since the only symmetry operations $\widehat{P}$ leaving the finite helices unaltered are identity, $\mathcal{E}$, and a $\pi$ rotation about the $x$ axis, $\mathcal{C}_{2x}$, the appropriate point group is $C_{2}$, whose character ($\chi$) table is given in Table~\ref{Table1}~\cite{Cornwell}. All modes must have the symmetry of the irreducible representations of this group, namely $Q_{1}$ and $Q_{2}$, while a mode is excited only if the electric field of the incident light has a nonvanishing projection onto the appropriate subspace. For the $C_{2}$ group the projection operator is \begin{equation}\label{Eq:projection} \mathcal{P}^{(\Gamma)} = \frac{1}{2} \sum_{\Gamma} \chi^{\ast} (\Gamma) \widehat{P} \mathbf{E}~, \end{equation} where $\Gamma = \{\mathcal{E}, \mathcal{C}_{2x}\}$ is the set of group operations. Since a $\pi$ rotation about the $x$ axis leads to the following transformations: $x \rightarrow x$, $y \rightarrow -y$, $z \rightarrow -z$, $\mathbf{k} \rightarrow -\mathbf{k}$, it is straightforward to show that an $x$-polarised plane wave projects only onto $Q_{1}$, while a $y$-polarised plane wave only onto $Q_{2}$, in agreement with the absorption spectra of Fig.~\ref{fig1}(d). For circularly polarised light, the electric field has both $x$ and $y$ components and both modes are efficiently excited. The reported optical response is rather robust against manufacturing deviations, facilitating realisation of the proposed experiment. We have already discussed how the number of NPs forming the helix affects the measured spectra, and concluded that short chains are much more preferable than longer ones. The number of helical revolutions is in fact the factor that calls for most attention. In the Supporting Information we present spectra for intermediate NP numbers (incomplete revolutions), increasing from 5 to 9 (Fig.~S1). It is shown that in those cases the spectral doublet is always efficiently excited. We also explore the role of the revolution angle, by considering $\pi/3$ steps (Fig.~S2), which are often more feasible with DNA-origami methods~\cite{kuzyk_nat483}. Similarly, the spectra are little affected by variations in the NP size and interparticle gap (we modelled up to $\pm 10 \%$ size variations in our calculations) within the same helix (Fig.~S3). On the other hand, intrinsic losses in the chosen material can be important, as we show through the corresponding spectra for helices of 9 gold NPs (Fig.~S4). There, the resonances are not only shifted to longer wavelengths, but also broadened so much that the doublet is not distinguishable in the absorption spectra; it might survive as a sign change in CD for some chain lengths. Finally, it is important to notice that in the following analysis, and in any experimental realisation, the exact values of resonance shifts or CD measurements are not important, and only the qualitative features of the spectra matter. \section{Quantum corrections in the helix response}\label{Sec:Quantum} In the remaining of the paper we explore how the absorption doublet discussed in the previous section, and the resulting CD signal (calculated here as the difference of absorption cross sections for LCP and RCP incident light), behave within the three different quantum-informed models presented above. We start with the HDM calculation, whose screening mechanism is expected to dominate for larger NP separations. We use $v_{\mathrm{F}} = 1.39 \times 10^{6}$\;m\;s$^{-1}$ for the Fermi velocity of silver, while $\varepsilon_{\infty}$ is obtained by subtracting from the experimental dielectric function a Drude permittivity with $\hbar \omega_{\mathrm{p}} = 8.99$\;eV and $\hbar \gamma = 0.25$\;eV~\cite{tserkezis_scirep6}. In Fig.~\ref{fig3}(a) we compare the LRA absorption spectra of Fig.~\ref{fig1}(c) (dashed lines) with those obtained by HDM (solid lines). As expected, the main difference is a large blueshift of the modes due to screening, which is the characteristic of HDM: since the induced charges are ``pushed'' inwards, the NPs behave as if they were effectively smaller, and therefore their separations larger, decreasing the strength of their interaction. This can be observed more clearly in CD spectra, through the transition from negative (short-wavelength branch) to positive (long-wavelength branch) peaks, as shown in Fig.~\ref{fig3}(b). The fingerprint of the blueshifting doublet is clearly visible even for $d = 14$\;nm, corresponding to an interparticle gap as wide as 2.2\;nm, for which individual absorption peaks are hard to resolve, as they nearly coincide in wavelength with the strong single-NP resonance (chain modes are only very weakly excited). \begin{figure}[h] \centerline{\includegraphics*[width=0.9\columnwidth]{figure3.eps}} \caption{(a) Normalised absorption cross section for the helix of Fig.~\ref{fig1}(b) (right-hand schematic), for LCP (red) and RCP (blue) incident light, within the LRA and HDM description (dashed and solid lines respectively). (b) CD spectra for the two models, when increasing the interparticle gap through $d = 12.5, 13, 13.5, 14$\;nm (red, green, light-blue and blue lines respectively, corresponding to gaps of 0.31\;nm, 0.93\;nm, 1.55\;nm, and 2.18\;nm). The arrows in both panels trace the transition from the double-peak feature in absorption to the positive-negative signal in CD, highlighted here for the $d = 12.5$\;nm case (LRA calculation).}\label{fig3} \end{figure} It is now interesting to apply the same analysis for the GNOR model. To that end, we perform FEM simulations to solve the coupled Eqs.~(\ref{Eq:CoupledHydro}) (as modified for GNOR), with a diffusion constant $D = 2.684 \times 10^{-4}$\;m$^{2}$\;s$^{-1}$, a value which allows direct comparison with previous studies of silver NP dimers and chains~\cite{tserkezis_prb96}. Since GNOR is specifically designed to account for increased absorptive losses when NP sizes and distances are small, and CD is by definition an observable depending on absorption, one expects substantial differences from the response predicted by LRA. Indeed, as seen in Fig.~\ref{fig4}(a), GNOR introduces such a degree of plasmon damping and broadening, that the two separate peaks of interest merge into a single broad resonance (blueshifted with respect to LRA due to nonlocal screening, already accounted for with HDM). In this case, the only way to trace the double resonance of the helix chain modes is through its CD signal, which, even in this case, retains the characteristic sign change. This is shown in Fig.~\ref{fig4}(b), for the same four NP separations as in Fig.~\ref{fig3}(b). One should notice, however, that broadening of the modes could have many different origins, and is not necessarily an indication of Landau damping (see also the spectra for gold NP helices in the Supporting Information, Fig.~S4). Consequently, in an experiment, one should take care to choose materials with low intrinsic loss, and NPs as smooth and with as homogeneous size distributions as possible. \begin{figure}[h] \centerline{\includegraphics*[width=0.9\columnwidth]{figure4.eps}} \caption{(a) Normalised absorption cross section for the helix of Fig.~\ref{fig1}(b) (right-hand schematic), for LCP (red) and RCP (blue) incident light, within the LRA and GNOR description (dashed and solid lines respectively). (b) CD spectra for the two models, when increasing the interparticle gap through $d = 12.5, 13, 13.5, 14$\;nm (red, green, light-blue and blue lines respectively, corresponding to gaps of 0.31\;nm, 0.93\;nm, 1.55\;nm, and 2.18\;nm). The arrows indicate the wavelength at which the CD signal changes sign, in the region of the double resonance of interest.}\label{fig4} \end{figure} Finally, for gaps as narrow as a few \AA, it is important to consider the possibility of direct electron tunnelling from one NP to another~\cite{savage_nat491,yan_prl115}. As described in Sec.~\ref{Sec:models}, QCM seeks to mimic this regime through a (usually multilayered) conductive junction, introduced in a classical EM calculation to connect neighbouring NPs, as shown schematically in Fig.~\ref{fig5}(a). For a strict, quantitative description, the conductivity of such a bridge must be obtained from quantum mechanical calculations (for reduced NP sizes that make it feasible), typically with TD-DFT~\cite{esteban_natcom3}. Nevertheless, since we are only interested in a qualitative description here, we use a homogeneous cylinder described by a Drude model, with $\hbar \omega_{\mathrm{g}} = 8.99$\;eV and $\hbar \gamma_{\mathrm{g}} = 0.025$\;eV (but with $\varepsilon_{\infty} = 1$, as in the previous calculations. Omitting the multilayered structure of the bridge can be compensated by considering different bridge diameters, which, through facilitating or hindering charge transfer between NPs, tune the optical response of the entire structure. On the other hand, the tunnelling conductivity might in fact be smaller than that of the bulk metal, and therefore, the Drude damping rate might be larger, according to Eq.~(\ref{Eq:QCMGamma}), leading to broadening of the modes. It turns out that in our case this broadening is rather small, as shown by the shaded blue line ($w = 4$\;nm) in Fig.~\ref{fig5}(c) which corresponds to $\hbar \gamma_{\mathrm{g}} = 0.04$\;eV, and is practically indistinguishable from the $\hbar \gamma_{\mathrm{g}} = 0.025$\;eV case (solid line). In the Supporting Information we show how the optical response of silver NP dimers is affected by these two parameters (Fig.~S5). \begin{figure}[h] \centerline{\includegraphics*[width=0.9\columnwidth]{figure5.eps}} \caption{(a) Schematic of the bridged NP helices used for a qualitative implementation of QCM: cylindrical bridges of diameter $w$, described by a Drude dielectric function with values for $\omega_{\mathrm{g}}$ and $\gamma_{\mathrm{g}}$ appropriate for silver, connect the NPs. (b) Absorption spectra for LCP and RCP (red and blue lines) incident light, calculated for the helix of Fig.~\ref{fig1}(b) within LRA (dashed lines) and QCM (solid lines) ($w = 4$\;nm). (c) CD spectra in QCM for varying $w$, from $w = 1$ to 4\;nm (from top to bottom). The shaded line for $w = 4$\;nm represents the spectra for a higher value of $\hbar \gamma_{\mathrm{g}} = 0.04$\;eV, corresponding to lower tunnelling conductivities in the gap. The arrows indicate the wavelength at which the CD signal changes sign in the region of the double resonance of interest.}\label{fig5} \end{figure} Fig.~\ref{fig5}(b) shows the absorption spectra obtained with QCM when Drude bridges of diameter $w = 4$\;nm connect each NP in the helix with its nearest neighbours. Once such junctions are created, threaded chain plasmons (TCPs) are excited in the infrared~\cite{herrmann_natcom5,tserkezis_oex22}, at longer wavelengths for thinner bridges in which charge transfer is more hindered, as can also be seen in the CD spectra of Fig.~\ref{fig5}(c). It is important to notice that these modes are not of the same nature as the chain modes discussed above and just redshifted, but hybrid, chain/rod (or, in our case, chain/spiral) modes. The unthreaded chain modes become screened chain plasmons, getting rapidly damped and blushifting due to the screening from the field accumulated around the bridge~\cite{perez_njp13}. The two limiting cases in threaded chains are therefore i) very weak, deep-infrared modes for $w \rightarrow 0$ and ii) single spiral (threaded helix) modes for $w \rightarrow 2R$~\cite{tserkezis_oex22}. Nevertheless, if tunnelling prevails, the spectra are dominated by a strongly redshifted double peak, predicting yet another qualitatively different optical response, characteristic of QCM, the exact quantitative aspects of which (in terms of resonance shift and strength) depend on the precise TD-DFT input~\cite{esteban_natcom3}. \section{Discussion}\label{Sec:discussion} Having explored three different quantum-informed models, one might ask what the real optical response of the helix will be, and what kind of resonances one should expect in experiments. The truth however is that, before measuring, there is no real way to know which predictions are more accurate. It can be expected that for larger NP separations, screening will prevail and spectra closer to those discussed in relation to Fig.~\ref{fig3} will be obtained. On the other hand, for small separations, both damping and tunnelling could become important, leading to broad, merged absorption peaks (but probably still traceable in CD), both at shorter (screened chain plasmons) and at longer (TCPs) wavelengths. Such a combination of effects could be predicted by more elaborate models, e.g. $d$-parameter based ones. But the situation is actually reminiscent of the debate about nonlocality a few years ago. The classical HDM predicts invariably resonance blueshifts. It took \emph{ab initio} calculations~\cite{teperik_prl110}, which for small monomers and dimers were feasible, together with sensitive experiments~\cite{rainers_prl74}, to establish the fact that for good free-electron metals (e.g. Na; in noble metals spill-out is much less important due to d-electron screening) spill-out prevails over screening, thus leading to resonance redshifts. Similarly, only fine experiments (the number of electrons involved here make \emph{ab initio} calculations impossible) can tell what the actual optical response of the helices is, and in doing so they will most likely call for the extension of the existing, or the development of new quantum-informed models. Finally, let us briefly comment on the feasibility of the proposed architectures. One way of fabricating advanced plasmonic devices that exhibit novel optical properties is utilising DNA origami \cite{liu_chemrev118}. Due to its programmability and specificity, DNA origami enables NP organisation at the sub-nm scale \cite{bidault_jacs130}, that is still a challenge for top-down techniques. For instance, a single stranded DNA could be folded by smaller single DNA strands, called ssDNAs, to build a pillar with a diameter 14-12.5\;nm. It has been shown that 10-15 nucleotides, approximately 1.3\;nm in length, are enough to provide this connection \cite{heck_acsphoton4}. The pillar would be robust enough to support metallic NPs of diameter 10\;nm \cite{kuzyk_nat483}, and NPs would stick onto their predefined locations by the use of capture strands protruding from the pillar. Hence, it should be possible to obtain interparticle gaps of 0.3-3.5\;nm. Another possibility could be directly growing NPs on a DNA template \cite{shemer_jacs128}. In the case of silver in particular, oxidation can also be an issue, and experiments should be performed as soon as carefully synthesised NPs are prepared \cite{scholl_nature483}, to reduce its influence \cite{weller_acsphoton3}. \section{Conclusions}\label{Sec:conclusion} In summary, we have shown that CD measurements in chiral chains of metallic NPs, experimentally feasible nowadays with DNA-origami techniques, serve as an excellent test bed for assessing quantum-informed models in plasmonics. By comparing HDM, GNOR, and QCM, we obtained three fundamentally different optical responses, each one characteristic of the specific model. The spectral doublet calculated within LRA shifts to the blue in HDM due to screening, merges into a single resonance due to broadening in GNOR, and strongly redshifts once tunnelling becomes important in QCM. CD measurements allow to monitor these changes simply through a change in sign. Even though our calculations only intend to qualitatively describe each model, we anticipate that quantitative differences should be large enough to be experimentally resolved, thus allowing to understand the limitations and range of validity of, in principle, any existing or new quantum-informed model. \begin{acknowledgements} We thank C. Wolff for discussions. N.~A.~M. is a VILLUM Investigator supported by VILLUM FONDEN (grant No. 16498). The Center for Nano Optics is financially supported by the University of Southern Denmark (SDU 2020 funding). Simulations were supported by the DeIC National HPC Centre, SDU. \end{acknowledgements}

4108470564862f749989c90bfa438374396820fe

\section{Introduction} As machine learning models are increasingly being used to make decisions that directly impact people, the risk of fraudulent attempts to fool these models is more present than ever. Let us consider the case of an insurance company with an automated service that allows customers to apply for credit insurance. Through a dedicated interface, clients have to provide personal details so a machine learning prediction model can assess their risks of default and decide automatically whether they are eligible for the insurance. A potential customer may want to improve his chances of getting the credit insurance and could be tempted to game the system by fooling the decision model. The malicious customer could lie to some of the questions he is being asked, such as his salary or age. These alterations will affect the input provided to the automatic classifier and may change the prediction. Such attacks could provoke major issues for an insurance company: an inaccurate risk assessment can lead to a possibility of bankruptcy. Then, how is it possible to prevent these attacks, and more precisely, what defense strategy should be applied against these attacks? In this paper, we address the following issue: given an automated classifier and a potential customer, our objective is to identify a sparse set of features that are the most important for the classifier's prediction made for this customer and thus the ones that could easily alter the prediction. Given these most important features, it is possible to ask the potential client for additional information or proof, in order to counter a potential adversarial attack by ascertaining the truthfulness of his declarations. The proposed approach is framed at the intersection of adversarial machine learning and human-interpretable machine learning. Given a point in a feature space (\emph{ie.} a potential client) we propose to search the closest decision boundaries of the classifier, which we assume to be the most relevant for the point's prediction, and fit an interpretable surrogate model to approximate these boundaries. The most important features for this prediction are then extracted from the surrogate. These features lead to checks in order to prevent a potential adversarial attack. This defense approach is local (\emph{ie} centered around the prediction to assess), focused on classical feature-based classification problems and model-agnostic (\emph{i.e.} independent to the choice of the black-box classifier), since we assume that no information about the classifier is made available for the proposed approach. The next section details the relevant background for this work. Section~\ref{sec:proposition} presents a formalization of the considered problem and a description of the principles of our proposition. Section~\ref{sec:experiment} proposes a first evaluation of our proposition on the German Credit dataset: we quantify the fidelity to the classifier and we provide an example of explanation that can be used as a defense against an attack by clarifying whether an input is an adversarial example or not. We conclude this paper by outlining the perspectives to pursue this work. \section{Background} In adversarial machine learning, evasion attacks~\cite{Biggio2013,Goodfellow2015,Papernot2016limitations} aim at generating adversarial examples that fool a machine learning classifier. Defense techniques have been developed to counter these attacks: either by changing the training process of the classifier to be attacked to make it more robust to small perturbations~\cite{Papernot2017practical,Bhagoji2017} or by adding a detection mechanism to identify potential adversarial examples~\cite{Cao2017}. However, both strategies generally rely on strong assumptions or knowledge, either about the dataset or about our capacity to identify the \textit{true} label of an instance. For instance, when dealing with images, a human is able to provide at test time the true label despite an adversarial perturbation. In the mentioned insurance context, the classifier's output for a potential adversarial example can't be compared with the true label since it is unknown (eg. a default would happen in the future). To mitigate this issue, we look into the classifier's inner working to assess the potential weaknesses of a client's application, and control the veracity of problematic information he may have provided. To do so, the field of machine learning interpretability aims at generating explanations to provide insights about a prediction made by a black-box classifier. Explanations can take the form of feature importance vectors that quantify the relative impact of each feature in the prediction~\cite{Baehrens2010,Ribeiro2016,Lundberg2016}. To extract these feature importances, we focus on surrogate model approaches which attempt to approximate the decision boundary of a black-box classifier with a simpler \textit{interpretable} model either at global scale (to mimic the global behavior of a black-box model)~\cite{Craven1996,Hara2016a} or at local scale (to mimic the black-box behavior locally, around a prediction for instance)~\cite{Ribeiro2016,laugel2018}. Surrogate model approaches can somewhat be related to the task of model theft in adversarial learning, which consists in building a substitute model to copycat the decisions of a black-box classifier, either to steal it and use its outputs~\cite{Tramr2016StealingML} or to use it for transferable evasion attacks~\cite{Papernot2017practical}. \section{Proposition: Local Adversarial Detection (LAD)}\label{sec:proposition} The following context is considered: a black-box classifier $b:\mathcal{X}\rightarrow \mathcal{Y}$ is trained on a dataset composed of an input feature space $\mathcal{X} = \mathcal{R}^D$ made of classical features and an output $Y$ where each instance $x \in \mathcal{X}$ is associated with a class label $y \in \mathcal{Y}$. The classifier $b$ is publicly available for querying and accepts an input vector $x$ (e.g. containing information such as age and income) to provide predictions $b(x)$ that lead to business decisions (such as insurance or credit acceptance). For security reasons, each input $x$ leading to a prediction $b(x)$ is considered suspicious: it is considered as being a potential adversarial example $\hat{x}$ (\emph{i.e.} fake information), generated to get a more favourable decision from the black-box classifier $b$ such that $b(x) \neq b(\hat{x})$. In the context of insurance, it is assumed that the provided input data is reasonable since it deals with real or physical variables. In this work, our objective is to design a security layer to provide an adaptive defense that returns the most important features to check in order to clarify whether or not an input $\hat{x}$ should be investigated. Given a prediction $b(\hat{x})$ and a potential adversarial example $\hat{x}$ to assess, we aim at providing the locally most important features for the black-box classifier that lead to the prediction $b(\hat{x})$. These features are also the most sensitive for someone willing to fool the classifier~$b$, and can be investigated by a human expert or a machine to ask for a relevant set of proofs to avoid a fraud. For instance, given $\hat{x}$, if the income is the only important feature that would significantly impacts the prediction $b(\hat{x})$, then a single proof of income would be necessary to counter an adversarial attack. Our objective is thus to identify a set $\mathcal{F}_{\hat{x}}$ of features from $\mathcal{X}$ that are key for the prediction $b(\hat{x})$ and should be checked to clarify whether $\hat{x}$ is a malicious customer or not. To solve that problem, we propose to fit a local surrogate in order to approximate the closest local black-box boundaries to the potential adversarial example $\hat{x}$. Then, the set $\mathcal{F}_{\hat{x}}$ of features that are key for the prediction $b(\hat{x})$ is extracted from the local surrogate. The proposed Algorithm~\ref{algo} and its different steps illustrated Figure~\ref{fig:algo} work as follows. To detect the closest decision boundaries, $N$ support points $x_{sp}^i$ are drawn in $\mathcal{X}$ with $b(x_{sp}) \neq b(\hat{x})$ (Figure~\ref{fig:algob}). These support points delimit segments $[\hat{x} ; x_{sp}^i]$~$ \forall i \in [1;N]$ on which the local black-box boundary should be sought: to do so, the maximum of the information gain is sought on every segment $[\hat{x} ; x_{sp}^i]$ based on $M$ points drawn on these segments then labeled using $b$ (Figures~\ref{fig:algoc} and~\ref{fig:algod}). The point on each segment where the information gain is maximal is called a boundary touchpoint $x_{bt}$. These points are on the black-box boundary or close to it: it is possible to outline locally the black-box boundary shape. Then, an interpretable local surrogate classifier $s_{\hat{x}}$ is trained on a small set of points generated in the immediate neighbourhood of the local black-box boundary (Figures~\ref{fig:algoe} and~\ref{fig:algof}), outlined by the boundary touchpoints $x_{bt}$. Finally, the set of key features $\mathcal{F}_{\hat{x}}$ for the prediction $b(\hat{x})$ is extracted from the surrogate $s_{\hat{x}}$ that approximates the local boundary of the classifier $b$. \begin{figure}[t] \centering \begin{subfigure}{0.31\textwidth} \includegraphics[width=\textwidth]{01_x_toexplain.png} \caption{Potential adversarial example $\hat{x}$ (red dot) over black-box classifier's boundaries} \label{fig:algoa} \end{subfigure} ~ \begin{subfigure}{0.31\textwidth} \includegraphics[width=\textwidth]{02_support_points.png} \caption{Random generation of support points $x_{sp}$ where $b(x_{sp}) \neq b(\hat{x})$} \label{fig:algob} \end{subfigure} ~ \begin{subfigure}{0.31\textwidth} \includegraphics[width=\textwidth]{03_segments.png} \caption{Random generation and classification of points on segments $[\hat{x};x_{sp}]$} \label{fig:algoc} \end{subfigure} \begin{subfigure}{0.31\textwidth} \includegraphics[width=\textwidth]{04_touchpoints.png} \caption{Boundary search: segment points that maximize the information gain} \label{fig:algod} \end{subfigure} ~ \begin{subfigure}{0.31\textwidth} \includegraphics[width=\textwidth]{05_touchpoint_augmented.png} \caption{Generation and classification of points around the boundary touchpoints} \label{fig:algoe} \end{subfigure} ~ \begin{subfigure}{0.31\textwidth} \includegraphics[width=\textwidth]{06_surrogate.png} \caption{Boundaries of the trained interpretable surrogate} \label{fig:algof} \end{subfigure} \caption{Principle of Local Adverse Detection (LAD)} \label{fig:algo} \end{figure} \begin{algorithm}[t] \begin{algorithmic} \STATE {\bfseries Input:} potential adversarial example $\hat{x} \in \mathcal{X}$, classifier $b:\mathcal{X}\rightarrow\mathcal{Y}$, untrained surrogate $s_x$, $N$, $M$ \STATE $x_{sp} \leftarrow$ Draw $N$ support points $x_{sp}^i \in \mathcal{X}, i \in [1...N]$ with $b(x_{sp}^i) \neq b(\hat{x})$ \FORALL{$x_{sp}^i \in x_{sp}$} \STATE $x_{seg}^i \leftarrow$ Draw $M$ segment points $x_{seg}^{i,j}, j \in [1...M]$ with $x_{seg}^{i,j} \in [\hat{x};x_{seg}^i]$ \STATE $y_{seg}^{i} \leftarrow b(x_{seg}^{i})$ label every segment points with the classifier $b$ \STATE $x_{bt}^{i} \leftarrow$ Find the touchpoint $x_{bt}^{i} \in [\hat{x};x_{seg}^i]$ that maximizes the info gain $IG(y_{seg}^{i})$ \STATE $X_{s_x^i} \leftarrow$ Draw $P$ points in an hypersphere of radius $r_{s_x}$ around the touchpoint $x_{bt}^{i}$ \STATE $Y_{s_x^i} \leftarrow$ Label every points $X_{s_x^i}$ with the classifier $b$ \ENDFOR \STATE $Y_{s_x} \leftarrow b(X_{s_x})$ \STATE Train the local surrogate $s_x$ on ($X_{s_x}, Y_{s_x}$) \STATE $\mathcal{F}_{\hat{x}} \leftarrow$ Extract the most important features from $s_x$ \STATE {\bfseries Return:} $\mathcal{F}_{\hat{x}}$ \end{algorithmic} \caption{Outline of Local Adverse Detection (LAD) algorithm} \label{algo} \end{algorithm} \vspace{-1.5em} \section{Case Study: German Credit Dataset}~\label{sec:experiment} We present a first use case towards designing a complete and robust experimental protocol to evaluate the quality of the proposed approach. This work is still in progress and requires further discussion. We apply the proposed approach to the German Credit dataset available from UCI. A classifier $b$ (a Random Forest with 200 estimators) is trained on 70\% of the data and acts as the automated decision model accepting or rejecting the customer's application. Considering information provided by customers $\hat{x}$ from the remaining test dataset (250 instances), the LAD algorithm is used (with $N=1000$, $M=100$ and the surrogate being a decision tree with a maximum tree depth of 5) to extract the key features for prediction $b(\hat{x})$. They constitute the features an expert should check to ascertain the prediction. Our first experiment assesses the accurate approximation of the black-box classifier's local decision boundary by the local surrogate trained with LAD. The \emph{Local Fidelity} metric described in~\cite{laugel2018} is used to assess locally, around the potential adversarial example, the fidelity of the surrogate $s_{\hat{x}}$ to the classifier $b$. The \emph{Local Fidelity} metric is defined as the fidelity of the local surrogate $s_{\hat{x}}$ to the black-box classifier $b$ within a neighborhood $\mathcal{V}_{\hat{x}}$ around $\hat{x}$. $|\mathcal{V}_x|$ points are therefore drawn uniformly in the neighbourhood $\mathcal{V}_{\hat{x}}$ of $\hat{x}$ bounded by an hypersphere of radius $r$. Then, to get the local fidelity of $s_{\hat{x}}$ to $b$, the classification accuracy is computed on these points labeled with both $s_{\hat{x}}$ and $b$, such as: \begin{equation} LocalFid(\hat{x}, s_{\hat{x}}) = Acc_{x_i \in \mathcal{V}_{\hat{x}}}(b(x_i), s_{\hat{x}}(x_i)) \end{equation} We set $|\mathcal{V}_x|=1000$ and the radius of the hyperspheres as a percentage (between $0.05\%$ and $0.5\%$) of the maximum distance between $\hat{x}$ and the test dataset. To challenge the proposed approach, we choose to use the substitute model method from \cite{Papernot2017practical}, based on data generated using Jacobian augmentation. The average local fidelity values over all the test dataset are shown Figure~\ref{fig:localfidelity_german}. As mentioned earlier, an important difference of \cite{Papernot2017practical} compared to the proposed approach lies in the fact that the trained substitute is global instead of focusing on a specific region. Then, as shown Figure~\ref{fig:localfidelity_german}, the average local fidelity of LAD is higher for smaller radius values as expected. However, as the radius increases and the evaluation of the fidelity of the substitutes is performed on a wider scale, the difference vanishes. \begin{figure}[t] \center \includegraphics[width=0.75\textwidth]{local_fidelity_agg.pdf} \caption{Local Fidelity for LAD and Papernot2017~\cite{Papernot2017practical} on German Credit dataset for several values of $r$} \label{fig:localfidelity_german} \end{figure} An important contribution of this proposition consists in the generation of explanations given a prediction for a potential adversarial example. A very first illustration is provided in Table~\ref{table:adverse_fi}. We consider a customer application to a credit described by vector $\hat{x}$, generated from $x$, the true unknown customer application: only the feature "Age in years" has been manipulated, from 47 to 46 in order to allow the applicant to get accepted for the credit by the automated classifier. In order to detect the fraud, the LAD algorithm has been used to generate a list of features that are locally important for the black-box model $b$, which could have been used to manipulate the outcome of $b$. As expected, the age feature has a high feature importance and is the first one a client could easily manipulate. LAD would tell a human expert to check in priority this feature for this illustrative example. \begin{table}[t] \centering \caption{LAD's feature importance for an adversarial example generated on German Credit given a classifier $b$. LAD is able to identify the importance of \textit{Age in years} for the classification} \begin{tabular}{|c||c|c|c|c|c|c|} \hline Features & Credit amount & Credit history & \textbf{Age in years} & Housing & Foreign worker & ... \\ \hline \begin{tabular}[c]{@{}c@{}}Importance\end{tabular} & 0.43 & 0.28 & \textbf{0.10} & 0.09 & 0.07 & ... \\ \hline \end{tabular} \label{table:adverse_fi} \end{table} \vspace{-1.5em} \section{Conclusion and Perspectives} This paper proposes to discuss a use-case for adversarial defense. It consists in the detection of local adversarial examples for feature-based automated decision problems. Credit insurance provides an example where some people may be tempted to manipulate information they have to provide to increase their chances to get accepted. While this paper describes a preliminary work, we proposed a method to detect potential adversarial examples by identifying features that have the most impact locally, over the prediction made by a black-box classifier. This method trains an interpretable surrogate to approximate the local decision boundary of the black-box classifier and extract feature importances. These features are provided to a human expert in charge of checking the corresponding information to ascertain the prediction. A first application on German Credit is used to illustrate our approach: the local fidelity of the method to the classifier it approximates is assessed and the important features to check are extracted for a generated adversarial example. Our on-going work aims first at consolidating the proposed use-case and formalization of the problem. We plan to improve and test in-depth our proposition (LAD) with a robust experimental protocol, in particular with adversarial examples generated by evasion attack methods to quantify to what extent the approach can help to detect adversarial examples in such context. \nomenclature{$\hat{x}$}{Potential adverse} \nomenclature{$b$}{Black-box classifier} \nomenclature{$\mathcal{X}$}{Feature space domain} \nomenclature{$X$}{Set of instances $X\in\mathcal{X}$} \nomenclature{$Y$}{Set of labels $Y = b(X)$} \nomenclature{$s_{\hat{x}}$}{Local surrogate for $\hat{x}$} \nomenclature{$X_{sp}$}{Set of support points $x_{sp}^i \in X_{sp}$} \nomenclature{$X_{seg}^i$}{Set of segment points $x_{seg}^{i,j} \in X_{seg}^i$ drawn on $[\hat{x};x_{sp}^i]$} \nomenclature{$X_{bt}$}{Set of boundary touchpoints $x_{bt}^k \in X_{bt}$} \nomenclature{$X_{s_{\hat{x}}}$}{Training set for $s_{\hat{x}}$ drawn in hyperspheres around each $x_{bt}^k$} \nomenclature{$Y_{s_{\hat{x}}}$}{Set of labels $Y_{s_{\hat{x}}} = b(X_{s_{\hat{x}}})$} \nomenclature{$\mathcal{F}_{s_{\hat{x}}}$}{Set of most important features from $s_{\hat{x}}$}

ac313811cc176db4f3f1e3812ff2e494e6994cfe

\section{\label{sec:intro}Introduction} The current state-of-the-art image recognition algorithms are based on convolutional neural networks \cite{he_deep_2015, krizhevsky_imagenet_2012, simonyan_very_2014}. These networks rely on large datasets of labeled images, to simultaneously generate a feature representation and a decision framework. This approach removes the need to handcraft features for any given problem but also gives rise to the question as to what feature representation the network has actually learned. These models are trained on large datasets that contain a number of biases \cite{tommasi_deeper_2015, torralba_unbiased_2011} or spurious variations, that are irrelevant, or even problematic, for a given task (e.g. discriminating by gender or ancestral origin). One such example is face recognition. Large publicly available face datasets are often composed of celebrities, such as \cite{cao_vggface2, msceleb1m, LFWTech, megaface, kostinger_annotated_2011, parkhi_deep_2015}. These can contain age, gender, and ancestral origin biases: for example, female celebrities tend to be younger than their male counterparts. This bias does not represent the real world outside of the movie business, reducing the usefulness of models trained on these datasets. Furthermore, in cases where large datasets are not available, training algorithms are initialized from networks that have been trained on similar tasks, for which more data is available \cite{girshick_rich_2013, long_learning_2015, oquab_learning_2014}. This method, called fine-tuning, carries two potential issues: 1) spurious variations are learned from the small dataset, and 2) inheriting biases present in the large dataset. The use of big data to train AI-models has, to name a few, been adopted by government agencies, institutions of law, human resources, and medical decision systems. A number of such models have been shown to make decisions based on the gender or ancestral origin of an individual, leading to concerns about their ``fairness'' \cite{pmlr-v81-buolamwini18a, ras_explanation_2018, sweeney_discrimination_2013}. With the recent enforcement of the General Data Protection Regulation laws\footnote{\url{https://www.eugdpr.org}}, individuals have the right to know the rationale behind an automated decision concerning them. To continue the adoption of deep learning, more needs to be done to make neural networks more transparent. One way to approach this issue is to prevent models from making decisions for the wrong reasons. To be sure that decisions are not being made due to biases, we must look beyond using accuracy as our only performance metric. An experiment by Zhao et al \cite{zhao_men_2017} showed that neural networks learn and amplify biases in the training data. Women were more often depicted in kitchens than men, so the network learned that being in the kitchen was a key feature for the identification of women. Though this may have been true for the dataset the classifier was trained on, in general, this is not a reliable indicator of the presence of a woman. In this paper, we introduce an algorithm, inspired by a domain and task adaptation approach \cite{tzeng_simultaneous_2015}, that can be used to 1) to ensure that a network is blind to a known bias in the dataset, 2) improve the classification performance when faced with an extreme bias, and 3) remove multiple spurious variations from the feature representation of a primary classification task. We use age, gender, ancestral origin, and pose information for facial images to demonstrate our framework. As discrimination by ancestral origin is a type of spurious variation in many tasks, we have created a new labeled ancestral origin dataset, ``Labeled Ancestral Origin Faces in the Wild (LAOFIW)'', from publicly available images. This dataset contains 14,000 images of individuals whose ancestral origin is sub-Saharan Africa, Indian Subcontinent, Europe, and East Asia, with a variety of poses, illumination, facial expressions. The rest of this paper is organized as follows. Section \ref{sec:data} introduces our LAOFIW dataset and other datasets we used for our experiments. Section \ref{sec:methods} discusses the methods to remove spurious variations from the feature representation of a network. In section \ref{subsec:expremovebias}, we present an experiment to remove a bias from the feature representation network. In section \ref{subsec:expremovepbias}, we investigate how removing an extreme bias can improve classification performance, and in section \ref{subsec:expremovemulti}, we investigate removing multiple spurious variations from a network. The results are detailed in section \ref{sec:results}. Finally, section \ref{sec:conclusion} summarises our findings. \section{\label{sec:relatedwork}Related Work} Image datasets are known to contain biases that cause models to generalize poorly to new, unseen data \cite{tommasi_deeper_2015, torralba_unbiased_2011}. This has been addressed by domain adaptation that aims to minimize the difference between the source and target domains \cite{gretton_covariate_2009}. Domain adaptation has been shown to improve the generalisability of classifiers for the same task across different domains \cite{tzeng_simultaneous_2015, tzeng_adversarial_2017}. We draw inspiration from domain adaptation methods to make the network agnostic to spurious variations. Learning feature representations that are invariant to certain spurious variations has been tackled in a number of ways. \cite{yin_multi-task_2018} take a multi-task learning approach to pose-invariant face recognition. They propose a novel approach to automatically weight each of the spurious variations during training. Another approach is to adjust the training data distribution at training time, to avoid learning biases \cite{zhao_men_2017}. This method relies on having labels for each of the spurious variations for each training instance. This is not feasible for most existing datasets, as they tend to be labeled with a single task in mind, where information about spurious variations is not available. Our method can make use of separate datasets, each labeled for distinct tasks, to remove multiple spurious variations simultaneously. Jia et al \cite{jia_right_2018} and Raff et al \cite{raff_gradient_2018} draw inspiration from \cite{ganin_domain-adversarial_2015} to remove a source of variation with the use of a gradient reversal layer to update the network in opposition of a task. Instead of applying gradient reversal on the output of the softmax layer, which penalizes correct classifications, we compute the cross-entropy of the output classifier and a uniform distribution, as in \cite{tzeng_simultaneous_2015, tzeng_adversarial_2017}. This ensures an equally uninformative classifier across all tasks. \section{\label{sec:data}Datasets} \subsection{Labeled Ancestral Origin Faces in the Wild} \label{subsec:LAOFIW} A new ancestral origin database was created as part of this work called ``Labeled Ancestral Origin Faces in the Wild (LAOFIW)''. The aim of this dataset was to 1) create a dataset for experimentation, and 2) be used as a spurious variation dataset for applications where training a network to be agnostic to ancestral origin is important. The database was assembled using the Bing Image Search API. Search terms based upon origin, e.g. ``German, English, Polish, etc'', were submitted in conjunction with the words ``man, woman; boy, girl''. Additionally, results were filtered to return photographic images of faces. In total, 43 origin search terms were queried returning 20,000 images. Duplicates were removed by comparing their Histogram of Oriented Gradients \cite{dalal_histograms_2005} encoded using a Gaussian mixture model for each image. The remaining 14,000 images, were manually divided into four broad ancestral origins: sub-Saharan Africa, Indian Subcontinent, Europe, and East Asia. These classes were selected on the basis of being visually distinct from each other. The database contains roughly the same number of male and female individuals. It also contains images with varied poses: more than a third of the images are non-frontal. Sample images are shown in figure \ref{fig:LAOFIWexamples}. \begin{figure}[h] \begin{center} \includegraphics[width=0.8\textwidth]{images/LAOFIW.jpg} \caption{Example images from the LAOFIW dataset for each of the four classes. In rows top to bottom: Sub-Saharan Africa, Europe, East Asia, Indian Subcontinent. The images are highly varied in age, gender, pose, lighting and expression.} \label{fig:LAOFIWexamples} \end{center} \vspace{-0.5cm} \end{figure} \subsection{Age and Gender Dataset} \label{subsec:imdb} The IMDB dataset is a large publicly available dataset of the images contained in the profiles of the 100,000 most popular actors on the IMDb website\footnote{\url{https://www.imdb.com/}} with gender and date of birth labels provided for each instance \cite{rothe_dex_2015}. We used this dataset to investigate the effects of age and gender bias in celebrity datasets. The labels in the IMDB dataset are noisy, with a number of individuals having both incorrect age and gender labels. This is due to the nature of how the data was collected: The authors assumed that images from an actor's profile that contained a single face would show the actor in question. As stated by the authors, however, these images often contain other actors, who co-starred in their movies \cite{rothe_dex_2015}. As gender and date of birth are taken from the profile of the actor, this causes erroneous labels for images that show co-stars. The age of an individual was calculated as the difference between the timestamp of the photo and the date of birth of the subject. In some cases, the time stamps of photos predate an actor's date of birth or are otherwise unreliable. To mitigate this problem, we used the Microsoft Azure Face API\footnote{\url{https://azure.microsoft.com/en-gb/services/cognitive-services/face/}} to extract gender and age estimates for an identity-balanced subset of 150,000 images. We rejected all images in which the predicted gender from Azure and the IMDB gender label disagreed. We then ran the analysis for age and removed images in which the Azure age prediction differed by more than 10 years from the IMDB labels. The resulting, cleaned dataset contained 60,000 images. In order to quantify the effectiveness of this data cleansing procedure, we trained a gender classification VGG-M network \cite{simonyan_very_2014} on both the original 150,000 images and the cleaned 60,000 images, withholding 20\% of the images for testing. The gender classification accuracies on the test images improved from 75\%, before cleaning, to 99\% after cleaning. The distribution of ages for each gender in the cleaned IMDB dataset is shown in figure \ref{fig:imdbhist}. We can observe a bias towards younger women and older men in the data. A subset of the cleaned data was used to create an unbiased test dataset, which contains equal numbers of men and women for each age category. This unbiased test dataset was used to evaluate network bias and was not used during training. \begin{figure}[h] \begin{center} \includegraphics[width=0.5\textwidth]{images/age_histograms.png} \caption{Age distributions and differences for women (red) and men (blue) in the IMDB dataset. The dataset bias towards younger women and older men is visible.} \label{fig:imdbhist} \end{center} \vspace{-1cm} \end{figure} \subsection{Pose Dataset} \label{subsec:aflw} We used the AFLW dataset \cite{kostinger_annotated_2011} categorized by yaw, to train the pose classifier mentioned in section \ref{subsec:LAOFIW}. The poses were binned into five categories, ``profile left, near-frontal left, frontal, near-frontal right, profile right''. The non-frontal images were duplicated and flipped to augment the dataset. A class-balanced subset of 24,000 images was selected for training, and a class-balanced set of 6,000 images was reserved for testing. \section{\label{sec:methods}Methods} We introduce a supervised-learning algorithm that aims to learn a single feature representation $\theta_{\text{repr}}$, that is informative for a primary classification task, whilst simultaneously being uninformative for a number of spurious variations, that represent undesirable sources of variation. For example, we may wish to create an age classifier, that does not base its decisions on any pose, ancestral origin, or gender information. We assume access to a primary dataset $\mathcal{D}_p = \left\{x_i, y_i\right\}_{i=1}^{n_p}$, containing $n_p$ images, $x_i$, with labels $y_i \in \{1, ..., K\}$. And similarly, we assume access to $M$ secondary datasets, $\mathcal{D}_s = \{\mathcal{D}_m\}_{m=1}^M$, each describing a single spurious variation. \subsection{Joint Learning and Unlearning} \label{subsec:learning} We introduce a joint learning and unlearning (JLU) algorithm to learn a primary classification task, whilst simultaneously unlearning multiple spurious variations. Our convolutional network (CNN) architecture is depicted in Figure \ref{fig:net}. The primary branch has a single classification loss, the primary classification loss, which assesses the ability of the network to accurately distinguish between classes of the primary task. Each secondary branch has two losses: a classification loss and a confusion loss. These losses are used to, in turn, assess the amount of spurious variation information left in the feature representation $\theta_\text{repr}$ and then remove it. \begin{figure}[h] \begin{center} \includegraphics[width=0.8\textwidth]{images/network.png} \caption{Overview of CNN Architecture: We use a joint loss over primary and secondary data to learn a feature representation that simultaneously learns to classify the primary task but becomes invariant to secondary tasks, the spurious variations. The spurious variation classification loss and confusion loss act in opposition to learn the classifier on the feature embedding and change the feature embedding to confuse the classifier, respectively. The base architecture is a VGG-M network \cite{simonyan_very_2014}.} \label{fig:net} \end{center} \vspace{-0.3cm} \end{figure} Let the classification objective for a generic task with $K$ classes and a corresponding classifier $\theta_c$, given a feature representation $\theta_{\text{repr}}$, be defined as the standard softmax loss: \begin{equation} L_{\text{softmax}}(x, y; \theta_{\text{repr}}, \theta_{C}) = - \sum_{k=1}^K \mathds{1} [y = k] \log p_k, \label{eq:class} \end{equation} where $p_k$ is the softmax of the classifier's output assigning the input $x$ to class $k$. We will refer to this loss evaluated on the primary task as $L_{P}(x_p, y_p; \theta_{\text{repr}}, \theta_{P})$ and on the $m$-th spurious variation as $L_{m}(x_{m}, y_{m}, \theta_{\text{repr}}; \theta_{m})$. Inspired by \cite{tzeng_simultaneous_2015}, we introduce a confusion loss for the $m$-th spurious variation, $L_{\text{conf}}(x_m, y_m, \theta_m; \theta_{\text{repr}})$, in (\ref{eq:conf}). Minimizing the confusion loss seeks to change the feature representation $\theta_{\text{repr}}$, such that it becomes invariant to the spurious variations. \begin{equation} L_{\text{conf,m}}(x_m, y_m, \theta_{m}; \theta_{\text{repr}}) = - \sum_{n_m} \frac{1}{n_m} \log p_{n_m} . \label{eq:conf} \end{equation} We compute the best classifier for each spurious variation, $\theta_{m}$, and then compute the entropy between the output predicted from each of these classifiers and a uniform distribution. The complete method minimizes the joint loss function: \begin{equation} \begin{aligned} L(x_p,y_p,x_s,y_s,\theta_p, \theta_s, \theta_{\text{repr}}) &= L_{p}(x_p, y_p; \theta_{\text{repr}}, \theta_{p}) \\ &+ L_{s} + \alpha L_{\text{conf}}, \label{eq:joint} \end{aligned} \end{equation} where $\alpha$ determines how strongly the confusion loss affects the overall loss, $\theta_s = \{\theta_1,...,\theta_M\}$, and \begin{align} L_{\text{conf}} &= \frac{1}{M} \sum_{m=1}^M L_{\text{conf}, m}(x_m, y_m, \theta_{m}; \theta_{\text{repr}}) \label{eq:secondary_conf_loss} \\ L_s &= \sum_{m=1}^M \beta_m L_m(x_m, y_m, \theta_{\text{repr}}; \theta_{m}), \label{eq:secondary_classif_loss} \end{align} where $\beta_m$ is a weight assigned to the $m-$th spurious variation. As mentioned in \cite{tzeng_simultaneous_2015}, the confusion loss (\ref{eq:secondary_conf_loss}) and the spurious variation classification loss (\ref{eq:secondary_classif_loss}) stand in opposition to one another, so they cannot be optimized in the same step. Therefore, we switch between training the spurious variation classification loss, $L_{s}$, and then the joint primary and confusion loss, $L_{p}(x_p, y_p; \theta_{\text{repr}}, \theta_{p}) + \alpha L_{conf}$. At each iteration, we find the best spurious variation classifier for the feature representation. The training procedure is shown in algorithm \ref{alg:joint}. We used the VGG-M architecture \cite{chatfield_return_2014}, which consists of five convolutional layers (conv1-5) and three fully connected layers (fc6-8). The feature representation parameters $\theta_{\text{repr}}$ represents layers conv1-fc7. \begin{algorithm}[h] \caption{Joint Learning and Unlearning}\label{euclid} \begin{algorithmic}[1] \Procedure{$\min$ L}{$x_p,y_p,x_s,y_s,\theta_p, \theta_s, \theta_{\text{repr}}$} \For{epochs} \While{$\frac{dL_s}{d\theta_s} \not= 0$} \State $\min L_s$ \EndWhile \State $\min L_p + \alpha L_\text{conf}$ \EndFor \EndProcedure \end{algorithmic} \label{alg:joint} \end{algorithm} \vspace{-0.5cm} \section{\label{sec:experiments}Experiments} In this section, we present experiments to demonstrate possible applications of our methodology and the datasets we tested them on. \begin{description} \item [Removal of a bias from a network ---] We train a gender-agnostic age classifier using the IMDB dataset, which contains a gender bias: female celebrities tend to be younger than their male counterparts in this dataset. \item [Removal of an extreme bias from a network ---] We train an age-agnostic gender classifier on subsets of the IMDB dataset that contain only young women and only old men, and vice versa. \item [Simultaneous removal of multiple spurious variations ---] We demonstrate our JLU algorithm's ability to simultaneously remove multiple spurious variations from the feature representation of a network trained for a primary classification task. \end{description} \subsection{Removal of a bias from a network} \label{subsec:expremovebias} We investigated the task of creating an age classifier using the cleaned IMDB dataset described in section \ref{subsec:imdb}. We hypothesize that the gender bias in the distribution, where men are generally older than their female counterparts in this dataset, will be learned by the network. We trained two networks to perform this task: 1) baseline --- trained solely on age data, 2) blind --- trained on age data, whilst removing gender-specific information from the network. We evaluated both networks on the unbiased test dataset detailed in section \ref{subsec:imdb}. We compared accuracies for age classification and prediction distributions for both genders. \subsection{Removal of an extreme bias from a network} \label{subsec:expremovepbias} We train gender classifiers using the following artificially biased subsets of the cleaned IMDB data from section \ref{subsec:imdb}: \begin{enumerate} \item Extreme bias 1 (EB1): women aged 0-30, men aged 40+ \item Extreme bias 1 (EB2): women aged 40+, men aged 0-30 \end{enumerate} We placed a buffer of 10 years without any individuals between the two sets to exaggerate the age difference between the genders. We hypothesize that the age bias in the distribution will be learned by the network, for example, if trained on EB1, it should falsely learn to predict that young men are in fact women. Histograms of these datasets are given in figure \ref{fig:poison}. We trained a baseline network on gender only, and a blind network using our JLU algorithm to unlearn age. The trained classifiers were evaluated on the unbiased test dataset detailed in section \ref{subsec:imdb}. \begin{figure}[h] \centering \begin{subfigure}[b]{0.48\textwidth} \includegraphics[width=\textwidth]{images/imdb_azure_gender_alt_co_4_b_2_poisoned_balanced.jpg} \caption{EB1: women aged 0-30, men aged 40+} \end{subfigure} ~ \begin{subfigure}[b]{0.48\textwidth} \includegraphics[width=\textwidth]{images/imdb_azure_gender_alt_co_4_b_2_poisoned_inv_balanced.png} \caption{EB2: women aged 40+, men aged 0-30} \end{subfigure} \caption{Histograms showing the ages of women (red) and men (blue) in the artificially biased datasets. In both cases, individuals aged under 35 were selected for the young, and individuals aged above 45 were selected for the old set.} \label{fig:poison} \end{figure} \subsection{Simultaneous removal of multiple spurious variations} \label{subsec:expremovemulti} We demonstrate our algorithm's ability to remove multiple spurious variations from the feature representation of a network. The primary task and spurious variations are selected from age, gender, ancestral origin and pose. In each case, one task is selected as the primary task, and the others make up the spurious variations that we wish to remove. A baseline network was trained for each primary task without using our JLU algorithm. We used the cleaned IMDB dataset described in section \ref{subsec:imdb} for age and gender labels. We used our LAOFIW dataset, described in section \ref{subsec:LAOFIW} to classify ancestral origin. Finally, we used the adapted AFLW dataset (Section \ref{subsec:aflw}) to classify pose. \subsection{\label{subsec:age}Age Classification} Since our dataset is relatively small for a deep learning task, we do not approach age estimation as a regression task. We conduct age classification by creating age bins of 5 years, or more years at either end of the distribution. The categories are shown in figure \ref{fig:imdbhist}. We define predictions within one class from the true age class as a positive classification to account for errors caused by edge cases in different bins. \subsection{Implementation details} Our base network is an adapted VGG-M network \cite{simonyan_very_2014}, pre-trained on the VGG-Face dataset \cite{parkhi_deep_2015}. In our experiments, we saw no significant improvement in updating the weights in the convolutional layers, therefore, the weights in layers conv1-conv5 were frozen for all experiments. This significantly increased the speed of the training algorithm. The confusion loss was approximated using the Kullback-Leibler divergence between the softmax output of a classifier and a uniform distribution. The network is trained using stochastic gradient descent (SGD) with a learning rate of $1\times10^{-4}$. Learning rates of classification layers for each task were boosted by a factor of 10. The hyperparameter in equation \ref{eq:joint} is set to a value of $\alpha=0.1$. The spurious variation classifier hyperparameters were all set to $\beta_m=1$. To address the imbalance in the class distributions of the training data, losses were weighted by the inverse of the relative frequency of that class. We conduct our experiments using the MatConvNet framework \cite{vedaldi15matconvnet}. \section{\label{sec:results}Results} \subsection{Removal of a bias from a dataset} We computed feature representations for the class-balanced test dataset from section~\ref{subsec:imdb} for both baseline and blind networks. T-SNE visualizations of these feature embeddings are shown in figure \ref{fig:tsne}. The feature representation of the baseline network that was trained to classify age is clearly separable by gender, demonstrating that the bias in the training data was learned. After unlearning gender, the feature representation is no longer separable by gender, demonstrating that this bias has been removed. \begin{figure}[h] \centering \begin{subfigure}[b]{0.4\textwidth} \includegraphics[width=\textwidth]{images/bias_tsne_clear_updated.png} \caption{Feature representation of baseline network} \label{fig:agebase} \end{subfigure} ~\qquad \begin{subfigure}[b]{0.4\textwidth} \includegraphics[width=\textwidth]{images/blind_tsne_clear_updated.png} \caption{Feature representation of blind network} \label{fig:ageunl} \end{subfigure} \caption{T-SNE visualizations of 4096-dimensional feature representation of the class-balanced test dataset for age classification networks trained on gender-biased data. (a) Baseline network without unlearning gender, (b) Blind network with gender unlearning. The feature representation is clearly separable by gender for the baseline network, showing that the network has learned the gender bias in the dataset. After unlearning, this bias is no longer pronounced.} \label{fig:tsne} \end{figure} The distributions of the age predictions and ground truth on the gender-balanced test set are shown in figure \ref{fig:singlesov}. The red and blue lines show the predicted ages for women and men, respectively. The black dashed line shows the ground truth for both genders. The distributions of predicted ages for females and males are different for the baseline network (Figure \ref{fig:agebase}): women's ages tend to be underestimated, whereas men's ages are generally overestimated. This behaviour mimics the bias within the training data, which confirms that the classification model is making age predictions that are, to a degree, dependent on gender. In order for us to be sure that the network is not using gender information to predict age, both female and male prediction distributions should be similar. Figure \ref{fig:ageunl} shows the distributions for the network that has been trained to no longer be able to differentiate between genders. The KL-divergence between age prediction distributions for men and women of the biased and unbiased networks are 0.049 and 0.027, respectively. The reduction in KL-divergence shows that the network has successfully unlearned gender, as the predicted age distributions for both genders are similar. Note, that we are not necessarily trying to perfectly predict the ground truth, but be confident that we are not treating men and women differently. The average prediction accuracy for age classification on the gender-balanced test dataset for the baseline network was 78.9\% (78.4\% for females and 79.4\% for males) and for the unbiased network was 78.1\% (77.4\% for females and 78.9\% for males). This corresponds to a reduction in accuracy of 0.8\% (1.2\% for females and 0.5\% for males). \begin{figure}[h] \centering \begin{subfigure}[b]{0.45\textwidth} \includegraphics[width=\textwidth]{images/distribution_before_line.png} \caption{Baseline network} \label{fig:agebase} \end{subfigure} ~ \begin{subfigure}[b]{0.45\textwidth} \includegraphics[width=\textwidth]{images/distribution_after_line.png} \caption{Unbiased network} \label{fig:ageunl} \end{subfigure} \caption{Prediction distributions of age for each gender compared to the ground truth (black) for men (blue) and women (red). The evaluation dataset was gender-balanced so that each class has an equal number of women and men. (a) Baseline network without unlearning gender, (b) unbiased network with gender unlearning. The red line shows the age predictions for women and the blue line shows them for men. The prediction distributions for women and men align after unlearning gender. The KL-divergence score between the prediction distributions of men and women are given for each case. The KL-Divergence score reduces, as two distributions become more similar.} \label{fig:singlesov} \end{figure} \subsection{Removal of an extreme bias from a network} The baseline gender classification accuracy on the gender-balanced test data was 70\% for a network trained on the EB1 dataset. This was improved by 16\% to 86\% for the blind network, where we simultaneously unlearned age information. The accuracies for networks trained on the EB2 dataset were 62\% for the baseline and 82\% for the blind network. This amounts to a 20\% increase in classification accuracy. \begin{figure}[h] \centering \begin{subfigure}[b]{0.9\textwidth} \includegraphics[width=\textwidth]{images/gender_poison_buff_balanced_dist.png} \caption{Male and female predictions by age by network trained on EB1 dataset} \label{fig:extremebiasresult1} \end{subfigure} \par\bigskip \begin{subfigure}[b]{0.9\textwidth} \includegraphics[width=\textwidth]{images/gender_poison_buff_inv_balanced_dist.png} \caption{Male and female predictions by age by network trained on EB2 dataset} \label{fig:extremebiasresult2} \end{subfigure} \caption{Age distributions of male and female predictions from networks trained on the biased (a) EB1 and (b) EB2 datasets, evaluated on the gender-balanced test data. The baseline network often wrongly predicts older individuals to be male and younger individuals to be female. Using the JLU algorithm reduces this bias and shifts the prediction distributions closer to the true distribution.} \label{fig:extremebiasresult} \end{figure} The age distributions of individuals that were predicted to be either male or female, trained on the EB1 dataset, are shown in figure \ref{fig:extremebiasresult}. The baseline network often wrongly predicts older individuals to be male, and younger individuals to be female, in line with the training data. The age distribution of gender predictions from the JLU network, however, is closer to the true distribution. We observed similar results for the baseline network trained on the EB2 dataset, where younger individuals were more often predicted to be male, and older individuals were more often predicted to be female. The JLU network prediction distributions were closer to the true distributions. \subsection{Simultaneous removal of multiple spurious variations} Figure \ref{fig:multi} shows how the accuracies on the test data for each attribute vary over the JLU training procedure when simultaneoulsy unlearning multiple spurious variations. For clarity, in the figure, spurious variation classification accuracies were rescaled using the equation below: \begin{equation} a = 1 - \frac{e}{e_{\text{max}}} , \label{eq:rescale} \end{equation} where $e$ is the mean-class error of the classifier and $e_{\text{max}}$ corresponds to the error-rate of a classifier that draws at random. A perfect classifier corresponds to a score of $a = 1$ and a random classifier corresponds to a score of $a = 0$. The primary classification accuracy was not rescaled. These results are summarised in table \ref{table:multi} for networks trained with and without JLU. The baseline column corresponds to the accuracy of the best classifier on a feature embedding trained without JLU. The ``blind'' column shows the same accuracies when using the JLU algorithm. When the classifier cannot recover meaningful information from the feature embedding the accuracy is equivalent to random chance. \begin{table}[h] \centering \caption{Classification accuracies for each spurious variation for a baseline network and a ``blind network''. The baseline accuracies show the best classifiers that can be learned on the feature representation of the primary task without the JLU algorithm. The ``Blind'' column shows the classification accuracies after JLU. The random chance column states the target accuracy for a spurious variation classifier after JLU, corresponding to an uninformative classifier's accuracy. The \% unlearned column shows the relative difference of baseline and blind accuracies to random chance.} \includegraphics[width=\textwidth]{images/table.png} \label{table:multi} \end{table} \textbf{Age ---} The primary classification accuracy of the blind network is 5\% less than the baseline network. Information about ancestral origin was removed from the network completely. Gender and pose information were removed by 88\% and 91\%, respectively. \textbf{Ancestral Origin ---} The primary classification accuracy of the blind network improved by 1\% compared to the baseline network. The proportions of age, gender, and pose information that were unlearned were 92\%, 98\%, and 86\%, respectively. \textbf{Gender ---} The primary classification accuracy of the blind network is the same as the baseline network. Age, ancestral origin, and pose information were removed by 94\%, 98\% and 95\%, respectively. \textbf{Pose ---} The primary classification accuracy of the blind network is the same as the baseline network. Information about age, ancestral origin, and gender were removed from the network by 93\%, 78\% and 86\%, respectively. Apart from the network that was trained on the primary task of age classification on data that contained a strong gender bias, all blind networks have the same or better primary classification accuracy compared to their baseline counterparts. Spurious variations were successfully removed in most cases, with classification accuracies reducing to within 5\% of random chance in 9/12 cases. The pose experiment proved the least effective, with ancestral origin information proving most difficult to remove. \\ \begin{figure}[h] \centering \begin{subfigure}[b]{0.47\textwidth} \includegraphics[width=\textwidth]{images/age_alt_dual.png} \caption{\textbf{Age} --- Ancestral Origin, Gender, Pose} \label{fig:multi_age} \end{subfigure} ~ \begin{subfigure}[b]{0.47\textwidth} \includegraphics[width=\textwidth]{images/ancestry_dual.png} \caption{\textbf{Ancestral Origin} --- Age, Gender, Pose} \label{fig:multi_anc} \end{subfigure} \par \begin{subfigure}[b]{0.47\textwidth} \includegraphics[width=\textwidth]{images/gender_cb_dual.png} \caption{\textbf{Gender} --- Age, Ancestral Origin, Pose} \label{fig:multi_gender} \end{subfigure} ~ \begin{subfigure}[b]{0.47\textwidth} \includegraphics[width=\textwidth]{images/pose_dual.png} \caption{\textbf{Pose} --- Age, Ancestral Origin, Gender} \label{fig:multi_pose} \end{subfigure} \caption{Classification accuracies on test data for the primary task and spurious variations during the training procedure of the JLU algorithm. The primary classification accuracy is compared to the baseline network accuracy (dotted). The spurious variation accuracies have been re-scaled using equation \ref{eq:rescale} so that zero accuracy corresponds to the accuracy of a random chance classifier.} \label{fig:multi} \end{figure} \section{\label{sec:conclusion}Conclusion} The paper proposes an approach for removing multiple known dataset biases or spurious variations from a deep neural network. Similar to previous work in the field, the algorithm is inspired by domain adaptation work. Drawing inspiration from \cite{tzeng_simultaneous_2015}, we compare the cross-entropy between the output distribution of classifiers that are trained to predict spurious variations and a uniform distribution. The resulting feature representation is informative for the primary task but blind to one or more spurious variations. We demonstrated our algorithm's efficacy on face classification tasks of age, ancestral origin, pose and gender. The resulting feature representations remained informative for one task, whilst simultaneously being uninformative for the others. When training a gender classification network on extremely age-biased data, our algorithm significantly improves (by up to 20\%) classification accuracies on an unbiased test dataset. This demonstrates that our algorithm allows networks trained on biased data to generalize better to unbiased settings, by removing each known bias from the feature representation of the network. This is a significant step towards trusting that a network definitely isn't basing its decisions on the wrong reasons. With increasing use of neural networks in government, law, and employment to make life-changing decisions, it is of great importance, that undesirable social biases are not encoded in the decision algorithm. We also created a dataset to detect ancestral origin from faces, which can be used to remove racial biases from the feature representation of a network. We will make this dataset available to the public. Some spurious variations are easier to remove than others --- applying different weights to each spurious variation classifier could account for these differences. A dynamic-weighting scheme, as in \cite{yin_multi-task_2018}, to automatically weight these different tasks at different time-points during training may improve convergence.

b766441b72890b1da99d51606f9465a191a00022

\section{Introduction} One of the purposes of any traffic situation analysis system, including driver assistance and fully operational autopilot, is avoidance of collisions with both static and moving objects. Currently the most common means for solving this task are radars, lidars and various stereo vision systems (including binocular stereo cameras, time-of-flight cameras, and structured light scanners). \cite{Mukhtar, Franke, Shepelev} Due to relatively low price and high spatiotemporal resolution binocular cameras are especially interesting as being limited primarily by the difficulty of 3D reconstruction which is a rapidly advancing field of study. They also are passive sensors, which is an advantage in under crowded conditions. Many new binocular stereo vision algorithms are published each year (e.g. \cite{Pang}, \cite{Saleem}) and it is important to evaluate quality of both collision avoidance algorithm as a whole and the underlying stereo disparity reconstruction in particular. Having sufficiently precise 3D world model the task of detecting and avoiding obstacles can be straightforwardly formalized thus allowing exact, rather than fuzzy heuristic verification. Binocular reconstruction algorithms are usually tested by the Middlebury dataset and toolset. \cite{Scharstein}. However, since it is impossible to create a general stereo matching algorithm for all practical purposes, problem-oriented criteria are gaining more attention \cite{Sidorchuk}. In particular, paper \cite{Geiger} describes the popular KITTI dataset for testing road situation analysis algorithms and demonstrates that performance ranking using KITTI is drastically different compared to Middlebury. Hence, for developing of obstacle detection systems one needs datasets and metrics specifically designed for road situation analysis. For instance, in the traffic sign recognition problem there are datasets with corresponding problem-oriented evaluation methods already proposed \cite{shakhuro}. However, using correct datasets is not sufficient for adequate applicability evaluation of 3D reconstruction algorithms. One also needs to analyze the impact of 3D reconstruction on the final outcomes, i.e. to perform a higher-level analysis. The authors of \cite{Schneider} suggest a three-level hierarchy of quality metrics for stereo vision algorithm evaluation: \begin{itemize} \item Low level -- correctness of point cloud and depth map -- evaluates the precision of the reconstruction algorithm itself but gives little information about the quality of the final problem and demands time-consuming preparation of highly detailed reference 3D data. \item Medium level -- the so called ``stixel'' representation -- the 3D world is presented by a simplified geometrical model allowing to work with reconstruction results on a level with less details yet relevant enough for the task being solved, thus simplifying the preparation of reference data. \item High level -- the final answer of the system under test. In \cite{Schneider} this was the position of the nearest vehicle, however the metric itself was mentioned rather superficially. \end{itemize} Apart from inability to produce the overall system quality, middle level metrics also have a costly procedure of obtaining test data (ground truth). The paper \cite{Schneider} studies the method of obtaining the correct answers in the form of stixel representation from radars or lidars followed by automatic conversion from depth maps or point clouds into stixel representation. However such conversion demands manual stixel editing to correct the ground truth still yielding high labor cost of the method. High labor cost is not an issue when using cameras with typical position relative to the vehicle and when target observation conditions are present in the existing open datasets, for example KITTI \cite{Geiger} and Cityscapes \cite{Marius}. However when using more exotic configurations like the one described in this paper (self-driving bus --- cameras are mounted high and tilted downwards), one needs to prepare a new dataset. In such cases there is an issue of minimizing the cost of its preparation while maintaining satisfactory detalization relatively to the problem being solved by the system. This paper considers obstacles quality evaluation as an estimation of the final answer the obstacle avoidance system generates, e.g. was decision to stop a vehicle generated correctly or not. Ground truth for such final answers are generated from low-detailed obstacles descriptions and do not lead to high cost while collecting a dataset. This scheme corresponds to medium and high abstraction levels in the \cite{Schneider} hierarchy and can be used for testing both the system as a whole (with a fixed decision taking module) and as a problem oriented comparison metric for stereo vision algorithms. The method can easily be generalized for systems which use other types of sensors providing an optical and depth image. Quality estimation of obstacle tracking is beyond the scope op this work. To introduce the problem context we provide an example of the obstacle detection system which was the initial target of the suggested quality evaluation approach. We study the approach to quality evaluation itself and present an example of using the described method for both optimizing the parameters of 3D reconstruction algorithm and analyzing the system behavior as a whole. \section{Obstacle detection system} Let us describe the architecture of the obstacle detection system (see fig. \ref{ODScheme}). It uses the disparity or depth map, RGB image (optionally), sensor calibration parameters and the information about driving corridor provided by the trajectory planning subsystem as its input. It is assumed that the radial distortion of input RGB images is already compensated. The correction might be performed with the use of calibration images or via blind radial distorsion compensation \cite{kunina}. This modular structure allows several options for the depth sensor: visible light stereo systems, time-of-flight cameras, structured light systems, etc. \begin{figure}[h] \begin{center} \includegraphics[width=1.0\linewidth]{images/scheme.jpeg} \end{center} \caption{\label{ODScheme}Obstacle detection system architecture.} \end{figure} In this work we obtained images from binocular camera with known calibration parameters mounted on a vehicle looking in direction of the motion positioned about 2.2 meters above the ground level. The depth map is computed using the stereo matching algorithm with subsequent filtering of the obtained disparity map. Implementation details are given below in section 4. The system output is a \textit{binary answer indicating obstacle presence or absence}. Thus the suggested scheme can be used both as a framework for prototyping an obstacle detection system and as a high level abstraction of a more sophisticated system that models the decision making procedure. An important feature of such system is the possibility to consider the stereo matching algorithm errors with different weights depending on object distance. Indeed, in our task the insensitivity to distant objects has no importance, while a failure to detect a closely located object may be fatal and the proximity measure is defined by the target properties of the autonomous system in development. \section{Quality Evaluation} Testing of the proposed obstacle detection system implies \textit{estimation of vehicle stops correctness} (see fig. \ref{QEScheme}). \begin{figure}[!h] \begin{center} \includegraphics[width=1.0\linewidth]{images/test_bench.jpeg} \end{center} \caption{\label{QEScheme}Structural scheme of obstacle detection quality evaluation with respect to stops.} \end{figure} Ground truth annotations should be provided as 3D-obstacles coordinates (x, y marked as bounding rectangle, z estimated from calibration) to avoid boundary effects. Any three-dimensional object can be represented by a bounding box or a set of stixels \cite{Pfeiffer}. However, generation of such markup is labor-intensive and can be simplified without loss of quality for the target task as will be shown later. We suggest representing the image markup and the result of obstacle detection by rectangles in a vertical plane. It is sufficient then to prepare the scene markup only in a problem-specific visibility zone, since activation of obstacle detector are relevant primarily for the vehicle driving corridor. In our work the object markup (further referred as \textit{marked obstacle}) is performed for single images from the stereo camera, obtained by virtual camera transformation to a viewpoint perpendicular to the road surface by means of a rectangle with vertical and horizontal sides. All large objects-obstacles including people, vehicles, walls, fences are marked in this manner. In fig. \ref{Matching}a and \ref{Matching}b the marked obstacles are represented by rectangles. Additional \textit{zones of indifference} can be marked (violet polygons in fig.\ref{Matching}a) for the case when we do not expect a confident obstacle detection, in presence of a mesh fence or bushes. Zones of indifference are usually needed in working with a wide driving corridor. Knowing intrinsic and extrinsic camera parameters one can compute 3D coordinates of the source 3D space box of the marked rectangle thus computing distance to it. \begin{figure}[!h] \begin{minipage}[h]{0.49\linewidth} \center{\includegraphics[width=0.95\linewidth]{images/fp_ex.jpeg} \\ a)} \end{minipage} \hfill \begin{minipage}[h]{0.49\linewidth} \center{\includegraphics[width=0.95\linewidth]{images/tp_ex.jpeg} \\ b)} \end{minipage} \caption{\label{Matching} The results of detected obstacles matching (marked red and orange inside and outside the driving corridor resp.) with the marked ones (dark and light blue inside and outside the driving corridor resp.): false-positive stop (a), true-positive stop (b). The driving corridor is shown green, zones of indifference are violet.} \end{figure} A similar representation can be obtained for the obstacle detector answer: for all outstanding objects a planar quadrangle is constructed which is tangent to the object front part and parallel to the image plane (further referred as \textit{detected obstacles}). In the proposed scheme shown in fig. \ref{QEScheme} the obstacles marked on a frame are proposed to be matched against the detected obstacles. The matched answers are counted only relatively to the obstacles falling inside the driving corridor. Each marked obstacle is considered as matched with detected obstacles if it satisfied two conditions. The first condition restricts distances to detected and marked obstacles by the following relation: $$\frac{|z_{ref} - z_{exp}|}{z_{ref}} < \mathrm{T},$$ where $z_{ref}$ is the distance to the marked obstacle, $z_{exp}$ is the distance to the detected obstacle, \mbox{$\mathrm{T}$ is the threshold} value (typically 0.25). The second condition is that the intersection of marked and detection obstacles rectangles should be non-empty. If detected obstacle cannot be matched with any marked one, including zones of indifference, then it is considered as a false positive obstacle. If marked obstacle cannot be matched with any of detected obstacle, it considered to be false negative, otherwise it is considered as a true positive obstacle. The correctness of vehicle stop is estimated by the following rule: \begin{itemize} \item the stop is considered true positive if the driving corridor contains at least one true positive obstacle (see fig. \ref{Matching}b), \item the stop is considered false positive, if the driving corridor contains only false positive obstacle (see fig. \ref{Matching}a), \item the stop is considered false negative, if the driving corridor contains only false negative obstacle, \item otherwise the stop is considered true positive. \end{itemize} \section{Experiment setup} Data acquisition was performed using a binocular camera mounted on a vehicle at 2.2 m height tilted $20^{\circ}$ downwards relatively to the horizon. The camera view angle was \mbox{about $80^{\circ}$}. The output image size was $1280 \times 1024$ pixels. The stereo base width was 75 cm. To demonstrate the applicability of the proposed quality evaluation method we implemented an obstacle detection algorithm which is schematically shown in fig. \ref{OD}. \begin{figure}[h] \begin{center} \includegraphics[width=1.0\linewidth]{images/scheme4.jpeg} \end{center} \caption{\label{OD} Stage based demonstration of obstacle detector operation. Dashed blocks are directly related to the obstacle detector.} \end{figure} The disparity map was constructed using the StereoBM local stereo matching algorithm from the OpenCV library \cite{Faugeras} with subsequent conversion to the the depth map. The depth map is fed as input to the obstacle detector filtered by spatial coordinates to cut out the road plane (presuming that the road plane cannot be tilted by more than $10^{\circ}$ relatively to the current vehicle position). The road plane cutoff height is an adjustable parameter of the obstacle detector. The 3D scene projection on the road plane is then constructed as a single channel image which is also further filtered (here we used morphological closure) and then on the obtained binary image connected components are extracted and the corresponding 3D coordinates are retrieved. The connected components obtained in this manner can be filtered by their area. The filter parameters are adjustable and demand optimization on the marked dataset to increase the target overall quality metric. \section{Evaluation results} The sequences we used for evaluation were recorded with the 3 Hz frequency and were consisted of 5-20 frames. In total the dataset contains 1794 frames, 65\% of which were recorded in simple conditions (day light, no precipitation) and 35\% in difficult conditions (rainy weather, night lighting, rooms with light reflecting floor), 840 frames contain marked obstacles in the driving corridor. We checked the detections of the obstacle detection system in the driving corridor of 2.5 m width and 7 m length. To test the suggested quality assessment method we optimized the system by the stereo algorithm and/or the obstacle detection algorithm parameters. So we produced a grid in a parameter space and for each point of this grid we calculated true positive stops rate (TPR) and false positive stops rate (FPR). An example of TPR vs FPR plot obtained in a such way is shown at the fig. \ref{Optim}. From that plot we choose optimal values of parameters that satisfied projects technical goal. \begin{figure}[!h] \includegraphics[width=0.5\linewidth]{images/optim.jpeg}\hspace{2pc} \begin{minipage}[b]{14pc}\caption{\label{Optim} Example of parameters optimization process. Every single point on a TPR-FPR plot corresponds to a single point in parameter space. With red ellipse the optimal parameters according to project technical task are marked.}\end{minipage} \end{figure} The results of the optimization process that described above are presented in table \ref{Table} and fig. \ref{Res}. As shown in table \ref{Table}, in this task the obstacle detector and stereo algorithm parameters optimization can produce a higher overall system quality compared to adjusting only the stereo matching algorithm. The optimization of obstacle detector parameters also produces comparable quality to stereo algorithm optimization but it is easier to compute since it has fewer parameters. \begin{table}[!b] \caption{\label{Table} Obstacle detection quality with different parameter values.} \begin{center} \begin{tabular}{lll} \br &True-positive & False-Positive \\& stops ratio & stops ratio\\ \mr Initial settings provided by an expert & 0.677 & 0.000 \\ Optimization of obstacle detector & 0.801 & 0.003 \\ Optimization of stereo algorithm & 0.819 & 0.013 \\ Optimization of both & 0.822 & 0.011 \\ \br \end{tabular} \end{center} \end{table} \begin{figure}[!] \begin{minipage}[h]{0.49\linewidth} \flushright{\includegraphics[width=0.65\linewidth]{images/res_a1.jpeg}} \end{minipage} \hfill \begin{minipage}[h]{0.49\linewidth} \flushleft{\includegraphics[width=0.65\linewidth]{images/res_a1_opt.jpeg}} \end{minipage} \vfill \begin{minipage}[h]{0.49\linewidth} \flushright{\includegraphics[width=0.65\linewidth]{images/res_a2.jpeg}} \end{minipage} \hfill \begin{minipage}[h]{0.49\linewidth} \flushleft{\includegraphics[width=0.65\linewidth]{images/res_a2_opt.jpeg}} \end{minipage} \vfill \begin{minipage}[h]{0.49\linewidth} \flushright{\includegraphics[width=0.65\linewidth]{images/res_a3.jpeg}} \end{minipage} \hfill \begin{minipage}[h]{0.49\linewidth} \flushleft{\includegraphics[width=0.65\linewidth]{images/res_a3_opt.jpeg}} \end{minipage} \center{a)}\vfill \vspace{0.2cm} \vfill \begin{minipage}[h]{0.49\linewidth} \flushright{\includegraphics[width=0.65\linewidth]{images/res_b1.jpeg}} \end{minipage} \hfill \begin{minipage}[h]{0.49\linewidth} \flushleft{\includegraphics[width=0.65\linewidth]{images/res_b1_opt.jpeg}} \end{minipage} \vfill \begin{minipage}[h]{0.49\linewidth} \flushright{\includegraphics[width=0.65\linewidth]{images/res_b2.jpeg}} \end{minipage} \hfill \begin{minipage}[h]{0.49\linewidth} \flushleft{\includegraphics[width=0.65\linewidth]{images/res_b2_opt.jpeg}} \end{minipage} \vfill \begin{minipage}[h]{0.49\linewidth} \flushright{\includegraphics[width=0.65\linewidth]{images/res_b3.jpeg}} \end{minipage} \hfill \begin{minipage}[h]{0.49\linewidth} \flushleft{\includegraphics[width=0.65\linewidth]{images/res_b3_opt.jpeg}} \end{minipage} \center{b)} \caption{\label{Res} Visualization of obstacle detection results in simple (a) and difficult (b) conditions in the example of a three-frame sequence. The results of detection with expert-adjusted parameters are on the left, the results after optimizing the obstacle detector and stereo algorithm parameters are on the right. } \end{figure} As shown in fig. \ref{Res}, proposed method for quality evaluation allows to tune parameters to be able to drive in a given driving corridor even in complex conditions with multiple reflections from the ground. \section{Conclusion} In this paper we study development of a stereo vision based obstacle detectors suitable for traffic situation analysis. The primary novelty of the described approach is the use of a quality metrics based on a high-level metric working with the final system outcome, i.e. required stops. We presumed this outcome to be binary but the proposed scheme can be expanded for the case of multiple answers, for example, when the obstacle detection subsystem is provided with decisions about avoiding an obstacle. The proposed metric can serve both as a robustness characteristic for the obstacle detector as a whole and as a problem-oriented quality evaluation for stereo vision algorithms in various problems of road situation analysis. The ground truth preparation in the dataset was reduced to object bounding boxes and indifference zone marking, thus having low labor cost and not demanding specific hardware. To demonstrate the proposed method applicability we have shown a specific implementation of the obstacle detection algorithm which even though simple demonstrates pretty good applicability even in complex conditions. Moreover, the suggested system can employ the obstacle detection algorithm based on stixel representation allowing to take into account road defects and to monitor the change in obstacle positions. \setcounter{secnumdepth}{0} \section{Acknowledgments} This research was supported by Russian Science Foundation grant \mbox{(No.~14-50-00150)}. \setcounter{secnumdepth}{1} \section*{References}

bb9a8e056f3b2195450aaaeadac471edef610909

\section{Introduction} \subsection{Context} A diffeomorphism $f:M\to M$ in a closed manifold $M$ is said to be \emph{partially hyperbolic} if the tangent bundle $TM$ decomposes as a direct sum of continuous and $Df$-invariant subbundles $$TM=E^s\oplus E^c \oplus E^u$$ such that vectors in $E^s$ are uniformly contracted by $Df$, vectors in $E^u$ are uniformly contracted by $Df^{-1}$ and vectors in $E^c$ have an intermediate behavior. The bundles $E^s$ and $E^u$ uniquely integrate to $f$-invariant foliations $\mathcal{W}^u$ and $\mathcal{W}^s$, respectively (see e.g. \cite{HPS}). The bundles $E^s\oplus E^c$ and $E^c\oplus E^u$ may or may not integrate to foliations $\mathcal{W}^{cs}$ and $\mathcal{W}^{cu}$. If they do integrate to $f$-invariant foliations the diffeomorphism is said to be \emph{dynamically coherent}. If $f$ is dynamically coherent, the bundle $E^c$ also integrates to an invariant foliation $\mathcal{W}^c$ whose leaves are the connected components of the intersections of leaves of $\mathcal{W}^{cs}$ and $\mathcal{W}^{cu}$. This work fits in the context of studying partially hyperbolic diffeomorphisms where $\mathcal{W}^c$ is a compact foliation (namely, all leaves of $\mathcal{W}^c$ are compact). Since Sullivan presented his example in \cite{S} of a foliation by cicles with unbounded lenght of leaves, compact foliations have been divided in whether the volume of leaves is uniformly bounded or not (see section \ref{compactfoliations} for a more detailed discussion). In particular, in the uniformly bounded case the leaf space is Hausdorff and has a nice orbifold structure, while in the non uniformly bounded scenario the leaf space is not Hausdorff and may have a complicate structure (see e.g \cite{Eps} and \cite{Vog}). Pugh posed the following questions (see ~\cite{RHRHU}, ~\cite{Gog}). \begin{q} \textit{Let $f:M\to M$ be a partially hyperbolic diffeomorphism with a compact center foliation $\mathcal{W}^c$. Is it true that the volume of center leaves is uniformly bounded? Is it true that $f$ can be finitely covered by a partially hyperbolic diffeomorphism $\tilde f \colon \widetilde M \to \widetilde M$ so that there is a fibration $p \colon \widetilde M \to N$ whose fibers are the center leaves and an Anosov diffeomorphism $\bar f \colon N\to N$ such that $p$ is a semiconjugacy between $\tilde f$ and $\bar f$?} \end{q} Progress on this questions has been done by Bonatti-Wilkinson \cite{BoW}, Bohnet \cite{Boh}, Bohnet-Bonatti \cite{BoBo}, Carrasco \cite{Car} and Gogolev \cite{Gog}. \subsection{Main result} The main theorem of this work is the following: \begin{teo}\label{main} Let $f:M\to M$ be a dynamically coherent partially hyperbolic diffeomorphism with compact center foliation $\mathcal{W}^c$. If $\dim(E^u)=1$ then the volume of the center leaves is uniformly bounded. \end{teo} In \cite{Boh} Bohnet has studied the case where the volume of the center leaves is uniformly bounded and $\dim(E^u)=1$. Combining her results with our main theorem yields: \begin{cor}\label{cor1} Let $f:M\to M$ be a dynamically coherent partially hyperbolic diffeomorphism with compact center foliation $\mathcal{W}^c$. If $\dim(E^u)=1$ then, modulo taking a double cover, the leaf space $M/\mathcal{W}^c$ is a torus $\mathbb{T}^d$ and the dynamics $F:M/\mathcal{W}^c\to M/\mathcal{W}^c$ induced by $f$ is topologically conjugate to an Anosov automorphism on $\mathbb{T}^d$, where $d=codim(\mathcal{W}^c)$. \end{cor} Observe that Theorem \ref{main} and Corollary \ref{cor1} are valid as well if $\dim(E^s)=1$ by working with $f^{-1}$ instead of $f$. In \cite{Gog} Gogolev proved that compact center foliations are uniformly compact under the assumptions $\dim(E^c)=1$, $\dim(E^s)\leq 2$ and $\dim(E^u)\leq 2$. Combining this result with our main theorem then yields: \begin{cor} Let $f:M\to M$ be a dynamically coherent partially hyperbolic diffeomorphism with compact center foliation. If $\dim(M)\leq 5$ then the volume of the center leaves is uniformly bounded. \end{cor} In particular, with the second corollary we can completely discard to be a center foliation the Sullivan's example ~\cite{S} of a foliation by circles in a 5-manifold with unbounded length of leaves. Note that for this specific example the result is, as far as we are aware, new. In Remark \ref{NoSullivan} we give a direct proof for this example without the need of the results from section \ref{four}. \subsection{Organization of the paper and structure of the proof} In section \ref{preliminaries} we give some preliminaries from partially hyperbolic dynamics and foliation theory. In particular, we address the topic of compact foliations and review some definitions and results that will be useful in the proof of the main theorem. In sections \ref{Section3} and \ref{four} we give the proof of the main theorem, and the structure is as follows. The proof will be by contradiction. Assume that the center foliation $\mathcal{W}^c$ has not uniformly bounded volume of leaves. This is equivalent with the \emph{bad set} $$\mathcal{B}=\{x:\text{ center leaf volume function is not locally bounded at } x\}$$ being not empty (see section \ref{compactfoliations}). The main result of section \ref{Section3} is that $\mathcal{B}$ is saturated by the center-unstable foliation: as the unstable holonomy of center leaves is trivial (see Lemma \ref{lemma1}), the stable holonomy groups of points in the same unstable leave are conjugated (see Lemma \ref{lemma2}). We deduce that $\mathcal{B}$ is an attactor. In particular, this implies that $f$ cannot be transitive. Section \ref{four} is dedicated to rule out the non transitive case. The attractor $\mathcal{B}$ induces an associated repeller $\mathcal{R}$. We first see that $\mathcal{R}$ is saturated by the center foliation, implying that the center leaf volume function is bounded in $\mathcal{R}$ (see section \ref{Rconstruction}). We then give a sort of topological description of center-stable leaves in $\mathcal{R}$, namely, all of them are bundles over a center leaf with stable manifolds as fibers, thus having trivial transverse holonomy (see section \ref{completeandbundle}). Finally, this allow us to adapt Hiraide arguments \cite{Hir} (see also Bohnet \cite{Boh}) in order to discard the existence of the codimension one transversally unstable repeller $\partial \mathcal{R}$. \subsection{Acknowledgements} The authors are greately thankful with their advisor Rafael Potrie. This work could not be possible without his advice, patience and generosity. The authors would also like to thank Christian Bonatti, Sylvain Crovisier, Sergio Fenley, Andrey Gogolev, Andy Hammerlindl and Martín Sambarino for their generous time and helpful conversations. \section{Preliminaries}\label{preliminaries} \subsection{Preliminaries from partially hyperbolic dynamics} Let $f:M\to M$ be a $C^1$ diffeomorphism on $M$ a closed Riemannian manifold. We say that $f$ is \emph{partially hyperbolic} if there exists a continuous and $Df$-invariant decomposition $$TM=E^s\oplus E^c\oplus E^u,$$ and some $\ell>0$ such that for every $x\in M$ and unit vectors $v^\sigma\in E^\sigma$ for $\sigma\in \{s,c,u\}$, one has that: \begin{center} $\|D_xf^{\ell}(v^s)\|<1$, \hspace{0.2cm} $\|D_xf^{-\ell}(v^u)\|<1$\hspace{0.2cm} and \end{center} \begin{center} $\|D_xf^{\ell}(v^s)\|<\|D_xf^{\ell}(v^c)\|<\|D_xf^{\ell}(v^u)\|.$ \end{center} We call \emph{stable}, \emph{center-stable}, \emph{center}, \emph{center-unstable} and \emph{unstable bundle} to $E^s,E^s\oplus E^c,E^c,E^c\oplus E^u$ and $E^u$, respectively. The stable and unstable bundles are known to be uniquely integrable (see e.g. \cite{HPS}) to foliations $\mathcal{W}^s$ and $\mathcal{W}^u$, respectively. However, the center, center-stable or center-unstable bundles may not integrate. We say that a partially hyperbolic diffeomorphism is \emph{dynamically coherent} if the center-stable and center-unstable bundles integrate to $f$-invariant foliations $\mathcal{W}^{cs}$ and $\mathcal{W}^{cu}$, respectively. In particular, this implies that the center bundle is also integrable: the center leaf through a point $x\in M$ being the connected component of $W^{cs}(x)\cap W^{cu}(x)$ that contains $x$. The resulting foliation, $\mathcal{W}^c$, is then also $f$-invariant and tangent to the center bundle. For more information and context on this topics see e.g. \cite{RHRHU}. Given a point $x\in M$ we denote as $W^\sigma(x)$ the leaf of $\mathcal{W}^\sigma$ through $x$ for $\sigma\in \{s,cs,cu,u\}$. We denote $\mathcal{C}_x$ the leaf of $\mathcal{W}^c$ through $x$. Given a center leaf $\mathcal{C}$, we denote $W^\sigma(\mathcal{C})$ the leaf of $\mathcal{W}^\sigma$ through $\mathcal{C}$ for $\sigma\in\{cs,cu\}$. For $x\in M$ and $r>0$, we denote $\Ball^\sigma_r(x)\subset W^\sigma(x)$ the intrinsic ball of center $x$ and radius $r>0$ in $W^\sigma(x)$ for $\sigma \in \{s,cs,c,cu,u\}$. If $\dim(E^{\sigma})=1$ we will simply denote it $(x-r,x+r)^{\sigma}$. If $\dim(E^{\sigma})=1$ and $\mathcal{W}^{\sigma}$ is oriented, we denote $(x,x+\delta)^u=\{ y\in \Ball^u_{\delta}(x):x<y\}$ and $W^u_+(x)=\{ y\in W^u(x):x<y\}$, and, if $y\in\mathcal{W}^u(x)$, we denote by $[x,y]^u$ and $(x,y)^u$ the oriented closed and open segments in $\mathcal{W} ^u(x)$, respectively, from $x$ to $y$. Leaves of $\mathcal{W}^s$ and $\mathcal{W}^u$ can be obtained as increasing union of balls, thus, are homeomorphic to $\mathbb{R}^{\dim E^s}$ and $\mathbb{R}^{\dim E^u}$, respectively. \subsection{Preliminaries from foliation theory} We will consider continuous foliations with $C^1$-leaves tangent to a continuous distribution. A general reference for this section is \cite{CC}. \subsubsection{Holonomy of a leaf} Let us briefly recall the definition of the holonomy group of a leaf. Consider $W$ a leaf in a foliation $\mathcal{W}$ of codimension $q$. Fix $x_0$ a point in $W$ and let $D$ be a disk of dimension $q$ transversal to $\mathcal{W}$ through $x_0$. For every loop $\gamma:[0,1]\to W$ based on $x_0$ one can consider $h_\gamma :D' \to D$ the \emph{holonomy return map} to $D$ of the leaves of $\mathcal{W}$ through points from a smaller transversal disk $D'\subset D$. Consider adequate small transversal disks $D_{\gamma(t)}$ through each point $\gamma(t)$ and, given $y$ in $D'=D_{\gamma(0)}$, define $h_\gamma(y)$ as the end point of the continuous curve $\gamma_y$ defined by $\gamma_y(t)\in D_{\gamma(t)}\cap W(y)$ and $\gamma_y(0)=y$. For a local homeomorphism $h:D'\subset D\to D$ fixing $x_0$ one defines the \emph{germ of $h$} as the class of all local homeomorphisms that coincides with $h$ in a neighborhood of $x_0$. With the operation given by the composition, this classes form $G(x_0,D)$ the \emph{group of germs of local homeomoprhisms at $x_0$}. One can see that for basepoint fixed homotopic curves in $W$ based at $x_0$ there corresponds the same holonomy germ. Since concatenation of curves corresponds to composition of holonomy maps (whenever well defined) one has a well defined homomorphism $$\phi:\pi_1(W,x_0)\to G(x_0,D)$$ where $\pi_1(W,x_0)$ is the fundamental group of $W$ based in $x_0$. Define the \emph{holonomy group of $W$ at $x_0$} as $\phi(\pi_1(W,x_0))$. The isomorphism class of $\phi(\pi_1(W,x_0))$ does not depend of $x_0$ or $D$, so we call it \emph{holonomy group of $W$} and denote it by $\Hol(W)$. \subsubsection{Reeb stability} This next theorem is a classical in foliation theory. See e.g. \cite[Theorem 2.4.3]{CC} and \cite[Theorem 3.]{CLN}. \begin{teo}[Generalized Reeb stability theorem]\label{Reeb} Let $W$ be a compact leaf in a foliation $\mathcal{W}$ such that $W$ has a finite holonomy group $\Hol(W)$. Then there exists $\mathcal{U}(W)$ a neighbourhood of $W$, saturated by leaves of $\mathcal{W}$, in which all leaves of $\mathcal{W}$ are compact with finite holonomy group. Moreover, $\mathcal{U}(W)$ has an associated projection $\pi:\mathcal{U}(W)\to W$ such that for every $W' \subset \mathcal{U}(W)$ the map $\pi|_{W'}:W'\to W$ is a finite covering with $k$ sheets, $k\leq |\Hol(W)|$, and for each $y\in W$ the set $\pi ^{-1}(\{y\})$ is a disk transversal to $\mathcal{W}$. The neighbourhood $\mathcal{U}(W)$ can be taken to be arbitrarily small. \end{teo} \subsubsection{Compact foliatons}\label{compactfoliations} We say that a foliation $\mathcal{W}$ is \emph{compact} if every leaf $W$ of $\mathcal{W}$ is compact. Given a compact foliation $\mathcal{W}$ and a Riemannian metric in $M$ we can consider the \emph{volume function} $$\mathrm{vol}:M\to [0,+\infty)$$ that assigns to each point $x\in M$ the volume of the leaf $W(x)$ with respect to the metric in $W(x)$ induced by the metric of $M$. It may be the case that a compact foliaton does not have uniformly bounded volume of leaves, meaning that $\mathrm{vol}$ is an unbounded function (see for example \cite{S} or \cite{EV}). Given a compact foliation $\mathcal{W}$, define the \emph{bad set of $\mathcal{W}$} as $$\mathcal{B}:=\{x\in M: \mathrm{vol} \text{ is not locally bounded at } x\}.$$ The set $\mathcal{B}$ does not depend on the choice of the metric in $M$. \begin{obs} The fact that $\mathcal{W}$ has uniformly bounded volume of leaves is equivalent to $\mathcal{B}$ being the empty set. \end{obs} \begin{obs} Observe that if $\mathcal{W}$ is invariant by a $C^1$ diffeomorphism $f$ then $\mathcal{B}$ is invariant by $f$ (namely, $f(\mathcal{B})=\mathcal{B}$). This is going to be used in our context where $\mathcal{W}$ will be the center foliation of a partially hyperbolic diffeomorphism. \end{obs} Let us see some of the properties of $\mathcal{B}$. For a proof of the following result see, for example, Epstein \cite{Eps} or Lessa \cite{Les}: \begin{prop} Let $\mathcal{W}$ be a compact foliation. Then the volume function is lower semicontinuous. That is, $\liminf_{x_n\rightarrow x} \mathrm{vol}(W(x_n)) \geq \mathrm{vol}(W(x))$. \end{prop} Using the past proposition we can prove (see e.g. \cite{EMS}): \begin{prop}\label{PropertiesB} The bad set $\mathcal{B}$ is closed with empty interior and is saturated by leaves of $\mathcal{W}$. \end{prop} \begin{proof} The set $\mathcal{B}$ is clearly closed. Semicontinuous functions are continuous in a residual set and $\mathcal{B}\subset\{x\in M: \mathrm{vol} \text{ is not continuous in }x\}$. Thus, from $\mathrm{vol}$ being lower semicontinuous we deduce that $\mathcal{B}$ has empty interior. The fact that the volume function is constant along leaves implies that $\mathcal{B}$ is saturated by leaves of $\mathcal{W}$: if $x\in \mathcal{B}$ then $x$ has arbitrarily large leaves arbitrarily close to it, so if $y\in W(x)$ then this arbitrarily long leaves pass close to $y$ as well (due to the continuity of the foliation $\mathcal{W}$), meaning that $y$ also belongs to $\mathcal{B}$. \end{proof} Leaves in $\mathcal{B}$ can be completely caracterized in terms of holonomy (see e.g. Epstein \cite[Theorem 4.2.]{Eps}): \begin{prop}\label{Holinfty} Let $\mathcal{W}$ be a compact foliation and $W$ a leaf of $\mathcal{W}$. Then $W\subset \mathcal{B}$ if and only if $|\Hol(W)|=\infty$. \end{prop} \subsection{Stable and unstable holonomies and product neighborhoods} In this section we see particular definitions and results we will use later. Throughout this section $f:M\to M$ will be a dynamically coherent partially hyperbolic diffeomorphism with compact center foliation. \subsubsection{Stable and unstable holonomies}\label{holonomies} A consequence of dynamical coherence is that each leaf $W$ of $\mathcal{W}^{cs}$ is foliated by center leaves. We denote the restricted foliation by $\mathcal{W}^c|_W$. Given a center leaf $\mathcal{C}$ denote by $\Hol(\mathcal{C})$ the holonomy group of $\mathcal{C}$ as a leaf of $\mathcal{W}^c$. Denote by $\Hol^s(\mathcal{C})$ the holonomy group of $\mathcal{C}$ as a leaf of $\mathcal{W}^c|_{W^{cs}(\mathcal{C})}$. If $W$ is a leaf of $\mathcal{W}^{cs}$ denote by $\Hol_{cs}(W)$ the holonomy group of $W$ as a leaf of $\mathcal{W}^{cs}$. More generally, if $V\subset W$, we denote $\Hol_{cs}(V)$ the subgroup of $\Hol_{cs}(W)$ that corresponds to holonomy return maps along closed curves inside $V$. We analogously define $\Hol^u$ and $\Hol_{cu}$. \begin{obs}\label{Hol_{cs}=Hol^u} We have that $\Hol_{cs}(\mathcal{C})\simeq \Hol^u(\mathcal{C})$ and $\Hol_{cu}(\mathcal{C})\simeq \Hol^s(\mathcal{C})$ for every center leaf $\mathcal{C}$. \end{obs} \begin{proof} We prove the first equality, the second one is analogous. Let $\gamma$ be a loop in a center leaf $\mathcal{C}$ based in some point $x_0\in \mathcal{C}$. Consider $D^u(x_0)$ a small unstable disk trough $x_0$. One can consider on one hand the holonomy return map $h_\gamma$ associated to $\gamma$ of the leaf $W^{cs}(\mathcal{C})$ in the foliation $\mathcal{W}^{cs}$ and, on the other hand, the holonomy return map $h_\gamma'$ of the leaf $\mathcal{C}$ in the foliation $\mathcal{W}^c|_{W^u(\mathcal{C})}$. Dynamical coherence gives us that $h_\gamma$ coincides with $h_\gamma'$. The thesis follows. \end{proof} \begin{obs} For a set $V$ inside a center-stable leaf $W$ that deformation retracts inside $W$ to some center leaf $\mathcal{C}$, we have that $\Hol_{cs}(V)\simeq\Hol_{cs}(\mathcal{C})\simeq\Hol^u(\mathcal{C})$. For such sets we write $\Hol^u(V)$ instead of $\Hol_{cs}(V)$. Analogously for $\Hol_{cu}(V)$ and $\Hol^s(V)$. \end{obs} \begin{obs} Each one of the previous holonomies is invariant by $f$. \end{obs} Dynamical coherence gives a relationship between the holonomy $\Hol(\mathcal{C})$ and the holonomies $\Hol^s(\mathcal{C})$ and $\Hol^u(\mathcal{C})$. In particular, we are going to need the following (see e.g. Carrasco \cite[Proposition 2.5.]{Car}): \begin{prop}\label{Holprod} For every center leaf $\mathcal{C}$ we have that $$\max\{|\Hol^s(\mathcal{C})|,|\Hol^u(\mathcal{C})|\}\leq |\Hol(\mathcal{C})| \leq |\Hol^s(\mathcal{C})||\Hol^u(\mathcal{C})|.$$ In particular, $\Hol(\mathcal{C})$ is finite if and only if $\Hol^s(\mathcal{C})$ and $\Hol^u(\mathcal{C})$ are finite. \end{prop} \begin{proof} Let $\mathcal{C}$ be a center leaf and $x_0$ a point in $\mathcal{C}$. Consider $D$ a small transversal disk to $\mathcal{W}^c$ and tangent to $E^s\oplus E^u$ in $x_0$. Let us denote $D^s$ and $D^u$ the connected components of $D \cap W^{cs}(x_0)$ and $D \cap W^{cu}(x_0)$ that contains $x_0$, respectively. More general, for every $x$ in $D$ denote $D^s(x)$ and $D^u(x)$ the connected components of $D \cap W^{cs}(x)$ and $D \cap W^{cu}(x)$ that contains $x$, respectively. We have then, in a neighborhood $D'\subset D$ of $x_0$, local stable/unstable coordinates. That is, for every $x$ in $D'$ there exists unique $x_s\in D^s$ and $x_u\in D^u$ such that $x=D^u(x_s)\cap D^s(x_u)$. Denote the coordinates of $x$ by $(x_s,x_u)$. Given a loop $\gamma$ based at $x_0$ we have well defined holonomy return maps, $h_\gamma^s:{D^s}'\to D^s$ and $h_\gamma^u:{D^u}'\to D^u$, for center leaves inside $W^{cs}(x_0)$ and $W^{cu}(x_0)$, respectively. Dynamical coherence says then that the holonomy return map associated to $\gamma$ in $D$ can be written in local cordinates as: $$h_\gamma=(h_\gamma^s,h_\gamma^u),$$ where $h_\gamma$ is given by $h_\gamma(x)=(h_\gamma^{s}(x_s),h_\gamma^{u}(x_u))$, with $x=(x_s,x_u)$ the stable/unstable coordinates defined before. Elements of $\Hol^s(\mathcal{C})$ and $\Hol^u(\mathcal{C})$ are germs of local homeomorphism fixing $x_0$ in $D^s$ and $D^u$, respectively. The previous discussion shows that the map $$\psi:\Hol^s(\mathcal{C})\times \Hol^u(\mathcal{C}) \to \Hol(\mathcal{C})$$ given by $\psi(h_\gamma^s,h_\gamma^u)=h_\gamma$ is surjective. This implies that $|\Hol(\mathcal{C})| \leq |\Hol^s(\mathcal{C})||\Hol^u(\mathcal{C})|$. Moreover, since $h_\gamma(x_s,x_0)=(h^s(x_s),x_0)$ and $h_\gamma(x_0,x_u)=(x_0,h^u(x_u))$, we obtain that $\max\{|\Hol^s(\mathcal{C})|,|\Hol^u(\mathcal{C})|\}\leq |\Hol(\mathcal{C})| $. \end{proof} \subsubsection{Product neighborhoods} Assume throughout the rest of this section that $\dim(E^u)=1$ and $\Hol^u(\mathcal{C})=\Id$ for every center leaf. This hypothesis will be fulfilled later during the proof of the main theorem. The following will be of good use many times (for similar results see e.g. \cite{CC}) or \cite{HH}: \begin{prop}[Product neighborhoods]\label{productnbhd} Let $\mathcal{C}$ be a center leaf and consider $E=\bigcup_{x\in \mathcal{C}}\Ball^s(x)$, where $\Ball^s(x)\subset W^s(x)$ is a disk such that $\Ball^s(x)\cap \mathcal{C} = \{x\}$ for every $x\in \mathcal{C}$ and $\Ball^s(x)$ varies continuously with $x$. Then there exists an homeomorphism over its image $\varphi:E\times [-1,1]\to M$ such that: \begin{enumerate} \item $\varphi(E\times \{0\})=E$. \item $\varphi(E \times \{y\})$ lies inside a $\mathcal{W}^{cs}$ leaf for every $y\in [-1,1]$. \item $\varphi(\{x\}\times [-1,1])$ lies inside a $\mathcal{W}^u$ leaf for every $x\in E$. \end{enumerate} \end{prop} In this case, $\varphi(E\times [-1,1])$ will be called as a \emph{$(\mathcal{W}^{cs},\mathcal{W}^u)$-product neighbourhood} of $E$ and, by an abuse of notation, we will simply refer to it as $E\times [-1,1]$. \begin{proof} We can consider $\epsilon>0$ such that for every distinct $x,x'\in E$ we get $(x-\epsilon,x+\epsilon)^u\cap (x'-\epsilon,x'+\epsilon)^u=\emptyset$. Let us fix $x_0\in \mathcal{C}$. By the continuity of $\mathcal{W}^{cs}$ we can consider $\delta>0$ such that for every curve $\gamma:[0,1]\to E$ with $\gamma(0)=x_0$ and $\length(\gamma)\leq 2\diam(E)$ we have a well defined holonomy map $$h_\gamma:(x_0-\delta,x_0+\delta)^u\to (\gamma(1)-\epsilon,\gamma(1)+\epsilon)^u.$$ Observe that if such a $\gamma$ is closed, since $E=\bigcup_{x\in \mathcal{C}}\Ball^s(x)$ with each $\Ball^s(x)$ a stable disk, then $\gamma$ would be homotopic to a loop in $\mathcal{C}$. Since $\Hol^u(\mathcal{C})=\Id$ we can take $\delta>0$ small enough as to assure that $h_\gamma$ coincides with the identity from $(x_0-\delta,x_0+\delta)^u$ to itself for every $\gamma$ such that $\length(\gamma)\leq 2\diam(E)$. For every $x\in E$ consider $\gamma_x:[0,1]\to E$ a curve from $x_0$ to $x$ such that $\length(\gamma_x)\leq \diam(E)$. Now, for every $y\in (x_0-\delta,x_0+\delta)^u$ define $$\varphi(x,y)=h_{\gamma_x}(y).$$ This definition is independent of the choice of the curve $\gamma_x$. Indeed, if $\gamma_x':[0,1]\to E$ is another such a curve, the concatenation $\gamma_x*\gamma_x'$ has length at most $2\diam(E)$ and then $id=h_{\gamma_x*\gamma_x'}=h_{\gamma_x}\circ h_{\gamma_x'}^{-1}$ implies $h_{\gamma_x}(y)=h_{\gamma_x'}(y)$ for every $y\in (x_0-\delta,x_0+\delta)^u$. The properties of the map $\varphi$ follow directly from its definition \end{proof} We will many times need a particular instance of the previous proposition: \begin{obs}\label{Reebproduct} For every center leaf $\mathcal{C}$ such that $|\Hol(\mathcal{C})|<\infty$ we can consider $\mathcal{U}^s(\mathcal{C})$ a neighbourhood of $\mathcal{C}$ in $W^{cs}(\mathcal{C})$ given by the Generalized Reeb stability theorem (see Theorem \ref{Reeb}) such that: \begin{itemize} \item The associated projection $\pi:\mathcal{U}^s(\mathcal{C})\to \mathcal{C}$ satisfies that $\pi^{-1}(x)$ is a disk in $W^s(x)$ for every $x\in \mathcal{C}$. \item There exists a $(\mathcal{W}^{cs},\mathcal{W}^u)$-product neighbourhood $\mathcal{U}^s(\mathcal{C})\times [-1,1]$ of $\mathcal{U}^s(\mathcal{C})$. \end{itemize} \end{obs} \begin{proof} By the transversality of the foliations $\mathcal{W}^s$ and $\mathcal{W}^c$ inside $W^{cs}(\mathcal{C})$ and the fact that $\mathcal{U}^s(\mathcal{C})$ can be taken arbitrarily small we obtain that the projection $\pi$ can be taken along leaves of $\mathcal{W}^s$. The neighbourhood $\mathcal{U}^s(\mathcal{C})$ fulfills the hypothesis of Proposition \ref{productnbhd} and that implies the existence of the product neighbourhood. \end{proof} \section{The bad set is saturated by the center-unstable foliation}\label{Section3} From now on, let $f:M\to M$ be a dynamically coherent partially hyperbolic diffeomorphism such that $\dim(E^u)=1$. As before, we assume that $\mathcal{W}^c$ is a compact foliation. We are going to see that $\mathcal{W}^c$ is in fact uniformly compact (meaning that the leaf volume function is bounded in $M$). \begin{obs} We can assume from now on that all bundles $E^s,E^{cs},E^c,E^{cu}$ and $E^u$ are orientable. \end{obs} \begin{proof} By taking a finite cover of $M$ we can lift all bundles $E^s,E^{cs},E^c,E^{cu}$ and $E^u$ to orientable bundles. Then $f$ lifts to a dynamically coherent partially hyperbolic diffeomorphism $\tilde{f}$ whose center foliation $\tilde{\mathcal{W}}^c$ is the lift of $\mathcal{W}^c$. The lifted center foliation $\tilde{\mathcal{W}}^c$ remains compact, and each one of its leaves is a finite cover of some leaf of $\mathcal{W}^c$. Then, if $\tilde{\mathcal{W}}^c$ is uniformly compact, so is $\mathcal{W}^c$. \end{proof} Let us see first a simple but yet crucial consequence of the codimension one hypothesis: \begin{lema}\label{lemma1} The group $\Hol^u(\mathcal{C})$ is trivial for every center leaf $\mathcal{C}$. \end{lema} \begin{proof} Let $\mathcal{C}$ be a center leaf. Recall that $\Hol^u(\mathcal{C})$ consists of the holonomy maps associated to $\mathcal{C}$ inside the center-unstable leaf $W^{cu}(\mathcal{C})$. Let $x$ be a point in $\mathcal{C}$ and $\gamma:[0,1]\to \mathcal{C}$ a closed curve based in $x$. Consider $\lambda' \subset\lambda\subset \mathcal{W}^u(x)$ small enough one-dimensional transversals through $x$ such that the holonomy return map associated to $\gamma$ is a well defined map $h_\gamma:\lambda' \to \lambda$. As we are assuming that the unstable foliation is orientable, the map $h_\gamma$ preserves the orientation of $\lambda$. Assume that $h_\gamma$ is not the identity. Then there exists $y\in\lambda'$ such that $\{h_\gamma^n(y)\}_{n\geq 0}$ or $\{h_\gamma^{-n}(y)\}_{n\geq 0}$ constitutes an infinite set of points lying in $\lambda$. Assume without loss of generality it is the former. Fix $U$ a small foliated neighborhood of $\mathcal{W}^c$ containing $\lambda$. Through every point of $\{h_\gamma^n(y)\}_{n\geq 0}$ corresponds a different plaque of $U$. Since each one of this plaques belong to $\mathcal{C}_y$ this contradicts the fact that $\mathcal{C}_y$ is compact. \end{proof} From the previous lemma, we deduce that each center leaf $\mathcal{C}$ has in $W^{cu}(\mathcal{C})$ a product neighbourhood of the form $\mathcal{C}\times (-\delta,\delta)^u$ , where each $\mathcal{C} \times \{y\}$ corresponds to a center leaf and each $\{x\} \times (-\delta, \delta)^u$ corresponds to an unstable arc (see Figure \ref{fig1}). This will allow us to use Remark \ref{Reebproduct}. This kind of ``stacking'' of center leaves along the unstable direction implies that stable holonomy is constant along unstable leaves (see Figure \ref{fig2}): \begin{lema}\label{lemma2} For every $x\in M$ and $y\in W^u(x)$ the groups $\Hol^s(\mathcal{C}_x)$ and $\Hol^s(\mathcal{C}_y)$ are isomorphic. \end{lema} \begin{proof} Obseve that it is enough to give a local argument: suppose that for every $x\in M$ there is an unstable arc $(x-\delta, x+\delta)^u$ such that $\Hol(\mathcal{C}_y)\simeq \Hol(\mathcal{C}_{x})$ for every $y\in (x-\delta, x+\delta)^u$. This implies that the set $\{y\in W^u(x)$ : $\Hol(\mathcal{C}_y)$ is isomorphic to $\Hol(\mathcal{C}_x)\}$ is an open subset of $W^u(x)$ as well as its complement, and thus the thesis follows. Let $x$ be a point in $M$ and denote by $\mathcal{C}$ the center leaf through $x$. Since $\mathcal{C}$ is compact there exists $\delta>0$ such that $\Ball^s_\delta(z)\cap \Ball^s_\delta(z')=\emptyset$ for every distinct $z,z'\in \mathcal{C}$. Thus $\bigcup_{z\in \mathcal{C}}\Ball^s_\delta(z)$ (denote it $\Ball^s_\delta(\mathcal{C})$) is in the hypothesis of the Proposition \ref{productnbhd} and we can consider a $(\mathcal{W}^{cs},\mathcal{W}^u)$-product neighbourhood $\Ball^s_\delta(\mathcal{C})\times (-1,1)$. Given $t\in (-1,1)$ we can define a projection $p_{t}:\Ball^s_\delta(\mathcal{C})\times \{0\}\to \Ball^s_\delta(\mathcal{C})\times \{t\}$ along unstable leaves inside $\Ball^s_\delta(\mathcal{C})\times (-1,1)$. Namely $$p_{t}(z,0)=(z,t).$$ The projection $p_{t}$ then indentifies $\mathcal{W}^c|_{\Ball^s_\delta(\mathcal{C})\times \{0\}}$ homeomorphically with $\mathcal{W}^c|_{\Ball^s_\delta(\mathcal{C})\times \{t\}}$ since the leaves of $\mathcal{W}^c$ are the connected components of the intersection of leaves of $\mathcal{W}^{cs}$ with leaves of $\mathcal{W}^{cu}$. This implies that \begin{center} $\Hol^s(\mathcal{C}_{y})\simeq\Hol^s(\mathcal{C}_{x})$ \end{center} for every $y\in \{ x\}\times (-1,1)$. \end{proof} \begin{figure}[!tbp] \centering \begin{minipage}[b]{.5\textwidth} \centering \input{Figure1.pdf_tex} \caption{}\label{fig1} \label{fig:test1} \end{minipage}% \hfill \begin{minipage}[b]{.5\textwidth} \centering \input{Figure2.pdf_tex} \caption{}\label{fig2} \label{fig:test2} \end{minipage} \end{figure} Recall from the preliminaries that we denote the bad set $\mathcal{B}$ of $\mathcal{W}^c$ as the points of $M$ in which the leaf volume function is not locally bounded. \begin{cor}\label{saturated} The bad set $\mathcal{B}$ of $\mathcal{W}^c$ is saturated by the center-unstable foliation. \end{cor} \begin{proof} From Proposition \ref{Holinfty} we have that a center leaf belongs to $\mathcal{B}$ if and only if $|\Hol(\mathcal{C})|=\infty$. By Proposition \ref{Holprod} and Lemma \ref{lemma1}, we have that $|\Hol(\mathcal{C})|< \infty$ if and only if $|\Hol^s(\mathcal{C})|< \infty$. The previous two lemmata then implies that $\mathcal{B}$ is saturated by the unstable foliation. As it is also saturated by the center foliation (see Remark \ref{PropertiesB}) the thesis follows. \end{proof} We obtain the following. \begin{cor} The bad set $\mathcal{B}$ of $\mathcal{W}^c$ is a proper attractor. In particular, there are no transitive, codimension one, dynamically coherent partially hyperbolic diffeomorphisms with compact center foliation and unbounded volume of leaves. \end{cor} \begin{proof} The set $\mathcal{B}$ is compact, $f$-invariant, has empty interior and is saturated by the center-unstable foliation (see Corollary \ref{saturated}). In particular, $\mathcal{B}$ is transversally stable and not all $M$, and thus a proper attractor. This implies that $f$ cannot be transitive. \end{proof} The aim of the rest of the work is to see that same thesis follows in the non transitive scenario. We finish this section noting that, by what we have done by this point and the work of Gogolev in \cite{Gog}, one is already able to discard the Sullivan foliation \cite{S} as being the center foliation of a dynamically coherent partially hyperbolic system: \begin{obs}\label{NoSullivan} The Sullivan foliation \cite{S} cannot be the center foliation of a dynamically coherent partially hyperbolic diffeomorphism. \end{obs} \begin{proof} The example given by Sullivan is a foliation by circles in a five dimensional compact space $M$ with unbounded length of leaves. Assume that Sullivan's foliation is the center foliation of a dynamically coherent partially hyperbolic diffeomorphism. As the center foliation is one dimensional (in a five dimensional manifold), Gogolev's work implies that it has uniformly bounded volume of leaves if $\dim(E^s)= 2$ and $\dim(E^s)= 2$ (see Main Theorem in \cite{Gog}). It remains to rule out the codimension one case. Assume without loss of generality that $\dim(E^u)=1$. Let us denote the Sullivan foliation by $\mathcal{F}$. In this particular case, the bad set $\mathcal{B}$ of $\mathcal{F}$ has the structure of the unitary tangent bundle $T^1S^2$ of a $2$-sphere. Moreover, the leaves of the foliation in $\mathcal{B}$ are exactly the fibers of this unitary tangent bundle. By the Corollary \ref{saturated} the set $\mathcal{B}$ is saturated by the center-unstable foliation, and so center-unstable leaves foliate $\mathcal{B}$. Given a center leaf $\mathcal{C}$ we have that $\Hol^u(\mathcal{C})=\Id$ and then $\mathcal{C}$ has a neighbourhood $\mathcal{C}\times (-\delta,\delta)^u$ in $W^{cu}(\mathcal{C})$ such that each $\mathcal{C}\times \{y\}$ is a center leaf (see Figure \ref{fig1}). This implies that the center-unstable foliation in $\mathcal{B}\simeq T^1S^2$ projects to a (topological) foliation without singularities in the base $S^2$. This cannot be. \end{proof} \section{Proof of the non transitive case}\label{four} We saw in the previous section that $f$ cannot be transitive if the bad set $\mathcal{B}$ of $\mathcal{W}^c$ is not empty. Lose the transitivity hypothesis and assume that $\mathcal{B}$ is not empty. We are going to see that this yields a contradiction. \subsection{Construction of the repeller $\mathcal{R}$}\label{Rconstruction} Let us consider $\mathcal{R}$ the repelling set induced by the attactor set $\mathcal{B}$ (see Figure \ref{figR}): $$\mathcal{R}=M \setminus \bigcup_{x \in \mathcal{B}}W^s(x).$$ We will next closely study $\mathcal{R}$. In particular, we are going to see that is saturated by center-stable leaves. \vspace{0.2cm} \begin{figure}[htb] \centering \def300pt{200pt} \input{Figure3.pdf_tex} \caption{}\label{figR} \end{figure} \vspace{0.2cm} For every $r>0$, let us denote $\Ball^s_r(\mathcal{B})=\bigcup_{x \in \mathcal{B}} \Ball^s_r(x)$. \begin{lema} The set $\mathcal{R}$ is non-empty, $f-$invariant, compact and saturated by the stable foliation. \end{lema} \begin{proof} Since $\mathcal{B}$ is saturated by leaves of $\mathcal{W}^{cu}$ then $\bigcup_{x \in \mathcal{B}}W^s(x)$ is open. This implies that $\mathcal{R}$ is compact. Let us see that $\mathcal{R}$ is not empty. The set $\bigcup_{x \in \mathcal{B}}W^s(x)$ is the increasing union of the open sets $\{\bigcup_{x \in \mathcal{B}} \Ball^s_n(x)\}_n$. If it were the case that $M=\bigcup_{x \in \mathcal{B}}W^s(x)$, then $M$ would coincide with $\Ball^s_{n_0}(\mathcal{B})$ for some $n_0$. Then $f^{-n}(M)$ would be contained in a small neighbourhood of the proper compact subset $\mathcal{B}$ for some big enough $n$ which of course cannot be. Since $\bigcup_{x \in \mathcal{B}}W^s(x)$ is $f$-invariant and saturated by $\mathcal{W}^s$ then the same is the case for $\mathcal{R}$. \end{proof} We are on the way to prove that $\mathcal{R}$ is also saturated by the center foliation. The proof will rely on the following three lemmata. The next lemma is the main observation that will allow us to continue to work in a neighbourhood of $\mathcal{R}$ as if the center foliation were uniformly compact. \begin{lema}\label{Blargas} For every $\epsilon >0$ there exists $K>0$ such that if $\mathcal{C}$ is a center leaf with $\mathrm{vol}(\mathcal{C})>K$ then $\mathcal{C}\subset \Ball^s_\epsilon(\mathcal{B})$. \end{lema} \begin{proof} Fix $\epsilon >0$ and suppose that there is no such $K$. Then there exists a sequence $(x_n)_{n\in \mathbb{N}}$ in $M\setminus \Ball^s_\epsilon(\mathcal{B})$ such that $\mathrm{vol}(\mathcal{C}_{x_n})\xrightarrow{n}\infty$. By taking a convergent subsequence $x_{n_k}\xrightarrow{k} x\in M\setminus \Ball^s_\epsilon(\mathcal{B})$ we obtain that $\mathrm{vol}$ is not locally bounded in $x$. So $x$ should be a point of $\mathcal{B}$ but that is impossible since $x\in M\setminus \Ball^s_\epsilon(\mathcal{B})$ \end{proof} The following lemma relates diameter and volume of leaves: \begin{lema}\label{voldiam} For every $K>0$ there exists $D>0$ such that if $\mathcal{C}$ is a center leaf with $\mathrm{vol}(\mathcal{C})<K$ then $\diam(\mathcal{C})<D$. \end{lema} \begin{proof} Take $\{U_i\}_{i=1,\dotsc,l}$ a finite covering of $M$ by foliated boxes of the center foliation. For each $i\in \{1,\dotsc,l\}$ denote $d_i$ the supremum of the diameter of center plaques in $U_i$ and $v_i$ the infimum of the volume of center plaques in $U_i$. Denote $d=\max_i \{d_i\}$ and $v=\min_i \{v_i\}$. Now let $\mathcal{C}$ be a center leaf with $\mathrm{vol}(\mathcal{C})<K$. We have that $\mathcal{C}$ has less than $\frac{K}{v}+1$ plaques in each $U_i$ and then $\diam(\mathcal{C})<ld (\frac{K}{v}+1)=D$. \end{proof} From the continuity of $\mathcal{W}^c$ we have: \begin{lema}\label{continuosW^c} Given $D>0$ and $\epsilon>0$ there exists $\delta>0$ such that for any $x$ and $y$ with $\mathrm{d}(x,y)<\delta$ we have that $\Ball^c_D(y)\subset \Ball_\epsilon (\mathcal{C}_x)$. \end{lema} We can now prove: \begin{prop}\label{csaturated} The set $\mathcal{R}$ is saturated by the center foliation. \end{prop} \begin{proof} Suppose that there exists $x\in \mathcal{R}$ and $y\in \mathcal{C}_x$ such that $y\in M\setminus \mathcal{R}$. Since $y\in M\setminus \mathcal{R}$ there exists $w\in \mathcal{B}$ such that $y\in W^s(w)$. Denote $d=\mathrm{d}(\mathcal{B},\mathcal{R})>0$. Note by Lemma \ref{Blargas} that $\mathrm{vol}$ is bounded in $\mathcal{R}$, say by some constant $K>0$. By Lemma \ref{voldiam} there exists $D>0$ such that for every center leaf $\mathcal{C}$ with $\mathrm{vol}(\mathcal{C})<K$ the diameter of $\mathcal{C}$ is less than $D$. So for every $z\in \mathcal{R}$ we have that $\Ball^c_D(z)=\mathcal{C}_z$. We can now consider $N$ large enough so that, by Lemma \ref{continuosW^c}, the points $f^N(w)$ and $f^N(y)$ are as close as to assure that $\Ball^c_D(f^N(y))\subset \Ball_{d/2}(\mathcal{B})$. This yields a contradiction since $f^N(y)\in \mathcal{C}_{f^N(x)}$ and $f^N(x)\in \mathcal{R}$ since $\mathcal{R}$ is $f$-invariant. This shows that for every $x\in \mathcal{R}$ the leaf $\mathcal{C}_x \subset \mathcal{R}$. \end{proof} \subsection{Completeness and trivial holonomy for center-stable leaves in $\mathcal{R}$}\label{completeandbundle} In this section we prove some properties of $\mathcal{R}$ in order to implement the proof of Hiraide in Section \ref{Hirai}. Let us first see in the following proposition that center-stable leaves in $\mathcal{R}$ are \textit{complete} (this terminology is used in \cite{BoW}, \cite{Car} and \cite{BoBo}). \begin{prop}\label{completeness} For every center leaf $\mathcal{C}$ in $\mathcal{R}$ we have that $W^{cs}(\mathcal{C})=\bigcup_{x\in \mathcal{C}}W^s(x)$. \end{prop} \begin{proof} Let $\mathcal{C}$ be a center leaf in $\mathcal{R}$. If we prove that $\bigcup_{x\in \mathcal{C}}W^s(x)$ is saturated by center leaves then $\bigcup_{x\in \mathcal{C}}W^s(x)$ will be a non empty open and closed subset of $W^{cs}(\mathcal{C})$ and then will coincide with $W^{cs}(\mathcal{C})$. Let us consider $y_0\in \bigcup_{x\in \mathcal{C}}W^s(x)$. We want to see that $\mathcal{C}_{y_0}\subset \bigcup_{x\in \mathcal{C}}W^s(x)$. Let $x_0$ be any point in $\mathcal{C}$ and let $\gamma:[0,1]\to W^{cs}(\mathcal{C})$ be a continuous path from $x_0$ to $y_0$. For every $t\in [0,1]$ denote by $\mathcal{C}_t$ the center leaf trough $\gamma(t)$. Recall that by Proposition \ref{Holinfty} the center leaves in $\mathcal{B}$ coincide with those with infinite holonomy group. Thus, since $\mathcal{R}\cap \mathcal{B}=\emptyset$, for every $t\in [0,1]$ we have that $\mathcal{C}_t\subset W^{cs}(\mathcal{C})$ satisfies $|\Hol^s(\mathcal{C}_t)|< \infty$. We can then take for each $\mathcal{C}_t$ a neighborhood $\mathcal{U}^s(\mathcal{C}_t)$ of $\mathcal{C}_t$ in $W^{cs}(\mathcal{C}_t)$ given by the Generalized Reeb stability theorem (see Theorem \ref{Reeb}). The associated projection $\pi_t:\mathcal{U}^s(\mathcal{C}_t)\to \mathcal{C}_t$ can be taken such that $\pi_t^{-1}(x)$ is a disk in $W^s(x)$ for every $x\in \mathcal{C}_t$. (See Remark \ref{Reebproduct}). Then $\{\mathcal{U}^s(\mathcal{C}_t)\}_{t\in [0,1]}$ is an open cover of $\gamma([0,1])$. Let us take a finite subcover $\{\mathcal{U}^s(\mathcal{C}_{t_0}),\dotsc,\mathcal{U}^s(\mathcal{C}_{t_k})\}$ such that $\mathcal{C}_{x_0}=\mathcal{C}_{t_0}$, $\mathcal{C}_{y_0}=\mathcal{C}_{t_k}$ and $\mathcal{U}^s(\mathcal{C}_{t_i})\cap \mathcal{U}^s(\mathcal{C}_{t_{i+1}})\neq \emptyset$ for every $0\leq i \leq k-1$. Observe that, if $\mathcal{C}'$ is a center leaf in some $\mathcal{U}^ s(\mathcal{C}_{t_i})$, then each stable disk of a point of $\mathcal{C}_{t_i}$ intersects $\mathcal{C}'$. Observe also that since each $\mathcal{U}^s(\mathcal{C}_{t_i})$ is saturated by center leaves, so it is each $\mathcal{U}^s(\mathcal{C}_{t_i})\cap \mathcal{U}^s(\mathcal{C}_{t_{i+1}})$. Then, by taking $\mathcal{C}'_i\subset \mathcal{U}^s(\mathcal{C}_{\gamma(t_i)})\cap \mathcal{U}^s(\mathcal{C}_{\gamma(t_{i+1})})$ we deduce that each stable leaf of $\mathcal{C}_i$ intersects $\mathcal{C}_{i+1}$. This implies that $\mathcal{C}_{y_0}\subset \bigcup_{x\in \mathcal{C}}W^s(x)$ as we wanted. \end{proof} The following is a mild extension of Lemma \ref{Blargas} that will come handy later. \begin{lema} There exists a constant $C>0$ such that for every $x$ in $\mathcal{R}$ we have $\#\{\mathcal{C}_x\cap W^s(x)\}< C$. \end{lema} \begin{proof} Cover $\mathcal{R}$ by a finite number $\{U_i\}_{1\leq i \leq k}$ of foliated boxes for the $\mathcal{W}^c$ foliation. Let $d>0$ be such that each plaque of each $U_i$ has volume larger than $d$. By Lemma \ref{Blargas} the volume function is bounded in $\mathcal{R}$, say by some constant $K$. Let us see that $\#\{\mathcal{C}_x\cap W^s(x)\}< K/d +1$ for every $x\in \mathcal{R}$. Suppose there exists $x\in \mathcal{R}$ and distinct points $\{x=x_0,\dotsc,x_l\}\subset \{\mathcal{C}_x\cap W^s(x)\}$ with $l \geq K/d +1$. Let $\gamma>0$ be a Lebesgue number for the covering $\{U_i\}_{1\leq i \leq k}$. Then we can consider $N$ large enough as to assure that $\diam(\{f^N(x_0),\dotsc,f^N(x_l)\})<\gamma$ in $W^s(f^N(x_0))$ with the intrinsic topology. So $\{f^N(x_0),\dotsc,f^N(x_l)\}$ is contained in some member $U_{i_0}$ of the covering. Since the points $\{f^N(x_0),\dotsc,f^N(x_l)\}$ are close in $W^s(f^N(x_0))$ in the intrinsic topology, then trough each one of them there correspond a different plaque of $U_{i_0}$. On the other hand, $\{f^N(x_0),\dotsc,f^N(x_l)\}\subset f^N(\mathcal{C}_x)$. This contradicts the fact that $\mathrm{vol} (f^N(\mathcal{C}_x)) \leq K$. \end{proof} We can now give some kind of a description of center-stable leaves in $\mathcal{R}$ (see Bohnet \cite[Corollary 4.10.]{Boh} for a similar result): \begin{prop}\label{csleavesR} Let $W$ be a center-stable leaf in $\mathcal{R}$. Then there exists a center leaf $\mathcal{C}$ in $W$ such that for every $x\in\mathcal{C}$ we have that $\mathcal{C} \cap W^s(x)=\{x\}$. Therefore, $W$ is a bundle with base $\mathcal{C}$ and fibers $\{W^s(x)\}_{x\in \mathcal{C}}$. \end{prop} \begin{proof} Observe that it is enough to prove what we want for some $f^N(W)$. Cover $\mathcal{R}$ by finite $(\mathcal{W}^{cs},\mathcal{W}^u)$-product neighborhood $\{\mathcal{U}^s(\mathcal{C}_i)\times [-1,1]\}_{1\leq i \leq k}$ with each $\mathcal{U}^s(\mathcal{C}_i)$ being a Generalized Reeb stability neighborhood of the center leaf $\mathcal{C}_i$ (see Remark \ref{Reebproduct}). Let $\gamma>0$ be a Lebesgue number for the covering. Let $W$ be a center-stable leaf in $\mathcal{R}$. By the previous lemma, we have that $\#\{\mathcal{C}_x\cap W^s(x)\}$ is bounded by a constant $C>0$ for every $x\in W$. So, let us consider $x_0\in W$ such that $l=\#\{\mathcal{C}_{x_0}\cap W^s(x_0)\}$ is maximal in $W$. Let $\{x_0,\dotsc,x_{l-1}\}=\mathcal{C}_{x_0}\cap W^s(x_0)$. Let $N>0$ be large enough so that $\diam(\{f^N(x_0),\dotsc,f^N(x_{l-1})\})<\gamma$ in $W^s(f^N(x_0))$ with the intrinsic topology. Then there exists $i\in \{1,\dotsc,k\}$ and $t\in [-1,1]$ such that $\{f^N(x_0),\dotsc,f^N(x_{l-1})\}\subset \mathcal{U}^s(\mathcal{C}_i)\times \{t\}$. Moreover, $\{f^N(x_0),\dotsc,f^N(x_{l-1})\}$ belongs to the same $s$-disk in the Generalized Reeb stability neighborhood $\mathcal{U}^s(\mathcal{C}_i)\times \{t\}$. So each $s$-disk of $\mathcal{U}^s(\mathcal{C}_i)\times \{t\}$ intersects $f^N(\mathcal{C}_{x_0})$ in at least $l$ distinct points. This implies that $(\mathcal{C}_i\times \{t\}) \cap W^s(x)=\{x\}$ for every $x\in \mathcal{C}_i\times \{t\}$. Otherwise, $f^N(\mathcal{C}_{x_0})$ would intersect some stable leaf in at least $2l$ distinct points and this cannot be since $l=\#\{\mathcal{C}_{x_0}\cap W^s(x_0)\}$ is maximal in $W$ and, therefore, also in $f^N(W)$. \end{proof} \begin{obs} For $W$ as in Proposition \ref{csleavesR} we have that the group $\Hol(W)$ is trivial. Indeed, any closed curve in $W$ is freely homotopic to a closed curve in $\mathcal{C}$, that has trivial unstable holonomy (see Lemma \ref{lemma1}). \end{obs} \subsection{Adapted Hiraide arguments to rule out the existence of $\mathcal{R}$}\label{Hirai} This last section is dedicated to prove that the set $\mathcal{R}$ as described before cannot exist. The proof we give is an adaptation of the work made by Hiraide in \cite{Hir} and Bohnet in \cite{Boh}. However, the proof itself is self-contained. The key advantage of Hiraide's proof over Newhouse's (Anosov case, see \cite{New}) is that the former takes place in a neighbourhood of the repeller while the latter makes a more global argument. As we want to avoid dealing with the bad set, we find it more convenient to follow Hiraide's proof. It is worth mention that because of the reasons just mentioned the authors could not directly adapt Newhouse's proof. From now on we will work with both $\mathcal{R}$ and its boundary $\partial \mathcal{R}$ in $M$. Note that, as well as $\mathcal{R}$, the set $\partial \mathcal{R}$ is non empty, closed, saturated by the center-stable foliation, has trivial transversal holonomy and the volume of its center leaves is uniformly bounded. The set $\partial \mathcal{R}$ has empty interior. Let us fix an orientation of $\mathcal{W}^u$. For every $x\in \mathcal{R}$ we can consider $\mathcal{U}^s(x)\times [-1,1]$ a $(\mathcal{W}^{cs},\mathcal{W}^u)$-product neighbourhood of $x$ with $\mathcal{U}^s(x)$ a small center-saturated Generalized Reeb stability neighborhood of $\mathcal{C}_x$ (see Remark \ref{Reebproduct}). Let $\{V_i=\mathcal{U}^s(x_i)\times (-1,1)\}_{0\leq i \leq k}$ be a finite cover of $\mathcal{R}$. Define $\mathcal{V}=\bigcup_{0\leq i\leq k} V_i$. We are going to see that for certain points near $\partial \mathcal{R}$ the center-stable leaf through this point must remain in $\mathcal{V}$ (see ``Sandwich Lemma'' \ref{Sandwich}) while it must also intersect $\mathcal{B}$, thus yielding a contradiction. \begin{obs} We can assume that: \begin{itemize} \item $\mathcal{V}\cap \mathcal{B} =\emptyset$ (this is obtained by taking each $\mathcal{U}^s(x)\times [-1,1]$ disjoint from $\mathcal{B}$). \item $\mathcal{U}^s(x_i)\times \{1\}\cap \partial \mathcal{R}=\emptyset$ (we can assume this since $\partial \mathcal{R}$ has empty interior). \end{itemize} \end{obs} For every $i\in \{0,\dotsc,k\}$ let $0<t_i<1$ be such that $\mathcal{U}(x_i)\times [t_i,1]\cap \partial \mathcal{R} = \mathcal{U}(x_i)\times \{t_i\}$. Denote each $\mathcal{U}(x_i)\times \{t_i\}$ as $P^+_i$. Informally speaking, $P^+_i$ is the last (according to the orientation of $\mathcal{W}^u$) center-stable plaque of $V_i$ that is contained $\partial \mathcal{R}$. \begin{lema} There exists a pair $(x_0,\delta)\in \partial \mathcal{R}\times \mathbb{R}^+$ such that either $(x_0,x_0+\delta)^u\cap \mathcal{R} =\emptyset$ or $(x_0-\delta,x_0)^u\cap \mathcal{R} =\emptyset$. \end{lema} \begin{proof} Observe first that the set $\mathcal{R}$ cannot be saturated by the unstable foliation (because in that case it would be all $M$), so there must exist a point $x \in \mathcal{R}$ such that $W^u(x)\cap M\setminus \mathcal{R} \neq \emptyset$. As $\mathcal{R}\cap W^u(x)$ is closed in $W^u(x)$ with the intrinsic topology then there must be at least one connected component $I$ of $W^u(x)\setminus \mathcal{R}$. Choose $x_0$ an endpoint of $I$. \end{proof} Fix $(x_0, \delta)$ given by the previous lemma. Assume that $(x_0,x_0+\delta)^u\cap \mathcal{R} =\emptyset$ (otherwise, simply change the orientation of $\mathcal{W}^u$). By Proposition \ref{completeness} and Proposition \ref{csleavesR} we have $W^{cs}(x_0)=\bigcup_{x\in \mathcal{C}}W^s(x)$ for some center leaf $\mathcal{C}\subset W^{cs}(x_0)$ such that $\mathcal{C}\cap W^s(x)=\{x\}$ for every $x\in \mathcal{C}$. Assume, without loss of generality, that $\mathcal{C}_{x_0}$ is such a center leaf. Some of the plaques $\{P^+_1,\dotsc,P^+_k\}$ could possibly be contained in $W^{cs}(x_0)$. Denote all of them as $\{P^+_{i_1},\dotsc,P^+_{i_m}\}$. We can consider now $N>0$ large enough as to assure that: $$(P^+_{i_1}\cup \dotsb \cup P^+_{i_m}) \subset \bigcup_{x\in \mathcal{C}_{x_0}}\Ball^s_N(x).$$ If $W ^{cs}(x_0)$ contains none of the plaques $\{P^+_1,\dotsc,P^+_k\}$ then take $N>0$ any positive number. For simplicity, let us denote $\bigcup_{x\in \mathcal{C}_{x_0}}\Ball^s_N(x)$ as $E$. The subset $E$ of $W^{cs}(x_0)$ is then in the hypothesis of the Proposition \ref{productnbhd} (in particular, $\Hol^u(E)=\Id$) so we can consider a $(\mathcal{W}^{cs},\mathcal{W}^u)$-product neighborhood $E\times [-1,1]$ of it. Recall that this means that each $E\times \{y\}$ lies in a center-stable leaf, each $\{x\}\times [-1,1]$ lies in an unstable leaf and $E\times \{0\}=E$. By eventually shriking it in the unstable direction, we can assume that $E\times [-1,1]$ is contained in $\mathcal{V}$. \begin{lema} Through every $x\in W^{cs}(x_0)\setminus E$ there exists $L_x=[x,x+\delta_x ]^u$ a closed, non trivial unstable segment that intersects $\mathcal{R}$ just in its endpoints. Moreover, $L_x$ varies continuously with $x$ in $W^{cs}(x_0)\setminus E$ and is contained in any element $V_i$ of the covering that contains $x$. \end{lema} \begin{proof} Let $x\in W^{cs}(x_0)\setminus E$. The point $x$ lies in some $V_i$. Then it must be that the connected component of $(W^u_+(x)\cup \{x\})\cap V_i$ that contains $x$ intersects $\mathcal{R}$ in at least some other point distinct from $x$ (since $x\notin \bigcup_{0\leq i\leq k}P^+_i$). Observe that the fact that $x$ lies inside $W ^{cs}(x_0)$ and $(x_0,x_0+\delta)^u$ is disjoint from $\mathcal{R}$ implies that $x$ cannot be accumulated in $W^u_+(x)\cup \{x\}$ by points of $W^u_+(x)\cap \mathcal{R}$. The existence of the stated $\delta_x>0$ follows. The definition of $L_x$ does clearly not depend on the choice of the $V_i$ containing $x$. The continuous dependence on $x$ follows. \end{proof} By shrinking $E\times [-1,1]$ even more (if necessary) in the unstable direction, we can assume that $\{x\}\times [0,1]\subset L_x$ for every $x\in \partial E$. Now, let $y\in (x_0,x_0+\delta)^u$ be close enough to $x_0$ such that $y\in E\times [0,1]$. Since $y\in M\setminus \mathcal{R}$ we have that $W^{cs}(y)\cap \mathcal{B}\neq \emptyset$. We are going to see that, on the other hand, $W^{cs}(y)\subset \mathcal{V}$ and this will yield a contradiction since $\mathcal{V}\cap \mathcal{B}=\emptyset$. The proof then is finished up to the following: \begin{lema}[Sandwich lemma]\label{Sandwich} For $y$ as above we have $W^{cs}(y)\subset \mathcal{V}$. \end{lema} \begin{proof} First, note that $\bigcup_{x\in W^{cs}(x_0)\setminus E}L_x$ is a foliated interval-bundle with base $W^{cs}(x_0)\setminus E$ and fibers the $L_x$'s that are transversal to $\mathcal{W}^{cs}$ for every $x$. We have then a well defined projection along fibers $\pi:\bigcup_{x\in W^{cs}(x_0)\setminus E}L_x\to W^{cs}(x_0)\setminus E$ given by $\pi([x,x+\delta_x))=x$ for $L_x=[x,x+\delta_x]$ . For every $\gamma:[0,1]\to W^{cs}(x_0)\setminus E$ and $z\in L_{\gamma(0)}$ we can lift $\gamma$ to $\gamma_z:[0,1]\to W^{cs}(z)$ such that $\pi\circ\gamma_z=\gamma$. This lift defines a projection $p_\gamma:L_{\gamma(0)}\to L_{\gamma(1)}$ given by $p_\gamma(z)= \gamma_z(1)$. Fix $x_0'\in \partial E$. \vspace{0.2cm} \emph{Claim.} For every curve $\gamma:[0,1]\to W^{cs}(x_0)\setminus E$ such that $\gamma(0)=\gamma(1)=x_0'$ there exists an endpoints fixed homotopy $\gamma_s:[0,1]\to W^{cs}(x_0)\setminus E$ such that $\gamma_0=\gamma$ and $\gamma_1([0,1])\subset \partial E$. \vspace{0.2cm} \emph{Proof of claim.} For every $s\in (0,1]$ we can consider a retraction $r_s:W^{cs}(x_0)=\bigcup_{x\in \mathcal{C}_{x_0}}W^s(x)\to \bigcup_{x\in \mathcal{C}_{x_0}}\Ball^s_{N/s}(x)$ such that $r_s$ varies continously with $s$ and is the identity in $E=\bigcup_{x\in \mathcal{C}_{x_0}}\Ball^s_{N}(x)$. For every $s\in (0,1]$ compose $\gamma$ with $r_s$ to get a curve $\gamma_s$ and set $\gamma_0$ as $\gamma$. Then the homotopy $\gamma_s$ is as desired and this proves the claim. \vspace{0.3cm} Now, for every $x\in W^{cs}(x_0)\setminus E$ consider $\gamma_x:[0,1]\to W^{cs}(x_0)\setminus E$ such that $\gamma_x(0)=x_0'$ and $\gamma_x(1)=x$. \begin{figure}[htb] \centering \def300pt{300pt} \input{Figure4.pdf_tex} \caption{}\label{figHir} \end{figure} \vspace{0.2cm} \emph{Remark.} The existence of such a $\gamma_x$ is guaranteed if $\mathrm{dim}(E^s)\geq 2$ for this implies that $W^{cs}(x_0)\setminus E$ is path connected. Otherwise, if $\mathrm{dim}(E^s)=1$ and $W^{cs}(x_0)\setminus E$ has two connected components simply add $x_0''\in \partial E$ not in the same connected component as $x_0'$ and reproduce the arguments that follow separately in each connected component with $x_0'$ and $x_0''$ playing the same role. \vspace{0.2cm} Denote $y=(x_0,t)$ in $E\times (0,1)$. Observe that we can extend $E\times (0,1)$ to a $(\mathcal{W}^{cs},\mathcal{W}^u)$-product neighborhood $(E\cup \partial E)\times (0,1)$. Denote $y'$ the point $(x_0',t)$ in this coordinates. Define $F:W^{cs}(x_0)\setminus E\to W^{cs}(y)$ by $$F(x)=p_{\gamma_x}(y').$$ \vspace{0.2cm} \emph{Claim.} The definition of $F$ does not depend on the choice of $\gamma_x$. \vspace{0.2cm} \emph{Proof of claim.} If $\gamma_x':[0,1]\to W^{cs}(x_0)\setminus E$ is another path such that $\gamma_x'(0)=x_0'$ and $\gamma_x'(1)=x$, we have to prove that $\gamma_x'^{-1}*\gamma_x$ lifts to a closed path $(\gamma_x*\gamma_x'^{-1})_{y'}$ from $y'$ to itself. By the first claim, $\gamma_x'^{-1}*\gamma_x$ is homotopic to some closed path $\alpha:[0,1]\to W^{cs}(x_0)\setminus E$ such that $\alpha([0,1])\subset \partial E$. This homotopy lifts to an homotopy contained in $W^{cs}(y')$ from $(\gamma_x'^{-1}*\gamma_x)_{y'}$ to $\alpha_{y'}$, with $\alpha_{y'}$ being the lift of $\alpha$ from $y'$. Recall that we have $y'\in (E\cup\partial E)\times \{t\}$. Then the lift $\alpha_{y'}$ has to be closed because $\alpha_{y'}([0,1])$ is contained in $W^{cs}(y')$ and this implies $\alpha_{y'}([0,1])\subset (\partial E\times \{t\})$ and then $\alpha_{y'}(1)=(\partial E\times \{t\})\cap L_{x_0'}=y'$. This proves the claim. \vspace{0.3cm} So we have a well defined map $F:W^{cs}(x_0)\setminus E\to W^{cs}(y)$. We can extend $F$ to $E$ in the natural way: $F(x)=(x,t)$ for every $x\in E$. Then we have $F:W^{cs}(x_0)\to W^{cs}(y)$. The map $F$ is clearly an injective local homeomorphism. Given a Lebesgue number $\eta>0$ for the covering $\mathcal{V}=\bigcup_{0\leq i\leq k} V_i$ there exists $\delta>0$ such that for every $x\in W^{cs}(x_0)$ we have that $\Ball^{cs}_\delta(F(x))\subset F(\Ball^{cs}_{\eta/2}(x))$. We deduce from this that $F$ is a proper map. Then $F$ is also surjective and this implies that $W^{cs}(y)\subset \mathcal{V}$ since $E\times \{t\}$ and every $L_x$ are contained in $\mathcal{V}$. \end{proof}

00015aff978906b25f49c4651a65aa94df399671

\subsection{ILP Tasks} Description of tasks and the proto-rule templates used during training. More details are in \cite{DBLP:journals/corr/abs-1711-04574} \subsubsection{Predecessor} In this task we aim to learn the predecessor relation, $predecessor(X,Y) \leftarrow succ(Y,X)$, from basic aritmetic facts $\{zer(0),succ(0,1),suc(1,2), ...\}$. We use the following template: \begin{align*} F(X,Y) &\leftarrow F(Y,X) & (9) \end{align*} \subsubsection{Even-Odd} In this task we aim to learn the $even$ predicate. Here the background knowledge is the same as in Predecessor (above). We must include an extra auxiliary predicate that learns to encode the relation $odd$. We use the following templates: \begin{align*} F(X) &\leftarrow F(X) & (1)\\ F(X) &\leftarrow F(Z), F(Z,X) &(2)\\ F(X) &\leftarrow F(Z), F(Z,X) &(2) \end{align*} \subsubsection{Even-succ2} Described in Experiments. \begin{align*} F(X) &\leftarrow F(X) & (1)\\ F(X) &\leftarrow F(Z), F(Z,X) &(2)\\ F(X,Y) &\leftarrow F(X,Z), F(Z,Y) & (3) \end{align*} \subsubsection{Less Than} Here we aim to learn the $lessThan$ relation. Background knowledge is the same as in the tasks above. A possible solution would be:\\ \\ $lessThan(X,Y) \leftarrow succ(X,Y)$ \\ $lessThan(X,Y) \leftarrow lessThan(X,Z), lessThan(Z,Y)$\\ \\ We used these templates: \begin{align*} F(X,Y) &\leftarrow F(X,Y) & (5)\\ F(X,Y) &\leftarrow F(X,Z), F(Z,Y) &(3) \end{align*} \subsubsection{Fizz} As in the children game Fizz-Buzz, numbers that are divisible by three should be classified as Fizz. For more details refer to \cite{DBLP:journals/corr/abs-1711-04574} These are the template protorules used during training: \begin{align*} F(X) &\leftarrow F(X) & (1)\\ F(X) &\leftarrow F(Z), F(Z,X) &(2)\\ F(X,Y) &\leftarrow F(X,Z), F(Z,Y) & (3)\\ F(X,Y) &\leftarrow F(X,Z), F(Z,Y) & (3) \end{align*} \subsubsection{Buzz} Following the same logic as in Fizz, we used the following templates: \begin{align*} F(X) &\leftarrow F(X) & (1)\\ F(X) &\leftarrow F(Z), F(Z,X) &(2)\\ F(X,Y) &\leftarrow F(X,Z), F(Z,Y) & (3) \end{align*} \subsubsection{Member} Here we aim to learn $member(X,Y)$, which is true if $X$ is an element of list $Y$. The background knowledge encodes values on a list by using two predicates: $cons(X,Y)$ which is true if node $Y$ is after list $X$ (lists are terminated with the null node 0); and $value(X,Y)$ which is true if the value of node $X$ is $Y$. One possible solution is:\\ \\ $member(X,Y) \leftarrow value(Y,X)$\\ $member(X,Y) \leftarrow cons(Y,Z), member(X,Z)$\\ \\ We used these templates: \begin{align*} F(X,Y) &\leftarrow F(Y,X) & (9)\\ F(X,Y) &\leftarrow F(Y,Z), F(X,Z) &(10) \end{align*} \subsubsection{Length} The $length(X,Y)$ relation is true if the length of list $X$ is $Y$. We represent lists in the same way as in the $Member$ task. We required at least one extra intentional predicate $pred1$. One possible solution would be:\\ \\ $Length(X,X)\leftarrow zero(X)$\\ $Length(X,Y)\leftarrow cons(X,Z), pred1(Z,Y)$\\ $pred1(X,Y) \leftarrow Length(X,Z), succ(Z,Y)$\\ \\ We used these templates: \begin{align*} F(X,X) &\leftarrow F(X) & (8)\\ F(X,Y) &\leftarrow F(X,Z), F(Z,Y) & (3)\\ F(X,Y) &\leftarrow F(X,Z), F(Z,Y) & (3) \end{align*} \subsubsection{Son} We aim to learn $sonOf(X,Y)$ relation from family-related facts involving $fatherOf$, $brotherOf$ and $sisterOf$. We required at least one extra intentional predicate that learns the relation $is-male$. One possible solution would be:\\ \\ $sonOf(X,Y)\leftarrow father(Y,X), isMale(X)$\\ $isMale(X)\leftarrow brother(X,Z)$\\ $isMale(X) \leftarrow father(X,Z)$\\ \\ We used these templates: \begin{align*} F(X,Y) &\leftarrow F(Y,X), F(X) & (11)\\ F(X) &\leftarrow F(X,Z) & (12)\\ F(X) &\leftarrow F(X,Z) & (12) \end{align*} \subsubsection{Grandparent} The goal of this task is to infer the grandparent relation from observed mother-of and father-of facts. Our templates were: \begin{align*} F(X,Y) &\leftarrow F(X,Z), F(Z,Y) & (3)\\ F(X,Y) &\leftarrow F(X,Y) & (5)\\ F(X,Y) &\leftarrow F(X,Y) & (5) \end{align*} \subsubsection{Relatedness} $related(X,Y)$ is true if there is an undirected path between $X$ and $Y$. Background knowledge contains family related facts as in the tasks Son and Grandparent. We used these templates: \begin{align*} F(X,Y) &\leftarrow F(X,Y) & (5)\\ F(X,Y) &\leftarrow F(X,Y) & (5)\\ F(X,Y) &\leftarrow F(X,Z), F(Z,Y) &(3)\\ F(X,Y) &\leftarrow F(Y,X) &(9) \end{align*} \subsubsection{Father} In this task we aim to learn the $Father$ relation in challenging set up (incomplete background knowledge and irrelevant facts). For details please refer to \cite{DBLP:journals/corr/abs-1711-04574}. We used these template: \begin{align*} F(X,Y) &\leftarrow F(X,Z), F(Z,Y) & (3) \end{align*} \subsubsection{Undirected Edge} In this task the background knowledge is composed of several $edge(X,Y)$ facts. The goal is to learn $undirectedEdge(X,Y)$ which is true if there is an edge between nodes $X$ and $Y$ regardless of the direction. We used the templates: \begin{align*} F(X,Y) &\leftarrow F(X,Y) & (5)\\ F(X,Y) &\leftarrow F(Y,X) &(9) \end{align*} \subsubsection{Adjacent to Red} In this example we extend the background knowledge of the example above with color facts: $green(C)$, $red(C)$, as well as $colour(X,C)$ which is true if node $X$ is of colour $C$. We included one auxiliary predicate that learns the relation $isRed(X)$. One possible solution would be:\\ \\ $adjToRed(X)\leftarrow edge(X,Y), isRed(Y)$\\ $isRed(X)\leftarrow colour(X,Y), red(Y)$\\ \\ We used these templates: \begin{align*} F(X) &\leftarrow F(X,Z), F(Z) &(13)\\ F(X) &\leftarrow F(X,Z), F(Z) &(13) \end{align*} \subsubsection{Two Children} Here we aimed to learn the \textit{has-at-least-two-children(X)} predicate, which is true if there are at least two facts of the form $edge(X,Z)$. The background knowledge includes $edge$ and $neq$ (not equals) relations. We included one auxiliary predicate $pred1$. One possible solution would be:\\ \\ $twoChildren(X)\leftarrow edge(X,Y), pred1(X,Y)$\\ $pred1(X,Y)\leftarrow edge(X,Z), neq(Z,Y)$\\ \\ We used these templates: \begin{align*} F(X,Y) &\leftarrow F(X,Z), F(Z,Y) &(3)\\ F(X,X) &\leftarrow F(X,Z), F(X,Z) & (15) \end{align*} \subsubsection{Graph Colouring} The task is to learn the \textit{adj-to-same(X,Y)} which is true if nodes $X,Y$ are of the same colour and there is an edge between them. The background knowledge is similar as in the task \textit{Adjacent to Red}. We included an auxiliary predicate that should learn the relation \textit{same-colour(X,Y)}.One possible solution would be:\\ \\ $adjToSame(X,Y)\leftarrow edge(X,Y), sameColour(X,Y)$\\ $sameColour(X,Y)\leftarrow colour(X,Z), colour(Y,Z)$\\ \\ \begin{align*} F(X,Y) &\leftarrow F(X,Z), F(Y,Z) &(10)\\ F(X,X) &\leftarrow F(X,Z), F(X,Z) & (15) \end{align*} \subsubsection{Connectedness} In this task we want to learn \textit{connected(X,Y)} which is true if there is a sequence of edges connecting nodes $X$ and $Y$. We used the templates: \begin{align*} F(X,Y) &\leftarrow F(X,Y) & (5)\\ F(X,Y) &\leftarrow F(X,Z), F(Z,Y) &(3) \end{align*} \subsubsection{Graph Cyclicity} In this task the algorithm should learn the cocept of cyclicity. This is true of a node when there is a path departing from it and arriving to itself. The templates used were: \begin{align*} F(X) &\leftarrow F(X,X) & (4)\\ F(X,Y) &\leftarrow F(X,Y) & (5)\\ F(X,Y) &\leftarrow F(X,Z), F(Z,Y) & (3) \end{align*} \subsection{Countries} We mimicked some of the templates found in the appendix E of \cite{DBLP:journals/corr/RocktaschelR17}. In order to closely follow their approach, only for this task we enforced some predicates to be the same, so that \#1 represents the same predicate across the rule. In detail for each task: \subsubsection{Countries S1} \begin{align*} \#1(X,Y) &\leftarrow \#1(Y,X) \\ \#1(X,Y) &\leftarrow \#2(X,Z), \#2(Z,Y) \end{align*} \textbf{Countries S2} \begin{align*} \#1(X,Y) &\leftarrow \#1(Y,X) \\ \#1(X,Y) &\leftarrow \#2(X,Z), \#3(Z,Y) \end{align*} \textbf{Countries S3} \begin{align*} \#1(X,Y) &\leftarrow \#1(Y,X) \\ \#1(X,Y) &\leftarrow \#2(X,Z), \#3(Z,W), \#4(W,Y) \end{align*} \subsection{Taxonomy and Kinship} \subsubsection{Taxonomy} This task was described in section \ref{taxonomy}, the dataset can be seen in figure \ref{fig:taxonomy-big}. We used the following templates to perform this task: \begin{align*} F(X,Y) &\leftarrow F(X,Z), F(Z,Y) & (3) \\ F(X,Y) &\leftarrow F(X,Z), F(Z,Y) & (3) \\ F(X,Y) &\leftarrow F(X,Z), F(Z,Y) & (3) \\ F(X,Y) &\leftarrow F(X,Z), F(Z,Y) & (3) \end{align*} \begin{figure}[!ht] \centering \includegraphics[scale=.3]{taxonomybig.png} \caption{Bigger Animal Taxonomy used for the tasks. Contains 4 predicates, 36 constants and 145 facts} \label{fig:taxonomy-big} \end{figure} \pagebreak \subsubsection{Kinship} This task was described in section \ref{kinshipSec}. The theory can be seen in figure \ref{fig:family} \begin{figure}[!ht] \centering \includegraphics[scale=.4]{family.png} \caption{Inferred family tree. Females shown in bold italics and males in ordinary font.} \label{fig:family} \end{figure} \begin{comment} \vspace{2cm} \pagebreak \subsection{Helper Functions for \texttt{infer}} \begin{algorithmic} \Function{\texttt{get\_constants}}{\texttt{u}, \texttt{head}} \State $\texttt{constants = []}$ \For{$i = 1 \texttt{to} \texttt{len}(\texttt{get\_args(head))})$} \State $\texttt{constants} \leftarrow \texttt{concat}(\texttt{constants}, \texttt{u}(\texttt{get\_arg}(\texttt{head}, i)))$ \EndFor \\ \Return \texttt{constants} \EndFunction \end{algorithmic} \begin{algorithmic} \Function{\texttt{unify}}{$(\texttt{f}_i), (\texttt{b}_i), \texttt{n}$} \State $\texttt{u} = \{\}$ \For{$\texttt{i} = 1 \texttt{to} \texttt{n}$} \State $\texttt{vars} \leftarrow \texttt{get\_vars}(\texttt{b}_i)$ \State $\texttt{cons} \leftarrow \texttt{get\_constants}(\texttt{f}_i)$ \For{$\texttt{j} = 1 \texttt{to} \texttt{len}(\texttt{vars})$} \State $\texttt{curr\_val} \leftarrow \texttt{u}(\texttt{vars}_j)$ \If {\texttt{curr}\_val == \texttt{null}} \State $\texttt{u} \leftarrow \texttt{append}(\texttt{u}, (\texttt{vars}_j, \texttt{cons}_j))$ \Else{\If {$\texttt{curr\_val} \neq \texttt{cons}_j$} \Return \texttt{null} \EndIf } \EndIf \EndFor \EndFor \\ \Return \texttt{u} \EndFunction \end{algorithmic} \end{comment} \subsection{Inductive Logic Programming} In ILP, the goal is to learn (induce) logical rules which can be chained to successfully answer queries about a target relation, given positive and negative examples of that relation and some background facts. Logical rules are of the form: \begin{equation} h \leftarrow b_1, b_2, \ldots, b_k \end{equation} where $h$ is an atom called the \emph{head} of the rule and $b_1, b_2, \ldots, b_k$ are atoms which constitute the \emph{body}. An \emph{atom} is a positive or negative literal. A literal is a predicate applied to terms which may, in our case, be either variables or constants. For example $grandfather(X,Y)$ is a (positive) atom whose predicate is $grandfather(\cdot, \cdot)$ and whose arguments (two in this case) are variables $X$ and $Y$. When the arguments of the atom are all constants (e.g. $parent(Tom, Bill)$ for constants $Tom$ and $Bill$) we call it a \emph{ground atom} which we also refer to as a \emph{fact} when it is given as true or its truth value is inferred from rules. Intuitively, the head of the rule is true if each of the $b_i$ in the body are. For example a rule might be: $$grandfather(X, Y) \leftarrow father(X, Z), parent(Z, Y)$$ Which is read: \emph{$X$ is the grandfather of $Y$ if $X$ is the father of $Z$ AND $Z$ is the parent of $Y$}. The head atom holds for any $X$ and $Y$ as long there is some individual $Z$ for which $father(X,Z)$ and $parent(Z,Y)$ are both true facts. Given background facts such as $father(Bill, Mary)$ and $parent(Mary, Liz)$, logical rules can be chained together to prove a goal fact like $grandfather(Bill, Liz)$. This is an example of a \emph{forward chaining} deduction because it starts from a set of facts and unifies (matches) them with the body of a rule to derive the consequences. It is also possible to do \emph{backward chaining} in which we start with a goal and work backwards by unifying it with the head of a rule, recursively trying to prove the body. Symbolic ILP systems have a rich history dating back decades. A common approach to inducing rules is called \emph{learning from entailment} \cite{DeRaedt:1997:LSC:270588.270596}, in which hypothesized rules are combined with background facts and trained to entail the positive and none of the negative concept examples. The FOIL algorithm \cite{reason:QuiCam95a} is an example of this approach. Our approach is also of this sort though we use neural networks to learn the rules. The classic ILP setting has been continually updated to handle richer knowledge. In \cite{de2008probabilistic} for example they provide several formulations of probabilistic ILP. We do not consider probabilistic interpretations in our approach, though that is an interesting avenue for future research. A related branch of research called Abductive Logic Programming attempts to learn consistent explanatory facts as well as rules \cite{kakas1992abductive}. For example, it might allow us to induce the fact that eagles are birds from the facts that eagles have wings and feathers together with the inheritance rule. Our approach for theory acquisition may be considered an example of this line of work. \subsection{Neuro-Symbolic Integration} Symbolic ILP systems do very well at generalizing from just a few examples. This is because they learn universal rules. They are, however, susceptible to noisy inputs and even a single bad fact can cause them to fail. On the other hand, neural systems generally are very robust to noisy input but are sample inefficient and prone to over-fitting on small amounts of data. Neuro-symbolic systems aim for the best of both worlds. They can be made robust to noisy inputs while still retaining some of the strong generalization properties typically associated with symbolic systems. There is a long history of research in neural-symbolic systems from which we choose just a small set to present here. For a recent survey see \cite{DBLP:journals/corr/abs-1711-03902}. In \cite{serafini2016logic} they introduce Logic Tensor Networks (LTN) and Real Logic whose semantics grounds the terms, atoms and clauses of the language as continuous functions. They demonstrate that the logic can be implemented using neural networks for the groundings of the symbols and apply it to solving a database completion problem. A follow-up paper applies LTN's to semantic image segmentation \cite{donadello2017logic}. Neither work considers the problems of rule induction or theory acquisition. A recent work \cite{manhaeve2018deepproblog} starts from a probabilistic logic programming language (problog) and extends it to handle neural predicates which compute probabilities. Like Prolog and ProbLog, DeepProbLog is a backward chaining approach. It leverages the automatic differentiation system of ProbLog to incorporate neural predicates and trains with gradient descent. Like Logic Tensor Networks, ProbLog can train neural network implementations of relations. ProbLog does not do rule induction. In Neural LP \cite{yang2017differentiable}, the system can learn chaining-type rules. It uses a neural controller built on top of TensorLog \cite{DBLP:journals/corr/Cohen16b} and is trained to learn rules to compute a ranked list of entities which satisfy a partially specified query. It differs from our approach in several respects. It requires a partially specified query. It represents predicates as TensorLog operators (matrices) whereas we represent them as parameter embeddings which can be associated with constants and a valuation to represent an atom. And it is not obvious how it could be applied to learn fact representations. \cite{sourek2015lifted} uses templates to create grounded networks that depend on the example. \cite{tran2016deep} studies the incorporation and extraction of knowledge into deep networks. \cite{kazemi2017relnn} focuses on predicting the properties of objects. Our approach is most directly related to two recent neural ILP approaches. In \cite{DBLP:journals/corr/abs-1711-04574} inference is done through the forward chained application of a set of logical rules. During learning, a set of all the possible candidate rules is generated according to a provided template. Parameters are weights associated with pairs of candidate rules. These weights are normalized to lie in $[0, 1]$ and interpreted as probabilities associated to the rule pairs as possible definitions of the concept. When there are a large number of rules, this method may suffer scalability issues. In addition it requires a representation of the truth values of all possible facts and non-facts. By contrast, \cite{DBLP:journals/corr/RocktaschelR17} construct a function representing a backward-chained proof of the goal and require only a representation of the true facts. A more conceptual distinction arises in their parameterizations. In \cite{DBLP:journals/corr/abs-1711-04574} the parameters are weights on rule pairs. In \cite{DBLP:journals/corr/RocktaschelR17} they start with a set of parameterized rules which, as in our approach, acquire their meaning as the predicate embeddings of their head and body atoms are trained through unification with the predicates of the facts. It is not obvious how these approaches could be applied to theory acquisition requiring fact induction. In our approach we follow \cite{DBLP:journals/corr/RocktaschelR17} in parameterizing with embeddings but use forward rather than backward chaining so that we don't have to represent a proof tree explicitly. Unlike \cite{DBLP:journals/corr/abs-1711-04574}, we don't have to generate all the candidate rules. Instead, learning is at the level of individual atoms. \subsection{Semantic Cognition} Semantic cognition concerns the acquisition and integration of knowledge. Previous work has modeled semantic cognition as a kind of logical dimensionality reduction \cite{katz2008,ullman2012theory} which uses probabilistic generative models that can simultaneously learn logical rules and a set of core relations that form a theory underlying the observed data. Like ILP approaches, these models can make deductive inferences through the application of logical rules. But unlike traditional ILP algorithms, these Bayesian models are also able to induce facts. This ability to apply both inductive and deductive reasoning at the level of both facts and rules provides humans with a rich space of techniques with which to tame the complexity of everyday experience. These approaches illuminated a promising direction but were severely challenged by scalability issues. With the success of neural techniques we believe it is useful to revisit these ideas. \subsubsection{Limitations} As in previous work, our model is provided with rules that conform to templates, ideally this should not be necessary. The network needs to consider the set of all possible facts when doing core fact induction (which is not necessary for the rule induction problems), this is not be scalable in practice. Forward chaining grows exponentially in the number of facts considered at each step, this can also present a problem when scaling to bigger datasets. From a cognitive science perspective, the model is still more limited than its Bayesian symbolic counterparts, specifically, while those models provide graded measures of confidence in their inferences, our neural logical reasoner does not currently provide meaningful consistent estimates of uncertainty. \subsubsection{Future Directions} One straightforward attempt of learning the structural information of the rules provided by the templates (arity and variable order) would be to encode it by adding dimensions to the embeddings and have the algorithm interpret them by using independent unifications in the desired way. This would constitute a slightly more complicated learning task but would maintain the same structure and mechanism that could be trained through gradient descent. It would also be interesting to explore richer sampling procedures and the integration of forward with backward chaining, this could perhaps yield regimes more similar to those of humans and could help scale to larger datasets. We would also like to investigate ways of providing better estimates of uncertainty -- from a full neural probabilistic formulation, to a heuristic metric based on the number of initializations and on the unification scores. \subsection{Rule Learning ILP Tasks} We tested the model in the ILP problems from \cite{DBLP:journals/corr/abs-1711-04574}\footnote{We only skip the Husband and Uncle tasks which require the datasets from \cite{wang2015soft}.}. The task of an ILP problem is to learn a target relation given a set of background knowledge facts $\mathcal{B}$ and a set of positive $\mathcal{P}$ and negative $\mathcal{N}$ examples of a target relation. As an example, consider the task of learning the predicate \textit{even(X)}. The background knowledge is defined using the \textit{zero(X)} and \textit{succ(X,Y)} predicates. $$\mathcal{B}=\{\textit{zero}(0),\textit{succ}(0,1),\textit{succ}(1,2),... ,\textit{succ}(9,10)\}$$ The target positive and negative predicate extensions are: $$\mathcal{P} = \{\textit{target}(0), \textit{target}(2),...,\textit{target}(10) \}$$ $$\mathcal{N} = \{\textit{target}(1),\textit{target}(9)\}$$ An example solution found by the algorithm is: $$target(X)\leftarrow zero(X)$$ $$target(X)\leftarrow target(Y), auxpred(Y,X)$$ $$auxpred(X,Y)\leftarrow succ(X,Z), succ(Z,Y)$$ Where \textit{auxpred} acquires the meaning of \textit{succ2} which is true of $X,Y$ whenever $X+2=Y$. Table \ref{evans_res} gives a performance comparison to \cite{DBLP:journals/corr/abs-1711-04574}. Since universal logical rules are perfectly generalizable, and to facilitate comparison, we use the same evaluation metric: the percentage of runs with different random weight initializations that successfully learn rules to solve the task with less than $1e-4$ mean squared error. To avoid local minima, we explored adding a decaying normal noise to the embeddings. This had a small positive effect in some of the tasks, reported results include the effect of the noise. Details of the problems are given in the Appendix. \begin{table}[!ht] \caption{ILP percentage of successful runs. $|I|$ is the number of intentional predicates.} \label{evans_res} \centering \begin{tabular}{lllll} \toprule \multirow{2}{*}{Task} & \multirow{2}{*}{$|I|$} & \multirow{2}{*}{Recursive} &\multirow{2}{*}{$\partial ILP$} &\multirow{2}{*}{Ours} \\ \\ \midrule Predecessor & 1 & No & 100 & 100\\ Even-Odd & 2 & Yes & 100 & 100\\ Even-succ2 & 2 & Yes & 48.5 & 100\\ Less than & 1 & Yes & 100 & 100\\ Fizz & 3 & Yes & 10 & 10\\ Buzz & 2 & Yes & 35 & 70\\ Member & 1 & Yes & 100 & 100\\ Length & 2 & Yes & 92.5 & 100\\ Son & 2 & No & 100 & 100\\ Grandparent & 2 & No & 96.5 & 100\\ Relatedeness & 1 & No & 100 & 100\\ Father & 1 & No & 100 & 100\\ Undirected Edge & 1 & No & 100 & 100\\ Adjacent to Red & 2 & No & 50.5 & 100\\ Two Children & 2 & No & 95 & 0\\ Graph Colouring & 2 & Yes & 94.5 & 0\\ Connectedness & 1 & Yes & 100 & 100\\ Cyclic & 2 & Yes & 100 & 100\\ \bottomrule \end{tabular} \end{table} We see that our algorithm performs equally or better in most of the tasks. In contrast to theirs, search is not made at the level of rules but at the more compositional level of atoms. In fact, when the embeddings of the dictionary of predicates are fixed as one-hot vectors, our procedure is very similar to theirs: the parameters that form the predicates of the rules can be treated as weights that select the correct predicate. Thus, like in their model, training consists on selecting the weights that make the embeddings look like the right one-hot vectors and the procedure becomes a symbol search, except at the level of the atoms instead of the rules. The more general case where the embeddings are trained and dense, opens the interesting direction to be explored of studying the learned vector embedding semantic space, as has been done for standard NLP tasks \cite{mikolov2013efficient}, which can potentially allow for similarity and analogical reasoning. In our table, we report the result with the trainable embeddings. It is worth noting the failure of our model in the Graph Colouring and in the Two Children tasks. A quick exploration suggests that the global optima has a very sharp neighborhood while the local minima are attractors in most of the space. This is reminiscent of the Terpret problem \cite{terpretproblem,gaunt2016terpret} and the local minima can only be avoided when the random initialization is very close to the correct rules. \begin{table*}[!ht] \hspace*{1.5cm} \begin{threeparttable} \caption{Performance on the COUNTRIES dataset} \label{sebastians} \centering \begin{tabular}{l|c c c|l} \toprule \multirow{2}{*}{Task} & \multicolumn{3}{c|}{Model AUC-PR} & \multirow{2}{*}{Rule examples and confidences} \\ \cmidrule(r){2-4} & NTP & NTP-$\lambda$ & Ours & \\ \midrule \multirow{2}{*}{S1} & \multirow{2}{*}{$90.83 \pm 15.4$ } & \multirow{2}{*}{$100.00 \pm 0.0$} & \multirow{2}{*}{$91.15 \pm 15.4$} & \multirow{2}{*}{0.85 loc(X,Y) $\gets$ loc(X,Z), loc(Z,Y)} \\ &&&&\\ \multirow{2}{*}{S2} & \multirow{2}{*}{$87.40 \pm 11.7$ } & \multirow{2}{*}{$94.04 \pm 0.4$ } & \multirow{2}{*}{$86.87 \pm 3.2$} & \multirow{2}{*}{0.57 loc(X,Y) $\gets$ neighbor(X,Z), loc(Z,Y)} \\ &&&&\\ \multirow{2}{*}{S3} & \multirow{2}{*}{$56.68 \pm 17.6$ } & \multirow{2}{*}{$77.26 \pm 17.0$ } & \multirow{2}{*}{$63.08 \pm 28.2$ } & 0.59 loc(X,Y) \\ &&&&$\gets$ neighbor(X,Z), loc(Z,W), loc(W,X) \\ \bottomrule \end{tabular} \end{threeparttable} \end{table*} Notice that since our approach is differentiable, it is not prone to some of the problems of symbolic ILP, and like in \citeauthor{DBLP:journals/corr/abs-1711-04574}, it can handle ambiguity and noise. \subsubsection{Countries} We are not focused specifically on knowledge base completion but use the COUNTRIES dataset \cite{bouchard2015approximate} to evaluate the scalability of our algorithm, comparing to other neural logical approaches. The dataset contains 272 constants, 2 predicates and 1158 true facts and is designed to explicitly test the logical rule induction and reasoning capabilities of link prediction models. We compare on the 3 tasks described in \cite{DBLP:journals/corr/RocktaschelR17}, requiring reasoning steps of increasing length and difficulty (S1,S2,S3 in table \ref{sebastians}). We report the Area Under the Precision-Recall-curve (AUC-PR) where results are comparable to the previous NTP approach. For completion, we also report $NTP-\lambda$ from the same paper which uses an additional neural link network as an auxiliary loss. Like them, we also show some example rules and a confidence score by taking the minimum similarity between the atoms of the rule and their decoded predicates. To perform the forward chaining during training, at each epoch we randomly sample from a section of the knowledge graph both the targets and a set of facts to form the background knowledge. Like the related work, our model can also suffer from scalability issues, as in forward chaining the size of the facts grows exponentially with the number of steps. Restricting the number of considered facts through sampling was sufficient for the task at consideration but this could show problems when scaling to much bigger datasets, we discuss some future directions in the conclusion. \begin{table*}[t] \hspace*{1.7cm} \begin{threeparttable} \caption{Theory Learning Results. Succ is the percentage of successful initializations; Acc stands for the accuracy of the recovered facts; Const is the number of constants. } \label{theory} \centering \begin{tabular}{ll|ccc|ccc} \toprule &&\multicolumn{3}{c|}{Taxonomy} &\multicolumn{3}{|c}{Family} \\ \midrule &&\# Preds & \# Const & \# Facts &\# Preds & \# Const & \# Facts\\ \cmidrule(r){3-8} \multicolumn{2}{l|}{\multirow{2}{*}{Observed Data} }&\multirow{2}{*}{4}&\multirow{2}{*}{36}&\multirow{2}{*}{145}&\multirow{2}{*}{6}&\multirow{2}{*}{10}&\multirow{2}{*}{30}\\ &&&&&& \\ \multicolumn{2}{l|}{Target Theory}&4&36&40&4&10&28\\ \midrule &&\% Succ & \% Acc & \# Induced Facts &\% Succ & \% Acc & \# Induced Facts\\ \cmidrule(r){3-8} \multicolumn{2}{l|}{\multirow{2}{*}{Algorithm}}&\multirow{2}{*}{70}&\multirow{2}{*}{99}&\multirow{2}{*}{69}&\multirow{2}{*}{100}&\multirow{2}{*}{96}&\multirow{2}{*}{30.8}\\ &&&&&&\\ \bottomrule \end{tabular} \end{threeparttable} \end{table*} \subsection{Learning Theories} We test the capability of our network to compress a set of observations in the form of a theory by learning a set of core facts in addition to the logical rules. We take the two domains considered by \cite{katz2008}: Taxonomy and Kinship. \subsubsection{Taxonomy}\label{taxonomy} A taxonomy is a set of observations structured into a tree where downstream nodes inherit the properties from the nodes above them in the tree. As shown in Figure 1, the tree constitutes a theory formed by two logical rules $IS(X,Y) \leftarrow IS(X,Z), IS(Z,Y)$; $HAS(Z,Y) \leftarrow IS(X,Z), HAS(Z,Y)$ capturing inheritance and by a set of core facts represented with the edges. All observed facts can be recovered by iterative application of the rules (if salmon are fish, and fish have gills, then salmon have gills). From a range of different possible taxonomies we report performance on the bigger original one from \cite{rogersm04}. This data contains 145 facts composed of 4 predicates and 36 constants. The facts can be compressed into a tree structured theory as shown in the Appendix that contains only 40 core facts. As shown in Table \ref{theory}, the algorithm is able to learn the theory in 70\% of the runs, achieving 99\% accuracy and compressing close to the optimal level (average of 69 compared to the optimal of 40). \subsubsection{Kinship} \label{kinshipSec} We also evaluated performance on the difficult kinship theory, which contains 10 constants and 6 observed predicates \textit{mother}, \textit{father}, \textit{daughter}, \textit{wife}, \textit{husband} (see figure in the Appendix). In this case the compression of the theory consists of a set of 4 auxiliary core predicates with 28 facts. The algorithm has to learn the concepts \textit{female}, \textit{male}, \textit{spouse}, \textit{child} which acquire their meaning through their extensions and the 6 logical rules that generate the observations: $$ mother(X,Y)\leftarrow female(X), child(Y,X)$$ $$ father(X,Y)\leftarrow male(X), child(Y,X)$$ $$ daughter(X,Y)\leftarrow female(X), child(X,Y)$$ $$ son(X,Y)\leftarrow male(X), child(X,Y)$$ $$ wife(X,Y)\leftarrow female(X), child(X,Y)$$ $$ husband(X,Y)\leftarrow male(X), child(X,Y)$$ Table 3 shows the statistics for the observed and target compressed data. The algorithm's performance is again quantified as the percentage of initializations where the rules are successfully learned, the accuracy of the recovered data and the number of learned core facts. The algorithm is able to perform this compression and learns a set of new predicates that conform the rules, recovering 96\% of the data correctly (the algorithm sometimes deduces that some facts are true when they aren't because it can induce incorrect core facts). \section{Introduction} Humans are continually acquiring, representing, and reasoning with new facts about the world. To make sense of the vast quantity of information with which we are presented, we must compress, structure and generalize from what we experience. This allows us to quickly understand new concepts and make useful predictions about them. For example, we might represent our knowledge of animals in a taxonomic hierarchy like that shown in Figure \ref{fig:taxonomy_small}. Using such a hierarchy coupled with an inheritance rule that specifies that the attributes of higher nodes are shared by lower ones, we can achieve exponential compression over a representation which just lists the facts. Even more exciting, it allows us to infer a whole range of new facts about an individual simply by observing where it fits in the hierarchy. For example, observing that a Harpy Eagle is a type of Eagle allows us to immediately deduce that a Harpy Eagle can fly and breathe.\footnote{There are some kinds of reasoning which are not easy to do with a taxonomy (for example, handling the exception that penguins are birds but don't fly) but our proposal is not limited to taxonomic representations.} But how can such representations be learned from raw observations? This has been a key problem in semantic knowledge acquisition going back to at least to the 1960's in the work of \citeauthor{collins69} \shortcite{collins69}, with symbolic, Bayesian, and neural approaches proposed \cite{rogersm04,hinton1986learning,katz2008}. In our view (and following \citeauthor{katz2008} \citeyear{katz2008}) there are three questions to be addressed in the development of a solution: (1) how can we induce logical rules from the observations? (2) how can we learn a small set of core facts (the taxonomy in the example) from which we can infer the observations (and more), and (3) how can this be done without explicit supervision on the structure of the rules? \begin{figure} \centering \includegraphics[width=0.48\textwidth]{taxonomy.png} \caption{Animal Taxonomy. Constants are in red and blue, relations are indicated with lines and arrows.} \label{fig:taxonomy_small} \end{figure} In this paper we propose a model which can be used for both Inductive Logic Programming (ILP) and theory acquisition/compression. The network is neuro-symbolic in the sense that it represents predicates for both rules and facts using dense vectors which can be trained using gradient descent towards representations of known predicates (including a fixed set of anonymous invented predicates). The network implements forward chaining and soft unification. \footnote{Soft unification relaxes the requirement that two predicate symbols must be identical for the rule to be applicable instead favoring a measure of the degree of similarity.} For ILP problems, the network is given a set of ``proto-rules" (rules with randomly initialized predicate parameters) and applies them using forward chaining to the background facts to produce consequent facts. After $K$ steps of forward chaining ($K$ is a hyper-parameter) the consequent facts are compared to the labeled target facts and the rule predicate parameters are trained towards representations of the predicates which yield all the true target facts and none of the false ones. Representations can be learned either for the known predicates used in the facts or for auxiliary invented predicates. In theory acquisition/compression, the network is given a set of fact observations and asked to learn a logical theory -- a set of core facts and a set of rules -- which together entail the observations. The ability to learn facts is an aspect that has not been emphasized in many ILP approaches but is present in the Bayesian literature. For example, when observing that salmon can swim, have fins and have gills, the model can learn the core fact that salmon are fish even though that is not deducible directly. By encouraging sparsity in the set of core learned facts with a penalty term in the loss, the model can be trained to try to minimize the size of the theory it learns. The remainder of the paper is structured as follows. In the next section we provide background material and related work. Then we present our model and training procedure. In Section 4 we present experiments which investigate different capabilities of the model and demonstrate the efficacy of our approach on a variety of ILP rule induction and domain theory learning datasets. We conclude with a discussion of limitations and future directions. \vspace{.3cm} \section{Background and Related Work} \input{background-andres} \begin{figure*}[t] \centering \includegraphics[scale=0.33]{Aldov3.png} \caption{Overview of the model, a step of forward chaining. Parameters are represented in green and constitute the trainable embeddings, orange arrows indicate paths on which gradients flow (in the opposite direction).} \label{fig:model} \end{figure*} \section{The Model} \input{logical-compressor-tim} \section{Experiments} \input{experiments-all} \section{Conclusions and Future Work} \input{conclusion}

709ab28c6a0ca4056370566c942b5922b27548de

\section{Introduction} \label{sec:intro} The need for highly compressed images continues to increase. A rapidly increasing portion of internet traffic is dominated by requests from mobile devices with limited and often metered bandwidth constraints. Efficient delivery of quality thumbnails is an active area of interest to some of the largest Internet companies, including Google~\cite{webp}, Facebook~\cite{cabral2015} and Apple~\cite{apple}. In addition to decreased download latency and bandwidth for end users, reducing image size also helps with storage requirements for billions of thumbnails that need to be rapidly accessed. JPEG has long been a standard approach for image compression. In this study, we examine compression in an operating regime where JPEG and other popular approaches do not fare well: under 200 bytes. Usually, when extreme compression is required, it is addressed with domain specific techniques, specialized for faces~\cite{bryt2008compression}, satellite imagery~\cite{huang2011satellite}, smooth synthetic images~\cite{orzan2013diffusion}, and surveillance~\cite{zhu2015dictionary}, among others. A powerful, recent, image compression approach is \emph{WebP}. Per~\cite{webp}, WebP lossless images are 26\% smaller in size compared to PNGs. WebP lossy images are 25-34\% smaller than comparable JPEG images at an equivalent SSIM quality index. This is the standard to which we will compare. The basis of many of the popular compression techniques is a subdivision of the image into a set of blocks. Our approach, which is based on triangulation, does not use a block approach nor a predefined, or uniform, spacing of triangles over the image. Instead, we use a limited set of vertices which are assigned a color index from a small colormap; simple color interpolation between each of the triangle vertices is used to fill in the triangles to create the resultant image. More triangles are devoted to the complex (high entropy) regions. Triangulation has previously been used in a diverse set of approaches for compression, see ~\cite{bougleux2009image,davoine1996fractal,demaret2006image}. Finally, other experimental compression approaches use deep neural networks~\cite{jiang1999image,toderici2015} and diffusion~\cite{schmaltz2009,hoeltgen2018}. \section{Triangulation of Images} \label{sec:algorithm} There are two broad components of our approach. The first is creating an effective triangulation and the second is efficiently encoding the triangulation. The triangulation component can be thought of as two pieces that must interact well: selecting where to place the triangulation's vertices and assigning a single color to each vertex. For transmission efficiency, we would like to minimize the number of vertices and the total number of unique colors. Rather than transmitting the connectivity matrix, we consider only Delaunay Triangulations that can be constructed in both the encoding and decoding stages given only the vertex coordinates. See also~\cite{galperin1983succinct,turan1984succinct}. Our approach follows a generate-and-test paradigm. We begin with a small thumbnail image, $I$ (usually $221 \times 221$). In the simplest version of our algorithm, shown in Figure~\ref{baseline}, we begin with an over-complete set of vertices on a fixed-size grid and prune them to a smaller set until the set and the color information can fit in 200 bytes. The grid-size is our only parameter for adjusting the compression rate. \begin{figure}[t!] \noindent\fbox{\parbox{\dimexpr\linewidth-2\fboxsep-2\fboxrule\relax}{ \begin{packed_enum} \small \item \emph{Insert initial, evenly-spaced candidate coordinate points from a subsampled version of $I$, into a set $P$.} By considering only a fixed subset $P$ of pixels locations for triangle coordinates, the encoding is faster and more compressible. \item \emph{Select a set $C$ of colors.} Cluster the colors in $I$ and select the 8 to 16 representative ones. Each vertex $p \in P$ is assigned a color in $C$. \item \emph{Triangulate.} Given $P$, create a Delaunay Triangulation. \item \emph{Generate $I'$ and test.} With $C$ and $P$, the triangles are filled using bilinear interpolation, yielding image $I'$. The difference between $I$ and $I'$ is tested with either SSIM or PSNR. \item \emph{Find least important vertex $p_x \in P$. } For each $p \in P$, in turn, remove that $p$, re-triangulate and fill triangles to create $I''$. Set $p_x$ to the point that produces the lowest error.\label{repeat} \item \emph {Remove least important vertex, $p_x$.} Continue from \ref{repeat}. \end{packed_enum} \vskip -0.05in }} \vskip -0.05in \caption{Baseline: A Deterministic, Greedy, Approach.} \label {baseline} \vskip -0.1in \end{figure} \normalsize Using only this simple procedure, the final triangulations and resulting images are shown in Figure~\ref{baselineResults}. The most salient observation is that the triangles cluster around the higher entropy regions. Homogeneous regions such as the sky have fewer triangles. Quantitatively, PSNR quality was close to WebP, but did not consistently out-perform WebP. Understanding the deficiencies is key to understanding the design of our improved approach. (1) The colors, based on global image characteristics, are selected once and never adapted. (2) The only allowed change is to remove vertices: they are never added back or moved slightly to find a better combinations of triangles. (3) The Baseline makes a greedy choice for every proposed mutation, limiting the effectiveness of local search once regions of high performance are found. \begin{figure}[b!] \centering \includegraphics[width=0.85\linewidth]{figs/baseline-performance.png} \caption{Results for baseline. From left: Original Image, 8 colors version, triangulation, and result. Note that the triangulation places more triangles in higher entropy regions.} \label{baselineResults} \end{figure} To address the difficulties, we turn to a stochastic variant. An often-used technique to search discrete, non-differentiable, optimization landscapes is with randomized search heuristics such as hill climbing~\cite{muhlenbein1992genetic}, evolutionary algorithms and strategies~\cite{dasgupta2013evolutionary,salimans2017} and simulated annealing~\cite{kirkpatrick1983optimization}. A stochastic hill climbing variant of the Baseline is given in Figure~\ref{stochastic}. At the expense of determinism, this allows us to more thoroughly explore the search space and adapt color settings. There are six possible operations other than the single ``remove vertex'' operator that the Baseline variant employed. Note that 4 of 7 mutation operators (Steps 5(d)-(g) in Figure~\ref{stochastic}) modify colors; no color modification was employed in the Baseline. Second, vertices can be added (even if they had been previously removed) if they are found to improve the score; vertices can also make ``localized'' moves to nearby grid points. Third, instead of always removing vertices until the desired byte-size was reached, the acceptance of a move is based on whether that move improves the quality-vs-size objective function. In contrast to the Baseline, the stochastic variant is initialized with only a small set of vertices $P$. As before, they are all on a pre-specified grid. The search progresses by \emph {mutating} the current solution set and evaluating the result with respect to the quality-vs-size objective function. \begin{figure}[t!] \noindent\fbox{\parbox{\dimexpr\linewidth-2\fboxsep-2\fboxrule\relax}{ \begin {packed_enum} \small \item \emph{Initialize $P$.} Greedily select $P$ as a set of 300 vertices that minimizes the difference with the input image. \item \emph{Initialize $C$.} Agglomerate the individual vertex colors down to 8 colors in the color table $C$. \item \emph{Initialize vertex colors.} For each vertex, assign the color index in $C$ closest to the vertex's color in the input image. \item \emph{Triangulate.} Given $P$, create a Delaunay triangulation. \item \emph{Mutate.} A mutation is a subset of the actions below. An action is included (once) with some probability. \begin {packed_enum} \item \emph{Displace a vertex.} Move a vertex either horizontally or vertically one grid point. \item \emph{Add a random vertex.} \item \emph{Remove a random vertex.} \item \emph{Re-assign vertex color randomly.} \item \emph{Add a color to $C$ and re-assign vertex colors.} \item \emph{Remove a color from $C$ and re-assign vertex colors.} \item \emph{Perturb a color entry.} Select a color entry and a channel and randomly change by ${\pm 1}$. \end {packed_enum} \item \emph{Re-triangulate and retain the mutation if it improved the trade-off between error and size. Otherwise discard the mutation.} Repeat from Step 5. \end {packed_enum} \vskip -0.05in }} \vskip -0.05in \caption{A Stochastic Approach. } \label {stochastic} \vskip -0.1in \end{figure} \begin{figure*}[t] \centering \begin{minipage}{1.2in} \caption{\small Visualizing the results vs. image byte size. For each of the 3 original images (left most column), the image at 2 WebP (roughly 400, 100 bytes) and 5 compression levels for our system (roughly 400, 300, 250, 200, 100) are given. Byte size (b) and grid size (g) given for each result from our system. } \label{mainimages} \end{minipage}~~~~~ \begin{minipage}{5.5in} ~~~original~~~~~~~~~~~~~~~~~~~~~WebP ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ our system \input {figs/visualization.txt} \end{minipage} \begin{minipage}{5in} \end{minipage} \vskip -0.1in \end{figure*} \subsection {Encoding the Triangulation} Once we have selected a triangulation, we need to losslessly compress that representation. In all cases discussed below, we use asymmetric numerical systems (ANS)~\cite{Duda2015} which is a computationally efficient method to achieve the compression rates of arithmetic encoding~\cite{mackay2003}. We send a ``header'' containing basic parameters such as the grid size, the number of vertices, and the color table. The color-table entries are sorted by frequency of use (starting with the most common). The color table is then coded by giving the number of colors, the color channel values (quantized to 6 bits per YCoCg channel), and the usage frequencies for each of the entries. We code the color channel values as a correction from the average of the previously transmitted values for that channel (e.g., the Y channel), with the first entry's prediction set to the mid-point gray. The histogram table for these color differences are sent as part of the header. The color-entry frequencies are transmitted using a binomial distribution model: we use a ``fair selection'' model on the remaining colors and vertices (i.e., $N = V_{C_r}$ trials, where $V_{C_r}$ is the number of remaining vertices not counted by the previous color entries, and a success probability of $p = 1/|C_r|$ where $|C_r|$ is the number of remaining color-table entries, $C_r$). We improve on this probability model by limiting the considered values of the distribution to be at least $V_{C_r}/|C_r|$ (based on knowing that the {\em upcoming} entries have been sorted) and at most the frequency count of the previous entry (based on knowing that the {\em previous} and {\em current} entries have been sorted). Using this approach to color-entry frequency coding, we save 1\%-3\% of the file size, on average, compared with using a simple uniform-probability model of the distributions. We also explore alternative approaches to compressing the vertex locations and their color assignments. For the vertex locations, the simplest alternative is to send one bit (without arithmetic coding) per grid point to mark whether it is used as a vertex. This takes $N_g$ bits, where $N_g$ is the number of grid points and corresponds to using a $0.5$ probability of occupancy. We can do better by using a fixed, but more accurate, probability of occupancy: $V_t / N_g$ where $V_t$ is the total number of vertices (sent in the header). In our experiments, this fixed-probability only saved an average of $1\%$ of the single-bit-per-grid-point approach. We also tried using run-length codes (coding the run length of ``unoccupied'' grid points between vertices), however the overhead of sending the distribution of run-lengths made this approach worse. The best compression in our tests uses an adaptive-probability approach to compressing the occupancy map. In this approach, we simply update our probability model as we progress through the grid, so that at each point the model probability is $V_r / N_r$ where $V_r$ and $N_r$ are the remaining vertex and grid point counts, respectively. In our tests, this gave a savings of $2.25\%$ over the single-bit-per-grid-point approach. We use a similar approach for coding the color-table index for each vertex. Since we know the color-entry frequencies, we can use adaptive models for these indices. Instead of simply using these remaining-count probabilities, we can explicitly treat the color index coding as a chain of Boolean encodings --- each with the probability indicated by the remaining-count probability for the corresponding color-table entry but with their order of encoding set by a spatially adaptive prediction. This helps even without changing any of the model probabilities because once we see a ``true'' Boolean value, we can stop encoding for that vertex and move to the next one. We determine the order of encoding by sorting the previously seen (and already transmitted) colors based on their Manhattan distance to the current vertex, with ties broken in favor of the more probable color. Using this ordering results in a $1\%-3\%$ file size reduction for the color-table frequencies. Using the spatially-adaptive ordering provides an additional $0.66\%$ file-size reduction. \section{Experiments} \label{sec:experiments} In this section, we present a summary of the experiments we performed. The performance is assessed with \emph{PSNR} (Peak-Signal-to-Noise Ratio) and \emph{SSIM} (Structural Similarity Index)~\cite{wang2004image}, which is based on the visible structures in the image, and is considered a perceptual metric. All of the results reported are the average of compressing 1,024 images, each consisting of $221 \times 221$ pixels. The images were randomly selected from the ImageNet training set~\cite{deng2014imagenet}. Figure~\ref{mainimages} shows the effects of final byte-size on the results obtained by the stochastic variant described in the previous section. The larger the allowed byte-size, the finer the initial grid can be. Our results are shown for grid sizes from $15 \times 15$ (around 100 bytes, compressed) up to $96 \times 96$ (around 400 bytes). Quantitative results are shown in Figure~\ref{mainresult}. When measured with either metric, PSNR or SSIM, the triangulation approach outperforms both WebP and JPEG. As can be seen, JPEG performs quite poorly in this operating range. Also note that to be as favorable to JPEG as possible, we set quality=20, which produced the best quality/size, and used headerless JPEG encoding. Adding a header significantly deteriorates performance. WebP is a close competitor at quality=10. At the range of interest (200 bytes) we outperform WebP in both metrics. By 400 bytes, WebP and the triangulation approach perform equally. Beyond 400 bytes, we again expect WebP to have an advantage. \begin{figure} \centering \includegraphics[width=\linewidth,height=1.5in]{figs/psnr-comparison.png} \includegraphics[width=\linewidth,height=1.5in]{figs/ssim-comparison.png} \caption{Measuring quality as a function of bytes. WebP becomes competitive at approximately 400 bytes, JPEG is competitive after 500 (not shown).} \label{mainresult} \vskip -0.2in \end{figure} Despite the promising results in Figure~\ref{mainresult}, we need to ensure that the average performance is indicative of expected performance. In Figure~\ref{hists}, we provide histograms of the PSNR and SSIM errors on the 1,024 image test set. In the same figure, we also look at the best and worst performing examples: intuitively, the worst cases have large entropy regions (similar to checkerboard patterns) while the best ones have large areas of similar colors. \begin{figure} \centering \includegraphics[width=0.495\linewidth]{figs/psnr-dist.png} \includegraphics[width=0.495\linewidth]{figs/ssim-dist.png} ~\\ \footnotesize \rotatebox{90}{~~~~~~$\approx 15$} \includegraphics[width=0.151\linewidth]{figs/in-depth/originals/0047.jpg} \includegraphics[width=0.151\linewidth]{figs/in-depth/compressed/0047_dec.png} \includegraphics[width=0.151\linewidth]{figs/in-depth/originals/0753.jpg} \includegraphics[width=0.151\linewidth]{figs/in-depth/compressed/0753_dec.png} \includegraphics[width=0.151\linewidth]{figs/in-depth/originals/0354.jpg} \includegraphics[width=0.151\linewidth]{figs/in-depth/compressed/0354_dec.png} \footnotesize \rotatebox{90}{$psnr\approx 35$} \includegraphics[width=0.151\linewidth]{figs/in-depth/originals/0054.jpg} \includegraphics[width=0.151\linewidth]{figs/in-depth/compressed/0054_dec.png} \includegraphics[width=0.151\linewidth]{figs/in-depth/originals/0188.jpg} \includegraphics[width=0.151\linewidth]{figs/in-depth/compressed/0188_dec.png} \includegraphics[width=0.151\linewidth]{figs/in-depth/originals/0444.jpg} \includegraphics[width=0.151\linewidth]{figs/in-depth/compressed/0444_dec.png} \caption{\small Histogram of Results (at bytes=200). Top Left: PSNR, Top Right: SSIM. 3 Cases with poor (middle row) and high PSNR (last row). Shown alternating: original, compressed. } \label {hists} \vskip -0.1in \end{figure} Finally, we would like to give a better understanding at the best steps in our pipeline. In the interest of space, we present one graph in Figure~\ref{ablative} that provides the most insight into the benefits of the stochastic approach over the baseline. We repeat all of our experiments using only a subset of the operators described in Figure~\ref{stochastic} Step 5. The biggest improvement is from vertex modfication and, next, from color-entry perturbation (PSNR shown, same for SSIM). The gains seen with color modification suggest that pre-computing the color table based on the color clusters in the thumbnail is not adequate for use in a triangulated approximation. \begin{figure}[b!] \vskip -0.1in \centering \includegraphics[width=0.9\linewidth]{figs/ablative2.png} \vskip -0.1in \caption{\small What makes the algorithm (Figure~\ref{stochastic}) work? Green: no search, only initialization. Blue: using only vertex-modification operators (steps~5 (a)-(c)). Red: adding color perturbation (steps~5 (a)-(c) plus (d) and~(g)). Yellow: all operators (steps~5 (a)-(g)). } \label{ablative} \end{figure} \section {Conclusions \& Future Work} We have presented an approach to compressing images to extremely small byte sizes. In the operating range of interest, 200-400 bytes, standard JPEG operates poorly. With tiny thumbnails under 400 bytes, we surpass the latest deployed version WebP, version 1.0.0, in both PSNR and SSIM. Only after $\sim$400 bytes, does WebP do better. Further, our representation is scale-free: the triangles can be easily scaled and simply rendered with low-level primitives. There are three immediate avenues for future work. The first is the large-scale human evaluation of the images to verify the SSIM and PSNR improvements. Second, the stochastic search process can be computationally expensive; domain-specific heuristics to narrow the search space may help. Third, we have found that adding synthetic noise in post-processing can enhance the perceived quality without always improving the quality/size score. Further research on a perceptual metric dedicated to thumbnails is needed.

fe2391971b0a1a360f19d87a24e3e0e779521ac1

\section{Introduction} The anomalous precession of Mercury's perihelion of approximately $43$ arc-seconds per century is one of the classic tests of the General Theory of Relativity, and a very significant support for the theory. This short note lays out a simple, exact series representation for calculating the value of this precession for a small mass orbiting a spherically symmetric body. We are here concerned with the relativistic effects only, and not the precession due to other orbiting masses \citep{Stewart}, solar oblateness \citep{xu} or any other effects \citep{Lo}. \section{Analytic calculation} \label{part3} In order to evaluate the angular change for a satellite going from perihelion to aphelion we begin with the Schwarzschild metric for the equatorial plane. Using the (-,+,+,+) convention, the corresponding line element equation is \begin{eqnarray} ds^2=-c^2 d\tau^2=-c^2 dt^2(1-r_s/r)+\frac{dr^2}{1-r_s/r}+r^2 d\phi^2\nonumber \end{eqnarray} where the central mass has Schwarzschild radius $r_s=2MG/c^2$. Parameterizing by proper time $\tau$, and indicating derivatives with respect to $\tau$ using dot notation, this may be written \begin{eqnarray} \dot s^2=-c^2=-c^2 \dot t^2(1-r_s/r)+\frac{\dot r^2}{1-r_s/r}+r^2 \dot\phi^2. \label{eqn1} \end{eqnarray} The equations of motion are derived by solving the Euler-Lagrange equations for extremizing proper distance, $s$. The Lagrangian has no direct dependence on $t$ or $\phi$ so the first integrals for each of these can be written immediately, giving the following two equations: \begin{eqnarray} \dot t&=&\frac{E}{1-rs/r}\nonumber\\ \dot \phi&=&\frac{c L}{r^2} \label{eqn2} \end{eqnarray} for constants of integration $E$ and $L$. Rather than solve the Euler-Lagrange equation for $r$, it is easier to substitute these equations for $\dot t$ and $\dot \phi$ into equation (\ref{eqn1}) and solve for $\dot r$. Dividing $\dot r$ by $\dot \phi$ (equation (\ref{eqn2})), gives a first order differential equation for the path of the planet: \begin{eqnarray} \frac{dr}{d\phi}=\pm r^2\sqrt{\frac{E^2-1}{L^2}+\frac{r_s}{L^2 r}-\frac{1}{r^2}(1-\frac{r_s}{r})}.\nonumber \end{eqnarray} The constants $E$ and $L$ are related to the energy and angular momentum of the satellite. However, for ease of calculations here, it is helpful to define new constants to bring the form of the equation to \begin{eqnarray} \frac{dr}{d\phi}=\pm r^2\sqrt{C_1+C_2 \frac{1}{r}-\frac{1}{r^2}(1-\frac{r_s}{r})}, \label{eqn4} \end{eqnarray} in which the new constants are $C_1=\frac{E^2-1}{L^2}$; $C_2=\frac{r_s}{L^2}$. Note that this differs from the Newtonian path equation $\frac{dr}{d\phi}=\pm r^2\sqrt{C_1+C_2 \frac{1}{r}-\frac{1}{r^2}}$ in the last term only. Next, these constants $C_1$ and $C_2$ may be determined in terms of known parameters. These are the point of closest approach to the sun (perihelion), denoted $r_p$, and the point furthest from the sun (aphelion), $r_a$. At these points, the radius $r$ is a minimum or maximum, so $\frac{dr}{d\phi}=0$. The equation for $r=r_p$ is then: \begin{eqnarray} \frac{dr}{d\phi}\vline_{(r=r_p)} =0=\pm r_p^2\sqrt{C_1+C_2 \frac{1}{r_p}-\frac{1}{r_p^2}(1-\frac{r_s}{r_p})},\nonumber \end{eqnarray} which may be solved for $C_1$. Replacing $C_1$ in equation (\ref{eqn4}) with this solution, we may then take this new equation and put $r=r_a$, which is the other turning point, so $\frac{dr}{d\phi}=0$ there also. Solving this for $C_2$ gives the path equation for the planet, fully specified in terms of the mass of the sun (as $r_s=2MG/c^2$), and the distances from the sun at perihelion and aphelion: \begin{eqnarray} \frac{dr}{d\phi}=\pm r^2\sqrt{\frac{1}{r_p}-\frac{1}{r}}\sqrt{\frac{1}{r}-\frac{1}{r_a}}\sqrt{1-r_s(\frac{1}{r}+\frac{1}{r_a}+\frac{1}{r_p})}.\nonumber \end{eqnarray} The `$\pm$' sign relates only to the direction of travel about the sun, which is unimportant for this calculation, being concerned only with the magnitude of the perihelion shift. As the planet moves from perihelion $r_p$, to aphelion $r_a$, the change in $\phi$ may be written as \begin{eqnarray} \triangle\phi=\int {d\phi}=\int_{r_p}^{r_a}{\frac{d\phi}{dr}dr}=\int_{r_p}^{r_a}{\frac{1}{dr/d\phi}dr}=\int_{r_p}^{r_a}{\frac{dr}{r^2\sqrt{\frac{1}{r_p}-\frac{1}{r}}\sqrt{\frac{1}{r}-\frac{1}{r_a}}\sqrt{1-r_s(\frac{1}{r}+\frac{1}{r_a}+\frac{1}{r_p})}}}\nonumber \end{eqnarray} This integrand is clearly singular at $r=r_a$ and $r=r_p$, but this may be resolved by two simple substitutions. First we let $r=\nicefrac{1}{u}$ and introduce the constants $u_a=\nicefrac{1}{r_a}$ and $u_p=\nicefrac{1}{r_p}$. In the resulting integrand, completing the square suggests the substitution $(u-\frac{u_a+u_p}{2})=(\frac{u_p-u_a}{2})\sin \theta$. Making the substitution and changing the limits accordingly we obtain \begin{eqnarray} \triangle\phi&=&\int_{-\pi/2}^{\pi/2}{\frac{d \theta}{\sqrt{1-r_s[\frac{u_p-u_a}{2}\sin \theta+3\frac{u_a+u_p}{2}]}}}.\nonumber \end{eqnarray} At this point, the integrand is made more readable by introducing the standard orbital parameters of eccentricity $e$, semi-major axis $a$, and semilatus rectum $p$, related to radii at perihelion and aphelion by $r_p=a(1-e)$, $r_a=a(1+e)$ and $p=a(1-e^2)$. Incidentally, we note that using these identities, we can work backwards to identify the original constants $E$ and $L$ by \begin{eqnarray} E^2&=&1-\frac{r_s}{2a}\frac{1-2r_s/p}{1-r_s(3+e^2)/(2p)}\nonumber\\ L^2&=&\frac{r_s p/2}{1-r_s(3+e^2)/(2p)},\nonumber \end{eqnarray} fully specifying the equations of motion in terms of the orbital parameters, as found in standard texts (for example, \citet{pois}, pp.272-274). In terms of these new constants, $\nicefrac{1}{2} (u_a+u_p)=\nicefrac{1}{p}$ and $\nicefrac{1}{2} (u_p-u_a)=\nicefrac{e}{p}$ so the integral becomes \begin{eqnarray} \triangle\phi&=&\int_{-\pi/2}^{\pi/2}{\frac{d \theta}{\sqrt{1-\frac{3r_s}{p}-\frac{r_s e}{p}\sin \theta}}}\nonumber\\ &=&\frac{1}{\sqrt{1-3r_s/p}}\int_{-\pi/2}^{\pi/2}{\frac{d \theta}{\sqrt{1-\frac{r_s e}{p}\frac{\sin \theta}{1-3 r_s/p}}}},\nonumber \end{eqnarray} where the factor in front of the integral is the precession in the small eccentricity limit (see \citet{pois}, p.275). We now have an integral which cannot be evaluated in terms of elementary functions (although it is a complete elliptic integral, \cite{abram}, p.589). However, this integral can be evaluated to arbitrary precision, by using for example, Gaussian quadrature. Alternatively, we can obtain a power series representation using the binomial expansion $(1-x)^{-1/2}=1+\frac{1}{2}x+\frac{3}{8}x^2+\frac{5}{16}x^3+\frac{35}{128}x^4+...$ Defining a new constant $\beta=\frac{r_s e}{p-3 r_s}$, this integral may be written as the series \begin{eqnarray} \triangle\phi&=&\frac{1}{\sqrt{1-3r_s/p}}\int_{-\pi/2}^{\pi/2}{\bigg[1+\frac{1}{2}\beta \sin \theta+\frac{3}{8}\beta^2 \sin^2 \theta+... \bigg]d\theta}.\nonumber \end{eqnarray} Each term in this series may be integrated analytically, with all the odd powers of $\sin\theta$ integrating to zero over the interval. After evaluating the integrals of $\sin^{2n}(\theta)$, the total change in azimuthal angle $\phi$ in going from perihelion to aphelion is thus given by \begin{eqnarray} \triangle\phi&=&\frac{\pi}{\sqrt{1-3r_s/p}} \bigg[1+\sum_{n=1}^{\infty}\frac{(4n)!}{n!^2(2n)! 2^{6n}} \beta^{2n}\bigg]. \label{eqn7} \end{eqnarray} The Newtonian value and Einstein's approximate correction are given by the expansion of equation (\ref{eqn7}) in powers of $r_s/p$ to give $\triangle\phi_0=\pi$ and $\triangle\phi_{1}=\frac{3 r_s \pi}{2 p}$ as given, for example, by \cite{chandra}, p.108. Note that this is for a half revolution, so multiplying by two gives the precession in radians per revolution. The approximate Schwarzschild radius of the sun is $r_s=\frac{2 G M}{c^2}=2.95$ km. For Mercury, $r_p=46,001,200$ km, $r_a=69,816,900$ km, and there are approximately $415.2$ revolutions per century. Converting from radians to arc-seconds gives the result $\approx 43$ arc-seconds per century. After subtracting the Newtonian value of $\pi$, the standard first order approximation $\frac{3 r_s \pi}{2 p}$ radians per half revolution gives the correct figure to seven significant figures, while the first term of equation (\ref{eqn7}),$\frac{\pi}{\sqrt{1-3r_s/p}}-\pi$ gives nine correct significant figures. Using just the first two terms \begin{eqnarray} \triangle\phi-\pi&\approx&\frac{\pi}{\sqrt{1-3r_s/p}}(1+\frac{3}{16}\beta^2)-\pi \label{eqn8} \end{eqnarray} gives 25 correct significant figures! The size of the constant $\beta$ is primarily a product of two dimensionless numbers: the eccentricity $e$ and the ratio $r_s/p$. In the case of Mercury's orbit, $e\approx0.2$ and $r_s/p\approx5\times10^{-8}$ so $\beta^2\approx10^{-16}$. Thus, each additional term adds approximately sixteen decimal places of accuracy. \section{Conclusion} The calculation of the anomalous precession of Mercury has been presented by means of a straight-forward derivation of the path equation in the Schwarzschild space-time utilizing the Euler-Lagrange equations. The result of this calculation is an integral that can be solved numerically to arbitrary precision. Alternatively, by expanding the integrand, the result can be expressed analytically as a simple, elegant infinite series. In the case of Mercury's orbit, this series converges to the correct value at a rate of approximately sixteen decimal places per term. Thus the simple two-term approximation (\ref{eqn8}) for $\triangle\phi$ gives an accessible and highly accurate formula for use by practitioners. \section*{Acknowledgements} I am indebted to Professor Larry Forbes and Professor Clifford Will for their helpful comments on earlier drafts of this paper.

849ba595e9141cb5dc437afca8955df6a6fde7cc

\section{Introduction} The results by \citet{Melendez09}, showing that the Sun departs from most solar twins in the solar neighbourhood in that the composition of its surface layers is comparatively rich in volatile elements and poor in refractory elements, have led to considerable discussion (for a recent review see \citet{Adibekyan17}). Although the finding, a correlation between the chemical abundances for the Sun relative to the twins with the condensation temperatures of the elements in a gas, has been verified by several independent studies \citep{Ramirez09, Gonzalez10, Adibekyan14, Nissen15, Spina16}, its interpretation is still disputed. The early suggestion by \citet{Melendez09} that it could be due to the formation of terrestrial planets in the proto-planetary nebula before the latter was dumped onto the Sun, polluting its convection zone by dust-cleansed gas, requires either an unusually long-lived nebula or a very rapid retraction of the solar convection zone to the surface (cf \citet{Gustafsson10}). As suggested by \citet{Onehag11} this early retraction might result from the episodic accretion scenario according to \citet{Baraffe10}, if applied to the Sun but not to the twins. The alternative, that is, that the effect is due to Galactic chemical evolution, and migration of the Sun from other galactic regions than those of the twins, was suggested and partly supported by tendencies found by \citet{Adibekyan14}, \citet{Maldonado15}, \citet{Maldonado16}, and \citet{Nissen15} for some correlation of the abundance effects to depend on stellar ages. Still another possibility was offered by \citet{Gustafsson10} who proposed that the pre-solar nebula was cleansed early on from dust by the radiation pressure from hot stars in the neighbourhood. This was supported by the results of \citet{Onehag11} and \citet{Onehag14} who carried out detailed analyses of solar-twins and solar analogues in the rich, old cluster M\,67, and found the cluster stars to have an abundance profile more similar to that of the Sun than most solar twins in the neighbourhood. However, \citet{ Gustafsson18} recently found that the cleansing mechanism proposed was unlikely to be efficient enough to explain the chemical peculiarities of a relatively massive cluster like M\,67. Also, the recent finding of binary stars with one component showing deviating abundances from that of the other, such as 16 Cyg \citep{Tucci14, Nissen17}, XO-2N/XO-2S \citep{Ramirez15, Biazzo15}, WASP 94 A/B \citep{Teske16a}, HD\,133131 A/B \citep{Teske16b}, HAT-P-4 \citep{Saffe17} and the co-moving pair HD240430/HD1240429 \citep{Oh17}, and sometimes with differences correlating with the condensation temperature of the element, suggests that the phenomenon may be local. An interesting fact, first noted by \citet{Gonzalez10}, is that the meteoritic elemental abundances show a similar tendency with the condensation temperature to the solar photospheric abundances, in the sense that the meteorites tend to have an even lower ratio of refractory elements relative to volatiles, than the photosphere. This also suggests that the phenomenon should at least partially be local. The issue studied in the present paper is whether radiative cleansing of the matter accreted onto the Sun could be provided by the proto-Sun itself, in its accretion flow. For this mechanism to work, the friction forces between the dust-grains and the gas as well as the gravity forces must be balanced by the radiative forces. Since the friction is dependent on the flow, this mechanism requires constraints on the density of the flow and on its infall velocity in order to work. Other constraints will be important, such as those on the luminosity of the proto-Sun and on the extinction in the cloud around it. We shall use elementary considerations to explore these constraints and shall then apply a somewhat more detailed model to quantify the possibilities of the suggested mechanism to explain the observed abundances in the presence of turbulence in the cloud. Our model mimics an initially homogeneous cloud of neutral atomic or molecular gas with dust, and with a light source representing the early Sun in its centre. The aim will be to find out of whether or not the radiative pressure from the Sun is able to prevent the dust, and thus the chemical elements in the dust, to fall onto the star. \section{The model and the results} We consider a specific dust grain, assumed to be spherical with radius $a$ and mass $m$. We set the vector, extending in space from the Sun to the grain, ${\bf r}$, with a unit vector ${\bf e_r}$ in the direction of ${\bf r}$. In interstellar clouds, the mean free path of the molecules is orders of magnitude longer than the size of the dust grains. In this case the equation of motion for a grain with mass $m$, illuminated by the Sun with luminosity $L$ and mass $M$, can be written (see \citet{Draine11}) \begin{equation} \begin{split} m \cdot \frac{{\rm d}^2 {\bf r{\rm (t)}}}{{\rm d}t^2} = \left[\frac{Le^{-\tau}\pi\,a^2\,\overline{q_{pr}}}{4\pi c} - G\,m\,M\right]\frac{\bf e_r}{r(t)^2}- \\ - 2\pi a^2 \,n(t)kT\, {\bf e}_{\bf v,v_g}B(s) \label{eq0} \end{split} .\end{equation} The three terms on the right-hand side (RHS) represent the radiative force, the gravity, and the friction force exerted by the gas on the grain, respectively. The electromagnetic radiation is attenuated by extinction, and the grain is located at an optical depth $\tau$ from the Sun at time $t$. The factor $q_{pr}$ is numerical, and is dependent on the grain composition and shape, and on wavelength. For the relevant dust and stellar light discussed here a mean $\overline{q_{pr}}\equiv q$ of about 1.5 may be adopted, see, however, Fig 1 of \citet{Ferrara91}, indicating that a higher value up to $q=2.5$ could be appropriate for minor graphite grains. Similarly, if the grains have a composite fluffy character, as suggested by \citet{Kimura17} on the basis of captured Local Interstellar Cloud grains, an increase of $q$ by up to a factor of two may be adequate for the more massive particles. The speed of light is $c$ and Newton's constant of gravity is $G$. The dust grain is retarded by a drag force, described by the last term of the right-hand side of Equation (\ref{eq0}). The number density of atoms or of molecules is $n(t)$, the kinetic temperature is $T,$ and $k$ is Boltzmann's constant. Following \citet{Draine11} we may write \begin{equation} B(s)\approx \frac{s}{\alpha}[1+(\alpha\cdot s)^2]^{1/2} + \left(\frac{eU}{kT}\right)^2 {\rm ln}\Lambda \frac{s}{3\sqrt{\pi}/4 + s^3}, \label{eq101} \end{equation} \begin{equation} s \equiv \frac{v_{\rm drift}}{\sqrt{2kT/m_{\rm H}}}, \,\,\,\alpha \equiv 3\sqrt{\pi}/8, \,\,\, \Lambda = \frac{3kT}{2a e |eU|}\left(\frac{kT}{\pi n_e}\right)^{1/2}, \label{eq102} \end{equation} where $v_{\rm drift} {\bf e_{v,v_g}}={\bf v}-{\bf v_{\rm g}}$ is the drift velocity of the grain relative to the gas in the direction ${\bf e_{\rm v,v_g}} $, with ${\bf v} = {\rm d}{\bf r}/{\rm d}t$ and ${\bf v}_g$ being the velocities of the grain and the gas, respectively, and $m_{\rm H}$ the mass of the hydrogen atoms or molecules, whatever species is assumed to dominate the gas. In the expression for $B(s)$ in Equation (\ref{eq101}), the last term represents the Coulomb forces, where $U$ is the grain potential, $e$ the electron charge, and $n_e$ the electron density in the gas. The potential is determined by the electric charge on the grain. Neglecting photoelectric emission, this charge is set by the requirement that the positive ions captured per time unit be balanced by the number of electron captured. For small degrees of ionization one then finds from Equations (9-15) of \citet{Spitzer78} that $eU/kT=-3.76$. One may schematically estimate the photoelectic emission (positive) contribution to the potential, as a result of the interstellar ultraviolet (UV) background. Following \citet{Spitzer78}, p. 144-146 and his Equation (6-20), setting the sticking probability of electrons striking the grain and the absorption efficiency factor equal to 1, and adopting a value of 0.1 for the photoelectric efficiency of radiation with wavelengths in the interval 91-110 nm, and else much smaller, and assuming the optical depth to be small, we obtain a photoelectric contribution to $eU/kT$ of: \begin{equation} \left(\frac{eU}{kT}\right)_{ph} \approx 2\cdot 10^6\left(\frac{\pi m_e}{8 kT}\right)^{1/2}\frac{1}{n_e} - 1. \label{eqnS1} \end{equation} To this we should add a contribution due to the radiation from the early Sun. We take the flux in the relevant UV interval of the present Sun from \citet{Lean87} to be representative, and then multiply the right-hand side of Equation (\ref{eqnS1}) by a factor $(1+2.2\cdot 10^{30}/r^2)$ where the distance to the Sun is $r$ in centimetres. For our neutral gas (with degrees of ionization $n(e)/n(H) =2.72 \cdot 10^{-4}$ at $T=160$\,K from Table 30.1 in \citet{Draine11b}) we find the Coulomb force term to be relatively small for the local contribution from the electron-ion balance. However, we shall see that the photoelectric charging of the grains has some significance. We note that the dependence of the drag force on the drift velocity is linear for small speeds -- in typical cases with $T = 10$\,K the shift to a mainly quadratic dependence occurs around $v_{\rm drift}\sim 400$ m/s which is higher than typical sound speeds in the cool neutral gas but smaller than typical free-fall velocities for gas at distances of several thousand astronomical units (AU) from the Sun. We assume that the gas velocity is not affected notably by the friction of the grains, and is set by \begin{equation} {\bf v_{\rm g}} = - \sqrt{\frac{2 G M}{r}}\times {\bf e_r} + {\bf v_{\rm turb}} \label{eq103} ,\end{equation} where the free-fall velocity in the first term is corrected by the fluctuating turbulent velocity of the gas, ${\bf v_{\rm turb}} = {\bf v_{\rm turb}}({\bf r},t)$. For the gas density we assume spherical symmetry around the Sun and a time-independent situation, that is, \begin{equation} n({\bf r},t) = n(r) = n_0 \cdot \left(\frac {r_0}{r }\right)^\beta \label{eq104} ,\end{equation} where $\beta$ is usually chosen to be $3/2$, corresponding to a stationary flow, or alternatively $2$, as suggested by \citet{Shu77}, and with the density profile of an isothermal sphere. There is observational support for such power-law density profiles for collapsing proto-stellar clouds and low-mass protostars (e.g. \citet{Loren83}, \citet{Jorgensen02}, \citet{Persson16}), but they are singular at $r=0$ and invalid close to the star. We have stopped the integration at $r=100$ AU to take the existence of an accretion disk into account. In solving Equation (\ref{eq0}), we alternatively set the initial velocity of the grain to ${\bf v}_{\rm 0} = {\bf 0}$ or ${\bf v}_{\rm 0} = - \sqrt{2GM/r_0}\cdot {\bf e_r}$, that is, the free-fall velocity, at $t = 0$ and $r = r_0$. The factor $r_0$ is set to 1000 AU and alternatively 10,000 AU. The number density of hydrogen atoms/molecules at $r_0$ is calculated to correspond to a given accretion rate $\dot{M}$ at $r_0$, calculated with the assumption of the free-fall speed from zero speed at infinity and a stationary flow which leads to \begin{equation} n_0 = \frac{\dot{M}}{4\pi \sqrt{2GM}r_0^{3/2}\mu m_{\rm H}}, \label{eq105} \end{equation} where $\mu$ is $1+(Y+Z)/X$ with $X$, $Y$ and $Z$ being the mass fractions of hydrogen, helium, and elements heavier than helium, respectively, in the gas. We end the calculation when the dust grain has reached a certain inner radius $r_1$; here, the neglect of a net angular momentum for the system is certainly inadequate, and the dust should in reality have ended up in an accretion disk. We also stop the integration at a time of 6 Myr. Then, in practice dust grains that have not reached the inner radius are pushed out to relatively large distances from the Sun by the radiative forces and are regarded as lost from the system. \subsection{The non-turbulent case} We shall first assume the turbulence velocities to be zero, that is, ${\bf v_{\rm turb}}$ is set equal to ${\bf 0}$ in Equation (\ref{eq103}). In this case the flows of grains and gas will constantly be directed towards the Sun, and the model will remain spherically symmetric. The equation of motion is simplified to \begin{equation} \frac{{\rm d^2} r}{\rm dt^2} =\left[\frac{3}{16 \pi} \frac{ Le^{-\tau}\,q}{c a \rho_d }-G\,M\right] \cdot \frac{1}{ r^2 }- \frac{3}{2} \frac{ \,n(t)kT}{a \rho_d }\cdot B(s) \label{eq1} ,\end{equation} where $\rho_d$ is the mass density in the dust grains. The present discussion is focussed on the conditions in the gas at a considerable distance from the Sun during the last accretion in its formation history, assumed to entail relatively thin gas. Therefore, an approximation made in our estimates below is the assumption of small optical depths; for example, we assume $\tau=0$ in Equations (\ref{eq0}) and (\ref{eq1}). In a subsection below, we explore the adequacy of this assumption. Two necessary conditions for the dust to be accelerated outwards are seen directly from Equation (\ref{eq1}). The radiative force has to be greater than the gravity force on the grain, that is, the large parenthesis on the RHS has to be positive, such that \begin{equation} a \leq \frac{3\,L \,q}{16 \pi \,\rho_d\,c\, G\,M} . \label{eq2} \end{equation} With characteristic numbers for the parameters ($L = 0.5 L_\odot$, $q=1.5$, $\rho_d=1$ g/cm$^3$ , and $M=1 M_\odot$) we find $a_{max} = 4.5\cdot 10^{-5}$ cm. This constraint implies that, with a dust size distribution $N(a) \sim a^{-3.5}$, as is often assumed following \citet{Mathis77}, see also \citet{Casuso10}, and extending from $a=10^{-7}$ to $10^{-4}$ cm$^{-3}$, more than half of the total dust mass in the accreting flows may be pushed out by radiation with this choice of parameters. The second necessary condition for an outward acceleration is that the radiative force in Equation (\ref{eq1}) is stronger than the friction force due to the down-streaming gas. Since $|v_{\rm drift}|>\sqrt{2GM/r}$ for outgoing grains and neglecting the electric forces, we find from Equations (\ref{eq101}), ({\ref{eq102}), and (\ref{eq1}) \begin{equation} \dot{M} \leq \frac{q\mu}{c\sqrt{2G}}\, \cdot L \left( \frac{r}{M}\right)^{1/2}; \,\,\, v_{\rm drift} \gg \sqrt{2kT/m_{\rm H}}. \label{eq33} \end{equation} With characteristic numbers as above, and with $T=10$K, we find from Equation (\ref{eq33}) a maximum accretion rate of 0.050 M$_\odot$/Myr at $r=10^4$ AU or 0.016 M$_\odot$/Myr at $r=10^3$ AU. For small graphite grains with $q \sim 2.5,$ maximum accretion rates that are almost double this latter rate may be possible. If we require a significant contribution to the star by dust-rarefied gas with a mass of at least the mass of the present convection zone (or more, since the convection zone during the first millions of years of the Sun should have been significantly more massive) we have to assume that the last percents of the solar mass were accreted slowly, during a time of several hundred thousand years or more in order to satisfy the condition of Equation (\ref{eq33}). We note that such accretion rates are at least two orders of magnitude smaller than those of the zoom-in simulations of forming accretion disks in giant molecular clouds (GMCs) by \citet{Kuffmeier17}, and that the timescale of our slow accretion extends far beyond that of these authors as well as the timescale of \citet{Shu77}. The relative significance of the radiative pressure and the gravity is not dependent on the distance from the star if the optical depth is small, as is demonstrated by the terms within the parentheses in Equation (\ref{eq1}). The $r^{1/2}$ dependence in Equation (\ref{eq33}) implies, however, that the upper limit on $\dot{M}$, set by the requirement that dust grains should be stopped from falling inwards by radiation pressure, becomes tighter as the dust grain is brought closer to the star (e.g. by initial momentum or by turbulence). This means that if moving inwards in a given free-fall flow of gas, the dust grains may finally come to a radius $r_c$ where they cannot escape from falling into the accretion disk. For the case with high drift speed, we obtain \begin{equation} r_c =\frac{1}{8\mu^2} \dot{M}^2 \cdot GM \cdot \left[\frac{Lq}{4c}-\frac{4\pi}{3} G M a\rho_d\right]^{-2}. \label{eq34} \end{equation} With characteristic parameters, as above, and with $\dot{M} = 0.01$ M$_\odot$/Myr and $a = 2\cdot 10^{-5}$ cm, we find $r_c=1.5\cdot 10^3\,$AU. In practice this means that the dust that separates from the infalling gas will have to decouple from the gas already at distances on the order of 1000 AU from the proto-Sun or further out. It is worth noting that these distances overlap with the location of the Oort cometary cloud. Also, the last factor of the last term in Equation (\ref{eq34}) suggests a possible fractionation of dust of different sizes: with the characteristic parameters chosen above, $r_c$ = 480\,AU for grains with $a= 0.35\cdot 10^{-5}$ cm, while $r_c$ = 13,000 AU for $a=3.5\cdot 10^{-5}$ cm. We have explored the possibilities for dust-grains to exist at various distances from the proto-Sun by setting the LHS equal to zero in Equation (\ref{eq1}) and then solving it for different values of $\dot{M}$, $a$ and $r$. The results for the non-electric case are displayed as full lines in Figure \ref{self_fig1}, where the regions in the diagram below each curve are the `allowed' regions, that is, where dust grains will be pushed out by radiation and not pulled by gravity and friction from the in-falling gas towards the stellar accretion disk. The corresponding curves for the case when electric forces are included (dashed blue lines in Fig. \ref{self_fig1}) clearly show that the critical gas accretion rates, able to sweep dust grains close to the star, are reduced by almost a factor of two, as compared with the case when coulomb forces are neglected. In particular, the results of considering the photo-electric effects on the grains are noteworthy. The effects are particularly important for the larger grains which contribute significantly to the totally accreted mass of dust. \begin{figure} \resizebox{\hsize}{!}{\includegraphics{Self_cleansing_comb1.eps}} \caption{The "allowed" parameter space (under each curve) for dust grains with radius $a$ that may exist around a star at a certain distance (indicated in AU at each curve) without being accreted to the star by gravity and friction from the in-falling gas with different accretion rates $\dot{M}$. The full lines correspond to cases with neutral grains while dashed blue curves indicate corresponding cases with electrically charged grains. The approximate upper limit on $a$, according to Equation (\ref{eq2}), is indicated by a vertical red line, while the approximate upper limit on ${\rm d}M/{\rm d}t \equiv \dot{M}$ for a distance from the star of $10^3$ AU, according to Equation (\ref{eq33}), is indicated by a horizontal red line. Also shown is the approximate critical radius $r_c$ as calculated by Equation (\ref{eq34}) as a dashed red curve. The fair agreement between the approximate (in red) and the more exact (black) curves is due to the fact that the quadratic term in the dependence of the friction force on drift speed, c.f. Equation (\ref{eq101}), is indeed dominating, as assumed in the approximations. The reduction of the "allowed" accretion rates for the charged grains relative to the corresponding cases with neutral ones is obvious.} \label{self_fig1} \end{figure} \subsection{The optical depth effects and the friction forces onto the gas} Above, we have assumed that the medium is transparent to the radiation from the star in the centre, thus assuming that the dust grains close to the star do not shadow the grains behind. In order to take such possible shadowing effects into consideration, we shall calculate the optical depth in the stellar envelope. We then disregard the effect of the dust in the accretion disk, which may well be optically thick in certain directions. The optical depth $\tau$ in the converging accretion flow, assumed to be spherically symmetric around the star, is \begin{equation} \tau = \int _{r_1} ^r {\rm d}r \int _{a_l }^{a_u} N_{\rm V}(a,r,t) \sigma(a) {\rm d}a , \label{eq1531} \end{equation} where $\sigma(a)$, which we here set equal to $\pi a^2$, is the absorption cross-section per grain, and $N_{\rm V}(a,r,t)$ is the number of dust grains with radii in the interval $(a,a+{\rm d}a)$ per volume unit at time $t$. When calculating $N_{\rm V}$ we have solved the equation of motion (Equation (\ref{eq1})) for a range of different $a$ values from $a_l=10^{-7}$ cm to $a_u = 10^{-4}$ cm, distributed in size in proportion to $\sim a^{-3.5}$ according to \citet{Mathis77}, in a convergent and stationary gas flow in free fall (i.e. applying Equation (\ref{eq103}) with ${\bf v_{\rm turb}=0}$ and with $\beta=3/2$ in Equation (\ref{eq104})). The gas-to-dust mass ratio was assumed to be 100, however, noting that there may be considerable variations in this ratio in interstellar clouds (see, e.g. \citet{Reach15}). As initial conditions we have alternatively chosen ${\bf v(r_{\rm 0}) = 0}$ and ${\bf v(r_{\rm 0}) = v_{\rm ff}(r_{\rm 0})}$, with ${\bf v_{\rm ff}}$ denoting the free-fall velocity with a start at rest in infinity, however with little effects on the end results. Typical values of $r_{\rm 0}$ and $r_1$ were 10,000 and 1000 AU, respectively. The individual grain orbits have been integrated for at least 1 Myr each in order to ascertain that a stationary situation has been achieved. When calculating $\tau$ we have used an iterative procedure, first setting $\tau(r) = 0$ when solving Equation (\ref{eq1}), subsequently calculating $\tau(r)$ by Equation (\ref{eq1531}), feeding the result into Equation (\ref{eq1}), and so on. In these calculations, we have neglected the coulomb forces in $B(s$). In practice, with the small resulting $\tau$ values, we have found that three iterations are enough to ascertain a satisfactory convergence.The resulting optical depths for three different accretion rates are shown in Fig. \ref{self_fig2}. As is seen, at these rates the depths are relatively small or even negligible. \begin{figure} \resizebox{\hsize}{!}{\includegraphics{Self_CleansingTau}} \caption{ The optical depth as a function of distance from the star for three different accretion rates, $10^{17}$ g/s, 3 x $10^{17}$ g/s and $10^{18}$ g/s, corresponding to 0.0016, 0.0047 and 0.016 M$_\odot$/Myr, respectively.} \label{self_fig2} \end{figure} \begin{figure} \resizebox{\hsize}{!}{\includegraphics{Self_Cleasning_FigDust1}} \caption{The contributions to the dust opacity for grains of different radii $a$, at different distances from the accreting star indicated at each curve in AU. The opacity contribution is given in cm$^{-1}$, per logarithmic interval in $a$. The dominating contribution from grains on the order of $a=a_{\rm max} $ in size, cf. Equation (\ref{eq2}), is seen in the inner regions, and is enhanced by the convergence of the flow, while at greater distances from the star the more numerous smaller grains contribute more, as a result of their initial motion inwards due to the starting condition $v(r_{\rm out}) = v_{\rm ff}$. The curves have been smoothed by a filter with a width of $5 \times10^{-6}$ cm, to compensate for fluctuations due to the stochastic nature of the method used. } \label{self_fig3} \end{figure} A basic reason for the small optical depths is the fact that the numerous grains with small radii are already pushed out by the stellar radiation at great distances from the star, where the gas flow has still not become strong and dense enough to pull them inwards by friction. The dominating dust opacity is instead contributed by the more massive grains, in particular those that have $a$ close to $a_{\rm max}$ as estimated in Equation (\ref{eq2}). The close balance between radiative forces and gravity makes them fall slower towards the star than the more heavy ones, which they also outnumber due to the $a^{-3.5}$ dependence of the initial number density. The situation is illustrated in Figure \ref{self_fig3}, where the opacity contribution by grains of different sizes and at different distances from the star is illustrated. In addition to neglecting dust opacities in the model calculations, we have neglected the friction forces from the dust onto the gas. Applying Newton’s third law, one easily finds in the high drift-velocity case that the friction force per gas mass, relative to the corresponding gravity force, is \begin{equation} \frac{F_{\rm fric}}{F_{\rm grav}} = \frac{3}{64\sqrt 2\pi} \cdot \frac{{\dot M} v_{\rm drift}^2}{(GM)^{3/2}(a_ua_l)^{1/2} \phi \rho_d \mu} r^{1/2}, \label{eq881} \end{equation} where $\phi$ is the gas-to-dust mass ratio. By setting in representative numbers one finds that for our models this ratio is on the order of $10^{-2}$ or less, which means that the friction forces on the gas can be neglected. It is obvious, however, that for considerably more intensive gas flows, such that ${\dot M} > 0.1 \,M_\odot$ per million years, or high luminosities increasing the radiative forces and thus $v_{\rm drift}$, the friction forces might become significant. \subsection{Conclusions from the spherically symmetric model} It is clear from the results presented above that a slow accretion of gas, with an accretion rate on the order of a few percent of a solar mass per million years, should enable the central star to push most of the grains with radii less than about $5\times 10^{-5}$ cm, corresponding to roughly half the grain mass, out of the accretion gas flow. A considerably higher accretion rate will both increase the friction forces and the optical depth of the envelope, due to relatively large grains, so much that the smaller grains fall inwards, and the cleansing becomes significantly weaker. Thus, a model with a moderate accretion flow seems viable to explain stellar atmospheres depleted in dust-forming elements for stars with convection zones of a few percent of the stellar mass, provided that such relatively thin convection zones exist at times when this mild accretion is still active. According to standard models of PMS evolution (\citet{Montalban08}) the convection zone remains massive, with masses greater than about 0.1 M$_\odot$ until about 20 Myr, while typical accretion gas disks only remain for less than 10 Myr (see, however, \citet{Pfalzner14}). Thus, we may have a timescale problem similar to that discussed for the planet-cleansing hypothesis for the abnormal volatiles/refractories ratio of the Sun \citep{Melendez09, Gustafsson10}. This could possibly be explained as a result of episodic accretion or more long-lived stages of the late accretion disks. Another possible caveat of the present model is, however, turbulence which was found by \citet{Gustafsson18} to be a major problem with the alternative hypothesis of massive dust-cleansing by bright stars as an explanation for the Mel{\'e}ndez effect. The effects of turbulence on the present self-cleansing model are discussed below. \subsection{The turbulent case} \begin{figure} \resizebox{\hsize}{!}{\includegraphics{3dvff.eps}} \caption{The orbits of grains over $10^5$ yr, from a point indicated by an `x' in black along the z axis at a distance of $10^4$ AU from the star (which is at the origin of the coordinate system), and with different grain radii $a$ as indicated by the different colours. The dust grains were given an initial velocity identical to the gas free-fall velocity. The turbulent mean speed was set to 0.1 km/s. We note that the scales of the x and y axes are different from that of the z axis. } \label{self_fig3a} \end{figure} \begin{figure} \resizebox{\hsize}{!}{\includegraphics{3dv00.eps}} \caption{The orbits of grains over $10^5$ yr, from a point indicated by an `x' in white along the z axis at a distance of $10^4$ AU from the star (which is at the origin of the coordinate system), and with different grain radii $a$ as indicated by the different colours. The initial velocity of the dust grains was set to zero. The turbulent mean speed was set to 0.1 km/s. We note that the scales of the x and y axes are different from that of the z axis. } \label{self_fig4} \end{figure} The origins and the consequences of the motions in giant molecular clouds are still discussed intensively. Observationally, however, it seems clear that turbulence is present in star-forming regions as well as quiet clouds. \citet{Solomon87} showed that the CO line widths of the GMCs scaled with the square root of their radii $R$, according to $\sigma_{\rm v} \approx 0.72 (R/{\rm pc})^{0.5}$ km/s, which seems to hold down to small scales (cf \citet{Bolatto08}, \citet{Rosolowsky08}, \citet{Heyer04}). The ideal value of $\sigma_{\rm v}$ to be chosen for a model of the solar birth cloud is therefore, however, highly uncertain. We therefor treat it as a free parameter here, varying it between 0.1 and 5 km/s. To handle the turbulence in the accretion flow of the gas in our model in full detail is beyond the scope of the present paper. Instead, we have followed the schematic recipe of \citet{Higham01}, by re-setting the turbulent velocity, ${\bf v}_{\rm turb}$, when solving Equation (\ref{eq0}) numerically every time the derivatives are to be calculated for this solution. We have used the MATLAB routine ODE23 for this purpose, which is based on a Runge-Kutta method with variable (and continuously adjusted) step size in time. The step size is typically 30 yr, which is less than one thousandth of the free-fall time from 10,000 AU. We thus set \begin{equation} {\bf v}_{\rm turb} = \xi_{\rm x}{\bf e}_{\rm x} + \xi_{\rm y}{\bf e}_{\rm y} + \xi_{\rm z}{\bf e}_{\rm z} \label{eq1541} \end{equation} in Equation (\ref{eq103}), where $\xi_{\rm x,y,z}$ are three independent random numbers drawn for each time step from a Gaussian distribution with a dispersion $\sigma_{\rm v}$. Another effect of turbulence is a fluctuating gas density. In order to study the effects of such fluctuations, we have run some models with the density perturbations guided by the log-normal distribution found from detailed simulations by \citet{Padoan02} (their Equation 6), around values calculated according to Equation (\ref{eq104}), with a characteristic fluctuation time of $10^4$ years. At the solution of Equation ({\ref{eq0}) we have set the optical depth equal to zero, guided by our results in Section 2.2. Subsequently, the solution entails solving three coupled second-order non-linear differential equations. For each run, this has been done for 1000 different dust grains of given radii $a$, initially located at a distance of $r_0=10,000$ AU from the star (along the positive z axis). In subsequent runs, $a$, $\sigma_{\rm v}$, ${\dot M}$, and the initial condition have been varied, one at a time. For the initial velocity of the dust grain at the starting distance we have alternatively used ${\bf v(r_{\rm 0})} = {\bf 0}$ and ${\bf v(r_{\rm 0})}=-\sqrt{2GM/r_0}\cdot {\bf e}_{\rm r}$. The integration proceeds for 5 Myr. If the orbit during this time gets closer to the star than 100 AU, we assume the grain to be caught in the accretion disk around the star, and finally fall into the star. Conversely, if it does not get closer to the star during this time we consider it lost. Grains that belong to this class are usually accelerated to very large distances from the star by the radiative pressure. One should note that in these simulations we totally neglect the outflows from the central disk, known observationally to carry away a considerable fraction of the mass; see \citet{Alexander14} and references therein. In most of these simulations for the turbulent case we have neglected the effects of coulomb forces on the grains. To check their significance, however, we made a series of simulations, taking the charging of the grains into account. With the parameters of these effects set as described above, they are found to have relatively limited effects on the results, the most important one being that due to the photoelectric charging. This is further illustrated below. Two typical runs, though limited to 100 grains for each $a$ value and with integration times of only 0.1 Myr for clarity, are illustrated in Figs. \ref{self_fig3a} and \ref{self_fig4}. The spread of the orbits into broom-like structures is entirely due to the turbulence; if ${\bf v}_{\rm turb}$ is set equal to ${\bf 0}$ the orbits collapse to a single vertical line. It is obvious from these figures that the choice of initial conditions for the velocity is important for the orbits, although the fraction of grains falling into the star as a function of grain radius is not as dependent on these conditions. The main results of these numerical experiments are displayed in Fig. \ref{self_fig5}. Here we have plotted a quantity called `the defect',} $ D(a),$ of the run, as a function of $a$ for different choices of the defining parameters. $D(a)$ measures the fraction of grains that leave the proto-solar cloud as a result of the radiative cleansing. It is defined as \begin{equation} D(a) = 1 - \frac{N_{\rm l}}{N_0}, \label{eq70} \end{equation} where $N_{\rm l}$ is the number of dust grains lost inside 100 AU, and $N_0$ is the corresponding number lost in a run with identical parameters with the exception that the stellar luminosity $L$ is set equal to zero. In this way, we try to correct for the fact that even at zero luminosity some of the gas may depart at high turbulence speeds from the stellar neighbourhood. It is seen from the figure that the variation of $D$ with $a$ is smoothed when $v_{\rm turb}$ increases, from a very steep change when $a$ passes the limiting value in Equation (\ref{eq2}) for small turbulence, to a less steep variation. We also note that an increased inflow of gas decreases the deficit at a given $a$, as does a change of the initial condition from ${\bf v(r_0)}={\bf 0}$ to setting ${\bf v(r_0)}$ equal to the free fall velocity of the gas flow. Also, the coulomb forces diminish $D$ but not drastically so. None of these effects is unexpected. Obviously, these variations of the deficit with the different parameters are not very dramatic: the conclusions concerning the cleansing effects in our model are clearly relatively robust with regard to its uncertain parameters. This may be seen by calculating the integrated deficit,} $\mathcal{D,}$ as \begin{equation} \begin{split} \mathcal{D} \equiv \left. \int_{a_l}^{a_u} D(a) \frac{4 \pi}{3} a^3 \rho_d N_0 a^{-3.5} da \middle/ \int_{a_l}^{a_u} \frac{4 \pi}{3} a^3 \rho_d N_0 a^{-3.5} da \right. \\ = \left. \int_{a_l}^{a_u} D(a) a^{-0.5} da \middle/ \int_{a_l}^{a_u} a^{-0.5} da \right.. \label{eq91} \end{split} \end{equation} $\mathcal{D}$ measures the fraction of the dust mass that is pushed away from the star instead of falling into it. In Table \ref{tab1} we present the values of $\mathcal{D}$ for our different models in Figure \ref{self_fig5}, and some other models. The inclusion of electrical forces and in particular the photo-electric charging obviously leads to a moderate decrease of $\mathcal{D}$. Also, turbulence-generated density variations lead to a not very dramatic shift of $D(a)$ to smaller values (model no. 10 in the table). As seen in the table (model no. 11), an adopted variation of the density as $1/r^2$, still assuming the inflow gas velocity to be roughly proportional to $1/r^{3/2}$ which means that the gas is piling up in the inner region of the gas envelope, leads only to small effects compared with the standard stationary flow models with $\rho \propto r^{-3/2}$. It seems that a rather robust conclusion may be drawn: about half the mass of dust in a late accretion flow towards a solar-type star is cleansed from the flow by the radiative forces of the star itself, as long as the in-falling gas mass rate is not higher than about 0.01 M$_\odot$/Myr. \begin{figure} \resizebox{\hsize}{!}{\includegraphics{Self_cleans_fig6_el}} \caption{The defect, $D(a),$ (cf. Equation (\ref{eq70}) as a function of grain radius $a$ for different values of the turbulence speeds $v_{\rm turb}$ and with initial dust velocities = ${\bf v}_0 = {\bf 0}$ and with infall rate of gas ${\dot M}=0.002$ M$_{\odot}$/Myr shown in thin lines. Red line: Coulomb forces, excluding photo-electric effect, taken into account with an ionization ratio of $n(e)/n(H) = 2.72\cdot 10^{-4}$.Thick line: doubled infall rate, $v_{\rm turb}=1$ km/s. Dashed line: initial speed ${\bf v}_0 =$ initial free-fall velocity of the gas, ${\bf v}_{\rm turb} = 1$ km/s, ${\dot M}=0.002$ M$_{\odot}$/Myr.} \label{self_fig5} \end{figure} \begin{table} \caption{Values of the integrated deficit, $\mathcal{D,}$ for runs of models with turbulence. The parameters of the models are the 1D turbulence parameter $v_{\rm turb}$ in km/s, and the gas-flow rate $\dot{M}$ in M$_\odot$/Myr. The initial condition for the dust grains at distance $r_0=$ 10,000 AU from the star is ${\bf v}_0$ = ${\bf 0}$ except for a model where ${\bf v}_0$ is set to the free-fall velocity ${\bf v}_{\rm ff}$ of the gas at $r_0$. All models were run neglecting the coulomb forces, except two, marked `coulomb' and `+ photo-el'. For these, the coulomb forces, without and with the photoelectric charging of the dust grains, respectively, were included schematically (see text). The model marked $\delta\rho$ in the last column was run with fluctuating density, see text, and the following model with a $\rho\propto r^{-2}$ dependence and $r_0=1000$ AU. The two last lines represent runs where the star was enclosed in a 0.125 pc$^3$ cubic box with six reflecting walls, to simulate the contributions of new grains from surrounding stars, and five reflecting walls, respectively. } \label{tab1} \begin{tabular}{c c c c c l} \hline\hline Model &${\bf v}_{\rm turb} $ & $\dot{M}$ & ${\bf v}_0 $& $\mathcal{D}$ & note \\ & km/s & M$_\odot$/Myr &&\\ \hline \\ $1$&$0$&$0.002$&${\bf 0}$&$0.61$\\ $2$&$1$&0.002&${\bf 0}$&$0.61$\\ $3$&$1$&0.002&${\bf 0}$&$0.60$& coulomb\\ $4$&$1$&0.002&${\bf 0}$&$0.47$& + photo-el.\\ $5$&$2$&0.002&${\bf 0}$&$0.64$\\ \hline $6$&$1$&$0.004$&${\bf 0}$&$0.55$\\ $7$&$1$&$0.010$&${\bf 0}$&$0.41$\\ $8$&$1$&0.002&${\bf v}_{\rm ff}$&$0.54$\\ $9$&$1$&0.002&${\bf 0}$&$0.60$& $r_o = 1000$AU\\ $10$&$1$&0.002&${\bf 0}$&$0.58$& $\delta \rho$\\ $11$&$1$&0.002&${\bf 0}$&$0.60$& $\rho^{-2}, r_o = 1000$AU\\ \hline $12$&$1$&0.002&${\bf 0}$&$0.60$ &closed box\\ $13$&$1$&0.002&${\bf 0}$&$0.61$ & semi-closed box\\ \hline \end{tabular} \end{table} The present model is unrealistic in several respects. One of the most important simplifications relative to reality is the neglect of magnetic fields. The star-forming clouds and the protostars are known to be magnetic (see e.g. \citet{Higuchi18}, \citet{Hull18} and references therein), which may have consequences for the dynamics of the in-falling gas. Also, since the grains may be charged, their dynamics may be directly affected by the fields. We have addressed the question of whether the fields threading the gas are swept along by it by estimating the timescale for ambipolar diffusion of the field out of the gas, following Equation (2.18) of \citet{Hennebelle12a}, which shows that the diffusion time varies as $n\cdot n_i \, (L/B)^2$ where $n$ and $n_i$ are the number densities of atoms and of ions, respectively, $L$ is a characteristic length scale and $B$ is the field strength. For our extended and thin flows ($L \sim 1000 \, {\rm AU}, n \sim 0.3 \, 10^3 {\rm cm}^{-3}$ and for characteristic numbers of the ionization fractions in the cold gas and with $B$ set to a few microGauss (\citet{Padoan18}), diffusion timescales are at least one order of magnitude greater than $10^6$ years. Therefore, the magnetic fields are swept along by the in-falling gas. Our estimates of the magnetic-field energy density $E_B$, as compared with the kinetic energy density $E_K$ of the gas and its gravitational energy density $E_G$, show that $E_B << E_K \sim E_G$, which suggests that the magnetic effects on the gas dynamics are only marginal. As regards the direct effects of the magnetic fields on the dynamics of the grains, they may, however be very considerable. According to \citet{Draine11} the dynamics is determined by the dimensionless quantity $\omega \cdot \tau_{\rm drag}$ where $\omega$ is the gyro-frequency (proportional to the field strength and the charge-to-mass ratio of the grain), and $\tau_{\rm drag}= (m\cdot v)/F_{\rm drag}$ is the drag timescale for the grain with a drag force $F_{\rm drag}$, a mass $m$ and a speed $v$. From Equation (30) of \citet{Draine11}, we find with our characteristic numbers that $\omega \tau_{\rm drag}$ may be typically in the range 1 - 10, which means that the magnetic fields, if oriented at angles relative to the forces on the grain, may significantly affect its motions. In particular, if the magnetic field is perpendicular to the direction from the grain towards the star, with the radiative force on the grain dominating relative to gravity and the gas drag force, the grain will nevertheless be dragged along in the inflow motion of the gas by the magnetic field. Key factors affecting the degree to which magnetic fields may thus hamper the radiative dust cleansing are therefore (1) the number of magnetic field lines in the outer regions of the gas (at $r \sim 10^3$ AU) in late and thin accretion flows that are close to perpendicular to the flow lines and (2) how stationary such regions are, that are locking the grains to the collapsing gas; that is, if the characteristic time for reorientation of the field in the gas due to turbulence is smaller or greater than the free-fall time of the gas towards the star. The study of such questions requires more detailed numerical MHD simulations. The effects of angular momentum are also neglected in the present simple model, which may be relevant since much of the dust-gas separation takes place at relatively great distances from the central star. However, closer to the star the orbits of the grains may be quite different due to their intrinsic angular momentum, and the non-radial drag of the rotating gas with the magnetic field. If the magnetic field drags the dust inwards from large distances, the winding-up of the field lines in the accretion disk may finally decelerate dust as it falls inwards, and outflows at angles with the disk field may even accelerate the grains selectively outwards again. For dust located in the inner part of the disk, the radiative field from the star may be very severely reduced by extinction. In this case, the pressure by infrared radiation from the hot disk may dominate and cause large dust grains to slide above the surface of the disk to larger radii (\citet{Vinkovic09}) while smaller grains are less affected. As suggested by the referee of the present paper, for grains orbiting the star, the Poynting-Robertson drag will tend to selectively decelerate the smaller grains and make them move inward towards the star. Both these effects may therefore favour larger grains at greater stellar distances as compared with smaller grains, which is contrary to the mechanism discussed in the present paper. As highlighted above, the composition and structure of the grains are also expected to affect the $q$ value in Equation (\ref{eq1}); flows with small graphite grains and relatively large fluffy composites may for that reason also be expected to be more effectively radiatively cleansed. \subsection{Stellar fields} Above, we have studied the case of a single star with a relatively mild flow of gas falling in towards it. The integrations of the dust orbits started at considerable distance (10,000 AU) from the star, and the gas flow was assumed to continue in a stationary way for several million years, which implies that regions far outside this distance from the star may also be involved. In reality, such an isolated situation should be rare: stars are known to be formed in associations or clusters, and the effects of the companion stars should be considered. In the case of the Sun, it is often argued that it had its origin in a relatively dense stellar environment on the basis of the abundances of some daughter elements of radioactive supernova products like $^{26}$Al, $^{41}$Ca and $^{60}$Fe. The situation has been summarised by \citet{Adams10} who also included various dynamic arguments in his discussion and concluded that the number of members of the solar birth cluster should fall in the range $N\sim 10^3 - 10^4$, which suggests a cluster radius of about 6 pc (cf. \citet{Joshi16}), and a typical distance to the closest star of about 0.5 pc. We note, however, the recent discussion by \citet{Fujimoto18} who argue that "normal" chemo-galactic evolution with new generations of stars preferentially forming in patches of the Galaxy that were contaminated by previous generations of stars can account for the observed radioisotopes without invoking a nearby supernova very close to the forming Solar system. Nearby stars and star passages may episodically affect the dynamics of the gas flows towards the accreting star. However, such effects on $\mathcal{D}$ are probably not very significant, in view of the relatively small effects of the persistent turbulence found above. However, if the neighbouring stars are assumed to actively cleans their envelopes, similarly to and simultaneously with the corresponding self-cleansing of the Sun, we expect flows of radiatively driven grains from those stars towards the Sun. Therefore, a reduction of $\mathcal{D}$ for the Sun is possible. We explored the effects of the dust from stellar neighbours by surrounding the star in the centre by a box of $d \times d \times d$ pc$^3$ in dimension, and then reflected the velocities of the outgoing particles when they hit the walls of the box. In order to avoid biases, we then shifted the starting point of the grain on the box side randomly (a precaution which was found to be of little significance for the result). The reflection was repeated if the grain hit another box surface until the full integration time of 5 Myr had elapsed. As is seen in Table \ref{tab1}, the effects of this closed box around the star are relatively small: the value of $\mathcal{D}$ is reduced by only one percent, and, in fact, does not significantly depend on the value of $d$. The reason for the small magnitude of these effects is the symmetry in the forces dominating at a given distance from the star, radiative pressure and gravity, which both scale as $1/r^2$, for a dust grain that departs from, or approaches, the star. Only when the grain is much closer (with the mild accretion rates adopted here), does friction introduce an asymmetry, so that the speed of the grain at a given distance from the star becomes different and higher in the approach than in the departure. For radiative cleansing to be significant at all, the accretion rates should not be too high, as is illustrated in Fig. \ref{self_fig1}, and this requirement prevents the asymmetries from becoming larger. The conclusion of these schematic experiments is that the effects on dust in a stellar envelope of the emission of dust from solar-type stars in a dense stellar environment should not be very significant in reducing the cleansing net-effects from those of isolated stars. One should note, however, that the radiation from neighbouring stars may be of considerable significance, not by pushing dust grains from their own envelopes towards the Sun but by helping in pushing the dust away from the accreting Sun and thus increasing $\mathcal{D}$. Nevertheless, effects of the gas flows from stellar winds and expanding H\,II regions may be equally significant, and in practice reduce the final accretion of dust-poor matter onto the Sun. Obviously, this complex situation requires more detailed modelling, and the results may well turn out to be quite dependent on initial conditions and the geometrical characteristics of the arrangement. \subsection{The inflow: its rate and anisotropy, and the accretion disk} In order for the proposed mechanism to lead to the consequences observed, that is, a depletion of the dust-forming elements in the solar atmosphere of $\sim 20\%$, the inflow rate of the gas (assumed here to correspond to the accretion rate via the accretion disk) multiplied by the time $t_{\rm flow}$when the flow is active must match the mass $m_{\rm conv}$ of the stellar convection zone such that \begin{equation} \dot{M} \times t_{\rm flow} > 0.2 m_{\rm conv}. \label{eq1010} \end{equation} Adopting the mass of the convection zone following Figure 4 of \citet{Baraffe10} for a non-accreting 1 M$_\odot$ model, we find typical convective masses of 0.5, 0.2, and 0.05 M$_\odot$ at 10, 20, and 30 Myr, respectively. With those numbers, the condition of Equation (\ref{eq1010}) would be fulfilled, even with significantly smaller flow rates $\dot{M}$ than 0.02 M$_\odot$/Myr, provided that these remain relatively constant during the respective times. Star-forming gas is far from homogeneous. Instead, these gas clouds show cores, sheets, and filaments. This is clearly displayed in the simulations by \citet{Kuffmeier17} which demonstrate that stars accrete gas often streaming in filaments that simply fill a part (typically about 10 \%)\ of the full space angle as seen from the central region with the protostar and its accretion disk (see their Figure 7). If such inhomogeneities are also characteristic of the much later accretion stages hypothesised here, the inflow rates in the filaments must be increased correspondingly, relative to those of the isotropic case discussed above. Such an increase is still possible within the conditions set by Equation ({\ref{eq33}) and would admit a small optical depth in the filament. The basic conclusion of the present study is that the early Sun may have radiatively cleansed its outer layers from refractory elements if the accretion rates in late phases were small but significant. An interesting question would then be whether or not an accretion disk should be observable around stars also at these stages, far beyond the generally quoted observed lifetimes of protoplanetary disks of, at the most, 10 Myr (see \citet{Mamajek09}, \citet{Williams11} and references therein, but also \citet{Pfalzner14} who point out that the lifetimes may be underestimated). \citet{Kraus14}, on the basis of IR observations, find an upper limit of 5\% on the fraction of K3.0-K7.9 stars in the Tuc-Hor moving group (with an estimated age of 30-45 Myr) to host significant amounts of warm circumstellar dust. According to \citet{Alexander14} "a small (but not negligible $\lessapprox 10\%$) sample of older $ \sim10-20$ Myr pre-main sequence stars ... retain their dust (and gas) signatures". The possible existence of disks at small accretion flows is dependent on the outflow rates from the disk region and the replenishment time for the disk. The outflow, whether due to photo-evaporation by EUV, FUV, X-rays, or MHD disk winds (see \citet{Alexander14} for a discussion on this controversial issue) may be varying from star to star, but is believed to finally disperse the disk and end the accretion. If the renewal time is also short for late stages, for example $\sim 30,000$ yr as found in the simulations by \citet{Kuffmeier17}, the global accretion rates of our model suggest disk masses of approximately $10^{-4}$ M$_\odot$ at most, while with the longer dissipation times, measured from the time when accretion stops as estimated by several authors (see \citet{Williams11} for references) we could find disks that are more massive by one order of magnitude. \citet{Hartmann16} (their Figure 8) plotted observed accretion rates based on spectroscopic observations of the Balmer continuum and emission lines versus estimated age. We note that for stars with ages between 10 and 30 Myr the rates are of the same order of magnitude as those needed for our mechanism to work, although the data have great scatter and probably suffer from severe bias. It may be rewarding to look further for accretion flows and disks around solar-type stars in associations with ages of 20-30 Myr. \section{Conclusions} In spite of the exploratory nature of our study of the possibility of `self-cleansing' of the dusty envelope around the early Sun, we can safely conclude that the proposed mechanism would cleanse the gas of about half its dust mass early on, provided that the incoming gas flow is not much more massive than 0.01 M$_\odot$/Myr, and is sufficiently homogeneous, non-magnetic, and non-rotating. In a dense cluster the situation is uncertain, however, and probably dependent on the geometry of the stellar field, the presence of luminous stars, and of outflows and winds, and so on. Also, the presence of outflows from the early Sun itself may affect the situation considerably. Moreover, we have neglected the inner regions of the solar nebula, where the proto-planetary disk certainly might complicate the situation, both dynamically, via the gas flow and the magnetic field and by shielding the outer envelope from stellar radiation in certain directions, as well as by promoting other dust-fractionation processes. In order to clarify the effects of such complications, detailed MHD simulations, partly analogous to those of more massive stars by \citet{Wibking18} and those allowing for angular momentum, should be performed. One may question whether the mechanism proposed could be used to explain the findings by several research groups mentioned in the Introduction, that is, of binary stars with significantly different compositions, and cases with significant trends in these differences with condensation temperature of the elements. A natural way of obtaining the observed differences within the framework of our cleansing paradigm would be to invoke a situation in which the two components of the binary had different accretion rates, one with Equation (\ref{eq33}) satisfied and one where the rate was higher, for example, as a result of an inhomogeneous distribution of the inflowing gas. An alternative explanation might be that the two components were in somewhat different evolutionary phases, with convection zones of significantly different depth, when this final accretion occurred, meaning that the accreted dust-cleansed material was mixed with different amounts of stellar material. The finding by \citet{Gonzalez10} that CI meteorites show even lower refractory/volatile abundance ratios than the solar photosphere would also be explained by the present mechanism if the radiatively cleansed matter from which the meteorites formed was not enough to replace all the convection zone but was mixed into it. The seemingly systematic low abundance ratio of refractory elements to volatile elements of the cluster stars in M\,67 seems, however, more difficult to relate to the present scenario. One hypothesis might be that the deep gravity well of the cluster retained gas for a long time, which could then slowly accrete onto the stars with self-cleansing in operation. Another possibility is that the radiative pressure from the bright stars in the cluster could drive away the dust grains from the late accretion flows of the less massive stars. In both these scenarios one would expect the ratio of refractories to volatiles in the stellar atmospheres to increase towards the lower end of the main sequence of the cluster, as a result of the deepening of the convection zone with decreasing temperature. \begin{acknowledgements} Martin Asplund, Alexis Brandeker, Jorges Mel{\'e}ndez, Dhruba Mitra, Poul Erik Nissen, {\AA}ke Nordlund, Henk Spruit and an anonymous referee are thanked for many valuable suggestions. \end{acknowledgements} \bibliographystyle{aa}

a661446765c3128a7b86a12b8503cc946d8b9cff

\section{Introduction} In 1925, Hopf stated the following topological spherical space form problem as follows:\par \textit{Classify all manifolds with universal cover homeomorphic to the $n$-dimensional sphere $S^n$ for $n>1$}.\par This is equivalent to determining all finite groups that can act freely on $S^n$. Smith \cite{smith} showed that such a group $G$ must have periodic cohomology, i.e., $ G $ does not contain a subgroup of the form $ \mathbb{Z}_p \oplus \mathbb{Z}_p$ for any prime $p$.\par A natural generalization of the above problem is to classify all finite groups that can act freely on products of finitely many spheres. More generally, one can ask a similar question for arbitrary topological spaces. This led to the concept of free rank of symmetry of a topological space, defined as follows: \begin{definition} Let $X$ be a finite-dimensional CW-complex and $p$ be a prime number. The free $p$-rank of $X$, denoted by $\textrm{frk}_p(X)$, is the largest $r$ such that $(\mathbb{Z}_p)^r$ acts freely on $X$. The smallest value of $r$ over all primes $p$ is called the free rank of $X$ and denoted by frk($ X $). \end{definition} By Smith's result, we get that \begin{displaymath} \textrm{frk}_p \,(S^ n)= \left\{ \begin{array}{lll} 1 & \textrm{if $n$ is odd and $ p $ is arbitrary,}\\ 1 & \textrm{if $n$ is even and $ p=2$,}\\ 0 & \textrm{if $n$ is even and $ p>2 $. }\\ \end{array} \right. \end{displaymath} For products of finitely many spheres, the following statement appears as a question \cite[Question 7.2]{adem-browder} or as a conjecture \cite[p. 28]{adem-davis}. \begin{conjecture} If $ (\mathbb{Z}_p)^r $ acts freely on $S^{n_1}\times S^{n_2} \times \cdots \times S^{n_k} $, then $ r\leq k$. \end{conjecture} \noindent Carlsson \cite{carlsson-nonexistence, carlsson-p>2} proved the above conjecture for homologically trivial actions on products of equidimensional spheres. Adem and Browder \cite{adem-browder} proved that $ \textrm{frk}_p((S^n)^k) =k$ with the only remaining cases as $ p=2 $ and $ n=1, 3, 7 $. Later, Yal\c{c}in \cite{yalcin-group-extensions} proved that $\textrm{frk}_2((S^1)^k)=k $. The most general result which settles a stable form of the above conjecture is due to Hanke \cite[Theorem 1.3]{hanke}, which states that $${\textrm{frk}_p}\big(S^{n_1} \times \cdots \times S ^{n_k}\big)=k_0 \hspace{4mm}\textrm{if}\hspace{2mm} p>3(n_1+\cdots+n_k),$$ where $k_0$ is the number of odd-dimensional spheres in the product.\par In a recent work \cite[Theorem 1.2]{okutan-yalcin}, Okutan and Yal\c{c}in proved the conjecture in the case when the the dimensions $\{n_i\}$ of the spheres are high compared to the differences $|n_i - n_j|$ between the dimensions. The general case is still open.\par For free actions of arbitrary finite groups, there is a more general conjecture due to Benson-Carlson\cite{benson-carlson}. For a finite group, define the rank of $G$, denoted by $\operatorname{rk}(G)$, as $$ \operatorname{rk}(G) = \textrm{ max} \: \big\{r \;|\; (\mathbb{Z}_p)^r \leqslant G \; \textrm{for some prime $p$ } \big\} $$ and define $$ h(G) = \textrm{ min} \:\big\{k\;|\; G\textrm{ acts freely on } S^{n_1} \times \cdots \times S ^{n_k} \big\} .$$ \begin{conjecture}\hspace{-1mm}(Benson-Carlson)\; If $G$ is a finite group then, $\textrm{rk}(G)= h(G)$. \end{conjecture} Note that, $h(G)$ is well-defined since a result of Oliver \cite{oliver} states that every finite group acts freely on some products of spheres.\par For products of even-dimensional spheres, Cusick \cite{ cusick-even-sphere, cusick-free-action-even-sphere} proved that if a finite group $G$ acts freely on $S^{2n_1} \times \cdots \times S ^{2n_k}$ such that the induced action on $ H_*(X;\mathbb{Z}_2) $ is trivial, then $G\cong (\mathbb{Z}_2)^r$ for some $r\le k$.\par Though an extensive amount of work has been done for spheres and their products, the free rank of many other interesting spaces is still unknown. A few results are known such as for products of lens spaces and for products of projective spaces. For lens spaces, Allday \cite[Conjecture 5.2]{allday} conjectured that $$ \textrm{frk}_p \,{ \big(L_p^{2n_{1}-1} \times \cdots \times L_p^{2n_{k}-1}\big)}=k .$$ Yal\c{c}in \cite{yalcin-lens} proved the equidimensional case of the above conjecture.\par For products of real proective spaces, Cusick \cite{cusick-realprojective} conjectured that if $( \mathbb{Z}_2)^r $ acts freely and mod 2 cohomologically trivially on $ \mathbb{R}P^{n_1}\times\cdots\times\mathbb{R}P^{n_k}, $ then $ r\leq \eta(n_1)+\cdots+\eta(n_k),$ where the function $ \eta $ on natural number is defined by \begin{equation*} \eta(n)= \left\{ \begin{array}{ll} 0 & \textrm{if $n$ is even},\\ 1 & \textrm{if $n\equiv1\!\!\!\!\pmod 4$, }\\ 2 & \textrm{if $n\equiv3\!\!\!\!\pmod 4$. }\\ \end{array} \right. \end{equation*} He proved the conjecture when $ n_i\not\equiv3\!\pmod4 $ for all $ i $. Later, Yal\c{c}in \cite[Theorem 8.3]{yalcin-group-extensions} proved the conjecture when $ n_i $ is odd for each $ 0\leq i\leq k $. The general case is still open.\par For products of complex projective spaces, Cusick \cite[Theorem 4.13]{cusick-euler} proved that if a finite group $G$ acts freely on $ (\mathbb{C} P^m)^k $, then $G$ is a 2-group of order atmost $ 2^k $ and the exponent of $ G $ cannot exceed $ 2k $.\par Viewing the product of two projective spaces as a trivial bundle, it is interesting to determine the free rank of symmetry of twisted projective space bundles over projective spaces. Milnor and Dold manifolds are fundamental examples of such spaces. It is well-known that Dold and Milnor manifolds give generators for the unoriented cobordism algebra of smooth manifolds. Therefore, determining various invariants of these manifolds is an interesting problem. The problem of determining the free 2-rank of symmetry of products of Milnor manifolds has been investigated in \cite{msinghmilnor}, wherein some bounds have been obtained for the same.\par In this paper, we investigate the free rank problem for products of Dold manifolds. A Dold manifold is the quotient of $ S^m \times \mathbb{C}P^n $ by a free involution (see Definition \ref{definition}) and it is denoted by $ P(m, n) $. To compute the free rank of ${\displaystyle \prod_{i=1}^{k} P(2m_i,n_i)}$, we first determine the free rank of ${\displaystyle \prod_{i=1}^{k} S^{2m_i}\times\mathbb{C}P^{n_i}}$. Since a Dold manifold is a twisted complex projective space bundle over a real projective space, our arguments also give the free rank of ${\displaystyle \prod_{i=1}^{k} \mathbb{R}P^{2m_i}\times\mathbb{C}P^{n_i}}$. To state our main results, we first define a function $ \mu:\mathbb{N} \to \{0,1\} $ by \begin{equation}\label{mu} \mu(n)= \left\{ \begin{array}{ll} 1 & \textrm{if $n$ is odd},\\ 0 & \textrm{if $n$ is even.}\\ \end{array} \right. \end{equation}\vspace{1mm} \begin{theorem}\label{free-rank-univ-cover-introduction} Suppose $G$ is a finite group acting freely and mod 2 cohomologically trivially on a finite-dimensional CW-complex $X$ homotopy equivalent to ${\displaystyle \prod_{i=1}^{k} S^{2m_i}\times\mathbb{C}P^{n_i}}$. Then $G\cong (\mathbb{Z}_2)^l$ for some $ l $ satisfying $$ l\leq k_0 + \mu(n_1)+\mu(n_2)+\cdots+\mu(n_k),$$ where $ k_0 $ is number of the positive-dimensional spheres in the product. \end{theorem}\medskip \begin{theorem}\label{free-rank-dold-introduction} Suppose $G$ is a finite group acting freely and mod 2 cohomologically trivially on a finite-dimensional CW-complex $X$ homotopy equivalent to ${\displaystyle \prod_{i=1}^{k} P(2m_i,n_i)}$. Then $G\cong (\mathbb{Z}_2)^l$ for some $l$ satisfying $$ l\leq \mu(n_1)+\mu(n_2)+\cdots+\mu(n_k).$$ \end{theorem}\medskip Since $ P(0, n)=\mathbb{C}P^n $, by taking all $ m_i =0$, we retrive the result of Cusick \cite[Theorem 4.15]{cusick-euler}.\vspace{1mm}\par For an arbitrary free finite $ G $-CW-complex, there is a more general conjecture due to Carlsson \cite[Conjecture I.3]{carlsson-conjecture}: \begin{conjecture}\label{conjecture} Suppose $G=(\mathbb{Z}_p)^k$ and $ X $ is a finite free $G$-complex. Then $$2^k\leq \sum_{i=0}^{\dim X} \dim_{\mathbb{Z}_p}\, H_i(X;\mathbb{Z}_p).$$ \end{conjecture} In particular, the above conjecture is applicable for CW-complexes. A direct check shows that \begin{displaymath} \sum_{i=0}^{\dim X} \dim_{\mathbb{Z}_p}\, H_i(X;\mathbb{Z}_p)\geq \left\{ \begin{array}{ll} 2^{k_0 + \mu(n_1)+\mu(n_2)+\cdots+\mu(n_k)} & \textrm{for $X=\displaystyle \prod_{i=1}^{k} S^{2m_i}\times\mathbb{C}P^{n_i}$},\\ 2^{ \mu(n_1)+\mu(n_2)+\cdots+\mu(n_k)} & \textrm{for $X=\displaystyle \prod_{i=1}^{k} P(2m_i,n_i)$},\\ \end{array} \right. \end{displaymath} for mod 2 cohomologically trivial actions. Thus, Conjecture \ref{conjecture} is verified for our spaces under these conditions.\par We also consider the problem of determining the possible mod 2 cohomology algebra of orbit spaces of free involutions on Dold manifolds. Some results related to the cohomology algebra of orbit spaces can be found in \cite{pinka,msinghprojective,msinghlens}. In a recent paper \cite[Theorem 1]{morita}, Morita et al. determined the possible mod 2 cohomology algebra of orbit spaces of free involutions on Dold manifolds $P(1,n)$, where $n$ is an odd natural number. In this paper, we prove the general case using slightly different techniques. \begin{theorem} Let $G=\mathbb{Z}_2$ act freely on a finite-dimensional CW-complex homotopy equivalent to the Dold manifold $P(m,n)$ with a trivial action on mod 2 cohomology. Then $H^*\big(P(m,n)/G; \mathbb{Z}_2\big)$ is isomorphic to one of the following graded algebras: \begin{enumerate} \item $\mathbb{Z}_2[x,y,z]/ \big\langle x^2, y^{\frac{m+1}{2}}, z^{n+1} \big\rangle,$\\ where $\deg(x)=1$, $\deg(y)=2$, $\deg(z)=2$ and $m$ is odd; \item $\mathbb{Z}_2[{x},{y},{z}]/\big\langle {f}, {g}, {z}^{\frac{n+1}{2}}+h\big\rangle,$\\ where $\deg({x})=1$, $\deg({y})=1$, $\deg({z})=4$, $n$ is odd, $$f= \big(x^{m+1}+ \alpha_1 x^my+\alpha_2x^{m-1}y^2\big) \hspace{2mm} \textrm{and}\hspace{2mm} g= \big(y^3+ \beta_1 xy^2+\beta_2x^2y\big) ,$$ where $\alpha{_i}$, $\beta_j\in \mathbb{Z}_2$, and $ h \in \mathbb{Z}_2[{x}, {y}, {z }]$ is either a zero polynomial or it is a homogeneous polynomial of degree $ 2n+2 $ with the highest power of $ {z} $ is less than or equal to $ \frac{n-1}{2} $. \end{enumerate} \end{theorem} This paper is organized as follows. In Section \ref{sec2}, we recall the definition and cohomology of Dold manifolds. In Section \ref{sec3}, we construct free involutions on these manifolds. In Section \ref{sec4}, we study the free rank of symmetry of products of these manifolds. In Section \ref{sec5}, we determine the possible mod 2 cohomology algebra of orbit spaces of free involutions on Dold manifolds. In Section \ref{sec6}, we give an application to $ \mathbb{Z}_2 $-equivariant maps.\par Throughout the paper, by $\mathbb{Z}_p$ we denote the cyclic group of order $p$. For topological spaces $ X $ and $ Y $, $ X\simeq Y $ means that they are homotopy equivalent. \bigskip \section{Definition and cohomology of Dold manifolds}\label{sec2} \begin{definition}\label{definition} Let $m, n$ be non-negative integers such that $m+n>0$. The Dold manifold $P(m,n)$ is a closed connected smooth manifold of dimension $m+2n$, obtained as the quotient of $S^m \times \mathbb{C}P^n$ under the free involution $$\Big((x_0, x_1, \dots, x_m),\,[z_0, z_1, \dots, z_n] \Big) \mapsto \Big((-x_0, -x_1, \dots, -x_m),\,[\overline{z}_0, \overline{z}_1, \dots, \overline{z}_n] \Big).$$ \end{definition} It was shown by Dold \cite{dold} that, for suitable values of $m$ and $n$, the cobordism classes of $P(m,n)$ serve as generators in odd degrees for the unoriented cobordism algebra of smooth manifolds.\par The projection $S^m \times \mathbb{C}P^n \rightarrow S^m$ induces the map $$p:P(m,n)\rightarrow \mathbb{R}P^ m,$$ which is a locally trivial fiber bundle with fiber $\mathbb{C}P^n$. It admits a cross-section $s: \mathbb{R}P^ m \rightarrow P(m,n)$ defined as $s([x])=\big[(x, \: [1,0,\dots,0])\big]$. Consider the cohomology classes $a=p^*(x)\in H^1(P(m,n);\mathbb{Z}_2)$, where $x$ is the generator of $H^1(\mathbb{R}P^ m; \mathbb{Z}_2)$, and $b\in H^2(P(m,n);\mathbb{Z}_2)$, where $ b $ is characterized by the property that the restriction to a fiber is non-trivial and $s^*(b)=0$. The classes $a\in H^1(P(m,n);\mathbb{Z}_2)$ and $b\in H^2(P(m,n);\mathbb{Z}_2)$ generate $H^*(P(m,n);\mathbb{Z}_2)$. In particular, Dold \cite{dold} proved that $$H^*(P(m,n); \mathbb{Z}_2) \cong \mathbb{Z}_2[a,b]/\big\langle a^{m+1}, b^{n+1} \big\rangle,$$ where $\deg (a)=1$ and $\deg (b) =2$.\par Dold manifolds have been well-studied in the past. Some results related to group actions on these manifolds can be found in \cite{morita,peltier,ucci}. In a very recent work \cite{nath-sankaran}, Nath and Sankaran defined generalized Dold manifolds and obtained results on stable parallelizability of such manifolds. \bigskip \section{Free involutions on Dold manifolds}\label{sec3} In this section, we give some examples of free involutions on Dold manifolds. By Euler characteristic argument one can see that a Dold manifold $ P(m, n) $ admits a free involution if at least one of $ m $ and $ n $ is odd.\par (i) When $n$ is odd: Define $T_1:\mathbb{C}P^n \rightarrow \mathbb{C}P^n $ as the free involution given by $$ T_1 \Big([z_0, z_1, \dots, z_{n-1}, z_n ]\Big) = [-\overline{z}_1, \overline{z}_0, \dots, -\overline{z}_n, \overline{z}_{n-1}].$$ Then $$id \times T_1 : S^m \times \mathbb{C}P^n \rightarrow S^m \times \mathbb{C}P^n $$ gives a free involution on $S^m \times \mathbb{C}P^n$. Since $T_1$ commutes with conjugation, we get a free involution on $P(m,n)$ by passing to the quotient.\vspace{1mm} \par (ii)\label{t2} When $m$ is odd: Let $T_2$ be the action on $S^m \times \mathbb{C}P^n$ defined by $$T_2\Big((x_0, x_1, \dots, x_{m-1}, x_{m}), \,[z_0, z_1, \dots, z_{n}]\Big) =\Big((-x_1, x_0, \dots, -x_{m}, x_{m-1}), \,[z_0, z_1, \dots, z_{n}]\Big).$$ Note that $T_2$ is free and it commutes with the action defining the Dold manifold. Hence, we get a free involution on $P(m,n)$ by passing to the quotient.\vspace{1mm} \par (iii) When both $m$ and $n$ are odd: In this case, define the action $T_3$ on $S^m \times \mathbb{C}P^n$ by $$T_3\Big((x_0, x_1, \dots, x_{m-1}, x_{m}), \,[z_0, z_1, \dots, z_{n}]\Big) =\Big((-x_1, x_0, \dots, -x_{m}, x_{m-1}),\,[-\overline{z}_1, \overline{z}_0, \dots, -\overline{z}_n, \overline{z}_{n-1}]\Big).$$ This gives another free involution on $P(m,n)$ by passing to the quotient. \bigskip \section{Free rank of symmetry}\label{sec4} In this section, we prove some results regarding the free rank of symmetry of ${\displaystyle \prod_{i=1}^{k} P(2m_i,n_i)}$. For this we first determine the free rank of symmetry of its universal cover ${\displaystyle \prod_{i=1}^{k} S^{2m_i}\times\mathbb{C}P^{n_i}}$. Our main computational tool is the Leray-Serre spectral sequence associated to the Borel fibration. For details, we refer the reader to \cite{allday-puppe,mccleary}.\medskip \begin{lemma}\label{lemma-zp-trivial-imply-z-trivial} Let $ p $ be an odd prime. If\, $ \mathbb{Z}_p$ acts freely and mod 2 cohomologically trivially on a finite-dimensional CW-complex $X$ homotopy equivalent to ${\displaystyle \prod_{i=1}^{k} S^{2m_i}\times\mathbb{C}P^{n_i}}$, then the induced action on $ H^*(X;\mathbb{Z}) $ is trivial. \end{lemma} \begin{proof} As the integral cohomology groups of $ X $ are torsion-free, we get $$ H^k(X;\mathbb{Z}_2)\cong H^k(X;\mathbb{Z}) \otimes \mathbb{Z}_2 .$$ Let $ g$ be a generator of $ \mathbb{Z}_p $ and $ \phi: \mathbb{Z} \to \mathbb{Z}_2 $ be the non-trivial homomorphism. Then we have the following commutative diagram: \[\begin{tikzcd}[column sep=4em, row sep=2em] H^k(X;\mathbb{Z}) \ar{r} {g_\mathbb{Z}^*} \ar[swap]{d}{\phi^*} & H^k(X;\mathbb{Z}) \ar{d}{\phi^*} \\ H^k(X;\mathbb{Z}_2) \ar[swap]{r}{g_{\mathbb{Z}_2}^*} & H^k(X;\mathbb{Z}_2), \end{tikzcd} \] \noindent where $ g_\mathbb{Z}^* $ and $ g_{\mathbb{Z}_2}^* $ are the induced maps on cohomology with $ \mathbb{Z} $ and $ \mathbb{Z}_2 $-coefficients, respectively. Let $ A\in \mathrm{GL}(n,\mathbb{Z}) $ be the matrix corresponding to ${g_{\small {\mathbb{Z}}}^*}$. Then the matrix $ \overline{A} $ corresponding to ${g_{\small{\mathbb{Z}_2}}^*}$ is obtained by reducing all the entries of $ A $ modulo 2. By assumption $\overline{A}=I$, the identity matrix. Consequently, $ A=2M+I$ for some matrix $ M $ and $ A^p=I $. Thus, \begin{equation}\label{equ-matrix} A^p-I = (A-I) \hspace{1mm}(A^{p-1} + A^{p-2} + \cdots + I )=0, \end{equation} and $$ (A^{p-1} + A^{p-2} + \cdots + I )= (2M + I)^{p-1} + (2M+I)^{p-2} + \cdots + I= 2N + I,$$ for some matrix $ N $. Since det$ (2N + I) $ is odd, $ (A^{p-1} + A^{p-2}+ \cdots + I )$ is invertible. It follows from equation \eqref{equ-matrix} that $ A=I $. \end{proof} By using Lefschetz fixed point theorem we get the following result. \begin{lemma}\label{lemma-zp-not-action-univ-Z-cohomlogy} If $ p $ is an odd prime, then $\mathbb{Z}_p$ cannot act freely and $ \mathbb{Z} $-cohomologically trivially on a finite-dimensional CW-complex $X$ homotopy equivalent to ${\displaystyle \prod_{i=1}^{k} S^{2m_i}\times\mathbb{C}P^{n_i}}$. \end{lemma} Lemma \ref{lemma-zp-trivial-imply-z-trivial} and Lemma \ref{lemma-zp-not-action-univ-Z-cohomlogy} together imply the following result. \begin{lemma}\label{lemma-zp} If $ p $ is an odd prime, then $\mathbb{Z}_p$ cannot act freely and mod 2 cohomologically trivially on a finite-dimensional CW-complex $X$ homotopy equivalent to ${\displaystyle \prod_{i=1}^{k} S^{2m_i}\times\mathbb{C}P^{n_i}}$. \end{lemma} \begin{lemma}\label{lemma-disjoint-union} A finite group $ G $ acts freely and mod 2 cohomologically trivially on the disjoint union $\displaystyle {\coprod _{i=1}^{k} X}$ of a space $ X $ if and only if it acts freely and mod 2 cohomologically trivially on $ X $. \end{lemma} \begin{proof} Suppose $ G $ acts freely and mod 2 cohomologically trivially on $ X $. Then we can define a free $ G $-action on $\displaystyle {\coprod _{i=1}^{k} X}=\displaystyle {\bigcup _{i=1}^{k}\,X \times \{i\}}$ by setting $$ g\cdot(x, i)=(gx, i) .$$ Since the cohomology algebra of a disjoint union is the direct product of the cohomology algebra of the components, it follows that the induced action of $ G $ on mod 2 cohomology algebra is also trivial. Conversely, if $ G $ acts freely and mod 2 cohomologically trivially on $\displaystyle {\coprod _{i=1}^{k} X}$, then the action preserves each component and acts as desired. \end{proof} \medskip Now we prove our main results. For the definition of the function $ \mu $ see equation \eqref{mu}. \begin{theorem}\label{univ-cover} Let $G$ be a finite group acting freely and mod 2 cohomologically trivially on a finite-dimensional CW-complex $X$ homotopy equivalent to ${\displaystyle \prod_{i=1}^{k} S^{2m_i}\times\mathbb{C}P^{n_i}}$. Then $G\cong (\mathbb{Z}_2)^l$ for some $ l $ satisfying $$ l\leq k_0 + \mu(n_1)+\mu(n_2)+\cdots+\mu(n_k),$$ where $ k_0 $ is the number of positive-dimensional spheres in the product. \end{theorem} \begin{proof} First, we consider the case where each $m_i>0 $. We prove the assertion by first showing that $\mathbb{Z}_4$ does not act freely and mod 2 cohomologically trivially on $X\simeq{\displaystyle \prod_{i=1}^{k} S^{2m_i}\times\mathbb{C}P^{n_i}}$. On the contrary, suppose $\mathbb{Z}_4$ acts freely and mod 2 cohomologically trivially on $ X $. By K\"{u}nneth formula we get $$H^*(X; \mathbb{Z}_2)=\mathbb{Z}_2[a_1,\dots, a_k]/\big\langle a_1^{n_1+1},\dots ,a_k^{n_k+1}\big\rangle \otimes \mathord{\adjustbox{totalheight=.85\baselineskip}{$\bigwedge$}} \big(b_1,\dots, b_k\big),$$ where $\deg(a_i)=2$ and $\deg(b_i)=2m_i$. Consider the Leray-Serre spectral sequence $ (E_r^{*,*},d_r) $ associated to the Borel fibration $$X\stackrel{}{\hookrightarrow} X_{\mathbb{Z}_4} \stackrel{\pi}{\longrightarrow} B_{\mathbb{Z}_4}.$$ The $ E_2$-page of this spectral sequence looks like \[ E_2^{*,*} \cong H^*(B_{\mathbb{Z}_4}; H^*(X;\mathbb{Z}_2)).\] Since the induced action on mod 2 cohomology is trivial, the system of local coefficients is simple. Thus, we have $$ E_2^{*,*} \cong H^*(B_{\mathbb{Z}_4};\mathbb{Z}_2) \otimes H^*(X; \mathbb{Z}_2). $$ Consider the subgroup inclusion $ i: \mathbb{Z}_2 \hookrightarrow \mathbb{Z}_4 $. By restricting the $ \mathbb{Z}_4 $ action to $ \mathbb{Z}_2 $, we get a free $ \mathbb{Z}_2 $ action on $ X $. Then consider the Leray-Serre spectral sequence $(\widetilde{E } _r^{*,*}, \widetilde{d}_r)$ associated to the Borel fibration $$X\stackrel{}{\hookrightarrow} X_{\mathbb{Z}_2} \stackrel{\pi}{\longrightarrow} B_{\mathbb{Z}_2}.$$ The naturality of the Leray-Serre spectral sequence gives the following commutative diagram: \[\begin{tikzcd}[column sep=4em, row sep=2em] E_r^{k,l} \ar{r} {f^*} \ar[swap]{d}{d_r} & \widetilde{E }_r^{k,l} \ar{d}{\widetilde{d}_r} \\ E_r^{k+r,l-r+1} \ar[swap]{r}{f^*} & \widetilde{E }_r^{k+r,l-r+1}, \end{tikzcd} \] \noindent where at the $ E_2 $-page the map $ f^*: (E_r^{*,*},d_r) \to (\widetilde{E } _r^{*,*}, \widetilde{d}_r)$ is defined by $$f^*= i^* \otimes id: H^*(B_{\mathbb{Z}_4}; \mathbb{Z}_2) \otimes H^*(X; \mathbb{Z}_2) \to H^*(B_{\mathbb{Z}_2};\mathbb{Z}_2)\otimes H^*(X; \mathbb{Z}_2).$$ We know that, $$H^*(B_{\mathbb{Z}_4}; \mathbb{Z}_2)\cong \mathbb{Z}_2[x]\otimes\mathord{\adjustbox{totalheight=.85\baselineskip}{$\bigwedge$}}(y),$$ where $\deg(x)=2$ and $\deg(y)=1 $. We also recall that $H^*(B_{\mathbb{Z}_2}; \mathbb{Z}_2)\cong \mathbb{Z}_2[t],$ where $\deg(t)=1$. It can be seen that the map $ i^*: H^*(B_{\mathbb{Z}_4}; \mathbb{Z}_2) \to H^*(B_{\mathbb{Z}_2}; \mathbb{Z}_2)$ is given by $ i^*(x)=t^2 $ and $ i^*(y)=0 $. Since $i^*(y)=0$, it follows that $f^*$ is zero on ${E_r^{odd, *}}$. Since $H^*(X; \mathbb{Z}_2)$ has non-zero cohomology only in even dimension, it follows that the non-zero differentials lie in $E_r$-page, where $r$ is odd. Consider an element $z\in H^*(X; \mathbb{Z}_2)$. Then $$\widetilde{d}_{2n+1}(1\otimes z)=\widetilde{d}_{2n+1}f^*(1\otimes z) =f^*d_{2n+1}(1\otimes z).$$ As $d_{2n+1}(1\otimes z) \in E_{2n+1}^{odd, *}$, it follows that $\widetilde{d}_{2n+1}(1\otimes z)=0$. Since this is true for any $z\in H^*(X; \mathbb{Z}_2)$ and for any differential $\widetilde{d}_r$, it follows that the corresponding spectral sequence collapses at the $E_2$-page. This contradicts the fact that $X/\mathbb{Z}_2$ is finite-dimensional. Thus, we have proved that $\mathbb{Z}_4$ cannot act freely on $X$.\par If $ G $ is not a finite 2-group, then there exists a prime $ p> 2$ such that $ p$ divides $|G|$. Consequently, $ \mathbb{Z}_p $ acts freely and mod 2 cohomologically trivially on $ X $, which contradicts Lemma \ref{lemma-zp}. Therefore, $ G $ is a finite 2-group.\par Now if $ n_i\equiv 1\pmod 4 $ for all $ i $, then the desired bound on the rank of $ G $ follows from Euler characteristic argument. But, if $n_i\equiv 3\pmod 4$ for some $ i $, then the Euler characteristic fails to give exact bound on the rank of $G$. To derive the exact bound, we first recall some properties of Lefschetz number. For a map $ f:X \to X $, we can define the Lefschetz number $ \tau(f) $ as $$ \tau(f)= \sum_{i} (-1)^i\; tr(f^*:H^n(X; \mathbb{Z})\to H^n(X; \mathbb{Z}) ).$$ Suppose $ f^* $ is an automorphism of $ H^*(X; \mathbb{Z}) $ such that $ f^*(a_i)=\xi_i a_i $ and $ f^*(b_i)=\lambda_i b_i $, where $ a_i, b_i $ are the generators of $ H^*(X;\mathbb{Z}) $ and $ \xi_i, \lambda_i \in \mathbb{Z}$. Then its Lefschetz number is precisely the following \begin{equation}\label{equ-lefschetz} \tau(f)=\prod_{i=1}^{k}(1 + \xi_i+\xi_i^2+\cdots+\xi_i^{n_i})(1+\lambda_i). \end{equation}\par Let $ g$ be an element of $G $ and $ g^*_{\mathbb{Z}} $ be the induced map on $ H^*(X; \mathbb{Z}) $. Since $ G $ acts freely, the Lefschetz number of $ g^*_{\mathbb{Z}} $ must vanish for all $ g\ne 1 $. This shows that $ G $ is isomorphic to some subgroup of $ Aut(H^*(X; \mathbb{Z})) $. Notice that, an element of $ Aut(H^*(X;\mathbb{Z})) $ is completely determined by its value on the generators $ a_1, a_2, \dots, a_k $ and $ b_1, b_2, \dots ,b_k $. Since $ g^*_\mathbb{Z} $ is an automorphism, it preserves degrees as well as cup-length, where for an element $ x\in H^*(X; \mathbb{Z}) $, the cup-length of $ x $ is the greatest integer $ n $ such that $ x^n \ne 0$. Thus, $ g^*_\mathbb{Z} $ can only permute between those generators which have same degrees and same cup-length (this follows because the cup-length of a sum of the generators is the sum of the cup-lengths of the individual generators). In other words, $ g^*_\mathbb{Z}(a_i)= \pm a_{\sigma(i)} $ and $ g^*_\mathbb{Z}(b_i)= \pm b_{\gamma(i)} $, where $ \sigma, \gamma \in S_k$ (the symmetric group on $ k $ elements) and $ a_\sigma(i) $ has same cuplength as $ a_i $. Let $ g^*_{\mathbb{Z}_2} $ be the induced map on $ H^*(X; \mathbb{Z}_2) $. By the naturality of the cohomology functor, we get the following commutative diagram: \[\begin{tikzcd}[column sep=4em, row sep=2em] H^k(X;\mathbb{Z}) \ar{r} {g_\mathbb{Z}^*} \ar[swap]{d}{} & H^k(X;\mathbb{Z}) \ar{d}{} \\ H^k(X;\mathbb{Z}_2) \ar[swap]{r}{g_{\mathbb{Z}_2}^*} & H^k(X;\mathbb{Z}_2). \end{tikzcd} \] \noindent Since we have assumed that $ G $ acts mod 2 cohomologically trivially, we get $ g^*_\mathbb{Z}(a_i)= \pm a_{i} $ and $ g^*_\mathbb{Z}(b_i)= \pm b_{i} $. This shows that $ G$ is a subgroup of $(\mathbb{Z}_2)^{2k} $. The Lefschetz number should be zero corresponding to each non-identity element of $ G $. From (\ref{equ-lefschetz}) we get that, $ \tau(f)=0 $ only if either $ \lambda_i=-1 $ for some $ i $ or $1 + \xi_j+\xi_j^2+\cdots+\xi_j^{n_j}=0 $ for some $ j $. The last case is possible only if $ n_j $ is odd and in that case we get $ \xi_j=-1 $. This implies that $$ G\subseteq (\mathbb{Z}_2)^{k + \mu(n_1)+\mu(n_2)+\cdots+\mu(n_k)} ,$$ where $ \mu(n)=0 $ if $ n $ is even and $ \mu(n)=1 $ if $ n $ is odd (note that if some $n_j $ is even, then an automorphism $ g^* $ that maps $ a_j \mapsto -a_j $ and fixes all other generators of $ H^*(X;\mathbb{Z}) $ is not possible). Consequently, $G\cong (\mathbb{Z}_2)^l$ for some $ l $ satisfying $$ l\leq k + \mu(n_1)+\mu(n_2)+\cdots+\mu(n_k).$$\par Next, we consider the case where $ m_i =0$ for some $ i $. Let $ k_0 $ be the number of positive-dimensional spheres in the product. Note that, ${\displaystyle \prod_{i=1}^{k} S^{2m_i}\times\mathbb{C}P^{n_i}}$ can be written as disjoint union of $ 2^{k-k_0} $ copies of $\Big({\displaystyle{ \prod_{m_i>0} S^{2m_i} \times \,\prod_{i=1}^{k} \mathbb{C}P^{n_i} }}\Big)$. By Lemma \ref{lemma-disjoint-union}, it is enough to prove the result for $\Big({\displaystyle{ \prod_{m_i>0} S^{2m_i} \times \,\prod_{i=1}^{k} \mathbb{C}P^{n_i} }}\Big)$. Applying similar argument as in the previous case, we get the desired result. \end{proof} \begin{corollary} For mod 2 cohomologically trivial actions \begin{equation*} \textrm{frk}_p(\, {\displaystyle \prod_{i=1}^{k} S^{2m_i}\times\mathbb{C}P^{n_i}} \,)= \left\{ \begin{array}{ll} k_0 + \mu(n_1)+\mu(n_2)+\cdots+\mu(n_k) & \textrm{if $p=2$},\\ 0 & \textrm{if $p>2$,}\\ \end{array} \right. \end{equation*} where $ k_0 $ is the number of positive-dimensional spheres in the product. \end{corollary} \vspace{.5mm} \begin{remark} If we take arbitrary dimensional spheres in the statement of Theorem \ref{univ-cover}, then our method does not work. Here we would like to emphasize that the problem of determining the free rank for products of different dimensional spheres is a long standing open problem. \end{remark} \vspace{1mm} In view of the preceding remarks, we would like to propose the following conjecture. \begin{conjecture}\label{conjectureforuniv} If $ (\mathbb{Z}_2)^l$ acts freely and mod 2 cohomologically trivially on a finite-dimensional CW-complex $X\simeq{\displaystyle \prod_{i=1}^{k} S^{m_i}\times\mathbb{C}P^{n_i}}$ where each $ m_i>0 $, then $$l \leq k + \mu(n_1)+\cdots+ \mu(n_k). $$ \end{conjecture} \medskip Next we prove a result related to the free rank of Dold manifolds. \begin{theorem}\label{doldfreerank} Suppose $G$ is a finite group acting freely and mod 2 cohomologically trivially on a finite-dimensional CW-complex $X$ homotopy equivalent to ${\displaystyle \prod_{i=1}^{k} P(2m_i,n_i)}$. Then $G\cong (\mathbb{Z}_2)^l$ for some $l$ satisfying $$ l\leq \mu(n_1)+\mu(n_2)+\cdots+\mu(n_k).$$ \end{theorem} \begin{proof} First, we prove that $\mathbb{Z}_4$ cannot act freely and mod 2 cohomologically trivially on $X\simeq{\displaystyle \prod_{i=1}^{k} P(2m_i,n_i)}$. On the contrary, suppose $\mathbb{Z}_4$ acts freely and mod 2 cohomologically trivially on $X$ and let $Y$ denote its orbit space. Then there is an exact sequence $$ 1\to (\mathbb{Z}_2)^k\to \pi_1(Y)\to \mathbb{Z}_4\rightarrow 1.$$ It follows from the exactness of the sequence that $\pi_1(Y)$ is a 2-group. Also, $\pi_1(Y)$ acts freely on the universal cover ${\displaystyle \prod_{i=1}^{k} S^{2m_i}\times\mathbb{C}P^{n_i}}$. By Theorem \ref{univ-cover}, $ \pi_1(Y) $ must be an elementary abelian 2-group and $\pi_1(Y)\cong(\mathbb{Z}_2)^{k+2}$. But this is not possible since $(\mathbb{Z}_2)^{k+2}\big/(\mathbb{Z}_2)^{k}$ is not isomorphic to $ \mathbb{Z}_4$. This implies that $\mathbb{Z}_4$ cannot act freely on ${\displaystyle \prod_{i=1}^{k} P(2m_i,n_i)}$. Similar proof also shows that there is no free $ \mathbb{Z}_p $-action on ${\displaystyle \prod_{i=1}^{k} P(2m_i,n_i)}$ for odd primes $ p $. Then using Theorem \ref{univ-cover}, one sees that $G\cong (\mathbb{Z}_2)^l$, where $ l $ is as desired. \end{proof} \medskip Examples from Section \ref{sec3} show that the preceding bound in Theorem \ref{doldfreerank} is sharp. As a consequence we get the following result on free rank. \begin{corollary} For mod 2 cohomologically trivial actions \begin{equation*} \textrm{frk}_p\,(\,{\displaystyle \prod_{i=1}^{k} P(2m_i, n_i)} \,)= \left\{ \begin{array}{ll} \mu(n_1)+\mu(n_2)+\cdots+\mu(n_k) & \textrm{if $p=2$},\\ 0 & \textrm{if $p>2$.}\\ \end{array} \right. \end{equation*} \end{corollary} \begin{remark} Note that, the mod 2 cohomology algebra of $\mathbb{R}P^m \times \mathbb{C}P^n$ and that of $P(m,n)$ are isomorphic. Also, they have the same universal cover $S^m \times \mathbb{C}P^n$. Hence, the proof of Theorem \ref{doldfreerank} also works for $\displaystyle{\prod_{i=1}^{k} \mathbb{R}P^{2m_i} \times \mathbb{C}P^{n_i}}$. \end{remark} Taking into account Theorem \ref{doldfreerank} and the free 2-rank of $ \prod_{i=1}^{k}\mathbb{R}P^{m_i} $, we conclude this section with the following: \begin{question} Does the equality \[\textrm{frk}_2\,(\prod_{i=1}^k P(m_i, n_i))= \eta(m_1)+\cdots+\eta(m_k)+ \mu(n_1)+\cdots+\mu(n_k) \] hold? Here, $ \mu(n_i)=1 $ for $ n_i $ is odd and $ \mu(n_i)=0 $ for $ n_i $ is even and $ \eta $ is defined by \begin{equation*} \eta(m)= \left\{ \begin{array}{ll} 0 & \textrm{if $m$ is even},\\ 1 & \textrm{if $m\equiv1\!\!\!\!\pmod 4$, }\\ 2 & \textrm{if $m\equiv3\!\!\!\!\pmod 4$. }\\ \end{array} \right. \end{equation*} \end{question} \bigskip \section{Cohomology algebra of orbit spaces of free involutions on Dold manifolds }\label{sec5} In this section, we determine the possible mod 2 cohomology algebra of orbit spaces of arbitrary free involutions on Dold manifolds. We require the following result in the proof of the main theorem. \begin{lemma}\label{lemma} Let $\mathbb{Z}_2 $ act freely on $ S^m \times \mathbb{C}P^n $. Then the induced action on $ H^*(S^m \times \mathbb{C}P^n;\mathbb{Z}_2) $ is trivial. \end{lemma} \begin{proof} Recall that, $$H^*(S^m\times \mathbb{C}P^n; \mathbb{Z}) \cong \mathbb{Z}[x,y]/\big\langle x^{2}, y^{\hspace{.3mm}n+1} \big\rangle,$$ where $\deg (x)=m$ and $\deg (y) =2$. Suppose $ \mathbb{Z}_2 $ acts freely on $S^m \times \mathbb{C}P^n $ and let $ g $ be the generator of $ \mathbb{Z}_2 $. Let $ g_\mathbb{Z}^* $ be the induced action on $ H^*(S^m \times \mathbb{C}P^n;\mathbb{Z}) $ corresponding to $ g $. If $ g_\mathbb{Z}^* $ is non-trivial then it has one of the following possibilities: \begin{enumerate} \item[] (i) $ x \mapsto -x $ and $ y\mapsto y ,$ \item[] (ii) $ x \mapsto x $ and $ y\mapsto -y ,$ \item[] (iii) $ x \mapsto -x $ and $ y\mapsto -y $. \end{enumerate} Since $ H^k(S^m \times \mathbb{C}P^n;\mathbb{Z}) $ is torsion-free for any $ k\geq 0 $, we get $$ H^k(S^m \times \mathbb{C}P^n;\mathbb{Z}_2)\cong H^k(S^m \times \mathbb{C}P^n;\mathbb{Z}) \otimes \mathbb{Z}_2 .$$ Let $ \phi: \mathbb{Z} \to \mathbb{Z}_2$ be the non-trivial homomorphism. Corresponding to $ \phi $ and $ g $ we get the following commutative diagram: \[\begin{tikzcd} H^k(S^m \times \mathbb{C}P^n;\mathbb{Z}) \ar{r} {g_\mathbb{Z}^*} \ar[swap]{d}{\phi^*} & H^k(S^m \times \mathbb{C}P^n;\mathbb{Z}) \ar{d}{\phi^*} \\ H^k(S^m \times \mathbb{C}P^n;\mathbb{Z}_2) \ar[swap]{r}{g_{\mathbb{Z}_2}^*} & H^k(S^m \times \mathbb{C}P^n;\mathbb{Z}_2), \end{tikzcd} \] where $ g_\mathbb{Z}^* $ and $ g_{\mathbb{Z}_2}^* $ are the induced maps on cohomology in $ \mathbb{Z} $ and $ \mathbb{Z}_2 $-coefficients, respectively. This shows that $ g_{\mathbb{Z}_2}^* =g_\mathbb{Z}^* \otimes \mathbb{Z}_2$ and hence we get $ g_{\mathbb{Z}_2}^* $ is the identity map. \end{proof} \medskip \begin{remark} The preceding lemma shows that if $\mathbb{Z}_2 \oplus \mathbb{Z}_2$ acts freely on $S^m \times \mathbb{C}P^n$, then the induced action on mod 2 cohomology is trivial. \end{remark} \begin{theorem}\label{coho-dold} Let $G=\mathbb{Z}_2$ act freely on a finite-dimensional CW-complex homotopy equivalent to the Dold manifold $P(m,n)$ with a trivial action on mod 2 cohomology. Then $H^*\big(P(m,n)/G; \mathbb{Z}_2\big)$ is isomorphic to one of the following graded algebras: \begin{enumerate}\label{case1} \item $\mathbb{Z}_2[x, y, z]/ \big\langle x^2, y^{\frac{m+1}{2}}, z^{n+1} \big\rangle,$ where $\deg(x)=1$, $\deg(y)=2$, $\deg(z)=2$ and $m$ is odd ; \item \label{case2} $\mathbb{Z}_2[{x}, {y}, {z}]/\big\langle {f}, {g}, {z}^{\frac{n+1}{2}}+h\big\rangle,$\\ where $\deg({x})=1$, $\deg({y})=1$, $\deg({z})=4$, $n$ is odd, $$f= \big(x^{m+1}+ \alpha_1 x^my+\alpha_2x^{m-1}y^2\big) \hspace{2mm} \textrm{and}\hspace{2mm} g= \big(y^3+ \beta_1 xy^2+\beta_2x^2y\big) ,$$ where $\alpha{_i}$, $\beta_j\in \mathbb{Z}_2$, and $ h \in \mathbb{Z}_2[{x}, {y}, {z }]$ is either a zero polynomial or it is a homogeneous polynomial of degree $ 2n+2 $ with the highest power of $ {z} $ is less than or equal to $ \frac{n-1}{2} $. \end{enumerate} \end{theorem} \begin{proof} Recall that, the cohomology algebra of the Dold manifold is $$H^*\big(P(m,n); \mathbb{Z}_2\big) \cong \mathbb{Z}_2[a,b]/\big\langle a^{m+1}, b^{n+1} \big\rangle,$$ where $\deg (a)=1$ and $\deg (b) =2$. Consider an arbitrary free involution on $P(m,n)$ and let $M$ denote its orbit space. Note that, the fundamental group of $M$ can be either $\mathbb{Z}_4$ or $\mathbb{Z}_2\oplus \mathbb{Z}_2$, and it acts on the universal cover $S^m \times \mathbb{C}P^n$ of $P(m,n)$.\par {Case (1)}: First, consider the case $\pi_1(M)=\mathbb{Z}_4 $. It follows that $H^1(M;\mathbb{Z}_2)= \mathbb{Z}_2$. Consider the Leray-Serre spectral sequence associated to the Borel fibration $$P(m,n)\stackrel{i}{\hookrightarrow} M \stackrel{\pi}{\longrightarrow} B_{\mathbb{Z}_2}.$$ Since we have assumed that $ \pi_1(B_{\mathbb{Z}_2})=\mathbb{Z}_2 $ acts trivially on $ H^*(P(m,n);\mathbb{Z}_2) $, the above fibration has a simple system of local coefficients. Hence, the spectral sequence has the form $$E_2^{p,q}\cong H^p(B_{\mathbb{Z}_2}; \mathbb{Z}_2) \otimes H^q(P(m,n);\mathbb{Z}_2).$$ Since $ M $ is finite-dimensional, the spectral sequence does not degenerate at the $E_2$-page. As $H^1(M;\mathbb{Z}_2)= \mathbb{Z}_2$, the only possibility of the differential is $d_2(1\otimes a)= t^2 \otimes 1$ and $d_2(1\otimes b)= 0$. This implies $m$ must be odd. By computing the ranks of $\operatorname{Ker} d_2$ and $\operatorname{Im} d_2$, we get rk$(E_3^{k,l})$ = 0 for all $k \geq 2$. Thus, the spectral sequence collapses at the $E_3$-page and we get that $ E_\infty^{*,*} \cong E_3^{*,*}$. Consequently, $$H^k(M;\mathbb{Z}_2)\cong \bigoplus_{i+j=k}E_{\infty}^{i,j}= E_{\infty}^{0,k} \oplus E_{\infty}^{1,k-1}$$ for $0 \leq k \leq m+2n$.\par We see that ${ 1 \otimes a^2 } \in E_2^{0,2}$ and $1 \otimes b \in E_2^{0,2}$ are permanent cocycles. Let $i^*$ and $\pi^*$ be the homomorphisms induced on cohomology, and they are precisely the edge maps (see \cite{mccleary}, Theorem 5.9). Choose $y \in H^2(M;\mathbb{Z}_2)$ such that $i^*(y)=a^2$. Then we have $y^{\frac{m+1}{2}}=0$. Similarly, choose $z \in H^2(M; \mathbb{Z}_2)$ such that $i^*(z)= b$. Then we have $z^{n+1}=0$. Let $x=\pi^*(t) \in E_\infty^{1,0} \subseteq H^1(M;\mathbb{Z}_2) $ be determined by $t \otimes 1 \in E_2^{1,0}$. As $E_\infty^{2,0} = 0$, we have $x^2 = 0$. In this case, we obtain \begin{align}\label{z4-coho} H^*(M;\mathbb{Z}_2) \cong \mathbb{Z}_2[x,y,z]/\big\langle x^2, y^{\frac{m+1}{2}}, z^{n+1}\big\rangle, \end{align} where $\deg(x)=1$, $\deg(y)=2$ and $\deg(z)=2$. \medskip {Case (2)}: Next, consider the case $\pi_1(M)=\mathbb{Z}_2 \oplus\mathbb{Z}_2$. In this case, it is not possible to determine the cohomology algebra of the orbit spaces directly from the Leray-Serre spectral sequence associated to the Borel fibration $$P(m,n)\stackrel{i}{\hookrightarrow} M \stackrel{\pi}{\longrightarrow} B_{\mathbb{Z}_2}.$$ So we proceed as follows. First, note that $\mathbb{Z}_2 \oplus\mathbb{Z}_2$ acts freely on the universal cover $S^m \times \mathbb{C}P^n$. Consider only those free $\mathbb{Z}_2 \oplus\mathbb{Z}_2$-actions on $S^m \times \mathbb{C}P^n$ that factor through the Dold manifold. Let $\alpha$ and $\beta$ be the generators of $\mathbb{Z}_2 \oplus\mathbb{Z}_2$. Then both $ \alpha $ and $ \beta $ give rise to some free involutions on $ S^m \times \mathbb{C}P^n $. Let $\alpha$ act on $S^m \times \mathbb{C}P^n$ by the free involution defining the Dold manifold, and the action of $\beta$ on $S^m \times \,\mathbb{C}P^n$ is an arbitrary free involution commuting with $\alpha$. By defining the action as above we get all possible free $\mathbb{Z}_2$-actions on the Dold manifold. Consider the subgroup $\mathbb{Z}_2$ generated by $\alpha$. There is an inclusion $i: \mathbb{Z}_2\stackrel{}{\hookrightarrow} \mathbb{Z}_2 \oplus\mathbb{Z}_2 $ given by $ \alpha \rightarrow (\alpha,0)$. This gives a map between the classifying spaces $$B_{\mathbb{Z}_2} \rightarrow B_{\:\mathbb{Z}_2 \:\oplus\:\mathbb{Z}_2}.$$ Recall that, $$H^*(B_{\mathbb{Z}_2 \:\oplus\: \mathbb{Z}_2}; \mathbb{Z}_2)= \mathbb{Z}_2[x,y],$$ where both $x$ and $y$ have degree one. The map between the classifying spaces induces a map $$H^*(B_{\mathbb{Z}_2 \:\oplus\: \mathbb{Z}_2}; \mathbb{Z}_2) \rightarrow H^*(B_{\mathbb{Z}_2}; \mathbb{Z}_2)$$ given by $x \rightarrow t, y\rightarrow 0$. In particular, we get a map of fibrations \begin{equation}\label{diagram} \begin{tikzcd} S^m\times\mathbb{C}P^n \arrow{r}{id} \arrow[swap]{d}{} & S^m\times\mathbb{C}P^n \arrow{d}{} \\ P(m,n) \arrow{r}{} \arrow[swap]{d}{}& S^m \times \mathbb{C}P^n/{\mathbb{Z}_2\oplus \mathbb{Z}_2} \arrow{d}{}\\ B_{\mathbb{Z}_2} \arrow {r}{i} & B_{\mathbb{Z}_2\:\oplus\: \mathbb{Z}_2}. \end{tikzcd} \end{equation} Consequently, we obtain a map between the Leray-Serre spectral sequences associated to the Borel fibrations $$i^*: E_2^{p,q}(\mathbb{Z}_2 \oplus \mathbb{Z}_2)\rightarrow E_2^{p,q}(\mathbb{Z}_2),$$ where by $E_2^{p,q}(G)$ we denote $$E_2^{p,q}(G)\cong H^p\big(B_G; H^q(S^m\times \mathbb{C}P^n;\mathbb{Z}_2)\big).$$ By Lemma \ref{lemma}, the action of $ \pi_1(B_G)$ on $ H^*(S^m \times \mathbb{C}P^n;\mathbb{Z}_2) $ is trivial for $G=\mathbb{Z}_2 $ or $\mathbb{Z}_2 \oplus \mathbb{Z}_2$. This implies the system of local coefficients is simple. Hence, we get $$E_2^{p,q}(G)\cong H^p(B_G; \mathbb{Z}_2) \otimes H^q(S^m\times \mathbb{C}P^n;\mathbb{Z}_2).$$ We know that, $$H^*(S^m\times \mathbb{C}P^n; \mathbb{Z}_2) \cong \mathbb{Z}_2[c,d]/\big\langle c^{2}, d^{\hspace{.3mm}n+1} \big\rangle,$$ where $\deg (c)=m$ and $\deg (d) =2$. Let $d'$ be the differential corresponding to $E_2^{p,q}(\mathbb{Z}_2 )$ and $d''$ be the differential associated to $E_2^{p,q}(\mathbb{Z}_2 \oplus \mathbb{ Z}_2)$.\par We first determine all possible non-zero differentials corresponding to the first fibration of diagram (\ref{diagram}). This is done by exploiting the cohomology algebra of $ P(m, n) $. Then we determine all possible non-zero differentials corresponding to the second fibration by comparing maps between the spectral sequences. In particular, this helps to show that the generator $c \in H^*(S^m, \mathbb{Z}_2)$ is transgressive in $E^{*,*}(\mathbb{Z}_2 \oplus \mathbb{Z}_2)$.\par Consider the Leray-Serre spectral sequence associated to $$S^m \times \mathbb{C}P^n \stackrel{}{\hookrightarrow} P(m,n) \stackrel{}{\longrightarrow} B_{\mathbb{Z}_2}.$$ We have $H^*(P(m,n); \mathbb{Z}_2) \cong \mathbb{Z}_2[a,b]/\big\langle a^{m+1}, b^{n+1} \big\rangle$, where $\deg (a)=1$ and $\deg (b) =2$. From this we observe that the differentials in $E^{*,*}(\mathbb{Z}_2)$ are as follows: $$d'_{m+1}(1\otimes c)= t^{m+1} \otimes 1 \hspace{2mm} \textrm{and }\hspace{2mm} d'_{r}(1\otimes d)=0 \textrm{ for all $r\geq 2$}.$$ Then by the naturality of the Leray-Serre spectral sequence we get $$ d''_{m+1}(1\otimes c) = i^* d''_{m+1}(1\otimes c)=d'(1\otimes c) = t^{m+1} \otimes 1.$$ From this it follows that the generator $c\in H^*(S^m; \mathbb{Z}_2)$ is transgressive in $E^{*,*}(\mathbb{Z}_2 \oplus \mathbb{Z}_2)$. This gives $$d''_{m+1}(1\otimes c) = f,\hspace{2mm} \textrm{where } f \in H^{m+1}(B_{\mathbb{Z}_2 \oplus \mathbb{Z}_2}; \mathbb{Z}_2)$$ is a non-zero homogeneous polynomial containing the term $ x^{m+1} $. Therefore, $ f $ must be of the following form: $$f= \big(x^{m+1}+ \gamma_1 x^my+\gamma_2x^{m-1}y^2+\dots+\gamma_ny^{m+1}\big) \otimes 1,$$ where $\gamma_i \in \mathbb{Z}_2$. Next we determine the possibilities for $d''_3(1 \otimes d)$. If $d''_3(1 \otimes d)=0$, then $ E^{i,0}_\infty(\mathbb{Z}_2 + \mathbb{Z}_2)$ is non-zero for infinitely many values of $ i $. This contradicts the fact that $S^m \times \mathbb{C}P^n/{\mathbb{Z}_2\oplus \mathbb{Z}_2}$ is a finite-dimensional CW-complex. Thus, \hspace{1mm}$d''_3(1 \otimes d)\neq0$ and this further imposes the condition that $n$ must be odd. Hence, $d''_3(1 \otimes d)=g$, where $g\in H^3(B_{\mathbb{Z}_2 \oplus \mathbb{Z}_2}; \mathbb{Z}_2)$ is a non-zero homogeneous polynomial which does not contain the term $ x^3 $. Therefore, it is of the following form: $$g= \big(y^3+ \beta_1 xy^2+\beta_2x^2y\big) \otimes1. $$ Thus, we get that $ 1 \otimes d $ transgresses to $ g $. Recall that, transgression operator commutes with the Steenrod operations, which implies $$ d_5''(1\otimes d^2)=0. $$ Using $ g $ we can reduce $ f $ to the following simpler form: $$f= \big(x^{m+1}+ \alpha_1 x^my+\alpha_2x^{m-1}y^2\big) \otimes1, $$ where $ \alpha_i \in \mathbb{Z}_2 $. Following all the preceding arguments we get that the spectral sequence $E^{*,*}(\mathbb{Z}_2 \oplus \mathbb{Z}_2)$ degenerates at $E_{m+2}$-page. Hence, $$ E_\infty^{*,*} \cong E_{m+2}^{*,*}.$$\par We see that $x \otimes 1 \in E_2^{1,0}$ and $y \otimes 1 \in E_2^{1,0}$ are permanent cocycles. Let $i^*$ and $\pi^*$ be the edge maps of the spectral sequence. Let $\tilde{x}=\pi^*(x) \in E_\infty^{1,0} \subseteq H^1\big(S^m \times \mathbb{C}P^n/{\mathbb{Z}_2\oplus \mathbb{Z}_2}; \mathbb{Z}_2\big) $ be determined by $x \otimes 1 \in E_2^{1,0}$. Similarly, let $\tilde{y}=\pi^*(y) \in E_\infty^{1,0}\subseteq H^1(S^m \times \mathbb{C}P^n/\mathbb{Z}_2 \oplus \mathbb{Z}_2;\mathbb{Z}_2)$ be determined by $y \otimes 1 \in E_2^{1,0}$. Then we get $ \pi^*(f)=\tilde{f}=0$ and $ \pi^*(g) =\tilde{g}=0$. Note that, $1 \otimes d^2 \in E_2^{0,4}$ is a permanent cocycle and it determines an element in $ E_{\infty}^{0,4}$. Choose $\tilde{z} \in H^4\big(S^m \times \mathbb{C}P^n/{\mathbb{Z}_2\oplus \mathbb{Z}_2};\mathbb{Z}_2\big)$ such that $i^*(\tilde{z})=d^2$. Since $i^*(\tilde{z } ^{\frac{n+1}{2}})=0$, we get $\tilde{z } ^{\frac{n+1}{2}}+h=0 $, where $ h \in \mathbb{Z}_2[\tilde{x}, \tilde{y}, \tilde{z }]$ is either a zero polynomial or it is a homogeneous polynomial of degree $ 2n+2 $ with the highest power of $ \tilde{z} $ is less than or equal to $ \frac{n-1}{2} $. So, in this case, we obtain \begin{align}\label{z2+z2-coho} H^*(M;\mathbb{Z}_2) \cong \mathbb{Z}_2[\tilde{x},\tilde{y},\tilde{z}]/\big\langle \tilde{f}, \tilde{g}, \tilde{z}^{\frac{n+1}{2}}+h\big\rangle, \end{align} where $\deg(\tilde{x})=1$, $\deg(\tilde{y})=1$ and $\deg(\tilde{z})=4$. Equations (\ref{z4-coho}) and (\ref{z2+z2-coho}) give all possible mod 2 cohomology algebra of orbit spaces of arbitrary free involutions on Dold manifolds. \end{proof} \medskip \begin{remark} Recall that, $ P(m,0)= \Bbb RP^m $ and $ P(0,n)= \Bbb CP^n $. The orbit space of a free involution on $\Bbb RP^m $ is the lens space as any arbitrary free involution on $\Bbb RP^m $ lifts to a free $ \Bbb Z_4 $-action on $ S^m $ and its cohomology algebra is consistent with Theorem \ref{coho-dold} (1). If $ n $ is odd, then the cohomology algebra of the orbit space of a free involution on $\Bbb CP^n $ is \begin{equation}\label{equ3} H^*( \mathbb{C}P^n/ \mathbb{Z}_2; \mathbb{Z}_2) \cong \mathbb{Z}_2[x,y]/\big\langle x^{3}, y^{\hspace{.3mm}\frac{n+1}{2}} \big\rangle, \end{equation} where $\deg (x)=1$ and $\deg (y) =4$ \cite[Corollary 4.7]{msinghprojective}. This is consistent with Theorem \ref{coho-dold} (2). \end{remark} \begin{remark} Conclusions of Theorem \ref{coho-dold} for $ m=1 $ are consistent with \cite[Theorem 1]{morita}. \end{remark} \begin{remark} Theorem \ref{coho-dold} implies that there are at least two conjugacy classes of free involutions on Dold manifolds. We note that, for an arbitrary free involution on a Dold manifold, the induced action on mod 2 cohomology is always trivial in the following cases: \begin{enumerate} \item[](i) $n$ is of the form $4k+1$ and $ m\geq4 $, \item[](ii) $n$ is even, \item[] (iii) $m\geq2n+2$. \end{enumerate} \end{remark} \begin{example} Suppose both $ m$ and $n $ are odd. Let $ g $ be a generator of $ \Bbb Z_4 $. Define the action of $ \mathbb{ Z}_4 $ on $ S^m \times \mathbb{C}P^n $ by $$g\Big((x_0, x_1, \dots, x_{m-1}, x_{m}), \,[z_0, z_1, \dots, z_{n-1}, z_{n}]\Big) =\Big((-x_1,x_0,\dots,-x_{m},x_{m-1}),\,[\overline{z}_1,\overline{z}_0,\dots,\overline{z}_n,\overline{z}_{n-1}]\Big).$$ Let $ \mathcal{Z} $ denote the orbit space. Further, the preceding action induces a free involution on the Dold manifold $ P(m,n) $ with the same orbit space $ \mathcal Z $. The projection $S^m \times \mathbb{C}P^n \rightarrow S^m$ induces the map $p:\mathcal{Z}\rightarrow L^m(4,1)$, which is a locally trivial fiber bundle with fiber $\mathbb{C}P^n$. Notice that, the system of local coefficients is trivial. Since the projection $ \pi $ has a section, the Leray-Serre cohomology spectral sequence associated to the fiber bundle $$\Bbb C P^n\stackrel{i}{\hookrightarrow} \mathcal{Z} \stackrel{\pi}{\longrightarrow}L^m(4,1) $$ degenerates at the $ E_2 $-page. Consequently, the map $ i^*:H^*(\mathcal{Z};\mathbb{Z}_2)\to H^*(\mathbb{C}P^n;\mathbb{Z}_2) $ is onto. Thus, the fiber is totally non-homologous to zero in $ \mathcal{Z} $ with respect to $ \mathbb{Z}_2 $ (see \cite[Chap. III, 7]{serre}). Then by \cite [Chap. III, Prop. 9]{serre}, we get $$ H^*(\mathcal{Z};\mathbb{Z}_2)\cong \mathbb{Z}_2[x, y, z]/\big\langle x^{2}, y^{\hspace{.3mm}\frac{m+1}{2}}, z^{n+1}\big\rangle,$$ where $\deg (x)=1$, $\deg (y) =2$ and $\deg (z)=2$. This realizes Theorem \ref{coho-dold} (1). \end{example} \medskip \begin{example} Suppose $n $ is odd. Let $ a $ and $ b $ be generators of $ \mathbb{Z}_2 \oplus \mathbb{Z}_2 $. Define a free $ \mathbb{Z}_2 \oplus \mathbb{Z}_2 $-action on $ S^m \times \mathbb{C}P^n $ as follows $$a\Big((x_0, x_1, \dots, x_{m-1}, x_{m}), \,[z_0, z_1, \dots, z_{n}]\Big) =\Big((-x_0, -x_1, \dots, -x_{m-1}, -x_{m}),\,[z_0, z_1, \dots, z_{n}]\Big)$$ and $$b\Big((x_0, x_1, \dots, x_{m}), \,[z_0, z_1, \dots, z_{n-1}, z_{n}]\Big) =\Big((x_0, x_1, \dots, x_{m}),\,[-\overline{z}_1, \overline{z}_0, \dots, -\overline{z}_n, \overline{z}_{n-1}]\Big).$$ Let $\mathcal T$ be the quotient of the above action. Note that, the preceding action induces a free involution on the Dold manifold with the same quotient space $ \mathcal{T} $. The projection $S^m \times \mathbb{C}P^n \rightarrow S^m$ induces the map $p:\mathcal T\rightarrow \Bbb R P^m$, which is a locally trivial fiber bundle with fiber the orbit space of $ \mathbb{C}P^n $ by the $ \mathbb{Z}_2 $-action given by $$\big\{[z_0,z_1,\dots,z_n]\mapsto [-\overline{z}_1,\overline{z}_0,\dots,-\overline{z}_n,\overline{z}_{n-1}]\big\}.$$ By equation (\ref{equ3}), we have $$H^*( \mathbb{C}P^n/ \mathbb{Z}_2; \mathbb{Z}_2) \cong \mathbb{Z}_2[x,y]/\big\langle x^{3}, y^{\hspace{.3mm}\frac{n+1}{2}} \big\rangle,$$ where $\deg (x)=1$ and $\deg (y) =4$. It can be seen that the system of local coefficients is trivial. Since the projection $ \pi $ has a section, the Leray-Serre cohomology spectral sequence associated to the fiber bundle $${\Bbb C P^n/ \mathbb{Z}_2}\stackrel{i}{\hookrightarrow} \mathcal T \stackrel{\pi}{\longrightarrow} \Bbb R P^m ,$$ degenerates at the $ E_2 $-page. Consequently, the map $ i^*: H^*(\mathcal T;\mathbb{Z}_2)\to H^*({\mathbb{C}P^n/\mathbb{Z}_2};\mathbb{Z}_2) $ is onto. Thus, the fiber is totally non-homologous to zero in $ \mathcal T $ with respect to $ \mathbb{Z}_2 $ (see \cite[Chap. III, 7]{serre}). Then by \cite [Chap. III, Prop. 9]{serre}, we get $$ H^*(\mathcal T;\mathbb{Z}_2)\cong \mathbb{Z}_2[x, y, z]/\big\langle x^{3}, y^{\hspace{.3mm}\frac{n+1}{2}}, z^{m+1}\big\rangle,$$ where $\deg (x)=1$, $\deg (y) =4$ and $\deg (z)=1$. This realizes Theorem \ref{coho-dold} (2). \end{example} \bigskip Since the mod 2 cohomology algebras of $\mathbb{R}P^m \times \mathbb{C}P^n$ and $P(m,n)$ are isomorphic, and both have the same universal cover $S^m \times \mathbb{C}P^n$, the above proof also works for $\mathbb{R}P^m \times \mathbb{C}P^n$. Hence, we get the following corollary. \begin{corollary} Let $G=\mathbb{Z}_2$ act freely on $\mathbb{R}P^m \times \mathbb{C}P^n$ with a trivial action on mod 2 cohomology. Then $H^*\big((\mathbb{R}P^m \times \mathbb{C}P^n)/G; \mathbb{Z}_2\big)$ is isomorphic to one of the following graded algebras: \begin{enumerate} \item $\mathbb{Z}_2[x,y,z]/ \big\langle x^2, y^{\frac{m+1}{2}}, z^{n+1} \big\rangle,$\\ where $\deg(x)=1$, $\deg(y)=2$, $\deg(z)=2$ and $m$ is odd; \item$\mathbb{Z}_2[{x},{y},{z}]/\big\langle {f}, {g}, {z}^{\frac{n+1}{2}}+h\big\rangle,$\\ where $\deg({x})=1$, $\deg({y})=1$, $\deg({z})=4$, $n$ is odd, $$f= \big(x^{m+1}+ \alpha_1 x^my+\alpha_2x^{m-1}y^2\big) \hspace{2mm} \textrm{and}\hspace{2mm} g= \big(y^3+ \beta_1 xy^2+\beta_2x^2y\big) ,$$ where $\alpha{_i}$, $\beta_j\in \mathbb{Z}_2$, and $ h \in \mathbb{Z}_2[{x}, {y}, {z }]$ is either a zero polynomial or it is a homogeneous polynomial of degree $ 2n+2 $ with the highest power of $ {z} $ is less than or equal to $ \frac{n-1}{2} $. \end{enumerate} \end{corollary} \medskip \section{Application to equivariant maps}\label{sec6} In this final section, by using the cohomology structure of $ H^*(P(m,n)/\mathbb{Z}_2;\mathbb{Z}_2) $, we determine when there exists a $ \mathbb{Z}_2 $-equivariant map from $ S^n $ to $ P(m,n) $. Let $X$ be a compact Hausdorff space with a free involution and consider the antipodal involution on ${S}^n$. Conner and Floyd \cite{conner-floyd} defined the index of the involution on $ X $ as $$\textrm{ind}(X) = \max ~ \{~ n ~|~ \textrm{there exists a}~ \mathbb{Z}_2 \textrm{-equivariant map}~ {S}^n \to X \}.$$ Using Borsuk-Ulam theorem one can see that ind($S^n)=n$. Using Characteristic classes we can derive some upper bound for the ind$ (X) $. Let $w \in H^1(X/G; \mathbb{Z}_2 )$ be the Stiefel-Whitney class of the principal $G$-bundle $X \to X/G$. Conner and Floyd also defined $$\textrm{{co-ind}}_{\mathbb{Z}_2}(X) = \max ~\{~n~|~ w^n \neq 0 \},$$ and showed that \cite[(4.5)]{conner-floyd} $$\textrm{ind}(X)\leq \textrm{{co-ind}}_{\mathbb{Z}_2}(X).$$\par Our final result is the following: \begin{proposition} If $ n $ is even, then there is no $\mathbb{Z}_2$-equivariant map ${S}^k \to P(m,n)$ for $k \geq 2$. \end{proposition} \begin{proof} Let $ X=P(m,n) $. Take $ f $ a classifying map $$f : X/\mathbb{Z}_2 \to B_{\mathbb{Z}_2}$$ for the principal $\mathbb{Z}_2$-bundle $X \to X/\mathbb{Z}_2$. Consider the Borel fibration $X\stackrel{}{\hookrightarrow} X_{\mathbb{Z}_2} \stackrel{\pi}{\longrightarrow} B_{\mathbb{Z}_2}$. We know for free actions $ X_{\mathbb{Z}_2} \simeq X/\mathbb{Z}_2 $. Let $\eta: X/\mathbb{Z}_2 \to X_{\mathbb{Z}_2}$ be the homotopy equivalence. Then $\pi \circ \eta : X/\mathbb{Z}_2 \to B_{\mathbb{Z}_2}$ also classifies the principal $\mathbb{Z}_2$-bundle $X \to X/\mathbb{Z}_2$, and hence it is homotopic to $f$. Therefore, it is sufficient to consider the map $$\pi^*: H^1(B_{\mathbb{Z}_2}) \to H^1(X_{\mathbb{Z}_2}).$$ The image of the Stiefel-Whitney class of the universal principal $\mathbb{Z}_2$-bundle $\mathbb{Z}_2 \hookrightarrow E_{\mathbb{Z}_2} \longrightarrow B_{\mathbb{Z}_2}$ is the Stiefel-Whitney class of $X \to X/\mathbb{Z}_2$. Since we have assumed that $ n $ is even, only case (1) of Theorem \ref{coho-dold} is applicable. It follows from the proof of Theorem \ref{coho-dold} (1) that $x \in H^1(X/\mathbb{Z}_2)$ is the Stiefel-Whitney class with $x \neq 0$ and $x^2=0$. This gives $\textrm{{co-ind}}_{\mathbb{Z}_2}(X) = 1$ and $\textrm{ind}(X)\leq 1$. Hence, there is no $\mathbb{Z}_2$-equivariant map ${S}^k \to P(m,n)$ for $k \geq 2$. \end{proof} \ack{I thank my Ph.D. advisor Dr. Mahender Singh for suggesting the problem and for many valuable discussions and corrections concerning this paper. I also thank UGC-CSIR for the financial support towards this work.} \bibliographystyle{plain}

e00bfbf47ae8ad96c4a0037559e7751ee6d990b9

adc83b19e793491b1c6ea0fd8b46cd9f32e592fc

\section{Introduction} Since the advent of commercially available, affordable, automated (computer controlled), basic fault--tolerant and user-friendly superconducting quantum interference device (SQUID) integral magnetometers in the early 90s, sensitive SQUID-based magnetometry has now established itself as an indispensable everyday characterization and experimental tool in modern science. It has become routinely performed on a wide range of different types of samples such as ultrathin films \cite{Ney:2001_EPL,Sawicki:2010_NP,Chibal:2016_SR,Gladczuk:2017_JPDAP,Hayassi:2018_APE,Sawicki:2018_PRB}, nanoparticles \cite{Sandaresan:2006_PRB,Sadowski:2011_PRB,Sueli:2016_JPCC,Rath:2011_JMMM,Augustyns:2017_PRB}, nanowires \cite{Siusys:2014_NL,Sadowski:2017_Nanoscale}, nanocomposites \cite{Peddis:2008_JPCC}, quantum dots \cite{Sun:2017_QM}, graphene \cite{LiuYuan:2013_SR,LiuYuan:2016_NC}, dilute magnetic \cite{Pereira:2011_JPCM,Sawicki:2013_PRB,Pereira:2013_JPCM,Henne:2016_PRB,NeyV:2016_PRB}, and ferromagnetic semiconductor epilayers \cite{Sawicki:2006_JMMM,Gas:2018_JALLCOM}. It currently enters the realms of organic spintronics \cite{Bujak:2013_CSR}, biology \cite{Kopani:2015_BioMetals}, and topological matter \cite{Zhao:2014_NM,Dutta:2017_SR}. Despite this broad range of subjects, the investigated materials share one crucial common aspect: the objects of interest come on bulky substrates, or at least have to be fixed to a kind of rigid carrier, permitting mounting them onto adequate sample holders suitable for withstanding measurement cycles in wide ranges of temperature $T$ and magnetic field $H$. However, the most popular multipurpose commercial SQUID magnetometers are not object- or element-selective probes, so the sought signal is very often (deeply) buried in the magnetic response of the ("nonmagnetic") carrier. The latter becomes dominant in the mid- to strong-magnetic-field range, when the signal from the investigated sample usually saturates, but that of the carrier continues to grow. To make things worse the signals exerted by such complex objects are detected and processed with some limited accuracy by the magnetometer hardware and software and finally can be, frequently, mishandled by a less experienced end user. Here, a lack of a proper laboratory procedure which may lead to magnetic contamination \cite{Abraham:2005_APL,Garcia:2009_JAP,Pereira:2011_JPDAP,Sawicki:2011_SST,Ney:2011_SST} and other experimental artifacts \cite{Ney:2008_JMMM,Pereira:2017_JPDAP,Buchner:2018_APL} has to be underlined. The latter can be sizably enlarged when the operator lacks the general recognition of how the measurement process and data processing and reduction is carried out \cite{Sawicki:2011_SST}. It has to be noted here that mitigating the commonly met problems, which is an indispensable step in responsible (SQUID) magnetometry, does not necessarily lead to the reduction of the \textsl{real} magnitude of the error bar towards the typically declared sensitivity by the manufacturers (usually around $10^{-8}$~emu). Such a precision indeed can be required in studies of very thin layers of magnetically diluted compounds or antiferromagnets, particularly when their magnetic anisotropy is targeted. The same challenge is faced when the critical exponents specific to magnetic phase transitions in very thin films are to be determined \cite{Stefanowicz:2013_PRB,WangMu:2016_PRB}. It has been established, following the best guidance and rules put forward to date in the literature, \cite{Stamenov:2006_RSI,Ney:2011_SST,Sawicki:2011_SST,Pereira:2017_JPDAP} that the most credible assessments of the sought moment $m_{\mathrm{X}}$ can be obtained upon a subtraction of the results of two separate measurements: that of the sample $m_{\mathrm{S}}$ and of its support $m_{\mathrm{R}}$, usually the bare substrate or the carrier. However, because both responses are mostly composed of the signal from the support, both the minuend and the subtrahend are very large numbers of comparable magnitude. The key point here is that the credibility of a difference between two such numbers directly depends on the absolute precision and accuracy with which \textsl{both} these numbers were established. Therefore, even minute magnetometer instabilities occasionally affecting the results of the measurement process make the required subtraction substantially erroneous. This in turn degrades the scientific soundness of the established results - frequently without issuing any clear warning sign to the researcher. In this report we put forward a carefully elaborated and experimentally validated experimental approach which increases the reliability of the above mentioned subtraction method by sizably reducing the vulnerability of the final results to the randomly appearing disturbances in the magnetometer's output. The provided below exact formulae allows vastly advance both the precision and reliability of the experimentally established signals of very weak magnetic sources which come on bulky substrates. Importantly, this approach is fully compatible with typically employed experimental methodology used on the everyday basis in magnetometry labs. Our report is organized as follows. First, we point out and exemplify the magnitude of the main system instabilities and spurious components contributing to the magnetometers' outputs which cannot be eliminated by the virtue of the subtraction method. Then we introduce the basics of the \textsl{in situ} compensation approach and detail the experimental method up to the point where the final expression for the desired moment is given. The report is ended by presenting examples of the sizable improvement in the consistency of results obtained by employing the compensational approach and by the final conclusions. \section{Equipment-related limitations to precise magnetometry} The study reported here is based on years of experience gained working with Quantum Design Magnetic Property Measurement System (MPMS) SQUID magnetometers. In the authors' view there are two most obvious \textsl{equipment-related} reasons which can affect not only the magnitude of the established magnetic moment $m$, but can also corrupt its dependence on $H$ and $T$, $m(H)$ and $m(T)$, respectively. The first problem arises from a non-perfect performance of the sample transport mechanism during scanning, that is, tripping the sample up and down along the SQUID pick-up coil axis $z$. This is the moment when the experimental response function $V(z)$, i.e. the dependence of the SQUID response \textsl{vs}. sample position, is established. It is important to be reminded here, that the magnitude of $m$ is established by a least-mean-square fitting of a reference $\Upsilon(z)$ function into $V(z)$. $\Upsilon(z)$ describes the response of the magnetometer's sensing coils to the position of an ideal point dipole of a constant magnetic moment along the $z$ axis. However, the MPMS system does not measure the actual position of the sample during scanning. The system only assigns a certain value of $z$ according to the time interval passed from the beginning of the scan. Therefore, any mechanical wear and tear in any of the mechanical components of the sample transport mechanism mars the shape of $V(z)$, and imprints on the magnitude of the established $m$. Obviously mechanical effects accumulate over time, and frequently evade the attention of the user(s). The main distortions of $m(H)$ originate from only approximately correct magnitudes of $H$ which are reported in the magnetometer's output files. This inaccuracy stems from the fact that the commercial SQUID magnetometers do not contain any built-in field sensors. Therefore, instead of the real magnitude of the magnetic field $H_{\mathrm{real}}$ - the field experienced by the sample during the measurement, only the value requested by the user $H_{\mathrm{set}}$ is reported. The point is that $H_{\mathrm{set}}$ only occasionally precisely corresponds to $H_{\mathrm{real}}$. There are many sources of $H_{\mathrm{set}} \leftrightarrow H_{\mathrm{real}}$ discrepancy \cite{Sawicki:2011_SST}. The main one for these considerations seems to be related to a limited accuracy of the digital-to-analog converter which controls the magnitude of the output current $J_H$ of the magnet power supply. In addition, the magnitude of $J_H$ may vary in response to an unstable environment, for example to variations of the ambient temperature in the lab. On the other hand, it has to be underlined that nowadays the available commercial magnetometers are very well engineered and are generally very well suited for the majority of tasks. It is only the necessity to assure a very high precision and reproducibility of $m$ that points to the above mentioned equipment insufficiencies. As shown below their presence precludes a fulfilment of some of the very demanding research tasks without employing more advanced measures to reduce the influence of the above mentioned main weaknesses in the design or performance. In order to exemplify the scale of the challenge addressed here and to show to what extent such issues can be mitigated, the measurements of very thin layers of the dilute magnetic semiconductor (Ga,Mn)N are considered. This magnetic form of GaN in which a small percentage of Ga atoms is randomly substituted by Mn, belongs to increasingly important group of Rashba materials \cite{Stefanowicz:2014_PRB,Adhikari:2016_PRB}. (Ga,Mn)N compounds are typically deposited by molecular beam epitaxy method \cite{Kunert:2012_APL,Gas:2018_JALLCOM} on GaN-buffered (2 -- 3~$\mu$m) sapphire substrates (300 -- 500~$\mu$m thick), whose strong diamagnetic response overwhelmingly dominates the signal of the dilute magnetic layer, particularly at large $H$ and for layers for which the thickness $d$ is in the single nanometers range. The problem of the correct extraction of the layers' signal remains sound even at very low temperatures, where high quality single phase (Ga,Mn)N layers exhibit low temperature ferromagnetic properties \cite{Sawicki:2004_ICPS,Sawicki:2012_PRB,Kunert:2012_APL,Gas:2018_JALLCOM}. The precision is of a great value here since, owing to the $L = 2$ and $S = 2$ configuration of the Mn$^{3+}$ species a field of the order of a few tens of kOe is required to reach the sufficient level of saturation, which would permit an establishment of Mn concentration $x$ with a satisfactory accuracy. It has to be noted that magnetic characterization by SQUID magnetometry is practically the last resort enabling a reasonable estimation of $x$ for a mere nanometre thin and/or magnetically diluted layers ($x < 0.1$). However, caution is required, the precision of the magnetometric determination of $x$ can be compromised by a lack of precise knowledge about the Mn oxidation state. A synchrotron based X-ray absorption near-edge spectroscopy method appears to be the most suitable for this purpose \cite{Stefanowicz:2010_PRB,Bonanni:2011_PRB,Gas:2018_JALLCOM}. To put the things into perspective we note that an expected magnitude of the saturation moment $m_{\mathrm{sat}}$ of a typical $\sim 5 \times 5$~mm$^2$ piece of $d=5$~nm (Ga,Mn)N layer containing about 5\% of Mn is about $5\times 10^{-6}$~emu or 5~$\mu$emu. Therefore, to stay below a decent 10\% relative error bar for the Mn concentration $x$ (it is the accuracy which is typically required as feedback information to the sample growers) the magnitude of $m_{\mathrm{sat}}$ in such layers has to be established with the absolute accuracy of better than 0.5~$\mu$emu, well above the manufacturer's declared floor level of 0.01~$\mu$emu. However, this simple evaluation does not mean one is safe here. On the contrary, the situation changes completely when two additional factors are taken into account. First is the presence of a bulky substrate, second that a large range of strong fields, $40 < H < 70$~kOe, is needed to assess $x$ reasonably well \cite{Bonanni:2011_PRB,Gas:2018_JALLCOM}. At this $H$-range a relevant 0.33~mm thick sapphire substrate exerts a flux corresponding to about -1000~$\mu$emu, so $m_{\mathrm{S}}(H)$ of both the whole sample (the layer plus the substrate) and of the bare substrate [the reference $m_R(H)$] must be established with the same absolute accuracy better than 0.5~$\mu$emu. This actually means that both these dependencies must be known with an absolute precision exceeding 0.05\%, or 3 1/2 digits. Evidence is given below that a typical MPMS system is not capable of maintaining such a stability of its output in a time frame sufficiently long to permit completion of the whole set of the required measurements. Far more severe constraints are imposed by studies of the magnetic anisotropy. It is of a single ion origin in (Ga,Mn)N and it does not "saturate" at any field attainable in SQUID magnetometers \cite{Stefanowicz:2010_PRB,Sztenkiel:2016_NC}. It actually means that at low temperatures magnetic moments established along both hard and easy axes (perpendicular and in-plane orientations, respectively) do not fall on top of each other even at $H = 70$~kOe - the maximum value of $H$ attainable in commercial SQUID magnetometers. Yet, the magnitude and the field dependence of the magnetic anisotropy provides an indispensable reference for the theoretical description of the magnetism in this \cite{Stefanowicz:2010_PRB,Sawicki:2012_PRB,Sztenkiel:2016_NC} or similar material systems \cite{Sawicki:2013_PRB}. In this case the magnitudes of the high field moments measured in both these orientations should be known with an absolute accuracy better than 0.1~$\mu$emu, or full 4 digits. \begin{figure}[t] \includegraphics[width=0.9\columnwidth]{fig_1-SzBrA-v5-Aver300i2K.pdf} \caption{\label{fig:1-SzBrA}% Magnitudes of typical instabilities of MPMS output exemplified by a comparison of the $m(H)$ dependencies obtained for the same reference $5 \times 5 \times 0.33$~mm$^3$ sapphire sample at 300 and 2~K, panels (a) and (b), respectively. The plotted differences $\Delta m(H) = m(H) - m_{avg}(H)$ are calculated for each $m(H)$ taking an average $m_{avg}(H)$ of all the $m(H)$ presented on the panel as the reference. Curves A -- E were obtained in a short time frame (one month), curve F a year later, and the curve G after two years from the first measurements. These measurements have been made using conventional sample holders, identical to that shown in Fig.~\ref{fig:CSH}~(a). The arrows indicate events of sudden shifts of $m$ ("jumps"). For the clarity of presentation the corresponding standard deviations $\sigma$ for these points are given in separate panels: (c) for 300 K and (d) for 2 K. The sizable magnitude of $\sigma$ at weak field is connected with a strong magnetic flux creep and escape in the superconducting magnet \cite{QD1070-207-MST}. The same $m(H)$ are presented in a different manner in Supplementary Figure S1.}% \end{figure} The scale of the typical output instabilities of a well maintained MPMS systems is exemplified in Figure~\ref{fig:1-SzBrA}, where results of $m(H)$ measurements of \textsl{the same} reference piece of a sapphire substrate acquired within a period of nearly two years are compared. Plotted in panels (a) and (b) are the differences $\Delta m(H) = m(H) - m_{avg}(H)$ calculated for each $m(H)$ taking an average $m_{avg}(H)$ of all the $m(H)$ from each panel as the reference. The aim of all these measurements was to validate, and, as it turned out -- to update, the substrate reference $m_{\mathrm{R}}(H)$, which must be subtracted from the samples' $m_{\mathrm{S}}(H)$ in order to establish $m_{\mathrm{X}}(H)$. As the data clearly illustrate for both $T = 2$ and 300~K, the observed instabilities ($\Delta m \simeq 1$~$\mu$emu) at high field only barely allow for basic magnetic characterization of very thin dilute films, and completely preclude investigations of their magnetic anisotropy. Another troublesome element of the results presented in Figure~\ref{fig:1-SzBrA} is the presence of "jumps" or "kinks". The most pronounced ones are indicated by the arrows in panel (a). It has been checked that the scans acquired before and after such an event are technically of the same high quality, which eliminates a systematic distortion of $V(z)$ as the culprit. This is further confirmed by a smooth behaviour of the standard deviation $\sigma$ at these regions of $H$ where the irregularities occur. We therefore assign their presence to minute inaccuracies in the field setting channel of the magnetometer. Actually, to reproduce the reported magnitude of the "jumps" of about 0.5~$\mu$emu at $H \simeq$~-40~kOe, a suddenly appearing difference between $H_{\mathrm{real}}$ and $H_{\mathrm{set}}$ of about 40~Oe is sufficient. This is only about 0.1\% of the intended magnitude of $H$. It has been observed that these two features affecting $m(H)$ come and go with different intensity and that they do not correlate either with a change of the noise level, or with other disturbances in the lab or its vicinity, or with a particular time of the day. Another estimation of the MPMS's output instability comes from a set of systematic measurements of a reference signal source at nominally identical settings. The interrogated object in Figure~\ref{fig:Czas} corresponds very well to the sapphire substrates considered in Figure~\ref{fig:1-SzBrA}. The main message from this two years long test is that although the long time average of the tested MPMS magnetometer stays at reasonable consistent level, at a short time frame, say 7-10 days, the output fluctuates with disappointingly high magnitude of $\Delta m \simeq 1.5$~$\mu$emu at 20~kOe and possibly at 5~$\mu$emu after extrapolation to 70~kOe. According to the estimates presented before, both these values are \textsl{prohibitive} for the reliable establishment of minute magnetic signals basing only on the subtraction approach. Two important conclusions can be drawn from these tests. Firstly, that in the subtraction method one cannot rely on a "universal background curve". Particularly, when the investigations are carried over an extended time frame. Secondly, that this is the magnitude of the output instability that is responsible for the sizable reduction of the attainable sensitivity of the SQUID magnetometry, even if it is performed in accordance with the best known experimental protocols. These two points immediately eliminate from consideration the automated background subtraction method provided by MultiVu software, since it also establishes the magnitude of $m_{\mathrm{X}}$ upon a difference of, essentially, same $m_S$ and $m_R$ measurements. It has therefore become obvious that in order to push down the limits of reliable magnetometry in the existing magnetometers one has to significantly reduce the contribution from the carrier of the researched object. However, for a day-to-day characterization and related basic research the mechanical or chemical thinning of sapphire down to the single $\mu$m range has not been a feasible option. Therefore, a method of an active elimination of the substrate signal remained the only viable and affordable solution. In the following parts of this report the basic concept of an active, \textsl{in situ} type of, substrate compensation method is presented, detailed, quantified and put to the test. \begin{figure}[t] \includegraphics[width=0.9\columnwidth]{fig_2-Czas-2.pdf} \caption{\label{fig:Czas} % MPMS SQUID magnetometer instability. All the measurements have been taken at nominally identical conditions ($T = 300$~K and $H = 20$~kOe, the latter ramped always up from zero) during the centering of one compensational sample holder (CSH) considered in this report [depicted in Figure~\ref{fig:CSH}~(c)]. In this case a 5~mm wide gap between sapphire compensating strips serves as the source of the reference signal. The time scale on the $x$ axis is counted from the date this CSH has been tested for the first time. The double arrow indicates the magnitude of the short time spread of the points (obtained, say, within 1-2 weeks), $\Delta m \simeq 1.5$~$\mu$emu, which matches well the observed spread of $\Delta m$(20~kOe) seen in Figure~\ref{fig:1-SzBrA}~(a). }% \end{figure} \section{Principles and challenges of the \textsl{in situ} compensation} The underlying concept of the \textsl{in situ} compensation of the spurious signal of the substrate arises from the fact that in these magnetometers in which the detection of the magnetic moment relies on (an axial) movement of the sample for a scan length of a few centimeters with respect to the self-balanced sensing coils, any infinitely long shape made of magnetically \textsl{homogeneous} material (even an iron bar) will not produce any signal. This stems from the fact that for each individual fragment of the infinite rod there exist a partner which produces a response of the same magnitude but of the opposite sign. In practice the phrase "infinitely long" can be replaced by "sufficiently long", with the required length becoming shorter the less magnetic is the considered material. Actually, the influence of a length of a specimen on the magnetometer's response can be viewed as an "ends issue". It is illustrated in Fig.~\ref{fig:ES(ends)}. We start our considerations from an "infinitely long" homogenously magnetized rod which is divided in its center into two semi-infinite parts [Fig.~\ref{fig:ES(ends)}~(a)]. In the main panel of Fig.~\ref{fig:ES(ends)} we plot the calculated SQUID responses $V(z)$ of each individual half-rod, taking $z = 0$ as the center of the sensing coils (red and black solid lines for the left and right halves, respectively). Throughout the whole study we assume the sensing coils to be arranged in the second order gradiometer fashion, as in the MPMS units, and for the numerical simulation we take the gradiometer's coils dimensions according to MPMS specification~\cite{QD1014-213-MST}. It is clearly seen, that the strongest signal of the individual half-rods are "generated" at the close vicinity to their origin, that is, at the place where the symmetry of the semi-infinite rod is the lowest. Such a resonant-like shape of $V(z)$ is the result of the particular gradiometer configuration and the two extrema existing on each $V(z)$ occur at the positions $z = \pm R$, where $R=9.7$~mm in MPMS magnetometer is the radius of the gradiometer coils. Actually, this is the most general rule for this kind of magnetometry - the strongest signals come from these places where the translational symmetry is reduced. Most importantly, the sum of these two $V(z)$ nulls out everywhere, as expected for an infinite homogenous object. \label{sec:basics} \begin{figure}[t] \includegraphics[width=0.7\columnwidth]{Fig_ES_ends_vfin2.pdf} \caption{\label{fig:ES(ends)} % (Main panel) Simulations of the changes of the signal $V$($z$) picked up by the $2^{nd}$ order gradiometer for the three cases (\textbf{a} -- \textbf{c}) depicted in upper part of the figure. The solid red and black lines in the graph represent individual responses of semi-infinite rods cut from a uniformly magnetized material at position $z = 0$, case (a). If moved together, the resulting $V(z) = 0$ everywhere. By shifting these halves from $z = 0$, by 2 mm against each other - case (b), their responces shift by the same amount, dashed lines in the graph of matching colors, so their sum acquires sizable amplitudes near the center (the blue dashed line), or near the bulge created by the overlapping rods. In fact the case (b) corresponds to the case (a) of an infinite rod with a muff located at its center (outlined in blue). This reflects the standard situation when samples are mounted on a very long sample holder. Indeed, the dashed blue sum corresponds precisely to the computed $V(z)$ of a 4~mm fragment of this rod, i.e. a sample, when it is tripped alone through the pick-up coils (the light blue background for the dashed blue sum). By pulling these halves away from each other, say, again by 2~mm each way from their joining point, case (c), a 4~mm wide gap is formed (green). Its magnetic response (dotted dark green line) can be obtained again as the sum of the individual $V(z)$ of both halves after shifting them to their new starting positions (dashed lines of matching colors). Importantly, this gap-related $V(z)$ has got exactly the same form, but is of an opposite sign to the sample-like $V(z)$ calculated for the case (b). This effective mutual cancelation of cases (b) and (c) constitutes the basis of the active \textsl{in situ} compensation of chunks of supporting material (substrates) accompanying the minute objects of investigations. The dark yellow short-dashed line exemplifies a response of only a 60~mm long rod. It corresponds to the approach exercised in ref.~\cite{Fukumura:2001_APL}. The thin grey solid line indicates the ideal point-dipole $V(z)$ of the material from which the rods are made and its maximum value $V(0)$ serves as the reference level for all the other curves plotted in this graph. In the considerations presented here the effects of a radial extent of the objects on $V(z)$ has been neglected. We touch upon this issue in section S4 of the Supplementary Information. The white background box marks the extent of the typical scanning window, when a high precision of the measurements is required.} \end{figure} By exactly the same token the system picks the signal from the sample - a typical short specimen can be viewed as two symmetry breaking ends located very close to each other (at the sample length distance), whose signals mutually cancel out away from their ends (i.e. the sample). This situation is simulated in the panel (b). Here, by pushing the same halves against each other, say by 2~mm from their origin, a 4~mm long bulge is created in the overlapping area, outlined in blue. And it is only the 4~mm long muff located symmetrically with respect to the center that breaks the translational symmetry of the whole structure, so its flux nulls out in the pick-up coils except only from the muff. The corresponding $V(z)$ for each of the halves are marked in dotted lines of matching colors and their sum, the dotted blue line, indeed nulls everywhere except of the (relatively) narrow region in the vicinity of the muff. Importantly, it is exactly the same $V(z)$ which will be induced by a 4~mm long fragment of this rod, i.e.~a sample, when tripped alone through the pick-up coils. The dotted grey line in Fig.~\ref{fig:ES(ends)} exemplifies the ideal point-dipole $V(z)$ of the material from which the rod is made of and its maximum value $V$(0) serves as the reference level for all the other curves plotted in this graph. And the comparison of the amplitudes of this $V(z)$ and that of the 4 mm long muff explains why in the conventional magnetometry it pays to increase the linear dimensions of the investigated (homogenous) samples - it reduces the effect of flux cancellation. Certainly, the net gain in the effective flux due to extending the length of the specimen has its limit (it has to null again for the infinitely long sample) and reaches its maximum when the length of the sample approaches $R$.\cite{Hayden:2017_RSI} More general considerations concerning the influence of samples' linear dimensions and their alignment can be found elsewhere \cite{Zieba:1993_RSI,Miller:1996_RSI,Stamenov:2006_RSI,Sawicki:2011_SST}. Both halves of the rod can also be pull apart from their joining point at the center, say, by the same 2~mm. It is illustrated in panel (c). In this case a 4~mm gap is formed (marked in green), again sizably breaking the symmetry of the rod near the center. The corresponding $V(z)$ for the separated semi-rods are marked by dashed lines in the graph. As in the previous case, the sum of the two dependencies (green solid line) nulls everywhere except of the (relatively) narrow region close to their ends. Most importantly, this sum corresponds precisely to a $V(z)$ of the symmetry breaking muff, but is of the opposite sign. So, both cancel each other completely. This fact constitutes the essence of the \textsl{in situ} compensation method: a long sample holder made with an adequate material has to posses a gap in which the investigated sample can be nested - ideally filling the gap completely. In this arrangement, the flux of such artificially elongated substrate will get nulled, leaving only the flux of the layers of interest for the detection. The latter can now be established with a far greater precision since the symmetry breaking by the layer is sizably stronger than the symmetry breaking by the elongated substrate. A very elegant attempt to realize the \textsl{in situ} compensation was proposed by T. Fukumura \textsl{et al.} \cite{Fukumura:2001_APL}, in their study of MBE grown ZnO epilayers with Mn. In order to eliminate the magnetic signal from the sapphire substrate, a mere few mm of the central part of about 60~mm long and 5~mm wide sapphire substrate strip was overgrown by about 2~$\mu$m thick (Zn,Mn)O film. As presented in Fig.~\ref{fig:ES(ends)}, in such a geometry only 70 -- 80\% of the substrate contribution could had been compensated. Nevertheless, the arrangement proved sufficient to observe the positive signal due to Mn ions in wide temperature range and in $H$ up to 50~kOe. Although very instructive and somehow effective, this solution is of a little practical value in typical laboratory practice, since most materials are deposited uniformly on large surfaces (substrate availability depending) and small specimens are cleft for further characterization or investigations. A more practical approach was implemented by M. Wang \textsl{et al.} \cite{WangMu:2016_PRB} for accurate studies of the critical region of the Curie transition in (Ga,Mn)As. This compound is grown on GaAs and the investigated $5 \times 5$~mm$^2$ samples were cleft from standard 2" substrates. For the SQUID magnetometry the samples were abutted on either side with strips of GaAs of the same width as the sample. All three pieces were glued to a length of 5~mm wide Si supporting stick. According to those authors both the Si support and the GaAs strips extended much further than the length of the detection coils, making the magnetometer effectively insensitive to the magnetic flux due to the layer's substrate and the Si support. The concept indeed proved very efficient in eliminating the spurious signal of the GaAs substrate. The authors assumed a perfect compensation in their particular case and did not consider a more general situation of a not perfect match of the sample to the abutting strips. This is, however, very likely because of the spread of substrates' thicknesses -- typically up to 30~$\mu$m, or 2 -- 5\% relative, even for wafers originating from the same batch. Such discrepancies could not have produce any noticeable signal in the very low field ($H \leq 300$~Oe) magnetometry exercised in \cite{WangMu:2016_PRB}. This is simply because in such a weak fields region the signals related to a small imbalance fell below the magnetometer's ultimate sensitivity. But in the high--$H$ region the situation changes completely. In section \ref{sec:wzor} we detail the method how to extract the moment of interest in the case of an arbitrary \textsl{in situ} compensation, which works at any $T$ and $H$ within the range available in the magnetometer. Another approach of background subtraction is based on a differential approach \cite{Cabassi:2010_MST}. It strictly relies on a perfectly symmetrical mounting of the sample and the relevant reference. However, the final $m_{\mathrm{X}}$ is obtained from a quite cumbersome six-parameters fitting of the experimental $V(z)$ to the relevant $\Upsilon$($z$) function. The differential approach seems to be a plausible solution for a range of specific samples (of chemical or biological interests), but it requires the slowest possible measurement mode (the "full scan" method) and its feasibility had not been put to scrutiny above 2~kOe, that is well below the full extent of the magnetic field available for experiments. In conclusion of this brief survey we note that neither of the approaches presented above fulfills the conditions required for precision magnetometry in the whole $T$ and $H$ ranges available in the MPMS magnetometers. Below, we provide a full account of the \textsl{in situ} compensation \textsl{scheme} allowing the elimination of the magnetic signal of bulky substrates and other unwanted components, frequently disturbing the integral magnetometry of nanoscale magnetic materials. In practical terms we expand and quantify the previous concept \cite{WangMu:2016_PRB} to the point that (i) makes routine measurements practically immune to the major systems instabilities, whereas, at the same time, (ii) permitting determination of absolute magnitudes of $m_{\mathrm{X}}$ even for samples of not strictly matching shapes and compositions to that of the compensating strips, all achieved with a better precision than in the conventional approach. \section{Implementation of the method} \subsection{Preparation of the sample holder for the \textsl{in situ} compensation} The actual form of the compensating sample holders (CSH) considered in this report are exemplified in Figure~\ref{fig:CSH}~(b-d). Since our main experimental concern has been focused on thin (Ga,Mn)N layers, which are routinely, and most economically, deposited on sapphire substrates, the main designs contain the substrate-compensating strips cut from 2" sapphire wafers, panels (b) and (c) of Figure~\ref{fig:CSH}. By the same token, for demanding studies of the magnetic layers grown on GaP substrates, another CSH presented in Figure~\ref{fig:CSH}~(d) has compensating strips cut from a 2" GaP wafer. It is most advantageous to prepare the compensation strips from the same batch of the wafers on which the investigated material is grown. This guarantees the closest possible levels of magnetic contamination present both in the investigated sample and in the compensating strips, and so a truly effective compensation. Actually, when 2" wafers are available, four ~4.5 cm long strips are needed to be cut to provide nearly full coverage of the available space sideways of the sample (two on each side of the sample). The mechanical constraints in MPMS SQUID magnetometers restrict the total length of the sample holder to less than 20 cm, and such a length, as established by numerical simulations similar to these presented in Fig.~\ref{fig:ES(ends)}, is more than sufficient to compensate most of the typically used diamagnetic substrates without introducing any experimentally noticeable modifications to the sample's $V(z)$, and so to established magnitudes of $m$. \begin{figure}[t] \includegraphics[width=0.9\columnwidth]{Fig_CSHs_v2.pdf} \caption{\label{fig:CSH}% Examples of the sample holders made by the authors: (\textbf{a}) an example of a conventional one made of 1.5 mm wide and about 20 cm long Si strips, which also can be a base for (\textbf{b-d}) assemblies allowing \textsl{in situ} compensation of unwanted flux of either the substrate or a bulk part of the sample. They are customized to work with Quantum Design MPMS magnetometers. In assemblies (b) and (d) the strips of the compensating materials (sapphire and GaP, respectively) differ from the material used for the supporting stick - Si, whereas in the assembly (c) all parts are made of sapphire. Most of the results presented in this report have been obtained using this particular sample holder. }% \end{figure} The compensating strips are then glued in pairs to the supporting stick in such a way to form an opening (a gap) in the middle of the holder of a width, preferably, a bit greater than the typical width of the investigated samples. This is the place where the investigated sample is placed. Somewhat enlarged width of the gap actually easies both mounting and removing the sample from the opening, without compromising the overall performance of the CSH. In fact, Eq.~(\ref{Eq:CQ}) given in section~\ref{sec:wzor} takes care of a lack of the perfect compensation, so a certain flexibility in this point will not do any harm. In the first two examples given in Figure~\ref{fig:CSH} the gaps are about $\sim 5.2 \times 5$~mm$^2$, i.e. a fraction of a mm wider than the typical width of the (Ga,Mn)N samples investigated in the authors' lab. The gap in the last CSH is wider due to different demands of GaP-based layers. It turns out that the role of the material serving as the mechanical support for the whole assembly is of a prime importance if precise $T$-dependent studies are planned. By far the most consistent performance is obtained when the supporting stick is made of the same material as the rest of the assembly. The example of such a CSH is given in Figure~\ref{fig:CSH}~(c), where instead of Si, the most easily available, clean, and affordable material for this purpose, the supporting stick is made of a 2~mm wide and 200~mm long sapphire strip, cut from a $200 \times 100 \times 0.5$~mm$^3$ $R$-plane sapphire plate. A strongly diluted GE varnish is used to firmly join the parts. The whole assemblies are finished with customized standard MPMS drinking straw adaptors [seen on the conventional sample holder in Figure~\ref{fig:CSH}~(a)], enabling an easy connection to the standard MPMS graphite probe. The design suggested by the authors, although quite natural, turns out to be also quite robust. Such assemblies as in Figure~\ref{fig:CSH} endure for quite a long time. The silent feature of Figure~\ref{fig:Czas} is that the interrogated CSH assembly has survived intact more than two years while serving for more than a hundred measurement runs performed usually in the full envelope of the environmental parameters attainable in the MPMS magnetometer. \subsection{Exemplary studies using CSH} \label{sec:exemplary} The main experimental benefit of working with CSH is presented in Figure~\ref{fig:KX}, where we compare unprocessed results of $m(H)$ of the same $d \simeq 5$~nm, $x \simeq 5$\% (Ga,Mn)N layer measured using a conventional sample holder [as seen in Figure~\ref{fig:CSH}~(a)] and with the \textsl{in situ} compensation, that is using a CSH similar to that depicted in Figure~\ref{fig:CSH}~(b). These experimental configurations are sketched in the respective panels of Figure~\ref{fig:KX}. Clearly, the abutting sapphire strips have reduced the flux of the substrate by a factor of $\sim 30$, which corresponds to mechanical thinning of the sapphire substrate down to about 10~$\mu$m - a process feasible but quite cumbersome due to the hardness of sapphire. Importantly, this also means that the instabilities reported in Figs.~\ref{fig:1-SzBrA} and \ref{fig:Czas} have been reduced in the same proportion, \textsl{i.e.} below 0.02~$\mu$emu in this case, opening wide the door for the most demanding experimental research in this field. Further examples of the qualitatively improved consistency of the results are given in Supplementary Figure S2. Another important fact shown there is a five-fold reduction of the experimental $\sigma$, indicating that the \textsl{in situ} compensation reduces significantly the noise level associated with the reduction of $V(z)$ to $m$. The data presented in Figure~\ref{fig:KX}~(b) indicate also that at such high levels of compensation ($\geq 97$\%) a nonlinear $m(H)$ of the (Ga,Mn)N layer is already clearly seen in the unprocessed results yielded by the magnetometer (red circles). However, despite such a strong compensation this $m_{\mathrm{exp}}(H)$ does not show a clear tendency to saturation, yet. Neither do the results obtained for a reference sapphire sample - marked by light-grey squares in the same panel. Importantly, neither does the latter $m(H)$ null out. This is an example of the typical experimental situation, when both the sample and its reference are of different dimensions and weights than is the opening (the gap) formed between the compensating strips. In the case presented here the reference sapphire sample is about 3\% heavier whereas the sapphire substrate of the magnetic layer is about 1.6\% lighter than the gap. So, it has to be said that the $m_{\mathrm{S}}(H)$ and $m_{\mathrm{R}}(H)$ obtained in a CSH, apart from possessing visibly increased experimental fidelity, do not permit one to accurately establish the magnitude of $m_{\mathrm{X}}(H)$. Further data manipulation is required and Eq.~(\ref{Eq:CQ}) introduced in the next section indeed allows to unambiguously establish the absolute magnitude of the $m_{\mathrm{X}}(H)$ specific to the investigated layer. The final results are presented in Figure~\ref{fig:Examples}~(a). \begin{figure}[t] \includegraphics[width=0.95\columnwidth]{Fig-KX_Metoda_v3fin.pdf} \caption{\label{fig:KX}% A general outline of the \textsl{in situ} compensation method, applied to magnetic field $H$ dependent studies. The specimen S investigated at $T = 5$~K consists of a 5~nm thin (Ga,Mn)N layer (marked in red) grown on about 300~$\mu$m thick sapphire substrate (grey). Both panels present the unprocessed magnetometer's output $m_{\mathrm{exp}}(H)$. Panel \textbf{(a)} depicts $m_{\mathrm{exp}}(H)$ when S is mounted on a conventional sample holder (a long Si stick, black). In this case the overwhelming majority of the flux contributing to $m_{\mathrm{exp}}(H)$ comes from the bulky (diamagnetic) substrate and so this part of $m_{\mathrm{exp}}(H)$ transmits all magnetometer performance flaws. When the same specimen is mounted into a compensational sample holder (CSH), i.e. is abutted by strips of matching substrate material, panel \textbf{(b)}, then the substrate's flux is nearly nulled out, so the remaining net flux comes predominantly from the (magnetic) layer of interest (red bullets). Note the 20 times expanded left Y scale with respect to (a). For comparison orange bullets represent the same data set but plotted on the same scale as in panel (a) (right Y scale). Light gray squares depict a reference sapphire sample measured in the same CSH, whereas the empty gray squares mark $m_{\mathrm{exp}}(H)$ of the empty CSH - here the source of the flux is the gap between compensating strips. }% \end{figure} The data selected for presentation in Figure~\ref{fig:KX} clearly demonstrate that even the preparation of a well-balanced CSH, in which the signals related to the magnetometer's instabilities would be reduced as much as 20 to 100 times, still does not guarantee an equivalent advancement in the absolute precision of the studies. Following the examples given in Figure~\ref{fig:KX}~(b) we note that only in the idealistic case of truly 100\% compensation could a single measurement of a sample in a matching CSH be sufficient to establish the moment of the researched magnetic layer $m_{\mathrm{X}}$. In this case $m_{\mathrm{X}} \equiv m_{\mathrm{exp}}$. However, in the real case the long compensating strips are not cut from exactly the same source from which came the substrate of the investigated layer, and the whole CSH assembly is not free from minor structural imperfections. Both these effects can exert a strong negative effect on the accuracy. \subsection{Determination of the magnetic moment} \label{sec:wzor} In order to eliminate these spurious fluxes we recall the subtraction approach. Assuming that both the sample and the reference are measured in \textsl{the same} CSH, then the difference of these moments will be devoid of any common signals, particularly those specific to possible imperfections of CSH, and should yield directly the sought $m_{\mathrm{X}}$. Since this is true for all types of measurements, to make our considerations more general, we spell out this difference as: $L_{\mathrm{X}}$ = $L_{\mathrm{S}} - L_{\mathrm{R}}$, where $L_{\mathrm{S}}$ and $L_{\mathrm{R}}$ stand for a full sets of relevant $T$-, $H$-, or time $t$- dependent measurements performed for the sample and the reference, respectively. $L_{\mathrm{X}}$ denotes the set of results for the subject of the research. Contrary to uncompensated measurements, the sought $L_X$ is established now as a difference of two rather small signals, i.e. comparable to the magnitude of the final $L_{\mathrm{X}}$, so the precision of the process is enhanced greatly. Moreover, since both $L_{\mathrm{S}}$ and $L_{\mathrm{R}}$ are practically devoid of magnetometer instability signals -- their detrimental effect on $L_{\mathrm{X}}$ is also reduced considerably. Furthermore, in the case considered here the difference $L_{\mathrm{S}} - L_{\mathrm{X}}$ is established accurately irrespective of the achieved compensation level. However, it is practically impossible to have so well matched the sample and the reference, without destroying the former to obtain the latter. Moreover, in everyday practice one reference specimen serves to provide the reference signal for a range of corresponding samples. In this case the difference $L_{\mathrm{S}} - L_{\mathrm{R}}$ will contain also a small portion of the signal of the empty sample holder due to the different degrees of compensation. Therefore, a third measurement of the empty sample holder $L_{\mathrm{G}}$ is needed. Practically, $L_{\mathrm{G}}$ is dominated by the signal exerted by the gap formed between the compensating strips (as presented in Fig.~\ref{fig:ES(ends)} and exemplified for the case considered in the previous section in Figure~\ref{fig:KX}~(b) by empty gray squares). In order to quantify the magnitude of sample holder contribution in the case of the non-perfect compensation, $L_{\mathrm{G}}$ has to be scaled down by a factor corresponding to the degree of the "magnetic" filling of the gap by the inserted specimens: $\beta_i =\mu_i \gamma_i/\mu_{\mathrm{G}} \gamma_{\mathrm{G}}$, where $\mu_i$ and $\gamma_{\mathrm{i}}$ are the corresponding masses and size correcting factors for the two specimens and the gap. The correcting factors $\gamma_i$ are needed to recover the proper magnitudes of $L_i$ in accordance with the real geometrical extent of the sources of the relevant fluxes. This is because the MPMS output is provided in the point dipole approximation. Following the notation adopted previously \cite{Sawicki:2011_SST}, MPMS output data have to be divided by the relevant $\gamma$ in order to recover the correct magnitude of $m$. More about the role of $\gamma$ and its relevance is discussed in section S4 of the Supplementary Information. Finally, we remark that $\mu_{\mathrm{G}}$, the mass of the missing material between the compensating strips, assumes \textsl{negative} magnitudes. A method of its precise determination is given in Supplementary Figure S3. Accordingly, the whole correction term taking care of the sample holder contribution to $L_{\mathrm{X}}$ for the case of non-identical shapes and masses of the sample and the reference assumes the form: \begin{equation}\label{Eq:BQ} \beta = \frac{\mu_{\mathrm{R}}\gamma_{\mathrm{R}} - \mu_{\mathrm{S}}\gamma_{\mathrm{S}}}{\mu_{\mathrm{G}} \gamma_{\mathrm{G}}}, \end{equation} and the full expression for $L_{\mathrm{X}}$ is: \begin{equation}\label{Eq:CQ} L_{\mathrm{X}}\gamma_{\mathrm{X}} = L_{\mathrm{S}} - L_{\mathrm{R}} + \beta L_{\mathrm{G}}. \end{equation} The form in which Eq.~(\ref{Eq:CQ}) is given underlines an important fact that it is valid for all types of $T$-, $H$-, $t$- dependent studies performed within the whole envelope of $T$ and $H$ ranges available in the magnetometer. Moreover, it also holds valid when the compensating strips and the substrates of the investigated samples are made of different materials. The differences in the magnetic susceptibilities are accommodated in the experimentally established magnitudes of $L_i$. So, the general form of the equation given here and the whole frame of the elaborated experimental method constitute a major qualitative advancement with respect to the other attempts \cite{Fukumura:2001_APL,Cabassi:2010_MST,WangMu:2016_PRB}. Furthermore, as laboriously verified, all the necessary values of $L_i$ needed to obtain the correct magnitude of $L_{\mathrm{X}}$ come directly from the values reported by the MultiVu software in the corresponding *.dat files. It is worth stressing that there is no need to numerically simulate the expected shape of $\Upsilon$($z$) for each of these complicated experimental configurations. Neither are any numerical fittings required to reduce the experimentally measured $V(z)$ into $m$. However, the latter remark remains valid only as long as the CSH is properly centered. The centering of the sample holder is one of the most important experiment-execution prerequisites, which have to be observed to achieve a truly sizable boost in accuracy and repeatability of the measurements. Here we strongly advise the reader to perform the pre-measurement centering of the \textsl{empty} CSH always at the same $T$ and $H$. The authors' choice has been 300~K and 20~kOe, respectively, the latter always ramped up from $H \simeq 0$. Actually, the data presented in Figure~\ref{fig:Czas} have been collected during the pre-measurement centering of the empty CSH, and the added value brought about by such a consecutive recording of the signal of the bare CSH at the repeatable conditions is a long term monitoring of both of the sample holder and of the condition of the magnetometer. Centering issues, as well as some other practical details concerning the setting of the measurements are given in the Supplementary Information section "Setting of the measurements". It is strongly suggested to collect all $L_i$ measurements along equivalent experimental sequences to assure the same thermal and magnetic history of the specimens and repeatable magnitudes of $H_{\mathrm{real}}$. Finally, we remark that the workload required for this experimental approach is less than is dictated by Eq.~\ref{Eq:CQ}. The weight of the $L_{\mathrm{G}}$ in the resulting magnitude of $L_{\mathrm{X}}$ is marginal and becomes the smaller the more the geometrical dimensions of the samples and the reference correspond to each other. This causes $\beta \rightarrow 0$. Therefore, it is sufficient to measure the full suite of $L_{\mathrm{G}}$ measurements once and repeat them rather infrequently, say even once a year, providing that neither the experimental configuration nor the CSH itself has changed in the meantime. For example, re-doing of the whole suite of $L_{\mathrm{G}}$ measurements, takes in particular care of the very weak trend down in time domain of the magnitude of the CHS response as can be noticed in Figure~\ref{fig:Czas}. The same applies to $L_{\mathrm{R}}$, though in this case more frequent checks are advisable due to the equal importance of $L_{\mathrm{R}}$ and $L_{\mathrm{S}}$ in the determination of $L_{\mathrm{X}}$. In any case, however, any newly performed $T$-, $H$-, or $t$-dependence of the sample calls for well-matched measurements of $L_{\mathrm{ref}}$ and $L_{\mathrm{G}}$. Therefore, this method works particularly well for a range of qualitatively similar samples for which the same suite of $L_{\mathrm{R}}$ measurements is needed. Then the main experimental effort can be directed to measure $L_{\mathrm{S}}$ only, as in the traditional approach to magnetometry. \section{Examples of application} As presented in Figure~\ref{fig:KX}~(b) a single measurement in a CSH is not sufficient to return the real magnitude of $m_{\mathrm{X}}$. In that particular example the second measurement $m_{\mathrm{R}}$ of about 4.6\% heavier reference yields the opposite slope of its $m(H)$. Therefore, their bare difference $L_{\mathrm{S}} - L_{\mathrm{R}}$ will yield $L_{\mathrm{X}}$ of an even greater slope, even if $L_{\mathrm{R}}$ is rescaled according to the relevant magnitudes of $\mu_{\mathrm{S}}\gamma_{\mathrm{S}}$ and $\mu_{\mathrm{R}}\gamma_{\mathrm{R}}$. It is actually the $ \beta L_{\mathrm{G}}$ term in Eq.~(\ref{Eq:CQ}), which, by bringing into consideration the extra flux generated by a not ideally filled gap of the CSH during $L_{\mathrm{S}}$ and $L_{\mathrm{R}}$ measurements, restores the real magnitude of $L_{\mathrm{X}}$, and eventually $m_{\mathrm{X}}$. The first example of the whole method in action is presented in Figure~\ref{fig:Examples}~(a), where magenta circles mark the results obtained by the application of Eq.~(\ref{Eq:CQ}) to the presented in Figure~\ref{fig:KX}~(b) set of $L_{\mathrm{S}}$, $L_{\mathrm{R}}$ and $L_{\mathrm{G}}$ measurements collected during the investigation of a nominally 5~nm thin (Ga,Mn)N layer of an expected Mn content between 3 and 5\%. The correctness of the result is corroborated by the following two notions. Firstly, even at the strongest available fields the established $m_{\mathrm{X}}(H)$ retains just a slightly convex character -- very alike to that of the previously investigated much thicker (Ga,Mn)N layers -- corresponding very closely at high field to the theoretical $m(H)$ for the Mn$^{3+}$ configuration in wurtzite GaN environment \cite{Stefanowicz:2010_PRB,Bonanni:2011_PRB,Sztenkiel:2016_NC}, yielding the Mn content $x$ = 3.4\%, a value close to the technological one. Secondly, the another set of results in Figure~\ref{fig:Examples}~(a) (marked by purple squares) saturates at a very similar level. This independently performed measurement has been obtained for the same sample mounted in the perpendicular orientation (different $L_{\mathrm{S}}$ and $\gamma_{\mathrm{S}}$, but the same $\mu_{\mathrm{S}}$, $L_{\mathrm{R}}$ and $L_{\mathrm{G}}$ and their $\mu_{\mathrm{i}}\gamma_{\mathrm{i}}$). So, both measurements constitute together the first experimental assessment of the magnetic anisotropy in such a very thin ferromagnetic (Ga,Mn)N layers, and this accomplishment would have not been possible without the elaboration of the \textsl{in situ} compensation method and the pertinent data reduction performed within the frame set by Eq.~(\ref{Eq:CQ}). \begin{figure} \includegraphics[width=0.9\columnwidth]{Fig-Examples-MA-ANQ.pdf} \caption{\label{fig:Examples}% \textbf{(a)} Results of in- and out- of plane (bullets and squares, respectively) magnetic measurement of a 5~nm thin (Ga,Mn)N layer obtained using \textsl{in situ} compensation and the data reduction according to Eq.~(\ref{Eq:CQ}). In case of the in plane configuration the $m(H)$ of the sample, reference and the empty compensational sample holder (CSH) are presented in Figure~\ref{fig:KX}~(b). \textbf{(b)} Magnetic responses corresponding to $\alpha$-Fe (open symbols) and $\gamma$'-Ga$_y$Fe$_{4-y}$N nanocrystals (closed symbols) embedded in a GaN matrix established at 2 and 300~K upon measurements in the same CSH as in (a) and data reduction according to Eq.~(\ref{Eq:CQ}). }% \end{figure} Another example, where the precise knowledge of the absolute values of magnetic moment proves fundamental is during investigations of structures expected to contain an antiferromagnetic AFM component, which usually exhibits a marginally weak magnetic response. In this view, CSHs have been used to study ensembles of Fe-rich, $\gamma$'-Ga$_y$Fe$_{4-y}$N nanometer-size nanocrystals (NCs) embedded in a GaN matrix, similar to those structurally characterized previously \cite{Navarro:2012_APL}. An example of the magnetic properties of such ensembles is given in Figure~\ref{fig:Examples}~(b), where magnetic responses corresponding to $\alpha$-Fe NCs (open symbols) and $\gamma$'-Ga$_y$Fe$_{4-y}$N ones (closed symbols) measured at 2 and 300 K are plotted. In this case it is not the establishment of the magnitude of the magnetic saturation of these two ensembles that defines the real technical merit of the result. The magnetic saturation in these easy saturating magnetic systems can be equally accurately established by the traditional approach from the measurements performed at the weak field region ($H < 10$~kOe), i.e. where the flux of substrate is small comparably to that of the NCs and so the magnetometer instabilities do not mar the $m_{\mathrm{X}}$ of the NCs. The real added value of the \textsl{in situ} compensation and the following data manipulation sets by Eq.~(\ref{Eq:CQ}) is the establishment that the magnetic response from these two NCs system does not show any high-field kink or an inclined $m(H)$ suggestive of a spin-flop transition or a spin canting, which are characteristic of AFM systems. On the other hand, the saturation levels established in these measurements allow one to verify other magnetic characteristics of these prospective phase separated materials \cite{NavarroQ:2019_PRB}. It has to be added, that the concept of the \textsl{in situ} compensation proves extremely useful in the studies of a broad range of systems. As an example, we take the determination of the level of magnetically responsive contaminants in bulk materials. In case referred to here the very same sample holder which is depicted in Figure~\ref{fig:CSH}~(b) was used to compensate the bulk diamagnetism in a search for possible superconducting precipitates in samples of topological Pb$_{0.20}$Sn$_{0.80}$Te and topologically trivial Pb$_{0.80}$Sn$_{0.20}$Te \cite{Mazur:2017_arXiv}. By adjusting the masses of the studied Pb$_{1-y}$Sn$_y$Te samples, and after correcting for the low-$T$ sapphire response, the increased resolution of the method allowed to establish that the relative weight of precipitates that could produce a response specific to superconducting Pb or Sn was below 0.1~ppm, a level that goes beyond the state of the art of the integral magnetometry, proving that this method can be employed widely as a very sensitive characterization tool in material science. \section{Conclusions} In this report a thorough method for mitigating signal instability problems in commercially available SQUID-based integral magnetometers has been put forward. The method is based on the \textsl{in situ} magnetic compensation, at the sample holder level, of the vast majority of the dominating unwanted signal of the sample substrate, bulkiness or a carrier, which normally accompanies the minute object of interest. Because the signal which is processed by the magnetometer is typically up to two orders of magnitude smaller, the output is much less dependable on the inevitable fluctuations of some environmental variables, that otherwise detrimentally reduce the real credibility of the outcome in the standard approach to precision magnetometry. In practice a two- to five-fold reduction in the absolute noise level has been observed. Practical solutions for the achievement of adequate compensation and proper expressions to evaluate the final results obtained in the compensational sample holder are given. The universal form of this expression allows to practically employ one design of the compensational sample holder in investigations of a range of specimens characterized by different sizes, shapes and compositions. Importantly, the method does not require any involving modelling of the magnetometer output signal and laborious fitting. All the required inputs to calculate the absolute magnitude of the net moment of the investigated object can be taken directly from the standard magnetometer output files. The method has been implemented in MPMS SQUID magnetometers, but it is by no means limited to this particular system. It is exemplified and put to the test on nanometer thin layers of dilute magnetic semiconductors, without and with embedded nanocrystals. The solution given here is of great relevance to numerous fields of material science (to a broad community), where magnetic investigations are becoming of prime importance: biophysics, organic spintronics, and in further emerging new fields dealing with topological insulators, 3D Dirac semimetals, and 2D materials. \section{Acknowledgements} The authors acknowledge funding from the National Science Centre, Poland through FUGA Grant DEC-2014/12/S/ST3/00549 and OPUS Grants DEC-2013/09/B/ST3/04175 and DEC-2017/27/B/ST3/02470, material issues assistance of Gerd Kunert, Alberta Bonanni, Vim van Roy, and Tomasz Baraniecki, and Andrea Navarro-Quezada (ANQ) for providing samples with embedded magnetic nanocrystals. The authors are further indebted to ANQ, and Andrew Rushford for critical reading of the manuscript. \section{References}

3b3b639b02abb85640e705baea9318297e065a07

\section{Introduction} It is a longstanding question how neutral interstellar gas is ionized. Massive stars and supernova remnants (SNRs) are thought to be the primary sources of gas ionization in galaxies. Their natal gas is rapidly disrupted by powerful UV radiation into the photoionized gas on timescales of $\sim$10 Myr after the formation of massive star clusters \citep[e.g.,][]{1999PASJ...51..745F,2009ApJS..184....1K}. Dense neutral gas within a low-density wind-blown bubble will be evaporated by supernova shocks, if the interacting time is longer than $\sim$$10^5$ yr \citep[e.g.,][]{1977ApJ...218..148M,2018arXiv180410579C}. Owing to their short time scales, it is difficult to examine an evolutionary process from the neutral gas to the ionized gas observationally. SNR N63A (also known as MCSNR 0535$-$6602, LHA~120-N~63A or SNR B0535$-$66.0) provides us an ideal laboratory for studying such evolutions of interstellar gas and the shock-cloud interaction. N63A is one of the brightest SNRs in the Large Magellanic Cloud (LMC), whose size is $81'' \times 67''$ or $\sim$18 pc in diameter assuming a distance of 50 kpc \citep[e.g.,][]{1993AA...275..265D,2003ApJ...583..260W,2017ApJS..230....2B}. The age of the SNR is estimated to be 2000--5000 yr \citep[][]{1998ApJ...505..732H,2003ApJ...583..260W}, indicating that the natal gas may still be associated with N63A. The SNR appears to be embedded within the H{\sc ii} region N63 coincident with the OB association NGC~2030 or LH83 \citep[][]{1988AJ.....96.1874C,1970AJ.....75..171L}. N63A is therefore believed to be the remnant of a massive star in the association \citep[e.g.,][]{1980PASP...92...32V, 1983ApJ...275..592S, 1998ApJ...505..732H}. The core-collapse origin was also confirmed by detailed measurements of Fe K$\alpha$ centroids \citep{2014ApJ...785L..27Y}. The SNR holds an optical nebula (diameter is $\sim6$ pc) within the shell, which comprises three prominent lobes \citep{1983ApJS...51..345M}. The two eastern lobes with high-intensity ratio of [S {\sc ii}]/H$\alpha$ \citep[0.7;][]{2008MNRAS.383.1175P} represent the shock-ionized gas, while the other western lobe with the low-intensity ratio corresponds to the photoionized gas \citep{1995AJ....110..739L}. Furthermore, all optical lobes show molecular shock properties with their near-infrared colors, suggesting that the shocked molecular gas dominates in the SNR \citep{2006AJ....132.1877W}. Subsequent detailed infrared spectroscopy confirmed that shock-excited molecular hydrogen lines are detected in all optical lobes \citep{2012ApJ...761..107C}. The imaging spectroscopy of X-rays also indicates the presence of dense interstellar gas with a mass of at least $\sim$450 $M_{\sun}$ \citep{2003ApJ...583..260W}. The CO clouds associated with the optical nebula are, however, not yet detected in spite of a number of efforts \citep[e.g.,][]{1988ApJ...331L..95C,1993A&A...276...25I,2010AJ....140..584D}. CO observations with the NANTEN2 4 m telescope and the Mopra 22-m telescope have detected a giant molecular cloud (GMC) only in the northeastern edge of the shell, whereas no significant CO emission lines have been detected toward the optical nebula \citep{2001ApJ...553L.185Y,2017AIPC.1792d0038S} In the present paper, we show the first detection of dense molecular clouds associated with the optical nebula in N63A using the $^{12}$CO($J$ = 1--0, 3--2) emission lines with the Atacama Large Millimeter / submillimeter Array (ALMA) and the Atacama Submillimeter Telescope Experiment (ASTE). Morphological Matching of molecular cloud with the optical, X-ray, or radio continuum nebula reveals new information on the origin of the ionized gas in the SNR N63A. Section \ref{sec:obs} describes observations and data reductions of CO, radio continuum, and X-rays. Subsection \ref{subsec:large} gives large-scale views of CO, H{\sc i}, and X-rays; subsections \ref{subsec:large} presents a detailed CO distribution with ALMA; subsections \ref{subsec:xspec} and \ref{subsec:comp} present a X-ray spectral analysis and comparisons of the X-ray absorbing column density with the interstellar medium (ISM). Discussion and conclusions are shown in sections \ref{discussion} and \ref{conclusions}, respectively. In a subsequent paper, we will present a detailed analysis of X-ray spectra for the whole SNR and compare them with CO and other available datasets. \begin{deluxetable*}{l l c c c c l c}[] \tablecaption{ATCA observations of N63A used in this study} \label{tbl-flux} \tablehead{Observing & Frequency & Array &Total Obs. & Band Width &Channels & References and notes \\ Date & (MHz) & Configuration&time (minutes)& (MHz) & &} \startdata \multicolumn{7}{|c|}{pre CABB} \\ \hline 1991 May~02 &4786 &6A &15.7 &128 &33 & \cite{1993AA...275..265D} \\ 1991 May~03 &4786 &6A &--- &128 &33 &\cite{1993AA...275..265D}\\ 1991 May~22 &5746 &1.5B &--- &128 &33 &\cite{1993AA...275..265D}\\ 1991 May~23 &5746--8640 &1.5B &1.3 &128 &33 &\cite{1993AA...275..265D}\\ 1991 May~24 &8640 &1.5B &--- &128 &33 &\cite{1993AA...275..265D}\\ 1992 August~17 &4786--8640 &1.5C &668 &128 &33 &\cite{1993AA...275..265D}\\ 1992 October~20 &4786--8640 &6C &573 &128 &33 &\cite{1993AA...275..265D}\\ 1997 April~06 &4786--8640 &375 &40 &128 &33 & \cite{2017ApJS..230....2B}\\ \hline \multicolumn{7}{|c|}{CABB} \\ \hline 2015 January~01 &5500--9000 &6A &42 &2048 &2049& This work \\ 2015 January~02 &2100 &6A &150 &2048 &2049& This work \\ 2018 March~27 &5500--9000 &EW352 &34 &2048 &2049& This work \\ \enddata \end{deluxetable*} \section{Observations and Data Reductions}\label{sec:obs} \subsection{CO}\label{subsec:co} Observations of the $^{12}$CO($J$ = 3--2) emission line were carried out in 16--24 November 2015 by using ASTE \citep{2004SPIE.5489..763E}, which is operated by the National Astronomical Observatory of Japan (NAOJ). We observed $3' \times 3'$ rectangular region centered at ($\alpha_\mathrm{J2000}$, $\delta_\mathrm{J2000}$) $\sim$ ($05^\mathrm{h}35^\mathrm{m}43$\farcs$7$, $-66\degr02\arcmin11\farcs8$) using the on-the-fly mapping mode with Nyquist sampling. The front end was ``DASH 345" receiver. The digital FX spectrometer ``MAC" \citep{2000SPIE.4015...86S} was used for a back end, whose bandwidth is 128 MHz with 1024 channels, corresponding to the velocity coverage of $\sim$111 km s$^{-1}$ and the resolution of $\sim$0.11 km s$^{-1}$. Typical system temperature is $\sim$300--400 K in the single-side band, including the atmosphere. We observed N159W [($\alpha_\mathrm{J2000}$, $\delta_\mathrm{J2000}$) $\sim$ ($05^\mathrm{h}40^\mathrm{m}3$\farcs$7$, $-68\degr47\arcmin00''$)] \citep{2011AJ....141...73M}, and then we estimate the main beam efficiency of $\sim$0.52. We also checked the pointing accuracy every half-hour to satisfy an offset within $2''$. After applying two-dimensional Gaussian smoothing, the final beam size is to be $\sim$$25''$. The noise fluctuation is $\sim$0.18 K at the velocity resolution of $\sim$0.4 km s$^{-1}$. Observations of the $^{12}$CO($J$ = 1--0) emission line were carried out in 31 January and 27 August 2016 using ALMA Band 3 (86--116 GHz) as a Cycle 3 project $\#$2015.1.01130.S. We utilized the mosaic mapping mode of a $100'' \times 100''$ rectangular region centered at ($\alpha_\mathrm{J2000}$, $\delta_\mathrm{J2000}$) $\sim$ ($05^\mathrm{h}35^\mathrm{m}46$\farcs$37$, $-66\degr02\arcmin04\farcs8$). The observations were conducted by using 38 antennas of the 12-m array. The baseline length ranges from 13.7 to 1551.1 m, corresponding to {\it{u-v}} distances from 4.6 to 596.0 $k\lambda$. The correlator was set up in dual polarization mode with a bandwidth of 58.59 MHz, corresponding to the velocity coverage of 152.5 km s$^{-1}$. Three quasars J0635$-$7516, J0519$-$4546, and J0529$-$7245 were observed as the complex gain calibrator, the flux calibrator, and the phase calibrator, respectively. The data reduction including the calibration was made by the Common Astronomy Software Application \citep[CASA;][]{2007ASPC..376..127M} package version 5.1.0. We utilized the multiscale CLEAN algorithm implemented in the CASA package \citep{2008ISTSP...2..793C}. The beam size of final datsets is $1\farcs93 \times 1\farcs71$ with a position angle of $66\fdg6$, corresponding to the spatial resolution of $\sim$0.4 pc at the LMC distance of $\sim$50 kpc \citep[e.g.,][]{2016A&A...585A.162M,2017ApJS..230....2B}. Typical noise fluctuation is $\sim$0.84 K at a velocity resolution of 0.4 km s$^{-1}$. To estimate the missing flux, we used the $^{12}$CO($J$ = 1--0) datasets obtained with Mopra \citep{2017AIPC.1792d0038S}. In the northeast of the SNR, we compared the integrated intensities of Mopra and ALMA CO data that are smoothed to match the FWHM resolution of $\sim$$45''$. We obtain the missing flux of $\sim$$10\%$ or less, and hence the missing flux is considered to be negligible. \subsection{X-rays}\label{subsec:xrays} We use archived X-ray data obtained with the {\it Chandra X-ray observatory}, for which the observation ID is 777 \citep[PI: Hughes][]{2003ApJ...583..260W}. The data were taken with the Advanced CCD Imaging Spectrometer S-array (ACIS-S3) on 16--17 October 2000. We used the Chandra Interactive Analysis of Observations \cite[CIAO;][]{2006SPIE.6270E..1VF} software version 4.10 with CALDB 4.7.8 for data reduction, imaging, and spectroscopic analysis. The data were reprocessed using the {\it chandra\_repro} procedure. We created energy-filtered, exposure-corrected images using the {\it fluximage} procedure in the energy bands of 0.3--0.6 keV, 0.6--1.1 keV, 1.1--6.0 keV, 4.3--6.0 keV, and 0.3--6.0 keV. The total effective exposure time is 43.4 ks. For the spectral analysis, we used HEASOFT (version 6.24), including spectral fitting with XSPEC (version 12.10.0c). We fit the spectrum in the energy band form 0.4--6.0 keV and the errors of model fit are quoted at 90\% confidence levels. We also used the ATOMDB version 3.0.9. \begin{figure*} \begin{center} \includegraphics[width=\linewidth]{n63a_chandra_radio_large_rgb_v2.eps} \caption{Three-colour image of N63A obtained with (a) {\it Chandra} X-rays (Red: 0.3--6.0 keV, Green: 0.6--1.1 keV, and Blue: 1.1-6.0 keV) and (b) ATCA (Red: 2100 MHz, Green: 4786 MHz, and Blue: 8640 MHz). Superposed contours indicate the broadband X-rays in the energy band from 0.3--6.0 keV, whose contour levels are 0.4, 0.8, 1.6, 3.2, 6.4, and $12.8 \times 10^{-6}$ counts s$^{-1}$ pixel$^{-1}$. (c) Fractional polarisation vectors at 5500~MHz overlaid on 4786~MHz ATCA (pre-CABB) image of N63A. The blue line represents a polarisation vector of 100 percent. We used {\texttt robust=0} weighting scheme to make this image. The peak fractional polarisation value is P = $5\pm1$ percent while the average polarisation is measured to be $\sim$3 percent.} \label{fig0} \end{center} \end{figure*}% \subsection{Radio Continuum}\label{subsec:radio} We make use of archival Australia Telescope Compact Array (ATCA) data obtained from the Australia Telescope Online Archive (ATOA). We analyzed data from projects C058, CX310 and C3229 that were taken in 1991, 1992, 1997, 2015, and 2018. These observations includes pre Compact Array Broadband Backend (pre--CABB) and CABB in various array configuration such as 6A, 1.5B, 1.5C, 6C, 375, and EW352 (for details see Table~\ref{tbl-flux}. The primary (flux density) calibration was done using source PKS~1934--638, while the secondary (phase) calibrators were PKS~0454--810 (in 1991), PKS~0407--658 (in 1992), and PKS~0530--727 (in 2015 and 2018). Data reduction and imaging were accomplished by using \textsc{miriad}\footnote{http://www.atnf.csiro.au/computing/software/miriad/} \citep{1995ASPC...77..433S} and \textsc{karma}\footnote{http://www.atnf.csiro.au/computing/software/karma/} \citep{1997PASA...14..106G} software packages. Images were formed using \textsc{miriad} multi--frequency synthesis \citep{1994A&AS..108..585S} and Briggs weighting of robust~=~0 and 1. They were deconvolved with primary beam correction applied. The same procedure was used for both \textit{Q} and \textit{U} stokes parameters. The pre-CABB images at 4786 and 8640~MHz have a resolution of 2.9\arcsec~$\times$~2.0\arcsec\ and 3.5\arcsec~$\times$~2.7\arcsec. However, our CABB images at all frequencies suffer from the insufficient {\it u-v} coverage but we still manage to achieve reasonable sensitivity and resolutions of 5.7\arcsec~$\times$~4.9\arcsec at 2100~MHz, 2.7\arcsec~$\times$~1.5\arcsec\ at 5500~MHz, and 1.2\arcsec~$\times$~0.82\arcsec\ at 9000~MHz (Table~\ref{tab0}). While our pre-CABB images are of better sensitivity then newer but incomplete {\it u-v} coverage CBB images, the new CABB polarisation images can show good polarisation regions \footnote{Note that ATCA polarisation capicity came online only from mid 1993 i.e. before here presented images at 4786 and 8640~MHz.} \begin{deluxetable}{@{}llllcccc@{}} \tablecaption{Details of ATCA radio continuum images of N63A \label{tab0}} \tablehead{Frequency&& Beam size && RMS ($\sigma$) &&PA\\ (MHz) && (arcsec) && (mJy~beam$^{-1}$) &&(degree)\\ } \startdata 2100 && 5.7$''$~$\times$~4.9$''$ && 0.10 && 0.4 \\ 4786 && 2.9$''$~$\times$~2.0$''$ && 0.31 && 80.4 \\ 5500 && 2.7$''$~$\times$~1.5$''$ && 0.25 && 11.2 \\ 8640 && 3.5$''$~$\times$~2.7$''$ && 0.89 && 78.8 \\ 9000 && 1.2$''$~$\times$~0.8$''$ && 0.11 && 11.4 \\ \enddata \end{deluxetable} \subsection{Astronomical data at the other wavelengths}\label{subsec:other} Optical data (H$\alpha$, [S {\sc ii}], and [O {\sc iii}]) are used to derive the spatial distribution and density of the ionized gas. We utilized the {\it Hubble Space Telescope} (HST) WFPC2 images of N63A, which are downloaded from the Hubble Legacy Archive. The observations were carried out using the F656N (H$\alpha$), F673N ([S {\sc ii}]), and F502N ([O {\sc iii}]) filters on 8 October 1997 and 12 September 2000. The exposure times of H$\alpha$, [S {\sc ii}], and [O {\sc iii}] are $\sim$1000 s, $\sim$1200 s, and $\sim$2400 s, respectively. For further details about the data reductions, see the HST Data Handbooks\footnote{http://www.stsci.edu/hst/HST\_overview/documents/datahandbook/}. \begin{figure*} \begin{center} \includegraphics[width=\linewidth]{n63a_large_scale_rgb_v5.eps} \caption{Integrated intensity maps of (a) ASTE $^{12}$CO($J$ = 3--2) and (b) ATCA \& Parkes H{\sc i} \citep{2003ApJS..148..473K} toward the SNR N63A. The integration velocity range is $V_\mathrm{LSR}$ = 276.8--288.0 km s$^{-1}$. Superposed white contours indicate the {\it Chandra} X-rays in the energy band of 0.3--6.0 keV as same as in Figure \ref{fig0}(b). (c) Enlarged view of the northeastern shell of the SNR inside the dashed box in Figure \ref{fig1}(b). Superposed magenta contours represent the $^{12}$CO($J$ = 3--2) integrated intensity. The lowest contour and contour intervals are 2 K km s$^{-1}$ and 1 K km s$^{-1}$, respectively. We also show the beam size of ASTE CO and ATCA \& Parkes H{\sc i}, and the scale bar in the top right corners.} \label{fig1} \end{center} \end{figure*}% \begin{figure} \begin{center} \includegraphics[width=\linewidth]{n63a_spec_rgb_v2.eps} \caption{Averaged line profiles of $^{12}$CO($J$ = 3--2) ({\it{red}}) and H{\sc i} ({\it{blue}} and {\it{cyan}}) toward the SNR N63A. The blue and cyan profiles are spatially averaged H{\sc i} spectra inside and outside of a circle with a radius of $30''$ as shown in Figure \ref{fig1}(b) (orange circle). The CO spectrum in red is averaged over the region where the integrated intensity is 2 K km s$^{-1}$ or higher ($> 5\sigma$). The velocity integration range used for the images in Figure \ref{fig1} is also shown by dashed lines.} \label{fig2} \end{center} \end{figure}% \begin{figure*} \begin{center} \includegraphics[width=150mm]{n63a_hst_alma_chandra_rgb_v8.eps} \caption{Three-colour image of the SNR N63A. The red, green, and blue colors represent the HST H$\alpha$ \citep{2001AIPC..565..409C}, ALMA $^{12}$CO($J$ = 1--0), and {\it Chandra} X-rays \citep[$E$: 0.3--6.0 keV,][]{2003ApJ...583..260W}, respectively. The velocity range of CO is from 276.8 to 288.0 km s$^{-1}$. We also show the beam size of ALMA CO observation and scale bar in the bottom right corner.} \label{fig3} \end{center} \end{figure*}% \begin{figure*} \begin{center} \includegraphics[width=150mm]{n63a_hst_alma_rgb_v13.eps} \caption{Three-colour images of the SNR N63A observed by HST [Image credit: NASA, ESA, HEIC, and The Hubble Heritage Team (STScI/AURA)]. The red, green, and blue colors represent the [S {\sc ii}], H$\alpha$, and [O {\sc iii}]. Superposed contours indicate the ALMA CO integrated intensities, whose velocity range is from 277.2 to 284.0 km s$^{-1}$. The contour levels are 6, 12, 18, 24, 30, 40, 50, 60, and 70 K km s$^{-1}$. The CO clouds A--K discussed in Section 3.2 are indicated.} \label{fig4} \end{center} \end{figure*}% We also use the archived H{\sc i} data obtained with the Australia Telescope Compact Array (ATCA) \& the Parkes radio telescope \citep{2003ApJS..148..473K}. The combined H{\sc i} image has an angular resolution of $\sim$1$'$, corresponding to the spatial resolution of $\sim$15 pc. Typical noise fluctuations are $\sim$0.3 K at a velocity resolution of 1.56 km s$^{-1}$. \section{Results}\label{sec:results} \subsection{Large-Scale views of X-rays, radio continuum, CO, and H{\sc i}}\label{subsec:large} Figures \ref{fig0}(a) and \ref{fig0}(b) show the three-colour images of X-rays and radio continuum, respectively. The X-ray shell shows an elliptical shape, slightly elongated in the northeastern direction, with a diffuse blow-out structure in southwest. The X-ray hole inside the SNR, also mentioned by \cite{2003ApJ...583..260W}, spatially coincides with the optical nebula at ($\alpha_\mathrm{J2000}$, $\delta_\mathrm{J2000}$) $\sim$ ($05^\mathrm{h}35^\mathrm{m}43$\farcs$1$, $-66\degr01\arcmin59\arcsec$). The radio continuum emission peaks at the same central place where the optical emission is detected and coincide with the feature X-ray hole. North, east and south-east area of N63A is closely follows X-ray emission. We also find clear indication of the steepening of the radio spectral index (redish color) at the south-east limb as well as up in the north, suggesting that the synchrotron radio emission dominates in these regions. However, towards the south-west side of the SNR we didn't detect any radio emission which is probably due to the insufficient sensitivity of our radio images [Figure \ref{fig0}(b)]. A linear polarisation image of N63A at 5500 MHz was created using the \textit{Q} and \textit{U} Stokes parameters and is shown in Figure \ref{fig0}(c). However, no reliable polarisation images could be created at 9000~MHz, due to the low signal--to--noise ratio caused by poor $uv$ coverage. The fractional polarisation has been evaluated using the standard \textsc{miriad} task \textrm{IMPOL}. Our estimated peak fractional polarisation value is P~=~5~$\pm$~1 percent, while average polarisation is about $\sim$3 percent. This is unusually weak for a young SNR and very similar to earlier \cite{1993AA...275..265D} results, especially when compared to the range of other LMC SNRs that we observed over the past decades with ATCA \citep[][]{2007MNRAS.378.1237B,2008SerAJ.177...61C,2008SerAJ.176...59C,2009SerAJ.179...55C,2010A&A...518A..35C,2010SerAJ.181...43B,2012MNRAS.420.2588B,2012RMxAA..48...41B,2012SerAJ.184...69B,2012A&A...543A.154H,2012SerAJ.185...25B,2013MNRAS.432.2177B,2014ApJ...780...50B,2014MNRAS.439.1110B,2014Ap&SS.351..207B,2014MNRAS.440.3220B,2015A&A...583A.121K}. Perhaps, significant depolarization within the 2~GHz bandwidth is present. Figure \ref{fig1} shows the large-scale distributions of $^{12}$CO($J$ = 3--2) and H{\sc i} toward the SNR N63A. We find three GMCs, whose size is $\sim$7--10 pc. Two of them extend toward the northeast and east of the X-ray shell (hereafter referred to as the ``NE cloud'' and the ``E cloud''), which correspond to the GMCs previously mentioned by \cite{2017AIPC.1792d0038S}. The other one lies just west of the center of the SNR. The GMC, hereafter referred to as the ``W cloud'', spatially coincides not only with the X-ray hole, but also with the radio continuum peak or optical nebula. In the H{\sc i} map, we find two cavity-like structures toward the northeast and center of the SNR. The former corresponds to the H{\sc i} shell GS~76 cataloged by \cite{1999AJ....118.2797K}. The latter represents an H{\sc i} absorption dip owing to the strong radio continuum emission from the optical nebula and SNR, which is similar to the case of LMC SNR N49 \citep{2018ApJ...863...55Y} or N103B \citep{2018ApJ...867....7S}. The enlarged view of the northeastern shell is shown in Figure \ref{fig1}(c). The northeastern X-ray shell appears to be associated not only with the NE and E clouds, but also with the H{\sc i} wall. Figure \ref{fig2} shows the averaged CO and H{\sc i} spectra toward the SNR N63A. We find significant differences in the H{\sc i} spectra toward the SNR (blue, inside the SNR) and its surroundings (cyan, outside the SNR). The negative H{\sc i} brightness temperature at the velocity of $\sim$278 km s$^{-1}$ is the absorption line, also suggesting that there is H{\sc i} located in front of the SNR. The CO spectrum has an intensity peak at $V_\mathrm{LSR}$ $\sim$281 km s$^{-1}$, which is slightly shifted from the central velocity of the H{\sc i} absorption line. \subsection{Detailed CO Distribution with ALMA}\label{subsec:ALMA} Figure \ref{fig3} shows a three-color image of N63A composed by the combination of HST H$\alpha$ \citep[red,][]{2001AIPC..565..409C}, ALMA $^{12}$CO($J$ = 1--0) (green), and {\it Chandra} X-rays in the energy band of 0.3--6.0 keV \citep[blue,][]{2003ApJ...583..260W}. We detect all the GMCs that we identified using the ASTE CO data. Since the E cloud is spatially separated from the X-ray shell, the GMC is probably not associated with the SNR. On the other hand, the NE cloud is elongated to the southwest direction, whose tip with a position at ($\alpha_\mathrm{J2000}$, $\delta_\mathrm{J2000}$) $\sim$ ($05^\mathrm{h}35^\mathrm{m}47\farcs8$, $-66\degr01\arcmin38\arcsec$) is adjacent to the northeastern X-ray shell. We also spatially resolved the W cloud into several CO clouds. The molecular clouds show clumpy distributions, which are likely embedded within the optical nebula and the X-ray shell. \begin{deluxetable*}{lcccccccccccc}[] \tablecaption{Properties of Molecular Clouds Embedded within the Optical Nebula in the SNR N63A} \tablehead{\multicolumn{1}{c}{Name} & $\alpha_{\mathrm{J2000}}$ & $\delta_{\mathrm{J2000}}$ && $T_{\rm b} $ && $V_{\mathrm{peak}}$ & $\Delta V$ & Size & Mass & $N$(H$_2$) & H$_2$ density & Ionization state\\ & ($^{\mathrm{h}}$ $^{\mathrm{m}}$ $^{\mathrm{s}}$) & ($^{\circ}$ $\arcmin$ $\arcsec$) && (K) && \scalebox{0.9}[1]{(km $\mathrm{s^{-1}}$)} & \scalebox{0.9}[1]{(km $\mathrm{s^{-1}}$)} & (pc) & ($M_\sun $) & ($\times 10^{22}$ cm$^{-2}$) & (cm$^{-3}$) &\\ \multicolumn{1}{c}{(1)} & (2) & (3) && (4) && (5) & (6) & (7) & (8) & (9) & (10) & (11)} \startdata A ........... & 5 35 40.52 & $-66$ 01 58.8 && 20.6 && 280.9 & 1.4 & 0.9 & 170 & 2.3 & \phantom{0}7700 & photoionized\\ B ........... & 5 35 40.13 & $-66$ 02 03.4 && 28.4 && 280.9 & 2.4 & 1.3 & 650 & 5.2 &12400 & photoionized\\ C ........... & 5 35 40.03 & $-66$ 02 06.9 && \phantom{0}5.4 && 281.6 & 1.8 & 0.8 & \phantom{0}60 & 0.9 & \phantom{0}4900 & photoionized\\ D ........... & 5 35 40.73 & $-66$ 02 03.0 && 27.8 && 280.4 & 2.1 & 1.5 & 730 & 4.6 & \phantom{0}7800 & photoionized\\ E ........... & 5 35 41.60 & $-66$ 01 57.3 && \phantom{0}9.9 && 278.3 & 0.8 & 0.7 & \phantom{0}30 & 0.6 & \phantom{0}3600 & photoionized\\ F ........... & 5 35 40.90 & $-66$ 02 10.0 && 14.5 && 283.2 & 1.2 & 0.9 & 100 & 1.3 & \phantom{0}5400 & photoionized\\ G ........... & 5 35 41.60 & $-66$ 02 10.8 && 26.1 && 282.6 & 1.1 & 1.5 & 360 & 2.2 & \phantom{0}4200 & shock-ionized \\ H ........... & 5 35 42.39 & $-66$ 02 06.1 && 17.1 && 282.2 & 1.3 & 1.4 & 270 & 1.6 & \phantom{0}4100 & shock-ionized\\ I ............. & 5 35 43.87 & $-66$ 02 05.9 && 22.5 && 280.9 & 0.8 & 1.0 & 110 & 1.3 & \phantom{0}4100 & shock-ionized\\ J ............ & 5 35 44.70 & $-66$ 02 06.8 && 15.3 && 280.0 & 0.9 & 0.9 & \phantom{0}80 & 1.0 & \phantom{0}3900 & shock-ionized\\ K ........... & 5 35 43.42 & $-66$ 02 23.3 && \phantom{0}3.5 && 282.0 & 1.7 & 0.4 & \phantom{0}10 & 0.5 & \phantom{0}6800 & shock-ionized\\ \enddata \tablecomments{Col. (1): Cloud name. Cols. (2--7): Properties of CO emission obtained by Gaussian fitting. Cols. (2)--(3): Position of peak intensity. Col. (4): Maximum brightness temperature. Col. (5): Center velocity. Col. (6): Line width (FWHM). Col. (7): Cloud size defined as $2 \sqrt{(S / \pi)}$, where $S$ is the total cloud surface area enclosed by the integrated intensity contour of $\sim5\sigma$. Col. (8): Cloud mass derived by using the relation between the molecular hydrogen column density $N$($\mathrm{H_2}$) and $^{12}$CO($J$ = 1--0) integrated intensity $W$(CO) as $N$($\mathrm{H_2}$) = 7.0 $\times$ $10^{20}$[$W$(CO) (K km $\mathrm{s^{-1}}$)] ($\mathrm{cm^{-2}}$) \citep{2008ApJS..178...56F}. (9) Maximum column density of molecular hydrogen. (10) Number density of molecular hydrogen. (11) Ionization state of optical nebula associated with the cloud.} \label{table1} \vspace*{-0.1cm} \end{deluxetable*} To estimate the mass of GMCs, we utilize the following equations: \begin{eqnarray} M = m_{\mathrm{H}} \mu \Omega D^2 \sum_{i} [N_i(\mathrm{H}_2)],\\ N(\mathrm{H}_2) = X \cdot W(\mathrm{^{12}CO}), \label{eq2} \end{eqnarray} where $m_\mathrm{H}$ is the mass of atomic hydrogen, $\mu = 2.72$ is the mean molecular weight relative to atomic hydrogen, $\Omega$ is the solid angle of each pixel, $D$ is the distance to the LMC, $N_i(\mathrm{H}_2)$ is the molecular hydrogen column density for each pixel $i$, $X$ is the CO-to-H$_2$ conversion factor, and $W(\mathrm{^{12}CO})$ is the integrated intensity of the $^{12}$CO($J$ = 1--0). We used the conversion factor $X = 7.0 \times 10^{20}$ cm$^{-2}$ (K km s$^{-1}$$)^{-1}$ \citep{2008ApJS..178...56F}. The physical properties of GMCs are estimated for the regions that are significantly detected by CO with a $\sim$5$\sigma$ or higher. We finally obtain the mass of GMC is $\sim$6400 $M_{\sun}$ for the NE cloud, $\sim$3600 $M_{\sun}$ for the E cloud, and $\sim$2600 $M_{\sun}$ for the W cloud. Figure \ref{fig4} shows an enlarged view of the optical nebula obtained by HST [S {\sc ii}] (red), H$\alpha$ (green), and [O {\sc iii}] (blue). The northeast and southeast lobes, comprising the shock-ionized gas, show a crescent-shape with many filamentary structures. By contrast, the western lobe is smoothly distributed in H$\alpha$ with optical dark lanes and a compact H{\sc ii} region. We identified eleven molecular clouds, A--K, toward the optical nebula as shown in white contours. Definitions and basic physical properties of each cloud are listed in Table \ref{table1}. The densest molecular clouds, named A, B, and D, show a good spatial correspondence with the optical dark lane, suggesting that these molecular clouds are located in front of the photoionized lobe. In point of fact, the peak proton column density of cloud B is $\sim$$1 \times 10^{23}$ cm$^{-2}$, corresponding to $A_\mathrm{V}$ $\sim$4 magnitude \citep[e.g.,][]{1974ApJ...187..243J}. On the other hand, there is no clear evidence of an optical dark lane toward the molecular clouds G--K, which are associated with the shock-ionized lobes. This implies that these molecular clouds are located inside or behind the shock-ionized lobes. We also note that bright optical filaments as shown in Figure \ref{fig4} are in spatial alignment (in projection) with the molecular cloud H, possibly suggesting the shape of the filamentary structure is reflected by that of natal molecular cloud before the shock-ionization. The total mass of molecular clouds is $\sim$1700 $M_{\sun}$ for the photoionized region (A--F) and $\sim$800 $M_{\sun}$ for the shock-ionized region (G--K). \begin{figure*} \begin{center} \includegraphics[width=\linewidth]{n63a_xray_closed_up_rgb_v16.eps} \caption{Three-colour images of the SNR N63A observed by {\it Chandra}. The red, green, and blue colors represent the energy bands of 0.3--0.6, 0.6--1.1, and 1.1--6.0 keV, respectively. Superposed contours represent (a) H$\alpha$, (b) [S {\sc ii}], (c) [O {\sc iii}], (d) CO, and (e) radio continuum with the center frequency of 4786 MHz. The optical contours are smoothed using a median-filter. The integration velocity range and contour levels of CO are the same as in Figure \ref{fig4}. The contour levels are 0.3, 0.5, 0.7, 1.0, 1.5, and 2.0 counts s$^{-1}$ pixel$^{-1}$ for H$\alpha$; 0.15, 0.35, 0.7, 1.0, 1.5, and 2.0 counts s$^{-1}$ pixel$^{-1}$ for [S {\sc ii}]; 0.06, 0.08, 0.1, 0.15, 0.2, 0.3, 0.4 for [O {\sc iii}]; 34, 38, 42, 46, 50, 54, 58, and 62 K for radio continuum. We also show boundaries of H$\alpha$ (red), radio continuum (green), and CO (cyan) in (f). Regions of X-ray enhancements are also shown.} \label{fig5} \end{center} \vspace*{0.5cm} \end{figure*}% We find that the molecular cloud is also depressed toward the exciting star HD~271389 with a position at ($\alpha_\mathrm{J2000}$, $\delta_\mathrm{J2000}$) $\sim$ ($05^\mathrm{h}35^\mathrm{m}40$\farcs$68$, $-66\degr02\arcmin05\farcs9$). Based on photometry of $U-B$ = $-0.853 \pm 0.082$ mag \citep{1996ApJS..104...71O} and $M_\mathrm{v}$ = $-3.6 \pm 0.5$ mag \citep{1986A&A...164...26L}, B0V--B1V is reasonable for the spectral type of the exciting star \citep[see also][]{2013ApJS..208....9P}. The UV radiation and stellar winds therefore must be powerful enough to ionize the molecular cloud. In fact, several pillar-like structures surrounding the exciting star are seen, which are also previously mentioned by \cite{2012ApJ...761..107C}. Further ALMA observations with fine angular resolution of sub-arcsecond will allow us to study the photon-dominated region (PDR) in detail. \begin{figure*} \begin{center} \includegraphics[width=\linewidth]{n63a_xray_spec_rgb_v9.eps} \caption{Background subtracted ACIS-S spectra of the Northeast, West, and Southeast regions with the best-fit models shown in top panels of Figures (a), (b), and (c, d), respectively. The magenta, blue, green, and purple lines represent the CIE (ISM), NEI/CIE (ejecta), PL, and another NEI (ISM) components, respectively. The black lines show the sum of the components. Bottom and middle panels in each figure indicate the residuals from each model.} \label{fig6} \end{center} \end{figure*}% Figure \ref{fig5} shows a comparison among the X-rays (red: 0.3--0.6 keV, green: 0.6--1.1 keV, and blue: 1.1--6.0 keV), H$\alpha$, [S {\sc ii}], [O {\sc iii}], CO, and radio continuum with a center frequency of 4786 MHz \citep{1993AA...275..265D}. The X-ray hole near the location of the optical nebula is considered to be the result of interstellar absorption by the dense gas cloud \citep{2003ApJ...583..260W}. We confirm that a particularly dark region in the X-ray hole shows a good spatial correspondence with the dense molecular clouds. Both the H$\alpha$ and [S {\sc ii}] emission also spatially coincide with the X-ray hole. Additionally, we find that the bright radio continuum emission shows a good spatial correspondence not only with the X-ray hole, but also with the H$\alpha$ nebula, suggesting that the radio continuum at $\sim$5 GHz is dominated by the free-free radiation. The spatial extent of the molecular clouds, optical nebula, and/or radio continuum therefore spatially accounts for $\sim$80\% area of the X-ray hole. The remaining $\sim$20\% is placed to the south of the molecular clouds G, F, and J. These regions are possibly associated with dense H{\sc i} clumps or diffuse ionized gas. In fact, if we take into account regions with low radio continuum emission of $\sim$10 K, the X-ray hole is completely filled by the neutral gas and/or ionized gas. We also note that X-rays are clearly enhanced around the northern shock-ionized lobe (see Figure \ref{fig5}a), and the edges of molecular clouds (see Figure \ref{fig5}d). The former regions are also bright in [O {\sc iii}] emission [Figure \ref{fig5}(c)], suggesting that the shock ionization occurred. The shock-velocities of the [O {\sc iii}] bright regions are thought to be $\sim$120 km s$^{-1}$ and below 200 km s$^{-1}$ \citep[e.g.,][]{1985ApJ...298..651C,2008ApJS..178...20A}. However, the quoted shock velocities may be too slow to produce X-rays in these regions. It possibly means that shockwaves are suddenly decelerated toward the region owing to the shock interaction with dense materials. We estimate the number density of ionized protons $n_\mathrm{p}$ using the radio continuum data and following equation: \begin{eqnarray} \tau_\mathrm{c}(\nu) = 8.213 \times 10^{-2} T_\mathrm{e}^{-1.35} (\nu \;/\; 1~\mathrm{GHz})^{-2.1} EM,\\ T_\mathrm{b} = T_\mathrm{e}[1- \exp{(\tau_\mathrm{c}(\nu))}], \label{eq4} \end{eqnarray} where $\tau_\mathrm{c}(\nu)$ is the optical depth of ionized gas, $T_\mathrm{e}$ is the electron temperature, $T_\mathrm{b}$ is the brightness temperature of the radio continuum emission, and $EM$ is the emission measure. $EM$ is defined as \begin{eqnarray} EM = \int_0^{L_1} n_\mathrm{e} n_\mathrm{p} dl \sim n_\mathrm{p}^2 L_1 \label{eq5} \end{eqnarray} where $n_\mathrm{e}$ is the number density of ionized electrons and $L_1$ is the thickness of the ionized gas in the unit of pc. The mean brightness temperature is $\sim$$43 \pm 7$ K for the shock-ionized region and $\sim$$34 \pm 1$ K for the photoinized region. We obtain the ionized proton density $\sim$$280 \pm 110$ cm$^{-3}$ for the shock-ionized region and $\sim$$560 \pm 100$ cm$^{-3}$ for the photoionized region, assuming $L_1 \sim$5 pc and $T_\mathrm{e} = 14743$ K for the shock-ionized lobe: $L_1 \sim$1 pc and $T_\mathrm{e} = 14620$ K for the photoionized lobes \citep{2012ApJ...761..107C}. These values are roughly consistent with estimates of $\sim$250 cm$^{-3}$ by the X-ray spectroscopy \citep{2003ApJ...583..260W} and of $\sim$50--300 cm$^{-3}$ by the optical studies (\citeauthor{1983ApJ...275..592S} \citeyear{1983ApJ...275..592S} and the references therein). \begin{deluxetable*}{llcccc}[ht] \tablecaption{Best-fit X-ray Spectral Parameters \label{table2}} \tablehead{ \colhead{Component} & \colhead{Parameter} & \colhead{Northeast} & \colhead{West} & \colhead{Southeast} & \colhead{Southeast} \\ &&&& {\scriptsize (2 CIE + PL)} & {\scriptsize (CIE + 2 NEI)}} \startdata Absorption & $N_{\rm H, LMC}~(10^{21}~\rm cm^{-2})$ & 1.9 $_{-0.4}^{+0.5}$ & 2.1 $_{-0.2}^{+0.6}$ & 5.4 $_{-0.6}^{+0.5}$ & 5.5 $_{-0.8}^{+0.5}$\\ ~ & $N_{\rm H, MW}~(10^{21}~\rm cm^{-2})$ & 0.6 (fixed) & 0.6 (fixed) & 0.6 (fixed) & 0.6 (fixed)\\ CIE (ISM) & $kT_e~\rm (keV)$ & 0.21 $_{-0.02}^{+0.02}$ & 0.12 $_{-0.01}^{+0.04}$ & 0.20 $_{-0.01}^{+0.01}$ & 0.20 $_{-0.01}^{+0.01}$\\ ~ & VEM~($\times 10^{59}~\rm cm^{-3}$) & 1.8 $_{-0.6}^{+1.0}$ & 1.5 $_{-0.8}^{+1.5}$ & 8.9 $_{-3.2}^{+4.3}$ & 9.9 $_{-4.2}^{+4.4}$\\ NEI/CIE (ejecta) & $kT_e~\rm (keV)$ & 0.70 $_{-0.01}^{+0.02}$ & 0.73 $_{-0.03}^{+0.01}$ & 0.71 $_{-0.03}^{+0.04}$ & 0.70 $_{-0.04}^{+0.04}$\\ ~ & $Z_{\rm O}~\rm (solar)$ & 1.72 $_{-0.46}^{+0.35}$ & 3.01 $_{-0.97}^{+3.73}$ & $\leq 2.64$ & $\leq 2.54$\\ ~ & $Z_{\rm Ne}~\rm (solar)$ & 1.16 $_{-0.34}^{+0.28}$ & 0.89 $_{-0.34}^{+0.75}$ & 3.07 $_{-0.73}^{+3.11}$ & 3.21 $_{-0.93}^{+1.26}$\\ ~ & $Z_{\rm Mg}~\rm (solar)$ & 0.66 $_{-0.14}^{+0.19}$ & 0.77 $_{-0.20}^{+0.67}$ & 1.22 $_{-0.36}^{+1.35}$ & 1.25 $_{-0.46}^{+0.75}$\\ ~ & $Z_{\rm Si}~\rm (solar)$ & 0.66 $_{-0.13}^{+0.14}$ & 0.63 $_{-0.16}^{+0.45}$ & 0.89 $_{-0.26}^{+0.95}$ & 0.91 $_{-0.26}^{+0.46}$\\ ~ & $Z_{\rm S} = Z_{\rm Ar} = Z_{\rm Ca}~\rm (solar)$ & 0.43 $_{-0.16}^{+0.18}$ & 0.87 $_{-0.28}^{+0.67}$ & 0.71 $_{-0.24}^{+0.44}$ & 0.56 $_{-0.30}^{+0.39}$\\ ~ & $Z_{\rm Fe} = Z_{\rm Ni}~\rm (solar)$ & 0.46 $_{-0.06}^{+0.07}$ & 0.42 $_{-0.09}^{+0.37}$ & 0.33 $_{-0.11}^{+0.39}$ & 0.36 $_{-0.12}^{+0.23}$\\ ~ & $n_et~\rm (10^{11}~cm^{-3}~s)$ & $\geq 6.5$ & 4.4 $_{-1.4}^{+4.1}$ & 100 (fixed) & $\geq 6.1$\\ ~ & VEM~($\times 10^{59}~\rm cm^{-3}$) & 1.0 $_{-0.1}^{+0.1}$ & 0.4 $_{-0.1}^{+0.1}$ & 1.1 $_{-0.3}^{+0.3}$ & 1.0 $_{-0.4}^{+0.3}$\\ NEI (ISM2) & $kT_e~\rm (keV)$ & $\cdots$ & $\cdots$ & $\cdots$ & $\geq 1.80$\\ ~ & $n_et~\rm (10^{11}~cm^{-3}~s)$ & $\cdots$ & $\cdots$ & $\cdots$ & 1.0 $_{-0.6}^{+1.4}$\\ ~ & VEM~($\times 10^{57}~\rm cm^{-3}$) & $\cdots$ & $\cdots$ & $\cdots$ & 7.1$_{-4.1}^{+10.6}$ \\ PL & $\Gamma$ & $\cdots$ & $\cdots$ & 1.7 $_{-1.5}^{+1.5}$ & $\cdots$\\ ~ & Flux$^{\ast}~\rm (erg~s^{-1}~cm^{-2})$ & $\cdots$ & $\cdots$ & $1.8 \times 10^{-13}$ & $\cdots$\\ \hline ~ & reduced-$\chi^2$ (d.o.f.) & 1.10 (141) & 1.17 (123) & 1.06 (156) & 1.04 (154)\\ \hline \enddata \tablecomments{ $^{\ast}$The flux is the unabsorbed flux in the 1--10 keV band.} \end{deluxetable*} \subsection{X-ray Spectral Analysis}\label{subsec:xspec} To compare the absorbing column density of X-rays with the interstellar gas density, we extract X-ray spectra from three regions as shown in Figure \ref{fig5}(d): the west (W), southeast (SE), and northeast (NE) regions. The W region covers the densest molecular clouds A, B, and D. The SE region represents the X-ray hole with clumpy molecular clouds, and NE corresponds to a reference region without dense clouds or ionized gas. Figure \ref{fig6} shows the background-subtracted ACIS-S spectra for each region. The background is selected as source-free region with a central position of ($\alpha_\mathrm{J2000}$, $\delta_\mathrm{J2000}$) $\sim$ ($05^\mathrm{h}35^\mathrm{m}31\farcs6$, $-66\degr03\arcmin30\arcsec$), whose position is outside of the SNR. Following the previous X-ray study in N63A \citep{2003ApJ...583..260W}, we first fitted the NE spectrum with a non-equilibrium ionization (NEI) plasma model using the VVRNEI in the XSPEC package. We separately set absorption column densities in the Milky Way ($N_{\rm H, MW}$) and the LMC ($N_{\rm H, LMC}$). For the absorption, we used the Tuebingen-Boulder ISM absorption model \citep[TBabs,][]{2000ApJ...542..914W} and fixed $N_{\rm H, MW}$ at $6.0 \times 10^{20}~\rm cm^{-2}$ \citep{1990ARA&A..28..215D}. We fixed the initial temperature ($kT_\mathrm{init}$) at 0.01 keV whereas the electron temperature ($kT_\mathrm{e}$), ionization parameter ($n_\mathrm{e}t$) and volume emission measure (VEM $= \int n_\mathrm{e} n_\mathrm{p} dV$) are free parameters. We allowed to vary the abundances of the elements O, Ne, Mg, Si, S, and Fe whose line emissions can be seen in the X-ray spectrum. The Ar and Ca abundances are linked to S, while Ni is linked to Fe. The other abundances fixed to the LMC values in literature \citep[He = 0.89, C = 0.45, N = 0.18, others = 0.50;][]{Russell1992}. During the analysis, we used the solar values of \cite{2000ApJ...542..914W}. The NE spectrum above 0.7~keV was be reproduced well by this model, but large residuals left in the 0.5--0.7~keV band ($\chi^2$/d.o.f = 264/143) as shown in the middle panel in Figure~\ref{fig6}a. \cite{2003ApJ...586..210P,2012ApJ...748..117P} performed spatial resolved spectral analysis with {\it Chandra} X-ray data of N49 which is an SNR in the LMC with a similar age and shock-cloud interaction to those of N63A. They reproduced the spectra with a two-component NEI model consisting of a higher-$kT_e$ and a lower-$kT_e$ components. Following their approach, we tried the two-component NEI model whose abundances were fixed to the LMC values in literature (O = 0.21, Ne = 0.28, Mg = 0.33, Si = 0.69, Fe = 0.35; \citeauthor{2016A&A...585A.162M} \citeyear{2016A&A...585A.162M}, He = 0.89, C = 0.45, N = 0.18, others = 0.50; \citeauthor{Russell1992} \citeyear{Russell1992}). In this fit, $n_et$ of the lower-$kT_e$ component became larger than $10^{13}~\rm cm^{-3}~s$, indicating that the plasma is in a collisional ionization equilibrium (CIE) state. Therefore, we fixed $n_et$ of the component at $10^{13}~\rm cm^{-3}~s$. The residuals above 0.7~keV can be improved by the fit but the other residuals left in the bands around the line emissions ($\chi^2$/d.o.f = 275/147; see Figure~\ref{fig6}a) because the abundance pattern of the higher-$kT_e$ component differ the LMC values. We therefore allowed to vary the abundances of O, Ne, Mg, Si, S, and Fe for the higher-$kT_e$ component. The NE spectrum is well reproduced ($\chi^2$/d.o.f = 155/141) by the model consisting of the lower-$kT_e$ CIE and higher-$kT_e$ NEI components with $kT_e = 0.21~_{-0.02}^{+0.02}~\rm keV$ and $kT_e = 0.70~_{-0.01}^{+0.02}~\rm keV$, respectively. The best-fit model and parameters are shown in Figure~\ref{fig6}a and Table~\ref{table2}, respectively. The best-fit values of the abundances of the higher-$kT_e$ NEI component are far from the LMC values, suggesting that the component is originated from the ejecta whereas the lower-$kT_e$ CIE plasma is a shocked ISM. We applied the same CIE+NEI model as NE to the W and SE spectra. The W spectrum can be fitted well (Figure~\ref{fig6}b and Table~\ref{table2}). On the other hand, the obtained a $\chi^2$/d.o.f. of the SE region was 202/157; hence, a two-component plasma model is rejected because the residuals left above 4 keV band (see middle panels in Figures~\ref{fig6}c and \ref{fig6}d). We therefore tried two different models; one includes a power-law (PL) component, and the other includes another NEI model. We first fitted the SE spectrum with the CIE+NEI+PL model. In this fit, $n_et$ of the NEI component became larger than $10^{13}~\rm cm^{-3}~s$, and therefore, we fixed it at $10^{13}~\rm cm^{-3}~s$. The fit significantly improved the residuals above 4~keV and reproduced the spectrum well ($\chi^2$/d.o.f. = 165/156). The best-fit model and parameters of the SE region are also shown in Figure~\ref{fig6}c and Table~\ref{table2}, respectively. In the CIE+2NEI model, $kT_e$, $n_et$ and VEM in the additional NEI component were allowed to vary. The abundances became large values (5--300 solar) when these were allowed to vary due to low photon statistics. In fact, fixing abundances at the LMC values, we found that our spectral model fit is equally good. Thus, we take this model fit with abundances fixed at the LMC values as the best-fit model for the CIE+2NEI scenario. The fit was significantly reduced to $\chi^2$/d.o.f. = 161/154 and obtained higher $kT_e$ of $\geq 1.80 $~keV and lower $n_et$ of $1.0 _{-0.6}^{+1.4} \times 10^{11}~\rm cm^{-3}~s$ than those of the lower-$kT_e$ NEI component (Figure~\ref{fig6}d and Table~\ref{table2}). The reduced-$\chi^2$ in the CIE+2NEI fit is a little smaller than that in the CIE+NEI+PL fit but this improvement is not statistically significant with a F-test probability of 0.097. Therefore, we consider two cases in later discussion of the hard X-ray component. We finally obtained the absorbing column density $N_{\rm H, LMC}$(X-ray) of (1.5--2.4$) \times 10^{21}~\rm cm^{-2}$ for the NE region, (4.7--6.0$) \times 10^{21}~\rm cm^{-2}$ for the SE region, and (1.9--2.7$) \times 10^{21}~\rm cm^{-2}$ for the W region. \subsection{Comparison of the ISM and Absorbing Column Density}\label{subsec:comp} To estimate the total interstellar proton column density $N_\mathrm{H}$(H$_2$ + H{\sc i}), we use the equation (\ref{eq2}) and following equations \citep[e.g.,][]{1990ARA&A..28..215D}: \begin{eqnarray} N_\mathrm{H}(\mathrm{H}_2 + \mathrm{H}{\textsc{i}}) = 2 \times N(\mathrm{H}_2) + N_\mathrm{H}(\mathrm{H}{\textsc{i}}),\\ N_\mathrm{H}(\mathrm{H}{\textsc{i}}) = 1.823 \times 10^{18} W(\mathrm{H}{\textsc{i}}), \label{eq7} \end{eqnarray} where $N_\mathrm{H}$(H{\sc i}) is the column density of atomic hydrogen and $W$(H{\sc i}) is the integrated intensity of H{\sc i}. In the SNR~N63A, it is difficult to derive both the $N_\mathrm{H}$(H{\sc i}) and $W$(H{\sc i}) owing to the strong absorption of H{\sc i}. Then, we assume $W$(H{\sc i}) = 300--500 K km s$^{-1}$ of the SNR from its surroundings. For $N$(H$_2$), we estimate the averaged values of $W$(CO) for each region. Then, we obtain $N_\mathrm{H}$(H{\sc i}) $\sim$5--$9 \times 10^{20}$ cm$^{-2}$ and $2 \times N$(H$_2) \sim$0.2--$2.7 \times 10^{22}$ cm$^{-2}$, indicating that the atomic hydrogen component is considered to be negligible in $N_\mathrm{H}$(H$_2$ + H{\sc i}). We finally obtain the total column density of $N_\mathrm{H}$(H$_2$ + H{\sc i}) is $\sim$$3 \times 10^{21}$ cm$^{-2}$ for the NE region, $\sim$$7 \times10^{21}$ cm$^{-2}$ for the SE region, and $\sim$$3 \times 10^{22}$ cm$^{-2}$ for the W region. These values are $\sim$1.5--15 times higher than the absorbing column densities obtained from the X-ray spectra for each region. \section{Discussion}\label{discussion} \subsection{Dense Molecular Clouds Engulfed by the Shock Waves}\label{subsec:dis1} N63A is a unique SNR embedded within the large H{\sc ii} region N63, which is also associated with the dense molecular clouds A--K, shock-ionized gas, and photoionized gas. As described in Section \ref{subsec:ALMA}, these dense clouds are certainly associated with both the shock-ionized and photoionized lobes. To be more precise, the eastern molecular clouds G--K are completely embedded within the shock-ionized lobes because the optical dark lane is not clearly seen despite their high-density ($\sim$4000--7000 cm$^{-3}$, see Table \ref{table1}). Spatial alignment of shock-ionized filaments with the molecular cloud H is a possible evidence for the physical relation among the ionized gas, natal dense gas, and the shock-survived clouds. By contrast, the western molecular clouds A--F are located just in front of, or partially embedded within the photoionized lobe because of the presence of optical dark lane and its pillar-like structures (Figure \ref{fig4}). Both the western and eastern clouds D, E, and J are rim-brightened in soft-band X-rays, suggesting that the surface of molecular clouds are selectively ionized by the shock. Considering the position of the optical nebula---near the center of the SNR---, the shock waves likely propagated from east to west and from far side to front side of the molecular clouds, if we assume that the supernova site is near the geometric center of the SNR. It is consistent with that the eastern-half of optical nebula is strongly shock-ionized and 70\% of GMC mass remains in the western photoionized lobe. We also confirm that the ionization time scale of shock-ionized lobes is three times longer than that of the photionized lobe by using the $n_\mathrm{e}t$ values and the ionized gas density (see Table \ref{table2} and Section \ref{subsec:ALMA}). Detailed spatially resolved observations using near infrared lines (e.g., H$_2$, [Fe {\sc ii}], [P {\sc ii}]) and numerical calculations are needed to derive shock parameters. We also argue that the molecular clouds have been completely engulfed by the shock waves. Because the column densities derived by X-rays ($\sim$1.3--$3.2 \times 10^{21}$ cm$^{-2}$) are significantly smaller than that of the total interstellar protons ($\sim$3--$30 \times 10^{21}$ cm$^{-2}$). This means that the X-ray emitters exist not only behind the molecular clouds, but also in front of the clouds. In addition to this, thermal plasma components with different velocities produced by the forward and reverse shock will be possibly detected. A further X-ray observation with high-spectral-resolution X-ray imaging instruments such as the X-Ray Imaging and Spectroscopy Mission (XRISM) will allow us to study the kinematics of thermal plasma components. Finally, we shall present a possible evolutionally scenario of N63A and its environments. In the northeastern edge of the LMC, the massive star cluster NGC~2030 was born $\sim$3--6 Myr ago \citep{1985A&A...152..427C}. There is a small amount of molecular gas owing to the edge of the galaxy \citep[see][]{2001ApJ...553L.185Y}, but is still rich in the atomic hydrogen \citep[c.f.,][]{2003ApJS..148..473K}. According to \cite{2017PASJ...69L...5F} and \cite{2018arXiv180300713T}, most of massive stars in the LMC have been possibly formed by the tidally-driven colliding H{\sc i} flows. Therefore, NGC~2030 was probably formed also by the H{\sc i} flows due to the tidal interactions between the LMC and SMC. Subsequently, massive stars including the progenitor of the SNR N63A and the exciting star HD~271389 started to evacuate the natal molecular and atomic gas by their strong UV radiation and stellar winds. About 3500 yrs ago, the massive progenitor of N63A exploded in the inhomogeneous density environment. In the large scale, the gas density of northeast is much higher than that of southwest (Figure \ref{fig1}). Therefore the southwestern X-ray shell shows diffuse and more expanded morphology, while the northeastern shell mainly collided with the dense HI wall (see Figure \ref{fig1}c). Then, the shock waves encountered the western molecular clouds, and now engulfed all the molecular clouds associated with the optical nebula. \subsection{Origin of Hard X-rays} In section \ref{subsec:xspec}, we presented that the hard X-ray component of SW can be described not only as the power-law model, but also as the high-temperature plasma model. In this section, we discuss both the cases and their strong relation with the interstellar environment. \subsubsection{Case 1: An Efficient Acceleration of Cosmic Ray Electrons via the Shock-Cloud Interaction}\label{subsec:dis2} Young SNRs are thought to be primary accelerators of cosmic rays not only in our Galaxy, but also in external galaxies such as the LMC. Supernova shockwave with a velocity of $\sim$3000--10000 km s$^{-1}$ provides an ideal site for the diffusive shock acceleration \citep[DSA;][]{1978ApJ...221L..29B,1978MNRAS.182..147B}. During the past twenty years, more efficient acceleration mechanisms of cosmic rays have been discussed from both the theoretical and observational studies (e.g., reverse shock acceleration, \citeauthor{2005A&A...429..569E} \citeyear{2005A&A...429..569E}; non-linear effect of DSA, \citeauthor{2001RPPh...64..429M} \citeyear{2001RPPh...64..429M}; magnetic reconnection in the turbulent medium, \citeauthor{2012PhRvL.108m5003H} \citeyear{2012PhRvL.108m5003H}). The shock-cloud interactions also have received attention as one of the efficient acceleration mechanisms of cosmic rays. \cite{2003PASJ...55L..61F} discovered the synchrotron X-ray enhancement toward the dense molecular clouds in the northwest of the Galactic young SNR RX~J1713.7$-$3946. Subsequent studies confirmed that many molecular clouds associated with the SNR are rim-brightened in synchrotron X-rays \citep{2010ApJ...724...59S,2013ApJ...778...59S,2016scir.book.....S}. Owing to interactions between the shock and inhomogeneous gas distribution---dense gas ($\sim$$10^3$ cm$^{-3}$) clumps in low-density environment ($\sim$0.01 cm$^{-3}$)---the magnetic field strength is significantly enhanced up to $\sim$1 mG via the strong turbulent motion around the dense gas clumps. Then, we observe bright synchrotron X-rays from the periphery of the molecular clouds. This interpretation is also consistent with the magnetohydrodynamic (MHD) simulations \citep[][]{2009ApJ...695..825I,2012ApJ...744...71I}. Additionally, similar observational trends are seen in other young SNRs both in the Galaxy and the LMC \citep[e.g.,][]{2015ASPC..499..257S,2017ApJ...843...61S,2017JHEAp..15....1S,2017arXiv171108165K,2018ApJ...863...55Y}. The X-ray hard spectra are reported toward the regions in which the shock-cloud interactions are strongly occurred, indicating that cosmic rays are efficiently accelerated to the higher maximum energy \citep[][]{2015ApJ...799..175S,2018arXiv180711050B}. The young SNR N63A possibly shows similar observational trends as described above, if the hard X-rays is dominated by the synchrotron X-rays. The synchrotron X-rays are significantly detected toward the southeast of the optical lobe [Figures \ref{fig5}(d) and \ref{fig6}(b)], corresponding to the shock-ionized region with tiny molecular clouds. This means that the synchrotron X-ray was enhanced via the interactions between the shocks and dense neutral clumps. To test our interpretation, we compare the hard X-ray image with the CO distribution. Figure \ref{fig7} shows the RGB image of hard X-rays ($E$: 4.3--6.0, red), CO (green), and broadband X-rays ($E$: 0.3--6.0 keV, blue). The energy band of hard X-rays has no line emission and dominantly consists of synchrotron X-rays relative to the thermal component [see Figure \ref{fig6}(b)]. We confirm that the hard X-ray peak A spatially corresponds to the shock-ionized region with clumpy neutral gas. We also note that the molecular cloud K is also associated with one of the minor peaks of hard X-rays, indicating that the shock-cloud interaction also occurred. However, there is no dense molecular cloud toward the peak B and the other two minor peaks of hard X-rays. We present a hypothesis that these hard X-ray peaks are possibly associated with a cold H{\sc i} clump with a density of a few 100 cm$^{-3}$. In fact, \cite{2003ApJ...583..260W} mentioned the presence of interstellar absorption toward the hard X-ray peak B. Moreover, it is known that the cold H{\sc i} clumps also enhance the synchrotron X-ray enhancement (e.g., RX~J1713.7$-$3946, \citeauthor{2013ApJ...778...59S} \citeyear{2013ApJ...778...59S}; RCW~86, \citeauthor{2017JHEAp..15....1S} \citeyear{2017JHEAp..15....1S}). To confirm this scenario, detailed H{\sc i} observations and X-ray imaging spectroscopy are needed. \subsubsection{Case 2: High Temperature Plasma toward the Shocked Molecular Clouds}\label{subsec:dis22} We here discuss an alternative idea that the hard X-rays are originated by high-temperature plasma of shocked ISM. In the SE region of N63A, the supernova shock may strongly interact with clumpy and dense molecular clouds, developing multiple reflected shock structures to heat the gas up to high temperature. In this scenario, the re-heated gas may be either low-abundant ISM or overabundant ejecta. Unfortunately, we could not distinguish them by X-ray spectroscopy alone due to the low-photon statistics. On the other hand, the morphological structure of hard X-rays (Figure \ref{fig7}) favors the origin of the shocked ISM. We note that the spatial extent of hard-X-ray toward peak A is very similar to that of the shock-ionized optical lobe (see also in Figures \ref{fig3} and \ref{fig4}), indicating that the hard X-rays are possibly same origin of the shock-ionized optical lobe. The $n_\mathrm{e}t$ value also supports the recent heating of the dense molecular clouds. Assuming the depth of the emitting region to be 3 pc, spatial extent of shock-ionized optical lobe, the emission volume of SE is estimated to be $V = 1.3 \times 10^{57}$ cm$^3$. Therefore the VEM of high-temperature plasma in SE corresponds to the electron density of $n_\mathrm{e} = 2.5$ $f^{-0.5}$ cm$^{-3}$, where $f$ is the filling factor for this component. We then derive the elapsed time since the dense molecular clouds was heated $t = 4.0 \times 10^{10}$ $f^{0.5}$ s $ < 1300$ yr This value is much less than the maximum age of 5000 yr, and hence the high temperature plasma likely has been heated recently. This situation is very similar to the Magellanic SNR N49 and Galactic SNR RCW~86 \citep{2003ApJ...586..210P,2008PASJ...60S.123Y}. To confirm the scenario, we need more detailed studies of spatially resolved X-ray spectroscopy for the whole remnant. \subsection{Ionized Gas as a Target of Cosmic Ray Protons}\label{subsec:dis3} Interstellar gas associated with SNRs also plays an important role in understanding the origin of gamma-rays. The gamma-rays from young SNRs are thought to be produced by two mechanisms: the hadronic process and the leptonic process. The former is due to the decay of neutral pions produced by the interactions between the accelerated cosmic-ray protons and interstellar protons, while the latter is where a cosmic-ray electron energizes a low-energy photon to gamma-ray energies via the inverse Compton effect. The non-thermal bremsstrahlung of cosmic-ray electrons is also one of the origins of leptonic gamma-rays. For the young SNRs, however, the non-thermal bremsstrahlung is negligible \citep[e.g.,][]{2018AA...612A...4H,2018AA...612A...6H,2018AA...612A...7H}. If the hadronic process is dominant, a good spatial correlation between the interstellar protons and gamma-rays is expected. \cite{2012ApJ...746...82F} demonstrated such spatial correspondence in the young SNR RX~J1713.7$-$3946 for the first time. The authors took into account both the molecular and atomic components as the interstellar protons, and then derived the averaged interstellar proton density of $\sim$130 cm$^{-3}$. The total energy of cosmic-ray protons is also estimated to be $\sim$10$^{48}$ erg, corresponding to $\sim$0.1\% of the typical kinematic energy of a supernova explosion. Subsequent studies for both the young and middle-aged SNRs show similar values of the total cosmic-ray energy $\sim$10$^{48}$--10$^{49}$ erg \citep[e.g.,][]{2013ApJ...768..179Y,2017AIPC.1792d0039Y,2014ApJ...788...94F,2017ApJ...850...71F,2017arXiv171108165K,2018arXiv180510647S}. \begin{figure} \begin{center} \includegraphics[width=\linewidth]{n63a_hard_xray_v4.eps} \caption{Same three-colour image of Figure \ref{fig2}, but the red color represents {\it Chandra} X-rays in the energy band of 4.3--6.0 keV.} \label{fig7} \end{center} \end{figure}% Most recently, \cite{2018Ap&SS.363..144C} presented a significant detection of GeV gamma-rays from the SNR N63A, which was confirmed with 9 year {\it Fermi} {\it Large}-{\it Area} {\it Telescope} (LAT) data. Owing to the coarse angular resolution, we could not compare the gamma-ray image with the gas distribution. The gamma-ray flux ($E$: 1--10 GeV) was estimated to be 6--$12 \times 10^{-7}$ MeV cm$^{-2}$ s$^{-1}$, corresponding to the gamma-ray luminosity $L_\gamma$(1--10 GeV) of $\sim$1.3--$5.8 \times 10^{35}$ ($d$ / 50 kpc)$^2$ erg s$^{-1}$, where $d$ is distance of the source. If the gamma-ray spectrum is dominated by the hadronic origin, the total energy of cosmic-ray protons $W_\mathrm{p}$ is given by the following equation \citep[e.g.,][]{2006A&A...449..223A}: \begin{eqnarray} W_\mathrm{p} \sim t_\mathrm{pp \rightarrow \pi_0} \times L_\gamma \label{eq8} \end{eqnarray} where $t_\mathrm{pp \rightarrow \pi_0} \sim$$4.5 \times 10^{15} (n$ / 1 cm$^{-3}$)$^{-1}$ s is the characteristic cooling time of protons, and $n$ is number density of interstellar protons. We then finally obtain the total energy of cosmic-ray protons $W_\mathrm{p}$(1--10 GeV) as follow: \begin{eqnarray} W_\mathrm{p} \sim 0.6\mathchar`-2.6 \times 10^{51} (n\;/\;1\;\mathrm{cm}^{-3})^{-1}(d\:/\:50\:\mathrm{kpc})^2\:\mathrm{erg} \label{eq9} \end{eqnarray} In general, $n$ is derived as ``neutral gas density'' consisting of both the molecular and atomic components \citep[e.g.,][]{2012ApJ...746...82F}, because the shock-ionization effect is negligible even for the middle aged SNRs W44 and IC443 \citep[e.g.,][]{2013ApJ...768..179Y,2017AIPC.1792d0039Y}. For the case of N63A, however, most of neutral molecular clouds have been ionized especially toward the eastern half of the optical nebula. In addition to this, the low-energy cosmic-ray protons traced by the {\it Fermi} data cannot penetrate into the dense molecular clouds. According to \cite{2012ApJ...744...71I}, the penetration depth $l_\mathrm{pd}$ of cosmic-ray protons is \begin{eqnarray} l_\mathrm{pd} \sim 0.002\; \eta^{0.5}\; (E / 1\;\mathrm{GeV})^{0.5}\; (B / 100\;\mathrm{\mu G})^{-0.5}\nonumber \\ (t_\mathrm{age} / 3500\;\mathrm{yr})^{0.5} \;\;(\mathrm{pc}), \label{eq10} \end{eqnarray} where $\eta$ $= B^2/\delta B^2 \ga 1$ is a turbulence-factor defined as the degree of magnetic-field fluctuations, $E$ is the energy of cosmic-ray protons, $B$ is the magnetic-field strength, and $t_\mathrm{age}$ is the age of the SNR. The magnetic field strength {\it B} in the Galactic molecular clouds is given by the following equation \citep{2010ApJ...725..466C}: \begin{eqnarray} B \sim 10\;(n / 300\; \mathrm{cm^{-3}})^{0.65} \;\; \mathrm{(\mu G)} \label{eq11} \end{eqnarray} where $n$ is the number density of protons in molecular clouds. In the SNR N63A, we estimate the magnetic field strength $B \sim$80--180 $\mu$G in the molecular clouds associated with N63A (see Table \ref{table1}), assuming the equation (\ref{eq11}) to hold for the LMC. GeV gamma-ray flux of N63A is measured from 1 to 10 GeV, corresponding to the cosmic-ray proton energy of 10--100 GeV if the hadronic process dominates. The penetration depth $l_\mathrm{pd}$ is therefore to be 0.005--0.03 pc if we assume $\eta = 1$ under the shock-cloud interaction \citep[e.g.,][]{2007Natur.449..576U}. The penetration depth is significantly smaller than the size of the molecular clouds (see Table \ref{table1}), indicating that the molecular cloud in N63A may not a target of the low-energy cosmic-ray protons. We therefore use the total ISM proton density consisting of both the ionized gas and neutral atomic hydrogen. By using the equations (\ref{eq5}), we derive the averaged ionized gas density to be $\sim$$110 \pm 70$ cm$^{-3}$ assuming $L_1 \sim$15 pc ($\sim$ diameter of the radio bright shell) and $T_e \sim$ 14743 K for the whole SNR. In Section \ref{subsec:comp}, we derived the atomic hydrogen column density as 5--$9 \times 10^{20}$ cm$^{-2}$ using the H{\sc i} data. We divided it by twice the shell thickness of $\sim$3 pc, which is estimated by a three-dimensional Gaussian fitting of the northeastern X-ray shell \citep[c.f.,][]{2017ApJ...843...61S}. We finally obtain the neutral atomic hydrogen density of $\sim$$80 \pm 20$ cm$^{-3}$ and the total ISM proton density of $\sim$$190 \pm 90$ cm$^{-3}$. The total energy of cosmic-ray protons is then estimated to be $\sim$0.3--$1.4 \times 10^{49}$ erg, corresponding to $\sim$0.3--1.4\% of typical kinematic energy released by a single supernova. These values are roughly consistent with the Galactic SNRs (e.g., \citeauthor{2018arXiv180510647S} \citeyear{2018arXiv180510647S} and references therein). Further gamma-ray observations using the Cherenkov Telescope Array (CTA) will allow us to study the gamma-ray morphology and spectrum of N63A in detail. \section{Conclusions}\label{conclusions} In the present study, we have carried out new $^{12}$CO($J$ = 1--0, 3--2) observations of the LMC SNR N63A by using ASTE and ALMA with angular resolutions of $\sim$0.4--6 pc. The primary conclusions are summarized as below. \begin{enumerate} \item We have found three GMCs toward the northeast, east, and near the center of the SNR N63A using ASTE $^{12}$CO($J$ = 3--2) data. The cloud size is $\sim$7--10 pc and the total mass of the GMCs is $\sim$$1 \times 10^4$ $M_{\sun}$. Using ALMA $^{12}$CO($J$ = 1--0) data, we spatially resolved the GMC into eleven molecular clouds, which are embedded within the optical nebula. The total mass of molecular clouds is $\sim$800 $M_{\sun}$ for the shock-ionized region and $\sim$1700 $M_{\sun}$ for the photoionized region. The densest molecular clouds A, B, and D show a good spatial correspondence with the optical dark lane, indicating that most of these clouds are located in front of the photoionized nebula. On the other hand, the molecular clouds G--K are placed inside or behind the shock-ionized lobes. The extent of the X-ray hole coincides with that of the CO clouds, H$\alpha$ nebula, and radio continuum, indicating that the interstellar absorption of X-rays is caused not only by the dense molecular clouds, but also by the ionized gas cloud. \item A spatially resolved X-ray spectroscopy has revealed that the absorbing column densities toward the molecular clouds are $\sim$1.5--$6.0 \times 10^{21}$ cm$^{-2}$, which are $\sim$1.5--15 times less than the averaged interstellar proton column densities. This indicates that all the dense molecular clouds have been completely engulfed by the shock waves, but are still survive from erosion owing to their high-density and short interacting time. The X-ray spectrum toward the shocked molecular clumps is also well fitted by the models consisting not only with the absorbed CIE, NEI, and power-law components, but also with the absorbed CIE and two NEI components. The former indicates that the shock-cloud interaction possibly enhances the synchrotron X-ray flux and/or the maximum energy of cosmic-ray electrons, through the amplifications of the magnetic field strength and turbulence motion. For the latter case, shock-cloud interaction develop multiple reflected shock structures to heat the gas up to high temperature roughly 1300 yrs ago or less. \item For the SNR N63A, the ionized gas may act as a target of the accelerated cosmic-ray protons because most of natal molecular clouds have been ionized by the shock. If the GeV gamma-rays from N63A are dominated by the hadronic origin, the total energy of cosmic-ray protons is calculated to be $\sim$0.3--$1.4 \times 10^{49}$ erg with the estimated ISM proton density of $\sim$$190 \pm 90$ cm$^{-3}$, containing both the shock-ionized gas and neutral atomic hydrogen. This value corresponds to $\sim$0.3--1.4\% of typical kinematic energy of a single supernova, roughly consistent with the Galactic SNRs. Further gamma-ray observations using the Cherenkov Telescope Array (CTA) will allow us to study the gamma-ray morphology and spectrum of N63A in detail. \end{enumerate} \acknowledgments {\small This paper makes use of the following ALMA data: ADS/JAO.ALMA $\#$2015.1.01130.S. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada) and NSC and ASIAA (Taiwan) and KASI (Republic of Korea), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ. The scientific results reported in this article are based on data obtained from the Chandra Data Archive (Obs ID: 777). This research has made use of software provided by the Chandra X-ray Center (CXC) in the application packages CIAO (v 4.10). This study was financially supported by Grants-in-Aid for Scientific Research (KAKENHI) of the Japanese Society for the Promotion of Science (JSPS, grant Nos. 15H05694, 16K17664, and 18J01417). HS is supported by ``Building of Consortia for the Development of Human Resources in Science and Technology'' of Ministry of Education, Culture, Sports, Science and Technology (MEXT, grant No. 01-M1-0305). HM is supported by World Premier International Research Center Initiative (WPI). K. Tokuda is supported by NAOJ ALMA Scientific Research Grant Number of 2016-03B. MS acknowledges support by the Deutsche Forschungsgemeinschaft (DFG) through the Heisenberg professor grant SA 2131/5-1 and the research grant SA 2131/4-1. We really appreciate the anonymous referee for useful comments and suggestions, which helped the authors to improve the paper.} \software{CASA \citep[v 5.1.0.:][]{2007ASPC..376..127M}, CIAO \citep[v 4.10:][]{2006SPIE.6270E..1VF}, MIRIAD \citep{1995ASPC...77..433S}, KARMA \citep{1997PASA...14..106G}.}

6a73fe977fe6404723a791258428dd3b4b14da15

\section{Introduction} \label{sec:intro} \noindent We consider the problem of graph-based hypothesis testing, which tests whether two sets of graph-valued observation samples are drawn from the same distribution. This is a topic of growing interest~\cite{li2011graph,asta2015geometric,muk2017,arrigo2017exponential}; however, there are only a few studies where the observations are {\em weighted graphs}. In this work, we address this gap, considering the problem of {\em hypothesis-testing for replicated weighted graph-valued data}. This is a challenging problem, since the average and atypical behavior of a sample of networks is difficult to characterize. Figure \ref{fig:net_twitter} illustrates two example domains with populations of weighted graphs. The top row illustrates the word co-occurrence networks of two Twitter users in Brazil, one that is {\em pro-government} and one that is {\em anti-government}. Here the entity corresponds to a user on social media, the nodes in the graph are vocabulary words, and the edges weights reflect co-occurrences of words in the posts of the users. The bottom row illustrates brain connectivity networks of two individuals, one {\em female} and one {\em male}. Here the entity corresponds to an individual, the nodes in the graph are brain regions, and the edges weights represent functional connectivity strength as measured by functional magnetic resonance imaging (fMRI). In both cases, we would like to investigate how populations and entities differ. For example, whether brain activity differs with respect to sex or whether word usage differs with respect to political views. There has been recent work on graph-based hypothesis testing (see \cite{akoglu2015graph} for a good survey). However, much of the work has focused on testing subgraphs {\em within} a larger graph (e.g., \cite{mongiovi}), or on one-sample tests comparing a single graph to a null model (e.g., \cite{moreno2}). Work focusing on {\em populations} of graphs has received considerably less attention and falls into one of two categories: that of~\cite{ginestet2014}, which introduces a geometric characterization of the network using the so-called Fr\'echet mean, and that of~\cite{durante2016}, who proposed a Bayesian latent-variable model for unweighted graphs. We focus on the latter, which allows us to bring the powerful machinery of probabilistic hierarchical modeling to the table, allowing noisiness and missingness, and providing interpretable confidence scores. Unfortunately, existing work along this second direction is limited to modeling binary graphs, so that in practice, a {\em threshold} must be used to transform counts or continuous weights to 0/1 values. Such a thresholding operation discards valuable information about the strength of the edge-weights, and can also exhibit sensitivity towards the choice of threshold. Too small a threshold can result in a graph that is too dense, and too large, too sparse. Often, there does not even exist a single appropriate threshold across the entire graph. \begin{figure} \centering \hspace{-12mm} \includegraphics[width=.55\textwidth]{figure/join_figures.pdf} \caption{Top row: word connectivity networks for two Brazilian Twitter users, (A) from the pro-government side and (B) from the anti-government side. Bottom row: brain connectivity network for two individuals, (C) female and (D) male.} \label{fig:net_twitter} \end{figure} In this paper, we address the issues of previous work and develop a hypothesis testing framework that facilitates testing over graphs populations with edge-weights, which can follow any parametric distribution. Specifically, we propose a Bayesian hypothesis testing framework that uses a mixture of latent space models for weighted networks to test for population-differences. Our framework is capable of population-level, entity-specific as well as edge-specific hypothesis testing. We consider testing in two broad scenarios: \begin{enumerate} \item When all {\em observations} from the same population follow the same distribution, we can ask: Are the population distributions identical? \item When all {\em entities} from the same population follow the same distribution, we can ask: Are the population distributions identical? Now, every entity has an associated set of graph-valued observations which are identically distributed, but are not exchangeable across entities. \end{enumerate} Observe the first case is equivalent to the second when each entity has only one associated graph; however the latter allows heterogeneity among entities in the same population. For instance an entity in a population that is politically {\em conservative} might frame an issue they discuss from an economic perspective, while another entity in a population that is politically {\em liberal} might focus on the social aspects of the same issue. \cite{peixoto2015inferring} proposed a strongly parametric time-varying framework to handle this important situation, our approach is significantly more flexible. We apply our testing framework to problems from the types of domains summarized in Figure \ref{fig:net_twitter}. First, we look at word co-occurrence network data from Twitter (on the political crisis in Brazil), as well as Instagram (on side effects of Adderall and Ritalin usage for Attention Deficit Hyperactivity Disorder~\cite{correia}). In both datasets, we investigate how populations and entities differ based on the way they communicate---specially in the manner in which the usage of pairs of key words differs between groups. Standard methods such as unigram mixture models, latent Dirichlet allocation (LDA)~\cite{blei2003latent} or N-gram language models~\cite{katz1987estimation}, which are based just on word-frequency, do not capture the kind of contextual information we are interested in. While these methods can identify words that `belong' to different groups, in our scenario there is a strong overlap in key words across groups, and such models will fail to differentiate between groups which share common themes and vocabularies. Second, we used functional magnetic resonance imaging (fMRI) data, to investigate how brain activity differs across populations like sex, age and personality traits like extroversion, conscientiousness and creativity. In both tasks, we show that graph-structure as well as graph weights are crucial to performance, and that we outperform baselines like latent Dirichlet allocation (LDA)~\cite{blei2003latent}, N-gram language models~\cite{katz1987estimation}, as well as thresholding methods like~\cite{durante2016}. \vspace{2mm} \noindent \textbf{Contributions:} Our contribution is a multi-level statistical hypothesis testing framework for populations of weighted networks, concerning both the overall graph distributions, as well as two types of local hypotheses: entity-specific and edge-specific. The latter are important since a population might have networks, or a network edges, that are statistically different, and that might escape detection by a global test. Our hierarchical Bayesian mixed membership model allows statistical information to be shared across groups, increasing accuracy of hypothesis tests without loss of statistical power. This allows practitioners to evaluate anomalies in a principled manner, using statistical significance. Notably, our framework is more robust than previous methods developed for binary graphs, which require thresholding of weighted data before application. \section{The model} We are given a set $\bo{A}$ of undirected graphs, with graph $A_{nt}$ belonging to entity $n$ at index $t$ (refered to as `time'). Here $n \in [1,...,N]$ and $t \in [1,...,T_n]$, with $A_{nt}[i,j]$ giving the link strength between vertices $i$ and $j$ of entity $n$ at time $t$ ($i,j \in [1,...,V]$). We also observe population information $y_{n} \in [1,...,G]$ for each entity. For instance, each network might represent word co-occurrences in a user's social media messages over some time period, while the population might indicate whether the user's political leanings are `Liberal' or `Conservative'. Underlying our testing framework is a probabilistic model which we now outline. We assume each observation $A_{nt}$ comes from one of $H$ clusters or mixture components, with cluster $h$ having parameter $\bm{\theta}^{(h)}$. Each cluster has a distribution over graphs which we write as $F(\bm{\theta}^{(h)})$ (we specify $\bm{\theta}^{(h)}$ and $F$ in the next paragraph). Clusters and cluster parameters are shared across populations, however each population $y$ has its own Dirichlet-distributed probability over clusters, $\bm{\beta}_y$. At a high-level ours is a Bayesian hypothesis-testing approach which tests whether the $\bm{\beta}_y$'s are identical across populations. For the case of two populations, we place equal {\em a priori} probability on the null hypothesis $H_0: \bm{\beta}_1 = \bm{\beta}_2$ and the alternative $H_1: \bm{\beta}_1 \ne \bm{\beta}_2$. Using the machinery of Bayesian inference, we evaluate the posterior probabilities of the two hypotheses given observations, and reject the null if its probability $P(H_0 | -)$ is less than some specified threshold (e.g., 0.05 or 0.1). We now describe the cluster-specific distribution over graphs, $F(\bm{\theta}^{(h)})$. For cluster $h$, $\bm{\theta}^{(h)}$ is a $V\times V$ matrix, whose $(ij)$th element parametrizes the probability of the weight on the edge between nodes $i$ and $j$. In our applications, we looked at count-valued edges, and so assumed $F\left(\bm{\theta}^{(h)}\right)$ to be Binomial or Poisson distributed with parameter $\bm{\theta}^{(h)}[i,j]$ on edge $(i,j)$. We define $\bm{\theta}^{(h)} = f\left(\bm{S}^{(h)}\right)$ where $f(\cdot)$ is some link function (e.g.\ the logistic or exponential function to ensure nonnegativity), and constrain $\bm{S}$ using a low-rank factorization scheme $\bm{S}^{(h)} = \bo{X}^{(h)} \bo{X}^{(h)T}$. Here $\bo{X}^{(h)} \in \mathcal{R}^{|V| \times R}$ and $R \ll |V|$, so that $\bo{X}_v^{(h)}$ gives the location of node $v$ in some low-dimension space, and $\bm{S}^{(h)}$ is the proximity of all nodes. The number of parameters thus grows linearly, rather than quadratically with the number of vertices. In equations, we expand the upper plates in Figure~\ref{mixGraph}(a) and (b), to get \begin{align} \bm{\theta}^{(h)}& = f(\bm{X}^{(h)}\bm{X}^{(h)T}), \quad \bo{X}_{v}^{(h)} \sim N_{R}\left(\bm{0},\mathds{I}\right), v=1\ldots,V \end{align} Each population $y$ has a distribution over clusters $\bm{\beta}_y$. With prior probability half, the null hypothesis is true (we indicate this with the variable $\mathcal{T}$), in which case all populations have the same distribution $\bm{\beta}$. Otherwise, each population $g$ has its own distribution, $\bm{\beta}_g$. Thus, \begin{align} \mathcal{T} \sim \text{Bern}&(1/2) \\ \text{If } \mathcal{T} = 0:& \quad \bm{\beta}_1 =\ldots,=\bm{\beta}_G \sim \text{Dir}(\alpha,...,\alpha) \nonumber \\ \text{Else:} &\quad \bm{\beta}_g \stackrel{iid}{\sim} \text{Dir}(\alpha,...,\alpha) \text{ for } g=1\ldots G.\nonumber \end{align} Now, consider the case where each entity has only a single associated graph. Then the $n$th entity (belonging to population $Y_n$) has a graph $A_n$ distributed as \begin{align} C_{n}|Y_{n} \sim \bm{\beta}_{Y_n} \quad \bo{A}_{n}|C_{n} \stackrel{}{\sim} F\left(\bm{\theta}^{(C_n)}\right) \label{eq:gen_model0} \end{align} Here $C_n$ refers to the latent variable that identifies the cluster membership of entity $n$, which depends on the population entity $n$ is drawn from. For the case where we have multiple network observations per entity, we add a layer to this hierarchical model. Now, each entity $n$ has their own distribution over clusters $\bm{\pi}_n$ centered around the population distribution: \begin{align} \bm{\pi}_{n} \sim Dir(\beta_{Y_n}). \end{align} The graph $t$ of this entity is independently distributed as \begin{align} C_{nt}|Y_{n} \sim \bm{\pi}_{n},\quad \bo{A}_{nt}|C_{nt} \stackrel{}{\sim} F\left(\bm{\theta}^{(C_{nt})}\right) \label{eq:gen_model} \end{align} Figure (\ref{mixGraph}) summarizes our generative process for both cases. \subsection{Model Inference} We are given a set of network observations ${\bm A}$, each written as $A_{nt}$ where $n$ indexes entities and $t$, `time'. For each $A_{nt}$, we are also given a population assignment $Y_n\in \{1,2\}$. Since we observe the population memberships $\bm{Y}$ and the networks $\bm{A}$, the inferential task is to learn $C_{nt}$, ${\bm\pi}_{n}$, $\bm{\beta}_y$ and $\bm{\theta}^{(h)}$. In the next section, we will use these variables as statistics in our hypothesis tests. For notational convenience, we will refer to a link between an arbitrary pair of nodes $i$ and $j$ with $l$, so that we can write $A_{n}[i,j]$ as $A_{n}[l]$. We also represent the weighted matrix with its vectorized lower triangular component $\mathcal{L}(\bo{A}_n) = (\mathcal{L}(A_{nt})_1,...,\mathcal{L}(A_{nt})_{V(V-1)/2})$. For the general model specified above, we carry out posterior inference via a Gibbs sampler, whose individual updates we outline next: \begin{enumerate} \item {\em Sample the cluster indicator for each graph.} \\This comes from the multinomial: \begin{multline*} P(C_{nt}=h|- ) =\\ \frac{\pi_{n_h} \beta_{h} \prod_l P\left( \calig{L}(A_{nt})_l = a_l|\theta_{l}^{(h)}\right)}{\sum_{m=1}^H \pi_{n_m} \beta_{m} \prod_l P\left( \calig{L}(A_{nt})_l = a_l |\theta_{l}^{(m)}\right)} \end{multline*} where $l \in [1,...,V(V-1)/2]$ and $a_l|\theta_l^{(h)} \sim \left[F(\bm{\theta}^{(h)})\right]_l$. \item {\em Sample the mixing probabilities for each entity $n$.} \\ With $\bm{{m}}_n$ a vector of cluster assignment counts of graphs of entity $n$ $$ \bm{\pi}_n \sim \text{Dir}(\bm{\beta}_{Y_n} + \bm{{m}}_n) $$ \item {\em Sample the locations of the nodes for each cluster.} \\With a Gaussian prior over the locations $\bm{X}_n$, and the weight-distribution parametrized after transforming through a link function $f$, this is a straightforward exercise in sampling from the posterior of a Gaussian with a nonlinear link function. Standard techniques exist to do this~\cite{duane1987hybrid,ellslisam}, though we followed a recent idea involving the Polya-Gamma data-augmentation scheme~\cite{polson2013}. \item {\em Sample the testing indicator $\mathcal{T}\sim \text{Bern}(P(\mathcal{H}_1)|-)$.} Since $\mathcal{T}$ is the test-statistic central to our methodology, we discuss this in a bit more detail in the next section. The update rule is given by Equation~\ref{probab}. \item {\em Sample the mixing probabilities for each population $y$.} \\ If $\mathcal{T}=0$ then for all $y$, $\bm{\beta}_y = \bm{\beta} \sim \text{Dir}(\alpha+n_1,...,\alpha+n_H)$ where $n_h$ is the number of graphs in cluster $h$. If $\mathcal{T}=1$ then $\bm{\beta}_y \sim \text{Dir}(\alpha+n_{1y},...,\alpha+n_{Hy})$ for each $y$, where $n_{iy}$ counts the number of graphs from population $y$ in cluster $i$. \end{enumerate} \section{Weighted-network comparison tests} \label{sec:datagen} For simplicity, we focus on the case where we have only two populations ($G=2$). Under our formulation, the problem of hypothesis testing amounts to testing whether the population-level cluster assignment probabilities $\bm{\beta}_1$ and $\bm{\beta}_2$ are significantly different under the posterior. We elaborate on this below. \begin{figure} \hspace{-8mm}\includegraphics[width=0.6\textwidth]{figure/graph_model.pdf} \caption{The graphical models are given by (A) fixed and (B) with time-varying structures. } \label{mixGraph} \end{figure} \subsection{Population-level network comparison test} This task involves comparing the posterior probabilities of the two hypotheses, $H_0: \bm{\beta}_1 = \bm{\beta}_2$ vs $H_1:\bm{\beta}_1\ne \bm{\beta}_2$. Since $H_0$ being true amounts to $\mathcal{T}=1$, our MCMC estimate of the probability equals the fraction of MCMC iterations where $\mathcal{T}=1$. We first describe how our Gibbs sampler updates this variable (step 4 of our Gibbs sampler). At any MCMC iteration, let $\bm{{m}}_y$ be the vector of cluster assignment counts for population $y$, with component $c$ giving the number of observations from population $y$ assigned to cluster $c$: $\bm{{m}}_y = \left(\sum_{n;y_{n}=y}\sum_t \mathds{I}_{C_{nt}=1},...,\sum_{n;y_{n}=y}\sum_t \mathds{I}_{C_{nt}=H}\right)$. We write $\bm{{m}}=\bm{{m}}_1 + \bm{{m}}_2$ (for the two populations in $G$, i.e., 1 and 2). Then, under the two hypotheses, these counts are distributed as \begin{equation} \begin{aligned} \mathcal{H}_0:& \ \bm{{m}}_1, \bm{{m}}_2 \stackrel{iid}{\sim} \text{Mult}(\bm{\beta}), \\ & \ \bm{\beta} \sim \text{Dirichlet}(\bm{\alpha}) \\ \mathcal{H}_1:& \ \bm{{m}}_1 \sim \text{Mult}(\bm{\beta}_1) \ \ and \ \ \bm{{m}}_2 \sim \text{Mult}(\bm{\beta}_2) \\ & \ \bm{\beta_1},\bm{\beta_2} \sim \text{Dirichlet}(\bm{\alpha}) \end{aligned} \label{overall_hyp} \end{equation} Marginalizing out the $\bm{\beta}'s$, and recalling that both hypotheses have the same prior probability, we can specify the posterior $$P(\mathcal{H}_1|-) = \frac{P(\bm{{m}}_1|\bm{\alpha})P(\bm{{m}}_2|\bm{\alpha})}{P(\bm{{m}}|\bm{\alpha})+P(\bm{{m}}_1|\bm{\alpha})P(\bm{{m}}_2|\bm{\alpha})}$$ From the Dirichlet-multinomial conjugacy, we can write down the marginal probabilities of the observations, giving \begin{equation} P(\mathcal{H}_1|-) = \frac{\prod_{y=1}^2 \frac{B(\bm{\alpha}+\bm{{m}}_y)}{B(\bm{\alpha})}}{\frac{B(\bm{\alpha}+ \bm{{m}})}{B(\bm{\alpha})}+ \prod_{y=1}^2 \frac{B(\bm{\alpha}+\bm{{m}}_y)}{B(\bm{\alpha})}} \label{probab} \end{equation} where $B(\cdot)$ is the multivariate beta function $B(x) = \prod_{i=1}^q\frac{\Gamma(x_i)}{\Gamma(\sum_{i=1}^q x_i)}$. Every Gibbs iteration samples $\mathcal{T}$ from this, with the posterior probability of the alternative hypothesis, $P(\mathcal{H}_1|-)$, being the fraction of MCMC samples where $\mathcal{T}$ equals $1$. If the estimate from Equation~\eqref{probab} is larger than a specified threshold {(e.g., 0.95)}, we reject the null hypothesis and conclude that the populations are significantly different. We can use this {\em network comparison} (NC) test for both models in Figure~\ref{mixGraph}. When we use the (fixed) model in \ref{mixGraph}a, we will refer to it as NC-F and when we use the (mixed-membership) model in \ref{mixGraph}b, we will refer to it as NC-M. \subsection{Entity-specific comparison test} This task refers to do following hypothesis test: $H_0^{n_1n_2}: \bm{\pi}_{n_1} = \bm{\pi}_{n_2}$ Vs $H^{n_1n_2}_1:\bm{\pi}_{n_1} \ne \bm{\pi}_{n_2}$ for any two users $n_1$ and $n_2$. Estimating this from our posterior samples is straightforward. Assuming multiple networks per entity, let $\bm{\tilde{m}}_{n}$ be a vector of counts for entity $n$, giving the number of observations assigned to each cluster. As mentioned earlier, the entity-specific distribution over clusters follows the distribution $\bm{\pi}_n \sim \text{Dir}(\bm{\beta})$. Following the earlier logic, Equation~\ref{local_hyp} gives the posterior probability two given entities have different cluster assignment probabilities: \begin{equation} P(\mathcal{H}^{n_1n_2}_{1}|-) = \frac{\prod_{i=1}^2 \frac{B(\bm{\beta}+\bm{\tilde{m}}_{n_i})}{B(\bm{\beta})}}{\frac{B(\bm{\beta}+ \bm{\tilde{m}})}{B(\bm{\beta})}+ \prod_{i=1}^2 \frac{B(\bm{\beta}+\bm{\tilde{m}}_{n_i})}{B(\bm{\beta})}} \label{local_hyp} \end{equation} It is important to note that Equation \ref{local_hyp} allows pairwise comparisons across populations, and therefore it is possible to have significantly similar entities from different populations and significantly different entities in the same population. \subsection{Edge-specific comparison test} This task refers to the following hypothesis test for an edge $l=(i,j)$, $H_0^l: \bm{{\theta}}_1[l] = \bm{{\theta}}_2[l]$ vs $H_1^l: \bm{{\theta}}_1[l] \ne \bm{{\theta}}_2[l]$. For the edge application we use an adjusted version of Cramer's V-statistic proposed by~\cite{dunson2009} given by Equation~\ref{edgeesp}: \begin{equation} p_l^2 = \sum_{y=1}^2 \mathbf{p}_{\mathcal{Y}} \sum_{a_l} \frac{ P(\calig{L}(A)_l = a_l|\bar{\theta}_{y_l}) - P(\calig{L}(A)_l = a_l|\bar{\theta}_{l})}{ P(\calig{L}(A)_l = a_l|\bar{\theta}_{l})} \label{edgeesp} \end{equation} where $\mathbf{p}_{\mathcal{Y}}$ is the sample size proportion of each population, $\bm{\bar{\theta}}_y =\sum_{n=1}^{N} \sum_{h=1}^H \beta_{yh} \bm{\theta}^{(h)} \mathds{I}_{y_n=y} $, and $\bm{\bar{\theta}} = \sum_{y=1}^2\frac{\bm{\bar{\theta}}_y}{2}$. If $p_l \approx 1$ then there is evidence that edge weights differ across the populations. \section{Related work} \label{sec:relwork} Graph-based hypothesis testing and anomaly detection are topics of growing interest with diverse applications~(see e.g., \cite{akoglu2015graph}). Many applications of hypothesis testing in network analysis focus on subgraphs {\em within} a larger graph (e.g., \cite{mongiovi}), or one-sample tests comparing a single graph to a null model (e.g., \cite{moreno2}). Work on populations of graphs can be divided on two areas: dynamic networks, in which one graph is replicated over time \cite{sarkar2012nonparametric}, \cite{charlin2015dynamic} and \cite{durante2014nonparametric}; and exchangeable graph modeling in which each graph is considered to be one observation for a single entity (see~\cite{ginestet2014,asta2015geometric,durante2015,durante2016} and~\cite{duvenaud2015convolutional}). Our paper generalizes work from the latter category by allowing within-population heterogeneity, with each entity having multiple graphs with similar statistical properties. Both~\cite{ginestet2014} and~\cite{asta2015geometric} deal with geometric characterizations of networks, and while their approaches are mathematically elegant, they are substantially less flexible than our work. \cite{duvenaud2015convolutional} take a convolution neural network approach for non-aligned graphs, where there is no known mapping between nodes in each graph. This, coupled with the fact that their method requires the presense of node features, makes it unsuitable for our applications. Most closely related to ours is the method presented in \cite{durante2016}. This method, which we will refer to DD, is a special case of our framework, where there is no within population variation, and where network edges are binary. For count or continuous-valued data, one might consider thresholding the edge weights of each entity and then applying the DD method. However this discards valuable information about the strength of the edge-weights, introduces sensitivity to threshold-level, and can reduce statistical power. Our method offers the ability to flexibly model such edge-weight information without any significant additional computational complexity. In particular, the computational time complexity of both our method and DD is $\mathcal{O}(NHR|V|^2)$ per iteration. {Here $N$ is the number of networks, $H$ is the number of clusters, $R$ is the dimensionality of the low rank approximation, and $|V|$ is the number of nodes in each graph. In practice, $H$ and $R$ are small constants that do not grow with the data.} In our experiments we compare with DD for different thresholds. \section{Experiments} \label{sec:syndata} \begin{figure*} \hspace{-8mm} \centering \includegraphics[width=1.04\textwidth]{figure/join_power_figures.pdf} \caption{Type I error and statistical power curves for the synthetic (left) and twitter (right) data for increasing sample sizes} \vspace{-4mm} \label{fig:syn_power} \end{figure*} In order to assess the efficacy of our method, we divided our analysis into four parts: statistical power and type-I error analyses, population-level hypothesis tests, edge-specific hypothesis tests, and additional exploratory analysis. We start with statistical power and type-I error analyses, the most important measures of assessing hypothesis tests. We investigate the efficacy of our (and competing) methods for varying sample sizes when the ground truth in known. We show that when the data are generated from a known two-population setup, our hypothesis testing framework produces significantly more accurate results and has lower variance, with respect to type-I error and statistical power, compared to a number of other baselines. We show that for time-varying data, the mixed membership extension of our model is essential for reliable inference. We also study the sensitivity of the method of~\cite{durante2016} (which requires unweighted graphs) to threshold settings, for population-level hypothesis tests. We show that for heterogeneous data, the hypothesis-test decisions are highly sensitive to threshold choice. We study the edge-specific hypothesis tests qualitatively, by visualizing the estimated model structure for each approach. We end by describing some additional insights that that our method gleans from the data. We start by describing the datasets. \subsection{Datasets} We generated synthetic weighted network data for two settings, the entity-homogeneous version from Figure~\ref{mixGraph}a (NC-F, where each entity is represented by one graph) and the entity-heterogeneous version from Figure \ref{mixGraph}b (NC-M, where each entity is represented by multiple graphs). We also applied our framework to three real-world applications: a Twitter dataset from the political crisis in Brazil, two datasets about drugs usage on Instagram, and fMRI recordings of brains of human subjects. \vspace{1mm} \noindent \textbf{Synthetic data (Homogeneous):} \label{syn:overlap} We generate synthetic data from two populations whose underlying weight probability matrices ${\bm \theta}$ overlap around the middle set of nodes, but where population $1$ has an elevated pattern of weight values for the first set of nodes, and population $2$ in the final set. Figure \ref{struc3} in Appendix \ref{sec:struc} shows this structure. We simulated 200 entities per population, with 100 nodes for each network. Given the structure, the weights of the edges were distributed according to a multivariate Zipf distribution \cite{paretomulti2002}. See Appendix section \ref{sec:struc} for more details. \vspace{1mm} \noindent \textbf{Synthetic data (Heteromogeneous):} \label{syn:time} Using the same population structure as above, we also construct a time-varying dataset where each individual has four time points, resulting in four different graphs per entity. In this case, we have 50 entities and 100 nodes for each network. Given the dependency structure, the weights for each entity at each time point were distributed according to a multivariate Zipf distribution. Figure \ref{struc_time} in Appendix \ref{sec:struc} shows these structures. We use this dataset to compare the behavior of the NC-F and NC-M models under different scenarios. \vspace{1mm} \noindent \textbf{Real data:} \noindent {\em Twitter:} Brazil has recently faced the worst economic/political crisis of its republic years. People were largely split into two sides: one who argued for the impeachment of the now former president, Dilma Rousseff, and the opposition who claimed that the process was a government coup. We crawled public Twitter posts from April 6th to May 31st 2016, using hashtags from both sides to collect tweets. The resulting dataset consists of $7,447$ users (entities), $4,233$ for the proposition and $3,214$ for opposition. In order to have appropriate data for the heterogeneous setting, we also split the dataset into time intervals, with each user having a network for every two weeks of tweets. We call this dataset ``Twitter time''. In this dataset, consisting of users with at least 15 days of tweets, we have a total of $2,098$ users, $1,255$ for proposition and $843$ for opposition. Figure \ref{fig:net_twitter} shows sample co-occurrence networks from the two sides: Proposition and Opposition. Each edge-weight indicates the number of tweets of a user $n$ containing two words (nodes) in a time interval $t$. \noindent {\em Instagram:} We collected public Instagram comments with hashtags referring to the two most common drugs to treat ADHD (Adderall and Ritalin) and Depression (Prozac and Zoloft). These medications all have additional uses (and consequently symptoms), for instance, Adderall is known for loss of appetite, and as an aid for academic performance. Our dataset consists of $65$ users with $44,408$ posts for \#adderall, $21$ with $17,466$ for \#ritalin, $111 $ with $129,405$ for \#prozac, and $35$ with $29,357$ for \#zoloft. \noindent {\em fMRI brain images:} Functional magnetic resonance imaging (fMRI) captures activity in the brain by measuring blood flow from one region of the brain to another. We used the MRN-111 dataset\footnote{\url{http://openconnecto.me/data/public/MR/}} which consists of functional magnetic resonance images (fMRI) for 114 subjects (entities). As in \cite{desikan2006automated} we used a total of 68 brain regions, 34 from the left hemisphere and 34 from the right. Nodes represent brain regions, and weights, white matter density across nodes. We compare brain activity across characteristics like Sex ($Male$ vs $Female$), and personality traits like creative level ($\le 90$ vs $\ge 111$), extroversion ($\le 30$ vs $\ge 35$). Values for creative level (CCI) and extroversion are given from a psychometric scale determined by the corresponding scientific literature, those thresholds were chosen to illustrate a clear \textit{LowVsHigh} setting. Figure \ref{fig:net_twitter} shows sample brain networks of the MRN111 dataset for female and male individuals. We observe significant variability in these weights, suggesting that thresholding can lead to loss of information. \subsection{Baselines} We compared our NC-F and NC-M methods with the following baselines: \begin{enumerate} \item LDA (topic modeling)~\cite{blei2003latent}: This treats each entity as a document made up of `topics' (each corresponding to a distribution over word-count patterns). \item N-gram language model \cite{katz1987estimation}: We use observed bigrams frequencies to estimate co-occurrence probabilities. \item DD network model~\cite{durante2016}: As stated previously, DD forms a special instance of our more general framework for unweighted networks. In order to apply DD, we need first threshold the weighted network observations. We do so using the following criterion \cite{simas2015distance}: $$ p_{ij} =\frac{\text{co-occurances between words $i$ and $j$}} {\text{min (counts of words $i$ and $j$)}}$$ Then $A_{n}[i,j]=1$ if $p_{ij}\!>\!threshold$ for a chosen threshold level. \end{enumerate} Since N-gram and LDA do not directly allow us to estimate $P(H_1)$, we use a Kolmogorov-Smirnov test on the words distribution to perform an overall hypothesis test between populations for the N-gram model, and a chi-square test for topic assignments across populations for LDA. \subsection{Hyperparameters tunning} DD and our method require setting the number of clusters $H$, the dimensionality of the low-rank factorization $R$, the Dirichlet concentration parameters $\alpha$ for $\bm{\beta}$, and the prior probability of $P(H_1)$. For our experiments, we fixed $H=15$, $R=10$, $\alpha = 1/H$ and $P(H_1) = 0.5$. We found that $H=15$ was more than enough clusters for all instances, larger numbers resulting in empty clusters. Most of our experiments focus on settings with count-valued weights, and in the case of word co-occurrences, the weight between words $i$ and $j$ is bounded by the smaller of the number of occurrences of the two words $i$ and $j$. For this setting, we therefore used the logit link and the binomial likelihood. On the other hand, the Brain dataset has count-valued weights with unbounded support, and we used Exponential link and the Poisson likelihood. Our results were fairly robust to hyperparameters settings. For our MCMC algorithm, we observed good mixing properties, and used $1300$ Gibbs samples with an extra $200$ burn-in samples. For LDA and N-Gram, we used settings following implementations from \cite{rehurek_lrec} and \cite{ngrampkg}, respectively. \subsection{Results} \vspace{1mm} \noindent \textbf{Type-I error and statistical power:} Type-I errors or false positives arise when a model incorrectly marks two populations as different when actually the null hypothesis is true, i.e., $P(H_1| H_1 \text{is false})$. Ideally, type-I error rates { should be } $0.05$ or less. Statistical power shows if the models can correctly determine when the populations are different, and $P(H_1 | H_1 \text{ is true})$ should be close to $1$. Measuring these quantities requires access to ground truth. For the Brazil dataset, the disparity of political tendencies between opposition and proposition is clear enough that we treat it as ground truth (For the other real datasets, we do not have ground truth available). We consider four sample sizes for the synthetic data: $50$, $100$, $200$ and $300$. For the twitter data, we consider three sample sizes: $105 \: (\sim\!\!5\%)$, $420 \: (\sim\!\!20\%)$ and $1049\: (\sim\!\!50\%)$. For DD which requires thresholding, we used four different threshold-levels, $0$, $25\%$, $50\%$ and $75\%$. For all methods, we compute $P(H_1| -)$ under different settings. In order to estimate variance $P(H_1| -)$, we generated 20 datasets for each sample size. Figure \ref{fig:syn_power} presents the Type-I error (i.e., $P(H_1| H_1 \text{is false})$) and power curves (i.e., $P(H_1| H_1 \text{is true})$) for increasing sample sizes for each method for synthetic data (left) and real world data (right). Each data point shown is the mean of the 20 trials for the respective sample size, we also present the range of $5\%$ to $95\%$ percentiles. We see that NC-F has the best overall performance both when $H_1$ is true and for $H_1$ false in the homogeneous scenario. N-gram has the worst performance overall so we do not consider this baseline in the real data. LDA have a good overall power, however it performs poorly as far as Type I error goes, with the largest variance. Further, LDA is not able to capture probabilistic structure underlying the data (we discuss this later). DD's performance is not so good for Type I error in the synthetic data, but it is good overall for power. This is not the case for the real data though, where power is poor and varies significantly with threshold. This sensitivity to threshold-level confirms the original motivation for this work. NC-M outperformed all other methods for the heterogeneous data, and NC-F was the second best even though it does not account for heterogeneity. This is due to the fact that our method accounts for the weight distributions, and thus can handle overdispersed counts relatively well. We also investigate how the number of active users in the dataset affects the overall power performance. First, we define ``activeness'' as a function of number of days a user tweeted. We created datasets restricting to users with at least $2$, $5$ and $10$ days of tweets. Thus, we assume that in the dataset with users having at least $10$ days of tweets, there are a significantly larger number of active users than in dataset with a minimum of $2$ days of tweets. Note that we are assessing only statistical power here, therefore we are only looking to the case that $H_1$ is true, i.e., ``PropositionVsOpposition''. Figure \ref{fig:power_active_sample} shows the power curve for each these datasets. As expected, our method needs fewer observations to find statistical difference between populations with more active users in the dataset, in other words NC-F has the ability to differentiate between populations easier as the proportion of active users increases. DD does not improve with larger sample size which suggests DD loses the ability to determine whether populations differ in a higher heterogeneous setting. \begin{figure}[htbp] \hspace{-8mm} \includegraphics[width=0.53\textwidth]{figure/power_across_samplesizes} \caption{Power curves of NC-F and DD (multiple threshold levels) for increasing proportion of active users} \label{fig:power_active_sample} \vspace{2mm} \end{figure} \vspace{1mm} \noindent \textbf{Population-level hypothesis test:} In the previous results, we had a glimpse of the sensibility of threshold choice in terms of decision making on the hypothesis testing procedure. Here, we aim to analyze this further. We estimate the posterior probability of $H_1$ for \textit{all observations} of all the datasets. We compare our results with DD for 10 different threshold levels $(0,10\%,20\%,...,90\%)$. Note that NC-F and NC-M do not vary with threshold level. Here, in addition to the Twitter data, we include results testing the fMRI dataset---comparing populations based on the creative index (CCI). \cite{durante2016} found a significant difference in Brain connectivity between non creative individuals ($CCI\le90$) vs creative subjects ($CCI\ge111$). In their tests, the graphs were thresholded at $0$. Figure \ref{fig:ph1_all} shows DD represented as red solid squares, NC-F and NC-M as blue and green lines, respectively. Again, DD exhibits sensitivity to the threshold choice making inferences unreliable. For instance, if we consider testing whether populations Proposition and Opposition are significantly different and use a 60\% threshold for Twitter time, we would reject the null, since the posterior of $P(H_1)\approx 1$. However, if we slightly change the threshold level to $50\%$, $P(H_1)\approx 0.5$ meaning that there is not enough evidence to support that they are statistically different and we accept the null. The same behavior can be seen on the fMRI dataset where the threshold at $0$ is statistical significant for \textit{LowVsHigh}, however it is not for any other threshold. Overall, we found that our methods NC-M and NC-F are more reliable, since they avoid the need for practitioners to make sensitive preprocessing choices. \begin{figure}[h] \centering \includegraphics[width=0.45\textwidth]{figure/ph1_all.pdf} \caption{$P(H_1 | - )$ across threshold levels for three datasets for when $H_1$ is false (left column) and $H_1$ is true (right column). NC-F and NC-M are represented as blue and green lines, respectively.} \label{fig:ph1_all} \vspace{2mm} \end{figure} \vspace{1mm} \noindent \textbf{Edge-specific level hypothesis test:} \label{sec:edge_results} Another important task is that of retrieving the structure of the co-occurrences probabilities. For better visualization, we generated a version of the synthetic homogeneous with 20 nodes, and we look at differences between true and predicted edge probability matrices for both populations, i.e.\ we compute the estimated difference $\bar{\bm{\theta}}_1 - \bar{\bm{\theta}}_2$ for each model and compared with the ground truth. In Figure~\ref{edgeprob}, we see that our proposed framework accurately recovers the structure of the ground truth. The DD0 also retrieves the structure of population 1, however it performs poorly for population 2. This is related to the sensitivity of results to the threshold-level, suggesting this needs to be chosen carefully across different scenarios. Our models NC-F and NC-M both do not require such hand-tuning, and further exploit values of the pre-thresholded counts for more accurate inference. Unsurprisingly, all the other models fail to correctly learn the structure used to simulate the data. For the Twitter and Instagram datasets, we looked at each edge, and identified those that are different using a $0.1$ significance level. Figure \ref{lochypbra} shows that the NC-F model was able to capture a clear pattern of significantly different use of words among populations for the Brazil dataset, as opposed to the other modeling schemes which look almost random. For the Instagram drug data, the edge-specific hypothesis testing matrix structure is much more significant compared with the Twitter case. One reason for this is that there is a group of Ritalin users that are German, and their words differ from others. \begin{figure}[h] \hspace{-8mm} \includegraphics[width=0.54\textwidth]{figure/edgeprob} \vspace{-6mm} \caption{Edge-specific probabilities difference} \label{edgeprob} \vspace{-4mm} \end{figure} \begin{figure}[h] \centering \begin{subfigure}{0.2\textwidth} \hspace{-1.1cm} \centering \includegraphics[trim={4mm 0 0 0},clip]{figure/lochypbra} \end{subfigure} \begin{subfigure}{0.2\textwidth} \centering \hspace{-1cm} \includegraphics[trim={4mm 0 0 0},clip]{figure/lochypdrug} \end{subfigure} \vspace{-2mm} \caption{Edge-specific tests for Twitter (left) and Instagram (right).} \label{lochypbra} \vspace{2mm} \end{figure} \vspace{1mm} \noindent \textbf{Exploratory analysis:} Here, we show some additional insights that our methods are capable of capturing. One interesting fact of the Brazil political scenario is that many high frequency words were extensively used across both populations, examples being ``impeachmentja'' (impeachment now), ``lavajato'' (carwash), ``golpenao'' (no coup), ``direitos'' (rights). However, using the probability structure $\bar{\bm{\theta}}$ estimated from our framework, we can make some interesting insights about how the two sides frame the issues differently. Figure \ref{swingwords} plots the difference of the link probability for each high frequency word used in conjunction (co-occurrence) with all other words, across the two sides---if the value is larger than zero then it is a `proposition expression', otherwise it is an `opposition expression'. For instance, ``lavajato'' is the name of the investigation and if it is used with ``motivo'' (motive) is a proposition statement where if it is used with ``luta'' (fight) then it is a clear opposition one. From Figure \ref{swingwords} it is clear that the two sides use sets of words (e.g., phrases) quite differently. \begin{figure}[!h] \vspace{3mm} \hspace{-2mm} \centering \includegraphics[width=.50\textwidth]{figure/swingwords_new} \caption{Swing words: differences in edge probabilities for example high frequency terms.} \label{swingwords} \vspace{3mm} \end{figure} Figure \ref{fig:ph1_all_explo} presents additional results for the Instagram and fMRI datasets. In this case, we look again to the behavior of $P(H_1|-)$ across threshold levels. It is important to highlight that we do not have a ground truth information to compare our findings with, however it is an additional set of results to explore the assessment of significant difference between two populations, and to note the lack of robustness of DD wrt threshold choice. \begin{figure}[h] \centering \includegraphics[width=.45\textwidth]{figure/ph1_all_exploratory.pdf} \caption{$P(H_1 | - )$ across threshold levels for Instagram and fMRI datasets.} \label{fig:ph1_all_explo} \vspace{-4mm} \end{figure} \section{Conclusion} \label{sec:con} This paper presents the first steps towards routine and systematic hypothesis testing on populations of weighted networks. Our statistical framework applies both to settings where entities from each population have single graphs associated with them, as well as settings where each entity has associated a set of graphs (we call these without and with within-population heterogeneity). Through a flexible and general clustering mechanism for replicated weighted networks, our framework offers a powerful and accurate hypothesis testing at three levels: population-level, entity-specific and edge-specific. We applied our model to study communication behavior on real social media data (Instagram and Twitter), as well as for brain connectivity data. We saw that by not relying on a a user-specified threshold, our proposed method offers robustness over the methodology of~\cite{durante2016}, besides outperforming other baselines like LDA, N-gram language models. \section{Appendix} \subsection{Synthetic data details} \label{sec:struc} \begin{figure}[tb] \centering \includegraphics[width=0.45\textwidth]{figure/struc3} \caption{Structure used to simulate homogeneous data } \label{struc3} \end{figure} \begin{figure} \centering \includegraphics[width=0.45\textwidth]{figure/struc_time} \caption{Structure used to simulate heterogeneous data } \label{struc_time} \end{figure} The synthetic data was generated to have the same construction of the Twitter dataset set, i.e. a set word co-occurrences networks. In this sense, each node is a word and each edge is a co-occurrence of two words. Moreover, each pair of words can co-occur at most the minimum occurrence of each individual word. In other words, say words $i$ and $j$ occurred $x$ and $y$ times, respectively, then the co-occurrence of words $i$ and $j$ is at most $\min(x,y)$. Hence, in order to generate graphs in this setting, we need to have the individual occurrences of all the words and probability structure for the co-occurrences. Since Zipf's law \cite{zipf} , or discrete Pareto, is almost always used to describe words frequencies, we used a multivariate Zipf generating process \cite{paretomulti2002} to generate the individual counts. Equation \ref{jeh} shows how to generate multivariate Zipf's values, $\bm{X}_n \in \mathbb{R}^{V}$ is the vector with all the individual counts of $V$ words for entity $n$. In our case, we considered the standard Zipf where $\lambda = 1$. \begin{equation} \begin{aligned}[c] p &\sim Beta(1,\lambda)\\ \bm{X}_n|p &\stackrel{ind}{\sim} Geo(p) \end{aligned} \Rightarrow \begin{aligned}[c] \bm{X}_n &\sim M^{(m)}Zipf(IV)(0,\lambda,1,1) \end{aligned} \label{jeh} \end{equation} The structures were arbitrarily chosen to have a clear difference across populations. Figure \ref{struc3} shows the structure used to simulate the homogeneous data used on the experiments section and Figure \ref{struc_time} shows the structure used to simulate for the heterogeneous data. Given the individual counts and the probability of each edge, we simulate edge count using Binomial distribution. Formally, $A_{nij} \sim Bin\left(\min\left({X_{ni},X_{nj}}\right),\bm{\theta}_{ij}\right)$, where $A_{nij}$ is the co-occurrence of words $i$ and $j$, $X_{ni}$ and $X_{nj}$ are their individual counts and $\bm{\theta}_{ij}$ is the probability of $i$ and $j$ co-occur once.

a945bfd2a851fff78eedc9e2e48266933fbd5a69

\section{Introduction} \label{intro} The celebrated \emph{hook-length formula} gives an elegant product expression for the number of standard Young tableaux of fixed shape $\lambda$: $$f^\lambda = \frac{|\lambda|!}{\prod_{u \in [\lambda]} h(u)}$$ \medskip Here $h(u)$ is the hook length of the cell $u$. \medskip The formula gives dimensions of irreducible representations of the symmetric group. The formula was discovered by Frame, Robinson and Thrall in~\cite{FRT} based on earlier results of Young \cite{You}, Frobenius \cite{Fro} and Thrall \cite{Thr}. Since then, it has been reproved, generalized and extended in several different ways, and applied in a number of fields ranging from algebraic geometry to probability, and from group theory to the analysis of algorithms. \medskip Interestingly, an identical formula (when hook lengths are defined appropriately) also holds for strict partitions and standard Young tableaux of \emph{shifted} shape. Even though it was discovered before the formula for straight shapes \cite{Thr}, it is typically considered the ``lesser'' of the two: less well known, less interesting, and with more complicated proofs. Shortly after the famous Greene-Nijenhuis-Wilf probabilistic proof of the ordinary hook length formula \cite{gnw}, Sagan \cite{sagan} extended the argument to the shifted case. The proof needs a careful analysis of special cases and a delicate double induction, and therefore lacks the intuitiveness of the original hook-walk proof. In 1995, Krattenthaler \cite{krat1} provided a bijective proof. While short, it is very involved, as it needs a variant of Hillman-Grassl algorithm, a bijection that comes from Stanley's $(P,\omega)$-partition theory, and the involution principle of Garsia and Milne. A few years later, Fischer \cite{fischer} gave the most direct proof of the formula, in the spirit of Novelli-Pak-Stoyanovskii's \cite{nps} bijective proof of the ordinary hook-length formula. At almost 50 pages in length, the proof is very involved. Bandlow \cite{Ban} gave a short proof via interpolation, and there is a variant of the hook-walk proof in \cite{kon:shifted}; again, special cases need to be considered, and the bijection is hard to describe succinctly. There are also many generalizations of both formulas, such as the $q$-version of Kerov~\cite{Ker1}, and its further generalizations and variations (see~\cite{GH,Ker2} and also \cite{ckp}). There are also a great number of proofs of the more general Stanley's hook-content formula (see e.g.~\cite[Corollary 7.21.4]{ec2}), see for example \cite{rw, krat2, krat3}. \medskip There is no (known) product formula for the number of standard Young tableaux of skew shape (straight or shifted), even though some formulas have been known for a long time. In 2014, Hiroshi Naruse \cite{naru} presented and outlined a proof of a remarkable cancellation-free generalization for skew shapes, both straight and shifted. Here we present the two formulas for shifted shapes. \medskip The \emph{shifted diagram of type B}, $[\lambda]^B$, of a strict partition $\lambda$ (i.e., a partition with distinct parts) is the set of cells $\{(i,j) \colon 1 \leq i \leq \ell(\lambda), i \leq j \leq i + \lambda_i - 1\}$. The \emph{shifted diagram of type D}, $[\lambda]^D$, of a strict partition $\lambda$ is the set of cells $\{(i,j) \colon 1 \leq i \leq \ell(\lambda), i + 1 \leq j \leq i + \lambda_i\}$. \medskip An \emph{excited move of type B} is the move of a cell $(i,j)$ of a shifted diagram to position $(i+1,j+1)$, provided that the cells $(i+1,j)$, $(i,j+1)$ and $(i+1,j+1)$ are not in the diagram. An \emph{excited move of type D} is the move of: \begin{itemize} \item a \emph{non-diagonal} cell $(i,j)$, $i + 1 < j$, of a shifted diagram to position $(i+1,j+1)$, provided that the cells $(i+1,j)$, $(i,j+1)$ and $(i+1,j+1)$ are not in the diagram, or \item a \emph{diagonal} $(i,i+1)$ cell of a shifted diagram diagonally to position $(i+2,i+3)$, provided that the cells $(i,i+2)$, $(i,i+3)$, $(i+1,i+2)$, $(i+1,i+3)$ and $(i+2,i+3)$ are not in the diagram. \end{itemize} Let $\p E^B(\lambda/\mu)$ (respectively, $\p E^D(\lambda/\mu)$) denote the set of all \emph{excited diagrams} of shifted shape $\lambda/\mu$, diagrams in $[\lambda]^B$ (resp., $[\lambda]^D$) obtained by taking the diagram of $\mu$ and performing series of excited moves of type B (resp., type D) in all possible ways. They were introduced by Ikeda and Naruse \cite{IN1}. \medskip Naruse's formula says that \begin{align} f^{\lambda/\mu} &= |\lambda/\mu|! \sum_{D \in \p E^B(\lambda/\mu)} \prod_{u \in [\lambda]^B \setminus D} \frac 1{h^B(u)} = \frac{|\lambda/\mu|!}{\prod_{u \in [\lambda]^B } h^B(u)} \sum_{D \in \p E^B(\lambda/\mu)} \prod_{u \in D} h^B(u) \label{eqB} \\ &= |\lambda/\mu|! \sum_{D \in \p E^D(\lambda/\mu)} \prod_{u \in [\lambda]^D \setminus D} \frac 1{h^D(u)} = \frac{|\lambda/\mu|!}{\prod_{u \in [\lambda]^D } h^D(u)} \sum_{D \in \p E^D(\lambda/\mu)} \prod_{u \in D} h^D(u) \label{eqD} \end{align} where all the hook lengths are evaluated in the diagram of $\lambda$. \medskip Our main result (Theorem \ref{thm1}) are the following formulas, valid for strict partitions $\lambda$, $\mu$ and for commutative variables $x_i$: \begin{align} \left( \sum_{k \in \p W^B(\mu,\lambda)} x_k \right) \sum_{D \in \p E^B(\lambda/\mu)} \prod_{(i,i) \in D} x_i \prod_{\substack{(i,j) \in D \\ i < j}} (x_i + x_j) &= \sum_{\mu \lessdot \nu \subseteq \lambda} \sum_{D \in \p E^B(\lambda/\nu)} \prod_{(i,i) \in D} x_i \prod_{\substack{(i,j) \in D \\ i < j}} (x_i + x_j) , \label{eqmainB} \\ \left( \sum_{k \in \p W^D(\mu,\lambda)} x_k \right) \sum_{D \in \p E^D(\lambda/\mu)} \prod_{(i,j) \in D} (x_i + x_j) &= \sum_{\mu \lessdot \nu \subseteq \lambda} \sum_{D \in \p E^D(\lambda/\nu)} \prod_{(i,j) \in D} (x_i + x_j), \label{eqmainD} \end{align} where $\p W^B(\mu,\lambda)$ and $\p W^D(\mu,\lambda)$ are certain finite subsets of positive integers. The formulas specialize to recursive versions of equations \eqref{eqB} and \eqref{eqD}. \medskip Like in \cite{kon:naruse}, the proof of both formulas is bijective and uses a simple bumping algorithm. Moreover, note that: \begin{itemize} \item if $\mu$ is the staircase shape $(\ell(\lambda),\ldots,1)$, the formulas \eqref{eqB} and \eqref{eqD} become the classical hook-length formula for straight (non-skew, non-shifted) shapes \item if $\mu$ is, more generally, a strict partition of the same length as $\lambda$, the formulas \eqref{eqB} and \eqref{eqD} become the hook-length formula for (straight) skew shapes (also due to Naruse) \item if $\mu = \emptyset$, the formulas \eqref{eqB} and \eqref{eqD} become the (also classical) hook-length formula for shifted shapes \end{itemize} Our algorithms thus also give bijective proofs of all these five important formulas. Let us emphasize that the non-skew cases (i.e, non-shifted and shifted) are not related in the sense that neither formula implies the other. \medskip The fact that the algorithm is so easy to describe (it takes less than a page!) is quite remarkable. One would hope that this fact will help the shifted hook-length formula overcome its ``lesser'' status; indeed, how can it be lesser if it is a straightforward generalization with a bijective proof that is not more complicated in any way? \medskip Of course, simplicity of definition does not imply that an algorithm is computationally efficient. We prove that the number of steps needed is $2^{\Omega(\sqrt n)}$ for partitions $\lambda,\mu$ of size at most $n$; see Theorem \ref{complexity}. \medskip This is a natural sequel to \cite{kon:naruse} and will only sketch some of the proofs that are similar. A reader who is familiar with \cite{kon:naruse} should note the following differences: \begin{itemize} \item there is only one set of variables, not two, see Definition \ref{xi}, \item the linear factors on the left of equations \eqref{eqmainB} and \eqref{eqmainD} are slightly different from the straight-shape case \item in \eqref{eqmainB}, there are monomials appearing in the product (not just binomials) \item the insertion algorithm always starts from the top row \end{itemize} \medskip In Section \ref{def}, we give basic definitions and notation. In Section \ref{identity}, we motivate equation \eqref{eqmainB} and \eqref{eqmainD} and show how they imply \eqref{eqB} and \eqref{eqD}. In Section \ref{bijection}, we use versions of the bumping algorithm on tableaux to prove the identities bijectively. In Section \ref{weighted}, we show weighted versions of formulas \eqref{eqB} and \eqref{eqD}. In Section \ref{proofs}, we show examples and sketch the proofs of the technical statements from Section \ref{bijection}. \section{Basic definitions and notation} \label{def} A \emph{partition} is a weakly decreasing finite sequence of positive integers $\lambda = (\lambda_1,\lambda_2,\ldots,\lambda_\ell)$. We say that $\lambda$ is a \emph{strict partition} if the sequence is strictly decreasing. We call $|\lambda| = \lambda_1 + \cdots + \lambda_\ell $ the \emph{size} of $\lambda$ and $\ell = \ell(\lambda)$ the \emph{length} of $\lambda$. We write $\lambda_i = 0$ for $i > \ell(\lambda)$. The \emph{shifted diagrams of types B and D} of $\lambda$ are $$[\lambda]^B = \{(i,j) \colon 1 \leq i \leq \ell(\lambda), i \leq j \leq \lambda_i + i - 1\}$$ and $$[\lambda]^D = \{(i,j) \colon 1 \leq i \leq \ell(\lambda), i < j \leq \lambda_i + i\}.$$ Note that the difference is that the columns start with $2$ in type D. \medskip We call the elements of $[\lambda]^B$ and $[\lambda]^D$ the \emph{cells} of $\lambda$. For partitions $\mu$ and $\lambda$, we say that $\mu$ is \emph{contained} in $\lambda$, $\mu \subseteq \lambda$, if $[\mu]^B \subseteq [\lambda]^B$ (which is equivalent to $[\mu]^D \subseteq [\lambda]^D$). In that case, we say that $\lambda/\mu$ is a \emph{skew shape} of size $|\lambda/\mu| = |\lambda| - |\mu|$, and the diagrams of $\lambda/\mu$ are $[\lambda/\mu]^B = [\lambda]^B \setminus [\mu]^B$ and $[\lambda/\mu]^D = [\lambda]^D \setminus [\mu]^D$. We write $\mu \lessdot \lambda$ if $\mu \subseteq \lambda$ and $|\lambda/\mu| = 1$. In this case, we also say that $\lambda$ \emph{covers} $\mu$. Note that diagrams and cover relations are different than in the context of (not necessarily strict) partitions, while containment is the same. \medskip We often represent a strict partition $\lambda$ graphically, with a cell $(i,j)$ in $[\lambda]^B$ or $[\lambda]^D$ represented by a unit square in position $(i,j)$. In this paper, we use English notation, so for example the Young diagram of the partition $\lambda = (6,5,3,2)$ is \medskip \begin{center} \ytableausetup{smalltableaux} \begin{ytableau} {} & {} & {} & {} & {} & {} \\ \none & {} & {} & {} & {} & {} \\ \none & \none & {} & {} & {} \\ \none & \none & \none & {} & {} \end{ytableau} \end{center} \medskip We often omit parenthesis and commas, so we could write $\lambda = 6532$. Also note that the cells appear in rows $1,\ldots,\ell(\lambda)$ and either in columns $1,\ldots,\lambda_1$ or in columns $2,\ldots,\lambda_1+1$, depending on the type we are interested in. \medskip A \emph{corner} of $\lambda$ is a cell that can be removed from $[\lambda]^B$, i.e., a cell $(i,j) \in [\lambda]^B$ satisfying $(i+1,j),(i,j+1) \notin [\lambda]^B$. An \emph{outer corner} of $\lambda$ is a cell that can be added to $[\lambda]^B$, i.e., a cell $(i,j) \notin [\lambda]^B$ satisfying $i = 1$ or $(i-1,j) \in [\lambda]^B$, and $j = 1$ or $(i,j-1) \in [\lambda]^B$. The partition $6532$ has (in type B) corners $(2,6)$ and $(4,5)$, and outer corners $(1,7)$, $(3,6)$ and $(5,5)$. In type D, add $1$ to the second coordinate in all cases. \medskip For a strict partition $\lambda$, denote by $\lambda_i'$ the number of cells in the $i$-th column of $[\lambda]^B$. For $\lambda = 6532$, we get $\lambda'_i = i$ for $i = 1,\ldots,4$, $\lambda'_5=4$, $\lambda'_6 = 2$, $\lambda'_i = 0$ for $i \geq 7$. Let us emphasize that $\lambda_i'$ retains the same meaning when we are in type D, so the number of cells in the $i$-th column of $[\lambda]^D$ is $\lambda'_{i-1}$. \medskip The \emph{type B hook length} of the cell $(i,j) \in [\lambda]^B$, $h^B(i,j)$, is defined as: \begin{itemize} \item the number of cells in row $i$ if $i = j$ \item the number of cells in rows $i$ and $j$ if $i < j \leq \ell(\lambda)$ \item the number of cells weakly to the right of $(i,j)$ and below $(i,j)$ if $i < j$ and $j > \ell(\lambda)$ \end{itemize} The definition for $i < j \leq \ell(\lambda)$ is often phrased as ``the number of cells weakly to the right of or below $(i,j)$, and in row $j$, where the cell $(j,j)$ is counted twice''. For example, the hook lengths of the cells $(2,2)$, $(1,4)$ and $(2,5)$ in $[6532]^B$ are $5$, $8$ and $4$: \medskip \begin{center} \ytableausetup{smalltableaux} \begin{ytableau} {} & {} & {} & {} & {} & {} \\ \none & *(gray!70) & *(gray!70) & *(gray!70) & *(gray!70) & *(gray!70) \\ \none & \none & & {} & {} \\ \none & \none & \none & & \end{ytableau} \qquad \qquad \begin{ytableau} {} & & & *(gray!70) & *(gray!70) & *(gray!70) \\ \none & {} & & *(gray!70) & & {} \\ \none & \none & & *(gray!70) & {} \\ \none & \none & \none & *(black) & *(gray!70) \end{ytableau} \qquad \qquad \begin{ytableau} {} & {} & {} & {} & {} & {} \\ \none & {} & {} & & *(gray!70) & *(gray!70) \\ \none & \none & {} & & *(gray!70) \\ \none & \none & \none & & *(gray!70) \end{ytableau} \end{center} \medskip The black square is counted twice. \medskip The \emph{type D hook length} of the cell $(i,j) \in [\lambda]^D$ is defined as the number of cells weakly to the right of $(i,j)$, below $(i,j)$, and in row $j$. For example, the hook lengths of the cells $(1,4)$ and $(2,5)$ in $[6532]^D$ are $8$ and $5$: \medskip \begin{center} \ytableausetup{smalltableaux} \begin{ytableau} {} & {} & *(gray!70) & *(gray!70) & *(gray!70) & *(gray!70) \\ \none & {} & *(gray!70) & {} & {} & {} \\ \none & \none & *(gray!70) & {} & {} \\ \none & \none & \none & *(gray!70) & *(gray!70) \end{ytableau} \qquad \qquad \qquad \begin{ytableau} {} & {} & {} & {} & {} & {} \\ \none & {} & {} & *(gray!70) & *(gray!70) & *(gray!70) \\ \none & \none & {} & *(gray!70) & {} \\ \none & \none & \none & *(gray!70) & {} \end{ytableau} \end{center} \medskip In both cases, the \emph{B/D hook} is the set of cells used to define the hook length. \medskip A \emph{standard Young tableau} (or \emph{SYT} for short) of shifted shape $\lambda$ is a bijective map $T$ from the diagram of $\lambda$ to $\{1,\ldots,|\lambda|\}$, $(i,j) \mapsto T_{ij}$, satisfying $T_{ij} < T_{i,j+1}$ if $(i,j),(i,j+1)$ are in the diagram and $T_{ij} < T_{i+1,j}$ if $(i,j),(i+1,j)$ are in the diagram. The number of SYT's of shifted shape $\lambda$ is denoted by $f^\lambda$, and the set of standard Young tableaux of shape $\lambda/\mu$ by $\SYT(\lambda/\mu)$. The type does not matter in this definition. The following illustrates $f^{42} = 5$: \medskip \begin{center} \ytableausetup{centertableaux} \ytableausetup{smalltableaux} \begin{ytableau} 1 & 2 & 3 & 4\\ \none & 5 & 6 \end{ytableau} \qquad \begin{ytableau} 1 & 2 & 3 & 5 \\ \none & 4 & 6 \end{ytableau} \qquad \begin{ytableau} 1 & 2 & 3 & 6 \\ \none & 4 & 5 \end{ytableau} \qquad \begin{ytableau} 1 & 2 & 4 & 5 \\ \none & 3 & 6 \end{ytableau} \qquad \begin{ytableau} 1 & 2 & 4 & 6 \\ \none & 3 & 5 \end{ytableau} \end{center} \medskip The \emph{shifted hook-length formula} gives a product expression for the number of standard Young tableaux: $$f^\lambda = \frac{|\lambda|!}{\prod_{u \in [\lambda]} h^B(u)} = \frac{|\lambda|!}{\prod_{u \in [\lambda]} h^D(u)}$$ The formulas are equivalent: one multiset of hook lengths is a permutation of the other. For example, $f^{42} = \frac{6!}{4 \cdot 6 \cdot 3 \cdot 1 \cdot 2 \cdot 1} = \frac{6!}{6 \cdot 4 \cdot 3 \cdot 1 \cdot 2 \cdot 1} = 5$. \medskip We can define a standard Young tableau of skew shifted shape $\lambda/\mu$ analogously. The number of SYT's of shifted shape $\lambda/\mu$ is denoted by $f^{\lambda/\mu}$. The following illustrates $f^{543/2} = 12$: \begin{center} \ytableausetup{centertableaux} \ytableausetup{smalltableaux} \begin{ytableau} *(gray!70) & *(gray!70) & 1 & 2 \\ \none & 3 & 4 & 5 \\ \none & \none & 6 & 7 \end{ytableau} \qquad \begin{ytableau} *(gray!70) & *(gray!70) & 1 & 2 \\ \none & 3 & 4 & 6 \\ \none & \none & 5 & 7 \end{ytableau} \qquad \begin{ytableau} *(gray!70) & *(gray!70) & 1 & 3 \\ \none & 2 & 4 & 5 \\ \none & \none & 6 & 7 \end{ytableau} \qquad \begin{ytableau} *(gray!70) & *(gray!70) & 1 & 3 \\ \none & 2 & 4 & 6 \\ \none & \none & 5 & 7 \end{ytableau} \qquad \begin{ytableau} *(gray!70) & *(gray!70) & 1 & 4 \\ \none & 2 & 3 & 5 \\ \none & \none & 6 & 7 \end{ytableau} \qquad \begin{ytableau} *(gray!70) & *(gray!70) & 1 & 4 \\ \none & 2 & 3 & 6 \\ \none & \none & 5 & 7 \end{ytableau} \vspace{0.3cm} \begin{ytableau} *(gray!70) & *(gray!70) & 1 & 5 \\ \none & 2 & 3 & 6 \\ \none & \none & 4 & 7 \end{ytableau} \qquad \begin{ytableau} *(gray!70) & *(gray!70) & 2 & 3 \\ \none & 1 & 4 & 5 \\ \none & \none & 6 & 7 \end{ytableau} \qquad \begin{ytableau} *(gray!70) & *(gray!70) & 2 & 3 \\ \none & 1 & 4 & 6 \\ \none & \none & 5 & 7 \end{ytableau} \qquad \begin{ytableau} *(gray!70) & *(gray!70) & 2 & 4 \\ \none & 1 & 3 & 5 \\ \none & \none & 6 & 7 \end{ytableau} \qquad \begin{ytableau} *(gray!70) & *(gray!70) & 2 & 4 \\ \none & 1 & 3 & 6 \\ \none & \none & 5 & 7 \end{ytableau} \qquad \begin{ytableau} *(gray!70) & *(gray!70) & 2 & 5 \\ \none & 1 & 3 & 6 \\ \none & \none & 4 & 7 \end{ytableau} \end{center} \medskip Suppose that $D \subseteq [\lambda]^B$. If $(i,j) \in D$ and $(i+1,j),(i,j+1),(i+1,j+1) \in [\lambda]^B \setminus D$, then an \emph{excited move of type B} with respect to $\lambda$ is the replacement of $D$ with $D' = D \setminus \{(i,j)\} \cup \{(i+1,j+1)\}$. For strict partitions $\lambda,\mu$, then an \emph{excited diagram} of type B and shape $\lambda/\mu$ is a diagram contained in $[\lambda]^B$ that can be obtained from $[\mu]^B$ with a series of excited moves. Let $\p E^B(\lambda/\mu)$ denote the (empty unless $\mu \subseteq \lambda$) set of all excited diagrams of type B and shape $\lambda/\mu$. The following shows $\p E^B(432/2)$: \medskip \begin{center} \ytableausetup{smalltableaux} \begin{ytableau} *(gray!70) & *(gray!70) & & \\ \none & & & \\ \none & \none & &\\ \end{ytableau} \qquad \begin{ytableau} *(gray!70) & & & \\ \none & & *(gray!70) & \\ \none & \none & &\\ \end{ytableau} \qquad \begin{ytableau} *(gray!70) & & & \\ \none & & & \\ \none & \none & & *(gray!70)\\ \end{ytableau} \qquad \begin{ytableau} {} & & & \\ \none & *(gray!70) & *(gray!70) & \\ \none & \none & & \\ \end{ytableau} \qquad \begin{ytableau} {} & & & \\ \none &*(gray!70) & & \\ \none & \none & & *(gray!70)\\ \end{ytableau} \qquad \begin{ytableau} {} & & & \\ \none & & & \\ \none & \none & *(gray!70) & *(gray!70)\\ \end{ytableau} \end{center} \medskip If $(i,j) \in D$ for $i + 1 < j$, $(i+1,j),(i,j+1),(i+1,j+1) \in [\lambda]^D \setminus D$, then an \emph{excited move of type D} with respect to $\lambda$ is the replacement of $D$ with $D' = D \setminus \{(i,j)\} \cup \{(i+1,j+1)\}$. On the other hand, if $(i,i+1) \in D$, $(i,i+2),(i,i+3),(i+1,i+2),(i+1,i+3),(i+2,i+3) \in [\lambda]^D \setminus D$, an excited move of type D with respect to $\lambda$ is the replacement of $D$ with $D' = D \setminus \{(i,i+1)\} \cup \{(i+2,i+3)\}$. If $\mu \subseteq \lambda$, then an \emph{excited diagram} of type D and shape $\lambda/\mu$ is a diagram contained in $[\lambda]^D$ that can be obtained from $[\mu]^D$ with a series of excited moves. Let $\p E^D(\lambda/\mu)$ denote the set of all excited diagrams of type D and shape $\lambda/\mu$. The following shows $\p E^D(432/2)$: \medskip \begin{center} \ytableausetup{smalltableaux} \begin{ytableau} *(gray!70) & *(gray!70) & & \\ \none & & & \\ \none & \none & &\\ \end{ytableau} \qquad \begin{ytableau} *(gray!70) & & & \\ \none & & *(gray!70) & \\ \none & \none & &\\ \end{ytableau} \qquad \begin{ytableau} *(gray!70) & & & \\ \none & & & \\ \none & \none & & *(gray!70)\\ \end{ytableau} \qquad \begin{ytableau} {} & & & \\ \none & & & \\ \none & \none & *(gray!70) & *(gray!70)\\ \end{ytableau} \end{center} \medskip Recall from the introduction that Naruse's formula says that \begin{equation*} f^{\lambda/\mu} = \frac{|\lambda/\mu|!}{\prod_{u \in [\lambda] } h^B(u)} \sum_{D \in \p E^B(\lambda/\mu)} \prod_{u \in D} h^B(u) = \frac{|\lambda/\mu|!}{\prod_{u \in [\lambda] } h^D(u)} \sum_{D \in \p E^D(\lambda/\mu)} \prod_{u \in D} h^D(u) \end{equation*} where all the hook lengths are evaluated in $[\lambda]^B$ (resp., $[\lambda]^D$). \medskip For example, the formula confirms that \begin{multline*} f^{432/2} = \frac{7!}{4 \cdot 7 \cdot 6 \cdot 3 \cdot 3 \cdot 5 \cdot 2 \cdot 2 \cdot 1} \left( 4 \cdot 7 + 4 \cdot 5 + 4 \cdot 1 + 3 \cdot 5 + 3 \cdot 1 + 2 \cdot 1 \right) \\ = \frac{7!}{7 \cdot 6 \cdot 4 \cdot 3 \cdot 5 \cdot 3 \cdot 2 \cdot 2 \cdot 1} \left( 7 \cdot 6 + 7 \cdot 3 + 7 \cdot 1 + 2 \cdot 1 \right) = 12 \end{multline*} As we can see, the formulas are no longer equivalent. \section{A polynomial identity} \label{identity} It is clear that both sides of \eqref{eqB} and \eqref{eqD} are equal to $1$ if $\lambda = \mu$. Since the minimal entry of a shifted standard Young tableau of shape $\lambda/\mu$ must be in an outer corner of $\mu$ which lies in $\lambda$, we have $f^{\lambda/\mu} = \sum_{\mu \lessdot \nu \subseteq \lambda} f^{\lambda/\nu}$, where $\sum_{\mu \lessdot \nu \subseteq \lambda}$ denotes the sum over all strict partitions $\nu$ that are contained in $\lambda$ and cover $\mu$. If we show that the right-hand sides of \eqref{eqB} and \eqref{eqD} satisfy the same recursion, we are done. Therefore the statements are equivalent to the following identities: \begin{align} (|\lambda| - |\mu|) \sum_{D \in \p E^B(\lambda/\mu)} \prod_{u \in D} h^B(u) & = \sum_{\mu \lessdot \nu \subseteq \lambda} \sum_{D \in \p E^B(\lambda/\nu)} \prod_{u \in D} h^B(u). \label{eq2B} \\ (|\lambda| - |\mu|) \sum_{D \in \p E^D(\lambda/\mu)} \prod_{u \in D} h^D(u) & = \sum_{\mu \lessdot \nu \subseteq \lambda} \sum_{D \in \p E^D(\lambda/\nu)} \prod_{u \in D} h^D(u). \label{eq2D} \end{align} \begin{definition}\label{xi} Define variables $x_i^B, x_i^D$, $i = 1,2,\ldots$, as follows: \begin{itemize} \item if $i \leq \ell(\lambda)$, then $x_i^B = x_i^D$ is the number of cells in the $i$-th row of diagram $[\lambda]$; in other words, $x_i^B = x_i^D = \lambda_i$; \item if $i > \ell(\lambda)$, then $x_i^B$ is the number of cells in the $i$-th column of diagram $[\lambda]^B$, decreased by $i$, and $x_i^D$ is the number of cells in the $i$-th column of diagram $[\lambda]^D$ (which is the same as the number of cells in the $(i-1)$-th column of $[\lambda]^B$), decreased by $i-1$; in other words, $x_i^B = \lambda_{i}' - i$ and $x_i^D = \lambda_{i-1}'-i+1$. \end{itemize} \end{definition} For example, for $\lambda = 432$, we have $x^B_1 = x^D_1 = 4$, $x^B_2 = x^D_2 = 3$, $x^B_3 = x^D_3 = 2$, $x^B_4 = -1$, $x^D_4 = 0$, $x^B_i = -i$ for $i \geq 5$, $x^D_5 = -1$, $x^D_i = -i+1$ for $i \geq -6$. It is clear that the sequences $x^B_1,x^B_2,\ldots$ and $x^D_1,x^D_2,\ldots$ are strictly decreasing. Also, $x^D_{\ell(\lambda) + 1}$ is always $0$. \medskip It is easy to see that for $(i,j) \in [\lambda]$, we have \begin{align} h^B_{ii} &= x^B_i, & & 1 \leq i \leq \ell(\lambda) \label{hooks1} \\ h^B_{ij} &= x^B_i + x^B_j, & & 1 \leq i \leq \ell(\lambda), \: i < j \leq \lambda_i + i - 1 \label{hooks2} \\ h^D_{ij} &= x^D_i + x^D_j, & & 1 \leq i \leq \ell(\lambda), \: i < j \leq \lambda_i + i \label{hooks3} \end{align} \begin{definition} \label{ws} For arbitrary strict partitions $\lambda,\mu$, let $\p W^B(\mu,\lambda)$ consist of \begin{itemize} \item $k \leq \ell(\lambda)$ for which $\lambda_k \neq \mu_i$ for all $i$, \item $k \geq \ell(\lambda) + 1$ for which $\lambda'_k - k \neq \mu'_i - i$ for all $i$, \end{itemize} and let $\p W^D(\mu,\lambda)$ consist of \begin{itemize} \item $k \leq \ell(\lambda)$ for which $\lambda_k \neq \mu_i$ for all $i$ \item $\ell(\lambda) + 1$ if $\ell(\lambda) - \ell(\mu)$ is odd \item $k \geq \ell(\lambda) + 2$ for which $\lambda'_k - k \neq \mu'_i - i$ for all $i$ \end{itemize} Furthermore, let $\p W(\lambda) = \{x_1,\ldots,x_{\lambda_1+1}\}$. \end{definition} A crucial role will also be played by the following lemma. \begin{lemma} \label{w} For arbitrary strict partitions $\lambda,\mu$ (not necessarily satisfying $\mu \subseteq \lambda$) and $x^B_k$, $x^D_k$ defined as above (for $\lambda$), we have $$|\lambda| - |\mu| = \sum_{k \in \p W^B(\mu,\lambda)} x^B_k = \sum_{k \in \p W^D(\mu,\lambda)} x^D_k.$$ \end{lemma} \begin{proof} Since we have $\lambda'_k -k = \mu'_k -k = -k$ for large $k$, the sum in the lemma is finite. Now note the following: if $j$ appears as a part of a strict partition $\lambda$, $j = \lambda_i$, then $\lambda'_{i+j-1} \geq i$ and $\lambda'_{i+j} < i$. That means that $x^B_{i+j-1} = \lambda_{i+j-1}-i-j+1 \geq -j + 1$ and $x^B_{i+j} = \lambda_{i+j}-i-j < -j$. In other words, $-j$ does \emph{not} appear among $x^B_k$'s. On the other hand, if $j$ does not appear as a part of $\lambda$, there exists $i \geq 0$ so that $\lambda_{i+1} < j < \lambda_i$. Then $\lambda'_{i+j} = i$ and $x^B_{i+j} = i - i - j = -j$. We have proved that $\{\lambda_1,\ldots,\lambda_{\ell(\lambda)}\} = \{-x^B_k \colon k > \ell(\lambda)\}^C$, where the complement is taken with respect to the set of non-negative integers.\\ Clearly, $$|\lambda| - |\mu| = \sum_{k=1}^{\ell(\lambda)} \lambda_k - \sum_{k=1}^{\ell(\mu)} \mu_k = \sum_{\substack{k \leq \ell(\lambda) \\ \nexists i \colon \lambda_k = \mu_i}} \lambda_k - \sum_{\substack{k \leq \ell(\mu) \\ \nexists i \colon \mu_k = \lambda_i}} \mu_k.$$ Now $$\{\mu_1,\ldots,\mu_{\ell(\mu)}\} \setminus \{\lambda_1,\ldots,\lambda_{\ell(\lambda)}\} = \{\lambda_1,\ldots,\lambda_{\ell(\lambda)}\}^C \setminus \{\mu_1,\ldots,\mu_{\ell(\mu)}\}^C = \{ -x^B_k \colon k > \ell(\lambda), \lambda'_k-k \neq \mu_i'-i\},$$ which proves the first statement. The proof for type D is very similar. As mentioned before, $x^D_{\ell(\lambda)+1} = 0$, so including it or not does not influence the sum. \end{proof} \begin{example} Take $\lambda = 432$ and $\mu = 2$. Since the parts $4$ and $3$ of $\lambda$ are not parts of $\mu$, while $2$ is, we have $1,2 \in \p W^B(432,2)$, $3 \notin \p W^B(432,2)$. Since $$(\lambda'_k-k)_{k \geq 4} = (-1,-5,-6,-7,\ldots), \qquad (\mu'_j-j)_{j \geq 1} = (0,-1,-2,-3,\ldots)$$ none of the integers $4,5,\ldots$ are in $\p W^B(432,2)$. Indeed, $|\lambda| - |\mu| = x^B_1 + x^B_2 = 7$. We similarly get $|\lambda| - |\mu| = x^D_1 + x^D_2 = 7$.\\ For $\lambda = 865321$ and $\mu = 431$, we have $\ell(\lambda)-\ell(\mu) = 3$, $$ (\lambda'_k-k)_{k \geq 7} = (-4,-7,-9,-10,-11,\ldots), \qquad (\mu'_j-j)_{j \geq 1} = (0,0,0,-2,-5,-6,-7,\ldots)$$ and $|\lambda| - |\mu| = x^B_1 + x^B_2 + x^B_3 + x^B_5 + x^B_7 = 8 + 6 + 5 + 2 + (-4) = 17$ and $|\lambda| - |\mu| = x^D_1 + x^D_2 + x^D_3 + x^D_5 + x^D_7 + x^D_8 = 8 + 6 + 5 + 2 + 0 + (-4) = 17$. \end{example} The following theorem is our main result. It gives two subtraction-free polynomial identities, which, by equations \eqref{hooks1}--\eqref{hooks3} and Lemma \ref{w}, specialize to equations \eqref{eq2B} and \eqref{eq2D} when $x_i = x_i^B$ (resp., $x_i = x_i^D$), and therefore imply the hook-length formulas for skew shifted diagrams. \begin{theorem} \label{thm1} For arbitrary strict partitions $\lambda$, $\mu$ and commutative variables $x_i$, we have \begin{align*} \left( \sum_{k \in \p W^B(\mu,\lambda)} x_k \right) \sum_{D \in \p E^B(\lambda/\mu)} \prod_{(i,i) \in D} x_i \prod_{\substack{(i,j) \in D \\ i < j}} (x_i + x_j) &= \sum_{\mu \lessdot \nu \subseteq \lambda} \sum_{D \in \p E^B(\lambda/\nu)} \prod_{(i,i) \in D} x_i \prod_{\substack{(i,j) \in D \\ i < j}} (x_i + x_j), \\ \left( \sum_{k \in \p W^D(\mu,\lambda)} x_k \right) \sum_{D \in \p E^D(\lambda/\mu)} \prod_{(i,j) \in D} (x_i + x_j) &= \sum_{\mu \lessdot \nu \subseteq \lambda} \sum_{D \in \p E^D(\lambda/\nu)} \prod_{(i,j) \in D} (x_i + x_j). \end{align*} \end{theorem} The theorem is trivially true for $\mu \not\subseteq \lambda$, as then both sides of both equations are equal to $0$. \begin{example} For $\lambda = 432$ and $\mu = 2$, we have the identities \begin{multline*} (x_1+x_2)(x_1 (x_1 + x_2) + x_1 (x_2 + x_3) + x_1 (x_3 + x_4) + x_2 (x_2 + x_3) + x_2 (x_3 + x_4) + x_3 (x_3 + x_4)) \\ = x_1 (x_1 + x_2) (x_1 + x_3) + x_1 (x_1 + x_2) (x_2 + x_4) + x_1 (x_2 + x_3) (x_2 + x_4) + x_2 (x_2 + x_3) (x_2 + x_4) \\ + x_1 (x_1 + x_2) x_2 + x_1 (x_1 + x_2) x_3 + x_1 (x_2 + x_3) x_3 + x_2 (x_2 + x_3) x_3\end{multline*} and \begin{multline*} (x_1+x_2) \left( (x_1 + x_2) (x_1 + x_3) + (x_1 + x_2) (x_2 + x_4) + (x_1 + x_2) (x_3 + x_5) + (x_3 + x_4) (x_3 + x_5) \right) \\ = (x_1 + x_2) (x_1 + x_3) (x_2 + x_3) + (x_1 + x_2) (x_1 + x_3) (x_1 + x_4)\\ + (x_1 + x_2) (x_1 + x_3) (x_2 + x_5) + (x_1 + x_2) (x_2 + x_4) (x_2 + x_5). \end{multline*} For $\lambda = 865321$ and $\mu = 431$, the first term on the left-hand side of the first equality is $x_1 + x_2 + x_3 + x_5 + x_7$, the second term is a sum of 4,992 monomials, and the right-hand side is a sum of 24,960 monomials. The first term on the left of the second equality is $x_1 + x_2 + x_3 + x_5 + x_7 + x_8$, the second term is a sum of 9,472 monomials, and the right-hand side is a sum of 56,832 monomials. \end{example} The (bijective) proof of Theorem \ref{thm1} is the content of the next section. \section{The bijections} \label{bijection} We interpret the two sides of equations \eqref{eqmainB} and \eqref{eqmainD} in terms of certain tableaux. To motivate the definition, look at the following excited diagrams of type B and D for $\lambda = 865321$ and $\mu = 431$: \medskip \begin{center} \ytableausetup{smalltableaux} \ytableausetup{aligntableaux=top} \begin{ytableau} *(gray!70) & *(gray!70) & & & & & & \\ \none & & & *(gray!70) & *(gray!70) & &\\ \none & \none & *(gray!70) & *(gray!70) & & & \\ \none & \none & \none & & & *(gray!70) \\ \none & \none & \none & \none & *(gray!70) & \\ \none & \none & \none & \none & \none & \\ \end{ytableau} \qquad\qquad \begin{ytableau} *(gray!70) & *(gray!70) & & & & & & \\ \none & *(gray!70)& & *(gray!70) & *(gray!70) & &\\ \none & \none & & *(gray!70) & & & \\ \none & \none & \none & & & *(gray!70) \\ \none & \none & \none & \none & *(gray!70) & \\ \none & \none & \none & \none & \none & \\ \end{ytableau} \end{center} \medskip Instead of actually moving the cells of $\mu$, write an integer in a cell of $\mu$ that indicates how many times it moves (diagonally) from the original position. For the above example, we get the following tableaux of shape $\mu = 431$: \medskip \begin{center} \ytableausetup{boxsize=1.25em} \ytableausetup{aligntableaux=top} \begin{ytableau} 0 & 0 & 1 & 1 \\ \none & 1 & 1 & 2 \\ \none & \none & 2 \end{ytableau} \qquad\qquad \begin{ytableau} 0 & 0 & 1 & 1 \\ \none & 0 & 1 & 2 \\ \none & \none & 2 \end{ytableau} \end{center} \medskip Rules for excited moves mean that the non-negative integer entries of the resulting tableau are weakly increasing along rows and columns. In type D, the entries on the diagonal must be even. Also, every tableau with non-negative integer entries and weakly increasing rows and columns corresponds to a valid excited diagram, provided that (in type D only) the entries on the diagonal are even, and the entry $r$ in row $i$ and column $j$ satisfies $j \leq \lambda_{i+r} + i - 1$ in type B and $j \leq \lambda_{i+r} + i$ in type D. Furthermore, it is enough to check this inequality only for the corners of $\mu$. \medskip The contributions $$ \prod_{(i,i) \in D} x_i \prod_{\substack{(i,j) \in D \\ i < j}} (x_i + x_j), \qquad \prod_{(i,j) \in D} (x_i + x_j),$$ of type B/D excited diagram $D$ can be written as $$ \prod_{(i,i) \in |\mu]^B} x_{i+T_{ii}} \prod_{\substack{(i,j) \in [\mu]^B \\ i < j}} (x_{i+T_{ij}} + x_{j+T_{ij}}), \qquad \prod_{(i,j) \in [\mu]^D} (x_{i+T_{ij}} + x_{j+T_{ij}}),$$ where $T$ is the corresponding tableau of shape $\mu$ with non-negative integer entries and weakly increasing rows and columns. To extract the monomials from the product, choose either $x_{i+T_{ij}}$ or $x_{j+T_{ij}}$ for each $(i,j)$ in $[\mu]^B$ or $[\mu]^D$, $i < j$, and multiply with $\prod_{(i,i) \in |\mu]^B} x_{i+T_{ii}}$ in type B. Write the number $T_{ij}$ in position $(i,j)$, $i < j$, in black if we choose $x_{i+T_{ij}}$, and in red if we choose $x_{j+T_{ij}}$. The number in $(i,i)$ in type B is always black. Call a tableau with non-negative integer black or red entries (red entries not allowed on the diagonal in type B, and odd entries not allowed on the diagonal in type D) and weakly increasing rows and columns a \emph{(shifted) bicolored tableau} of type B/D. Denote by $\p B^B(\mu)$ and $\p B^D(\mu)$ the (infinite unless $\mu = \emptyset$) set of shifted bicolored tableaux of type B/D and shape $\mu$, and denote by $\p B^B(\mu,\lambda)$ and $\p B^D(\mu,\lambda)$ the (finite) set of shifted bicolored tableaux $T$ of type B/D and shape $\mu$ that satisfy $j \leq \lambda_{i+T_{ij}} + i - 1$ in type B and $j \leq \lambda_{i+T_{ij}} + i$ in type D. \medskip The weight of a shifted bicolored tableau $T$ of shape $\mu$ is \begin{align*} w^B(T) &= \prod_{(i,j) \in b(T)} x_{i + T_{ij}} \prod_{(i,j) \in [\mu]^B \setminus b(T)} x_{j + T_{ij}}\\ w^D(T) &= \prod_{(i,j) \in b(T)} x_{i + T_{ij}} \prod_{(i,j) \in [\mu]^D \setminus b(T)} x_{j + T_{ij}} \end{align*} where $b(T)$ is the set of black entries of $T$. \begin{example} The following are some bicolored tableaux in $\p B^B(431)$. A bicolored tableau is in $\p B^B(431,865321)$ if and only if $T_{25} \leq 2$, $T_{34} \leq 2$, so the first two are in $\p B^B(431,865321)$ and the third is not. \medskip \begin{center} \ytableausetup{boxsize=1.25em} \ytableausetup{aligntableaux=top} \ytableaushort {0{\rd 0}0{\rd 1},{\none}001,{\none}{\none}{0}} \qquad \qquad \ytableaushort {{0}{\rd 0}11,{\none}11{\rd 2},{\none}{\none}2} \qquad \qquad \ytableaushort {{0}{\rd 0}{\rd 0}{\rd 2},{\none}{0}{1}{3},{\none}{\none}{2}} \end{center} \medskip \noindent The weights of these tableaux are $x_1^2x_2^3x_3^2x_5$, $x_1x_2^3x_3^2x_5x_6$, and $x_1x_2^2x_3^2x_5^2x_6$, respectively.\\ The following are some bicolored tableaux in $\p B^D(431)$. A bicolored tableau is in $\p B^D(431,865321)$ if and only if $T_{26} \leq 2$, $T_{35} \leq 2$, so the first two are in $\p B^D(431,765521)$ and the third is not. \medskip \begin{center} \ytableausetup{boxsize=1.25em} \ytableausetup{aligntableaux=top} \ytableaushort {0{\rd 0}0{\rd 1},{\none}001,{\none}{\none}{\rd 0}} \qquad \qquad \ytableaushort {{\rd 0}{\rd 0}11,{\none}01{\rd 2},{\none}{\none}2} \qquad \qquad \ytableaushort {{0}{\rd 0}{\rd 0}{\rd 2},{\none}{0}{1}{3},{\none}{\none}{2}} \end{center} \medskip \noindent The weights of these tableaux are $x_1^2x_2^2x_3^2x_4x_6$, $x_2^4x_3^2x_5x_7$, and $x_1x_2x_3^2x_4x_5^2x_7$, respectively. \end{example} \medskip The left-hand sides of equations \eqref{eqmainB} and \eqref{eqmainD} are the enumerators (with respect to weight $w$) of $\p B^B(\mu,\lambda) \times \p W^B(\mu,\lambda)$ and $\p B^D(\mu,\lambda) \times \p W^D(\mu,\lambda)$. The right-hand sides are the enumerators of the sets $\bigcup_{\nu} \p B^B(\nu,\lambda)$ and $\bigcup_{\nu} \p B^D(\nu,\lambda)$, where the union is over all strict partitions $\nu$ that cover $\mu$ and are contained in $\lambda$. In the remainder of this section, we present a weight-preserving bijection between the two sides in both cases. \medskip Let us first describe the \emph{bump of type B}. This is an algorithm that takes as input a shifted bicolored tableau $S$ of shape $\mu$, a cell $(i,j) \in [\mu] \cup \{(0,0)\}$, a direction $d \in \{ \downarrow, \rightarrow\}$, and a variable index $k$. Except when $i = j = 0$, we assume that $k$ can be in position $(i,j)$, either as a black $k-i$ if $d = \: \rightarrow$ or a red $k-j$ if $d = \: \downarrow$. As output, it produces a new tableau $S'$ (that differs from $S$ in at most one cell), a cell $(i',j')$ (which is either one column to the right or one row below $(i,j)$), a new direction $d' \in \{\downarrow, \rightarrow, \circ\}$, and a new variable index $k'$ ($\infty$ if $d' = \circ$). New direction being $\circ$ means that we must terminate the bumping process. Note that if $s = (0,0)$, $d$ will always be $\downarrow$. Without loss of generality, we can assume that $S$ fills the entire plane: write $\infty$ to the right and below the diagram, and $0$ above and to the left. \medskip If direction $d$ is $\rightarrow$, we move one column to the right; in other words, we set $j' = j+1$. Let $i'$ be the largest possible index so that we can insert variable $x_k$ as a black number $k-i'$ in position $(i',j')$; in other words, we must have $S_{i'-1,j'} \leq k - i' \leq S_{i'+1,j'}$, and $i'$ is the largest index with this property. Such $i'$ always exists and is at most $i$, as we will see in Section \ref{proofs}; hence $i' \leq i \leq j < j'$ and $(i',j')$ is never on the diagonal. It can happen that $(i',j')$ lies outside of $|\mu]$. In that case, $S'$ is the tableau we get if we write a black $k-i'$ in position $(i',j')$, $d' = \circ$, and $k = \infty$. If $(i',j') \in [\mu]$, $x_k$ bumps out whatever is in position $(i',j')$ as follows: $S'$ is the tableau we get if we replace the entry in position $(i',j')$ by a black $k-i'$. If $S_{i'j'}$ is black, the new direction $d'$ is $\rightarrow$, and the new variable index $k'$ is the entry in position $(i',j')$ of $S$, i.e.~$S_{i'j'} + i'$. If $S_{i'j'}$ is red, the new direction $d'$ is $\downarrow$ and the new variable index is $k' = S_{i'j'}+j'$. \medskip If direction $d$ is $\downarrow$, we move one row down; in other words, we set $i' = i+1$. Let $j' \geq i'$ be the largest possible index satisfying $S_{i',j'-1} \leq k - j' \leq S_{i',j'+1}$. It can happen that $(i',j')$ lies outside of $|\mu]$. In that case, $S'$ is the tableau we get if we write a red $k-j'$ in position $(i',j')$ (black $k-j' = k - i'$ if $i' = j' = \ell(\mu) +1$), $d' = \circ$, and $k = \infty$. If $(i',j') \in [\mu]$, $x_k$ bumps out whatever is in position $(i',j')$ as follows: $S'$ is the tableau we get if we replace the entry in position $(i',j')$ by a red $k-j'$ (black $k-j' = k - i'$ if $i' = j'$). If $S_{i'j'}$ is black, the new direction $d'$ is $\rightarrow$, and the new variable index $k'$ is the entry in position $(i',j')$ of $S$, i.e.~$S_{i'j'} + i'$. If $S_{i'j'}$ is red, the new direction $d'$ is $\downarrow$ and the new variable index is $k' = S_{i'j'}+j'$. \medskip In summary, a black number is bumped one column to the right and stays black, and a red number is bumped one row down and stays red unless it lands on the diagonal. \medskip The \emph{bump in type D} is very similar; now the diagonal entries can be red, but they must be even. Of course, in type D, the diagonal cell is of the form $(i,i+1)$, so a black entry in an even row gives an even index variable and a red entry in an even row gives an odd index variable (and the other way round for odd rows). So if we bump the index variable $k$ to the right or down and land on $(i,i+1)$, we write a black $k-i$ if that is an even number, and a red $k-i-1$ otherwise. Also, while red numbers are generally bumped down, a red number on the diagonal is bumped to the right. We leave out the details (but see a detailed example in Section \ref{proofs}). \medskip The \emph{insertion process} (in type B or D) takes as input a shifted bicolored tableau $S$ of appropriate type and shape $\mu$ and a variable index $k$. It first performs the bump (of type B or D) for $S$, $(i,j) = (0,0)$, $d = \: \downarrow$, and $k$, with output $S'$, $(i',j')$, $d'$, and $k'$. If $d' = \circ$, the insertion process returns $S'$; otherwise, it performs the bump for $S'$, $(i',j')$, $d'$, and $k'$, with output $S''$, $(i'',j'')$, $d''$ and $k''$. It continues in this way until the direction is $\circ$. \medskip Finally, the \emph{repeated insertion process} (in type B or D) takes as input a bicolored tableau in $\p B^B(\mu,\lambda)$ or $\p B^D(\mu,\lambda)$ and a variable index $k$ in $\p W^B(\mu,\lambda)$ or $\p W^D(\mu,\lambda)$. It performs the insertion process with $S$ and $k$. If the resulting tableau $S'$ is in $\p B(\lambda,\nu)$ for $\nu$, $\mu \lessdot \nu \subseteq \lambda$, the repeated insertion process outputs $S'$. Otherwise, let $k'$ be the variable index represented by the entry in $[\nu] \setminus [\mu]$. Perform the insertion process with $S'$ and $k'$ as inputs. If the resulting tableau $S''$ is in $\p B(\lambda,\nu)$ for $\nu$, $\mu \lessdot \nu \subseteq \lambda$, the repeated insertion process outputs $S''$. Otherwise, we continue in this manner. \medskip It is not obvious that the bump, the insertion process, and the repeated insertion process are well defined. However, we will sketch the proof of the following theorem in Section \ref{proofs}. \begin{theorem} \label{welldefined} The indices $i'$, $j'$ and $k'$ in the bump of type B or D are well defined, and $S'$ is a well-defined bicolored tableau.\\ The insertion processes of type B and D described above always terminate and are weight-preserving bijections $$\psi^B_\mu \colon \p B^B(\mu) \times \{x_1,x_2,x_3,\ldots \} \longrightarrow \bigcup_{\nu} \p B^B(\nu),$$ $$\psi^D_\mu \colon \p B^D(\mu) \times \{x_1,x_2,x_3,\ldots \} \longrightarrow \bigcup_{\nu} \p B^D(\nu),$$ where the unions are over all strict partitions $\nu$ which cover $\mu$.\\ The repeated insertion processes of type B and D described above always terminate and are weight-preserving bijections $$\Psi^B_{\mu,\lambda} \colon \p B^B(\mu,\lambda) \times \p W^B(\mu,\lambda) \longrightarrow \bigcup_{\nu} \p B^B(\nu,\lambda),$$ $$\Psi^D_{\mu,\lambda} \colon \p B^D(\mu,\lambda) \times \p W^D(\mu,\lambda) \longrightarrow \bigcup_{\nu} \p B^D(\nu,\lambda),$$ where the unions are over all strict partitions $\nu$ which cover $\mu$ and are contained in $\lambda$. \end{theorem} This theorem proves \eqref{eqmainB} and \eqref{eqmainD} and hence the hook-length formulas for skew shapes, equations \eqref{eqB} and \eqref{eqD}. \medskip While both the bump and the insertion process are efficient, repeated insertion can be very slow. Indeed, in Section \ref{proofs}, we prove the following. \begin{theorem} \label{complexity} For every positive integer $m$, there exists a shifted bicolored tableau $S$ of type B and shape $\mu = (m+2,m,m-1,m-2\ldots,1)$ so that repeated insertion of $2$ into $S$ with respect to $\lambda = (m+2,m+1,m-1,m-2\ldots,1)$ needs $2^m$ insertions. \end{theorem} It is worthwhile to consider what happens when $\ell(\lambda) = \ell(\mu)$. In that case, $\lambda/\mu$ can be interpreted as a skew non-shifted shape, and we can compute $f^{\lambda/\mu}$ using Naruse's formula for straight shapes. Of course, the formulas \eqref{eq2B} and \eqref{eq2D} give an equivalent result: we can have no excited moves on or above the diagonal, so the contribution of these cells on both sides can be canceled. \ref{w} \section{A weighted generalization of the hook-length formula} \label{weighted} After reading \cite{kon:naruse}, Darij Grinberg (personal communication) pointed out that a certain weighted gen\-er\-al\-i\-za\-tion (\cite[Theorem 18]{hopkins}) of the hook-length formula has a Naruse-style extension to skew shapes, and that it can be proved by taking a different specialization of the main polynomial identity from \cite{kon:naruse}. It turns out that the skew shifted shapes also allow for this generalization. \medskip Recall that the \emph{content} of a cell $(i,j)$ in a Young diagram of a strict partition $\lambda$ is the integer $j-i$, which we denote $c_\lambda(i,j)$ (or simply $c(i,j)$ if it is clear what partition we are considering). Note that in type B, the diagonal cells have content $0$, while in type D, they have content $1$. Let $z_k$, $k \in {\mathbb Z}_{\geq 0}$, be some commutative variables. Define weighted hook lengths $h^B(u;{\bf z})$ and $h^D(u;{\bf z})$ of a cell $u$ in a diagram of $\lambda$ of type B/D as the sum of $z_{c_\lambda(i,j)}$ over all cells $(i,j)$ in the type B/D hook of $u$. For example, the weighted hook lengths of the cells $(2,2)$, $(1,4)$ and $(2,5)$ in $[6532]^B$ are $z_0 + z_1 + z_2 + z_3 + z_4 + z_5$, $2z_0 + 2z_1 + z_2 + z_3 + z_4 + z_5$, and $z_1 + z_2 + z_3 + z_4$. On the other hand, the weighted hook lengths of the cells $(1,4)$ and $(2,5)$ in $[6532]^D$ are $2z_1 + 2z_2 + z_3 + z_4 + z_5 + z_6$ and $z_1 + z_2 + z_3 + z_4 + z_5$. Of course, if $z_k = 1$ for all $k$, the weighted hook length becomes the usual hook length. \medskip Furthermore, for a standard Young tableau $T$ of skew shifted shape $\lambda/\mu$, $|\lambda|-|\mu| = n$, define $$T^B_{\bf z} = \prod_{k=1}^n \frac 1{z_{c_\lambda(T^{-1}(n))} + z_{c_\lambda(T^{-1}(n-1))} + \cdots + z_{c_\lambda(T^{-1}(k))}},$$ where $T^{-1}(k)$ is the unique cell of the tableau with entry $k$, and we take contents in type B. We define $T^D_{\bf z}$ analogously (and we obtain it from $T^B_{\bf z}$by shifting all indices by $1$). If $z_k = 1$ for all $k$, $T^B_{\bf z} = T^D_{\bf z} = 1/n!$. For example, if $T$ is the tableau $$\begin{ytableau} *(gray!70) & *(gray!70) & 2 & 3 \\ \none & 1 & 4 & 5 \\ \none & \none & 6 & 7 \end{ytableau}$$ then $$T^B_{\bf z} = \frac 1{(2z_0+2z_1+2z_2+z_3)(z_0+2z_1+2z_2+z_3)(z_0+2z_1+z_2+z_3)(z_0+2z_1+z_2)(z_0+z_1+z_2)(z_0 + z_1)z_1},$$ $$T^D_{\bf z} = \frac 1{(2z_1+2z_2+2z_3+z_4)(z_1+2z_2+2z_3+z_4)(z_1+2z_2+z_3+z_4)(z_1+2z_2+z_3)(z_1+z_2+z_3)(z_1 + z_2)z_2}.$$ \begin{theorem} \label{z} For a strict partition $\lambda$, we have \begin{align} \sum_{T \in \SYT(\lambda/\mu)} T^B_{\bf z} &= \sum_{D \in \p E^B(\lambda/\mu)} \prod_{u \in [\lambda]^B \setminus D} \frac 1{h^B(u;\bf z)} = \frac{1}{\prod_{u \in [\lambda]^B } h^B(u;\bf z)} \sum_{D \in \p E^B(\lambda/\mu)} \prod_{u \in D} h^B(u;\bf z) \\ \sum_{T \in \SYT(\lambda/\mu)} T^D_{\bf z}&= \sum_{D \in \p E^D(\lambda/\mu)} \prod_{u \in [\lambda]^D \setminus D} \frac 1{h^D(u;\bf z)} = \frac{1}{\prod_{u \in [\lambda]^D } h^D(u;\bf z)} \sum_{D \in \p E^D(\lambda/\mu)} \prod_{u \in D} h^D(u;\bf z) \end{align} \end{theorem} \begin{proof}[Sketch of proof] We only sketch the proof for type B; type D and all details are left as an exercise for the reader. the term in $T^B_{\bf z}$ corresponding to $k = 1$ equals $1/(\sum_{u \in [\lambda/\mu]^B} z_{c_\lambda(u)}))$ and is therefore independent of the tableau $T$ (it depends only on $\lambda$ and $\mu$). That means that the equivalent recursive form of the first equality is $$\sum_{u \in [\lambda/\mu]^B} z_{c_\lambda(u)}) \sum_{D \in \p E^B(\lambda/\mu)} \prod_{u \in D} h^B(u;\bf z) = \sum_{\mu \lessdot \nu \subseteq \lambda} \sum_{D \in \p E^B(\lambda/\nu)} \prod_{u \in D} h^B(u;\bf z).$$ It turns out that this is just another specialization of the first polynomial identity in Theorem \ref{thm1}.\\ Define $x_i^B(\bf z) = z_0 + \ldots + z_{\lambda_i-1}$ for $i \leq \ell(\lambda)$ and $x_i^B(\bf z) = - z_0 - \ldots - z_{i-\lambda'_i-1}$ for $i > \ell(\lambda)$. Note that when $z_k = 1$ for all $k$, $x_i^B(\bf z) = x_i^B$. It is easy to check that $h^B(i,i;\bf z) = x_i^B(\bf z)$ for $i \leq \ell(\lambda)$, and $h^B(i,j;\bf z) = x_i^B(\bf z) + x_j^B(\bf z)$ for $i \leq \ell(\lambda)$, $i < j \leq \lambda_i + i - 1$. Furthermore, it is a straightforward generalization of Lemma \ref{w} that $$\sum_{u \in [\lambda/\mu]^B} z_{c_\lambda(u)} = \sum_{k \in \p W^B(\mu,\lambda)} x^B_k(\p z).$$ The result follows. \end{proof} \begin{example} For $\lambda = 321$ and $\mu = 1$, the theorem says that \begin{multline*} \frac{1}{z_0 (z_0+z_1) (2 z_0+z_1) (2 z_0+z_1+z_2) (2 z_0+2 z_1+z_2)} \\ + \frac{1}{z_0 (z_0+z_1) (z_0+z_1+z_2) (2 z_0+z_1+z_2) (2 z_0+2 z_1+z_2)} \\ = \frac{1}{z_0 (z_0+z_1) (2 z_0+z_1) (z_0+z_1+z_2) (2 z_0+z_1+z_2) (2 z_0+2 z_1+z_2)} \cdot \Big((z_0+z_1+z_2) + (z_0+z_1) + z_0\Big). \end{multline*} Note that the terms on the left correspond to standard Young tableaux \begin{center} \ytableausetup{centertableaux} \ytableausetup{smalltableaux} \begin{ytableau} *(gray!70) & 1 & 2 \\ \none & 3 & 4 \\ \none & \none & 5 \end{ytableau} \qquad \mbox{and} \qquad \begin{ytableau} *(gray!70) & 1 & 3 \\ \none & 2 & 4 \\ \none & \none & 5 \end{ytableau}, \end{center} while the three terms in the sum on the right correspond to the excited diagrams \begin{center} \ytableausetup{smalltableaux} \begin{ytableau} *(gray!70) & & \\ \none & & \\ \none & \none &\\ \end{ytableau} \qquad \begin{ytableau} {} & & \\ \none & *(gray!70) & \\ \none & \none &\\ \end{ytableau} \qquad \begin{ytableau} {} & & \\ \none & & \\ \none & \none & *(gray!70)\\ \end{ytableau} \qquad \end{center} \end{example} Finally, let us mention that an interesting specialization of Theorem \ref{z} is $z_i = q^i$, which gives a $q$-version of the skew shifted hook-length formula. \section{Examples and proofs} \label{proofs} \subsection*{Examples of bumps, insertion, and repeated insertion} As our first example in type B, take $\lambda = 865321$, $\mu = 431$, and perform repeated insertion with the tableau $$\ytableausetup{boxsize=1.25em} \ytableausetup{aligntableaux=top} \ytableaushort {0{\rd 0}{\rd 1}{\rd 1},{\none}{1}2{2},{\none}{\none}{2}}$$ and $k = 1$. Note that $\lambda_1 = 8$ is not a part of $\mu$, so $1$ is indeed in $\p W^B(\mu,\lambda)$. \medskip By the definition of the insertion process, we first take $(i,j) = (0,0)$ and $d = \: \downarrow$. Because direction is downward, we move one row down, to row $i' = 1$, and find the largest $j'$ so that the number $1-j'$ can be inserted in position $(1,j')$. Of course, the only option is $j' = 1$, and because $(1,1)$ is on the diagonal, we insert a black $0$ in that position. A black $0$ is also bumped out, so the new direction is $\rightarrow$, and the new variable index is $1$. We move one column to the right. The only possible position is $(1,2)$, where we insert a black $0$, and bump the variable $x_2$ (represented by the red $0$) downward. It lands in position $(2,2)$ and bumps out the black $1$. We move to the right. The variable $x_3$ that was bumped can either be represented as a black $2$ in position $(3,1)$ or a black $2$ in position $(3,2)$. Both are possible, so we choose the latter option. A black $2$ is bumped to the right, land in $(2,4)$ and bumps out a black $2$. This one lands in position $(1,5)$ as a black $3$. The bumps are represented in the following figure. The gray cells show the positions of bumped cells. \medskip \begin{center} \ytableausetup{boxsize=1.25em} \ytableausetup{aligntableaux=top} \ytableaushort {{*(gray!25) 00}{\rd 0}{\rd 1}{\rd 1},{\none}{1}2{2},{\none}{\none}{2}} \quad \ytableaushort {{0}{*(gray!25) 0 \rd 0}{\rd 1}{\rd 1},{\none}{1}2{2},{\none}{\none}{2}} \quad \ytableaushort {{0}{0}{\rd 1}{\rd 1}{\none},{\none}{*(gray!25)01}2{2},{\none}{\none}{2}} \ytableaushort {{0}{0}{\rd 1}{\rd 1},{\none}{0}{*(gray!25)12}{2},{\none}{\none}{2}} \quad \ytableaushort {{0}{0}{\rd 1}{\rd 1},{\none}{0}{1}{*(gray!25)22},{\none}{\none}{2}} \quad \ytableaushort {{0}{0}{\rd 1}{\rd 1}{*(gray!70) 3},{\none}{0}{1}{2},{\none}{\none}{2}} \end{center} \medskip Since $(1,5) \notin [431]^B$, the insertion process is done. On the other hand, the repeated insertion process is not, since the entry in position $(1,5)$ violates condition $\leq 2$. Therefore we remove this cell, and insert the variable $x_4$ into the resulting tableau. The insertion process is represented in the following figure. \medskip \begin{center} \ytableausetup{boxsize=1.25em} \ytableausetup{aligntableaux=top} \ytableaushort {{0}{0}{\rd 1}{\rd 1},{\none}{0}{1}{2},{\none}{\none}{2}} \quad\ytableaushort {{0}{0}{*(gray!25) \rd 1\rd 1}{\rd 1},{\none}{0}{1}{2},{\none}{\none}{2}} \quad\ytableaushort {{0}{0}{\rd 1}{\rd 1},{\none}{0}{*(gray!25) {\rd 1}1}{2},{\none}{\none}{2}} \quad\ytableaushort {{0}{0}{\rd 1}{\rd 1},{\none}{0}{\rd 1}{*(gray!25) 12},{\none}{\none}{2}} \quad\ytableaushort {{0}{0}{\rd 1}{\rd 1}{*(gray!70) 3},{\none}{0}{\rd 1}{1},{\none}{\none}{2}} \end{center} \medskip Again, the repeated insertion process is not finished, since the entry in position $(1,5)$ is $> 2$. We remove the cell, and insert $x_4$ into the remaining tableau. The insertion process is now as follows: \medskip \begin{center} \ytableausetup{boxsize=1.25em} \ytableausetup{aligntableaux=top} \ytableaushort {{0}{0}{\rd 1}{\rd 1},{\none}{0}{\rd 1}{1},{\none}{\none}{2}} \quad \ytableaushort {{0}{0}{*(gray!25) \rd 1 \rd 1}{\rd 1},{\none}{0}{\rd 1}{1},{\none}{\none}{2}} \quad \ytableaushort {{0}{0}{\rd 1}{\rd 1},{\none}{0}{*(gray!25) \rd 1 \rd 1}{1},{\none}{\none}{2}} \quad \ytableaushort {{0}{0}{\rd 1}{\rd 1},{\none}{0}{\rd 1}{1},{\none}{\none}{*(gray!25) 12}} \quad \ytableaushort {{0}{0}{\rd 1}{\rd 1},{\none}{0}{\rd 1}{1},{\none}{\none}{1}{*(gray!70) 2}} \end{center} \medskip The last shifted bicolored tableau is in $\p B^B(431,865321)$, so we are done with the repeated insertion process. \medskip Let us mention that if we perform the repeated insertion process on all 24,960 pairs $(S,k)$ in $$\p B^B(431,865321) \times \p W^B(431,865321),$$ the process needs just one insertion in 17,398 cases, two insertions in 6,080 cases, three insertions in 977 cases, four insertions in 455 cases, five insertions in 25 cases, and six insertions in 25 cases. \medskip Based on this, one might conjecture that the number of insertions in the repeated insertion process is bounded by $|\mu|$, $\ell(\lambda)$ or something similar. The next example shows that this is not the case. Take $\lambda = 65321$, $\mu = 64321$, $k = 2$, and the tableau \medskip \begin{center} \ytableausetup{boxsize=1.25em} \ytableausetup{aligntableaux=top} \ytableaushort {{0}{0}{0}{0}{0}, {\none}{\rd 0}{\rd 0}{\rd 0}, {\none}{\none}{0}{0}, {\none}{\none}{\none}{0}} \end{center} \medskip Now we represent each insertion in one diagram, with the gray cells indicating the insertion path. We see that we need 16 insertions. In fact, we will use a generalization of this example to construct an example with $\lambda$ and $\mu$ of (approximately) size $m^2/2$ and length $m$, and $(S,k) \in \p B^B(\mu,\lambda) \times \p W^B(\mu,\lambda)$ that needs $2^m$ steps of the insertion process. \medskip \begin{center} \ytableausetup{boxsize=1.25em} \ytableausetup{aligntableaux=top} \ytableaushort {{0}{*(gray!25) \rd 0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!70)0}, {\none}{0}{\rd 0}{\rd 0}{\rd 0}, {\none}{\none}{0}{0}{0}, {\none}{\none}{\none}{0}{0}, {\none}{\none}{\none}{\none}{0}} \qquad \ytableaushort {{*(gray!25)0}{*(gray!25) 0}{0}{0}{0}{0}, {\none}{*(gray!25)0}{*(gray!25) 0}{\rd 0}{\rd 0}{*(gray!70)1}, {\none}{\none}{*(gray!25)0}{*(gray!25) 0}{*(gray!25) 0}, {\none}{\none}{\none}{0}{0}, {\none}{\none}{\none}{\none}{0}} \qquad \ytableaushort {{0}{0}{*(gray!25) \rd 0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!70)0}, {\none}{0}{0}{\rd 0}{\rd 0}, {\none}{\none}{0}{0}{0}, {\none}{\none}{\none}{0}{0}, {\none}{\none}{\none}{\none}{0}} \qquad \ytableaushort {{*(gray!25)0}{*(gray!25) 0}{*(gray!25) 0}{0}{0}{0}, {\none}{0}{*(gray!25) \rd 0}{*(gray!25) 0}{\rd 0}{*(gray!70)1}, {\none}{\none}{0}{*(gray!25) \rd 0}{*(gray!25) 0}, {\none}{\none}{\none}{0}{0}, {\none}{\none}{\none}{\none}{0}} \vspace{0.3cm} \ytableaushort {{0}{0}{*(gray!25) \rd 0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!70)0}, {\none}{0}{\rd 0}{0}{\rd 0}, {\none}{\none}{0}{\rd 0}{0}, {\none}{\none}{\none}{0}{0}, {\none}{\none}{\none}{\none}{0}} \qquad \ytableaushort {{*(gray!25)0}{*(gray!25) 0}{*(gray!25) 0}{0}{0}{0}, {\none}{0}{*(gray!25) \rd 0}{0}{\rd 0}{*(gray!70)2}, {\none}{\none}{*(gray!25)0}{*(gray!25) 0}{0}, {\none}{\none}{\none}{*(gray!25)0}{*(gray!25) 0}, {\none}{\none}{\none}{\none}{0}} \qquad \ytableaushort {{0}{0}{0}{*(gray!25) \rd 0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!70)0}, {\none}{0}{\rd 0}{0}{\rd 0}, {\none}{\none}{0}{0}{0}, {\none}{\none}{\none}{0}{0}, {\none}{\none}{\none}{\none}{0}} \qquad \ytableaushort {{*(gray!25)0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!25) 0}{0}{0}, {\none}{0}{\rd 0}{*(gray!25) \rd 0}{*(gray!25) 0}{*(gray!70)1}, {\none}{\none}{0}{0}{*(gray!25) \rd 0}, {\none}{\none}{\none}{0}{0}, {\none}{\none}{\none}{\none}{0}} \vspace{0.3cm} \ytableaushort {{0}{0}{*(gray!25) \rd 0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!70)0}, {\none}{0}{\rd 0}{\rd 0}{0}, {\none}{\none}{0}{0}{\rd 0}, {\none}{\none}{\none}{0}{0}, {\none}{\none}{\none}{\none}{0}} \qquad \ytableaushort {{*(gray!25)0}{*(gray!25) 0}{*(gray!25) 0}{0}{0}{0}, {\none}{0}{*(gray!25) \rd 0}{\rd 0}{0}{*(gray!70)2}, {\none}{\none}{*(gray!25)0}{*(gray!25) 0}{*(gray!25) 0}, {\none}{\none}{\none}{0}{*(gray!25) \rd 0}, {\none}{\none}{\none}{\none}{0}} \qquad \ytableaushort {{0}{0}{0}{*(gray!25) \rd 0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!70)0}, {\none}{0}{\rd 0}{\rd 0}{0}, {\none}{\none}{0}{0}{0}, {\none}{\none}{\none}{0}{\rd 0}, {\none}{\none}{\none}{\none}{0}} \qquad \ytableaushort {{*(gray!25)0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!25) 0}{0}{0}, {\none}{0}{\rd 0}{*(gray!25) \rd 0}{0}{*(gray!70)1}, {\none}{\none}{0}{*(gray!25) \rd 0}{*(gray!25) 0}, {\none}{\none}{\none}{0}{\rd 0}, {\none}{\none}{\none}{\none}{0}} \vspace{0.3cm} \ytableaushort {{0}{0}{*(gray!25) \rd 0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!70)0}, {\none}{0}{\rd 0}{\rd 0}{0}, {\none}{\none}{0}{\rd 0}{0}, {\none}{\none}{\none}{0}{\rd 0}, {\none}{\none}{\none}{\none}{0}} \qquad \ytableaushort {{*(gray!25)0}{*(gray!25) 0}{*(gray!25) 0}{0}{0}{0}, {\none}{0}{*(gray!25) \rd 0}{\rd 0}{0}{*(gray!70)3}, {\none}{\none}{*(gray!25)0}{*(gray!25) 0}{0}, {\none}{\none}{\none}{*(gray!25)0}{*(gray!25) 0}, {\none}{\none}{\none}{\none}{0}} \qquad \ytableaushort {{0}{0}{0}{0}{*(gray!25) \rd 0}{*(gray!25) 0}{*(gray!70)0}, {\none}{0}{\rd 0}{\rd 0}{0}, {\none}{\none}{0}{0}{0}, {\none}{\none}{\none}{0}{0}, {\none}{\none}{\none}{\none}{0}} \qquad \ytableaushort {{*(gray!25)0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!25) 0}{*(gray!25) 0}{0}, {\none}{0}{\rd 0}{\rd 0}{*(gray!25) \rd 0}{*(gray!70)0}, {\none}{\none}{0}{0}{0}, {\none}{\none}{\none}{0}{0}, {\none}{\none}{\none}{\none}{0}} \end{center} \medskip Let us show an example in type D. Again, take $\lambda = 865321$, $\mu = 431$, and insert variable $x_1$ (i.e.~$k=1$) into the tableau $$\ytableausetup{boxsize=1.25em} \ytableausetup{aligntableaux=top} \ytableaushort {{\rd 0}{\rd 0}{\rd 0}{\rd 2},{\none}{0}{\rd 1}{2},{\none}{\none}{\rd 2}}$$ Note that $\lambda_1 = 8$ is not a part of $\mu$, so $1$ is indeed in $\p W^D(\mu,\lambda)$. \medskip By the definition of the insertion process, we first take $(i,j) = (0,0)$ and $d = \: \downarrow$. Because direction is downward, we move one row down, to row $i' = 1$, and try to insert variable $x_1$ as a red entry. However, in row $j' \geq 2$, $x_1$ is represented by a red $1-j'$, which is always negative. It is also odd for $j' = 2 = i'+1$, so we can insert it (only) as a black $k-i' = 0$ in position $(i',j') = (1,2)$. The index variable that is bumped out is $k' = 2$ (represented by the red $0$ in column $2$). Even though the entry in $(i',j')$ is red, we bump it to the right (because $j' = i'+1$). In other words, $d' = \: \rightarrow$. The variable $x_2$ can be represented as a black $1$ in row $1$ (impossible, as then the entry in position $(1,3)$ would be larger than the one in position $(2,3)$) or a black $0$ in row $2$. We choose the latter option. In other words, $(i'',j'') = (2,3)$, and the variable bumped out is $x_2$ (note that we just saw that it can happen that the bumped entry remains unchanged), which gets bumped to the right. It lands in position $(2,4)$ and bumps out the red $1$, representing the variable $x_5$. This variable travels downward; it can be represented as a black $2$ in position $(3,4)$ (not red $1$, since odd numbers are forbidden on the diagonal!) or red $0$ in position $(3,5)$. The latter is not allowed, as the entry in $(3,5)$ cannot be smaller than the one in $(3,4)$. So the variable bumps out the red $2$. Since the red $2$, representing the variable $x_6$, was on the diagonal, it goes to the right, and lands in position $(3,5)$ as a black $3$. The bumps are represented in the following figure. The gray cells show the positions of bumped cells. \medskip \begin{center} \ytableausetup{boxsize=1.25em} \ytableausetup{aligntableaux=top} \ytableaushort {{\rd 0}{\rd 0}{\rd 0}{\rd 2},{\none}{0}{\rd 1}{2},{\none}{\none}{\rd 2}} \qquad \ytableaushort {{*(gray!25) 0 \rd 0}{\rd 0}{\rd 0}{\rd 2},{\none}{0}{\rd 1}{2},{\none}{\none}{\rd 2}} \qquad \ytableaushort {{0}{\rd 0}{\rd 0}{\rd 2},{\none}{*(gray!25) 00}{\rd 1}{2},{\none}{\none}{\rd 2}} \qquad \ytableaushort {{0}{\rd 0}{\rd 0}{\rd 2},{\none}{0}{*(gray!25) 0 \rd 1}{2},{\none}{\none}{\rd 2}} \qquad \ytableaushort {{0}{\rd 0}{\rd 0}{\rd 2},{\none}{0}{0}{2},{\none}{\none}{*(gray!25) 2 \rd 2}} \qquad \ytableaushort {{0}{\rd 0}{\rd 0}{\rd 2},{\none}{0}{0}{2},{\none}{\none}{2}{*(gray!70) 3}} \end{center} \medskip Since $(3,5) \notin [431]^D$, the insertion process is done. On the other hand, the repeated insertion process is not, since the entry in position $(3,5)$ violates condition $\leq 2$. Therefore we remove this cell, and insert the variable $x_6$ into the resulting tableau. The insertion process is represented in the following figure. \medskip \begin{center} \ytableausetup{boxsize=1.25em} \ytableausetup{aligntableaux=top} \ytableaushort {{0}{\rd 0}{\rd 0}{\rd 2},{\none}{0}{0}{2},{\none}{\none}{2}} \qquad \ytableaushort {{0}{\rd 0}{\rd 0}{*(gray!25) \rd 1 \rd 2},{\none}{0}{0}{2},{\none}{\none}{2}} \qquad \ytableaushort {{0}{\rd 0}{\rd 0}{\rd 1},{\none}{0}{0}{*(gray!25) \rd 2 2},{\none}{\none}{2}} \qquad \ytableaushort {{0}{\rd 0}{\rd 0}{\rd 1}{*(gray!70) 3},{\none}{0}{0}{\rd 2},{\none}{\none}{2}} \end{center} \medskip Again, the repeated insertion process is not finished, since the entry in position $(1,6)$ is $> 2$. We remove the cell, and insert $x_4$ into the remaining tableau. The insertion process is now as follows: \medskip \begin{center} \ytableausetup{boxsize=1.25em} \ytableausetup{aligntableaux=top} \ytableaushort {{0}{\rd 0}{\rd 0}{\rd 1},{\none}{0}{0}{\rd 2},{\none}{\none}{2}} \qquad \ytableaushort {{0}{\rd 0}{*(gray!25) \rd 0 \rd 0}{\rd 1},{\none}{0}{0}{\rd 2},{\none}{\none}{2}} \qquad \ytableaushort {{0}{\rd 0}{\rd 0}{\rd 1},{\none}{0}{*(gray!25) \rd 0 0}{\rd 2},{\none}{\none}{2}} \qquad \ytableaushort {{0}{\rd 0}{\rd 0}{*(gray!25) 1 \rd 1},{\none}{0}{\rd 0}{\rd 2},{\none}{\none}{2}} \qquad \ytableaushort {{0}{\rd 0}{\rd 0}{1},{\none}{0}{\rd 0}{*(gray!25) \rd 1 \rd 2},{\none}{\none}{2}} \qquad \ytableaushort {{0}{\rd 0}{\rd 0}{1},{\none}{0}{\rd 0}{\rd 1},{\none}{\none}{2}{*(gray!70) \rd 2}} \end{center} \medskip The last shifted bicolored tableau is in $\p B^D(431,865321)$, so we are done with the repeated insertion process. \medskip If we perform the repeated insertion process on all 56,832 pairs $(S,k)$ in $$\p B^D(431,865321) \times \p W^D(431,865321),$$ the process needs just one insertion in 42,672 cases, two insertions in 11,087 cases, three insertions in 2,182 cases, four insertions in 741 cases, five insertions in 88 cases, and six insertions in 62 cases. \medskip We can easily adapt the second example for type B to show that repeated insertion process in type D may also need many steps. Note also that the complexity does not come from the fact that the diagram is shifted: there are also examples with many steps for the repeated insertion process in \cite{kon:naruse}. \subsection*{Properties of the bump, the insertion process, and the repeated insertion process} Say that we have a tableau $S$, $(i,j)$, $d$ and $k$, and that $d = \: \rightarrow$. The bump algorithm in type B says that we should take $j' = j + 1$, and find the largest possible $i'$ so that we can write $k - i'$ in position $(i',j')$ while keeping the column $j'$ weakly increasing. Since the sequence $(S_{i'j'})_{i' \geq 0}$ is weakly increasing, $(S_{i'j'} + i')_{i' \geq 0}$ is strictly increasing. There are two options. One is that $k = S_{i''j'} + i''$ for some $i''$. In that case, we have just one possible choice for $i'$, and that is $i''$. The entry $S_{i'j'}$ remains unchanged, so the new tableau is still a valid shifted bicolored tableau. Also, we cannot have $i' > i$, since then $k - i' \geq S_{ij'}$, and that contradicts the assumption that we can write $k - i > k - i'$ in position $(i,j)$ in $S$. If, on the other hand, we have $S_{i''-1,j'} + i'' - 1 < k < S_{i''j'} + i''$ for some $i''$, then $S_{i''-1,j'} < k - (i''-1) \leq S_{i''j'}$ and $S_{i''-1,j'} \leq k - i'' < S_{i''j'}$. So we can either write $k - (i''-1)$ in position $(i''-1,j')$ or $k - i''$ in position $(i'',j')$. We select $i' = i''$. Now the entry in position $(i',j')$ becomes smaller, and we have to check that $S_{i',j'-1} = S_{i'j} \leq k - i'$. Again, we cannot have $i' > i$, so the fact that $k-i$ can be written in position $(i,j)$ means that $k - i' \geq k - i \geq S_{i'j}$. \medskip The analysis when $d = \: \downarrow$ is very similar and we leave it as an exercise for the reader. In summary, the process either moves one column to the right and weakly up, or one row down and weakly to the left, and numbers in the tableau can only get smaller. \medskip In order to prove that the insertion process terminates, it is enough to prove that a certain (integer) quantity with an upper bound increases at each step. We claim that such a quantity is $$\left\{ \begin{array}{ccl} k + j & \colon & d = \: \rightarrow \\ k + i & \colon & d = \: \downarrow \end{array}\right.$$ When we go to the right, $j' = j + 1$, and when we go down, $i' = i + 1$. In each case we bump a weakly larger number, so $k' \geq k$. It is also clear that this quantity is bounded by $\ell(\mu) + \mu_1 + \max T_{ij}$. \medskip The inverse of the bump is easy to construct, and repeated application of it gives the inverse of the insertion process. The fact that the bump (and consequently $\psi_\mu$) is weight preserving is obvious. We omit the details for type D bump and insertion. \medskip The fact that the repeated insertion process terminates (and is a weight-preserving bijection) can be proved in a very similar way as the proof of \cite[Theorem 11]{kon:naruse} using sieve equivalence. More precisely, in order for that lemma to apply and complete the proof of Theorem \ref{thm1}, we need the following result. \begin{lemma} Let $\psi_\mu(T,k)$ denote the result of type B insertion of variable index $k$ into the tableau $T \in \p B^B(\mu)$. If $S \in \p B^B(\nu)$ for $\mu \lessdot \nu$, $T$ is the tableau of shape $\mu$ that we obtain if we remove the unique cell of $S$ outside of $\mu$, and $k$ is the variable index of the removed entry, then write $\varphi_\mu(S) = (T,k)$. Denote the (finite) set $\p B^B(\mu,\lambda) \times \p W(\lambda)$ by $A$, its image under $\psi_\mu$ by $B$, its subset $\p B^B(\mu,\lambda) \times \p W^B(\mu,\lambda)$ by $X$, and the set $\bigcup_\nu \p B^B(\nu,\lambda)$ by $Y$. Then the restriction of $\varphi_\mu$ to $B \setminus Y$ is an injection with image $A \setminus X$. \end{lemma} The proof is analogous to the proof of \cite[Lemma 12]{kon:naruse}, and proves that the repeated insertion process is a weight-preserving bijection that completes the proof of the main theorem for type B. Again, we omit the details for type D. \subsection*{Complexity of the algorithm} Let $m \geq 1$. Let $S^{m,0}$ be the shifted bicolored tableau of type B and shape $\mu^{(m)} = (m+2,m,m-1,m-2,\ldots,1)$ with black $0$'s everywhere except in columns $3,\ldots,m+1$ of row $2$, where it has red $0$'s, let $k_0 = 2$ and $\lambda^{(m)} = (m+2,m+1,m-1,m-2,\ldots,1)$. Our second example in Section \ref{proofs} is the repeated insertion of $(S^{4,0},k_0)$. The following is defined for $m \geq 1$, $n \geq 0$, and is clearly a generalization of the definition of $S^{m,0}$: $$S^{m,n}_{ij} = \left\{ \begin{array}{ccl} 0 & \colon & i = j \\ {\rd 0} & \colon & j > i = 1, \, n \mbox{ odd and } (n + 1)/2^{j-2} \in \{2,3,5,7,9,11,\ldots\} \\ 0 & \colon & j > i = 1, \, n \mbox{ even or } (n + 1)/2^{j-2} \notin \{2,3,5,7,9,11,\ldots\} \\ 0 &\colon & j > i = 2, \, j = 1 + \mbox{number of digits in binary expansion of } n \\ {\rd 0} &\colon & j > i = 2, \, j \neq 1 + \mbox{number of digits in binary expansion of } n \\ {\rd 0} & \colon & j > i \geq 3, \, (\lfloor \frac{n+2}{2^{i-2}} \rfloor - 1)/2^{j-i} \mbox{ odd positive integer}\\ 0 & \colon & j > i \geq 3, \, (\lfloor \frac{n+2}{2^{i-2}} \rfloor - 1)/2^{j-i} \mbox{ not odd positive integer}\end{array} \right.$$ $$k_n = \left\{ \begin{array}{ccl} 1 & \colon & n \mbox{ odd} \\ j+1 & \colon & n = 2^j - 2 \\ j + 2 & \colon & n \mbox{ even, }(n+2)/2^j \mbox{ odd integer} \end{array} \right.$$ \begin{prop} For $0 \leq n < 2^m$, the insertion process of type B with input $(S^{m,n},k_n)$ ouputs the tableau obtained from $S^{m,n+1}$ by adding a black $k_{n+1} - 1 = 0$ in position $(1,m+3)$ if $n$ is even, and a black $k_{n+1} - 2$ in position $(2,m+2)$ if $n$ is odd, except when $n = 2^m - 1$, where the output has a black $0$ in position $(2,m+2)$. Therefore the repeated insertion process for $(S^{m,0},k_0)$ stops after precisely $2^m$ steps. \end{prop} The proof is easy for $n$ even, and a tedious book-keeping exercise for $n$ odd. We leave it is an exercise for the reader. It is clear that the proposition proves Theorem \ref{complexity}. \section*{Acknowledgments} The author would like to thank Alejandro Morales, Igor Pak and Greta Panova for telling him about the problem and the interesting discussions that followed. Many thanks to Darij Grinberg for suggesting the appropriate weighted generalization and other useful comments, and to Sara Billey for fruitful talks about the subject. \bibliographystyle{siam} \def$'${$'$}

396051107be31e5f04bb9aaedaffaafce9a9d2bf

\section*{APPENDIX} \section{EXPERIMENTAL RESULTS} \label{sec:EXPERIMENTAL_RESULTS} In this section, we present experiments obtained with several implementations of the control architecture shown in Fig.~\ref{fig:controller_architecture}. We use the iCub humanoid robot \cite{Metta2010} to carry out the experimental activities. Let us recall that iCub is $\SI{104}{\centi \meter}$ tall, with a foot length and width of $\SI{19}{\centi \meter}$ and $\SI{9}{\centi \meter}$, respectively. \begin{table}[b] \vskip-0.5cm \centering \caption{Maximum forward straight walking velocities achieved using different implementations of the control architecture. } \begin{tabular}{cc|c} \begin{tabular}{@{}c@{}}Simplified Model \\ Control\end{tabular} & \begin{tabular}{@{}c@{}}Whole-Body \\ QP Control\end{tabular} & \begin{tabular}{@{}c@{}}Max Straight \\ Velocity (m/s)\end{tabular}\\ \hline Predictive & Velocity & 0.19\\ Predictive & Position & 0.20\\ Instantaneous & Velocity & 0.22\\ Instantaneous & Position & 0.41 \end{tabular} \label{tab:max_velocity} \end{table} The control architecture runs on the iCub head's computer, namely a 4-th generation Intel \textsuperscript{\tiny\textregistered} Core i7 @ $\SI{1.7}{\giga \hertz}$. In any of its implementations, the architecture takes (in average) less than $\SI{3}{\milli \second}$ for evaluating its outputs. OSQP \cite{osqp} library was used to solve the optimization problems. Tab.~\ref{tab:max_velocity} summarizes the maximum velocities achieved using the different implementations of the control architecture. In particular, the labels \emph{instantaneous} and \emph{predictive} mean that the associated layer generates its outputs considering inputs and references either at the single time $t$ or for a time window, respectively. The labels, \emph{velocity} and \emph{position} control, instead, mean that the layer outputs are either desired joint velocities or position, respectively -- see Sec.~\ref{subsubsec-pos-vel-control}. Let us remark that all the implemented control architectures exploit the controller presented in section~\ref{ZMP-CoM-Controller} to attempt the stabilization of the desired center-of-pressure and desired center-of-mass position and velocity. The performances of this controller much depend on the gains $K_{zmp}$ and $K_{com}$. In particular, we observed that the gains achieving good tracking during standing and walking were not the same. For this reason, we implemented a gain-scheduling technique depending on the robot is walking or standing. The transition between the two sets of gains is smoothed with a minimum jerk trajectory \cite{Pattacini2010a}. Comparing different control architectures, however, is a far cry from being an easy task. To do so, we decided to perform two main experiments only, which are used as benchmarks for all control architecture implementations. These two experiments represent the maximum robot velocity that has been achieved with all architectures, and the maximum velocity achieved with a specific architecture only -- see Tab.~\ref{tab:max_velocity}. Namely, \begin{itemize} \item[\textbf{- Experiment 1}] a forward robot speed of $\SI{0.19}{\meter \per \second}$; \item[\textbf{- Experiment 2}] a forward robot speed of $\SI{0.41}{\meter \per \second}$. \end{itemize} \begin{figure*}[t] \centering \begin{myframe}{Instantaneous + Position Control} \begin{subfigure}[b]{0.329\textwidth} \centering \includegraphics[width=\textwidth]{figures/inst_pos-min_vel-dcm.pdf} \caption{DCM} \label{fig:inst_pos-min_vel-dcm} \end{subfigure} \hfill \begin{subfigure}[b]{0.329\textwidth} \centering \includegraphics[width=\textwidth]{figures/inst_pos-min_vel-com.pdf} \caption{CoM} \label{fig:inst_pos-min_vel-com} \end{subfigure} \hfill \begin{subfigure}[b]{0.329\textwidth} \centering \includegraphics[width=\textwidth]{figures/inst_pos-min_vel-zmp.pdf} \caption{ZMP} \label{fig:inst_pos-min_vel-zmp} \end{subfigure} \end{myframe} \vskip-0.5cm \hfill \begin{myframe}{Predictive + Position Control} \begin{subfigure}[b]{0.329\textwidth} \centering \includegraphics[width=\textwidth]{figures/mpc_pos-min_vel-dcm.pdf} \caption{DCM} \label{fig:mpc_pos-min_vel-dcm} \end{subfigure} \hfill \begin{subfigure}[b]{0.329\textwidth} \centering \includegraphics[width=\textwidth]{figures/mpc_pos-min_vel-com.pdf} \caption{CoM} \label{fig:mpc_pos-min_vel-com} \end{subfigure} \begin{subfigure}[b]{0.329\textwidth} \centering \includegraphics[width=\textwidth]{figures/mpc_pos-min_vel-zmp.pdf} \caption{ZMP} \label{fig:mpc_pos-min_vel-zmp} \end{subfigure} \end{myframe} \caption{ Tracking of the DCM (a), CoM (b) and ZMP (c) using the instantaneous controller with the whole-body controller as position control. Tracking of the DCM (d), CoM (e) and ZMP (f) using the MPC and the whole-body controller as position control. Walking velocity: $\SI{0.19}{\meter \per \second}$.} \end{figure*} \begin{figure*}[t] \vspace*{-0.1cm} \begin{myframe}{Instantaneous + Position Control} \begin{subfigure}[b]{0.329\textwidth} \centering \includegraphics[width=\textwidth]{figures/inst_pos-max_vel-dcm.pdf} \caption{DCM} \label{fig:inst_pos-max_vel-dcm} \end{subfigure} \hfill \begin{subfigure}[b]{0.329\textwidth} \centering \includegraphics[width=\textwidth]{figures/inst_pos-max_vel-com.pdf} \caption{CoM} \label{fig:inst_pos-max_vel-com} \end{subfigure} \hfill \begin{subfigure}[b]{0.329\textwidth} \centering \includegraphics[width=\textwidth]{figures/inst_pos-max_vel-zmp.pdf} \caption{ZMP} \label{fig:inst_pos-max_vel-zmp} \end{subfigure} \end{myframe} \vskip-0.5cm \hfill \begin{myframe}{Predictive + Position Control} \begin{subfigure}[b]{0.329\textwidth} \centering \includegraphics[width=\textwidth]{figures/mpc_pos-max_vel-dcm.pdf} \caption{DCM} \label{fig:mpc_pos-max_vel-dcm} \end{subfigure} \hfill \begin{subfigure}[b]{0.329\textwidth} \centering \includegraphics[width=\textwidth]{figures/mpc_pos-max_vel-com.pdf} \caption{CoM} \label{fig:mpc_pos-max_vel-com} \end{subfigure} \hfill \begin{subfigure}[b]{0.329\textwidth} \centering \includegraphics[width=\textwidth]{figures/mpc_pos-max_vel-zmp.pdf} \caption{ZMP} \label{fig:mpc_pos-max_vel-zmp} \end{subfigure} \end{myframe} \caption{Tracking of the DCM (a), CoM (b) and ZMP (c) with the instantaneous and whole-body QP control as position. Tracking of the DCM (d), CoM (e) and ZMP (f) with the predictive and whole-body QP control as position control. At $t\approx \SI{2}{\second}$, the robot falls down. Walking velocity: $\SI{0.41}{\meter \per \second}$.} \vskip-0.5cm \end{figure*} \begin{figure*}[t] \centering \begin{subfigure}[b]{\columnwidth} \begin{myframe}{Instantaneous + Position Control} \begin{subfigure}[t]{0.49\columnwidth} \centering \includegraphics[width=\textwidth]{figures/inst_pos-min_vel-lf.pdf} \caption{ } \label{fig:inst_pos-min_vel-lf} \end{subfigure} \begin{subfigure}[t]{0.49\columnwidth} \centering \includegraphics[width=\textwidth]{figures/inst_pos-max_vel-lf.pdf} \caption{ } \label{fig:inst_pos-max_vel-lf} \end{subfigure} \end{myframe} \end{subfigure} \begin{subfigure}[b]{\columnwidth} \begin{myframe}{Instantaneous + Velocity Control} \begin{subfigure}[t]{0.49\columnwidth} \centering \includegraphics[width=\textwidth]{figures/inst_vel-min_vel-lf.pdf} \caption{ } \label{fig:inst_vel-min_vel-lf} \end{subfigure} \begin{subfigure}[t]{0.49\columnwidth} \centering \includegraphics[width=\textwidth]{figures/inst_vel-max_vel-lf.pdf} \caption{ } \label{fig:inst_vel-max_vel-lf} \end{subfigure} \end{myframe} \end{subfigure} \caption{Tracking of the left foot position using the instantaneous simplified model control. Whole-body QP control as position control (a) and velocity control (c), walking velocity: $\SI{0.19}{\meter \per \second}$. Whole-body QP control as position control (b) and velocity control (d), walking velocity: $\SI{0.41}{\meter \per \second}$. } \vskip-0.25cm \end{figure*} \subsection{Simplified model control: Predictive versus Instantaneous} \label{sub_sec:Simplified Models Controller} In this section, we compare the control laws~\eqref{eq:reactive_dcm} and~\eqref{MPC_solution}, which both generate a (desired) center-of-pressure that attempt at stabilizing a desired DCM. To simplify the comparison, the controller of the \emph{whole-body QP layer} is kept fixed in this section, and we show and discuss only the results when the robot is position controlled. The time horizon of the predictive control is $\SI{2}{\second}$. \subsubsection{Experiment 1} Figs. \ref{fig:inst_pos-min_vel-dcm} and \ref{fig:mpc_pos-min_vel-dcm} depict the DCM tracking performances obtained with the instantaneous and predictive controllers, respectively. Both controllers seem to show good tracking performances, and the DCM error is kept below $\SI{5}{\centi \meter}$ in both cases. Note that the instantaneous controller induces, however, faster variations of the measured DCM. This contributes to overall higher vibrations of the robot. One of the reasons for this variation is that the instantaneous controller~\eqref{eq:reactive_dcm} injects a (desired) center-of-pressure proportional to the measured DCM $\xi = x + \dot{x}/\omega$, which in turns contains the center-of-mass velocity. To mitigate this, we suggest to filter joint velocities appropriately. In our case, however, joint velocities were not filtered to avoid delays in the measured DCM. Our experience showed that adding a filter on joint velocities is not an easy task, and we did not find the right trade off for obtaining overall performance improvements. Figs.~\ref{fig:inst_pos-min_vel-com} and \ref{fig:mpc_pos-min_vel-com}, depict the CoM tracking performances, which are mainly dependent on the ZMP-CoM controller~\eqref{eq:zmp_controller}. This controller receives desired DCM values from the \emph{simplified model control} layer, which are obtained with either the instantaneous or predictive controllers. In both cases, CoM tracking performances are good, and the CoM error is kept below $\SI{2}{\centi \meter}$. Figs.~\ref{fig:inst_pos-min_vel-zmp} and~\ref{fig:mpc_pos-min_vel-zmp} depict the ZMP tracking performances, which are still mainly dependent on the ZMP-CoM controller~\eqref{eq:zmp_controller}. It is important to observe that the desired ZMP is smoother when the \emph{simplified model control} uses the predictive law~\eqref{MPC_solution} to generate it. This is a tunable property that depends on the associated weight in the cost function of the MPC problem. Although this smoother behaviour does contribute to less robot vibrations, the overall robot performance became less reactive and, consequently, less robust to robot falls. Although the extensive hand-made tuning, we were not able to increase the robot velocity when the \emph{simplified model control} used the predictive law~\eqref{MPC_solution}. \subsubsection{Experiment 2} At a robot desired walking speed of $\SI{0.41}{\meter \per \second}$, there is initially no significant difference between the DCM tracking obtained with instantaneous and predictive control laws -- see Fig.~\ref{fig:mpc_pos-max_vel-dcm} and~\ref{fig:inst_pos-max_vel-dcm} for $t < \SI{1.5}{\second}$. However, fast robot walking velocities require fast variations of the desired CoM and ZMP. In the case of the predictive controller, this fast variation induces a not-very-good performance of the desired ZMP around $t = \SI{1.5}{\second}$ -- see Fig.~\ref{fig:mpc_pos-max_vel-zmp}. Clearly, these bad performances in turn induce a bad tracking of the DCM shown in Fig.~\ref{fig:mpc_pos-max_vel-dcm} at $t\approx \SI{2}{\second}$, and consequently a robot fall. At this point, one is tempted to increase the gain $K_{zmp}$ of the controller~\eqref{eq:zmp_controller}, which shall induce a better tracking of the ZMP. This unfortunately gives raise to higher robot oscillations, which in turn degrades the ZMP-CoM tracking, and still a robot fall. In a nutshell, the \emph{predictive simplified control} is much less robust than the \emph{instantaneous simplified control} with respect to ZMP tracking errors. We suggest to filter the measurements from force sensors to obtain less noisy ZMP measurement, which should allow one to increase the gains for ZMP tracking. In our case, adding filters led to slower system response and, consequently, a robot fall. \subsection{Whole-Body QP Control: Position versus Velocity} In this section, we compare the performances of the three layer architecture when the robot is either position or velocity controlled -- see Sec.~\ref{subsubsec-pos-vel-control} for the meaning of these modes. For comparison purposes, the \emph{simplified model control} is achieved via the instantaneous control~\eqref{eq:reactive_dcm}. Also, to emphasize the comparison, we stress the importance of the tracking of the desired feet positions. \subsubsection{Experiment 1} Figs. \ref{fig:inst_pos-min_vel-lf} and \ref{fig:inst_vel-min_vel-lf} depict the tracking of desired left foot positions when the robot is either position or velocity controlled, respectively. The position controller ensures better tracking performance than the velocity one. One is then tempted to increase the gains of the * quantities~\eqref{feetVelocitiesStar}, which should increase the tracking performances of the velocity control. However, we observed that the noise due to numerical derivative is harmful for the overall performance, and makes the robot shake and then fall. Again, filtering the taken joint measures may be helpful, but it introduced a delay that did not allow us to increase the walking speed. \subsubsection{Experiment 2} The aforementioned foot position tracking problem worsens at higher walking velocity. Fig.~\ref{fig:inst_pos-max_vel-lf} shows that the feet tracking error is lower than $\SI{5}{\centi \meter}$ on the $x$ axis and $\SI{1}{\centi \meter}$ on the $z$ one for position control. Instead, the velocity control in Fig.~\ref{fig:inst_vel-max_vel-lf} keeps the error always lower than $\SI{6}{\centi \meter}$ on the $x$ component and $\SI{3}{\centi \meter}$ on the $z$ direction. \section{INTRODUCTION} Bipedal locomotion of humanoid robots remains an open problem despite decades of research in the subject. The complexity of the robot dynamics, the unpredictability of its surrounding environment, and the low efficiency of the robot actuation system are only a few problems that complexify the achievement of robust robot locomotion. In the large variety of robot controllers for bipedal locomotion, the Divergent-Component-of-Motion (DCM) is an ubiquitous concept used for generating walking patterns. This paper presents and compares different DCM based control architectures for humanoid robot locomotion. \par During the DARPA Robotics Challenge, a common approach for humanoid robot control was that of defining an hierarchical architecture composed of several layers~\cite{feng2015optimization}. Each layer generates references for the layer below by processing inputs from the robot, the environment, and the outputs of the layer before. From top to bottom, these layers are here called: \emph{trajectory optimization}, \emph{simplified model control}, and \emph{whole-body quadratic programming (QP) control}. The \emph{trajectory optimization} layer often generates desired foothold locations by means of optimization techniques. To do so, both kinematic and dynamical robot models can be used~\cite{dai2014whole,herzog2015trajectory}. When solving the optimization problem associated with the \emph{trajectory optimization} layer, computational time may be a concern especially when the robot surrounding environment is not structured. There are cases, however, where simplifying assumptions on the robot environment can be made, thus reducing the associated computational time. For instance, flat terrain allows one the view the robot as a simple unicycle \cite{PascalHandbook,flavigne2010reactive}, which enables quick solutions to the optimization problem for the walking pattern generation~\cite{Dafarra2018}. \par \begin{figure}[t] \centering \begin{subfigure}[b]{0.29\columnwidth} \centering \includegraphics[width=\textwidth]{figures/walk1.png} \end{subfigure} \begin{subfigure}[b]{0.29\columnwidth} \centering \includegraphics[width=\textwidth]{figures/walk2.png} \end{subfigure} \begin{subfigure}[b]{0.29\columnwidth} \centering \includegraphics[width=\textwidth]{figures/walk3.png} \end{subfigure} \caption{iCub walks with the presented controller architecture. \label{fig:icub}} \vskip-0.5cm \end{figure} The \emph{simplified model control} layer is in charge of finding feasible center-of-mass (CoM) trajectories and is often based on simplified dynamical models, such as the Linear Inverted Pendulum Model (LIPM) \cite{Kajita2001} and the Capture Point (CP) \cite{Pratt2006}. These models have become very popular after the introduction of the Zero Moment Point (ZMP) as a stability criterion~\cite{Vukobratovic1969a}. To obtain feasible CoM trajectories, the \emph{simplified model control} layer often combines the LIPM with Model Predictive Control (MPC) techniques, also known as the Receding Horizon Control (RHC) \cite{Kajita2003,diedam2008online}. Another model that is often exploited in the \emph{simplified model control} layer is the Divergent Component of Motion (DCM) \cite{Englsberger2015}. The DCM can be viewed as the extension of the capture point (CP) to the three dimensional case, however always under the assumption of a constant Virtual Repellent Point (VRP) to Enhanced Centroidal Moment Pivot point (eCMP) height difference \cite{Englsberger2015}. Attempts at loosening this latter assumption and extending the DCM to more complex models have also been presented \cite{Hopkins2015}. The \emph{whole-body QP control} layer generates robot positions, velocity or torques depending on the available control modes of the underlying robot. These outputs aim at stabilizing the references generated by the layers before. It uses whole-body kinematic or dynamical models, and very often instantaneous optimization techniques, namely, no MPC methods are here employed. Furthermore, the associated optimisation problem is often framed as an hierarchical stack-of-tasks, with strict or weighted hierarchies \cite{Stephens2010,Nava2016}. This paper presents and compares several DCM based implementations of the above layered control architecture. In particular, the \emph{trajectory optimization} layer is kept fixed with a unicycle based planner that generates desired DCM and foot trajectories. The \emph{simplified model control} layer, instead, implements two controllers for the tracking of the DCM: an instantaneous and an MPC controller. In the same layer, we also present a controller that ensures the tracking of the CoM and the ZMP, which exploits 6-axes Force Torque sensors (F/T). Finally, the \emph{whole-body QP control} ensures the tracking of the desired CoM and feet trajectories. It achieves this by presenting velocity and inverse kinematic based controllers. The several combinations of the control architecture are tested on the iCub humanoid robot \cite{Nataleeaaq1026} . \par The paper is organized as follows. Sec.~\ref{sec:BACKGROUND} introduces notation, the humanoid robot model, and some simplified models used for locomotion. Sec.~\ref{sec:ARCHITECTURE} describes each layer of the control architecture, namely the trajectory optimization, the simplified model control and and the whole-body QP control layer. Sec.~\ref{sec:EXPERIMENTAL_RESULTS} presents the experimental validation of the proposed approach, and shows an explanatory table comparing the different control approaches. Finally, Sec.~\ref{sec:CONCLUSION_AND_FUTURE_WORK} concludes the paper. \section{CONCLUSION AND FUTURE WORK} \label{sec:CONCLUSION_AND_FUTURE_WORK} This paper contributes towards the benchmarking of different implementations of state-of-the-art control architectures for humanoid robots locomotion. In particular, the control architecture is composed of three layers, which all exploit the concept of the Divergent Component of Motion. The three layers are here called: trajectory optimization, simplified model control, the whole-body QP control. A key feature of this paper is that we compare walking results obtained with predictive and instantaneous controllers for the simplified model layer. Also, we compare position and velocity robot control modes for the whole-body QP control layer. We show that instantaneous controllers combined with robot position control allowed us to achieve a desired walking speed of $\SI{0.41}{\meter \per \second}$, which is the highest walking velocity ever achieved for the iCub humanoid robot. As future work, we plan to extend the proposed benchmarking to torque-control algorithms (e.g.~\cite{7803266}), and to propose architecture implementations allowing the robot walking on inclined planes. Another interesting future work is the implementation of a dynamic footstep planner. Indeed, currently, the footsteps are generated independently from the state of the robot. A dynamic planner can plan the footprints according to the external disturbance acting on the humanoid, i.e. an unknown contact force acting on the robot. \section{BACKGROUND} \label{sec:BACKGROUND} \subsection{Notation} \begin{itemize} \item $I_n$ and $0_n$ are used to denote respectively the $n\times n$ identity and zero matrices; \item $\mathcal{I}$ denotes an inertial frame; \item $e_i \in \mathbb{R}^n$ is the canonical vector, consisting of all zeros but the $i$-th component that is equal to one; \item given two frames, $\mathcal{A}$ and $\mathcal{B}$, $\prescript{\mathcal{A}}{}{R}_B \in SO(3)$ represents the rotation matrix between the frames, i.e. given two vectors $\prescript{\mathcal{A}}{}{p}, \prescript{\mathcal{B}}{}{p} \in \mathbb{R}^3$ respectively expressed in $\mathcal{A}$ and $\mathcal{B}$, the rotation matrix $\prescript{\mathcal{A}}{}{R}_\mathcal{B}$ is such that $\prescript{\mathcal{A}}{}{p} = \prescript{\mathcal{A}}{}{R}_\mathcal{B} \prescript{\mathcal{B}}{}{p}$; \item given a skew-symmetric matrix $W\in \mathfrak{so}(3)$ the \emph{vee operator} is $.^\vee : \mathfrak{so}(3) \to \mathbb{R}^3$; \item $\prescript{\mathcal{A}}{}{\omega}_\mathcal{B} \in \mathbb{R}^3$ denotes the angular velocity between the frame $\mathcal{B}$ and the frame $\mathcal{A}$ expressed in the frame $\mathcal{A}$; \item given a square matrix $A \in \mathbb{R}^{3 \times 3}$ the \emph{skew} operator is $\text{sk} :\mathbb{R}^{3 \times 3} \to \mathfrak{so}(3)$, $\text{sk}(A) := (A - A^\top)/2$; \item the subscripts $\mathcal{T}$, $\mathcal{LF}$, $\mathcal{RF}$ and $\mathcal{C}$ indicates the frames attached to the torso, left foot, right foot and CoM. \end{itemize} \subsection{Humanoid robot models} A humanoid robot is an example of floating base multibody systems composed of $n + 1$ links connected by $n$ joints with one degree of freedom. Since none of the links of the robot has an a priori constant position and orientation with respect to a global frame $\mathcal{I}$, its configuration can be determined by the position and the orientation of the base frame $\mathcal{B}$ (e.g. the stance foot during the locomotion) and the joints values. Thus the configuration space is defined by $\mathbb{Q} = SO(3) \times \mathbb{R}^3 \times \mathbb{R}^n$. An element of $\mathbb{Q}$ is then a triplet $q = (\prescript{\mathcal{I}}{}{p}_\mathcal{B}, \prescript{\mathcal{I}}{}{R}_\mathcal{B}, s)$, where $(\prescript{\mathcal{I}}{}{p}_\mathcal{B}, \prescript{\mathcal{I}}{}{R}_\mathcal{B}) \in (SO(3) \times \mathbb{R}^3)$ is used to represent the position and the orientation of the base frame, $\mathcal{B}$, expressed with respect to the inertial frame $\mathcal{I}$; and $s \in \mathbb{R}^n$ represents the joint angles. Furthermore, according to the group theory, $\mathbb{Q}$ is a Lie group. Indeed given two elements $a = (p_a, R_a, s_a) \in \mathbb{Q}$ and $b = (p_b, R_b, s_b)\in \mathbb{Q}$ the group multiplication $a \cdot b$ is defined as $a \cdot b = (p_a + p_b, R_a R_b, s_a + s_b) \in \mathbb{Q}$, while the inverse of $a$ is given by: $a^{-1} = (-p_a, R_a^\top, -s_a) \in \mathbb{Q}$. Since $\mathbb{Q}$ is a Lie group, the velocity of the multibody system is represented by the Lie algebra $\mathbb{V}$. An element of the Lie algebra $\mathbb{V}$ is a triplet $ \nu = (\prescript{\mathcal{I}}{}{\dot{p}}_\mathcal{B}, \prescript{\mathcal{I}}{}{\omega}_\mathcal{B}, \dot{s})$ where $\prescript{\mathcal{I}}{}{\omega}_{\mathcal{B}}$ is the angular velocity of the base with respect the inertial frame whose coordinates are written in the inertial frame, i.e. $\prescript{\mathcal{I}}{}{\dot{R}}_\mathcal{B} = \prescript{\mathcal{I}}{}{\hat{\omega}}_\mathcal{B} \prescript{\mathcal{I}}{}{R}_\mathcal{B}$. \par Given a link of the floating base system its position and orientation with respect to the inertial frame is uniquely identified by an homogeneous transformation, $\prescript{\mathcal{I}}{}{H}_\mathcal{A} \in SE(3)$, between the inertial frame, $\mathcal{I}$, and the frame attached to that link, $\mathcal{A}$. $\prescript{\mathcal{I}}{}{H}_\mathcal{A}$ is the classical homogeneous transformation containing the rotation matrix, $\prescript{\mathcal{I}}{}{R}_\mathcal{A}$, and the vector connecting the origin of the inertial frame to origin of the frame $\mathcal{A}$ expressed in the inertial frame $\prescript{\mathcal{I}}{}{p}_\mathcal{A}$. Given a link of the floating base system the velocity of a rigid body can be represented by $\prescript{\mathcal{I}}{}{v}_\mathcal{A} = \begin{bmatrix} \prescript{\mathcal{I}}{}{\dot{p}}_\mathcal{A}^\top & \prescript{\mathcal{I}}{}{\omega}_{\mathcal{A}}^\top \end{bmatrix}^\top$. \par The Jacobian $J_{\mathcal{A}}(q)$ is the map between the robot velocity and the linear and angular velocities of the frame $\mathcal{A}$, i.e. $\prescript{\mathcal{I}}{}{v}_\mathcal{A} = J_{\mathcal{A}} \nu$ where $J_{\mathcal{A}}$ can be split into two parts, one related to the velocity of the base and the other one multiplying the joints velocity, $J_\mathcal{A}(q) = \begin{bmatrix} J^b_\mathcal{A} & J^j_\mathcal{A} \end{bmatrix}$. \begin{figure*}[!t] \centering \includegraphics[scale=0.89]{figures/entire_controller_architecture.pdf} \caption{The control architecture is composed of three layers: the \emph{trajectory optimization}, the \emph{simplified model control}, and the \emph{whole-body QP control}. \label{fig:controller_architecture}} \vskip-0.5cm \end{figure*} \subsection{Simplified models} \label{sub_sec:simplified_models} For the purpose of this work, the motion of the humanoid robot is approximated by means of the well known \emph{Linear inverted pendulum model} (LIPM) \cite{Kajita2001}. By using the LIPM, in case of walking on a flat surface the CoM belongs to an horizontal plane with a constant height $z_0$. The simplified CoM dynamics is given by \cite{Kajita2001}: \begin{equation} \label{eq:3d-lipm} \ddot{x} = \omega^2(x - r^{zmp}), \end{equation} where $x \in \mathbb{R}^2$ is the vector containing the projection of the CoM on the walking surface, $r^{zmp}\in \mathbb{R}^2$ is the position of the zero moment point (ZMP) and $\omega$ is the inverse of the pendulum time constant, i.e. $\omega = \sqrt{g/z_0}$ where $g$ is the gravity constant. Analogously, one can define the divergent component of motion (DCM) as $\xi = x + \dot{x}/\omega$ \cite{Englsberger2015}. Clearly, the DCM time derivative is given by: \begin{equation} \label{eq:dcm_dynamics} \dot{\xi} = \omega(\xi - r^{zmp}). \end{equation} Using the DCM as state variable, the LIPM dynamics \eqref{eq:3d-lipm} can be decomposed into two parts: \begin{equation} \label{eq:simplified-model} \begin{bmatrix} \dot{x}\\ \dot{\xi} \end{bmatrix} = \begin{bmatrix} -\omega I_2 & \omega I_2\\ 0_2 & \omega I_2 \end{bmatrix} \begin{bmatrix} x\\ \xi \end{bmatrix} + \begin{bmatrix} 0_2\\ \omega I_2 \end{bmatrix} r^{zmp}. \end{equation} Performing the state space decomposition we can easily show that the CoM tends the DCM, while the DCM dynamics has a strictly positive real part eigenvalue. \section{ARCHITECTURE} \label{sec:ARCHITECTURE} This section summarizes the components of the control architecture presented in Fig. \ref{fig:controller_architecture}. In particular, the control architecture is composed of three main layers. The first layer is represented by the \emph{trajectory generator}, whose main purpose is to generate desired footstep positions and orientation and also the desired DCM trajectory. The second layer employs \emph{simplified robot models} to track the desired DCM, CoM and ZMP trajectories. Finally, the third control layer is given by the \emph{whole-body QP} block. It has the main purpose of ensuring the tracking of the desired feet positions and orientations and also the CoM trajectories. This, differently from the previous control layer, exploits whole-body robot kinematic models and feedback. \subsection{Trajectory optimization layer} \label{sub_sec:Trajectory_optimization} The main purpose of this layer is to evaluate the desired footstep positions and the desired feet and DCM trajectories. \par To plan the desired footstep positions, the humanoid robot is approximated as an unicycle \cite{Dafarra2018}. The feet are represented by the unicycle wheels, and the footsteps can be obtained through sampling of the unicycle trajectories. In particular, given the finite set of unicycle positions, we can sample particular feet positions. Each position is associated with a time instant $k$. This time instant is considered as the foot impact time $t_{imp}$. Thus, the impact time can be used as a decision variable, which allows us to select the feet position and avoid fast/slow step duration and long/short step length. Furthermore, the footsteps are planned to avoid excessive rotation between the two feet positions that could be unfeasible because of joint limits. Once the footsteps are planned, the desired feet trajectory is obtained by cubic spline interpolation. \par The DCM is chosen so as to satisfy the following time evolution: \begin{equation} \label{eq:dcm_solution_ios} \xi = r^{zmp} + e ^{\omega t} (\xi_0 - r^{zmp}), \end{equation} where $\xi_0$ is the initial position of the DCM, $r^{zmp}$ is the position of the ZMP (placed on the center of the stance foot) and $t$ has to belong to the step domain $t \in [0, \; t^{step}_i]$ where $t^{step}_i$ is the duration of the $i$-th step. Assuming that the final position of the DCM coincides with the ZMP at last step (i.e. $\xi_{N-1}^{eos} = r_{N}^{zmp}$), \eqref{eq:dcm_solution_ios} can be used to find the desired DCM position at the end of each step \cite{Englsberger2014} \begin{equation} \label{eq:dcm_eos_evaluation} \begin{cases} \xi_{N-1}^{eos} = r_{N}^{zmp} \\ \xi_{i-1}^{eos} = \xi_{i}^{ios} = r_{i}^{zmp} + e^{-\omega t^{step}_i} (\xi_i^{eos} - r_i^{zmp}), \end{cases} \end{equation} where $\xi_{i}^{ios}$ and $\xi_{i}^{eos}$ are respectively the desired DCM initial and final positions for the $i$-th step. \par By substitution of \eqref{eq:dcm_eos_evaluation} into \eqref{eq:dcm_solution_ios}, one obtains the reference DCM trajectory: \begin{equation} \label{eq:dcm_solution_eos} \xi_i(t) = r^{zmp}_i + e ^{\omega (t - t^{step}_i)} (\xi^{eos}_i - r^{zmp}_i). \end{equation} The DCM velocity can be easily evaluated via differentiation of \eqref{eq:dcm_solution_eos}: $\dot{\xi}_i(t) =\omega e ^{\omega (t - t^{step}_i)} (\xi^{eos}_i - r^{zmp}_i)$. In light of the above, the DCM trajectory along the walking pattern can be computed recursively. The presented planning method is very powerful and it allows us to generate the desired DCM trajectory in real-time. Nevertheless, it has the main limitation of taking into account single support phases only. Indeed, by considering instantaneous transitions between two consecutive single support phases, the ZMP reference is discontinuous. This leads to the discontinuity of the external forces and also of the desired joint torques. The development of a DCM trajectory generator that handles non-instantaneous transitions between two single support phases becomes pivotal~\cite{Englsberger2014}. \par Since the ZMP is related to the DCM position and velocity, namely $r^{zmp} = \xi - \dot{\xi}/\omega$, a DCM reference trajectory with continuous derivative ($C^1$ continuity) guarantees a continuous ZMP trajectory. This can be easily ensured using a third order polynomial to smooth the edges of the DCM reference trajectory evaluated using the exponential technique \cite{Englsberger2014}: \[ \xi^{DS} = a_3 t^3 + a_2 t^2 + a_1 t + a_0, \] where the parameters $a_i$ for $i=0:3$ have to be chosen in order to satisfy the velocity and position boundary conditions, namely: \[ \begin{cases} \xi^{DS_i}_i = r^{zmp}_{i-1} + e ^{\omega (t^{DS_i}_{i-1} - t^{step}_{i-1})} (\xi^{eos}_{i-1} - r^{zmp}_{i-1}) \\ \dot{\xi}^{DS_i}_i = \omega e ^{\omega (t^{DS_i}_{i-1} - t^{step}_{i-1})} (\xi^{eos}_{i-1} - r^{zmp}_{i-1}) \\ \xi^{DS_e}_i = r^{zmp}_i + e ^{\omega (t^{DS_e}_i - t^{step}_i)} (\xi^{eos}_i - r^{zmp}_i) \\ \dot{\xi}^{DS_e}_i = \omega e ^{\omega (t^{DS_e}_i - t^{step}_i)} (\xi^{eos}_i - r^{zmp}_i). \end{cases} \] Here $\xi^{DS_i}$ and $\dot{\xi}^{DS_i}$ are the desired DCM position and velocity at the beginning of the double support phase, $\xi^{DS_e}$ and $\dot{\xi}^{DS_e}$ are the desired DCM position and velocity at the end of the double support phase, $t^{DS_i}$ and $t^{DS_e}$ are the initial and final instant of the double support phase. \par In Fig. \ref{fig:dcm_ds} the whole DCM trajectory is shown. During the single support phase (orange segment) the exponential interpolation technique \eqref{eq:dcm_solution_eos} is used; while during the double support phase (light blue curves) the trajectory is obtained using the polynomial technique described above. \begin{figure}[tpb] \centering \includegraphics[angle =-90, scale=0.5]{dcm_ds.pdf} \caption{Planning of 2D-DCM on a flat terrain. The single support trajectories are represented by orange segments. The double support trajectories are represented by light blue curves.} \label{fig:dcm_ds} \vskip-0.5cm \end{figure} \subsection{Simplified model control layer} As discussed in Sec. \ref{sub_sec:simplified_models}, the simplified model~\eqref{eq:simplified-model} shows that the CoM asymptotically converges to a constant DCM, while the DCM has an unstable first order dynamics. Then, we present in this section control laws that aim at stabilizing the DCM dynamics only assuming the $r^{zmp}$ as control input. This stabilization problem has been tackled by designing and comparing \emph{instantaneous} and \emph{ predictive} (MPC) control laws. Differently from the MPC, the instantaneous control law does not solve any optimization problem and it uses only the current desired and actual position of the DCM to evaluate its output. \par \subsubsection{DCM instantaneous control} \label{instantaneous-control} Differently from \cite{Englsberger2015}, we propose an instantaneous control law that does not perform a cancellation of the unstable system dynamics, which, in practice, is never perfect and may lead to system instabilities. More precisely, we choose \begin{equation} \label{eq:reactive_dcm} \begin{split} r^{zmp}_{ref} {=} \xi_{ref} {-} \frac{\dot{\xi}_{ref}}{\omega} {+} K^{\xi}_{p} (\xi {-} \xi_{ref}) + K^{\xi}_{i} \int{(\xi {-} \xi_{ref}) \diff t} , \end{split} \end{equation} where $K^{\xi}_{p}> I_2$ and $K^{\xi}_{i} > 0_2$ \par Applying the control input \eqref{eq:reactive_dcm} to system \eqref{eq:dcm_dynamics} leads to the following closed loop dynamics: \begin{equation} \label{eq:reactive_cl} \begin{split} (\dot{\xi} {-} \dot{\xi}_{ref}) {=} \omega(I_2 {-} K^{\xi}_{p}) (\xi {-} \xi_{ref}) {-}\omega K^{\xi}_{i} \int{(\xi {-} \xi_{ref}) \diff t}. \end{split} \end{equation} It is easy to show the DCM error and its integral converge asymptotically to zero. The above law~\eqref{eq:reactive_dcm} is very simple to implement and guarantees the tracking of the desired DCM. The main limitation is that may not ensure the feasibility of the gait since the position of the ZMP may exit the support polygon. Furthermore, the above instantaneous control law does not take into account the future planned desired DCM trajectory. \subsubsection{DCM predictive control} \label{predictive-control} In order to overcome these issues, a model predictive controller can be designed~\cite{Krause2012}. \par In the MPC framework, the DCM dynamics \eqref{eq:dcm_dynamics} is used as a prediction model and it is discretized supposing piecewise constant ZMP trajectories: \begin{equation} \label{eq:dcm_discrete} \begin{split} \xi_{k+1} &{=} F \xi_k + G r_k^{zmp} \\ &{=} \begin{bmatrix} e^{\omega T} & 0 \\ 0 & e^{\omega T} \end{bmatrix} \xi_k {+} \begin{bmatrix} 1 - e^{\omega T} & 0\\ 0 & 1 - e^{\omega T} \end{bmatrix} r_k^{zmp}. \end{split} \end{equation} In order to ensure that the stance foot does not rotate around one of its edges, the desired ZMP must not exit the support polygon \cite{Vukobratov2004}. This is verified by means of a set of linear inequality constraints: \begin{equation} \label{eq:inequality_constraints} A_{c_k} r^{zmp}_k \le b_{c_k}, \end{equation} where $A_{c_k}$ and $b_{c_k}$ are time variant and their dimension depends on the type of support. \par The cost function shall then ensure the tracking of the desired trajectory. In particular, we choose the following cost: \begin{IEEEeqnarray}{LL} \IEEEyesnumber \phantomsection \label{eq:mpc_const_function} J_k = & \sum\limits_{j=k}^{N + k-1} ({\xi}_j-{\xi}_j^{ref})^\top Q ({\xi}_j-{\xi}_j^{ref}) + \IEEEyessubnumber \label{eq:mpc_cost_state}\\ & ({r}^{zmp}_j - {r}^{zmp}_{j-1})^\top R ({r}^{zmp}_j - {r}^{zmp}_{j-1}) + \IEEEyessubnumber \label{eq:mpc_cost_input}\\ &({\xi}_{k+N}-{\xi}_{k+N}^{ref})^\top Q_N ({\xi}_{k+N}-{\xi}_{k+N}^{ref}) \IEEEyessubnumber \label{eq:mpc_cost_FinalTerm} \end{IEEEeqnarray} where $Q$, $Q_N$ and $R$ are positive $2 \times 2$ symmetric matrices and $N$ is length of the preview window. \par The terms \eqref{eq:mpc_cost_FinalTerm} \eqref{eq:mpc_cost_state} guarantee the tracking of the given reference trajectory, while \eqref{eq:mpc_cost_input} is added for obtaining a smoother ZMP trajectory. Even if the cost function $J_k$ is time-dependent, it is always positive and convex. \par The MPC problem can be summarized as follows: \begin{IEEEeqnarray}{LCL} \label{MPC_solution} r_k^{{zmp}^{*}} =& \argmin\limits_{\substack{\xi_k, \dots, \xi_{k + N}\\ r^{zmp}_k, \dots, r^{zmp}_{k + N-1}}} & J_k \\ \nonumber \\ &\text{s.t.} &{\xi}_{i+1} = F {\xi}_i + G {r}^{zmp}_i \nonumber \\ & & A_{c_k} {r}^{zmp}_{k} \leq {b}_{c_k} \nonumber \\ & & {\xi}_k = \bar{{\xi}} \nonumber \\ & & {r}^{zmp}_{k-1} = \bar{{r}}^{zmp}, \nonumber \end{IEEEeqnarray} where $i$ satisfies $k\le i \le N + k - 1$, $\bar{{r}}^{zmp}$ is the desired ZMP computed at the previous control iteration and $\bar{{\xi}}$ is the estimated DCM position. \par Since the cost function is a quadratic positive function and the constraints are linear, the optimal control problem is quadratic and it can be converted into a strictly convex quadratic programming problem (QP) of the form: \begin{IEEEeqnarray*}{CLLL} \min\limits_{w_k} & \IEEEeqnarraymulticol{3}{C}{\frac{1}{2} w_k^\top H_k w_k + g_k^\top w_k}\\ \text{s.t.} & A_{c_k}^i w_k &\le &b_{c_k}^i\\ & A_{c_k}^e w_k &= &b_{c_k}^e. \end{IEEEeqnarray*} Here the optimization variables are stacked inside the vector $w_k = \begin{bmatrix} \xi_k ^\top & \hdots & \xi_{k + N} ^\top & r_{k}^{zmp ^\top} & \hdots & r_{k + N - 1} ^{zmp ^\top} \end{bmatrix}^\top$. The Hessian matrix $H_k$ and the gradient vector $g_k$ can be obtained from \eqref{eq:mpc_const_function}. The inequality constraint matrix $A_{c_k}^i$ and vector $b_{c_k}^i$ embed the ZMP constraint \eqref{eq:inequality_constraints}. While the equality constraints, $A_{c_k}^e$ and $b_{c_k}^e$ are determined using the prediction model \eqref{eq:dcm_discrete}. \subsubsection{ZMP-CoM Controller} \label{ZMP-CoM-Controller} Independently from the chosen DCM controller, namely either the controller in Sec.~\ref{instantaneous-control} or in~\ref{predictive-control}, one obtains a desired ZMP and CoM position and velocity to be stabilised. As consequence, another control loop is needed after the DCM controller. In this paper, we choose the proposed control law \cite{Choi2007}, i.e.: \begin{equation} \label{eq:zmp_controller} \dot{x}^* = \dot{x}_{ref} - K_{zmp}(r^{zmp}_{ref} - r^{zmp}) + K_{com} (x_{ref} - x), \end{equation} where $K_{com} > \omega I_2$ and $0_2 < K_{zmp} < \omega I_2.$ \subsection{Whole-body QP control layer} The main control objective for the whole-body QP control layer is to stabilise some robot kinematic quantities by using the entire robot body. We here choose to track the position of the CoM, the torso orientation, and the left and right feet position and orientation. To do so, we use a stack of tasks formulation. The tracking of the feet poses and of the CoM position is considered as high priority tasks (hard constraint), while the torso orientation is considered as a low priority task (soft constraint). Furthermore, a postural condition is also added as a low-priority task. Then, the following cost function is defined: \begin{IEEEeqnarray}{LL} \IEEEyesnumber \phantomsection \label{eq:velocity_control_cost} f(\nu) {=} \frac{1}{2} [ & (v ^ * _ {\mathcal{T}}{-}J_{\mathcal{T}} \nu)^\top K_{\mathcal{T}} (v ^ * _ {\mathcal{T}}{-}J_{\mathcal{T}} \nu) {+} (\dot{s} {-} \dot{s}^ * ) ^\top \Lambda (\dot{s} {-} \dot{s}^*)]. \label{eq:velocity_control_regularization_cost} \IEEEeqnarraynumspace \end{IEEEeqnarray} with $K_{\mathcal{T}} > 0$ and $ v^*_{\mathcal{T}} = -K _{\omega _{\mathcal{T}}} \text{sk}(\prescript{\mathcal{I}}{}{R} _{\mathcal{T}} {\prescript{\mathcal{I}}{}{R}_{\mathcal{T}} ^{*}}^{\top})^{\vee}, $ where $K _{\omega _{\mathcal{T}}} > 0$. This latter control law guarantees almost-global stability and convergence of ${}^\mathcal{I}R _{\mathcal{T}}$ to ${}^\mathcal{I}R _{\mathcal{T}}^*$ \cite{Olfati-Saber:2001:NCU:935467}. \par The postural task \eqref{eq:velocity_control_regularization_cost}, with $\Lambda> 0$, is achieved by asking for a desired joints velocity that depends on the error between the desired and measured joints position \begin{equation} \label{eq:regularization_term} \dot{s}^ * = -K_{s} (s - s^d), \end{equation} where $K_{s}$ is a positive definite matrix. \par Concerning the hard constraints, we have: \begin{equation} \label{eq:velocity_control_jacobians} J_{\mathcal{C}}(\nu) \nu = v^ * _ {\mathcal{C}}, \quad J_{\mathcal{LF}}(\nu) \nu = v^ * _ {\mathcal{LF}}, \quad J_{\mathcal{RF}}(\nu) \nu = v^ * _ {\mathcal{RF}}, \end{equation} where $v^*_{\mathcal{C}}$ is the linear velocity of the CoM, $v^*_{\mathcal{LF}}$ and $ v^*_{\mathcal{RF}}$ are respectively the desired left foot and right foot velocities. More specifically $v^*_{\mathcal{\#F}}$, where $\#=[\mathcal{R}, \mathcal{L}]$, is chosen as: \begin{equation} \label{feetVelocitiesStar} v^*_ {\# \mathcal{F}} = \prescript{\mathcal{I}}{}{\dot{p}}^*_{\# \mathcal{F}} - \begin{bmatrix} K^p _{x _{\# f}} e^p_{\#f} + K^i _{x _{\# f}}\int{ e^p_{\#f} \diff t}\\ K _{\omega _{\# f}} \text{sk}(\prescript{\mathcal{I}}{}{R}_{\#\mathcal{F}} \prescript{\mathcal{I}}{}{R} _{\# \mathcal{F}} ^{*^\top} )^{\vee} \end{bmatrix}. \end{equation} Here $e^p_{\#f} = \prescript{\mathcal{I}}{}{p}_{\# \mathcal{F}} - \prescript{\mathcal{I}}{}{p}^*_{\# \mathcal{F}}$, while the gains $K^p _{x _{\# f}}$, $K^i _{x _{\# f}}$, $K _{\omega _{\# f}}$ are positive definite matrices. \par Finally, the desired velocity of the CoM $v^ * _{\mathcal{C}}$ is chosen as: \[ v^*_{\mathcal{C}} = \dot{x}^* - K^p _{\mathcal{C}}(x - x^*) - K^i _{\mathcal{C}}\int{ x - x^* \diff t}, \] where the gain matrices are positive definite positive $\dot{x}^*$ is the output of the ZMP-CoM \eqref{eq:zmp_controller} controller and $x^*$ is the integrated signal. Finally, we add constraints on the maximum velocity \begin{equation} \label{eq:ik_jacobian_limits} \dot{s}^- \le \dot{s} \le \dot{s}^+. \end{equation} \par The above hierarchical control objectives can be cast into a whole-body optimization problem: \begin{IEEEeqnarray}{LCL} \nu ^* =& \argmin\limits_{\nu} \frac{1}{2} [&(v^* _ {\mathcal{T}}-J_{\mathcal{T}} \nu)^\top K_{\mathcal{T}} (v^*_{\mathcal{T}}-J_{\mathcal{T}} \nu) + \nonumber \\ & & (\dot{s} - \dot{s}^ * )^\top \Lambda (\dot{s} - \dot{s}^*) ] \label{velocityQP}\\ &\text{s.t.} & \dot{s} = S \nu \nonumber \\ & & \dot{s}^ * = -K_{s} (s - s^d) \nonumber \\ & & J_{\mathcal{C}} \nu = v^ * _ {\mathcal{C}} \nonumber \\ & & J_{\mathcal{LF}} \nu = v^ * _ {\mathcal{LF}} \quad J_{\mathcal{RF}} \nu = v^ * _ {\mathcal{RF}} \nonumber \\ & & \dot{s}^- \le \dot{s} \le \dot{s}^+ \nonumber \end{IEEEeqnarray} Since the decision variable is the robot velocity $\nu$ and the body velocity depends, through the Jacobian matrices, linearly on $\nu$ the optimization problem can be converted to the QP problem of the form: \begin{IEEEeqnarray*}{CLL} \min\limits_{\nu} & \IEEEeqnarraymulticol{2}{C}{\frac{1}{2}\nu ^\top H \nu + g^\top \nu} \\ \mbox{s.t.} & A_{c} \nu = b_{c}\\ &\dot{s}^- \le \dot{s} \le \dot{s}^+. \end{IEEEeqnarray*} The Hessian matrix $H$ and the gradient vector $g$ are evaluated from \eqref{eq:velocity_control_cost}. The constraint matrix and vector $A_c$ and $b_c$ are obtained from \eqref{eq:velocity_control_jacobians}. Using this formulation the optimization problem can be solved using a standard numerical QP solver. \subsubsection{Position and velocity controlled robot} \label{subsubsec-pos-vel-control} It is important to notice that the outcome of~\eqref{velocityQP} is (also) the robot joint velocity. When a robot velocity controller is available, one can set these joint velocities to the low level robot controller. In this case, the \emph{*} quantities in~\eqref{velocityQP} can be evaluated by using robot sensor feedback, and the robot is said to be \emph{velocity controlled}. On the other hand, if the robot velocity control is not available, one may integrate the outcome of~\eqref{velocityQP} to obtain desired joint position to be set to a low level robot position controller. In this case, the \emph{*} quantities in~\eqref{velocityQP} can be evaluated by using the desired integrated quantities instead of sensor feedback, and~\eqref{velocityQP} behaves as an inverse kinematics module, and the robot is said to be \emph{position controlled}.

05eef441738c4fc55806f472d783f7bf3b88912f

\section{Introduction} We investigate in this note the reconstructibility of a homogeneous polynomial from its partial derivatives. The study can date back to J.~Carlson and Ph.~Griffiths, who in \cite{CG} showed that a generic homogeneous polynomial could be reconstructed, up to a nonzero constant multiple, from its Jacobian ideal, or equivalently, from its first order partial derivatives; in that paper, they used this result to study variation of Hodge structures and proved the global Torelli theorem for hypersurfaces. For further developments of the determination of a homogeneous polynomial by its Jacobian ideal, see \cite{ZW} and references therein. In the classical theory of variation of Hodge structures for smooth hypersurfaces, as in \cite{CG}, only first order derivatives of the defining homogeneous polynomials are involved. We can also construct higher order versions of this classical theory; see for instance \cite{DGI}. In this higher order analogous theory, a problem arises concerning the reconstructibility of a homogeneous polynomial from its higher order partial derivatives. In this paper, we will solve this problem and prove that a generic homogeneous polynomial has the desired property. Let $S=\bb{C}[x_0,x_1,\cdots,x_n]$ be the graded polynomial ring in $n+1$ variables with coefficients in $\bb{C}$ $$ S=\bigoplus_{d=0}^\infty S_{n,d}, $$ where $S_{n,d}$ is the vector space of homogeneous polynomials of degree $d$. Given $f\in S_{n,d}$ and a natural number $k\geq 0$. Denote by $J_k(f)$ the graded ideal of $S$ generated by all partial derivatives of $f$ of order $k$ and by $E_k(f)$ the degree $d-k$ homogeneous component of $J_k(f)$, that is, the vector space spanned by all $k$-th order partial derivatives of $f$. We will prove the following theorem. \begin{thm}\label{thm2} Given $n\geq1$ and $d\geq 3$, and $k\geq 1$ a natural number such that $k\leq \frac{d}{2}-1$. Suppose $f$ is a generic homogeneous polynomial in $S_{n,d}$. Let $g$ be another homogeneous polynomial in $S_{n,d}$ such that $E_k(f)=E_k(g)$, then $g\in\bb{C}^*f$. \end{thm} The underlying idea in the proof is very simple, so we give an outline here. We will show that $E_{k-1}(g)=E_{k-1}(f)$, then apply induction on $k$ to obtain $E_r(g)=E_r(g)$ for all $0\leq r\leq k$. Since $E_0(f)$ is essentially nothing but $\bb{C}f$, the conclusion follows immediately. Note that we already have a more precise result, Theorem 1.1 in \cite{ZW}, for the case $k=1$. But we do not need to use it to prove Theorem \ref{thm2}; instead, we will use induction on $k$ until the case $k=0$ is reached. In addition, the restriction $k\leq\frac{d}{2}-1$ is given in order to ensure that $\dim E_{k+1}(f)=\dim S_{n,k+1}$ for a generic $f$, see Lemma \ref{lem1} below. \bigskip We almost proved the above theorem when we wrote the paper \cite{ZW}, but we did not write down the complete proof due to the lack of adequate knowledge of its applications. We would like to thank Professor A. Dimca for his kindly pointing out the applications to the study of higher Jacobians, associated polar maps, and so on. We thank the referee for valuable comments. We also thank Yau Mathematical Sciences Center for their financial support and wonderful working atmosphere. \section*{Notations} As in the introduction, $S_{n,d}$ denotes the vector space of homogeneous polynomials of degree $d$. The multi-index set $$ \bb{N}^{n+1}=\{(i_0,i_1,\cdots, i_n)\,:\,i_j\geq 0\ \text{for } j=0,1,\cdots, n\}. $$ We denote by $I$ an element of $\bb{N}^{n+1}$. We shall see $\bb{N}^{n+1}$ as a subset of the vector space $\bb{R}^{n+1}$; among operations on $\bb{N}^{n+1}$ are addition, subtraction and multiplication by a positive integer: $$ I\pm I'=(i_0\pm i_0',i_1\pm i_1',\cdots, i_n\pm i_n') $$ and $$ mI=(mi_0,\cdots,mi_n) $$ for $I=(i_0,\cdots, i_n), I'=(i_0',\cdots, i_n')$, and $m\in\bb{N}$. Denote by $\e_j,j=0,\cdots, n$ the canonical basis of $\bb{R}^{n+1}$: $$ \e_j=(0,\cdots, 0,1,0,\cdots,0), $$ where 1 lies in the $j$-th entry. Using this basis, we may write $I=(i_0,\cdots, i_n)$ as $I=\sum_{j=0}^ni_j\e_j$. Moreover, there is an obvious partial ordering $"\geq"$ on $\bb{N}^{n+1}$, with $$ I=(i_0,i_1,\cdots,i_n)\geq I'=(i_0',\cdots,i_n')\Leftrightarrow i_j\geq i_j',\quad j=0,\cdots, n, $$ or more concisely, $$ I\geq I'\Leftrightarrow I-I'\in\bb{N}^{n+1}. $$ The order of $I=(i_0,\cdots, i_n)$: $$ |I|=i_0+\cdots+i_n. $$ For $f\in S_{n,d}$, the partial derivative of $f$ of type $I$ is $$ D_If=\frac{\p^{|I|}f}{\p x_0^{i_0}\p x_1^{i_1}\cdots\p x_n^{i_n}}. $$ By definition, $E_k(f)$ is the vector subspace of $S_{n,d-k}$ spanned by $D_If, |I|=k$; thus we have $$ E_k(f)=\langle D_If\,:\,|I|=k\rangle. $$ \section{Polynomials determined by higher order derivatives} In this section, we will give the proof of Theorem \ref{thm2}. We begin our proof with the following lemma. \begin{lem}\label{lem2} Let $f\in S_{n,d}$. If $k\geq 1$ and $\dim E_k(f)=\dim S_{n,k}$, then $\dim E_{k-1}(f)=\dim S_{n,k-1}$. \end{lem} \begin{proof} The proof is almost obvious: if we are given a linear relation $$ \sum_{|I|=k-1}a_ID_If=0, $$ by taking differentiation with respect to the variable $x_0$, it follows that $$ \sum_{|I|=k-1}a_ID_{I+\e_0}f=0. $$ On the other hand, the assumption of $E_k(f)$ implies that $\{D_{I+\e_0}f\,:\,|I|=k-1\}$ are linearly independent, so $a_I=0$ for all $I$. \end{proof} An induction on $k$ gives the following corollary. \begin{cor}\label{cor1} Let $f\in S_{n,d}$. If $k\geq 1$ and $\dim E_k(f)=\dim S_{n,k}$, then $\dim E_r(f)=\dim S_{n,r}$ for all $0\leq r\leq k$. \end{cor} As a second step to the proof of Theorem \ref{thm2}, we show the following proposition. \begin{prop}\label{prop1} Given $n\geq 1$, $d\geq 3$, and $k\geq 1$. Let $f,g\in S_{n,d}$ be such that $E_k(g)=E_k(f)$ and $\dim E_{k+1}(f)=\dim S_{n,k+1}$, then $E_{k-1}(g)=E_{k-1}(f)$. \end{prop} \begin{proof} We will show $E_{k-1}(g)\subseteq E_{k-1}(f)$. This is sufficient for our purpose because the two vector spaces have the same dimension by Corollary \ref{cor1}. From $E_k(g)=E_k(f)$, we have a system of linear relations as follows: for all $I\in\bb{N}^{n+1}$ such that $|I|=k$, we have \begin{equation}\label{eq2} D_Ig=\sum_{|I'|=k}a_{I,I'}D_{I'}f. \end{equation} for some $a_{I,I'}\in\bb{C}$. Our discussions in the sequel will be divided into two steps. {\bf Step 1: Differentiating equations. } Fix $I$ and $0\leq p\leq n$ such that $I\geq \e_p$. For any $0\leq q\leq n$, we will apply the equality $D_{\e_q}D_Ig=D_{\e_p}D_{(I-\e_p)+\e_q}g$ to equation \eqref{eq2}; to this end, we obtain first \begin{eqnarray*} D_{\e_q}D_Ig&=&D_{\e_q}\biggl(\sum_{|I'|=k}a_{I,I'}D_{I'}f\biggr)\\ &=&\sum_{|I'|=k}a_{I,I'}D_{I'+\e_q}f;\\ \end{eqnarray*} second, \begin{eqnarray*} D_{\e_p}D_{(I-\e_p)+\e_q}g&=&D_{\e_p}\biggl(\sum_{|I'|=k}a_{(I-\e_p)+\e_q,I'}D_{I'}f\biggr)\\ &=&\sum_{|I'|=k}a_{I-\e_p+\e_q,I'}D_{I'+\e_p}f. \end{eqnarray*} From our assumption $\dim E_{k+1}(f)=\dim S_{n,k+1}$, it follows that $\{D_Jf\,:\, |J|=k+1\}$ are linearly independent. Therefore, using $D_{\e_q}D_Ig=D_{\e_p}D_{(I-\e_p)+\e_q}g$ and comparing the coefficients of each term $D_Jf$, we obtain that $$ a_{I,J-\e_q}=a_{I-\e_p+\e_q,J-\e_p} $$ for all $|J|=k+1$. Here we used the convention that $a_{I,J-\e_q}=0$ if $J\not\geq \e_q$. Since the above conclusion holds for all $I,J,p,q$ satisfying $I\geq\e_p$, it follows that for all $I,I',p,q$ such that $|I|=|I'|=k$ and $I\geq\e_p$, \begin{equation}\label{eq3} a_{I,I'}=a_{I-\e_p+\e_q,I'-\e_p+\e_q}. \end{equation} {\bf Step 2: Considering $(k-1)$-th order partial derivatives. } Let $K\in\bb{N}^{n+1}$ be such that $|K|=k-1$, then the Euler formula for $D_Kg$ gives \begin{eqnarray}\label{eq4} (d-k+1)D_Kg=\sum_{p=0}^nx_pD_{K+\e_p}g. \end{eqnarray} Substituting \eqref{eq2} into \eqref{eq4}, we have $$ (d-k+1)D_Kg=\sum_{p=0}^n\sum_{|I'|=k}x_pa_{K+\e_p,I'}D_{I'}f. $$ By \eqref{eq3}, we deduce first of all that $a_{K+\e_p,I'}=0$ if $I'\not\geq \e_p$, and thus $$ (d-k+1)D_Kg=\sum_{p=0}^n\sum_{I'\geq\e_p}x_pa_{K+\e_p,I'}D_{I'}f=\sum_{p=0}^n\sum_{I'\geq\e_p}x_pa_{K+\e_p,(I'-\e_p)+\e_p}D_{I'}f, $$ or written in a more convenient way, \begin{eqnarray*} (d-k+1)D_Kg&=&\sum_{p=0}^n\sum_{|K'|=k-1}x_pa_{K+\e_p,K'+\e_p}D_{K'+e_p}f\\ &=&\sum_{|K'|=k-1}\biggl(\sum_{p=0}^nx_pa_{K+\e_p,K'+\e_p}D_{K'+e_p}f\biggr). \end{eqnarray*} Now the relations \eqref{eq3} imply that $a_{K+\e_p,K'+\e_p}=a_{K+\e_q,K'+\e_q}$ for any $p,q=0,\cdots, n$; therefore, we obtain $$ (d-k+1)D_Kg=\sum_{|K'|=k-1}a_{K+\e_0,K'+\e_0}\biggl(\sum_{p=0}^nx_pD_{K'+e_p}f\biggr). $$ By the Euler formula for $D_{K'}f$, we have that $$ \sum_{p=0}^nx_pD_{K'+e_p}f=(d-k+1)D_{K'}f, $$ so $$ D_Kg=\sum_{|K'|=k-1}a_{K+\e_0,K'+\e_0}D_{K'}f. $$ Since this holds for all $K$ satisfying $|K|=k-1$, it follows that $E_{k-1}(g)\subseteq E_{k-1}(f)$. \end{proof} \subsection{Linear independence of partial derivatives} As a final step to our proof of Theorem \ref{thm2}, we need the following lemma, which is interesting in its own right; see also \cite{DGI}, Proposition 3.4. \begin{lem}\label{lem1} Given $n\geq 1$ and $d\geq 3$, suppose $0\leq k\leq\frac{d}{2}$. Then for a generic $f\in S_{n,d}$, we have $$ \dim E_k(f)=\dim S_{n,k}. $$ \end{lem} \begin{proof} In fact, this result has already been proved in \cite{IE} as well as in \cite{Ia}. For completeness, we give a detailed proof in our setting. Suppose given a linear relation \begin{equation}\label{eq1} \sum_{|I|=k}a_ID_If=0. \end{equation} Let $R=\bb{C}[y_0,\cdots,y_n]$ be the polynomial ring in variables $y_0,\cdots, y_n$. The apolar action of $R$ on $S$ is given by $$ R\times S\to S, (F,f)\mapsto\langle F,f\rangle=F\bigl(\frac{\p}{\p x_0},\cdots,\frac{\p }{\p x_n}\bigr)f(x_0,\cdots, x_n); $$ see \cite{Ia}. Denote $$ P(y_1,\cdots,y_n)=\sum_{|I|=k}a_I y^I, $$ then the relation \eqref{eq1} can be simply written as $\langle P,f\rangle=0$. From \cite[Proposition 3.4]{Ia}, it follows that for a generic $f$, the map $R_k \to S_{d-k}, Q\mapsto\langle Q,f\rangle$ is injective since $k\leq d/2$ implies that $\dim S_{d-k}\geq\dim R_k$. Hence from relation \eqref{eq1}, we obtain $P=0$ when $f$ is generically chosen, i.e., $a_I=0$ for all $I$ with $|I|=k$. We are done. \end{proof} \begin{rk} In view of the obvious bound for $\dim E_k(f)$ given by $$ \dim E_k(f)\leq\min\{\dim S_{n,k},\dim S_{n,d-k}\}, $$ the condition on $k$ in Lemma \ref{lem1} is optimal. \end{rk} \subsection{Proof of Theorem \ref{thm2}} Let $f$ be a generic polynomial in $S_{n,d}$ and $E_k(g)=E_k(f)$. Under the assumption $k\leq\frac{d}{2}-1$, it follows that $k+1\leq\frac{d}{2}$, hence, by Lemma \ref{lem1}, we have $\dim E_{k+1}(f)=\dim S_{n,k+1}$; therefore the requirements in Proposition \ref{prop1} are satisfied. By Proposition \ref{prop1}, it follows that $E_{k-1}(g)=E_{k-1}(f)$. Note that by Corollary \ref{cor1}, we have $\dim E_k(f)=\dim S_{n,k}$, so the requirements in Proposition \ref{prop1} are satisfied with $k$ replaced by $k-1$ and we obtain $E_{k-2}(g)=E_{k-2}(f)$. These arguments can be repeated until we obtain $E_0(g)=E_0(f)$. By definition, we have $E_0(g)=\bb{C}g$ and $E_0(f)=\bb{C}f$, therefore $g$ is a constant multiple of $f$. \section{Applications} As pointed out in the introduction, the most remarkable application of the results in this paper lies in the study of higher order analogue of variation of Hodge structures for hypersurfaces; see \cite{DGI}. In this section, we give some other applications in the study of deformations of homogeneous polynomials. For $k\geq 0$, denote by $\mathcal{U}_{n,d}(k)$ the following set $$ \mathcal{U}_{n,d}(k)=\{\,f\in S_{n,d}\,:\,\dim E_k(f)=\dim S_{n,k}\}. $$ From semi-continuity of $\dim E_k(f)$ with respect to $f$, we see that $\mathcal{U}_{n,d}(k)$ is a Zariski open subset of $S_{n,d}$. Obviously, we have $\mathcal{U}_{n,d}(k)=\emptyset$ if $k>\frac{d}{2}$. From Lemma \ref{lem1}, we have the following result. \begin{cor} Given $n\geq 1$ and $d\geq 3$. For $k\leq\frac{d}{2}$, the set $\mathcal{U}_{n,d}(k)$ is a Zariski open dense subset of $S_{n,d}$. \end{cor} In addition, for any $f\in\mathcal{U}_{n,d}(k)$, we have by definition that $\dim E_k(f)=\dim S_{n,k}$; by Lemma \ref{lem2}, we deduce that $\dim E_{k-1}(f)=\dim S_{n,k-1}$, that is $f\in\mathcal{U}_{n,d}(k-1)$. In other words, for fixed $n$ and $d$, the sequence of sets $\{\mathcal{U}_{n,d}(k)\}$ satisfies the following relations $$ \mathcal{U}_{n,d}(0)\supseteq\mathcal{U}_{n,d}(1)\supseteq\cdots\supseteq\mathcal{U}_{n,d}(k)\supseteq\mathcal{U}_{n,d}(k+1)\supseteq\cdots. $$ Note that $\mathcal{U}_{n,d}(k)$ is a cone in $S_{n,d}$, hence we can consider its projectivization, denoted by $\bb{P}(\mathcal{U}_{n,d}(k))$, in $\bb{P}(S_{n,d})$. Similar to the construction in \cite{ZW}, the assignment $$ [f]\mapsto\bb{P}(E_k(f)) $$ gives a well-defined map, denoted by $\varphi_k$, from $\bb{P}(\mathcal{U}_{n,d}(k))$ to an obvious Grassmannian for $k\leq\frac{d}{2}$. Using Proposition \ref{prop1} and Lemma \ref{lem2}, we prove the following result which gives an extension of Corollary 7.7 in \cite{ZW}. \begin{cor} For $k\leq\frac{d}{2}-1$, the map $\varphi_k: \bb{P}(\mathcal{U}_{n,d}(k))\ni[f]\mapsto\bb{P}(E_k(f))$ is injective when restricted to $\bb{P}(\mathcal{U}_{n,d}(k+1))$. In particular, it is generically injective. \end{cor} \begin{proof} To begin the proof, suppose $[f]$ and $[g]$ are two elements of $\bb{P}(\mathcal{U}_{n,d}(k+1))$ such that $\varphi_k([f])=\varphi_k([g])$. By the definition of $\varphi_k$, this means that $E_k(f)=E_k(g)$. Now the assumption $[f]\in\bb{P}(\mathcal{U}_{n,d}(k+1))$ implies that $\dim E_{k+1}(f)=\dim S_{n,k+1}$, hence by Proposition \ref{prop1}, we obtain $E_{k-1}(f)=E_{k-1}(g)$. An induction argument on $k$ gives $[f]=[g]$, which goes exactly the same as the proof of Theorem \ref{thm2} where only the properties $\dim E_{k+1}(f)=\dim S_{n,k+1}$ and $E_k(f)=E_k(g)$ are essentially used. Thus, $\varphi_k$ is injective on $\bb{P}(\mathcal{U}_{n,d}(k+1))$. \end{proof} \begin{rk} We do not know whether $\varphi_k$ is injective on $\bb{P}(\mathcal{U}_{n,d}(k))$ or not, except the case $k=\frac{d}{2}$ where $\varphi_{\frac{d}{2}}$ is a constant map, because in this case $E_k(f)=S_{n,k}$ for any $f\in\mathcal{U}_{n,d}(k)$. \end{rk}

7a04bfdd1caaaf0f088cf0bcc557d619ae0f240e

\section{A Heuristic Approach} \noindent We first discuss the limitation of the solution discussed in the previous subsection. It is easy to see that Step \ref{step:predecessor} of Algorithm \ref{algo:graphConversion}, where the set of predecessors of a node is constructed, may explode. Consider the following example. \begin{example} Consider a node $v^{p}_i$ of LPG requires 10 inputs and each input is provided by 10 nodes of the dependency graph. The number of possible predecessor nodes of $v^{p}_i$ is $10^{10}$. \hfill$\blacksquare$ \end{example} \noindent If the number of inputs of a node or the number of nodes providing an input, increases, the number of predecessor nodes also increases exponentially. In our heuristic, we try to address the above issue. The main motivation of this algorithm is to reduce the search space of the original problem. On one hand, we attempt to reduce the number of combinations generated at Step \ref{step:predecessor} of Algorithm \ref{algo:graphConversion}. On the other hand, we try to restrict the number of nodes generated at a particular level of the LPG. Our approach is based on the notion of anytime algorithms\cite{yan2015anytime} using beam search. Beam search uses breadth-first search to build its search space. However, at each level of the graph, we store only a fixed number of nodes, called the beam width. The greater the beam width, the fewer the number of nodes pruned. While constructing $G_P$, all predecessor nodes of the set of nodes at a particular level are computed, as earlier. However, only a subset of the nodes is stored depending on the beam width of the algorithm. Consider $\{v^{(p)}_{i_1}, v^{(p)}_{i_2}, \ldots, v^{(p)}_{i_k}\} \subset V_P$ be the set of nodes generated at level $i$ and the beam width is $k_1 \le k$. Therefore, only $k_1$ out of $k$ nodes are stored. The nodes are selected based on the values of the cumulative Pareto optimal tuples computed till $v^{(p)}_{i_j}$, for $j = 1, 2, \ldots, k$. At each level, the selected set of nodes is ranked between $1, 2, \ldots, k_1$. A node with rank $i$ has higher priority than a node with rank $j$, where $j > i$. Consider $l$ be the number of levels in the dependency graph, where the level $l$ consists of $V_{End}$. The selection criteria for choosing $k_1$ nodes from $(l - 1)^{th}$ level is discussed below. \begin{itemize} \item The feasible non dominated tuples $V_{(l-1)}^*$ corresponding to the cumulative Pareto optimal tuples computed till each $v^{(p)}_{{(l-1)}_j}$, for $j = 1, 2, \ldots, k$ are computed first. \item If $|V_{(l-1)}^*| = k_1$, $V_{(l-1)}^*$ is returned. \item If $|V_{(l-1)}^*| > k_1$, the following steps are performed: \begin{itemize} \item The utility $U(t_j)$ corresponding to each tuple $t_j \in V_{(l-1)}^*$ is computed as follows: {\scriptsize{ \begin{equation} A_k = Max_{t_j \in V_{(l-1)}^*}{\cal{P}}_k^{(t_j)} \end{equation} \begin{equation} B_k = Min_{t_j \in V_{(l-1)}^*}{\cal{P}}_k^{(t_j)} \end{equation} \begin{equation} NV({\cal{P}}_k^{(t_j)}) = \left\{ \begin{array}{ll} \frac{A_k - {\cal{P}}_k^{(t_k)}}{A_k - B_k} & ,\text{for a negative QoS} \\ \frac{{\cal{P}}_k^{(t_j)} - B_k}{A_k - B_k} & ,\text{for a positive QoS} \\ 1 & ,A_k = B_k \\ \end{array} \right. \end{equation} \begin{equation}\label{equ:utility} U(t_j) = \sum_{{\cal{P}}_k \in t_j} NV({\cal{P}}_k^{(t_j)}) \end{equation} }}where $NV({\cal{P}}_k^{(t_j)})$ is the normalized value of ${\cal{P}}_k$ for $t_j \in V_{(l-1)}^*$. \item Tuples in $V_{(l-1)}^*$ are sorted in descending order based on utility and finally, the first $k_1$ tuples are chosen, which are returned. \end{itemize} \item If $|V_{(l-1)}^*| < k_1$, $(k_1 - |V_{(l-1)}^*|)$ more tuples are chosen from $\{v^{(p)}_{{(l-1)}_1}, v^{(p)}_{{(l-1)}_2}, \ldots, v^{(p)}_{{(l-1)}_k}\} \setminus |V_{(l-1)}^*|$ using the same procedure discussed above. \item Finally, the selected tuples are ranked based on their utility value. The rank of the tuple with the highest utility value is set to 1. \end{itemize} \noindent Consider ${\cal{TP}}_{(i + 1)} = \{t_{(i + 1)_1}, t_{(i + 1)_2}, \ldots, t_{(i + 1)_{k'}}\}$ be the set of tuples selected at level $(i + 1)$ of $G_P$. It may be noted, $|{\cal{TP}}_{(i + 1)}| = k' \le k_1$. Without loss of generality, we assume that the priority of $t_{(i + 1)_{j1}}$ is greater than that of $t_{(i + 1)_{j2}}$, where $j_1 < j_2$. Also consider $\{v^{(p)}_{(i + 1)_1}, v^{(p)}_{(i + 1)_2}, \ldots, v^{(p)}_{(i + 1)_{k''}}\} \subset V_P$ be the set of nodes corresponding to the tuples in ${\cal{TP}}_{(i + 1)}$. We further assume that at least one tuple corresponding to $v^{(p)}_{(i + 1)_{j1}}$ has higher priority than any tuple corresponding to $v^{(p)}_{(i + 1)_{j2}}$, where $j_1 < j_2$. We now discuss the selection criteria for choosing $k_1$ nodes from level $i$, where $i < (l - 1)$: \begin{itemize} \item A cumulative Pareto optimal tuple $t$ corresponding to a node $v$ is selected for the rank position 1, if the following conditions are satisfied: \begin{itemize} \item $v$ belongs to the set of predecessor nodes of $v^{(p)}_{(i + 1)_1}$. \item $t$ has the highest utility value among all the cumulative Pareto optimal tuples corresponding to all the predecessor nodes of $v^{(p)}_{(i + 1)_1}$. \end{itemize} \item In general, a cumulative Pareto optimal tuple $t$ corresponding to a node $v$ is selected for the rank position $j$, where $j \le k'$, if: \begin{itemize} \item $v$ belongs to $\cup_{x=1}^{j} Pred(Corr(t_{(i + 1)_x}))$, where $Corr(t_{(i + 1)_x})$ is the node corresponding to $t_{(i + 1)_x}$ and $Pred(Corr(t_{(i + 1)_x}))$ is the set of predecessor nodes of $Corr(t_{(i + 1)_x})$. \item $t$ has the highest utility value among all the cumulative Pareto optimal tuples corresponding to $\cup_{x=1}^{j} Pred(Corr(t_{(i + 1)_x})) \setminus \{\text{set of already selected tuples}\}$. \end{itemize} \item A cumulative Pareto optimal tuple $t$ corresponding to a node $v$ is selected for rank $j$ ($k' \le j \le k_1$), if: \begin{itemize} \item $t$ has the highest utility value among all the cumulative Pareto optimal tuples corresponding to $\cup_{x=1}^{k''} Pred(v^{(p)}_{(i + 1)_x}) \setminus \{\text{set of already selected tuples}\}$. \end{itemize} \end{itemize} \noindent Clearly, the search space of the algorithm is determined by the beam width. The main motivation of this algorithm is to be able to improve the solution quality monotonically with the increase in beam width. The selection procedure enforces this criteria. As the beam size increases, the number of pruned nodes decreases and the solution quality of the algorithm either remains same or improves, as formally stated below. \begin{lemma} The solution quality of the heuristic algorithm monotonically improves with increase in beam width. \hfill$\blacksquare$ \end{lemma} \begin{lemma}\label{lemma:AnytimeToOptimal} With an infinite beam width, the algorithm is identical to the Pareto optimal algorithm. \hfill$\blacksquare$ \end{lemma} \noindent Since at each level, a finite number of nodes are generated, the above lemma holds. Moreover, if the beam width of the heuristic algorithm is greater than or equal to the maximum number of nodes belonging to a level of $G_P$, where $G_P$ represents the complete LPG constructed by the optimal algorithm, no nodes are required to be pruned in this algorithm and thereby, the algorithm is identical to the optimal one. \section{Conclusion and future directions} \label{sec:conclusion} \noindent This paper addresses the problem of multi-constrained multi-objective service composition in IOM based on the Pareto front construction. Experimental results on real benchmarks show the effectiveness of our proposal. We believe that our work will open up a lot of new research directions in the general paradigm of multi-objective service composition. Going forward, we wish to come up with a theoretical bound on the solution quality degradation of our heuristic algorithms. \subsection{Dependency graph construction} \noindent The composition solutions are generated at run-time in response to a query. To find a response to a query, a dependency graph is constructed first. The dependency graph ${\cal{D}} = (V, E)$ is a directed graph, where $V$ is the set of nodes and $E$ is the set of edges. Each node $v_i \in V$ corresponds to a service ${\cal{W}}'_i \in W'$ that is eventually activated by the query inputs and each directed edge $(v_i, v_j) \in E$ represents a direct dependency between two services, i.e., the service corresponding to the node $v_i$ produces an output which is an input of the service corresponding to the node $v_j$. Each edge is annotated by the input-output of the services. Each solution to a query is either a path or a subgraph of ${\cal{D}}$ \cite{chattopadhyay2015scalable}. The dependency graph is constructed using the algorithm illustrated in \cite{DBLP:journals/tweb/ChattopadhyayBB17}. While constructing the dependency graph, here we additionally validate the local and global constraints. While the local constraints are validated once, when a service is selected for the first time, the global constraints are validated in each step of the solution construction. While an activated service is selected for node construction, the service is first validated against the set of local and global constraints. Each service ${\cal{W}}'_i$ corresponds to a set of skyline services. If any service from the skyline services violates any local / global constraint, we disregard that service by removing its corresponding QoS tuple from ${\cal{P}}'_i$ of ${\cal{W}}'_i$. If ${\cal{P}}'_i$ is empty, we do not construct any node corresponding to ${\cal{W}}'_i$. It may be noted, if a service ${\cal{W}}'_i$ violates any of the global constraints, any solution that includes ${\cal{W}}'_i$ also violates the global constraint. \begin{example} Consider Example \ref{exm:overview}. To respond to the query, while constructing the dependency graph, four services ${\cal{W}}'_1, {\cal{W}}'_2, {\cal{W}}'_3$ and ${\cal{W}}'_4$ are activated from the query inputs at first. It may be noted, ${\cal{W}}'_1$ is associated with three QoS tuples (500, 7, 93\%), (350, 4, 97\%) and (600, 13, 69\%), out of which one tuple (600, 13, 69\%) violates ${\cal{LC}}_1$, since its reliability is less than 70\%. Therefore, while validating ${\cal{LC}}_1$, the third tuple, i.e., (600, 13, 69\%) is removed from ${\cal{P}}^{'(1)}$ corresponding to ${\cal{W}}'_1$. \hfill$\blacksquare$ \end{example} \noindent During dependency graph construction, the set of services that can be activated by the query inputs are identified first. With the set of identified services, the dependency graph is constructed. Finally, backward breadth first search (BFS) is used in ${\cal{D}}$ to identify the set of nodes that are required to produce the set of query outputs. The remaining nodes are removed from the graph. \begin{figure}[!htb] \centering \includegraphics[width=\linewidth]{example.pdf} \caption{Dependency Graph in response to a query (a) generated from query inputs (b) after backward traversal} \label{fig:dependency} \end{figure} \begin{example} Consider the query in Example \ref{exm:overview}. Figure \ref{fig:dependency} shows the dependency graph constructed over the services described in Table \ref{tab:exampleModifiedServices} in response to the query. Figure \ref{fig:dependency}(a) shows the dependency graph constructed from the query inputs, while Figure \ref{fig:dependency}(b) shows the one generated after removal of unused nodes. In Figure \ref{fig:dependency}(a), the nodes marked with red represent the services that do not take part to produce the query outputs. \hfill$\blacksquare$ \end{example} \noindent If the dependency graph consists of a loop, we identify the loop and break the cycle \cite{rodriguez2015hybrid}. Finally, we partition the dependency graph into multiple layers using the approach used in \cite{DBLP:journals/tweb/ChattopadhyayBB17}, where a node $v_i$ belongs to a layer $L_k$, if for all the edges $(v_j, v_i)$, $v_j$ belongs to any layer $L_{k'}$, where $k' < k$. The first layer $L_0$ consists of a single node $V_{Start}$. Finally, we introduce dummy nodes in each layer if necessary as demonstrated in \cite{DBLP:journals/tweb/ChattopadhyayBB17}, to ensure that each node in a layer $L_i$ is connected only to the nodes in either its immediate predecessor layer $L_{(i - 1)}$ or its immediate successor layer $L_{(i + 1)}$. We assume that each dummy node has a QoS tuple with the best value for each QoS parameter. If a solution to a query consists of any dummy node, the dummy node is removed from the solution while returning the solution. The above assumption ensures that after removal of the dummy nodes, the QoS values of the solution remain unchanged. In the next subsection, we discuss the feasible Pareto optimal solution frontier generation technique. \section{Solution Architecture}\label{sec:method} \noindent In the following, we first define a few terminologies to build up the foundation of our work. \input{definition} \input{preprocessing} \input{dependencyGraphConstruction} \input{paretoOptimal} \input{anytime} \input{ga} \section{Solution Generation Using NSGA}\label{sec:ga} \noindent In this section, we present a different approach based on the non-dominated sorting genetic algorithm (NSGA) \cite{deb2002fast}. While the previous algorithms are deterministic algorithms, this is a randomized algorithm. This algorithm is basically an adaptive heuristic search algorithm based on the evolutionary ideas of natural selection and genetics \cite{srinivas1994genetic}. In this algorithm, a population of candidate solutions, called phenotypes \cite{srinivas1994genetic}, to a query is evolved toward better solutions. Each candidate solution is encoded into a binary string, called chromosome or genotype, which can be mutated and altered. The algorithm starts from a population of randomly generated chromosomes. In each iteration (called a generation) of the algorithm, a new set of chromosomes are generated and the fitness of every chromosome in the population is evaluated. The fitness value of a chromosome usually refers to the value of the objective function in the optimization problem being solved. In this paper, the fitness value of a chromosome is computed (as discussed later) from the QoS values of its corresponding phenotype. The more fit chromosomes are selected from the current population to form the next generation. Each chromosome in the current population is modified using different genetic operators (namely crossover, mutation) to construct its off-string for the next generation. The algorithm terminates after a fixed number of iterations, which is provided externally. Once a query comes to the system, the dependency graph is constructed first. The dependency graph contains all possible solutions to the query. The genetic algorithm is applied on the dependency graph to generate high-quality solutions. Before discussing the details of the algorithm, we first define the notion of a chromosome used in this paper. \begin{definition}{\em [\bf Chromosome:]} A chromosome is a binary string ${\cal{B}} = b_1 b_2 \ldots b_n$, where each bit $b_i \in \{b_1, b_2, \ldots, b_n\}$ of the string represents a node $v_i \in V$ in the dependency graph ${\cal{D}} = (V, E)$ constructed in response to a query and is defined as: $b_i = \left\{ \begin{array}{ll} 1 & \text{if } v_i \text{ is present in the solution sub-graph }\\ 0 & \text{otherwise } \end{array} \right. $ \hfill$\blacksquare$ \end{definition} \noindent It may be noted, each binary string is not a valid solution in terms of functional dependency (e.g., the string of all $0$'s). Therefore, in our algorithm, we consider only those chromosomes which represent a valid solution to a query in terms of functional dependency. Later in this section, we present an algorithm for constructing a chromosome. In the next subsection, we present an overview of our algorithm. \subsection{Algorithm to generate Pareto optimal solution} \noindent Our algorithm has the following steps. \begin{enumerate} \item {\bf{Initial population construction}}: First $n$ chromosomes are chosen randomly to construct the initial population. \item {\bf{Iteration of the algorithm}}: In each iteration of the algorithm, a new generation is created. The algorithm terminates after a fixed number of iterations $k$, where $k$ is provided externally. The following steps are performed in each iteration of the algorithm. \begin{enumerate} \item \label{step:selection} {\bf{Selection of parent chromosomes}}: Two chromosomes are chosen randomly from the current population using Roulette wheel selection \cite{srinivas1994genetic}. \item \label{step:crossover} {\bf{Crossover}}: With probability $p_c$ (i.e., crossover probability), the crossover operation is performed between the two chromosomes selected in the earlier step to generate two off-springs. Steps \ref{step:selection} and \ref{step:crossover} are performed $n$ number of times. \item {\bf{Mutation}}: Each off-spring, generated in the earlier step, is mutated to generate a new off-spring. \item {\bf{Fitness value computation}}: The fitness value for each chromosome in the population is computed. \item {\bf{Construction of new generation}}: The population now consists of more than $n$ chromosomes. Based on the fitness value of the chromosomes in the population, the best $n$ chromosomes are selected from the population and the rest of the chromosomes are removed. \end{enumerate} \item {\bf{Solution construction}}: Once the algorithm terminates, the best chromosomes, obtained from the final population, are returned as the solutions of our algorithm. \end{enumerate} \noindent We now briefly demonstrate each step of this algorithm. \subsection{Chromosome Construction} \noindent To construct a chromosome, we traverse the dependency graph backward starting from $V_{End}$. Each bit of a chromosome is first initialized by 0. During the traversal of the dependency graph, whenever a node of the dependency graph is visited, the corresponding bit in the chromosome is marked as 1. \begin{itemize} \item We start from $V_{End}$ and mark the corresponding bit in chromosome as 1. \item For each input $io$ of a visited node $v_i$, we randomly select a node $v_j$ that produces $io$ as output and mark the corresponding bit for $v_j$ as 1. \item Each node is processed only once. In other words, if a node $v_i$ is encountered more than once, we do not reprocess the node. \item The algorithm terminates when there is no node left for processing. \end{itemize} \noindent It may be noted, once the algorithm terminates, we get a valid solution in terms of functional dependency corresponding to the generated chromosome. \begin{example} Consider the example shown in Figure \ref{fig:chromosome}. Figure \ref{fig:chromosome}(a) shows a dependency graph constructed in response to a query, while Figure \ref{fig:chromosome}(b) presents an example candidate solution (i.e., phenotype) constructed from the dependency graph. \begin{figure}[!htb] \centering \includegraphics[width=\linewidth]{chromosome.pdf} \caption{Example of a chromosome} \label{fig:chromosome} \end{figure} As shown in the figure, $V_{End}$ takes two inputs $o_{12}$ and $o_{13}$. For each input, one node is randomly selected (as shown in the figure by bold line). The algorithm terminates when there is no node left for processing. Figure \ref{fig:chromosome}(c) represents the chromosome corresponding to the phenotype shown in Figure \ref{fig:chromosome}(b). \hfill$\blacksquare$ \end{example} \noindent We now discuss the two main operators of the genetic algorithm, namely, crossover and mutation. \subsection{Crossover} \noindent Crossover is a binary genetic operator used to obtain chromosomes from one generation to the next. In this paper, we consider only single point crossover \cite{srinivas1994genetic}. However, the crossover mechanism applied here is different from the normal single point crossover, in order to ensure that the off-springs generated after the crossover operation are valid solutions in terms of functional dependency. Consider two parent chromosomes $CR_1$ and $CR_2$ selected to participate in a crossover. Also consider $D_{CR1} = (V_{CR1}, E_{CR1})$ and $D_{CR2} = (V_{CR2}, E_{CR2})$ are two subgraphs (i.e., phenotypes) of the dependency graph $D$ corresponding to $CR_1$ and $CR_2$ respectively. The main intuition of this step is to randomly choose a common node $v_i$ between $D_{CR1}$ and $D_{CR2}$ at first and then exchange the set of nodes in $D_{CR1}$ and $D_{CR2}$ that are responsible to activate $v_i$ for both $D_{CR1}$ and $D_{CR2}$. In this way we ensure to obtain functionally valid off-springs. We now formally demonstrate the crossover operation. \begin{itemize} \item At first, the set of common nodes from $D_{CR1}$ and $D_{CR2}$ are identified. Out of them, one common node $v_i$ is chosen randomly. \item For each of the subgraphs $D_{CR1}$ and $D_{CR2}$, we compute the subgraphs $D'_{CR1} = (V'_{CR1}, E'_{CR1})$ containing all the paths from $V_{Start}$ to $v_i$ of $D_{CR1}$ and $D'_{CR2} = (V'_{CR2}, E'_{CR2})$ containing all the paths from $V_{Start}$ to $v_i$ of $D_{CR2}$. \item Finally, $D'_{CR1}$ and $D'_{CR2}$ are exchanged to generate two new off-springs $D_{CR3} = (V_{CR3}, E_{CR3})$ and $D_{CR4} = (V_{CR4}, E_{CR4})$, where, \\ {\small $V_{CR3} = (V_{CR1} \setminus \{v\}) \cup V'_{CR2}$}, where $v \in V'_{CR1}$ and $v$ does not belong to any path other than $V_{Start}$ to $v_i$ in $E_{CR1}$;\\ {\small $E_{CR3} = (E_{CR1} \setminus \{e\}) \cup E'_{CR2}$}, where $e \in E'_{CR1}$ and $e$ does not belong to any path other than $V_{Start}$ to $v_i$ in $E_{CR1}$; \\ {\small $V_{CR4} = V'_{CR1} \cup (V_{CR2} \setminus \{v\})$}, where $v \in V'_{CR2}$ and $v$ does not belong to any path other than $V_{Start}$ to $v_i$ in $E_{CR2}$; \\ {\small $E_{CR4} = E'_{CR1} \cup (E_{CR2} \setminus \{e\})$}, where $e \in E'_{CR2}$ and $e$ does not belong to any path other than $V_{Start}$ to $v_i$ in $E_{CR2}$. \end{itemize} \noindent Each off-spring is generated after scanning two subgraphs $D_{CR1}$ and $D_{CR2}$ once. For example, during the construction of $D_{CR3}$, $D_{CR1}$ is traversed backward starting from $V_{End}$. Once a node of $D_{CR1}$ is encountered, the node is copied in $D_{CR3}$. For each node $v_j$ (other than $v_i$) in $D_{CR3}$, all the nodes that provide at-least one input to $v_j$ in $D_{CR1}$ are copied in $D_{CR3}$. Once the traversal of $D_{CR1}$ is done, the traversal of $D_{CR2}$ starts backward starting from $v_i$. Similar to the previous, once a node of $D_{CR2}$ is encountered, the node is copied in $D_{CR3}$. For each node $v_j$ in $D_{CR3}$, all the nodes that provide at-least one input to $v_j$ in $D_{CR2}$ are copied in $D_{CR3}$. \begin{example} Consider two subgraphs of the dependency graph corresponding to two parent chromosomes as shown in Figure \ref{fig:crossover}(a) and (b). The node corresponding to ${\cal{W}}'_{12}$ is chosen as the common node. Finally, the off-springs generated using the above procedure are shown in Figure \ref{fig:crossover}(c) and (d). \begin{figure}[!htb] \centering \includegraphics[width=\linewidth]{crossover.pdf} \caption{Example of crossover} \label{fig:crossover} \end{figure} \hfill$\blacksquare$ \end{example} \subsection{Mutation} \noindent Mutation is a unary genetic operator used to obtain a chromosome from one generation to the next. Like crossover, mutation is also performed in a different way than normal mutation operation in our approach to ensure functional dependency preserving solutions are generated. Consider a parent chromosome $CR_1$ to participate in mutation. Also consider $D_{CR1} = (V_{CR1}, E_{CR1})$ is the subgraph of the dependency graph $D$ corresponding to $CR_1$ and $p_m$ is the mutation probability. The intuition of this step is as follows. We traverse $D_{CR1}$ in backward. While traversing, with probability $p_m$ we choose a node $v_i$ to be mutated. If a node $v_j$ is selected instead of $v_i$ after mutation, we randomly generate the nodes that are needed to activate $v_j$ from the query inputs. In this manner, we ensure to obtain a functionally valid off-spring. We now formally discuss the mutation operation. \begin{itemize} \item We start traversing $D_{CR1}$ backward starting from $V_{End}$ to generate a new off-spring $D_{CR2} = (V_{CR2}, E_{CR2})$. We first add $V_{End}$ in $V_{CR2}$. \item For each input $io$ of a node $v_i \in V_{CR2}$, if there is only one node $v_j \in V$ in the dependency graph $D$ to produce $io$ as output, we add $v_j$ in $V_{CR2}$ and $(v_j, v_i)$ in $E_{CR2}$. \item For each input $io$ of a node $v_i \in V_{CR2}$, if $v_i \in V_{CR1}$ and there exists multiple nodes in the dependency graph $D$ to produce $io$ as output, with probability $(1 - p_m)$ we add $v_j$ in $V_{CR2}$, where $(v_j, v_i) \in E_{CR1}$ and $v_j$ produces $io$ as output. We also add $(v_j, v_i)$ in $E_{CR2}$. \item For each input $io$ of a node $v_i \in V_{CR2}$, if $v_i \in V_{CR1}$ and there exists multiple nodes in the dependency graph $D$ to produce $io$ as output, with probability $p_m$, we perform the following operation: we randomly choose a node $v_j$ from $D$ such that $(v_j, v_i) \notin E_{CR1}$ and $v_j$ produces $io$. We add $v_j$ to $V_{CR2}$ and $(v_j, v_i)$ to $E_{CR2}$. \item For each input $io$ of a node $v_i \in V_{CR2}$, if $v_i \notin V_{CR1}$, we randomly select a node $v_j$ that produces $io$ as output and add $v_j$ to $V_{CR2}$ and $(v_j, v_i)$ to $E_{CR2}$. \item Each node in $V_{CR2}$ is processed only once. In other words, if a node $v_i$ is encountered more than once, we do not reprocess the node. \item The algorithm terminates when there is no node left for processing. \end{itemize} \begin{example} Consider the dependency graph shown in Figure \ref{fig:mutation}(a) and the subgraph of the dependency graph corresponding to a chromosome shown in Figure \ref{fig:mutation}(b). In the figure, for input $io_9$ of the node corresponding to ${\cal{W}}'_8$, the mutation operation is performed. Finally, the off-spring generated using the above procedure is shown in \ref{fig:mutation}(c). \begin{figure}[!htb] \centering \includegraphics[width=\linewidth]{mutation.pdf} \caption{Example of mutation} \label{fig:mutation} \end{figure} \hfill$\blacksquare$ \end{example} \subsection{Fitness function} \noindent In this section, we discuss the computation of the fitness function for each chromosome in the population. For each chromosome in the population, we first compute the QoS tuple. The chromosomes are then divided into different levels based on domination. A chromosome $CR_i$ belongs to level $l_j$, if there exists at-least one chromosome in each level $l_k$, where $k < j$, that dominates $CR_i$ and no chromosome belonging to level $l_m$, where $m \ge j$ dominates $CR_i$. For chromosomes belonging to the same level, we compute the crowding distance $Distance (CR_i)$ as illustrated in \cite{deb2002fast}. Crowding distance ensures solution diversity. This essentially measures how close the individual chromosome is from its neighbors with respect to each QoS parameter. Finally, the chromosomes are ranked as follows: Between two chromosomes $CR_i$ and $CR_j$, rank of $CR_i$ is higher than the rank $CR_j$ if $(Level (CR_i) < Level (CR_j))$ or ($(Level (CR_i) == Level (CR_j))$ and $(Distance (CR_i) > Distance (CR_j))$). The rank of a chromosome is treated as the fitness value of the chromosome. Less fitness value represents better solution quality. When algorithm terminates, all the phenotypes corresponding to the chromosomes with level 0 are returned as the solutions. In the next section, we experimentally shows the comparative study of our approaches. \section{Introduction}\label{sec:intro}} \noindent In recent times, web services have become ubiquitous with the proliferation of Internet usage. A web service is a software component, that takes a set of inputs, performs a specific task and produces a set of outputs. A set of non-functional quality of service (QoS) parameters (e.g., response time, throughput, reliability etc.) are associated with a web service. These QoS parameters determine the performance of a service. Sometimes, a single web service falls short to respond to a user query. Therefore, service composition~\cite{el2010tqos,oh2008effective} is required. During service composition, multiple services are combined in a specific order based on their input-output dependencies to produce a desired set of outputs. While providing a solution in response to a query, it is also necessary to ensure fulfillment of end-to-end QoS requirements~\cite{Zeng:2003:QDW:775152.775211}, which is the main challenge in QoS aware service composition \cite{7226855,bartalos2012automatic,peer2005web}. A large body of literature in service composition deals with optimization of a single QoS parameter \cite{rodriguez2015hybrid,6009375}, especially, response time or throughput. However, a service may have multiple QoS parameters; therefore, the service composition problem turns out to be a multi-objective optimization problem. Though optimality of the end solution is the primary concern in multi-objective service composition, computing the optimal solution is time consuming. This has led to another popular research theme around multi-constrained service composition \cite{Alrifai:2012:HAE:2180861.2180864,1357986}, where a constraint is specified on each QoS parameter and the objective is to satisfy all the QoS constraints in the best possible way. Two different models have been considered in service composition literature, namely, workflow based model (WM) \cite{guidara2015heuristic,qi2010combining} and input-output dependency based model (IOM) \cite{rodriguez2015hybrid}. The salient features of both the models are discussed in Table \ref{tab:features}. Most of the research proposals on multiple-QoS aware service composition have considered WM \cite{ba2016exact,zhang2013selecting}. In general, the methods proposed in WM cannot solve the problem in IOM, since in IOM, in addition to the QoS values of a service, the input-output dependencies between the service also need to be considered. Approaches \cite{feng2007model,hu2005quality}, that consider IOM typically transform the multiple objectives into a single objective and generate the optimal solution instead of the Pareto optimal solutions \cite{6985731}. A weighted sum is used to convert multiple objectives into a single objective. However, finding the weights is a challenging task. \input{features} In this paper, we study the multi-objective QoS-aware web service composition problem in IOM. To the best of our knowledge, there is not much work in IOM considering multiple QoS aware service composition based on Pareto front construction. However, considering the parameters individually instead of a weighted sum combination, has a major significance, since it can deal with the users having various QoS preferences. Additionally, we have considered multiple local and global constraints on different QoS parameters. In this paper, our major contributions are as follows: \iffalse \begin{itemize} \item We first propose an optimal algorithm, that constructs the Pareto optimal solution frontier satisfying all QoS constraints for the multi objective problem in IOM. We theoretically prove the soundness and completeness of our algorithm. \item Additionally, we propose two heuristics. The first one employs a beam search strategy \cite{russell1995modern}, while the other is based on non deterministic sorting genetic algorithm (NSGA) \cite{deb2002fast}. \item To demonstrate the time-quality trade-off, we perform extensive experiments on the benchmarks ICEBE-2005 \cite{icebe2005} and WSC-2009 \cite{bansal2009wsc}. Additionally, we compare our proposed methods with \cite{yan2015anytime}, which proposes the composition problem in IOM using a single objective weight transformation. \end{itemize} \fi \noindent ~\textbullet~ We first propose an optimal algorithm, that constructs the Pareto optimal solution frontier satisfying all QoS constraints for the multi objective problem in IOM. We theoretically prove the soundness and completeness of our algorithm. \noindent ~\textbullet~ Additionally, we propose two heuristics. The first one employs a beam search strategy \cite{russell1995modern}, while the other is based on non deterministic sorting genetic algorithm (NSGA) \cite{deb2002fast}. \noindent ~\textbullet~ To demonstrate the time-quality trade-off, we perform extensive experiments on the benchmarks ICEBE-2005 \cite{icebe2005} and WSC-2009 \cite{bansal2009wsc}. Additionally, we compare our proposed methods with \cite{yan2015anytime}, which proposes the composition problem in IOM using a single objective weight transformation. \noindent This paper is organized as follows. In Section \ref{sec:related}, we compare and contrast our model and proposed approaches with the existing literature. Section \ref{sec:preliminaries} presents some background, the next outlines our problem. Section \ref{sec:method}-\ref{sec:ga} present our proposal, Section \ref{sec:result} presents results. Section~\ref{sec:conclusion} concludes the work. \subsection{Running Example} \noindent We now present an illustrative example for our problem. \begin{example}\label{exm:overview} Table \ref{tab:exampleServices} shows a brief description of the services in a service repository, their inputs, outputs and values of response time (in ms), throughput (number of service invocations per minute) and reliability (in percentage) in the form of a tuple (RT, T, R). \begin{table}[!ht] \tiny \caption{{Description of example services} \centering \begin{tabular}{c|c|c|c} Services & Inputs & Outputs & (RT, T, R)\\ \hline\hline ${\cal{W}}_1$ & $\{i_1\}$ & $\{io_4, io_5\}$ & ${\cal{P}}^{(1)}:$ (500, 7, 93\%)\\ \hline ${\cal{W}}_2$ & $\{i_1\}$ & $\{io_4, io_5\}$ & ${\cal{P}}^{(2)}:$ (600, 13, 69\%)\\ \hline ${\cal{W}}_3$ & $\{i_1\}$ & $\{io_4, io_5\}$ & ${\cal{P}}^{(3)}:$ (350, 4, 97\%)\\ \hline ${\cal{W}}_4$ & $\{i_1\}$ & $\{io_4, io_5\}$ & ${\cal{P}}^{(4)}:$ (475, 3, 85\%)\\ \hline ${\cal{W}}_5$ & $\{i_2\}$ & $\{io_6\}$ & ${\cal{P}}^{(5)}:$ (1300, 15, 81\%)\\ \hline ${\cal{W}}_6$ & $\{i_2\}$ & $\{io_6\}$ & ${\cal{P}}^{(6)}:$ (700, 19, 90\%)\\ \hline ${\cal{W}}_7$ & $\{i_1, i_2\}$ & $\{o_{14}\}$ & ${\cal{P}}^{(7)}:$ (1100, 9, 80\%)\\ \hline ${\cal{W}}_8$ & $\{i_3\}$ & $\{io_4\}$ & ${\cal{P}}^{(8)}:$ (1100, 6, 73\%)\\ \hline ${\cal{W}}_9$ & $\{i_3\}$ & $\{io_4\}$ & ${\cal{P}}^{(9)}:$ (300, 13, 79\%)\\ \hline ${\cal{W}}_{10}$ & $\{i_3\}$ & $\{io_4\}$ & ${\cal{P}}^{(10)}:$ (800, 9, 78\%)\\ \hline ${\cal{W}}_{11}$ & $\{io_4\}$ & $\{io_8, io_9\}$ & ${\cal{P}}^{(11)}:$ (1300, 3, 65\%)\\ \hline ${\cal{W}}_{12}$ & $\{io_4\}$ & $\{io_8, io_9\}$ & ${\cal{P}}^{(12)}:$ (900, 7, 83\%)\\ \hline ${\cal{W}}_{13}$ & $\{io_4\}$ & $\{io_8, io_9\}$ & ${\cal{P}}^{(13)}:$ (400, 9, 93\%)\\ \hline ${\cal{W}}_{14}$ & $\{io_4\}$ & $\{io_8, io_9\}$ & ${\cal{P}}^{(14)}:$ (750, 5, 79\%)\\ \hline ${\cal{W}}_{15}$ & $\{io_5, io_6, io_7\}$ & $\{io_9, io_{10}, io_{11}\}$ & ${\cal{P}}^{(15)}:$ (700, 17, 91\%)\\ \hline ${\cal{W}}_{16}$ & $\{io_5, io_6, io_7\}$ & $\{io_9, io_{10}, io_{11}\}$ & ${\cal{P}}^{(16)}:$ (500, 13, 90\%)\\ \hline ${\cal{W}}_{17}$ & $\{io_8\}$ & $\{o_{12}\}$ & ${\cal{P}}^{(17)}:$ (150, 5, 86\%)\\ \hline ${\cal{W}}_{18}$ & $\{io_8\}$ & $\{o_{12}\}$ & ${\cal{P}}^{(18)}:$ (400, 2, 73\%)\\ \hline ${\cal{W}}_{19}$ & $\{io_8\}$ & $\{o_{12}\}$ & ${\cal{P}}^{(19)}:$ (300, 3, 81\%)\\ \hline ${\cal{W}}_{20}$ & $\{io_9\}$ & $\{o_{13}\}$ & ${\cal{P}}^{(20)}:$ (1500, 12, 94\%)\\ \hline ${\cal{W}}_{21}$ & $\{io_9\}$ & $\{o_{13}\}$ & ${\cal{P}}^{(21)}:$ (900, 14, 97\%)\\ \hline ${\cal{W}}_{22}$ & $\{io_{10}\}$ & $\{o_{12}\}$ & ${\cal{P}}^{(22)}:$ (1700, 14, 87\%)\\ \hline ${\cal{W}}_{23}$ & $\{io_9, io_{10}\}$ & $\{o_{14}\}$ & ${\cal{P}}^{(23)}:$ (1100, 10, 80\%)\\ \hline ${\cal{W}}_{24}$ & $\{io_9, io_{10}\}$ & $\{o_{14}\}$ & ${\cal{P}}^{(24)}:$ (1700, 12, 81\%)\\ \hline ${\cal{W}}_{25}$ & $\{io_{10}\}$ & $\{o_{12}\}$ & ${\cal{P}}^{(25)}:$ (1400, 13, 83\%)\\ \hline ${\cal{W}}_{26}$ & $\{io_{10}\}$ & $\{o_{12}\}$ & ${\cal{P}}^{(26)}:$ (1900, 7, 80\%)\\ \hline ${\cal{W}}_{27}$ & $\{io_{11}\}$ & $\{o_{13}\}$ & ${\cal{P}}^{(27)}:$ (1500, 11, 92\%)\\ \hline ${\cal{W}}_{28}$ & $\{io_{11}\}$ & $\{o_{13}\}$ & ${\cal{P}}^{(28)}:$ (1100, 15, 94\%)\\ \hline ${\cal{W}}_{29}$ & $\{io_{10}, io_{11}\}$ & $\{o_{15}\}$ & ${\cal{P}}^{(29)}:$ (500, 17, 72\%)\\ \hline ${\cal{W}}_{30}$ & $\{io_{10}, io_{11}\}$ & $\{o_{15}\}$ & ${\cal{P}}^{(30)}:$ (350, 12, 74\%)\\ \hline \end{tabular}\label{tab:exampleServices} \end{table} Consider a query with inputs $i_1, i_2, i_3$ and desired outputs $o_{12}, o_{13}$. The objective is to find a solution to the query in such a way that the values of the QoS parameters are optimized (i.e., minimizing response time, maximizing throughput and reliability). It may be noted, a single solution may not be able to optimize all the QoS parameters. Therefore, multiple solutions need to be generated optimizing different QoS parameters. The services that are eventually activated by the query inputs are shown in Figure \ref{fig:activation}. The services at $L_1$ of Figure \ref{fig:activation} are directly activated by the query inputs, while the services at $L_2$ and $L_3$ are indirectly activated by the query inputs. Each ellipse represents the input parameters available in the system at a particular point of time. Additionally, we have the following set of constraints. \begin{itemize} \item ${\cal{LC}}_1:$ Each service participating in the solution must have a reliability value greater than $70\%$. \item ${\cal{GC}}_1:$ The reliability of the solution must be more than $60\%$. \item ${\cal{GC}}_2:$ The response time of the solution must be less than $2.5$s. \end{itemize} \begin{figure}[!htb] \centering \includegraphics[height=4cm,width=\linewidth]{exampleActivation.pdf} \caption{Response to the query} \label{fig:activation} \end{figure} \noindent In this paper, we demonstrate the Pareto optimal solutions construction method given the above scenario using this example. \hfill$\blacksquare$ \end{example} \section{Pareto Front Construction} \noindent \noindent To find the feasible Pareto optimal solutions, we transform the dependency graph into a layered path generation graph (LPG). LPG $G_P = (V_P, E_P)$ is a directed acyclic graph, where $V_p$ is a set of nodes and $E_p$ is a set of edges. Each node $v^{(p)}_i \in V_p$ consists of a set of nodes $\{v_{i_1}, v_{i_2}, \ldots, v_{i_l}\} \subseteq V$ of ${\cal{D}}$. A directed edge from $v^{(p)}_i$ to $v^{(p)}_j$ exists, if each service corresponding to a node $v_{j_k}$, belonging to $v^{(p)}_j$ is activated by the outputs of the services corresponding to the nodes, belonging to $v^{(p)}_i$. Similar to the dependency graph, the LPG also consists of two dummy nodes: a start node $V^{(p)}_{Start}$ and an end node $V^{(p)}_{End}$ consisting of the start node and the end node of ${\cal{D}}$ respectively. We assume that each dummy node has a QoS tuple with the best value for each QoS parameter. While constructing $G_P$, we simultaneously compute the Pareto optimal solution frontier and validate the global constraints. We define the notion of a cumulative Pareto optimal tuple. \begin{definition}{\em [\bf Cumulative Pareto Optimal Tuple:]} A set of non dominated QoS tuples, generated due to the composition of a set of services during an intermediate step of the solution construction, is called a cumulative Pareto optimal tuple. \hfill$\blacksquare$ \end{definition} \noindent The cumulative Pareto optimal tuples, generated at the final step of the solution construction, is the Pareto front. For each node $v^{(p)}_i \in V_P$, we maintain two sets of QoS tuples: a set of non dominated tuples and a set of cumulative Pareto optimal tuples. We now discuss the construction of $G_P$. \begin{algorithm} \scriptsize \caption{Graph Conversion and Solution Generation} \begin{algorithmic}[1] \State {Input: ${\cal{D}} = (V, E)$, ${\cal{GC}}$} \State {Output: $G_P = (V_P, E_P)$} \State Queue $Q =$ Insert$(V^{(P)}_{End})$; \Repeat \State $v^{p}_i = $ Remove($Q$);\Comment{$v_i$ corresponding to ${\cal{W}}'_i$} \State ${\cal{PW}} \leftarrow$ the set predecessor nodes of $v^{p}_i$;\label{step:predecessor} \For {$v^{p}_j \in {\cal{PW}}$} \If {$v^{p}_j$ is not constructed earlier} \State ${\cal{P}}^{''(j)} \leftarrow$ the set of non dominated tuples corresponding to $v^{p}_j$; \If {$p \in {\cal{P}}^{''(j)}$ does not satisfy any global constraints} \State Remove $p$ from ${\cal{P}}^{''(j)}$; \EndIf \State {\bf{if}} {${\cal{P}}^{''(j)} = \phi$}, {\bf{then}} continue; \State $Q =$ Insert$(v^{p}_j)$; \EndIf \State Construct an edge $(v^{p}_j, v^{p}_i)$; \State ${\cal{CP}}^{(j)} \leftarrow$ The cumulative Pareto front of $v^{p}_j$ is constructed from \indent\indent \indent (${\cal{CP}}^{(j)} \cup $ Combination of (${\cal{P}}^{''(j)}, {\cal{CP}}^{(i)}$)); \If {$p \in {\cal{CP}}^{(j)}$ does not satisfy any global constraints}\label{step:constraintCheckStart} \State Remove $p$ from ${\cal{CP}}^{(j)}$; \EndIf\label{step:constraintCheckEnd} \EndFor \Until{$(Q \ne \phi)$} \end{algorithmic} \label{algo:graphConversion} \end{algorithm} \begin{figure}[!htb] \centering \includegraphics[width=\linewidth]{exampleConversion.pdf} \caption{Conversion of Dependency Graph to LPG} \label{fig:conversion} \end{figure} To construct $G_P$, we traverse the graph ${\cal{D}}$ in a backward direction, starting from the node $V_{End}$. We start the transformation from dependency graph to LPG by constructing a dummy node $V^{(p)}_{End}$ of $G_P$ consisting of $V_{End}$ of ${\cal{D}}$. During the procedure, we maintain a FIFO (i.e., First In First Out) queue. The following steps convert ${\cal{D}}$ to $G_P$: \begin{figure*}[!htb] \centering \includegraphics[height=4cm,width=\linewidth]{exampleQoSComputation.pdf} \caption{Pareto Optimal Front Construction} \label{fig:pareto} \end{figure*} \begin{itemize} \item The first node $v^{{p}}_i$ is removed from the queue. \item The set of predecessor nodes of $v^{{p}}_i$ is constructed. \item For each predecessor node $v^{{p}}_j$ of $v^{{p}}_i$, the temporary Pareto optimal front till $v^{{p}}_j$ is constructed or modified (for already existing nodes). \item For each Pareto optimal QoS tuple till $v^{{p}}_j$, the global constraints are validated. If any global constraint is violated, the tuple from the Pareto front is removed. \item Each predecessor node is inserted in the queue, if the queue does not already hold the same. \end{itemize} \noindent We briefly elaborate each step below. We first insert $V^{(p)}_{End}$ in the queue and then continue the procedure until the queue becomes empty. In each step, we remove a node from the queue, say $v^{{p}}_i$ (in FIFO basis) and construct its predecessor nodes as described below. Consider a node $v^{(p)}_i \in V_P$ consisting of a set of nodes $\{v_{i_1}, v_{i_2}, \ldots, v_{i_l}\} \subseteq V$ of ${\cal{D}}$. Also, consider $io_1, io_2, \ldots, io_x$ be the set of inputs that are required to activate the services corresponding to $\{v_{i_1}, v_{i_2}, \ldots, v_{i_l}\}$. For each $io_j \in \{io_1, io_2, \ldots, io_x\}$, we compute a set of nodes $V^{(io_j)} \subset V$, such that an edge $e \in E$ annotated by $io$ is incident to at least one node in $\{v_{i_1}, v_{i_2}, \ldots, v_{i_l}\}$. We then compute a set of combinations of nodes in ${\cal{D}}$ consisting of a node from each $V^{(io_j)}$, for all $io_j \in \{io_1, io_2, \ldots, io_x\}$. We now define the notion of a redundant service. \begin{definition}{\em [\bf Redundant Service:]} A service ${\cal{W}}_k$ belonging to a solution $S_i$ in response to a query ${\cal{Q}}$ is redundant, if $S_i \setminus \{{\cal{W}}_k\}$ is also a solution to ${\cal{Q}}$. \hfill$\blacksquare$ \end{definition} \noindent We consider the following assumption: a solution $S_i$ with some redundant services $\{{\cal{W}}_{i_1}, {\cal{W}}_{i_2}, \ldots, {\cal{W}}_{i_k}\}$ cannot be better, in terms of QoS values, than $S_i \setminus \{{\cal{W}}_{i_1}, {\cal{W}}_{i_2}, \ldots, {\cal{W}}_{i_k}\}$. Two sets $V^{(io_j)}$ and $V^{(io_k)}$ may not be mutually exclusive, for $io_j, io_k \in \{io_1, io_2, \ldots, io_x\}$, since the service corresponding to one node may produce more than one output from $\{io_1, io_2, \ldots, io_x\}$. Therefore, if we consider a combination $c$ consisting of one node from each $V^{(io_j)}$, where $io_j \in \{io_1, io_2, \ldots, io_x\}$ , we do not need to consider any combination which is a superset of $c$. For each combination, we construct a node $v^{(p)}_y$ and an edge $(v^{(p)}_y, v^{(p)}_i)$ of $G_P$. \begin{example} Fig. \ref{fig:conversion}(b) shows the LPG generated from the dependency graph in Fig. \ref{fig:conversion}(a). Consider the end node $V^{(p)}_{End}$ of Fig.\ref{fig:conversion}(b). $V^{(p)}_{End}$ consists of $V_{End}$. $\{o_{12}, o_{13}\}$ is the required set of inputs (i.e., query outputs). $V^{(o_{12})} = \{{\cal{W}}'_7, {\cal{W}}'_9\}$ and $V^{(o_{13})} = \{{\cal{W}}'_8, {\cal{W}}'_{11}\}$. We get 4 combinations from $V^{(o_{12})}, V^{(o_{13})}$ and construct a node for each combination and the corresponding edges. Consider another node of $G_P$ consisting of $\{{\cal{W}}'_7, {\cal{W}}'_8\}$. The required set of inputs is $\{io_8, io_9\}$. $V^{(io_{8})} = \{{\cal{W}}'_5\}$ and $V^{(io_9)} = \{{\cal{W}}'_5, {\cal{W}}'_6\}$. We get 2 combinations from $V^{(io_8)}, V^{(io_9)}$. However, one combination $\{{\cal{W}}'_5, {\cal{W}}'_6\}$ is a superset of another $\{{\cal{W}}'_5\}$. Therefore, we disregard the combination $\{{\cal{W}}'_5, {\cal{W}}'_6\}$ and construct a node corresponding to $\{{\cal{W}}'_5\}$ and the corresponding edge. \hfill$\blacksquare$ \end{example} \noindent We now prove the following lemma. \begin{lemma} Each path from $V^{(p)}_{Start}$ to $V^{(p)}_{End}$ in $G_P$ represents a solution to the query in terms of functional dependencies. \hfill$\blacksquare$ \end{lemma} \iffalse \begin{proof} {\small{ The proof of the lemma follows from the construction of $G_P$. Consider a path $\rho: V^{(p)}_{Start}-v^{(p)}_{i_1}-v^{(p)}_{i_2}-\ldots-v^{(p)}_{i_k}-V^{(p)}_{End}$ of $G_P$. The services corresponding to $v^{(p)}_{i_1}$ are directly activated by the query inputs, while the services corresponding to $v^{(p)}_{i_2}$ are activated by outputs of the services corresponding to $v^{(p)}_{i_1}$. Similarly, the services corresponding to $v^{(p)}_{i_j}$ for $j= 2, 3 \ldots, k$ are activated by the outputs of the services corresponding to $v^{(p)}_{i_{(j - 1)}}$. Finally, the services corresponding to $v^{(p)}_{i_k}$ produce the query outputs. Therefore, the services associated with the path $\rho$ forms a composition solution in terms of functional dependency.}} \end{proof} \fi \noindent Once a node of $G_P$ is constructed, we construct the set of Pareto optimal tuples corresponding to the node. Consider a node $v^{(p)}_i \in V_P$ consisting of a set of nodes $\{v_{i_1}, v_{i_2}, \ldots, v_{i_l}\} \subseteq V$ of ${\cal{D}}$. The QoS tuples corresponding to $\{v_{i_1}, v_{i_2}, \ldots, v_{i_l}\}$ are combined and a new set of tuples, ${\cal{P}}^{''(i)}$, is constructed. Each tuple in ${\cal{P}}^{''(i)}$ is then validated against the set of global constraints ${\cal{GC}}$. If any tuple violates any of the global constraints, the tuple is removed from ${\cal{P}}^{''(i)}$. If no tuple from ${\cal{P}}^{''(i)}$ satisfies the global constraints, we disregard the node $v^{(p)}_i$. Otherwise, we compute the set of non dominated tuples from ${\cal{P}}^{''(i)}$ and associate these with $v^{(p)}_i$. Consider $v^{(p)}_j$ is removed from the queue and $v^{(p)}_i$ is created as the predecessor of $v^{(p)}_j$. If $v^{(p)}_i$ already exists in the queue, we do not need to recompute the set of non dominated tuples of $v^{(p)}_i$. Once the set of non dominated tuples corresponding to $v^{(p)}_i$ are constructed, we construct the cumulative Pareto optimal solutions till $v^{(p)}_i$. In order to find the Pareto front till $v^{(p)}_i$, we combine the tuples in the Pareto front constructed till $v^{(p)}_j$ with the set of non dominated tuples of $v^{(p)}_i$. The combined tuples are verified against the global constraints and if any tuple violates any of the global constraints, we remove the tuple from the combined set. Finally, we compute the cumulative Pareto optimal solutions till $v^{(p)}_i$ from the set of combined tuples and the cumulative Pareto front of $v^{(p)}_i$. The Pareto front constructed in $V^{(p)}_{Start}$ constitutes the feasible Pareto optimal solutions to the query. \begin{example} Fig. \ref{fig:pareto} shows the Feasible Pareto front generation method on a LPG. The set of initial non dominated tuples of $V^{(p)}_{End}$ consists of one tuple $(0, \infty, \infty)$, initialized with the best values of these parameters. The cumulative Pareto optimal front of $V^{(p)}_{End}$ also consists of the same tuple. Now consider a node $v^{(p)}_5$. When $v^{(p)}_5$ is created as a predecessor of $v^{(p)}_8$, the set of non dominated tuples ${\cal{P}}^{''(5)}$ corresponding to $v^{(p)}_5$ are constructed first. The cumulative Pareto front ${\cal{CP}}^{(5)}$ till $v^{(p)}_5$ is constructed next by combining the cumulative Pareto front till $v^{(p)}_8$ and the set of non dominated tuples of $v^{(p)}_5$, followed by selecting the Pareto front from the combined set. In the next iteration, when $v^{(p)}_5$ is constructed as a predecessor of $v^{(p)}_9$, the set of non dominated tuples are not recomputed. However, ${\cal{CP}}^{(5)}$ is modified. The cumulative Pareto front till $v^{(p)}_9$ and the set of non dominated tuples of $v^{(p)}_5$ are combined first and then the Pareto front is selected from the combined set and the already existing set ${\cal{CP}}^{(5)}$. It may be noted, the cumulative Pareto front till $v^{(p)}_2$ violates the global constraints. Hence, the node is disregarded from the graph. The final solution path is marked by the bold line. \hfill$\blacksquare$ \end{example} \noindent Algorithm \ref{algo:graphConversion} presents the formal algorithm for constructing the feasible Pareto optimal solution in response to a query. We now prove the following lemma. \begin{lemma}\label{lemma:complete} Algorithm \ref{algo:graphConversion} is complete. \hfill$\blacksquare$ \end{lemma} \iffalse \begin{proof} {\small{ In order to prove the completeness of Algorithm \ref{algo:graphConversion}, we show Algorithm \ref{algo:graphConversion} generates all distinct feasible solutions in terms of QoS values belonging to the Pareto front in response to a query ${\cal{Q}}$. We prove the above lemma by contradiction. We first assume that a unique (in terms of QoS values) feasible solution $S_i$ belonging to the Pareto front is not generated by Algorithm \ref{algo:graphConversion}. This implies, one of the following condition holds: \begin{itemize} \item Case 1: $S_i$ cannot be generated through any path of $G_P$, even if no tuple or node is removed during solution construction. \item Case 2: $S_i$ is removed while constructing the set of non dominated tuples corresponding to a node in $G_P$. \item Case 3: $S_i$ is removed while constructing the cumulative Pareto optimal solution till a node $v^{(p)}_i \in G_P$. \end{itemize} Since we have already proved in Lemma \ref{lemma:pareto} that the preprocessing step is Pareto optimal solution preserving in terms of QoS values, we can now assume that $S_i$ consists of the services in $W'$. We now analyze each case separately below. Consider Case 1. Below, we prove that Case 1 cannot be true by the following argument. While constructing the set of predecessor nodes of a node $v^{(p)}_i \in G_P$, we consider all possible combinations of nodes in ${\cal{D}}$ which can activate the nodes corresponding to $v^{(p)}_i$. Consider the subgraph ${\cal{D}}_{Sub}$ of ${\cal{D}}$ corresponding to the solution $S_i$ and the partition ${\cal{D}}_{Sub}$ into multiple layers as defined by Equations \ref{equ:layerGeneral} and \ref{equ:layerBasic}. The last layer (say $L_k$) of ${\cal{D}}_{Sub}$ consists of $V_{End} \in {\cal{D}}$. Consider each layer $L_i$, for $i = 0, 1, \ldots, k$ in ${\cal{D}}_{Sub}$ consists of a set of nodes $V_i \subset V$. It may be noted, the set of services corresponding to $V_{(k-1)}$ in $L_{(k - 1)}$ produce the query outputs. $V^{(p)}_{End} \in G_P$ consists of $V_{End} \in {\cal{D}}$. While constructing the set of predecessor nodes of $V^{(p)}_{End}$, since all possible combinations of nodes in ${\cal{D}}$ which can activate $V_{End}$ are considered, $V_{(k-1)}$ is also considered. Therefore, a node $v^{(p)}_{i_1} \in G_P$ must correspond to $V_{(k-1)}$. In general, $v^{(p)}_{i_j} \in G_P$ corresponds to $V_{(k-j)}$. Since all possible combinations of nodes in ${\cal{D}}$ which can activate $V_{(k-j)}$ are considered while constructing the set of predecessor nodes of $V_{(k-j)}$, $V_{(k - j - 1)}$ is also considered. Hence, a path corresponding to ${\cal{D}}_{Sub}$ belongs to $G_P$, which contradicts our assumption. We now consider Case 2. Consider $v^{(p)}_i \in G_P$ consists of $\{v_{i_1}, v_{i_2}, \ldots, v_{i_k}\} \in V$ of ${\cal{D}}$. The services corresponding to $\{v_{i_1}, v_{i_2}, \ldots, v_{i_k}\}$ are executed in parallel. It may be noted, each service corresponding to $v_{i_j} \in \{v_{i_1}, v_{i_2}, \ldots, v_{i_k}\}$ may be associated with more than one QoS tuple. Therefore, after composition, more than one tuple is generated. We consider only the set of non dominated tuples {\small{${\cal{TP}}^*$}} from the set of generated tuples {\small{${\cal{TP}}$}}. If $S_i$ is removed due to removal of {\small{$({\cal{TP}} \setminus {\cal{TP}}^*)$}}, this implies $S_i$ consists of one QoS tuple {\small{$t \in ({\cal{TP}} \setminus {\cal{TP}}^*)$}}. This further implies, another tuple $t' \in$ {\small{${\cal{TP}}^*$}} must dominate $t$. By a similar argument, as in the proof of Lemma \ref{lemma:pareto}, we can say another solution $S_j$ generated by Algorithm \ref{algo:graphConversion} either has exactly the same QoS tuple as $S_i$, which implies $S_i$ is not unique, or $S_j$ dominates $S_i$, which contradicts our assumption. The argument of Case 3 is same as in Case 2. Therefore, none of the above cases hold, which contradicts our assumption. Thus Algorithm \ref{algo:graphConversion} is complete.}} \end{proof} \fi \begin{figure*}[!htb] \centering \includegraphics[height=3.5cm,width=\linewidth]{exampleQoSComputationHeuristic.pdf} \caption{Heuristic solution Construction} \label{fig:heuristic} \end{figure*} \begin{lemma}\label{lemma:sound} Algorithm \ref{algo:graphConversion} is sound. \hfill$\blacksquare$ \end{lemma} \iffalse \begin{proof} {\small{ In order to prove the soundness of Algorithm \ref{algo:graphConversion}, we need to show each solution $S_i$, generated by Algorithm \ref{algo:graphConversion} in response to a query ${\cal{Q}}$, is a feasible solution and belongs to the Pareto front. If $S_i$ is generated by Algorithm \ref{algo:graphConversion}, it must be a feasible solution, as it is ensured by Steps \ref{step:constraintCheckStart}-\ref{step:constraintCheckEnd} of Algorithm \ref{algo:graphConversion}. We now prove by contradiction that $S_i$ belongs to the Pareto front. We first assume $S_i$ does not belong to the Pareto front. This implies, there exists another solution $S_j$ which dominates $S_i$. However, being complete, Algorithm \ref{algo:graphConversion} generates all the feasible solutions belonging to the Pareto front. Therefore, Algorithm \ref{algo:graphConversion} generates $S_j$ as one of the solutions. Therefore, $S_i$ cannot be generated by Algorithm \ref{algo:graphConversion}, since $S_j$ dominates $S_i$, which contradicts our assumption. Thus, Algorithm \ref{algo:graphConversion} is sound.}} \end{proof} \fi \noindent The search space of this algorithm is exponential in terms of the number of services required to serve a query. This limits its scalability to large service repositories. In the next subsection, we propose two scalable heuristics. \section{Background and Problem Formulation}\label{sec:preliminaries} \noindent In this section, we discuss some background concepts for our work. We begin with a classification of a QoS parameter. \begin{definition}{\em [\bf Positive / Negative QoS parameter:]} A QoS parameter is called a positive (negative) QoS parameter, if a higher (lower) value of the parameter implies better performance. \hfill$\blacksquare$ \end{definition} \noindent Reliability, availability, throughput are examples of positive QoS parameters, while response time, latency are examples of negative QoS parameters. Consider two services ${\cal{W}}_i$ and ${\cal{W}}_j$ being compared with respect to a QoS parameter ${\cal{P}}_k$. We have the following cases: \begin{itemize} \item ${\cal{W}}_i$ is {\em{better}} than ${\cal{W}}_j$ with respect to ${\cal{P}}_k$ implies, \begin{itemize} \item If ${\cal{P}}_k$ is a positive QoS, ${\cal{P}}^{(i)}_k > {\cal{P}}^{(j)}_k$, where ${\cal{P}}^{(i)}_k$ and ${\cal{P}}^{(j)}_k$ are the respective values of ${\cal{P}}_k$ for ${\cal{W}}_i$ and ${\cal{W}}_j$. \item If ${\cal{P}}_k$ is a negative QoS parameter, ${\cal{P}}^{(i)}_k < {\cal{P}}^{(j)}_k$. \end{itemize} \item ${\cal{W}}_i$ is {\em{as good as}} ${\cal{W}}_j$ with respect to ${\cal{P}}_k$ implies, ${\cal{P}}^{(i)}_k = {\cal{P}}^{(j)}_k$, irrespective of whether ${\cal{P}}_k$ is positive or negative. \item ${\cal{W}}_i$ is {\em{at least as good as}} ${\cal{W}}_j$ with respect to ${\cal{P}}_k$ implies, either ${\cal{W}}_i$ is {\em{better}} than ${\cal{W}}_j$ or ${\cal{W}}_i$ is {\em{as good as}} ${\cal{W}}_j$ with respect to ${\cal{P}}_k$. \end{itemize} \noindent The QoS parameters are further classified into four categories based on the aggregate functions used for composition: maximum, minimum, addition, multiplication. A query is specified in terms of a set of input-output parameters. We now present the concept of eventual activation of a web service for a given query. A web service is {\em activated}, when the set of inputs of the service is available in the system. As an example, consider ${\cal{W}}_1$ in Table \ref{tab:exampleServices}, ${\cal{W}}_1$ is activated when its input $i_1$ is available. A service ${\cal{W}}_i$ is {\em eventually activated} by a set of input parameters $I$, if ${\cal{W}}_i$ is either directly activated by $I$ itself or indirectly activated by the outputs of the set of services that are eventually activated by $I$, as shown in Example~\ref{exm:overview}. In the next subsection, we formally discuss the model considered in this paper and our objective. \subsection{Problem Formulation}\label{subsec:problemformulation} \noindent The service composition problem considered in this paper can be formally described as below: \begin{itemize} \item A set of web services $W = \{{\cal{W}}_1, {\cal{W}}_2, \ldots, {\cal{W}}_n\}$ \item For each service ${\cal{W}}_i \in W$, a set of inputs ${\cal{W}}_i^{ip}$ and a set of outputs ${\cal{W}}_i^{op}$ \item A set of QoS parameters ${\cal{P}} = \{{\cal{P}}_1, {\cal{P}}_2, \ldots, {\cal{P}}_m\}$ \item For each service ${\cal{W}}_i \in W$, a tuple of QoS values ${\cal{P}}^{(i)} = ({\cal{P}}^{(i)}_1, {\cal{P}}^{(i)}_2, \ldots, {\cal{P}}^{(i)}_m)$ \item A set of aggregation functions ${\cal{F}} = \{f_1, f_2, \ldots, f_m\}$, where $f_i$ is defined for a QoS parameter ${\cal{P}}_i \in {\cal{P}}$ \item A query ${\cal{Q}}$, specified by a set of inputs ${\cal{Q}}^{ip}$ and a set of requested outputs ${\cal{Q}}^{op}$ \item Optionally, a set of local QoS constraints ${\cal{LC}} = \{{\cal{LC}}_1, {\cal{LC}}_2, \ldots, {\cal{LC}}_{k_l}\}$ and a set of global QoS constraints ${\cal{GC}} = \{{\cal{GC}}_1, {\cal{GC}}_2, \ldots, {\cal{GC}}_{k_g}\}$ \end{itemize} \noindent A constraint denotes a bound on the worst case value of a QoS parameter. While the {\em{local constraints}} are applicable on a single service (${\cal{LC}}_1$ in Example \ref{exm:overview}), the {\em{global constraints}} are applicable on a composition solution, (${\cal{GC}}_1$ in Example \ref{exm:overview}). The objective of multi-objective QoS constrained service composition is to serve ${\cal{Q}}$ in a way such that the QoS values are optimized, while ensuring functional dependencies are preserved, and all local and global QoS constraints are satisfied. Since multiple (and often disparate) QoS parameters are involved, this calls for a classical multi-objective optimization, and we address this challenge in this work. In this paper, we propose an optimal solution construction methodology. Often, a single solution may not be the best with respect to all the QoS parameters. Therefore, instead of producing a single solution, our method generates a set of Pareto optimal solutions as described in the following sections. \subsection{Preprocessing phase} \noindent The motivation behind preprocessing the web services is to reduce the run-time computation. We first define the notion of equivalent services, which serve as the foundation. \begin{definition}{\em [\bf Equivalent Services:]}\label{def:equivalent} Two services ${\cal{W}}_i$ and ${\cal{W}}_j$ are equivalent (${\cal{W}}_i \simeq {\cal{W}}_j$), if the inputs of ${\cal{W}}_i$ are same as in ${\cal{W}}_j$, and the outputs of ${\cal{W}}_i$ are same as in ${\cal{W}}_j$. \hfill$\blacksquare$ \end{definition} \begin{example} Consider the first two services of Table \ref{tab:exampleServices}: ${\cal{W}}_1$ and ${\cal{W}}_2$ with input $\{i_1\}$ and outputs $\{io_4, io_5\}$. Here, ${\cal{W}}_1$ and ${\cal{W}}_2$ are equivalent, since, they have identical inputs and outputs. \hfill$\blacksquare$ \end{example} \noindent Here, we apply the clustering technique proposed in \cite{7933195}. As the first step of preprocessing, we compute the set of equivalent services. Each equivalence class forms a cluster, while the set of equivalence classes of a given set of web services forms a partition of $W$. Therefore, the clusters are mutually exclusive and collectively exhaustive. We first divide the services in the service repository into multiple clusters and represent each cluster by a single representative service. Once the services are clustered, we find the skyline service set for each cluster. The set of skyline services is used for our run-time service composition step. The representative service corresponding to each cluster is associated with multiple QoS tuples corresponding to each service of the skyline service set. The main aim of this step is to prune the search space. Since the number of clusters must be less than or equal to the number of services in the service repository, the number of services reduces by this preprocessing. After preprocessing, we now have the following set of services: $W' = \{{\cal{W}}'_1, {\cal{W}}'_2, \ldots, {\cal{W}}'_{n'}\}$, where $n' \le n$ and each service ${\cal{W}}'_i \in W'$ consists of a set of QoS tuples ${\cal{P}}^{'(i)}$. Equality condition holds, only when the service repository does not contain any equivalent services and the preprocessing phase cannot reduce the number of services. \begin{example} If we cluster the services shown in Table \ref{tab:exampleServices}, the number of services reduces from 30 to 12. Table \ref{tab:exampleModifiedServices} shows the clustered service set. The first column of Table \ref{tab:exampleModifiedServices} presents the representative service for each cluster, while the second column shows the cluster itself. Finally, the third column indicates the set of QoS tuples corresponding to each service of the skyline services corresponding to a cluster. Consider the first cluster $C_1: \{{\cal{W}}_1, {\cal{W}}_2, {\cal{W}}_3, {\cal{W}}_4\}$, shown in the first row of Table \ref{tab:exampleModifiedServices}. ${\cal{W}}'_1$ is the representative service corresponding to $C_1$. The skyline service set of $C_1$ is $C^{*}_1: \{{\cal{W}}_1, {\cal{W}}_2, {\cal{W}}_3\}$. Therefore, ${\cal{P}}^{'(1)}$ consists of the QoS tuple corresponding to each service in $C^{*}_1$. \begin{table}[!ht] \tiny \caption{\scriptsize{Description of services after preprocessing}} \centering \begin{tabular}{c|c|c} Representative & Cluster & (RT, T, R)\\ Web Service & & \\ \hline\hline ${\cal{W}}'_1$ & $\{{\cal{W}}_1, {\cal{W}}_2, {\cal{W}}_3, {\cal{W}}_4\}$ & (500, 7, 93\%), (350, 4, 97\%), (600, 13, 69\%)\\ \hline ${\cal{W}}'_2$ & $\{{\cal{W}}_5, {\cal{W}}_6\}$ & (700, 19, 90\%)\\ \hline ${\cal{W}}'_3$ & $\{{\cal{W}}_7\}$ & (1100, 9, 80\%)\\ \hline ${\cal{W}}'_4$ & $\{{\cal{W}}_8, {\cal{W}}_9, {\cal{W}}_{10}\}$ & (300, 13, 79\%)\\ \hline ${\cal{W}}'_5$ & $\{{\cal{W}}_{11}, {\cal{W}}_{12}, {\cal{W}}_{13}, {\cal{W}}_{14}\}$ & (400, 9, 93\%)\\ \hline ${\cal{W}}'_6$ & $\{{\cal{W}}_{15}, {\cal{W}}_{16}\}$ & (700, 17, 91\%), (500, 13, 90\%)\\ \hline ${\cal{W}}'_7$ & $\{{\cal{W}}_{17}, {\cal{W}}_{18}, {\cal{W}}_{19}\}$ & (150, 5, 86\%)\\ \hline ${\cal{W}}'_8$ & $\{{\cal{W}}_{20}, {\cal{W}}_{21}\}$ & (900, 14, 97\%)\\ \hline ${\cal{W}}'_9$ & $\{{\cal{W}}_{22}, {\cal{W}}_{25}, {\cal{W}}_{26}\}$ & (1700, 14, 87\%), (1400, 13, 83\%)\\ \hline ${\cal{W}}'_{10}$ & $\{{\cal{W}}_{23}, {\cal{W}}_{24}\}$ & (1100, 10, 80\%), (1700, 12, 81\%)\\ \hline ${\cal{W}}'_{11}$ & $\{{\cal{W}}_{27}, {\cal{W}}_{28}\}$ & (1100, 15, 94\%)\\ \hline ${\cal{W}}'_{12}$ & $\{{\cal{W}}_{29}, {\cal{W}}_{30}\}$ & (500, 17, 72\%), (350, 12, 74\%)\\ \end{tabular}\label{tab:exampleModifiedServices} \end{table} \noindent Consider the query in Example \ref{exm:overview}. The number of services reduces from 30 to 12. The number of QoS tuples reduces from 30 to 18.\hfill$\blacksquare$ \end{example} \noindent The preprocessing step helps to prune the search space by removing some services. No useful solution in terms of QoS values is lost in preprocessing, as stated formally below. \begin{lemma}\label{lemma:pareto} The preprocessing step is Pareto optimal solution preserving in terms of QoS values. \hfill$\blacksquare$ \end{lemma} \noindent All proofs are compiled in Appendix. \iffalse \begin{proof} {\small{ We prove the above lemma by contradiction. Consider a solution $S_i$ to a query ${\cal{Q}}$ is lost due to preprocessing. We show either of the following two conditions must hold: \begin{itemize} \item A solution $S_j$ to the query ${\cal{Q}}$, generated after preprocessing, dominates $S_i$. \item A solution $S_j$ to the query ${\cal{Q}}$, generated after preprocessing, has exactly the same QoS tuple as $S_i$. \end{itemize} \noindent We first assume that none of the above conditions hold for $S_i$. Since $S_i$ is lost due to preprocessing, there must exist a solution consisting of at least one service ${\cal{W}}_k$, which is removed during preprocessing. This further implies, presence of another service ${\cal{W}}_l$, belonging to the same cluster as ${\cal{W}}_k$, such that ${\cal{W}}_l$ dominates ${\cal{W}}_k$. We now analyze different cases for different aggregation functions for different QoS parameters. Consider a QoS parameter ${\cal{P}}_x$ with aggregation function $f_x$. If ${\cal{P}}_x$ is a positive QoS, ${\cal{P}}^{(l)}_x \ge {\cal{P}}^{(k)}_x$. If ${\cal{P}}^{(l)}_x = {\cal{P}}^{(k)}_x$, $f_x(P^*, {\cal{P}}^{(l)}_x) = f_x(P^*, {\cal{P}}^{(k)}_x)$, where $P^*$ denotes the values of ${\cal{P}}_k$ for the set of services with which ${\cal{P}}^{(k)}_x$ is combined in $S_i$. We, therefore, consider the cases where ${\cal{P}}^{(l)}_x > {\cal{P}}^{(k)}_x$. \begin{itemize} \item $f_x$ is addition: $Sum(P^*, {\cal{P}}^{(l)}_x) > Sum(P^*, {\cal{P}}^{(k)}_x)$. \item $f_x$ is product: $Prod(P^*, {\cal{P}}^{(l)}_x) > Prod(P^*, {\cal{P}}^{(k)}_x)$. \item $f_x$ is maximum: \begin{itemize} \item If $Max(P^*, {\cal{P}}^{(k)}_x) = {\cal{P}}^{(k)}_x$, \\ $Max(P^*, {\cal{P}}^{(l)}_x) > Max(P^*, {\cal{P}}^{(k)}_x)$. \item If ${\cal{P}}^{(k)}_x < Max(P^*, {\cal{P}}^{(k)}_x) < {\cal{P}}^{(l)}_x$, \\ $Max(P^*, {\cal{P}}^{(l)}_x) > Max(P^*, {\cal{P}}^{(k)}_x)$. \item If {\small{${\cal{P}}^{(k)}_x < Max(P^*, {\cal{P}}^{(k)}_x)$ and ${\cal{P}}^{(l)}_x < Max(P^*, {\cal{P}}^{(k)}_x)$}}, \\ $Max(P^*, {\cal{P}}^{(l)}_x) = Max(P^*, {\cal{P}}^{(k)}_x)$. \end{itemize} Therefore, $Max(P^*, {\cal{P}}^{(l)}_x) \ge Max(P^*, {\cal{P}}^{(k)}_x)$. \item $f_x$ is minimum: \begin{itemize} \item If {\small{$Min(P^*, {\cal{P}}^{(k)}_x) = {\cal{P}}^{(k)}_x$ and $Min(P^*) > {\cal{P}}^{(k)}_x$}},\\ $Min(P^*, {\cal{P}}^{(l)}_x) > Min(P^*, {\cal{P}}^{(k)}_x)$. \item If $Min(P^*) \le {\cal{P}}^{(k)}_x$,\\ $Min(P^*, {\cal{P}}^{(l)}_x) = Min(P^*, {\cal{P}}^{(k)}_x)$. \end{itemize} Therefore, $Min(P^*, {\cal{P}}^{(l)}_x) \ge Min(P^*, {\cal{P}}^{(k)}_x)$. \end{itemize} Using a similar argument, it can be shown that the above lemma also holds for any negative QoS parameter. Therefore, the solution consisting of ${\cal{W}}_l$ is {\em{at least as good as}} the solution consisting of ${\cal{W}}_k$, which contradicts our assumption.}} \end{proof} \fi \section{Related Work}\label{sec:related} \noindent Automatic service composition \cite{7027801,hwang2008dynamic} is a fundamental problem in services computing. A significant body of research has been carried out on QoS-aware service composition considering a single QoS parameter, especially, response time and throughput \cite{chattopadhyay2015scalable,xia2013web,chen2014qos}. Multiple-QoS aware service composition has been discussed in \cite{mostafa2015multi}. We first discuss related work regarding the models followed by the solution approaches for multiple QoS aware web service composition. \subsection{Problem Models}\label{subsec:problemmodel} \noindent The two most popular models considered in literature are the workflow model (WM) and input-output dependency based model (IOM). The salient features of these two models are discussed in Table \ref{tab:features}. In WM, it is assumed that a task can be accomplished by a single web service. However, in practice, it may not be the case always. Some times more than one service may be required to perform a particular task. Therefore, the input-output dependency based model becomes popular. It may be noted, existing solution approaches for WM are unable to solve the composition problem in IOM. This is mainly because of the following reasons: \begin{itemize} \item In WM, a workflow is provided as an input, whereas, in IOM, no workflow is provided, rather the aim of IOM is to find out a flow of services to serve the query so that the overall QoS values are optimized. \item In general, while selecting a service in WM, only the QoS values of the service need to be taken care of. In contrast to the former case, while selecting a service in IOM, not only the QoS values of the services need to be considered but also its input-output dependencies on the other services need to be taken into account. \item In WM, the number of services does not vary across all solutions, while, in IOM, the number of services varies across the solutions to a query. \end{itemize} However, methods that can solve the composition problem in IOM can solve the composition problem in WM. Moreover, the search space of WM is a subset of the search space of IOM. We now discuss different solution models. \subsection{Solution Models and Approaches}\label{subsec:solutionmodel} \noindent We classify below the different solution models and discuss the approaches existing in literature. \noindent {\bf{Scalarization (SOO)}}: To deal with multiple QoS aware service composition, \cite{qi2010combining,wagner2011qos,DBLP:journals/tweb/ChattopadhyayBB17} have resorted to scalarization techniques to convert multiple objectives into a single objective using the weighted sum method. In \cite{yan2015anytime}, the authors have proposed a planning graph based approach and an anytime algorithm that attempts to maximize the utility in IOM. A scalarization technique in WM was proposed in \cite{qi2010combining}. Though scalarization techniques are simple and easy to implement, however, some information may be lost due to the transformation from multiple objectives to a single objective. Moreover, finding the weights of the parameters is difficult. User preferences are required to decide the weights of the parameters, which is not always easy to identify. Even though the preferences of the parameters are obtained, it is not easy to quantify the preferences to find the weights of the parameters, which has a great impact on finding the optimal solution. \noindent {\bf{Single-objective multi-constrained optimization (SOMCO)}}: To overcome the shortcomings of scalarization techniques, researchers have looked at another popular approach, namely, single-objective multi-constrained optimization \cite{Alrifai2010,alrifai2009combining,DBLPChattopadhyayB16}. In this approach, one parameter is selected as the primary parameter to be optimized, while for the rest of the parameters, a worst case bound is set (often termed as constraints). For example, in \cite{cao2007service}, the authors analyzed the relation between multi-objective service composition and the Multi-choice, Multi-dimension 0-1 Knapsack Problem (MMKP) in WM and used the weighted sum approach to compute the utility function. The objective of \cite{cao2007service} is to maximize the total utility while satisfying different QoS constraints. In \cite{Alrifai2010} and \cite{alrifai2009combining}, authors proposed a multi constrained QoS aware service composition approach, instead of finding the optimal solutions in WM. In \cite{Alrifai2010}, an Integer Linear Programming (ILP) based approach was proposed, where ILP is used to divide the global constraints into a set of local constraints and then using the local constraints, the service selection is done for each task in WM. In \cite{schuller2012cost,Zeng:2003:QDW:775152.775211}, ILP based methods are used to solve multi-constrained service composition in WM. Dynamic binding is the main concern of \cite{alrifai2009combining}, where authors proposed to generate the skyline services for each task in WM and cluster the services using the K-means algorithm. In \cite{DBLPChattopadhyayB16}, an ILP based multi-constrained service composition was proposed in IOM. In this class of methods as well, selecting the primary parameter to optimize (and rest to put constraints on) is a challenging problem and often depends on user preferences. Moreover, determining the constraint values is not an easy task and may this often lead to no solution being generated (i.e., no solution exists to satisfy all the constraints). \noindent {\bf{Pareto optimal front construction (POFC)}}: To address the above challenges, another research approach based on constructing the Pareto optimal frontier has been proposed. A Pareto front consists of the set of solutions where each solution is either same or better in at-least one QoS value than rest of the solutions belonging to the Pareto front. This approach does not require identifying the user preferences of the QoS parameters. Therefore, this approach can easily deal with users having different preferences. To the best of our knowledge, most of the work based on Pareto front construction \cite{schuller2012cost} focus on WM. For example, in \cite{6985731}, the authors proposed to generate the Pareto optimal solutions in a parallel setting. In \cite{trummer2014multi}, the authors proposed a fully polynomial time approximation method to solve the problem. A significant amount of work has been done based on evolutionary algorithms \cite{cremene2016comparative,li2010applying}, such as Particle Swarm Optimization \cite{liao2013multi}, Ant Colony Optimization \cite{shanshan2012improved}, Bee Colony Optimization \cite{liu2014parameter}, Genetic Algorithms \cite{zhang2013genetic,wang2008optimal}, NSGA2 \cite{hashmi2013automated,wagner2012multi}. In this paper, we consider the Pareto front construction model on IOM. Table \ref{tab:relatedModel} summarizes the state-of-the-art methods considering different models that have been discussed above. \begin{table}[!ht] \scriptsize \caption{State of the Art regarding Models} \centering \begin{tabular}{c|c} \hline Models & Methods \\ \hline WM-SOO & \cite{qi2010combining}\\\hline WM-SOMCO & \cite{Alrifai2010,alrifai2009combining,cao2007service,Zeng:2003:QDW:775152.775211,zhang2013genetic,wang2008optimal,liu2014parameter,liao2013multi,shanshan2012improved,schuller2012cost} \\\hline WM-POFC & \cite{6985731,mostafa2015multi,cremene2016comparative,hashmi2013automated,wagner2012multi,trummer2014multi,li2010applying} \\\hline IOM.SOO & \cite{yan2015anytime,wagner2011qos,DBLP:journals/tweb/ChattopadhyayBB17} \\\hline IOM-SOMCO & \cite{DBLPChattopadhyayB16} \\\hline IOM-POFC & - \\ \hline \end{tabular}\label{tab:relatedModel} \end{table} \subsection{Novelty of our work and contributions} \noindent In contrast to the above, we consider this problem in IOM-POFC setting. In addition, we have considered local and global constraints on QoS parameters. In Section \ref{sec:preliminaries}, we formally describe our model. The search space of the composition problem addressed in this paper is exponential as discussed earlier. Therefore, we first try to reduce the search space of our algorithm using clustering as demonstrated in \cite{7933195,wagner2011qos}. On the reduced search space, we propose an optimal algorithm using a graph based method. In literature, the graph based methods \cite{DBLP:journals/tweb/ChattopadhyayBB17,xia2013web,chen2014qos} are mainly applied either to solve the service composition problem for single parameter optimization or to solve the multiple QoS aware problem using scalarization. In this paper, we apply the graph based approach to construct a Pareto optimal solution frontier. The optimal algorithm is an exponential time procedure and often does not scale for large scale composition. Therefore, we further propose two heuristic algorithms. Our first heuristic algorithm is based on beam search technique. Beam search technique is applied in \cite{yan2015anytime} to solve the multiple QoS aware problem using scalarization. Here, we use the beam search technique to find Pareto optimal solutions. Since our algorithm is a heuristic approach, it does not generate the optimal solutions. However, we have shown that the solution quality monotonically improves with increase in the size of the beam width. Our second heuristic algorithm is based on NSGA. Though multiple evolutionary algorithms exist in literature \cite{zhang2013genetic,wang2008optimal,hashmi2013automated,wagner2012multi} to solve multiple QoS aware optimization, however, all these methods, to the best of our knowledge, solve the problem in WM. We use it to find the solutions for IOM. Moreover, in each step of the algorithm based on NSGA, we ensure that the solutions generated by the algorithm is a functionally valid solution. \section{Experimental Results}\label{sec:result} \noindent We implemented our proposed framework in Java. Experiments were performed on an i3 processor with 4GB RAM. The algorithms were evaluated on the 19 public repositories of the ICEBE-2005 Web Service Challenge \cite{icebe2005} and the 5 public repositories of the 2009-2010 Web Service Challenge (WSC 2009-10) \cite{bansal2009wsc} and the dataset demonstrated in \cite{7933195}. The WSC 2009-10 dataset contains only the values of two QoS parameters (response time and throughput) for each service. Additionally, we randomly generated the values for reliability and availability for each service. However, the ICEBE-2005 dataset does not contain the value of any QoS parameter for the services. We randomly generated values of all 4 QoS parameters (response time, throughput, reliability and availability) for our experiments with the ICEBE-2005 dataset. \noindent {\textbf{Configurations of our algorithms}}: No configuration parameter is required for the optimal algorithm. For our first heuristic approach, we varied the beam width from 100 to 500. Finally, we compare our first heuristic with \cite{yan2012anytime} considering the beam width as 500. For our second heuristic algorithm, we chose the following configuration parameters: population size as 100, the mutation probability as 0.01 and crossover probability as 0.85 and the number of iterations as 10000. \input{resultSub} \subsection{Different metrics to compare results} \noindent We use the following metrics to compare our results with other approaches in literature, as defined below. \begin{itemize} \item $n$: The cardinality of the solutions obtained from an algorithm. \item Commonality Ratio ${\cal{CR}}({\cal{\hat{T}}}_1, {\cal{\hat{T}}}_2)$: Given two sets of solution tuples ${\cal{\hat{T}}}_1$ and ${\cal{\hat{T}}}_2$, {\scriptsize{\[ {\cal{CR}}({\cal{\hat{T}}}_1, {\cal{\hat{T}}}_2) = \frac{|{\cal{\hat{T}}}_1 \cap {\cal{\hat{T}}}_2|} {|{\cal{\hat{T}}}_1 \cup {\cal{\hat{T}}}_2|} \]}} \item Commonality Non Dominated Solution Ratio ${\cal{CN}}({\cal{\hat{T}}}_1, {\cal{\hat{T}}}_2)$: Given two sets of solution tuples ${\cal{\hat{T}}}_1$ and ${\cal{\hat{T}}}_2$, {\scriptsize{\[ {\cal{CN}}({\cal{\hat{T}}}_i, {\cal{\hat{T}}}_3) = \frac{|{\cal{\hat{T}}}_i \cap {\cal{\hat{T}}}_3|}{|{\cal{\hat{T}}}_3|}; i=1,2; \]}} ${\cal{\hat{T}}}_3 = $ set of non dominated tuples obtained from ${\cal{\hat{T}}}_1 \cup {\cal{\hat{T}}}_2$; \item Average Distance Ratio ${\cal{AD}}({\cal{\hat{T}}}_1, {\cal{\hat{T}}}_2)$: Given two sets of solution tuples ${\cal{\hat{T}}}_1$ and ${\cal{\hat{T}}}_2$ {\scriptsize{\[ {\cal{AD}}({\cal{\hat{T}}}_1, {\cal{\hat{T}}}_2) = \frac{\frac{1}{|{\cal{\hat{T}}}_1|}\sum_{t_i \in {\cal{\hat{T}}}_1}U(t_i)}{\frac{1}{|{\cal{\hat{T}}}_2|}\sum_{t_i \in {\cal{\hat{T}}}_2}U(t_i)} \]}} $U(t_i)$ is calculated as in Eq. \ref{equ:utility} for all tuples in ${\cal{\hat{T}}}_1 \cup {\cal{\hat{T}}}_2$. \item Speed up $S(A_1, A_2)$: Given two algorithms $A_1$ and $A_2$, {\scriptsize{\[ S(A_1, A_2) = \frac{\text{Computation time for } A_2}{\text{Computation time for } A_1} \]}} \end{itemize} \noindent These metrics are used to compare the solutions obtained by two different algorithms or the same algorithm with different configurations. The {\em{average distance ratio}} metric shows the quality difference between two solutions. It may be noted, ${\cal{AD}}({\cal{\hat{T}}}_1, {\cal{\hat{T}}}_2) > 1$ means ${\cal{\hat{T}}}_1$ provides better result than ${\cal{\hat{T}}}_2$, while ${\cal{AD}}({\cal{\hat{T}}}_1, {\cal{\hat{T}}}_2) < 1$ means ${\cal{\hat{T}}}_2$ provides better result than ${\cal{\hat{T}}}_1$. The {\em{commonality ratio}} indicates how many tuples are common in two solutions, where as the {\em{commonality non dominated ratio}} shows how many non-dominated tuples are common. \subsection{Analysis on a synthetic dataset} \noindent We first analyze the performance of our proposed algorithms on a synthetic dataset demonstrated in \cite{7933195}. Here we compare the performance of the heuristic methods with respect to the optimal one. The total number of services in the service repository was 567. We used the QWS \cite{Al-Masri:2007:DBW:1242572.1242795} dataset to assign the QoS values to the services. The QWS dataset has 8 different QoS parameters and more than 2500 services. From the QWS dataset, we randomly selected 567 services and the corresponding QoS values were assigned to the services in our repository. After preprocessing, we had only 164 different services. The Pareto optimal algorithm with preprocessing achieved $6.36$ times speed up in comparison to the Pareto optimal algorithm without any preprocessing. The number of solutions obtained by the Pareto optimal algorithm ($PO$), our first heuristic ($H_1$) and our second heuristic ($H_2$) are 7, 8, 7 respectively. While $H_1$ generated 6 non dominated tuples, $H_2$ generated 5 non dominated tuples common with the tuples generated by $PO$. Average distance ratio between $PO$ and $H_1$ is 1.38, whereas, the average distance ratio between $PO$ and $H_2$ is 1.63. Our first heuristic method with beam width 500 achieved $31$ times speed up in comparison to the Pareto optimal algorithm with preprocessing, whereas, our second heuristic method achieved $43$ times speed up in comparison to the Pareto optimal algorithm with preprocessing. It may be noted, though our first heuristic algorithm generated better quality result with respect to our second heuristic algorithm, however, the second one provided higher speed up than our first one. \subsection{Analysis on public datasets} \noindent We now show the experimental results obtained by our algorithm on ICEBE-2005 and WSC 2009-10. \subsubsection{Results of Preprocessing} \noindent Given a service repository as an input to our problem, we first present the details of the reduction obtained in the number of services after the preprocessing step in our methodology on the benchmark datasets. Table \ref{tab:reduction} presents the summary on the ICEBE-2005 dataset. However, it is interesting to note that no reduction could be obtained for the WSC 2009-10 dataset, {\textcolor{black}{since the dataset does not contain any equivalent service.}} \begin{table}[!ht] \scriptsize \caption{Reduction after preprocessing for ICEBE-2005} \centering \begin{tabular}{l|c|c|l|c|c} Data Set & $N_1$ & $N_2$ & Data Set & $N_1$ & $N_2$ \\ \hline Out Composition & 143 & 79 & Composition1-20-4 & 2156 & 2032 \\ Composition1-20-16 & 2156 & 2143 & Composition1-20-32 & 2156 & 2152 \\ Composition1-50-4 & 2656 & 2596 & Composition1-50-16 & 2656 & 2652 \\ Composition1-50-32 & 2656 & 2656 & Composition1-100-4 & 4156 & 4081 \\ Composition1-100-16 & 4156 & 4150 & Composition1-100-32 & 4156 & 4155 \\ Composition2-20-4 & 3356 & 3195 & Composition2-20-16 & 6712 & 6678 \\ Composition2-20-32 & 3356 & 3347 & Composition2-50-4 & 5356 & 5239 \\ Composition2-50-16 & 5356 & 5346 & Composition2-50-32 & 5356 & 5349 \\ Composition2-100-4 & 8356 & 8233 & Composition2-100-16 & 8356 & 8347 \\ Composition2-100-32 & 8356 & 8354 & & &\\ \hline \multicolumn{6}{c}{$N_1$ and $N_2$ represent the number of services before and after preprocessing}\\ \end{tabular}\label{tab:reduction} \end{table} \subsubsection{Runtime performance analysis} \noindent We now analyze the performance of our heuristic algorithms and show the trade-off between computation time and solution quality. Tables \ref{tab:compareAnyTimeWSC} and \ref{tab:compareAnyTimeICEBE} show the comparative results obtained by the first heuristic algorithm ($H_1$) for WSC-2009 and ICEBE-2005 datasets respectively. We gradually increased the size of the beam width of the algorithm and generated the solution. We consider three different cases depending on the size of the beam width: 100, 300 and 500 respectively. The last column of Table \ref{tab:compareAnyTimeWSC} and \ref{tab:compareAnyTimeICEBE} present the computation time required by $H_1$ to generate the solution. We compare the results obtained by the algorithm in Case $i$ with the same in Case $(i + 1)$, for $i = 1, 2$. For example, the results obtained by $H_1$ with beam width 100 is compared with the same with beam width 300. As we have already discussed in Section \ref{sec:method}, as the beam size increases, either the solution quality remains same or improves. In this comparative study, we show the degradation of the solution quality when the beam size decreases in terms of ${\cal{AD}}$, ${\cal{CR}}$ and ${\cal{CN}}$. Finally, we compare our heuristic algorithms with \cite{yan2012anytime} which transforms the multiple objectives to a single objective to generate a single solution. In contrast, our method is able to generate the multiple non-dominated feasible solutions in a comparable time limit. Table \ref{tab:compareOurAnyTimeICEBE} shows the speed up achieved by both the methods with respect to \cite{yan2012anytime}. As is evident from Columns 2, 3, 5 and 6 of Table \ref{tab:compareOurAnyTimeICEBE} that in some cases, we have achieved speed-up more than 1 in 15 and 14 cases for $H_1$ and $H_2$ respectively out of 23 cases, which implies our method performs better than the method in \cite{yan2012anytime} in terms of computation time. The Pareto optimal algorithm, being compute intensive, does not produce any result and encounters a memory out error on both the datasets, on our machine with 4GB RAM. We now show a comparative performance analysis for both the heuristic methods. Table \ref{tab:compareHeuristicAnyTimeICEBE} shows the comparison between both the methods for the ICEBE-2005 and WSC-2009 datasets respectively. As evident from the tables, $H_1$ and $H_2$ generated the same results in 8 cases, $H_1$ generated better results than $H_2$ in 7 cases, while $H_2$ generated better results than $H_1$ in 9 cases. In 15 cases $H_1$ achieved more speed up than $H_2$, while in 9 cases $H_2$ outperformed $H_1$ in terms of speed up. As evident from our result, both the heuristic performed well and achieved more speed-up. \begin{table}[!ht] \scriptsize \caption{{\scriptsize{Comparison between our methods and \cite{yan2012anytime} for ICEBE-2005 (Rows 2-10) and WSC 2009 (Rows 11-13)}}} \centering \begin{tabular}{c|c|c|c} Dataset & $S_1, S_2$ & Dataset & $S_1, S_2$\\ \hline Composition1-20-4 & 0.90, 0.18 & Composition2-20-4 & 1.75, 2.54\\ Composition1-20-16 & 1.09, 0.13 & Composition2-20-16 & 1.48, 1.72\\ Composition1-20-32 & 0.78, 0.2 & Composition2-20-32 & 0.68, 0.85\\ Composition1-50-4 & 1.51, 0.52 & Composition2-50-4 & 0.70, 0.38\\ Composition1-50-16 & 0.68, 0.4 & Composition2-50-16 & 0.85, 2.83\\ Composition1-50-32 & 1.03, 0.27 & Composition2-50-32 & 7.85, 11.06\\ Composition1-100-4 & 0.57, 2.48 & Composition2-100-4 & 1.17, 0.6\\ Composition1-100-16 & 8.85, 0.9 & Composition2-100-16 & 3.8, 7.04\\ Composition1-100-32 & 3.55, 1.8 & Composition2-100-32 & 7.11, 13.67\\ \hline WSC-01 & 0.84, 1.3 & WSC-04 & 1.68, 0.45 \\ WSC-02 & 7.09, 3.97 & WSC-05 & 5.57, 1.67\\ WSC-03 & 11.11, 3.88 &&\\\hline \multicolumn{4}{l}{$S_1 = S(H_1,$\cite{yan2012anytime} $)$, $S_2 = S(H_2,$ \cite{yan2012anytime}$)$} \end{tabular}\label{tab:compareOurAnyTimeICEBE} \end{table} \begin{table}[!ht] \scriptsize \caption{\scriptsize{Comparison between our first (with beam width 500) and second heuristic methods for ICEBE-2005 (Rows 2-20) and WSC-2009 (Rows 21-25)}} \centering \begin{tabular}{c|c|c|c|c|c} Dataset & n & ${\cal{AD}}$ & ${\cal{CR}}$ & ${\cal{CN}}$ & Speed-up\\ \hline Out Composition & (3, 4) & 0.96 & 1 & (0.23, 0.39) & 1.86\\ Composition1-20-4 & (7, 2) & 2.56 & 0.59 & (0.81, 0.69) & 4.93\\ Composition1-20-16 & (5, 9) & 1.71 & 0.63 & (0.9, 0.58) & 0.69\\ Composition1-20-32 & (4, 4) & 1 & 1 & (1,1) & 8.61\\ Composition1-50-4 & (9, 7) & 0.82 & 0.67 & (0.91, 0.96) & 0.86\\ Composition1-50-16 & (6, 5) & 1.92 & 0.33 & (0.68, 0.57) & 3.91\\ Composition1-50-32 & (2, 2) & 1 & 1 & (1,1) & 0.8\\ Composition1-100-4 & (11, 11)& 1 & 1 & (1,1) & 2.93\\ Composition1-100-16 & (8, 6) & 0.9 & 0.67 & (0.33, 0.33) & 1.85\\ Composition1-100-32 & (3, 3) & 1 & 1 & (1,1) & 1.68\\ Composition2-20-4 & (7, 9) & 0.69 & 0.42 & (0.72, 0.96) & 0.3\\ Composition2-20-16 & (9, 6) & 0.91 & 0.85 & (0.63, 0.76) & 3.86\\ Composition2-20-32 & (11,9) & 0.68 & 0.98 & (0.61, 0.8) & 0.71\\ Composition2-50-4 & (2, 5) & 3.93 & 0.67 & (0.85, 0.35) & 0.23\\ Composition2-50-16 & (8, 11) & 2.91 & 0.35 & (0.96, 0.62) & 1.96\\ Composition2-50-32 & (3, 7) & 0.86 & 0.85 & (0.65, 0.5) & 9.8\\ Composition2-100-4 & (6, 6) & 1 & 1 & (1,1) & 0.54\\ Composition2-100-16 & (8, 6) & 0.8 & 0.69 & (0.85, 0.92) & 1.97\\ Composition2-100-32 & (5, 4) & 1.93 & 0.71 & (0.96, 0.81) & 0.52\\ \hline WSC-01 & (2, 2) & 1 & 1 & (1, 1) & 0.63\\ WSC-02 & (2, 2) & 1 & 1 & (1, 1) & 1.79\\ WSC-03 & (2, 2) & 1 & 1 & (1, 1) & 2.86\\ WSC-04 & (5, 6) & 0.92 & 0.83 & (0.83, 1) & 3.71\\ WSC-05 & (6, 4) & 1.87 & 0.27 & (0.8, 0.6) & 3.33\\\hline \multicolumn{6}{l}{~}\\ \multicolumn{6}{l}{${\cal{CR}}(Tuple_{H_1}, Tuple_{H_2})$, (${\cal{CN}}(Tuple_{H_1}, {\cal{\hat{T}}}_3), {\cal{CN}}(Tuple_{H_2},{\cal{\hat{T}}}_3))$}\\ \multicolumn{6}{l}{${\cal{AD}}(Tuple_{H_1}, Tuple_{H_2})$, $n$ is written as $(|Tuple_{H_1}|, |Tuple_{H_2}|)$, $S({H_1}, {H_2})$} \end{tabular}\label{tab:compareHeuristicAnyTimeICEBE} \end{table} \section{Motivating Example}\label{sec:example} \noindent In this section, we explain the motivation of our work on a simple intuitive example. Consider the services defined in Table~\ref{tab1}. The service dependency graph is shown in Figure \ref{Composition}. Consider a service query request with ${\cal{R}}_i = \{i_1, i_2, i_3\}$ with an objective to obtain ${\cal{R}}_o = \{o_{12}, o_{13}\}$ as response. A desired solution for the query is the sequence of services to be composed to arrive at the desired query output starting from the query inputs. Evidently, a number of solutions are possible as discussed below, each having different QoS value for the composite solution. Consider, for example, the solution ${\cal{Q}}_1 : S_1; S_4; [S_6 || S_7]$, where $S_1$, $S_4$ and the parallel combination $[S_6 || S_7]$ are executed sequentially ($S_6$ and $S_7$ are executed in parallel). This solution has the best QoS in terms of the number of invoked services and the throughput as 30000. In the solution ${\cal{Q}}_1$, the throughput is calculated as, min (50000, 30000, min (40000, 45000)) = 30000. Considering response time as the primary objective, the best solution is achieved by ${\cal{Q}}_2 : [S_1 || S_2 || S_3]; S_5; [S_7 || S_8]$, which has the minimum response time value as 800 ms. In the solution ${\cal{Q}}_2$, $S_1, S_2, S_3$ are executed in parallel, $S_7, S_8$ are executed in parallel and the combination $[S_1 || S_2 || S_3]$, $S_5$ and the combination $[S_7 || S_8]$ are executed sequentially. Therefore, the response time is max(100, 150, 100) + 200 + max(400, 450) = 800ms. \begin{table}[htb] \scriptsize \caption{\small{Definition of services}} \begin{center} \begin{tabular}{|c|c|c|c|c|} \hline \textbf{Service} & \textbf{Input} & \textbf{Output} & \textbf{RT(ms)} & \textbf{TR}\\ \hline \hline $S_1$ & $\{i_1\}$ & $\{i\-o_4, i\-o_5\}$ & 100 & 50000\\ \hline $S_2$ & $\{i_2\}$ & $\{i\-o_6\}$ & 150 & 45000\\ \hline $S_3$ & $\{i_3\}$ & $\{i\-o_7\}$ & 100 & 10000\\ \hline $S_4$ & $\{i\-o_4\}$ & $\{i\-o_8, i\-o_9\}$ & 250 & 30000\\ \hline $S_5$ & $\{i\-o_5, i\-o_6, i\-o_7\}$ & $\{i\-o_9, i\-o_{10}, i\-o_{11}\}$ & 200 & 10000\\ \hline $S_6$ & $\{i\-o_8, i\-o_9\}$ & $\{o_{12}\}$ & 500 & 40000\\ \hline $S_7$ & $\{i\-o_9\}$ & $\{o_{13}\}$ & 400 & 45000\\ \hline $S_8$ & $\{i\-o_{10}\}$ & $\{o_{12}\}$ & 450 & 40000\\ \hline $S_9$ & $\{i\-o_{11}\}$ & $\{o_{13}\}$ & 550 & 50000\\ \hline \end{tabular} \end{center}\label{tab1} \end{table} \begin{figure}[!ht] \centering \includegraphics[height=2.5cm,width=5cm]{fig/2.pdf} \caption{Service dependency graph} \label{Composition} \end{figure} \noindent An obvious solution for optimal Q0S-based composition is to explore all paths (or possible sub graphs) of the dependency graph (Figure \ref{Composition}), calculate the values of the QoS parameters for each possibility and choose the path (or subgraph) which has the best QoS. If different QoS parameters give different solution, we can choose the solution which has the best QoS value for one of the parameters and the values for the rest of the parameters as good as possible. Evidently the solution may not be unique and may vary depending on the choice of the parameter taken as the primary objective. We first consider {\em response time} as the primary QoS parameter. \cite{6649580} proposed an iterative solution, where using a forward QoS computation technique, for each node ${\cal{N}}_i$ of the dependency graph, the best response time for the composition solutions till the node ${\cal{N}}_i$ is calculated. Following this construction, we get ${\cal{Q}}_2$ as the solution for the above example. On the other hand, if we consider {\em throughput} as the primary QoS parameter, following the same construction, we get ${\cal{Q}}_1$ as the optimal solution for the above example. If we consider both simultaneously, the answer is not obvious. Finding the optimal solution considering multiple QoS parameters is a nontrivial challenge, and we address this in this paper. \\ \noindent {\textbf {Our contribution :}} Contemporary approaches existing in literature take considerable time to produce the best solution at runtime. In this proposal, on one hand, we use a single QoS aware path selection technique for resolving a given service query in a fast and efficient way. This is a modest elaborate version of our contribution in \cite{chattopadhyay2015scalable}. On the other hand, we additionally address the problem of multiple QoS parameters and propose a path selection strategy. The essential idea behind our approach is to select and construct a query solution path beginning with the desired output and progressively traverse backward on the service dependency graph. Though our scheme does not guarantee the best solution in terms of QoS, we show that we can achieve a close to optimal solution in considerably less time than the contemporary approaches. We also propose a parameterized search algorithm that improves on solution quality at the cost of computation time. In the next section, we discuss the detailed methodology behind our proposals.

49abf5b649a622df77aca9f953141946184db4fe

\section{Introduction} \fi \iffalse \section{Presentation} \fi The quantum Rabi model (QRM) describes the interaction of a two-level emitter with a single quantized mode of the electromagnetic field or, more generally, of a two-level system (TLS) with a bosonic mode. Together with its extension for an ensemble of emitters, i.e., the Dicke model (DM), it constitutes a cornerstone of quantum optics~\cite{Braak16}. The physics predicted by the QRM and the DM strongly depends on the relative values of the mode frequency, $\omega$, and the coupling strength between the TLS and the bosonic mode, $g$. For weak coupling, i.e., $g/\omega \ll 1$, the rotating wave approximation (RWA) applies. In this case, the QRM and the DM reduce to the Jaynes-Cummings and the Tavis-Cummings models, respectively. The RWA breaks down in the ultra-strong coupling regime (USC), i.e., for $g/\omega \gtrsim 0.1$. When increasing the coupling strength further, one enters the deep-strong coupling regime (DSC)~\cite{Rossatto17}. For such high values of $g/\omega$, new phenomena are expected~\cite{Ashhab10, Casanova10, Garziano16, ArXiv_forn-diaz18, ArXiv_Kockum18}. The existence of a quantum phase transition in the thermodynamic limit adds to the richness of the DM~\cite{Hepp73, Wang73, Emary03b}. Furthermore, USC and DSC may enable novel protocols for quantum communication and quantum information processing~\cite{Nataf11,Romero12,Kyaw15}. Over the last decade, USC was reached using various experimental platforms~\cite{Anappara09,Guenter09,Bourassa09,Todorov10,Niemczyk10,Forn-Diaz10,Schwartz11,Zhang16,George16}. More recently, DSC was achieved in circuit quantum electrodynamics~\cite{Yoshihara17,Forn-Diaz17} as well as by coupling a THz metamaterial with cyclotron resonances in a two-dimensional electron gas~\cite{Bayer17}. While these systems reach record-high ratios of $g/\omega$, the large coupling strengths make state preparation and read-out challenging. For this reason, alternative routes were proposed to achieve large coupling in experimental platforms that, at the same time, offer a high level of control and tunability. Following this path, the dynamics of the QRM in the USC and DSC regimes was studied, respectively, with analog and digital quantum simulations using circuit quantum electrodynamics~\cite{Braumueller17,Langford17}, and DSC was reached with single trapped ions~\cite{Lv18}. Here, we implement a mechanical analogue of the Dicke model by coupling the spin of individual neutral atoms to their quantized motion in a trapping potential. In our approach, the coupling is enabled by spatial gradients of the vector light shift inherent to optical microtraps. Fluorescence spectroscopy grants access to the energy spectrum of the system, revealing ultra-strong spin-motion coupling in our experiment, i.e., the coupling strength is a significant fraction of the mode frequency. Furthermore, we demonstrate that the coupling strength can be readily and independently tuned in situ. \iffalse \section{Experimental setup and coupling scheme} \fi \begin{figure} \includegraphics[width=0.9\columnwidth]{ExpSetup_combined_LO3D.pdf} \caption{\textbf{Experimental setup}. \textbf{a}, Individual cesium atoms are trapped near the surface of the nanofiber-section of a tapered optical fiber. They are exposed to a near-resonant excitation laser field (frequency $\omega_\mathrm{I}$), propagating along the $+y$-direction. A fraction of the atomic fluorescence is scattered into the guided mode of the nanofiber (frequency $\omega_\mathrm{S}$). This light is superposed with a reference beam that is derived from the excitation laser and frequency-shifted by $+10\,\mathrm{MHz}$. The resulting beat note is recorded with a single photon counting module (SPCM). Its Fourier analysis yields the fluorescence spectrum, which grants access to the energy spectrum of the trapped atoms. \textbf{b}, The strong spatial confinement of the trapping light fields results in large fictitious magnetic field gradients that give rise to a coupling between the atomic spin and its motional degree of freedom. Contours of the scalar part of the trapping potential are indicated by black lines. A yellow dot marks the position of the atom at one trapping site. The amplitude of the main component of the field, $B_\mathrm{fict}^{(x)}$, is shown as a density plot (logarithmic color scale). The gray disk indicates the cross section of the nanofiber.} \label{fig:ExpSetup_combined} \end{figure} Our implementation employs laser-cooled individual cesium atoms trapped in the evanescent light field surrounding the nanofiber-section of a tapered optical fiber, see Fig.~\ref{fig:ExpSetup_combined}a and supplemental material (SM). The strong transverse confinement of the trapping light fields results in a strong polarization gradient in the azimuthal direction. In addition to the scalar light shift that gives rise to trapping, atoms in the evanescent field then experience a spatially-varying vector light shift~\cite{Albrecht16, Meng18}. This shift can be thought of as arising from the Zeeman interaction with a position-dependent fictitious magnetic field, $\bm{B}_\mathrm{fict}$~\cite{Cohen-Tannoudji72}. For our configuration, $\bm{B}_\mathrm{fict}$ mainly points along the $x$-direction, and its amplitude exhibits strong spatial gradients, see Fig.~\ref{fig:ExpSetup_combined}b. Near the trap minimum, the $x$-component of the fictitious magnetic field varies approximately linearly along $y$, so that $\bm{B}_\mathrm{fict} \approx b_y y \,\bm{e}_x$, with $\bm{e}_x$ the unit vector along $x$ and $b_y \approx \SI{1.9}{\gauss\per\micro\meter}$. The Zeeman interaction of a trapped atom with this fictitious magnetic field results in a coupling between the atomic spin and motional degrees of freedom (DOF). Here, we assume a harmonic trapping potential, with a set of frequencies $\{\omega_i\}$ and annihilation operators $\{\hat{a}_i\}$ ($i=x,y,z$). In addition to $\bm{B}_\mathrm{fict}$, we apply a homogeneous offset magnetic field, $\bm{B}_0 = B_0 \, \bm{e}_{y}$, along the $y$-direction. The dynamics of a trapped atom is then described by the following Hamiltonian: \begin{equation} \hat{H} = \sum\limits_{i=x,y,z} \hbar \omega_i \hat{a}^\dagger_i \hat{a}_i + g_F \mu_B \bm{\hat{F}}\cdot(\bm{B}_0+\bm{B}_{\text{fict}}), \label{eq:Hat_ini} \end{equation} \noindent with $g_F$ the hyperfine Land\'e factor and $\mu_B$ the Bohr magneton. Assuming that the fictitious magnetic field consists of a linear gradient along $y$, and only considering the $y$ motional DOF, we can rewrite~\eqref{eq:Hat_ini} as (see SM): \begin{equation} \hat{H}_y = \hbar \omega_y \hat{a}^\dagger_y \hat{a}_y + \hbar\Delta\hat{F}_{y} + \frac{\hbar g_y}{\sqrt{2F}} \left( \hat{a}^\dagger_y + \hat{a}_y \right) \left( \hat{F}_{+} + \hat{F}_{-} \right), \label{eq:Hat_QRM} \end{equation} \noindent where $\hat{F}_{+}$ ($\hat{F}_{-}$) is the spin raising (lowering) operator for the eigenstates of $\hat{F}_{y}$ with eigenvalues $\hbar m_{F}$. For $F=1/2$, Hamiltonian~\eqref{eq:Hat_QRM} corresponds to the QRM, while for $F>1/2$, as is the case for cesium, it corresponds to the DM. The physics is governed by three parameters: The bosonic mode frequency, $\omega_y$, the Zeeman splitting between adjacent $m_{F}$-states, $\Delta \propto B_0$, and the spin-motion coupling strength, $g_y \propto b_y$. For our configuration, we expect $g_y \approx 2\pi\times\SI{19}{\kilo\hertz}$ for a calculated trap frequency $\omega_y \approx 2\pi\times\SI{95}{\kilo\hertz}$, i.e., $g_y/\omega_y \approx 0.2$. The low-energy eigenstates of $\hat{H}_y$ are illustrated in Fig.~\ref{fig:2D_Spectrum}a,b. We consider the case of cesium in the $F=4$ hyperfine ground state. In the absence of spin-motion coupling ($g_y=0$), the eigenstates are the bare states $\ket{m_F, n_y}$, where $n_y$ corresponds to a Fock state of the harmonic trapping potential. In the presence of spin-motion coupling, the new eigenstates are dressed states. When the coupling is resonant ($\Delta = \omega_y$), the degeneracy of the bare states $\ket{-4,1}$ and $\ket{-3,0}$ is lifted, and the new eigenstates are $\ket{\pm}=(\ket{-4,1} \mp \ket{-3,0})/\sqrt{2}$, separated in energy by $\hbar\Omega_y$, where $\Omega_y>0$ is the Rabi frequency. Here, we expect $\Omega_y = 2 g_y\approx 2\pi\times\SI{38}{\kilo\hertz}$ (see SM). \iffalse \section{Fluorescence spectroscopy and 2D spectrum} \fi \begin{figure*} \includegraphics[width=0.9\textwidth]{2D_Spectrum.pdf} \caption{\textbf{Experimental signature of ultra-strong spin-motion coupling}. \textbf{a}, The bare eigenstates in the harmonic trapping potential are $\ket{m_F, n_i}$ with eigenenergies $\hbar(m_F \Delta+n_i \omega_i)$, where $\hbar\Delta$ is the Zeeman splitting between two neighboring $m_F$-states and $\hbar\omega_i$ is the energy of one motional quantum. The spin-motion coupling is resonant for $\Delta = \omega_i$. \textbf{b}, At resonance, the spin-motion coupling lifts the degeneracy between the bare states $\ket{-4,1}$ and $\ket{-3,0}$, and the new eigenstates $\ket{+}$ and $\ket{-}$ (the dressed states) are split by $\hbar \Omega_i$. \textbf{c}, Fluorescence spectra for different values of $\Delta$. Avoided crossings occur when the resonance condition is fulfilled for the $x$- and $y$-DOFs. Dashed white lines indicate the frequencies of the transitions between bare states and are derived from a fit of the experimental data far from resonance (see SM). Black lines correspond to transitions between dressed states, where the coupling strengths are derived from a fit of the experimental data at resonance (see SM). The values of the fit parameters are given in the main text. Solid black lines correspond to transitions from and to the ground state and dashed black lines to transitions between excited states. \textbf{d}, \textbf{e}, Fluorescence spectra, measured for two Zeeman splittings, $\Delta_1$ and $\Delta_2$, respectively (cf.~solid horizontal white lines in \textbf{c}). Far from resonance, three pairs of motional sidebands are apparent (\textbf{d}). When the coupling is resonant, one motional sideband is split (\textbf{e}). We also observe sidebands corresponding to transitions between the excited states (green and purple triangles).} \label{fig:2D_Spectrum} \end{figure*} In order to probe the low-energy part of Hamiltonian~\eqref{eq:Hat_QRM}, we perform a heterodyne fluorescence spectroscopy measurement~\cite{Jessen92, Meng18}. The experimental setup is sketched in Fig.~\ref{fig:ExpSetup_combined}a. The atoms are exposed to a laser light field propagating along the $+y$-axis and $\sigma^-$-polarized with respect to the propagation direction. The laser is red-detuned with respect to the cycling transition of the $D_2$ line of Cesium, and its intensity is kept low enough to ensure that it is scattered coherently by the atoms (see SM). This laser provides degenerate Raman cooling~\cite{Meng18} and optical pumping, so that most of the atoms populate the low-lying energy states depicted in Fig.~\ref{fig:2D_Spectrum}a,b. Part of the fluorescence light is scattered into the guided mode of the optical nanofiber~\cite{Mitsch14b}. This light is superposed with a reference beam, derived from the excitation laser and frequency-shifted by $+\SI{10}{\mega\hertz}$. The resulting beat note is recorded using a single photon counting module (SPCM). Post-processing the SPCM data yields the intensity power spectral density (PSD). This heterodyne setup enables a precise measurement of the frequency difference between the incoming photons from the excitation beam (frequency $\omega_I$) and the photons scattered by the atoms (frequency $\omega_S$). In the case of elastic scattering, the atomic state and the frequency of the photons are unchanged ($\omega_I = \omega_S$), yielding the carrier peak in the PSD. In the case of inelastic scattering, the atomic state is changed and the difference of energy between the incoming and scattered photons has to match the difference of energy between the initial and final atomic states. This gives rise to sidebands around the carrier peak, the positions of which grant access to the energy spectrum of the atoms. We record fluorescence spectra for different values of the Zeeman splitting, $\Delta\propto B_0$, see Fig.~\ref{fig:2D_Spectrum}c-e. Far from resonance, i.e., for $|\Delta - \omega_i| \gg \Omega_i$, transitions between the bare states result in three pairs of motional sidebands, see Fig.~\ref{fig:2D_Spectrum}d. These transitions change the motional state of the atom but not its spin. These sidebands do not depend on $\Delta$, and their positions can be used to infer the trap frequencies. We find $\left\{\omega_x,\, \omega_y,\, \omega_z\right\} = 2\pi\times \left\{149(2),\, 93(2),\, 243(5)\right\}\,\mathrm{kHz}$. The strong asymmetry of the amplitudes of the positive- and negative-frequency peaks indicates that the atoms are close to the motional ground state~\cite{Meng18}. A fourth peak is also visible in the upper left part of Fig.~\ref{fig:2D_Spectrum}c. It corresponds to a transition between adjacent $m_F$-states for a given motional state. Its position depends linearly on $\Delta$. Close to resonance, we observe a splitting of the motional sideband corresponding to the resonantly coupled DOF. This is clearly visible in Fig.~\ref{fig:2D_Spectrum}e, which is measured close to the resonance of the $y$-DOF, i.e., for $\Delta \approx \omega_y$. The width of the splitting already indicates that we operate in the USC regime. When scanning $\Delta$ around resonance, an avoided crossing is observed. Such an avoided crossing is also visible for the $x$-DOF, indicating that strong spin-motion coupling is present for this DOF, too. This additional coupling could arise from the polarizations of the trapping light fields not being perfectly aligned and/or from a spurious vector light shift originating from the interference of the probe light with its reflection on the nanofiber. Both effects should lead to a significant $x$-gradient of the fictitious magnetic field and, thus, to a strong spin-motion coupling for this DOF. \begin{figure} \includegraphics[width=0.9\columnwidth]{880_effect.pdf} \caption{\textbf{Experimental demonstration of the tunability of the spin-motion coupling strength}. A nanofiber-guided tune-out laser field at a wavelength of $\lambda=\SI{880}{nm}$ allows one to modify the fictitious magnetic field gradient along $y$ and, thus, the corresponding spin-motion coupling strength, without changing the scalar part of the trapping potential. \textbf{a}, Fluorescence spectra, taken for resonant coupling of the $y$-DOF ($\Delta = \omega_y$) and different values of the tune-out laser power, $P_{880}$. We show the positive-frequency peak corresponding to the transition between the dressed states. Its position corresponds to the Rabi splitting, $\Omega_y$. This peak shifts towards the carrier for increasing values of $P_{880}$, indicating a reduction of the coupling strength. \textbf{b}, As expected from theory, $\Omega_y$ depends linearly on $P_{880}$, see red circles. A fit (black dashed line) yields $\mathrm{d}\Omega_y/\mathrm{d}P_{880} = \SI{-120(10)}{\hertz\per\micro\watt}$. For $P_{880} > \SI{100}{\micro\watt}$, the proximity of the carrier impedes a precise measurement of the peak position.} \label{fig:880_effect} \end{figure} Besides the Rabi splitting, a new pair of peaks is apparent close to resonance. These sidebands, labeled by triangles in Fig.~\ref{fig:2D_Spectrum}e, are located at $\pm \Omega_i$ around the carrier and correspond to transitions between the dressed states. The observation of transitions from the ground state to the lowest pair of dressed states, giving rise to the celebrated vacuum Rabi splitting, and the simultaneous observation of direct transitions between dressed states is enabled by two features of our system: First, although we achieve cooling close to the motional ground state~\cite{Meng18}, there is a finite population of the first excited states and, therefore, we can observe transitions starting from these states; second, the energy gap, $\hbar\Omega_i$, between the dressed states is comparable to the energy gap between the ground state and the first excited state manifold, $\hbar\omega_i$, so that the corresponding transitions have similar energy and can be detected by the same method. The position of these sidebands allows us to precisely measure the Rabi splitting, $\Omega_i$. We find $\Omega_y = 2\pi \times \SI{35(1)}{kHz}$ and $\Omega_x = 2\pi \times \SI{36(1)}{kHz}$. This corresponds to coupling strengths of $g_y/\omega_y = 0.19(1)$ and $g_x/\omega_x = 0.12(1)$, respectively. Thus, we clearly reach the ultra-strong coupling regime for both DOFs~\footnote{Here, we choose the $g_i/\omega_i$ ratio as a figure of merit for USC. Another common choice consists in using the $\Omega_i/\omega_i$ ratio~\cite{Guenter09,Todorov10,Schwartz11,George16}, which would yield, in our case, $\Omega_y/\omega_y = 0.38(1)$ and $\Omega_x / \omega_x = 0.24(1)$.}. \iffalse \section{Tuning the coupling } \fi Another feature of our setup is the possibility to tune the coupling strength \textit{in situ}. For this purpose, we use an additional fiber-guided light field at the so-called tune-out wavelength, near \SI{880}{\nano\meter}~\cite{Arora11}. At this wavelength, the scalar polarizability vanishes, so that this laser field only induces a vector light shift. This field propagates in the same direction as the blue-detuned trapping light field and has the same polarization. For this configuration, we expect a partial compensation of the fictitious magnetic field gradient~\cite{Albrecht16} and, thus, a reduction of the coupling strength. To experimentally quantify this effect, we measured the Rabi splitting, $\Omega_y$, for different powers of the tune-out laser, $P_{880}$, see Fig.~\ref{fig:880_effect}. As expected in our regime, $\Omega_y$ decreases linearly with $P_{880}$. The measured slope is $\mathrm{d}\Omega_y/\mathrm{d}P_{880} = \SI{-120(10)}{\hertz\per\micro\watt}$. From an \textit{ab initio} calculation, taking into account the vector polarizability of cesium and the mode function of the nanofiber-guided tune-out light, we expect $\SI{-100}{\hertz\per\micro\watt}$, in reasonable agreement with the experimental value. We note that, for a different experimental configuration, the tune-out laser field may also enhance the coupling strength. In this case, a power of $P_{880} \approx \SI{800}{\micro\watt}$ should be sufficient to induce a coupling on the order of the trap frequency. Furthermore, by modulating the tune-out laser field intensity, one may dynamically adjust the coupling strength, even on timescales shorter than the Rabi oscillation period. This might enable, e.g., adiabatic USC/DSC ground-state preparation or the study of quench dynamics. \iffalse \section{Conclusion} \fi In summary, the demonstrated implementation of a mechanical analogue of the Dicke model with cold atoms constitutes a novel route to explore ultra-strong and, potentially, even deep-strong coupling phenomena with unprecedented level of control. While our implementation takes advantage of the specific polarization gradients intrinsically present in our nanofiber-based optical trap, other optical micropotentials, such as optical lattices~\cite{Bloch05}, also qualify for implementing our scheme~\cite{Schneeweiss18}. Possible future research directions include the study of the dynamical Casimir effect via a modulation of the system parameters~\cite{Liberato07} or of the role of dissipation in the USC/DSC regime~\cite{Beaudoin11}. Understanding these effects will be beneficial, e.g., for the realization of ultra-fast quantum gates~\cite{Romero12,Kyaw15} or of qubit protection protocols~\cite{Nataf11} relying on USC. Finally, a suitably tailored real and fictitious magnetic field pattern can be used to realize generalizations of the quantum Rabi model or of the Dicke model, such as the driven QRM, or to implement ultra-strong two-photon coupling~\cite{Schneeweiss18}. \section*{Acknowledgments} Financial support by the European Research Council (CoG NanoQuaNt) and the Austrian Science Fund (FWF, SFB NextLite Project No. F 4908-N23 and DK CoQuS project No. W 1210-N16) is gratefully acknowledged. \section{Experimental setup and sequence} Laser-cooled cesium atoms are trapped in the evanescent field surrounding a silica optical nanofiber of nominal radius $a = \SI{250}{nm}$. The trapping potential is created by sending a blue-detuned running wave field with a free-space wavelength of \SI{783}{\nano\meter} and a power of $\SI{17.8}{\milli\watt}$ and a red-detuned standing wave field at \SI{1064}{\nano\meter} wavelength with a total power of $\SI{2.80}{\milli\watt}$ into the nanofiber. The blue- and the red-detuned fields are guided in the quasi-linearly polarized fundamental HE$_{11}$ modes and the polarizations of the two fields are orthogonal. Two diametric arrays of trapping sites are formed~\cite{Vetsch10}. The calculated trap frequencies are $(\omega_x, \,\omega_y, \,\omega_z) = 2\pi\times(145,\,95,\,228)\,\mathrm{kHz}$ in the radial, azimuthal, and axial directions, respectively, for a total trap depth of about $\SI{200}{\micro\kelvin}$. The trap minima are located about $\SI{270}{\nano\meter}$ away from the nanofiber surface. The atoms are loaded into the nanofiber-based trap from a magneto-optical trap via an optical molasses stage~\cite{Vetsch10}. In this process, the collisional blockade effect~\cite{Schlosser02} limits the maximum number of atoms per trapping site to one. After loading, atoms are distributed over the two diametric arrays of trapping sites. In order to reduce possible inhomogeneities, we remove atoms from one of the two arrays. This is achieved by performing degenerate Raman cooling using a fiber-guided light field, as first demonstrated in~\cite{Meng18}. The fiber-guided light at the position of the atoms is almost perfectly $\sigma^-$- and $\sigma^+$-polarized for one array and the other, respectively. This leads to degenerate Raman cooling of the atoms in one array and to heating of the atoms in the other array. After a few millisecond of optical pumping, all the atoms on the $\sigma^+$-side are lost, while the atoms on the $\sigma^-$-side remain trapped and are cooled close to the ground state. The heterodyne fluorescence spectroscopy setup is described in~\cite{Meng18}: The excitation laser beam features a waist of about $\SI{1.2}{\milli\meter}$ at the position of the atoms and a total power of $\SI{120}{\micro\watt}$. Its frequency is red-detuned by about $\SI{55}{MHz}$ with respect to the $F=4 \to F^\prime=5$ transition of the $D_2$ line of cesium. The corresponding peak intensity is around $5\,\mathrm{I_\mathrm{sat}}$, where $I_\mathrm{sat}$ is the saturation intensity of the considered transition. The resulting saturation parameter is on the order of $10^{-2}$, meaning that incoming photons are mostly coherently scattered. Atoms are exposed to the excitation beam for $\SI{100}{\milli\second}$. A typical spectrum as shown in Fig.~2 is obtained after averaging over about 7000 experimental runs. For an offset magnetic field smaller than \SI{0.2}{G}, degenerate Raman cooling is no longer effective as the Zeeman states become degenerate in energy. The atoms are then heated out of the trap by the excitation laser in about \SI{10}{\milli\second} during heterodyne detection due to recoil heating, which lowers the signal to noise ratio. To mitigate this problem, we use an interleaved detection scheme alternating \SI{5}{\milli\second} of heterodyne measurement and \SI{2}{\milli\second} of optical molasses cooling. We perform between 20 to 40 cycles for each experimental run. This interleaved detection scheme was used for the spectra corresponding to Zeeman splittings of $\SI{0}{\kilo\hertz} \leq \Delta/(2\pi) \leq \SI{70}{\kilo\hertz}$ in Fig.~2c, i.e., for $\SI{0}{G} \leq B_0 \leq \SI{0.2}{G}$. We compared the spectra obtained with this interleaved scheme with the ones measured with the continuous scheme for ranges of offset magnetic field where both methods work. We did not observe a significant difference on the peak positions for the two schemes. \section{Detailed derivation of the spin-motion coupling} Here, we detail the mapping of Hamiltonian~(1) to Hamiltonian~(2). Assuming a linear gradient for the fictitious magnetic field ($\bm{B}_\mathrm{fict} \approx b_y y \,\bm{e}_x$) and using that $\bm{B}_0 = B_0\,\bm{e}_y$, Hamiltonian~(1) becomes: \begin{equation} \hat{H}_y = \hbar \omega_y \hat{a}^\dagger_y \hat{a}_y + g_F \mu_B B_0 \hat{F}_{y} + g_F \mu_B b_y \hat{y}\hat{F}_{x}. \label{eq:Hat_QRM_spin_position} \end{equation} \noindent We can write the position operator $\hat{y}$ in terms of the raising and lowering operators using $\hat{y} = y_0 (\hat{a}^\dagger_y + \hat{a}_y)$, where $y_0 = \sqrt{\hbar/(2M\omega_y)}$ is the size of the harmonic oscillator ground state. Here $M$ denotes the atomic mass. We also introduce $\hat{F}_{+}$ and $\hat{F}_{-}$, which are the raising and lowering operators for the eigenstates of $\hat{F}_{y}$, respectively, so that $\hat{F}_{x} = (\hat{F}_{+} + \hat{F}_{-}) / 2$. Hamiltonian~(\ref{eq:Hat_QRM_spin_position}) can then be written as: \begin{equation} \hat{H}_y = \hbar \omega_y \hat{a}^\dagger_y \hat{a}_y + g_F \mu_B B_0 \hat{F}_{y} + \frac{g_F \mu_B b_y y_0}{2} \left( \hat{a}^\dagger_y + \hat{a}_y \right) \left( \hat{F}_{+} + \hat{F}_{-} \right). \end{equation} \noindent This is identical to Hamiltonian~(2), where we identify the Zeeman splitting, $\Delta$, and the spin-motion coupling strength, $g_y$, as: \begin{align} \hbar\Delta & = g_F \mu_B B_0, \\ \frac{\hbar g_y}{\sqrt{2}} & = \frac{g_F \mu_B b_y y_0}{2}. \end{align} We denote $\ket{m_F, n_y}$ the eigenstates of Hamiltonian~(2) in the absence of spin-motion coupling, where $m_F$ and $n_y$ label eigenstates of $\hat{F}_{y}$ and Fock states of the harmonic potential along $y$, respectively. We are interested in the Rabi frequency, $\Omega_y$, between the low-energy states $\ket{a} = \ket{m_F=-F, n=1}$ and $\ket{b} = \ket{m_F=-F+1, n=0}$. For $\Delta = \omega_y$, these states are resonantly coupled and we find: \begin{equation} \frac{\hbar\Omega_y}{2} = \left|\bra{b} \hat{H}_y \ket{a} \right| = \frac{\hbar g_y}{\sqrt{2F}} \left|\bra{b} \hat{a}_y \hat{F}_{+} \ket{a} \right| = \hbar g_y. \end{equation} \noindent Hence, we find $\Omega_y = 2 g_y$. \section{Models for numerical simulations} \subsection{Simplified model} Here, we describe the derivation of the expected transitions frequencies, indicated by black lines in Fig.~3c. In order to gain more intuition, we used a simplified model, only considering the low-energy states shown in Fig.~3b. This model turns out to be sufficient to describe the observed spectra. Indeed, because of the degenerate Raman cooling and optical pumping originating from the probing laser field, most of the atoms populate these low-lying energy states. The simplified model only considers the following bare states: \begin{align} \ket{g} & = \ket{m_F=-4, n_x=0, n_y=0}, \\ \ket{e} & = \ket{m_F=-3, n_x=0, n_y=0}, \\ \ket{x} & = \ket{m_F=-4, n_x=1, n_y=0}, \\ \ket{y} & = \ket{m_F=-4, n_x=0, n_y=1}. \end{align} \noindent The corresponding Hamiltonian reads: \begin{equation} \begin{split} \hat{H}/\hbar = & \Delta \ketbra{e}{e} + \omega_x \ketbra{x}{x} + \omega_y \ketbra{y}{y} \\ + & (\Omega_x/2) \left(\ketbra{e}{x} + \ketbra{x}{e} \right) \\ + &(\Omega_y/2) \left(\ketbra{e}{y} + \ketbra{y}{e} \right),\\ \end{split} \label{eq:simple_Ham} \end{equation} \noindent where $\Omega_{x,y} = 2 g_{x,y}$ are the Rabi frequencies for the $x\leftrightarrow e$ and $y\leftrightarrow e$ transitions, respectively. A numerical diagonalization of~(\ref{eq:simple_Ham}) in the basis $\{ \ket{g}, \ket{e}, \ket{x}, \ket{y} \}$ for different values of the Zeeman splitting, $\Delta$, yields sets of four eigenergies. The fluorescence spectra, as presented in Fig.~2, are expected to show peaks at frequencies corresponding to the difference of the obtained eigenenergies. In order to confirm the validity of this simplified model, we compared its results with the ones of a numerical simulation considering the full Hamiltonian, see supplementary material. In our range of parameters and for the typical temperatures measured in our setup, the two models show a good agreement. The fit of the experimental data with this model is performed as follows: Trap frequencies are inferred by fitting the peaks corresponding to motional sidebands for large values of the offset magnetic field, $B_0$, where the effect of spin-motion coupling is negligible. More in detail, we perform a fit of the spectra shown in Fig.~2c for $\SI{250}{\kilo\hertz} \leq \Delta/(2\pi) \leq \SI{330}{\kilo\hertz}$. The Zeeman splitting, $\Delta$, was calibrated as follows: The coils generating the offset magnetic field, $B_0$, were calibrated independently using microwave spectroscopy. From this calibration, we inferred the linear relation between the current circulating in the coils and the amplitude of the resulting magnetic field. In order to account for shifts induced by spurious real and fictitious magnetic fields originating, e.g., from the excitation light field, we performed a fit on the peak corresponding to the $\ket{g} \to \ket{e}$ transition in Fig.~2c. Its position corresponds to $\omega_S-\omega_I = \Delta$. Again, this was done for large values of $B_0$, corresponding to $\Delta / (2\pi) \geq \SI{270}{\kilo\hertz}$. Finally, the coupling strengths $g_{x,y}$ were obtained by fitting the position of the peak corresponding to the transition between dressed states at resonance, $\Delta = \omega_{x,y}$, as explained in the main text of the manuscript. \subsection{Full Dicke model} \begin{figure} \includegraphics[width=\columnwidth]{Methods_simulation_comparison.pdf} \caption{\footnotesize \textbf{Comparison between simplified and full theoretical model}. The black dashed lines correspond to the simplified model discussed above, only considering low-energy states. They are identical to the ones shown in Fig.~2c. The result of a numerical diagonalization of the full Hamiltonian is shown with a colormap for the parameters given in Table.~\ref{table:simulation_parameters}. The two models agree very well.} \label{fig:Methods_simulation_comparison} \end{figure} \begin{table} \footnotesize \vspace{0.5cm} {\renewcommand{\arraystretch}{1.2} \begin{tabular}{|l|c|c|} \hline \textbf{Parameter} & \textbf{Notation} & \textbf{Value}\\ \hline \multicolumn{3}{|c|}{$y$-DOF} \\ \hline trap frequency & $\omega_y$ & $2\pi \times \SI{93}{kHz}$ \\ coupling strength & $g_y$ & $2\pi \times \SI{17.5}{kHz}$ \\ mean number of motional quanta & $\mean{n_y}$ & $0.5$ \\ Lamb-Dicke parameter & $\eta_y$ & $0.15$ \\ \hline \multicolumn{3}{|c|}{$x$-DOF} \\ \hline trap frequency & $\omega_x$ & $2\pi \times \SI{149}{kHz}$ \\ coupling strength & $g_x$ & $2\pi \times \SI{18}{kHz}$ \\ mean number of motional quanta & $\mean{n_x}$ & $0.5$ \\ Lamb-Dicke parameter & $\eta_x$ & $0.1$ \\ \hline \end{tabular} } \caption{\footnotesize Parameters used for the numerical simulations shown in Fig.~\ref{fig:Methods_simulation_comparison} and inferred from the experimental spectra shown in Fig.~2c.} \label{table:simulation_parameters} \end{table} Here, we compare the simplified model, described in the methods and used to compute the expected transitions frequencies in Fig.~2c, with a full diagonalization of Hamiltonian~(2), taking into account the $x$ and $y$ motional DOFs as well as all the Zeeman states. The result is shown in Fig.~\ref{fig:Methods_simulation_comparison}. We performed a numerical diagonalization of~(2) using the QuTiP python toolbox~\cite{Johansson13}, for the parameters shown in Table~\ref{table:simulation_parameters}. The simulation was considering a $F=4$ spin, and a set of $5$ motional states for each motional DOF. We assumed a temperature corresponding to a mean number of motional quanta of $\mean{n} = 0.5$~\cite{Meng18}. From the numerical diagonalization, we obtain a set of eigenenergies $\{E_i\}$ and corresponding eigenstates $\{\ket{i}\}$. The transition $\ket{i}\to \ket{j}$ gives rise to a peak centered at a frequency $\hbar\omega_{ij} = E_i - E_j$. For the spectra plotted in Fig.~\ref{fig:Methods_simulation_comparison}, we assumed the peaks to be Gaussian with a constant width of \SI{2}{kHz}. The amplitude of the peak is given by the initial (thermal) population of the state $\ket{i}$ and a Frank-Condon factor $F_{ij} = \bra{j}\hat{V}\ket{i}$. In the Lamb-Dicke regime, we can write $\hat{V} = (1 + \eta_x [\hat{a}^\dagger_x + \hat{a}_x] + \eta_y [\hat{a}^\dagger_y + \hat{a}_y]) \otimes \hat{S}_- \hat{S}_-^\dagger$, where $\eta_{x}$ and $\eta_{y}$ are the corresponding Lamb-Dicke parameters. Here, $\hat{S}_-$ ($\hat{S}_-^\dagger$) is a generalized lowering (raising) operator for the emission (absorption) of a $\sigma^-$-polarized photon, taking into account the Clebsch-Gordan coefficients of the considered optical transition~\cite{Gangl02}. As shown in Fig.~\ref{fig:Methods_simulation_comparison}, the agreement between the simplified model (dashed lines) and the full model (colormap) is very good for the range of parameters considered here.

e33e85440e3a5aa7a373147e262fc67e918c750a

\section{Introduction} Recently, \cite{Hernan2018b} classified Data Science into three tasks: description, prediction, and causal inference. The first two fall firmly in the realm of statistical inference in that they are purely data-driven tasks, while the last requires something more than the observed data alone \citep{Pearl2016}. Consider, for example, the target population of HIV-infected women of child-bearing age (15-49 years old) in East Africa. After obtaining measurements on sample of women from this population, we could provide some basic descriptive statistics on demographic and clinical variables, such as age, education, use of antiretroviral therapy, pregnancy, and viral suppression, defined as plasma HIV RNA <500 copies/mL. Likewise, we could use these variables to build a predictor of viral suppression. This predictor could rely on parametric logistic regression or more advanced machine learning algorithms, such as Super Learner \citep{SuperLearner, Petersen2015SL}. Now consider the potential impact of pregnancy on clinical outcomes in this population. While optimizing virologic outcomes is essential to preventing mother-to-child-transmission of HIV, the prenatal period could plausibly disrupt or enhance HIV care for a pregnant woman \citep{UNAIDSGap2014}. We can then ask, \textit{"what is the effect of pregnancy on HIV RNA viral suppression among HIV-positive women of child-bearing age in East Africa?"}. While the exposure of pregnancy is not a traditional treatment as commonly considered in a randomized trial, this question is still causal in that we are asking about the outcomes of patients under two different conditions and to answer this question, we must go beyond the observed data set. In particular, causal inference requires an \emph{a-priori} specified set of, often untestable, assumptions about the data generating mechanism. Once we posit a causal model, often encoded in the language of causal graphs, we can express our scientific question in terms of a causal quantity. Under explicit assumptions, we can then translate that causal quantity into a statistical estimand, a function of the observed data distribution. This translation, called \textit{identifiability}, is not guaranteed, as it depends on the underlying scientific question, the structure of the causal model, and the observed data. Lack of identifiability, however, provides us guidance on further data collection efforts and the additional assumptions needed for such translation. Altogether we obtain a statistical estimand that as closely as possible matches the underlying scientific question and thereby ensures that our objective is driving the statistical analysis, as opposed to letting the statistical analysis determine the question asked and answered \citep{Petersen2014roadmap, Hernan2008}. Once the estimand has been specified, we return to realm of statistics and the purely data-driven exercises of point estimation, hypothesis testing, and creating $95\%$ confidence intervals. Interpretation of the resulting values, however, requires us again to consider our causal assumptions In this primer, we review the Causal Roadmap of \cite{Petersen2014roadmap} to (1) specify the scientific question; (2) build an accurate causal model of our knowledge; (3) define the target causal quantity; (4) link the observed data to the causal model; (5) assess identifiability; (6) estimate the resulting statistical parameter; and (7) appropriately interpret the results. This Roadmap borrows the general logic from Descartes's Scientific Method \citep{Descartes1637} and shares a common flow of other causal frameworks \citep{Neyman1923, Rubin1974, Holland1986, Robins1986, Rubin1990, Spirtes93, Pearl2000, Little2000, Dawid2000, Heckman2007, Robins2009, MarkBook, Richardson2013, Hernan2016}. In particular, all approaches demand a clear statement of the research objective, including the target population and interventions of interest \citep{Hernan2018, Ahern2018}. All approaches also provide guidance for conducting a statistical analysis that best answers the motivating question. Unlike some of the other frameworks, however, the Roadmap emphasizes the use of non-parametric or semi-parametric statistical methods, such as targeted maximum likelihood estimation (TMLE), to avoid unwarranted parametric assumptions and harness recent advances in machine learning. As a result this framework has sometimes been called the Targeted Learning Roadmap \citep{MarkBook, Tran2016, Kreif2017}. \section{The Roadmap for Causal Inference} \subsection{Specify the Scientific Question} \label{sec:question} The first step is to specify our scientific question. This helps frame our objective in a more detailed way, while incorporating knowledge about the study. In particular, we need to specify the target population, the exposure, and the outcome of interest. As our running example, we ask, what is the effect of becoming pregnant on HIV RNA viral suppression ($<$500 copies/mL) among HIV-positive women of child-bearing age (15-49 years) in East Africa? This question provides a clear definition of the study variables and objective of our research. It also makes explicit that the study only makes claims about the effect of a specific exposure, outcome, and target population. Any claims outside this context, such as a different exposure, outcome, or target population, represent distinct questions and would require going through the Roadmap again from the start. The temporal cues present in the research question are of particular importance. They represent the putative cause, here pregnancy, and effect of interest, here viral suppression. The temporal cues, together with background knowledge, are frequently used as a basis for specifying the causal model, our next step. \subsection{Specify the Causal Model} \label{sec:causalmodel} One of the appealing features of causal modeling, and perhaps the reason behind its success, is the rich and flexible language for encoding mechanisms underlying a data generating process. Here, we focus on \cite{Pearl2000}'s structural causal models, which unify causal graphs and structural equations \citep{Pearl1988, Goldberger1972, Duncan1975}. Structural causal models formalize our knowledge, however limited, of the study, including the relationships between variables and the role of unmeasured factors. Let us consider again our running example of the impact of pregnancy on HIV viral suppression among women in East Africa. Let $W_1$ denote the set of baseline demographic covariates, such as age, marital status, and education level, and $W_2$ denote the set of pre-exposure HIV care variables, such as prior use of antiretroviral therapy. The exposure $A$ is a binary variable indicating that the woman is known to be pregnant, and the outcome $Y$ is a binary indicator of currently suppressing HIV viral replication: $<$500 copies per mL. These constitute the set of \textit{endogenous} variables, which are denoted $X = \{W_1, W_2,A,Y\}$ and are essential to answering the research question. Each endogenous variable is associated with a latent background factor $U_{W_1}, U_{W_2},U_A,$ and $U_Y$, respectively. The set of background factors are called \textit{exogenous} variables and denoted $U=(U_{W_1}, U_{W_2},U_A,U_Y)$. These variables account for all other unobserved sources that might influence each of the endogenous variables and can share common components. In our example, unmeasured background factors $U$ might include socioeconomic status, the date of HIV infection, the date of conception, her partner's HIV status, and her genetic profile. \paragraph{Causal Graphs:} The ``causal story'' of the data can be conveyed using the language of graphs \citep{Pearl2000, Pearl2016}. Graphical models consist of a set of nodes representing the variables, and a set of directed or undirected edges connecting these nodes. Two nodes are \emph{adjacent} if there exists an edge between them, and a \emph{path} between two nodes $A$ and $B$ is a sequence of adjacent nodes starting from $A$ and ending in $B$. If an edge is \emph{directed} from node $A$ to node $B$, then $A$ is the \emph{parent} of $B$, and $B$ is the \emph{child} of $A$. More generally, for any path starting from node $A$, the set of nodes included in this path are \emph{descendants} of $A$, and $A$ is the \emph{ancestor} of all the nodes included in this set. Here, we are interested in Directed Acyclic Graphs (DAGs), which are fully directed graphs with no path from a given node to itself. DAGs provide a mechanism to explicitly encode our causal assumptions about the underlying data generating process. Specifically, a variable $A$ is a \emph{direct cause} of another variable $B$, if $B$ is the child of $A$ in the causal graph. Also, a variable $A$ is a \emph{cause} of another variable $B$, if $B$ is a descendant of $A$ in the causal graph \citep{Pearl2000}. \begin{figure}[ht] \centering \subfloat[]{{ \input{dag_ex1} }}% \qquad \subfloat[]{{ \includegraphics[width=0.4\linewidth]{alt_dag2.png} }}% \caption{Encoding the underlying causal mechanisms with graphical models. Shaded nodes represent exogenous variables, and unshaded nodes are endogenous variables. Directed edges represent a direct cause between a pair of variables. Double-headed dashed arrows represent potential correlation between the exogenous factors (i.e., unmeasured common causes of the endogenous variables). In (a) we give a directed acyclic graph (DAG) with a single node $U$ representing all the common unmeasured sources. In (b) we provide an alternative representation to make explicit the relationships between the unmeasured background factors $U=\{U_{W_1}, U_{W_2}, U_A,U_Y\}$ and each endogenous variable. } \label{fig:dag1}% \end{figure} To illustrate, Figure~\ref{fig:dag1}(a) provides a DAG corresponding to our running example. From this graph, we can make the following statements: \begin{packed_enum} \item Baseline demographics $W_1$ may affect a woman's pre-exposure HIV care $W_2$, her pregnancy status $A$, and her HIV viral suppression status $Y$. \item Prior care $W_2$ may affect her pregnancy status $A$, and her HIV viral suppression status $Y$. \item Being pregnant $A$ may affect her HIV viral suppression status $Y$. \item Unmeasured factors $U=(U_{W_1}, U_{W_2},U_A,U_Y)$ may affect a woman's baseline characteristics, her prior care, her fertility, and her suppression outcome. \end{packed_enum} In Figure~\ref{fig:dag1}(a), a single node $U$ represents all the common, unmeasured factors that could impact the pre-exposure covariates, the exposure, and the outcome. In an alternative representation in Figure~\ref{fig:dag1}(b), we have explicitly shown each exogenous variable ($U_{W_1}, U_{W_2},U_A,U_Y$) as a separate node and as parent to its corresponding endogenous variable $(W_1,W_2,A,Y)$, respectively. In the latter, dashed double-headed arrows denote correlation between the exogenous factors. Both representations make explicit that there could be unmeasured common causes of the covariates $W=(W_1,W_2)$ and the exposure $A$, the exposure $A$ and the outcome $Y$, and the covariates $W$ and the outcome $Y$. In other words, there is measured and unmeasured confounding present in this study. Altogether, we have avoided many unsubstantiated assumptions about the causal relationships between the variables. This causal model is, thus, non-parametric beyond the assumed time-ordering between variables. Causal graphs can be extended to accommodate more complicated data structures. Suppose, for example, plasma HIV RNA viral levels are missing for some women in our population of interest. We could modify our causal model to account for incomplete measurement \citep{Robins2000, Robins1994, Scharfstein1999, Daniel2012, Mohan2013, Balzer2017CascadeMethods}. Specifically, we redefine the exposure node for pregnancy as $A_1$ and introduce a new intervention node $A_2$ defined as indicator that her viral load is measured. The resulting causal graph is represented in Figure~\ref{fig:dag2}. We refer the readers to \cite{Mohan2013} for detailed discussion of formulating a causal model for the missingness mechanism and to \cite{PetersenCascade2017} for a real world application handling missingness on both HIV status and viral loads \citep{Balzer2017CascadeMethods}. For the remainder of the primer, we assume, for simplicity, there are no missing data and Figure~\ref{fig:dag1} holds. As discussed in the Appendix, other extensions can also be made to account for common complexities, such as longitudinal data and effect mediation. \begin{figure}[ht] \begin{center} \input{dag_ex2} \end{center} \caption{Causal graph extending the running example to account for missingness on the outcome. Along the baseline demographic $W_1$, clinical covariates $W_2$, and suppression outcome $Y$, we now have two intervention nodes $A_1$ for pregnancy status and $A_2$ for measurement of plasma HIV RNA level.} \label{fig:dag2} \end{figure} In the subsequent steps, we discuss how altering the causal graph, particularly by removing edges, is equivalent to making additional assumptions about the data generating process. Before doing so, however, we present the causal model in its structural form. \paragraph{Non-Parametric Structural Equations:} Structural causal models also encode information about the data generating process with a set of non-parametric equations. Like the causal graph, these equations describe how ``nature'' would deterministically generate the variables in our study \citep{Pearl2000,Pearl2016}. Use of the equations can be preferable in longitudinal settings when causal graphs can become unwieldily. Formally, we define a structural causal model, denoted $\mathcal{M}^*$, by the set of exogenous variables $U$, the set of endogenous variables $X$, and a set of functions $\mathcal{F}$ that deterministically assign a value to each variable in $X$, given as input the values of other variables in $X$ and $U$. These non-parametric structural equations allow us to expand our definition of causal assumptions \citep{Pearl2000,Pearl2016}. Variable $A$ is considered to be a \textit{direct cause} of variable $B$, if $A$ appears in the function assigning a value to $B$. Variable $A$ is also a \textit{cause} of variable $B$, if $A$ is direct cause of $B$ or any causes of $B$. In our HIV viral suppression example, the corresponding structural equations are \begin{align} \label{Eq:SCM} W_1 &= f_{W_1}(U_{W_1}) \nonumber \\ W_2 &= f_{W_2}(W_1, U_{W_2}) \\ A &= f_A(W_1, W_2,U_A) \nonumber \\ Y &= f_Y(W_1, W_2,A,U_Y) \nonumber \end{align} where the set of functions $\mathcal{F}=\{f_{W_1}, f_{W_2},f_A,f_Y\}$ encode the mechanism deterministically generating the value of each endogenous variable. The exogenous variables $U=\{U_{W_1}, U_{W_2},U_A,U_Y\}$ have a joint probability distribution $\mathbb{P}_U$ and coupled with the set of structural equations $\mathcal{F}$ give rise to a particular data generating process that is compatible with the causal assumptions implied by $\mathcal{M}^*$. In our example, for a given probability distribution $\mathbb{P}_U$ and set of structural equations $\mathcal{F}$, the structural causal model $\mathcal{M}^*$ describes the following data generating process. For each woman, \begin{packed_enum} \item \textit{Draw the exogenous variables $U$ from the joint probability distribution $\mathbb{P}_U$}. Intuitively, when we sample a woman from the population, we obtain all the unmeasured variables that could influence her baseline covariates, prior care, pregnancy status, and suppression outcome. \item \textit{Generate demographic covariates $W_1$ deterministically using $U_{W_1}$ as input to the function $f_{W_1}$}; the demographic covariates include her age, marital status, education attained, and socioeconomic status. \item \textit{Generate past HIV care covariates $W_2$ deterministically using $U_{W_2}$ and the woman's demographic covariates $W_1$ as input to the function $f_{W_2}$}; the measured clinical factors include history of antiretroviral therapy use and prior HIV suppression status. \item \textit{Generate pregnancy status $A$ deterministically using $U_A$, $W_1$, and $W_2$ as inputs to function $f_A$}. Recall $A$ is an indicator equaling 1 if the woman is known to be pregnant and 0 otherwise. \item \textit{Generate HIV suppression outcome $Y$ deterministically using $U_Y$, $W_1$, $W_2$, and $A$ as inputs to function $f_Y$}. Recall $Y$ is an indicator equaling 1 if her HIV RNA viral level is less than 500 copies per mL and 0 otherwise. \end{packed_enum} It is important to note that the set of structural equations are non-parametric. In other words, the explicit relationship between the system variables, as captured by the set of functions $\mathcal{F}$, are left unspecified. If knowledge is available regarding a relationship of interest, it can be readily incorporated in the structural equations. For instance, in a two-armed randomized trial with equal allocation probability, the function that assigns a value to the exposure variable $A$ can be explicitly encoded as $A = f_A(U_A) = \mathbb{I}(U_A < 0.5)$, where $\mathbb{I}$ is an indicator function and $U_A$ assumed to be drawn from a $Uniform(0,1)$. \subsection{Define the Target Causal Quantity} \label{sec:causalparameter} Once the causal model is specified, we may begin to ask questions of causal nature. The rationale comes from the observation that the structural causal model $\mathcal{M}^*$ is not restricted to the particular setting of our study, but can also describe the same system under changed conditions. The structural equations are \emph{autonomous}, which means that modifying one function does not change another. Therefore, we can make targeted modifications to our causal model to evaluate hypothetical, counterfactual scenarios that would otherwise never be realized, but correspond to our underlying scientific question. In our running example, we are interested in the effect of pregnancy on viral suppression. In the original causal model (Figure~\ref{fig:dag1} and Equation~\ref{Eq:SCM}), a woman's pregnancy status is determined by her baseline demographics $W_1$, prior care status $W_2$, and unmeasured factors $U_A$, such as contraceptive use. However, our objective is to determine the probability of viral suppression if all women in the target population were pregnant versus if the same women over the same time-frame were not pregnant. The autonomy of the structural equations allows us to modify the way in which the exposure, here pregnancy, is determined. In particular, we can intervene on the exposure $A$ to deterministically set $A = 1$ in one scenario, and then set $A = 0$ in another, while keeping the other equations constant. \begin{figure}[ht] \begin{center} \input{mut_dag_ex1} \end{center} \caption{Causal graph after intervention on the exposure, pregnancy status, to set $A=a$. Since this is done deterministically and independently of other variables in the system, the only node $causing$ a change in $A$ is the intervention node $a \in \{0,1\}$. } \label{fig:mut_dag1} \end{figure} The post-intervention causal graph is given in Figure~\ref{fig:mut_dag1} and the structural equations become \begin{align*} W_1 = f_{W_1}(U_{W_1}) &\quad \quad \quad W_1 = f_{W_1}(U_{W_1}) \\ W_2 = f_{W_2}(W_1, U_{W_2}) &\quad \quad \quad W_2 = f_{W_2}(W_1, U_{W_2}) \\ A = 1 &\quad \quad \quad A = 0 \\ Y(1) = f_Y(W_1, W_2, 1,U_Y) &\quad \quad \quad Y(0) = f_Y(W_1, W_2,0,U_Y) \end{align*} These interventions generate counterfactual outcomes $Y(a)$ for $a\in \{0,1\}$, whose distribution is denoted $\mathbb{P}^*$. These causal quantities are indicators that a participant would have suppressed viral replication, if possibly contrary to fact, her pregnancy status were $A=a$ In this case, it is both physically impossible and unethical to design a randomized trial for pregnancy. In other words, we cannot directly intervene on a woman's pregnancy status. Likewise in Figure~\ref{fig:dag2}, enforcing measurement of the outcome, which translates into setting $A_2=1$, is impossible. While neither intervention is plausible, we believe counterfactuals provide a language to express many questions in Data Science in a mathematically tractable way. Nonetheless, we note that there has been an lively debate about defining and interpreting analyses of variables on which one cannot directly intervene \citep{Pearl1995, Hernan2005, vanderLaan2005resp, Petersen2014roadmap, Hernan2016}. Given the counterfactual outcomes and their distribution $\mathbb{P}^*$, we can express our scientific question as a mathematical quantity. One common choice is the Average Treatment Effect (ATE): \begin{equation} \Psi^*(\mathbb{P}^*) := \mathbb{E}^*[Y(1)-Y(0)], \end{equation} where the expectation is taken with respect to $\mathbb{P}^*$. Since the causal model $\mathcal{M}^*$ provides the set of possible probability distributions for the exogenous and endogenous factors $(U,X)$ and thus the counterfactual outcomes $(Y(1),Y(0))$, $\Psi^*$ is a mapping from $\mathcal{M}^*$ to the real numbers. The target causal parameter $\Psi^*(\mathbb{P}^*)$ represents the difference in the expected counterfactual outcome if all units in the target population were exposed and the expected counterfactual outcome if the same units were not exposed. For the running example, $\Psi^*(\mathbb{P}^*)$ can be interpreted as the difference in the counterfactual probability of viral suppression if all women in the target population were pregnant versus if all women were not. Before discussing how these causal quantities can be identified from the observed data distribution, we emphasize that for simplicity we have focused on a binary intervention, occurring deterministically at a single time point. Scientific questions corresponding to categorical, continuous, stochastic, and longitudinal exposures are also encompassed in this framework, but beyond the scope of this primer and are briefly discussed in the Appendix. We also note that other summaries, such as relative measures, the sample average effect, or marginal structural models, may better capture the researcher's scientific question. \subsection{Link the Observed Data to the Causal Model} \label{Sec:link} Thus far, we have defined our scientific question, specified a structural causal model $\mathcal{M}^*$ to represent our knowledge of the data generating process, intervened on that causal model to generate counterfactual outcomes, and used these counterfactuals to express our scientific question as a causal quantity. The next step is to provide an explicit link between the observed data and the specified structural causal model. Returning to our running example, suppose we have a simple random sample of $N$ women from our target population. On each woman, we measure her baseline demographics $W_1$, prior HIV care $W_2$, pregnancy status $A$, and suppression outcome $Y$. These measurements constitute our observed data for each woman in our sample: $O=\{W_1, W_2,A,Y\}$. Therefore, we have $N$ independent, identically distributed copies of $O$, which are drawn from some probability distribution $\mathbb{P}$. Other sampling schemes, such as case-control, are accommodated by this framework, but are beyond the scope of this primer. If we believe that our causal model accurately describes the data generating process, we can assume that the observed data are generated by sampling repeatedly from a distribution compatible with the structural causal model. In other words, the structural causal model $\mathcal{M}^*$ provides a description of the study under existing conditions (i.e., the real world) and under specific intervention (i.e., the counterfactual world). As a result, the observed outcome $Y$ equals the counterfactual outcome $Y(a)$ when the observed exposure $A$ equals the exposure of interest, $a$; this is commonly called the \textit{consistency assumption}. In our example, all the endogenous variables are observed: $X=O$; therefore, we can write \begin{equation} \mathbb{P}(O=o) = \sum_{u} \mathbb{P}^*(X=x|U=u)\mathbb{P}^*(U=u), \end{equation} where an integral replaces the summation for continuous variables. This, however, might not always be the case. Suppose, for example, we only measured demographics, pregnancy status, and viral suppression, but failed to measure variables related to prior HIV care. Then the observed data would be $O=(W_1,A,Y)$ and are a subset of all the endogenous variables $X$. In either case, we see that the structural causal model $\mathcal{M}^*$, defined as the collection of all possible joint distributions of the exogenous and endogenous variables $(U,X)$, implies the statistical model $\mathcal{M}$, defined as the collection of all possible joint distributions for the observed data $O$. The structural causal model $\mathcal{M}^*$ rarely implies restrictions on the resulting statistical model $\mathcal{M}$, which is thereby often non-parametric. An important exception is a completely randomized trial, where the unmeasured factors determining the treatment assignment $U_A$ are independent of the others and results in a semi-parametric statistical model. The \emph{D-separation} criteria of \cite{Pearl2000} can be used to evaluate what statistical assumptions, if any, are implied by the causal model. The true observed data distribution $\mathbb{P}$ is an element of the statistical model $\mathcal{M}$. \subsection{Assessing Identifiability} \label{sec:ident} In the previous section, we established a bridge between our structural causal model $\mathcal{M}^*$ and our statistical model $\mathcal{M}$. However, we have not yet discussed the conditions under which causal assumptions and observed data can be combined to answer causal questions. Structural causal models provide one way to assess the assumptions needed to express our target causal quantity as a statistical estimand, which is a well-defined function of the observed data distribution $\mathbb{P}$. Recall in Section~\ref{sec:causalparameter} that we defined our target causal parameter as the average treatment effect $\Psi^* (\mathbb{P}^*)= \mathbb{E}^*[Y(1) - Y(0)]$: the difference in the expected viral suppression status if all women were pregnant versus if none were. If given a causal model and its link to the observed data, the target causal parameter can be expressed as a function of the observed data distribution $\mathbb{P}$, then the causal parameter is called $identifiable$. If not, we can still explicitly state and evaluate the assumptions needed to render the target causal parameter identifiable from the observed data distribution. One of the main tools for assessing identifiability of causal quantities is a set of criteria based on causal graphs. In general, these criteria provide a systematic approach to identify an appropriate adjustment set. Here, we focus on identifiability for the effect of a single intervention at one time, sometimes called ``point-treatment effects''. For these problems, we first present the back-door criterion and the front-door criterion. For a detailed presentation of graphical methods for assessing identifiability in causal graphs, the reader is referred to \cite{Pearl2000, Pearl2016}. Formally, we say that a path is \emph{blocked} if at least one variable in that path is conditioned on, and we define a \emph{back-door path} from a given node $A$ as any path that contains an arrow into node $A$. Then, given any pair of variables ($A,B$), where $A$ occurs before $B$ in a directed acyclic graph, a set of variables $C$ is said to satisfy the \emph{back-door criterion} with respect to ($A,B$) if (1) the descendants of $A$ do not include any node in $C$, and (2) $C$ blocks every back-door path from $A$ to $B$ . The rationale behind this criterion is that, for $C$ to be the appropriate adjustment set that isolates the causal effect of $A$ on $B$, we must block all spurious paths between $A$ and $B$, and leave directed paths from $A$ to $B$ unblocked. This criterion does not, however, cover all possible graph structures. Alternatively, a set of variables $C$ satisfies the \emph{front-door criterion} with respect to a pair of variables ($A,B$) if (1) all directed paths from $A$ to $B$ are blocked by $C$, (2) all paths from $A$ to $C$ are blocked, and (3) all paths from $C$ to $B$ containing an arrow into $C$ are blocked by $A$. We note that the front-door criterion is more involved than its back-door counterpart, in the sense that it requires more stringent conditions to hold for a given adjustment set to satisfy identifiability. In practice, it is often the case that the back-door criterion is enough to identify the needed adjustment set, especially in point-treatment settings. When the back-door criterion holds, the observed association between the exposure and outcome can be attributed to the causal effect of interest, as opposed to spurious sources of correlation. In our running example, the set of baseline covariates $W=(W_1, W_2)$ will satisfy the back-door criterion with respect to the effect of pregnancy $A$ on HIV viral suppression $Y$, if the following two conditions hold: \begin{packed_enum} \item No node in $W$ is a descendant of $A$. \item All back-door paths from $A$ to $Y$ are blocked by $W$. \end{packed_enum} Looking at the posited causal graph from Figure \ref{fig:dag1}(a), we see that the first condition holds, but the second is violated. There exists a back-door path from $A$ to $Y$ through the unmeasured background factors $U$. Intuitively, the unmeasured common causes of pregnancy and HIV viral suppression obstruct our isolation of the causal effect of interest and thus ``confound'' our analyses. Therefore, our target causal quantity is not identifiable in the original causal model $\mathcal{M}^*$. Nonetheless, we can explicitly state and consider the plausibility of the causal assumptions needed for identifiability. In particular, the following independence assumptions are sufficient to satisfy the back-door criterion and thus identify the causal effect in this point-treatment setting. \begin{packed_enum} \item There must not be any unmeasured common causes of the exposure and the outcome: $U_A \rotatebox[origin=c]{90}{$\models$} U_Y$ $and$, \begin{packed_enum} \item There must not be any unmeasured common causes of the exposure and the baseline covariates: $U_A \rotatebox[origin=c]{90}{$\models$} U_{W_1}$ and $U_A \rotatebox[origin=c]{90}{$\models$} U_{W_2}$ \item[$or$] \item There must not be any unmeasured common causes of the baseline covariates and the outcome: $U_{W_1} \rotatebox[origin=c]{90}{$\models$} U_Y$ and $U_{W_2} \rotatebox[origin=c]{90}{$\models$} U_Y$. \end{packed_enum} \end{packed_enum} These criteria are reflected in the causal graphs shown in Figure~\ref{fig:ident_dag}. In the running example, assumption 1.a states that there are no unmeasured common causes of pregnancy status and demographic or clinical factors, while 1.b assumes that there are no unmeasured common causes of viral suppression and demographic or clinical factors. The independence assumptions in 1.a hold by design in a stratified, randomized trial, where the unmeasured factors determining the exposure assignment are independent of all other unmeasured factors. As a result, these independence assumptions (1.a and/or 1.b) are sometimes called the \emph{randomization assumption} and equivalently expressed as $Y(a) \rotatebox[origin=c]{90}{$\models$} A \mid W$. These assumptions are also referred to as ``unconfoundedness'', ``selection on observables'', and ``conditional exchangeability'' \citep{Robins1986}. \begin{figure}[ht]% \centering \subfloat[]{{ \input{ident1_dag} }}% \qquad \subfloat[]{{ \input{ident2_dag} }}% \caption{Causal graphs corresponding the identifiability assumptions (1.a) and (1.b), respectively. Here, we have explicitly shown that the unmeasured factors contributing to the exposure $U_A$ in (a) and the outcome $U_Y$ in (b) are independent of the others.} \label{fig:ident_dag}% \end{figure} With these assumptions, we can express the distribution of counterfactual outcomes in terms of the distribution of the observed data: \begin{align*} \mathbb{P}^*(Y(a)) &= \sum_{w} \mathbb{P}^*(Y(a) | W=w) \mathbb{P}^*(W=w) \\ &= \sum_{w} \mathbb{P}^*(Y(a) | A=a, W=w) \mathbb{P}^*(W=w) \\ &= \sum_{w} \mathbb{P}(Y | A=a, W=w) \mathbb{P}(W=w) \end{align*} where $W=(W_1, W_2)$ denotes the pre-exposure covariates, including both demographic and clinical factors, and where the summation generalizes to an integral for continuous covariates here and in all subsequent expressions. The first equality is by the law of iterated expectations. The second equality holds by the randomization assumption, and the final by the established link between the causal and statistical model (Section~\ref{Sec:link}). Under these assumptions, we can express the average treatment effect $\Psi^*(\mathbb{P}^*)= \mathbb{E}^*[Y(1) - Y(0)]$, as a statistical estimand, often called the \emph{G-computation identifiability result} \citep{Robins1986}: \begin{equation} \label{stat-estimand} \Psi(\mathbb{P}) := \sum_{w} \big[ \mathbb{E}(Y|A=1,W=w) - \mathbb{E}(Y|A=0,W=w) \big] \mathbb{P}(W=w) \end{equation} Thus, our statistical target is the difference in the expected outcome, given the exposure and covariates, and the expected outcome, given no exposure and covariates, averaged with respect to the distribution of the baseline covariates $W$. In our example, $\Psi(\mathbb{P})$ is the difference in the probability of viral suppression, between pregnant and non-pregnant women with the same values of the covariates, standardized with respect to the covariate distribution in the population. The same quantity can be expressed in inverse probability weighting form: \begin{equation}\label{ipw-estimand} \Psi(\mathbb{P}) := \mathbb{E}\left[ \left( \frac{\mathbb{I}(A=1)}{\mathbb{P}(A=1\mid W)} - \frac{\mathbb{I}(A=0)}{\mathbb{P}(A=0\mid W)}\right) Y \right] \\ \end{equation} The latter representation highlights an additional data support condition, known as \emph{positivity}: \begin{equation*} min_{a\in \mathcal{A}}\ \mathbb{P}(A=a|W=w) > 0, \ \text{for all $w$ such that } \mathbb{P}(W=w)>0. \end{equation*} Each exposure level of interest must occur with a positive probability within the strata of the discrete-valued adjustment set $W$. This assumption is also called ``overlap'' and ``the experimental treatment assignment assumption''. We refer the reader to \cite{Petersen2012} for a discussion of this assumption and approaches when it is theoretically or practically violated. For continuous covariates, it is not straightforward to evaluate the positivity assumption. One commonly used approach is to transform these continuous variables into categories or quantiles, and assess positivity violations as with categorical variables \citep{Cole2008, Messer2010}. Overall, the identifiability step is essential to specifying the needed adjustment set, and thereby statistical estimand to link our causal effect of interest to some function of the observed data distribution. Above, we focused on a simple point-treatment setting with measured and unmeasured confounding, but without mediation, biased sampling, or missing data. In more realistic settings, there many other sources of association between our exposure and outcome, including selection bias, direct and indirect effects, and the common statistical paradoxes of Berkson's bias and Simpson's Paradox \citep{Hernan2004, HernandezDiaz2006}. Furthermore, in the setting of longitudinal exposures with time-dependent confounding, the needed adjustment set may not be intuitive and the short-comings of traditional approaches become more pronounced \citep{Robins1986, Robins2000, Robins2009, Pearl2016}. Indeed, methods to distinguish between correlation and causation are crucial in the era of ``Big Data'', where the number of variables is growing with increasing volume, variety, and velocity \citep{Rose2012problems, Marcus2014problems, Balzer2016BigData}. Nonetheless, it is important to note that specifying a causal model (Section~\ref{sec:causalmodel}) does not guarantee the identification of a causal effect. Causal frameworks do, however, provide insight into the limitations and full extent of the questions that can be answered given the data at hand. They further facilitate the discussion of modifications to the study design, the measurement additional variables, and sensitivity analyses \citep{Robins1999sensitivity, Imai2010, Vanderweele2011sensitivity, Diaz2012Sensitivity}. In fact, even if the causal effect is not identifiable (e.g., Figure~\ref{fig:dag1}), the Causal Roadmap still provides us with a statistical estimand (e.g., Equation~\ref{stat-estimand}) that comes as close as possible to the causal effect of interest given the limitations in the observed dataset. In the next sections, we discuss estimation of this statistical parameter and use identifiability results, or lack there of, to inform the strength of our interpretations. \subsection{Estimate the Target Statistical Parameters} Once the statistical model and estimand have been defined, the Causal Roadmap returns to traditional statistical inference to estimate functions of a given observed data distribution. Here, we focus on estimators based on the G-computation identifiability result $\Psi(\mathbb{P})$. Popular methods for estimation and inference for $\Psi(\mathbb{P})$, which would equal the average treatment effect if the identifiability assumptions held, include parametric G-computation, Inverse Probability Weighting (IPW), and Targeted Maximum Likelihood Estimation (TMLE) \citep{Robins1986, Horvitz1952, Rosenbaum83, vanderLaan2006, MarkBook}. Below we briefly outline the implementation of each estimator and refer the reader to \cite{PetersenIntroCI} for worked \texttt{R} code for each algorithm. We emphasize that while each algorithm is targeting a causally motivated statistical estimand, these algorithms are not directly estimating causal effects, and therefore it is a misnomer to call them ``causal estimators''. \paragraph{Parametric G-computation} is an algorithm that simply estimates the quantities needed to calculate the statistical estimand defined in Equation~\eqref{stat-estimand} and then substitutes those quantities into the G-computation formula \citep{Robins1986, Taubman2009, Young2011, Westreich2012, Zhang2018}. As a result, this algorithm is sometimes called the \emph{simple substitution estimator} and is implemented with the following steps. \begin{packed_enum} \item Regress the outcome on the exposure and covariate adjustment set to estimate the conditional expectation $\mathbb{E}(Y|A,W)$. \item Based on the estimates from the Step 1, generate the predicted outcomes for each individual in the sample while deterministically setting the value of the exposure to the levels of interest, but keeping the covariates the same: \begin{equation*} \hat{\mathbb{E}}(Y_i | A_i=1, W_i) \text{ and } \hat{\mathbb{E}}(Y_i | A_i=0, W_i) \text{ for all observations} \ i=1, ..., N. \end{equation*} For a binary outcome this step is corresponds to generating the predicted probabilities $\hat{\mathbb{P}}(Y_i=1 | A_i=a, W_i)$ for exposure levels $a \in\{0,1\}$. \item Obtain a point estimate by taking a sample average of the difference in the predicted outcomes from Step 2: \begin{equation*} \hat{\Psi}_{Gcomp}(\hat{\mathbb{P}}) = \frac{1}{N} \sum_{i=1}^N \big[\hat{\mathbb{E}}(Y_i | A_i=1, W_i) -\hat{\mathbb{E}}(Y_i | A_i=0, W_i) \big] \end{equation*} where $\hat{\mathbb{P}}$ denotes the empirical distribution, which is the non-parametric maximum likelihood estimator of the covariate distribution $\mathbb{P}(W)$. The sample proportion is a simple non-parametric estimator of $\mathbb{P}(W=w)$: $\frac{1}{N} \sum_i \mathbb{I}(W_i=w)$. \end{packed_enum} \paragraph{Inverse Probability Weighting (IPW)} is an estimator based on an alternative form of the G-computation identifiability result defined in Equation~\eqref{ipw-estimand} \citep{Horvitz1952, Rosenbaum83, Robins2000, Bodnar2004, Cole2008}. In this form, the statistical estimand is a function of the conditional probability of being exposed, given the adjustment covariates $\mathbb{P}(A=1|W)$, which is often called the \emph{propensity score} \citep{Rosenbaum83}. IPW controls for confounding by up-weighting rare exposure-covariate subgroups, which have a small propensity score, and down-weighting more common subgroups, which have a larger propensity score. The IPW estimator is implemented with the following steps. \begin{packed_enum} \item Regress the exposure on the covariate adjustment set to estimate the propensity score $\mathbb{P}(A=1|W)$. \item Based on the estimates from Step 1, predict each individual's probability of receiving her observed exposure, given the adjustment covariates: \begin{equation*} \hat{\mathbb{P}}(A_i | W_i) \text{ for all observations} \ i=1, ..., N. \end{equation*} \item Obtain a point estimate by taking the empirical mean of the outcome weighted by the inverse of the conditional exposure probabilities: \begin{equation*} \hat{\Psi}_{IPW}(\hat{\mathbb{P}}) = \frac{1}{N} \sum_{i=1}^N \left(\frac{\mathbb{I}(A_i=1)}{\hat{\mathbb{P}}(A_i=1 | W_i)} -\frac{\mathbb{I}(A_i=0)}{\hat{\mathbb{P}}(A_i=0 | W_i)} \right) Y_i. \end{equation*} Thus, individuals who are exposed receive weight as one over the estimated propensity score $\hat{\mathbb{P}}(A_i=1 | W_i)$, while individuals who are not exposed receive weight as negative one over the estimated probability of not being exposed, given the covariates $\hat{\mathbb{P}}(A_i=0 | W_i)$. \end{packed_enum} The performance of the parametric G-computation depends on consistent estimation of the conditional expectation of the outcome, given the exposure and covariates $\mathbb{E}(Y|A,W)$, and the performance of IPW relies on consistent estimation of the propensity score $\mathbb{P}(A=1|W)$. Traditionally, both estimators have relied on parametric regression models to estimate these quantities. If sufficient background knowledge is available to support using such a regression, it should have already been encoded in the causal model, yielding parametric structural equations in Section~\ref{sec:causalmodel}, and can be incorporated during estimation. However, in most real-world studies with a large number of covariates and potentially complicated relationships, we usually do not have the knowledge support using such parametric regressions. More often, our statistical model $\mathcal{M}$ for the set of possible distributions of the observed data is non-parametric or semi-parametric (Section~\ref{Sec:link}). Furthermore, we want to avoid introducing new and unsubstantiated assumptions during estimation. Reliance on poorly specified parametric regressions can result in biased point estimates and misleading inference (e.g., \cite{Benkeser2017, LuqueF2018}). At the same time, non-parametric methods, such as stratification, will break down due to sparsity. Here, recent advances in machine learning can help us estimate $\mathbb{E}(Y|A,W)$ and $\mathbb{P}(A=1|W)$ without introducing new assumptions. \paragraph{Data-adaptive estimation} or machine learning techniques can be used to effectively estimate the \emph{nuisance parameters}, which are the quantities needed to compute our statistical estimand: $\mathbb{E}(Y|A,W)$ and $\mathbb{P}(A=1|W)$. We focus our discussion on \emph{ensemble learning methods}, which ``stack'' or combine several prediction algorithms together and can be implemented as follows \citep{Wolpert92, Breiman96}. First, we pre-specify a library of candidate algorithms, such as generalized linear models, splines, random forests, neural networks, or support vector machines. We also define a measure of performance through an appropriate loss function, such as the mean squared error or the negative log-likelihood. Next, we randomly split the observed data into training and validation sets to assess the performance of each algorithm in the library. We then fit each algorithm using only data from the training set and predict the outcomes for the units in validation set. Each algorithm's performance is quantified by deviations, corresponding to the loss function, between the actual and predicted outcomes for the units in the validation set. Repeating the process $V$ times amounts to performing \emph{V-fold cross-validation}. We could then select the algorithm with the best performance, corresponding to the smallest cross-validated \emph{risk} estimate. This procedure, called Discrete Super Learner \citep{SuperLearner}, effectively sets up a competition between the algorithms specified in the library, and selects the one with the best performance. It naturally follows then that Discrete Super Learner can only perform as well as the best performing algorithm specified in its library. The full Super Learner algorithm improves upon its discrete version by taking a weighted combination of the algorithm-specific predictions to create a new prediction algorithm. We refer the reader to \cite{Chpt3} for further discussion of Super Learner and its properties and to \cite{Naimi2018} for worked examples and \texttt{R} code. The algorithm is available in the \texttt{SuperLearner} package in \texttt{R} \citep{SLRpkg}. The goal of Super Learner is to do the best possible job, according to the specified loss function, of predicting the outcome $Y$, given the exposure $A$ and covariates $W$, or predicting the exposure $A$, given the covariates $W$. As a result, Super Learner-based estimators of the nuisance parameters $\mathbb{E}(Y|A,W)$ or $\mathbb{P}(A=1|W)$ have the wrong bias-variance tradeoff for our statistical estimand $\Psi(\mathbb{P})$, which is a single number as opposed to a whole prediction function. TMLE, discussed next, provides one way to integrate data-adaptive algorithms, such as Super Learner, and still obtain the best possible bias-variance tradeoff for the statistical estimand of interest. Indeed, a particular appeal of the Targeted Learning framework is the use of flexible estimation methods to respect the statistical model, which is often non-parametric, and to minimize the risk of bias due to regression model misspecification. \paragraph{Targeted Maximum Likelihood Estimation (TMLE)} provides a general approach to constructing semi-parametric, efficient, substitution estimators \citep{vanderLaan2006, MarkBook, Petersen2014ltmle}. Here, we provide a very brief overview and refer the reader to \cite{Schuler2017} for a thorough introduction to the algorithm, which is available in the \texttt{tmle}, \texttt{ltmle}, and \texttt{drtmle} packages in \texttt{R} \citep{tmlePackage, ltmlepackage, Benkeser2017}. To implement TMLE for the G-computation identifiability result $\Psi(\mathbb{P})$, given in Equation~\ref{stat-estimand}, we take the following steps. First, we use Super Learner provide an initial estimator of the conditional mean outcome, given the exposure and covariates $\mathbb{\hat{E}}^0(Y|A,W)$. Next, we ``target'' this initial estimator using information from the propensity score $\hat{\mathbb{P}}(A=1|W)$, also estimated with Super Learner. Informally, this targeting step can be thought of as a second chance to control confounding and serves to reduce statistical bias for the $\Psi(\mathbb{P})$. We denote the updated estimator of the conditional mean outcome as $\mathbb{\hat{E}}^1(Y|A,W)$ and use it to obtain targeted predictions of the outcome setting the exposures of interest, but keeping the covariates the same: $\mathbb{\hat{E}}^1(Y_i | A_i=1,W_i)$ and $\mathbb{\hat{E}}^1(Y_i | A_i=0,W_i)$ for all observations $i=1,\ldots, N$. Finally, we obtain a point estimate by taking the average difference in these targeted predictions. \begin{equation*} \hat{\Psi}_{TMLE}(\hat{\mathbb{P}}) = \frac{1}{N} \sum_{i=1}^N \big[ \hat{\mathbb{E}}^1 (Y_i|A_i=1,W_i) - \hat{\mathbb{E}}^1 (Y_i|A_i=0,W_i) \big]. \end{equation*} TMLE's updating step also serves to endow the algorithm with a number of theoretical properties, which often translate into superior performance in finite samples. First, under regularity and empirical process conditions detailed in \cite{MarkBook}, TMLE follows the Central Limit Theorem and thus the normal distribution can be used for constructing 95\% confidence intervals and hypothesis testing, even if machine learning is used for estimation of the nuisance parameters $\mathbb{E}(Y|A,W)$ or $\mathbb{P}(A=1|W)$. Furthermore, the estimator is \emph{double robust} in that it will be consistent if either $\mathbb{E}(Y|A,W)$ or $\mathbb{P}(A=1|W)$ is consistently estimated. Collaborative TMLE further improves upon this robustness result \citep{vanderLaan2010ctmle, Gruber2015ctmle}. Finally, if both nuisance parameters are estimated consistently and at fast enough rates, the estimator will be locally efficient and in large samples attain the minimal variance in a semi-parametric statistical model. We refer the reader to \cite{Kennedy2017} for an introduction to semi-parametric, efficiency theory. Finally, we note that there is nothing inherent in the TMLE algorithm that demands the use of Super Learner. However, its implementation with machine learning algorithms avoids introducing new unsubstantiated assumptions during estimation and improve our chances for consistent results. Again, relying on misspecified parametric regressions can induce statistical bias and yield misleading statistical inference. \subsection{Interpretation of Results} The final step of the Roadmap is to interpret our results. We have seen that the causal inference framework clearly delineates the assumptions made from domain knowledge (Section~\ref{sec:causalmodel}) from the ones desired for identifiability (Section~\ref{sec:ident}). In other words, this framework ensures that the assumptions needed to augment the statistical results with a causal interpretation are made explicit. In this regard, \cite{Petersen2014roadmap} argue for a hierarchy of interpretations with ``increasing strength of assumptions''. First, we always have a statistical interpretation as an estimate of the difference in the expected outcome between exposed and unexposed units with the same covariate values, standardized over the covariate distribution in the population. We can also interpret $\hat{\Psi}(\hat{\mathbb{P}})$ as an estimate of the marginal difference in the expected outcome associated with the exposure, after controlling for measured confounding. To interpret our estimates causally, we need the identifiability assumptions (Section~\ref{sec:ident}) to hold in the original causal model (Section~\ref{sec:causalmodel}). If either graphs in Figure~\ref{fig:ident_dag} represented the true causal structure that generated our data and the positivity assumption held, then we could interpret $\hat{\Psi}(\hat{\mathbb{P}})$ as the average treatment effect or for a binary outcome the causal risk difference. Now, recall that the counterfactual outcomes were derived through intervening on the causal model (Section~\ref{sec:causalparameter}). The selected intervention should match our underlying scientific question (Section~\ref{sec:question}) and does not have to correspond to a feasible or realistic intervention. If the identifiability assumptions (Section~\ref{sec:ident}) held and the intervention could be conceivably implemented in the real world, then we could further interpret $\hat{\Psi}(\hat{\mathbb{P}})$ as an estimate of the intervention's impact if it had been implemented in the population of interest. Finally, if the identifiability assumptions were met and the intervention implemented perfectly in a study sample, whose characteristics exactly matched those of our population and who were fully measured, then we could interpret $\hat{\Psi}(\hat{\mathbb{P}})$ as replicating the results of the randomized trial of interest. We note this hierarchy represents a divergence from the Target Trial framework of \cite{Hernan2006b}, who suggest causal inference with observational data can be thought of as ``emulating'' a randomized trial. In our running example, the causal model shown in Figure~\ref{fig:dag1} represents our knowledge of the data generating process; there are measured ($W_1,W_2$) as well as unmeasured $U$ common causes of the exposure $A$ and the outcome $Y$. Thus, the lack of identifiability prevents any interpretation as a causal effect or further along the hierarchy. Thus, we can interpret a point estimate of $\Psi(\mathbb{P})$ as the difference in the probability of HIV RNA viral suppression associated with pregnancy after controlling for the measured demographic and clinical confounders. \section{Conclusion} The objective of statistical analyses is to make inferences about the data generating process underlying a randomized trial or an observational study. In practice, statistical inference is concerned with purely data-driven tasks, such as prediction, estimation and hypothesis testing. In recent decades, the advent of causal inference has triggered a shift in focus, particularly within the data analysis community, toward a territory that has traditionally evaded statistical reach: the causal mechanism underlying a data generating process. Statistical inference relies on patterns present in the observed data, such as correlation, and therefore is unable, alone, to answer questions of causal nature \citep{Pearl2010, Pearl2016}. Nonetheless, questions about cause and effect are of prime importance in all fields including Data Science \citep{Pearl2018, Hernan2018b}. We have presented an overview of one framework for causal inference. We emphasized how the Causal Roadmap helps ensure consistency and transparency between the imperfect nature of real world data, and the complexity associated with questions of causal nature. Of course, this work serves only as a primer to causal inference in Data Science, and we have only presented the fundamental concepts and tools in the causal inference arsenal Indeed, this framework can be extended to richer and more complicated questions. For instance, our running example for average treatment effect only focused on a single exposure at a single time point. However, as demonstrated in \cite{Tran2016, Kreif2017}, the Causal Roadmap can also handle multiple intervention nodes with time-dependent confounding. Other recent avenues of research in causal inference are discussed in the Appendix. As a final note, a Data Scientist may debate the usefulness of applying the causal inference machinery to her own research. We hope to have clarified that if appropriately followed, the Causal Roadmap forces us to think carefully about the goal of our research, the context in which data were collected, and to explicitly define and justify any assumptions. It is our belief that conforming to the rigors of this causal inference framework will improve the quality and reproducibility of all scientific endeavors that rely on real data to understand how nature works. \section*{Appendix} Here, we briefly highlight some extensions to more advanced settings. For each, we provide a broad definition and a few examples with citations to some relevant works. \begin{enumerate} \item \textbf{Marginal structural models} provide a summary of how the distribution of the counterfactual outcome changes as a function of the exposure and possibly pre-exposure covariates \citep{Robins1999, Robins2000, Bodnar2004, Neugebauer2007, Robins2009, Chpt24, Zheng2016MSM}. Marginal structural models are another way to define our target causal parameter and especially useful when the exposure is continuous or has many levels. \\ \textit{Examples:} \cite{Robins2000} specified a logistic regression model to summarize the dose-response relation for the cumulative effect of zidovudine (AZT) treatment on the counterfactual risk of having undetectable HIV RNA levels among HIV-positive patients. For a time-to-event outcome, \cite{Cole2012} used a Cox proportional hazard model to summarize the association between treatment initiation and the counterfactual hazard of incident AIDS or death among persons living with HIV. \item \textbf{Longitudinal exposures}, corresponding to interventions on multiple treatment nodes, allow us to assess the cumulative effect of an exposure or exposures over time \citep{Robins2000, Bang&Robins05, Robins2009, Chpt24, vanderLaan2012towertmle, Westreich2012, Petersen2014ltmle}. Examining the effects of longitudinal exposures is complicated by time-dependent confounding, when a covariate is affected by a prior treatment and confounds a future treatment. In these settings, causal frameworks have been especially useful for identifying the appropriate adjustment sets and thereby statistical analysis.\\ \textit{Examples:} \cite{Schnitzer2014} sought to assess the effect of breastfeeding duration on gastrointestinal infections among new borns, while \cite{Decker2014} investigated the effects of sustained physical activity and diet interventions on adolescent obesity. \item \textbf{Effect mediation} refers to a general class of causal questions seeking to distinguish an exposure's direct effect on the outcome from its indirect effect through an intermediate variable \citep{Robins1992, Pearl2001, Petersen2006b, vanderLaan2008DE, VanderWeele2009, Imai2010, Zheng2012, Tran2016}. There are several types of direct and indirect effects. For example, the controlled direct effect refers to the contrast between the expected counterfactual outcomes under two levels of the exposure, but when the mediator is fixed at a constant level. The natural direct effect, also called the pure direct effect, refers to the contrast between the expected counterfactual outcomes under two levels of the exposure, but when the mediator remains at its counterfactual level under the reference value of the exposure. Indirect effects can be defined analogously. \\ \textit{Examples:} \cite{Naimi2016} examined the disparity in infant mortality due to race that would remain if all mothers breastfeed prior to hospital discharge. More recently, \cite{Rudolph2018} investigated how the impact of neighborhood disadvantage on adolescent substance use was mediated by school and peer environment. \item \textbf{Dynamic treatment regimes} are personalized rules for assigning the exposure or treatment as a function of an individual's covariate history \citep{Murphy2003, Hernan2006dynamic, vanderLaan2007indvtxt, Kitahata2009, Hernan2009b, Cain2010, Kreif2017}. They are also called ``adaptive treatment strategies'' and ``individualized treatment rules''. Static interventions, which assign a single level of the exposure to all individuals regardless of their covariate values, can be considered a special case of dynamic interventions.\\ \textit{Examples:} \cite{Cain2010} and \cite{Young2011} both considered CD4-based thresholds for initiating antiretroviral therapy and their impact on mortality among persons living with HIV. Recently, \cite{Kreif2017} compared static and dynamic regimes to understand the optimal timing and level of nutritional support for children in a pediatric intensive care unit. \item \textbf{Stochastic interventions} aim to change or shift the distribution of the exposure \citep{Korb2004, Taubman2009, Cain2010, Diaz2012, Diaz2013, Rudolph2017}. Stochastic interventions are especially useful when the exposure of interest can not be directly manipulated and can help alleviate violations to the positivity assumption. Deterministic interventions, which assign a given level of the exposure with probability one, can be considered a special case of stochastic interventions. \\ \textit{Examples}: \cite{Diaz2012} asked what is the impact of a policy encouraging more exercise, according to health and socioeconomic factors, on mortality in a population of older adults? \cite{Danaei2013} examined the impact of various lifestyle interventions, such as eating at least 2 servings of whole grain per day, on the risk of type 2 diabetes in women. \item \textbf{Clustered data} occur when there is dependence or correlation between individuals within some grouping, such as a clinic, school, neighborhood, or community. Such correlation can arise from shared cluster-level factors, including the exposure, and from social or biological interactions between individuals with a cluster \citep{Halloran1991, Halloran1995, Oakes2004, Tchetgen2012, vanderLaan2014Networks, Schnitzer2014, Prague2016, Balzer2018Hierarchical, Morozova2018, Buchanan2018}. This dependence must be accounted when specifying the causal model and often demands relaxing the stable unit treatment value assumption, which prohibits one unit's exposure from impacting another's outcome \citep{Rubin1978}. \\ \textit{Examples:} \cite{Balzer2018Hierarchical} examined the impact of household socioeconomic status, a cluster-level variable, on the risk of failing to test for HIV. Likewise, \cite{Buchanan2018} investigated both the individual and disseminated effects of a network-randomized intervention among people who inject drugs. \item \textbf{Missing data, censoring, and losses to follow up} can all be treated as additional intervention nodes in a given causal framework \citep{Robins2000, Robins1994, Scharfstein1999, Daniel2012, Mohan2013, Balzer2017CascadeMethods}. Thereby, we can treat missing data as a causal inference problem - as opposed to causal inference as a missing data.\\ \textit{Examples}: When estimating the effect of iron supplementation during pregnancy on anemia at delivery, \cite{Bodnar2004} used inverse probability of censoring weights to adjust for the measured ways in which the women who were censored could differ from those who were not. Likewise, \cite{PetersenCascade2017} estimated the probability of HIV RNA viral suppression over time among a closed cohort of HIV-infected adults, under a hypothetical intervention to prevent censoring and ensure complete viral load measurement. \item \textbf{Transportability}, a subset of generalizability, aims to apply the effect for a given sample to a different population or setting \citep{Cole2010, Stuart2011, Hernan2011transport, Petersen2011transport, Bareinboim2013, Pearl2015, Lesko2017, Balzer2017general}. \\ \textit{Examples:} \cite{Rudolph2017transport} examined whether the reduction in school dropout observed in the Moving to Opportunity trial was consistent between Boston and Los Angeles. Recently, \cite{Hong2018} investigated whether the reductions in cardiovascular risk from rosuvastatin as observed in the JUPITER trial would also have been observed in the UK population who were trial eligible. \end{enumerate}

a4b67edfa521cd80a1d1698ddfbe66a6c63abe1b

\section{Introduction} \label{sec:int} Douglas-Rachford method~\cite{MR0084194}, originally proposed for solving discretized heat equations, was extended by Lions and Mercier~\cite{MR0551319} for finding a zero of the sum of two maximal monotone operators. This method is, presently, the subject of intense research due to its efficiency, its use for solving PDEs, large-scale optimization problems (even some non-convex ones), imaging problems, and its connection with the alternating direction method (see~\cite{gabay,MR2858834,MR3134441,MR3260973,MR3260382,MR3394965,MR3587821,MR3516863,MR3535928} and the references therein). Eckstein and Bertseka~\cite{MR1168183} proved that Douglas-Rachford method can be regarded as an instance of the proximal point method applied to an implicitly defined operator. Recently, Eckstein and Yao~\cite{EckYao} cleverly used this result to propose an inexact version of Douglas-Rachford method with relative error tolerance based on Solodov and Svaiter hybrid proximal-extragradient method~\cite{hpe1,hpe2,hpe-unified,hpe-breg}. Each iteration of Douglas-Rachford method requires the sequential solution of two proximal subproblems. Eckstein-Yao algorithm is a semi-inexact version of this method in the sense that it allows for inexactness on the first proximal subproblem and requires the second one to be solved exactly. Complexity of Eckstein-Yao inexact Douglas-Rachford method was derived by Marques Alves and Geremia~\cite{AlvGer}. The main contribution of this work is the introduction and analysis of a fully inexact version of Douglas-Rachford method wherein both proximal subproblems are solved approximately within a relative error tolerance. To the best of our knowledge, this is the first fully inexact version of Douglas-Rachford method with a relative error tolerance. We also propose a semi-inexact version for the cases where one of the proximal subproblems can be solved exactly. In the semi-inexact case of Douglas-Rachford, our proposal is to solve the first subproblem exactly and the second one inexactly, so that the error criterion to be satisfied is immediately available (during the computation of the approximate solution of the second subproblem). Douglas-Rachford method for solving the inclusion problem \begin{align*} 0 \in A(x)+B(x), \end{align*} where $A$ and $B$ are maximal monotone operators in a Hilbert space $H$, can be written as: choose $x_0,b_0\in H$ and for $k=1,2,\dots$ \begin{equation*} \begin{alignedat}{3} &\text{compute }y_k,\;a_k\in H &\;\;\text{such that}\quad&a_k\in A(y_k), &\quad&\lambda a_k+y_k=x_{k-1}-\lambda b_{k-1},\\ &\text{compute }x_k,\;b_k\in H &\text{such that}\quad&b_k\in B(x_k), &&\lambda b_k+x_k=y_k+\lambda b_{k-1}. \end{alignedat} \end{equation*} Since $x_k=x_{k-1}-\lambda(a_k+b_k)$ and $b_k=b_{k-1}-\lambda^{-1}(x_k-y_k)$, we propose the following (fully) inexact version of Douglas-Rachford method: choose $z_0,w_0\in H$ and for $k=1,2,\dots$ \begin{align*} \begin{aligned} &\text{compute }y_k,\;a_k\in H\text{ and }x_k,\;b_k\in H \text{ such that:} \\[.2em] &\begin{alignedat}{3} &y_k,a_k \;\;\text{is an approximate solution of }&\;\;&a\in A(y), \;\; \lambda a+y={z}_{k-1}-\lambda {w}_{k-1},\\ &x_k,b_k\;\;\text{is an approximate solution of }&&b\in B(x), \;\;\lambda b+x=y_k+\lambda {w}_{k-1}, \end{alignedat} \\[.2em] &\text{set }{z}_k={z}_{k-1}-(1-t_k)\lambda (a_k+b_k),\; {w}_k={w}_{k-1}-(1-t_k)\lambda^{-1}(x_k-y_k), \end{aligned} \end{align*} where $t_k$ is an under-relaxation parameter which is zero whenever it is ``safe'' to do so (more of which latter). If both subproblems are exactly solved and $t_k=1$, at the $k$-th iteration of this generic method, then $z_k=x_k$ and $w_k=b_k$. Following~\cite{hpe2,hpe-unified}, we will allow errors in the inclusions \emph{and} in the equations in the computation of $y_k,a_k$ and $x_k,b_k$. Errors in the inclusion will be considered using the $\varepsilon$-enlargement~\cite{bis} of the maximal monotone operator $B$. Since our semi-inexact method is related to~\cite{EckYao}, it is worth discussing their differences. Eckstein-Yao semi-inexact version of Douglas-Rachford method~\cite{EckYao} does not require the introduction of a relaxation factor, as ours do, so that their inexact version is formally closer than ours to the exact method. Their version allows for errors in the second subproblem, while our semi-inexact version allows for errors in the second one. Weak convergence, in infinite dimensional spaces, of their version is an interesting open question, while we prove here weak convergence of our version. Their version has a very good practical performance~\cite[Section 7]{EckYao} and nice complexity properties~\cite{AlvGer}, while the practical performance and complexity properties of our version are, as of now, open questions. Weak convergence on Douglas-Rachford method was established by the author in~\cite{MR2783211} for the inexact version with the summable error tolerance. Here we use the techniques and ideas of that work to prove weak convergence under a relative error tolerance. This work is organized as follows. In Section \ref{sec:bd} we establish the notation and prove some basic results. In Section \ref{sec:idr} we the present a fully inexact Douglas-Rachford method and analyze its convergence properties. In Section \ref{sec:sidr} we present a semi-inexact Douglas-Rachford method and analyze its convergence properties. In Appendices \ref{sec:comp} and \ref{sec:tech} we prove some technical results. \section{Basic definitions and results} \label{sec:bd} From now on $H$ is a real Hilbert space with inner product $\inner{\cdot}{\cdot}$ and induced norm $\norm{x}=\sqrt{\inner{x}{x}}$. In $H \times H$ we consider the canonical inner product and associated norm of Hilbert spaces products. We are concerned with the inclusion problem \begin{align} \label{eq:p} 0 \in A(x)+B(x) \end{align} where $A:H\rightrightarrows H$ and $B:H\rightrightarrows H$ are maximal monotone operators. The \emph{extended solution set}~\cite{EckSva} of this problem is \begin{align} \label{eq:se} S_e(A,B)=\set{(z,w)\;:\;w\in B(z),\, -w\in A(x)}. \end{align} It is trivial to verify that \begin{align*} 0 \in A(z)+B(z) \iff (z,w)\in S_e(A,B)\text{ for some }w\in H. \end{align*} The $\varepsilon$-enlargement~\cite{bis} of a maximal monotone operator $T:H\rightrightarrows H$ is defined as \begin{align} \label{eq:teps} T^{[\varepsilon]}(z)=\set{v\in H\;:\; \inner{z-z'}{v-v'}\geq -\varepsilon \;\; \forall z'\in H,\;v'\in T(z') } \qquad (\varepsilon\geq0,\;\; z\in H). \end{align} The following elementary properties of the $\varepsilon$-enlargement will be used in the sequel. \begin{proposition} \label{pr:teps} If $T:H\rightrightarrows H$ is maximal monotone, then \begin{enumerate} \item\label{it:0teps} $T^{[\varepsilon=0]}=T$; \item\label{it:lteps} $\lambda(T^{[\varepsilon]})=(\lambda T)^{[\lambda\varepsilon]}$ for any $\varepsilon\geq 0$ and $\lambda>0$. \end{enumerate} \end{proposition} \begin{proof} Item~\ref{it:0teps} follows trivially from~\eqref{eq:teps} and the maximal monotonicity of $T$ while item~\ref{it:lteps} follows directly from \eqref{eq:teps}. \end{proof} In each iteration of Douglas-Rachford method, one shall compute $(I+\lambda T)^{-1}\zeta$, first with $T=A$ and $\zeta=x_{k-1}-\lambda b_{k-1}$ and them with $T=B$ and $\zeta= y_k+\lambda b_{k-1}$. Let $T$ be a maximal monotone operator in $H$. Computation of $z=(I+\lambda T)^{-1}\zeta$ can be decoupled in an inclusion and an equation, which we call the \emph{proximal inclusion-equation system}: \begin{align} \label{eq:psys} v\in T(z),\qquad \lambda v+z-\zeta=0. \end{align} If we allow errors in the inclusion, by means of the $\varepsilon$-enlargement of $T$, and error in the equation we get \begin{align*} v\in T^{[\varepsilon]}(z),\qquad \lambda v+z-\zeta=r \end{align*} where $\varepsilon$ is the \emph{error in the inclusion} and $r$ is the \emph{residual in the equation} at \eqref{eq:psys}. In some sense, $\norm{r}^2+2\lambda\varepsilon$ quantifies the overall error in the inexact solution of \eqref{eq:psys}, due to the next result proved in~\cite{error-bound}. \begin{proposition} [\mbox{\cite[Corollary 1]{error-bound}}] \label{pr:eb} Suppose $T:H\rightrightarrows H$ is maximal monotone, $\lambda<0$, and $\zeta\in H$. If \begin{align*} v^*\in T(z^*),\quad \lambda v^*+z^*-\zeta=0, \qquad v\in T^{[\varepsilon]}(z),\quad \lambda v+z-\zeta=r \end{align*} then $\norm{\lambda(v^*-v)}^2+\norm{z^*-z}^2 \leq \norm{r}^2+2\lambda\varepsilon$. \end{proposition} The next lemma will be instrumental in the convergence analysis of the inexact methods we propose in this work. \begin{lemma} \label{lm:bas} Let $z,w\in X$ and $\lambda>0$. Suppose \begin{alignat*}{2} \lambda a+y & = z-\lambda w+r,\quad & a&\in A^{[\varepsilon]}(x),\\ \lambda b+x & = y+\lambda w+s, & b&\in B^{[\mu]}(y), \end{alignat*} and define $z'=z-\lambda(a+b)$, $w'=w-\lambda^{-1}(x-y)$. For any $(z^*,w^*)\in S_e(A,B)$, \begin{multline*} \norm{(z^*,\lambda w^*)-(z,\lambda w)}^2 \geq \norm{(z^*,\lambda w^*)-(z',\lambda w')}^2 +\norm{\lambda (a+w)}^2\\ -[\norm{s}^2 +2\lambda(\inner{r}{a+b}+\varepsilon+\mu)]. \qquad \end{multline*} \end{lemma} \begin{proof} In view of item \ref{it:lteps} on Proposition~\ref{pr:teps}, it suffices to prove the lemma for the case $\lambda=1$, which we assume now. In this case \begin{align*} a+y=z-w+r,\;\; b+x=y+w+s,\;\;z'=z-(a+b),\;\;w'=w-(x-y). \end{align*} Fix $(z^*,w^*)\in S_e(A,B)$ and let \begin{align*} \pi=\inner{z^*-z'}{z'-z}+\inner{w^*-w'}{w'-w}. \end{align*} In view of definition~\eqref{eq:se}, $w^*\in B(z^*)$ and $-w^*\in A(z^*)$. It follows from these inclusions, the inclusions $a\in A^{[\varepsilon]}(y)$ and $b\in B^{[\mu]}(x)$, and definition~\eqref{eq:teps}, that \begin{align*} \inner{z^*-z'}{z'-z} &=\inner{z^*-z'}{w^*-b}+\inner{z^*-z'}{-w^*-a} \\ &=\inner{z^*-x}{w^*-b}+\inner{x-z'}{w^*-b} +\inner{z^*-y}{-w^*-a}+\inner{y-z'}{-w^*-a} \\ &\geq \inner{x-z'}{w^*-b}+\inner{y-z'}{-w^*-a}-\varepsilon-\mu \,. \end{align*} Direct combination of this inequality with the definitions of $\pi$ and $w'$ yields \begin{equation*} \pi \geq \inner{x-z'}{w^*-b}+\inner{y-z'}{-w^*-a}-\varepsilon-\mu +\inner{w^*-w'}{y-x}. \end{equation*} Since the expression at the right hand-side of the above inequality does not depend on $w^*$, we can substitute $b$ for $w^*$ is this expression and use the definitions of $z'$ and $w'$ to conclude that \begin{align*} \pi & \geq \inner{y-z'}{-b-a}+\inner{b-w'}{y-x} -\varepsilon-\mu \\ &=\inner{y-x+r+s}{-a-b}+ \inner{s}{y-x}-\varepsilon-\mu \\ &=\inner{x-y}{a+b}-[ \inner{r}{a+b}+\inner{s}{a+b+x-y}+ \varepsilon+\mu]. \end{align*} Therefore, \begin{align*} \norm{(z^*, w^*)-(z, w)}^2 &= \norm{(z^*, w^*)-(z', w')}^2 + \norm{(z', w')-(z, w)}^2 +2\pi \\ &\geq \norm{(z^*, w^*)-(z', w')}^2 +\norm{a+b}^2+\norm{x-y}^2 \\ &\qquad \mbox{} +2\inner{x-y}{a+b} -2[\inner{r}{a+b}+\inner{s}{a+b+y-x}+\varepsilon+\mu] \\ & = \norm{(z^*, w^*)-(z', w')}^2 +\norm{a+b+x-y}^2 \\ &\qquad \mbox{} -2[\inner{r}{a+b}+\inner{s}{a+b+y-x}+\varepsilon+\mu]. \end{align*} To end the proof, observe that \begin{equation*} \norm{a+b+x-y}^2-2\inner{s}{a+b+x-y} =\norm{a+b+x-y-s}^2-\norm{s}^2=\norm{a+w}^2-\norm{s}^2 \end{equation*} and combine the two above equations. \end{proof} \section{A fully inexact Douglas-Rachford method with relative error tolerance} \label{sec:idr} In this section we present an Inexact Douglas-Rachford method wherein both proximal subproblems are to be solved within a relative error tolerance. We prove that the sequences generated by this method converge weakly to a point in $S_e(A,B)$, whenever the solution set of \eqref{eq:p} is non-empty. \vspace{.1in} \vspace{.1in} \noindent \fbox{ \begin{minipage}[h]{6.4 in} {\bf Algorithm I: inexact DR method with relative error tolerance}. \begin{itemize} \item[(0)] Let $z_0,w_0\in H$, $\lambda>0$, $0<\sigma<\nu<1$. For $k=1,2,3,\dots$, \item[(1)] find $a_k,y_k,\varepsilon_k, b_k,x_k,\mu_k$ such that \begin{align} \label{eq:s1} \begin{split} & a_k\in A^{[\varepsilon_k]}(y_k), \quad b_k\in B^{[\mu_k]}(x_k), \\ & \begin{aligned}\norm{\lambda a_k+y_k-(z_{k-1}-\lambda w_{k-1})}^2 +\norm{\lambda b_k+x_k-(y_k+\lambda w_{k-1})}^2 & \\ +2\lambda(\varepsilon_k+\mu_k) & \leq \dfrac{\sigma^2}{4} (\norm{a_k+b_k}^2+\norm{x_k-y_k}^2); \end{aligned} \end{split} \end{align} \item [(2)] set \begin{align} \label{eq:s2} \begin{split} &\delta_k=\norm{\lambda a_k+y_k-(z_{k-1}-\lambda w_{k-1})}^2 +\norm{\lambda b_k+x_k-(y_k+\lambda w_{k-1})}^2 +2\lambda(\varepsilon_k+\mu_k), \\ &\rho_k=\norm{\lambda(a_k+b_k)}^2+\norm{x_k-y_k}^2, \\ &t_k= \begin{cases} 0 & \text{if }a_k+b_k= x_k-y_k=0\\ \nu\max\Set{0, \sqrt{\dfrac{4\delta_k}{\sigma^2\rho_k}} -\dfrac{\norm{\lambda(a_k+w_{k-1})}^2}{\rho_k} }&\text{otherwise}, \end{cases} \\[.6em] & z_k=z_{k-1}-(1-t_k)\lambda(a_k+b_k),\quad w_k=w_{k-1}-(1-t_k)\lambda^{-1}(x_k-y_k); \end{split} \end{align} \end{itemize} \noindent \end{minipage} } \vspace{.1in} \vspace{.1in} We did not specify how to compute $a_k$, $y_k$, $\varepsilon_k$ ,$b_k$, $x_k$, and $\mu_k$, which adds generality to the method. In Appendix~\ref{sec:comp} we show that under some mild conditions (on $A$ and $B$) step (2) is computable. Whenever $t_k=0$, $z_k=z_{k-1}-\lambda(a_k+b_k)$ and $w_k=w_{k-1}-\lambda^{-1} (x_k-y_k)$, so that formally we retrieve a Douglas-Rachford iteration. In those iterations where $t_k\neq 0$, we will have an under-relaxed Douglas-Rachford iteration. To simplify the convergence analysis, let $r_k$ and $s_k$ denote the residuals in the equations of the proximal inclusion-equations systems to be solved for $A$ and $B$ at the $k$-th iteration, that is, \begin{align} \label{eq:rsk} \begin{aligned} r_k&=\lambda a_k+y_k-(z_{k-1}-\lambda w_{k-1}), \quad s_k=\lambda b_k+x_k-(y_k+\lambda w_{k-1}), \end{aligned} \qquad (k=1,2,\dots). \end{align} With this notation, \eqref{eq:s1} and the first line of \eqref{eq:s2} writes \begin{align} \label{eq:syn} \begin{split} \begin{aligned} &\begin{alignedat}{2} a_k&\in A^{[\varepsilon_k]}(y_k), \quad &\lambda a_k+y_k&=z_{k-1}-\lambda w_{k-1}+r_k, \\ b_k&\in B^{[\mu_k]}(x_k), &\lambda b_k+x_k&=y_k+\lambda w_{k-1}+s_k, \end{alignedat} \\ &\delta_k=\norm{r_k}^2+\norm{s_k}^2 +2\lambda(\varepsilon_k+\mu_k), \qquad\delta_k\leq\dfrac{\sigma^2}{4}\rho_k. \end{aligned} \end{split} \end{align} It follows from the definition of $t_k$ at \eqref{eq:s2} that for all $k$, \begin{alignat}{1} \label{eq:boundt} 0\leq t_k\leq \nu<1, \qquad \dfrac{\nu}{\sigma}\sqrt{4\delta_k\rho_k} -\nu\norm{\lambda(a_k+w_{k-1})}^2 \leq t_k\rho_k ,\qquad t_k^2\rho_k\leq\dfrac{4\nu^2}{\sigma^2}\delta_k. \end{alignat} From now on in this section, \begin{align} \label{eq:pk} \qquad p_k =(z_k,\lambda w_k),\qquad \qquad \qquad\qquad (k=0,1,2,\dots). \end{align} First we will prove that the sequence $\set{p_k}$ converges F\'ejer to the set of points $(z,\lambda w)$ where $(z,w) \in S_e(A,B)$. \begin{lemma} \label{lm:fej} For any $({z},{w})\in S_e(A,B)$ and all $k$ \begin{align*} \norm{({z},\lambda {w})-p_{k-1}}^2 & \geq \norm{({z},\lambda {w})-p_k}^2 +(1-t_k) \left\{\dfrac{\nu-\sigma}{\sigma}\sqrt{4\delta_k\rho_k} + (1-\nu)\norm{\lambda(a_k+w_{k-1})}^2 \right\}, \\ \norm{({z},\lambda {w})-p_0}^2 & \geq \norm{({z},\lambda {w})-p_k}^2 +\sum_{i=1}^k(1-t_i) \left\{\dfrac{\nu-\sigma}{\sigma}\sqrt{4\delta_i\rho_i} + (1-\nu)\norm{\lambda(a_i+w_{i-1})}^2 \right\}. \end{align*} \end{lemma} \begin{proof} Fix $({z},{w})\in S_e(A,B)$ and let $p^*=({z},\lambda {w})$. Define, for $k=1,2,\dots$, \begin{align} \label{eq:p'k} z'_k=z_{k-1}-\lambda(a_k+b_k), \quad w'_k=w_{k-1}-\lambda^{-1}(x_k-y_k), \quad p'_k=(z'_k,\lambda w'_k). \end{align} It follows from \eqref{eq:pk}, \eqref{eq:syn} and Lemma~\ref{lm:bas} that \begin{align*} \norm{p^*-p_{k-1}}^2 &\geq \norm{p^*-p'_k}^2+\norm{\lambda(a_k+w_{k-1})}^2 -[\norm{s_k}^2+2\lambda(\inner{r_k}{a_k+b_k}+\varepsilon_k+\mu_k)] . \end{align*} Using Cauchy-Schwarz inequality, the definition of $\rho_k$ in \eqref{eq:s2}, and the third line on \eqref{eq:syn} we conclude that \begin{align*} \norm{s_k}^2+2(\inner{r_k}{a_k+b_k}+\varepsilon_k+\mu_k) & \leq \norm{s_k}^2+2\lambda(\varepsilon_k+\mu_k) +2\norm{r_k}\sqrt{\rho_k}\\ &=\delta_k-\norm{r_k}^2+2\norm{r_k} \sqrt{\rho_k} \end{align*} and $\norm{r_k}\leq \sqrt{\delta_k}\leq\sigma\sqrt{\rho_k}/2$. Since the expression on the right hand-side of the above equality is increasing for $\norm{r_k}\leq\sqrt{\rho_k}$, its maximum value on $[0,\sqrt{\delta_k}]$ is attained at $\norm{r_k}=\sqrt{\delta_k}$. Combining these observations with the two above inequalities we conclude that \begin{align*} \norm{p^*-p_{k-1}}^2 &\geq \norm{p^*-p'_k}^2+\norm{\lambda(a_k+w_{k-1})}^2 -2\sqrt{\delta_k\rho_k}. \end{align*} Since $p_k=t_kp_{k-1}+(1-t_k)p'_k$ and $\norm{p_k-p'_k}^2=\rho_k$, \begin{align*} \norm{p^*-p_{k-1}}^2 &=t_k\norm{p^*-p_{k-1}}+(1-t_k)\norm{p^*-p_{k-1}}^2 \\ & \geq t_k\norm{p^*-p_{k-1}}+ (1-t_k)\left[ \norm{p^*-p'_k}^2+\norm{\lambda(a_k+w_{k-1})}^2 -2\sqrt{\delta_k\rho_k}\right] \\ & = \norm{t_k(p^*-p_{k-1})+(1-t_k)(p^*-p'_k)}^2+ (1-t_k)t_k\norm{p_{k-1}-p'_k}^2\\ &\quad\mbox{}+ (1-t_k)\left[\norm{\lambda(a_k+w_{k-1})}^2 -2\sqrt{\delta_k\rho_k}\right] \\ & = \norm{p^*-p_k}+ (1-t_k)\left[t_k\rho_k+\norm{\lambda(a_k+w_{k-1})}^2 -2\sqrt{\delta_k\rho_k}\right] . \end{align*} To end the proof of the first inequality, use the above inequality and the next to last inequality in \eqref{eq:boundt}. The second inequality of the lemma follows trivially from the first one. \end{proof} \begin{lemma} \label{lm:prec} If $S_e(A,B)$ is nonempty, then \begin{enumerate} \item \label{it:bd} $\set{z_k}$ and $\set{w_k}$ are bounded; \item \label{it:rsem} $r_k\to 0$, $s_k\to 0$, $\varepsilon_k\to 0$ and $\mu_k \to 0$ as $k\to \infty$; \item \label{it:ay} $a_k+w_{k-1}\to 0$ and $y_k-z_{k-1}\to 0$ as $k\to \infty$; \item \label{it:bx} $b_k-w_k \to 0$ and $x_k-z_k \to 0$ as $k\to \infty$. \end{enumerate} \end{lemma} \begin{proof} Take $({z},{w})$ in $S_e(A,B)$. It follows from Lemma~\ref{lm:fej} that $\set{(z_k,\lambda w_k)}$ is bounded, which proves item~\ref{it:bd}, and that \begin{align*} \sum_{k=1}^\infty\sqrt{\delta_k\rho_k} < \infty,\qquad \sum_{k=1}^\infty \norm{a_k+w_{i-1}}^2 <\infty. \end{align*} Since $\delta_k\leq\sigma^2\rho_k/4$, $\sqrt{\delta_k\rho_k}\geq 2 \delta_k/\sigma$ and it follows from the first above inequality that $\delta_k\to 0$ as $k\to \infty$. This result, together with the third line of \eqref{eq:syn} proves item \ref{it:rsem}. The first limit in item~\ref{it:ay} follows trivially form the second above inequality while the second limit follows from the first one, the first limit in item~\ref{it:rsem} and the definition of $r_k$ in \eqref{eq:rsk}. It follows from the update formula for $z_k$ and $w_k$, and from \eqref{eq:syn} that \begin{align*} z_k-x_k+r_k+s_k=t_k\lambda(a_k+b_k),\qquad w_k-b_k+s_k=t_k(x_k-y_k). \end{align*} Squaring both sides of each one of the above equations and adding them we conclude that \begin{align*} \norm{z_k-x_k+r_k+s_k}^2+\norm{w_k-b_k+s_k}^2 =t_k^2\rho_k. \end{align*} Since $\delta_k\to 0$ as $k\to \infty$, it follows from the above equations and the last inequality in~\eqref{eq:boundt} that \begin{align*} t_k^2\rho_k\to 0,\quad z_k-x_k+r_k+s_k\to 0,\quad w_k-b_k+s_k\to 0\qquad \text{as }k\to \infty. \end{align*} Item \ref{it:bx} follows from the above result and from item~\ref{it:rsem}. \end{proof} \begin{theorem} If $S_e(A,B)\neq \emptyset$, then $\set{(x_k,b_k)}$, $\set{(y_k,-a_k)}$ and $\set{(z_k,w_k)}$ converge weakly to a point in this set. \end{theorem} \begin{proof} Suppose a subsequence $\set{(z_{k_n},w_{k_n})}$ converges weakly to some $(z,w)$. It follows from Lemma~\ref{lm:prec} that \begin{alignat*}{4} &x_{k_n}\tow z, &\quad& y_{k_n-1}\tow z,\quad &&x_{k_n}-y_{k_n-1}\to 0, \\ &b_{k_n}\tow w, && a_{k_n-1}\tow -w,&\quad& a_{k_n-1}+b_{k_n}\to 0, & \\ &\mu_{k_n}\to 0,&&\varepsilon_{k_n-1}\to 0,&& &\qquad \text{as }k\to \infty. \end{alignat*} Since $b_{k_n}\in B^{[\mu_{k_n}]}(x_{k_n})$, $a_{k_n-1}\in A^{[\varepsilon_{k_n-1}]}(y_{k_n-1})$ for $n=1,2,\dots$, it follows from Lemma~\ref{lm:limit} that $w\in B(z)$ and $-w\in A(w)$. Therefore, $(z,w)\in S_e(A,B)$. We have proved that all weak limit points of $\set{(z_k,w_k)}$ are in $S_e(A,B)$. Therefore, all weak limit points of $\set{p_k=(z_k,\lambda w_k)}$ are in $\Omega$, \begin{align*} \Omega=\set{(z,\lambda w)\;:\;(z,w)\in S_e(A,B)}. \end{align*} In Lemma~\ref{lm:fej} we proved that $\set{p_k}$ is Fej\'er convergent to $\Omega$. Since $\Omega\neq\emptyset$, it follows from these results and from Opial's Lemma~\cite{opial} that the bounded sequence $\set{p_k}$ has a unique weak limit point and such a point belongs to $\Omega$. Therefore, $\set{p_k}$ converges weakly to a point in $(z,\lambda w)$ where $(z,w)\in S_e(A,B)$. To end the proof, use items~\ref{it:ay} and \ref{it:bx} of Lemma~\ref{lm:prec}. \end{proof} \section{A semi-inexact Douglas-Rachford method} \label{sec:sidr} In this section we present an inexact Douglas-Rachford method wherein, in each iteration, the first proximal subproblem is to be exactly solved and the second proximal subproblem is to be solved within a relative error tolerance. A possible advantage of solving the second subproblem approximately, instead of the first one, is that the error criterion to be satisfied is readily available during the computation of the second step, thereby obviating the necessity of computing more than once the approximate solution per iteration. Since one of the subproblems is to be solved exactly, following~\cite{EckYao}, we call this method semi-inexact. \vspace{.1in} \vspace{.1in} \noindent \fbox{ \begin{minipage}[h]{6.4 in} {\bf Algorithm II: A semi-inexact Douglas-Rachford method with relative error tolerance}. \begin{itemize} \item[(0)] Let $z_0,w_0\in H$, $\lambda>0$, $0<\sigma<\nu<1$. For $k=1,2,3,\dots$, \item[(1)] compute $a_k,y_k$ such that \begin{subequations} \begin{align} \label{eq:s1a} &a_k\in A(y_k),\quad \lambda a_k + y_k = z_{k-1} - \lambda w_{k-1}, \intertext{compute $b_k,x_k,\mu_k$ such that} \label{eq:s1b} &b_k\in B^{[\mu_k]}(x_k),\; \norm{\lambda b_k+x_k-(y_k+\lambda w_{k-1})}^2 +2\lambda\mu_k\leq \sigma^2 (\norm{a_k+b_k}^2 +\norm{x_k-y_k}^2); \end{align} \end{subequations} \item [(2)] set \begin{align} \label{eq:s2.si} \begin{split} &\delta_k= \norm{\lambda b_k+x_k-(y_k+\lambda w_{k-1})}^2 +2\lambda \mu_k, \\ &\rho_k=\norm{\lambda(a_k+b_k)}^2+\norm{x_k-y_k}^2, \\ &t_k= \begin{cases} 0 & \text{if }a_k+b_k= x_k-y_k=0\\ \nu^2\max\Set{0, \dfrac{\delta_k}{\sigma^2\rho_k} -\dfrac{\norm{\lambda(a_k+w_{k-1})}^2}{\rho_k} }&\text{otherwise}, \end{cases} \\[.6em] & z_k=z_{k-1}-(1-t_k)\lambda(a_k+b_k),\quad w_k=w_{k-1}-(1-t_k)\lambda^{-1}(x_k-y_k); \end{split} \end{align} \end{itemize} \noindent \end{minipage} } \vspace{.1in} \vspace{.1in} From now on in this section, $\set{z_k}$, $\set{w_k}$, $\set{a_k}$ etc. are sequences generated by Algorithm {\bf II}. To simplify the convergence analysis, let $s_k$ denote the residuals in the equation of the inclusion-equation system to be solved for $B$ at the $k$-th iteration, that is, \begin{align} \label{eq:rsk.si} s_k=\lambda b_k+x_k-(y_k+\lambda w_{k-1}), \qquad\qquad (k=1,2,\dots). \end{align} With this notation, \eqref{eq:s1a}, \eqref{eq:s1b} and the first line of \eqref{eq:s2.si} writes \begin{align} \label{eq:syn.si} \begin{split} \begin{aligned} &\begin{alignedat}{2} a_k&\in A(y_k), &\lambda a_k+y_k&=z_{k-1}-\lambda w_{k-1}, \\ b_k&\in B^{[\mu_k]}(x_k), \;\;\; &\lambda b_k+x_k&=y_k+\lambda w_{k-1}+s_k, \end{alignedat} \\ &\delta_k=\norm{s_k}^2 +2\lambda\mu_k, \qquad\delta_k\leq\sigma^2\rho_k. \end{aligned} \end{split} \end{align} It follows from the definition of $t_k$ at \eqref{eq:s2.si} that for all $k$, \begin{alignat}{1} \label{eq:boundt.si} 0\leq t_k\leq \nu^2<1, \qquad \dfrac{\nu^2}{\sigma^2}\delta_k-\nu^2\norm{\lambda(a_k+w_{k-1})}^2 \leq t_k\rho_k ,\qquad t_k^2\rho_k \leq \dfrac{\nu^4}{\sigma^2}\delta_k. \end{alignat} From now on in this section, \begin{align} \label{eq:pk.si} p_k =(z_k,\lambda w_k),\qquad\qquad (k=0,1,2,\dots). \end{align} We will prove that the sequence $\set{p_k}$ converges to a point $(z,\lambda w)$ where $(z,w) \in S_e(A,B)$. \begin{lemma} \label{lm:fej.si} For any $(z^*,w^*)\in S_e(A,B)$ and all $k$ \begin{align*} \norm{(z^*,\lambda w^*)-p_{k-1}}^2 & \geq \norm{(z^*,\lambda w^*)-p_k}^2 +(1-t_k) \left\{\dfrac{\nu^2-\sigma^2}{\sigma^2}\delta_k + (1-\nu^2)\norm{\lambda(a_k+w_{k-1})}^2 \right\}, \\ \norm{(z^*,\lambda w^*)-p_0}^2 & \geq \norm{(z^*,\lambda w^*)-p_k}^2 +\sum_{i=1}^k(1-t_i) \left\{\dfrac{\nu^2-\sigma^2}{\sigma^2}\delta_i + (1-\nu^2)\norm{\lambda(a_i+w_{i-1})}^2 \right\}. \end{align*} \end{lemma} \begin{proof} Fix $(z^*,w^*)\in S_e(A,B)$ and let $p^*=(z^*,\lambda w^*)$. Define, for $k=1,2,\dots$, \begin{align} \label{eq:p'k.si} z'_k=z_{k-1}-\lambda(a_k+b_k), \quad w'_k=w_{k-1}-\lambda^{-1}(x_k-y_k), \quad p'_k=(z'_k,\lambda w'_k). \end{align} It follows from \eqref{eq:pk.si}, \eqref{eq:syn.si} and Lemma~\ref{lm:bas} that \begin{align*} \norm{p^*-p_{k-1}}^2 &\geq \norm{p^*-p'_k}^2+\norm{\lambda(a_k+w_{k-1})}^2 -[\norm{s_k}^2+2\lambda\mu_k] \\ &\geq \norm{p^*-p'_k}^2+\norm{\lambda(a_k+w_{k-1})}^2 -\sigma^2\rho_k. \end{align*} Since $p_k=t_kp_{k-1}+(1-t_k)p'_k$ and $\norm{p_k-p'_k}^2=\rho_k$, \begin{align*} \norm{p^*-p_{k-1}}^2 &=t_k\norm{p^*-p_{k-1}}+(1-t_k)\norm{p^*-p_{k-1}}^2 \\ & \geq t_k\norm{p^*-p_{k-1}}+ (1-t_k)\left[ \norm{p^*-p'_k}^2+\norm{\lambda(a_k+w_{k-1})}^2 -\delta_k\right] \\ & = \norm{t_k(p^*-p_{k-1})+(1-t_k)(p^*-p'_k)}^2+ (1-t_k)t_k\norm{p_{k-1}-p'_k}^2\\ &\quad\mbox{}+ (1-t_k)\left[\norm{\lambda(a_k+w_{k-1})}^2 -\delta_k\right] \\ & = \norm{p^*-p_k}+ (1-t_k)\left[t_k\rho_k+\norm{\lambda(a_k+w_{k-1})}^2 -\delta_k\right]. \end{align*} To end the proof of the first inequality, use the above inequality and the next to last inequality in \eqref{eq:boundt.si}. The second inequality of the lemma follows trivially from the first one. \end{proof} \begin{lemma} \label{lm:prec.si} If $S_e(A,B)$ is nonempty, then \begin{enumerate} \item \label{it:bd.si} $\set{z_k}$ and $\set{w_k}$ are bounded; \item \label{it:rsem.si} $s_k\to 0$ and $\mu_k \to 0$ as $k\to \infty$; \item \label{it:ay.si} $a_k+w_{k-1}\to 0$ and $y_k-z_{k-1}\to 0$ as $k\to \infty$; \item \label{it:bx.si} $b_k-w_k \to 0$ and $x_k-z_k \to 0$ as $k\to \infty$. \end{enumerate} \end{lemma} \begin{proof} Take $(z^*,w^*)$ in $S_e(A,B)$. It follows from Lemma~\ref{lm:fej.si} that $\set{(z_k,\lambda w_k)}$ is bounded, which proves item~\ref{it:bd.si}, and that \begin{align*} \sum_{k=1}^\infty\delta_k < \infty,\qquad \sum_{k=1}^\infty \norm{a_k+w_{i-1}}^2 <\infty. \end{align*} It follows from the first above inequality that $\delta_k\to 0$ as $k\to \infty$. This result, together with the third line of \eqref{eq:syn.si} proves item \ref{it:rsem.si}. Item~\ref{it:ay.si} follows trivially form the second above inequality and the equality in \eqref{eq:s1a}. It follows from the update formulas for $z_k$ and $w_k$, and from \eqref{eq:syn.si} that \begin{align*} z_k-x_k+s_k=t_k\lambda(a_k+b_k),\qquad w_k-b_k+s_k=t_k(x_k-y_k). \end{align*} Squaring both sides of each one of these equalities and adding them we conclude that \begin{align*} \norm{z_k-x_k+s_k}^2+\norm{w_k-b_k+s_k}^2 =t_k^2\rho_k. \end{align*} Since $\delta_k\to 0$ as $k\to \infty$, it follows from the above equations and the last inequality in~\eqref{eq:boundt.si} that \begin{align*} t_k^2\rho_k\to 0,\quad z_k-x_k+s_k\to 0,\quad w_k-b_k+s_k\to 0\qquad \text{as }k\to \infty. \end{align*} Item \ref{it:bx.si} follows from the above result and from item~\ref{it:rsem.si}. \end{proof} \begin{theorem} If $S_e(A,B)\neq \emptyset$, then $\set{(x_k,b_k)}$, $\set{(y_k,-a_k)}$ and $\set{(z_k,w_k)}$ converge weakly to a point in this set. \end{theorem} \begin{proof} Suppose a subsequence $\set{(z_{k_n},w_{k_n})}$ converges weakly to some $(z,w)$. It follows from Lemma~\ref{lm:prec.si} that \begin{align*} \begin{alignedat}{3} &x_{k_n}\tow z, &\quad& y_{k_n-1}\tow z,\quad &&x_{k_n}-y_{k_n-1}\to 0, \\ &b_{k_n}\tow w, && a_{k_n-1}\tow -w,&\quad& a_{k_n-1}+b_{k_n}\to 0, \\ &\mu_{k_n}\to 0&&\varepsilon_{k_n-1}\to 0&&\qquad \end{alignedat} \qquad \text{as }k\to \infty. \end{align*} Since $b_{k_n}\in B^{[\mu_{k_n}]}(x_{k_n})$, $a_{k_n-1}\in A(y_{k_n-1})$ for $n=1,2,\dots$, it follows from Lemma~\ref{lm:limit} that $w\in B(z)$ and $-w\in A(w)$. Therefore, $(z,w)\in S_e(A,B)$. Define, again, \begin{align*} \Omega=\set{(z,\lambda w)\;:\;(z,w)\in S_e(A,B)}. \end{align*} We have proved that all weak limit points of $\set{(z_k,w_k)}$ are in $S_e(A,B)$. Therefore, all weak limit points of $\set{p_k=(z_k,\lambda w_k)}$ are in $\Omega$. In Lemma~\ref{lm:fej.si} we proved that $\set{p_k}$ is Fej\'er convergent to $\Omega$. Since $\Omega\neq\emptyset$, it follows from these results and from Opial's Lemma~\cite{opial} that the bounded sequence $\set{p_k}$ has a unique weak limit point and such a point belongs to $\Omega$. Therefore, $\set{p_k}$ converges weakly to a point in $(z,\lambda w)$ where $(z,w)\in S_e(A,B)$. To end the proof, use items~\ref{it:ay.si} and \ref{it:bx.si} of Lemma~\ref{lm:prec.si}. \end{proof}

a0fe353ffabf876fee085229151188e2262d57e4

\section{Abstract} { The comparison of observed brain activity with the statistics generated by artificial intelligence systems is useful to probe brain functional organization under ecological conditions. Here we study fMRI activity in ten subjects watching color natural movies and compute deep representations of these movies with an architecture that relies on optical flow and image content. The association of activity in visual areas with the different layers of the deep architecture displays complexity-related contrasts across visual areas and reveals a striking foveal/peripheral dichotomy. } \begin{quote} \small \textbf{Keywords:} deep learning; video encoding; brain mapping; \end{quote} \section{Introduction} The understanding of brain functional architecture has long been driven by subtractive reasoning approaches, in which the activation patterns associated with different experimental conditions presented in event-related or block designs are contrasted in order to yield condition-specific maps \cite{poline2012}. A more ecological way of stimulating subjects consists in presenting complex continuous stimuli that are much more similar to every-day cognitive experiences. The analysis of the ensuing complex stimulation streams proceeds by extracting relevant features from the stimuli and correlating the occurrence of these features with brain activity recorded simultaneously with the presentation of the stimuli. The analysis of video streams has been carried in \cite{eickenberg2017seeing} or \cite{gucclu2015deep} using a deep convolutional network trained for image classification. More recently, \cite{gucclu2017increasingly} has used a deep neural network trained for action recognition to analyze video streams. Like \cite{gucclu2017increasingly}, we use a deep neural network trained for action recognition to extract video features and train a linear model to predict brain activity from these features. In contrast, our study is not restricted to dorsal stream visual areas but involves the whole brain, and the deep neural network we use is pretrained on the largest action recognition dataset available \cite{kay2017kinetics}. From the different layers of the deep neural networks, we build video representations that allow us to segregate (1) occipital and lateral areas of the visual cortex (reproducing the results of \cite{gucclu2015deep}) and (2) foveal and peripheric areas of the visual cortex. We also introduce an efficient spatial compression scheme for deep video features that allows us to speed up the training of our predictive algorithm. We show that our compression scheme outperforms PCA by a large margin. \section{Methods} \subsection{Deep video representation} We use a deep neural network trained for action recognition to build deep representations of the Berkeley Video stimuli \cite{nishimoto2011reconstructing}. This material consists of more than four hours of color natural movies built by mixing video blocks of $5$-$15$ seconds in a random fashion. The deep network we use is called Temporal Segment Network (TSN) \cite{wang2016temporal}. Following an idea introduced in 2014 \cite{simonyan2014two} it was intended to mimic the dorsal and ventral stream by separately processing raw frames and optical flow fields. We chose TSN for our experiments because it uses a much larger number of layers than the original network (which results in higher accuracy in action recognition) and that a version of TSN pretrained on Kinetics -- a massive video dataset (300~000 unique video clips) with 400 different classes all describing a human and at least 400 videos per class -- is publicly available. The network is trained to recognize human actions such as slack-lining, skateboarding, massaging feet, dancing zumba and dining. The version of TSN we use in our experiments is based on Inception v3 \cite{szegedy2016rethinking} for both streams where small networks are used as building blocks of the main large network \cite{lin2013network}. Each stream in the TSN Network is composed of more than 40 convolution layers and a fully connected layer. The activities after the last layer represent the probability of belonging to each action class. \subsection{Feature extraction} The raw frames encode information about pixels, and flow fields encode information about pixels displacements. Although flow fields and raw frames streams do not precisely disentangle spatial content and motion information in videos, we may expect that the raw frames stream better represent local spatial features while the flow fields stream more efficiently convey dynamic information. Following \cite{eickenberg2017seeing} we consider that the activation statistics in the first layers (the ones closer to those of the input) have a low level of abstraction, whereas the last layers (closer to the labels) represent high-level information. Therefore each activity in both streams can be considered as specific features or representations of the video. If we were to extract all network activities of the Berkeley Video Dataset we would need to store more than 6 millions floats per frame in the dataset. Such a representation would be highly redundant. In order to keep the volume of data reasonable, in each stream we only focus on four convolutional layers $L_1, L_2, L_3, L_4$ ranked by complexity. We further compress the data using spatial smoothing, and use temporal smoothing so that we get one representation every two seconds of video, which allows us to match the acquisition rate of fMRI scanners. \subsection{Regression} 10 subjects were scanned while watching the color natural movies of the Berkeley Video Dataset. The fMRI images were acquired at high spatial resolution (1.5mm), from a Prisma Scanner, using Multi-band and IPAT accelerations (mb factor=3, ipat=2). These data are part of a large-scale mapping project on a limited number of participants, called Human Brain Charting. Data acquisition procedures and initial experiments run in this project are described in \cite{ibc}. In order to link extracted deep video features to the internal representation of videos in each subject we use a simple linear model to fit their brain activity in each voxel. The use of a very simple model allows us to posit that the performance of the predictive model from a particular video representation is mostly linked to the suitability of the video representation. Hence the performance of the algorithm can be seen as a measure of the biological suitability of the video representation. We use a kernel ridge regression with an hyper-parameter setting the magnitude of the l2-penalization on the weights. The resulting prediction is obtained using a cross validation procedure (11 sessions are used for train, 1 for test and at least 5 different splits are considered). To set the value of the hyper-parameter, we use a 5-fold cross validation on the train set and consider 20 different values. During hyper parameter selection, we only focus on the visual cortex to make this computation efficient. The chosen measure of performance of our prediction algorithm is the coefficient of determination $m_{cv}$. Let $\mathbf{y}_{pred}$ and $\mathbf{y}_{real}$ be the respectively the prediction of a voxel activity and the real voxel activity. Then \[ m_{cv}(\mathbf{y}_{pred}, \mathbf{y}_{real}) \;=\; 1 - \frac{\sum_{t=1}^{n_b} (\mathbf{y}_{pred}[t] - \mathbf{y}_{real}[t])^2}{\sum_{t=1}^{n_b}(\mathbf{y}_{real}[t] - \overline{\mathbf{y}_{real}})^2} \] The metric used to select the best parameter is the number of voxels having a coefficient of determination $m_{cv}$ greater than $0.1$. This procedure leads to different parameter values depending on the chosen layer activities. Figure~\ref{fig:feature_extraction} gives an overview of the pipeline used to extract and process deep video features to estimate the brain activity of subjects. \begin{figure} \centering \includegraphics[scale=0.4]{conceptual_figure.pdf} \caption{ Feature extraction and regression scheme: at each time frame we compute and extract the activities of four layers $L_1, \cdots, L_4$ of the temporal segment network on a single frame and on a stack of 5 consecutive optical flow fields. The extracted activities are spatially and temporally down-sampled and then used to predict brain activity of subjects exposed to the video stimuli.} \label{fig:feature_extraction} \end{figure} \section{Results} The extracted deep network features lead to different prediction performance depending on the down-sampling procedure, the stream used and the localization of target voxels. \subsection{An efficient spatial compression scheme} We show that preserving the channel structure of the network during spatial compression procedure is key for developing an efficient compression scheme. We compare three spatial compression schemes for network activities: (1) Standard principal component analysis (PCA) with $2000$ components; the transformation is learned on training sessions before it is applied to all sessions. (2) Average pooling inside channels (APIC) which computes local means of activities located in the same channel. (3) Average pooling inside and between convolution layers (APBIC) which is used to get the same number of output features for all layers while minimizing the number of convolutions between channels. It allows us to check that the performance of the predictive algorithm is not merely driven by the number of features. The procedure for activities extraction, temporal down-sampling and brain activity prediction is not changed while the spatial compression scheme varies. The benchmark is performed using a leave-one-out cross-validation procedure with two splits in three subjects. Figure~\ref{fig:spatial_compression} shows that both approaches preserving channel organization structure outperform PCA by a large margin. \begin{figure} \hspace*{-3mm} \includegraphics[scale=0.49]{spatial_compression.pdf} \caption{ Comparison of the different neural network compression streams. The APIC approach slightly outperforms APIBC, and both strongly outperform PCA. When using APIC or APIBC we predict correctly up to 850 times more voxels than when using PCA.} \label{fig:spatial_compression} \end{figure} These results suggest that data stored in the same channel are similar and that mixing data between channels tends to destroy valuable information. In our pipeline, we average only inside same channels (APIC) because it yields the best performance. Choosing APBIC would be trading performance for computation speed since its high compression rate enables a much faster training of the prediction algorithm. \subsection{Data based parcellation of the brain using deep video representation} Depending on the considered region of the brain, the best fitting representation varies. We show that the compressed activities of different layers show contrasts between low-level (retinotopic) versus high-level (object-responsive) areas, but also between foveal and peripheral areas. The difference between the prediction score from high level layer activity and low level layer activity of both streams ($L_4^{flow} - L_2^{flow}$ and $L_4^{rgb} - L_2^{rgb}$) yields a clear contrast between occipital (low-level) and lateral (high-level) areas (see Fig~\ref{fig:L4-L1}). This highlights a gradient of complexity in neural representation along the ventral stream which was also found in \cite{gucclu2015deep}. \begin{figure}[t] \centering \includegraphics[scale=0.3]{layer4-layer1-rgb.png} \includegraphics[scale=0.3]{layer4-layer1-flow.png} \caption{ High level and low level areas contrasts: Difference between predictions score from high level layer activity and low level activity of the raw frames stream $L_4^{rgb} - L_2^{rgb}$ (top) and flow fields stream $L_4^{flow} - L_2^{flow}$ (bottom). The results show a clear contrast between occipital areas better predicted by lower level layers (blue) and lateral areas better predicted from highest level layers (red), illustrating a gradient of complexity across areas.} \label{fig:L4-L1} \end{figure} The difference between predictions score from low-level layers activity of flow fields stream and high level layers activity of raw frames stream ($L_1^{flow} - L_4^{rgb}$) yields a contrast that does not match boundaries between visual areas; instead, it does coincide with the retinotopic map displaying preferred eccentricity (see Figure~\ref{fig:eccentricity-comparison}). Intuitively this means that regions where brain activity is better predicted from the highest layer of optical flow fields than from the lowest layer of raw frames stream are involved in peripheric vision whereas regions where activity is better predicted from the lowest layer of raw frames stream than from the highest layer of optical flow fields are mainly foveal. \begin{figure}[t] \centering \begin{tabular}{cc} \includegraphics[scale=0.30]{layer1-flow-layer4-rgb2.pdf}&\\ \includegraphics[scale=0.30]{eccentricity.pdf}\\ \end{tabular} \caption{ The difference between predictions score from low-level layers activity of flow fields and high-level layers activity of raw frames stream $L_1^{flow} - L_4^{rgb}$ (top) resembles the preferred eccentricity map of the same subject (bottom). Areas that are better predicted from low level flow fields streams are mostly involved in peripheric vision whereas areas better predicted from high level raw frames stream are mainly foveal.} \label{fig:eccentricity-comparison} \end{figure} We use the contrasts between high level layers and low level layers, and the eccentricity related contrast to construct a parcellation of the brain based on these contrasts (see Figure~\ref{fig:resulting-parcelisation.png}). From the 8 possible resulting profiles, three major clusters stand out allowing us to successfully depict a clustering of the voxels using contrasts from deep representation of the stimuli. \begin{figure}[t] \centering \includegraphics[scale=0.39]{areas_vertical.png} \caption{Parcellation summarizing artificial-biological correspondences: The set of active voxels were split into subgroups according to their differential response to three contrasts: $L2^{flow} - L4^{flow}$, $L2^{rgb} - L4^{rgb}$, and $L1^{flow}-L4^{rgb}$. From the 8 possible resulting profiles, 3 major clusters stand out: deep blue, $L2^{flow} > L4^{flow}$, $L2^{rgb} > L4^{rgb}$, and $L1^{flow} < L4^{rgb}$; it corresponds to a voxel set in primary visual areas that has low eccentricity (foveal regions); green, $L2^{flow} > L4^{flow}$, $L2^{rgb} > L4^{rgb}$, and $L1^{flow} > L4^{rgb}$, it corresponds to the same visual areas, but for voxels with higher eccentricity (peripheric voxels); yellow , $L2^{flow} < L4^{flow}$, $L2^{rgb} < L4^{rgb}$, and $L1^{flow} < L4^{rgb}$, it corresponds to lateral and lateral visual areas that encode more abstract representations of the objects. } \label{fig:resulting-parcelisation.png} \end{figure} \begin{comment} \subsection{Picture of the resulting response patterns} In order to improve our understanding, and obtain a complete data-drive characterization of these reponses, we further try and predict across voxels the response to three key contrasts ($L2^{flow} - L4^{flow}$, $L2^{rgb} - L4^{rgb}$, and $L2^{flow} < L2^{rgb}$) using 52 functional contrasts acquired on the same subjects in classical brain mapping experiments \cite{ibc}, plus three retinotopic maps (retinotopicity, exentricity and polar angle). This experiment is performed in a leave-one-subject-out setting and relies on a RandomForest Classifier. It turns out that the three contrast can be predicted well above chance ($L4^{flow} - L2^{flow}$: $R^2_{cv}=.4$, $L4^{rgb} - L2^{rgb}: \, R^2_{cv}=.22$, and $L1^{flow} - L4^{rgb}: \; R^2_{cv}=.04$ -- chance is negative). In Figure~\ref{fig:feature_importance}, we show which ones, among the 52 input activity maps are actually used to predict each contrast: $L4^{flow} - L2^{flow}$ is mostly predicted by retinotopicity, directly showing that this quantifies the abstraction of the representation in each location; $L4^{rgb} - L2^{rgb}$ is mostly explained by two contrasts: retinotopicity, recalling the abstract character of the representation, and also the response to places versus other visual categories, hinting at the importance of large-angle, background representations as opposed to objects in this stream, and $L1^{flow} - L4^{rgb}$ is best explained by eccentricity, showing that it represents mostly a foreground/background segregation. \begin{figure} \centering \includegraphics[scale=0.3]{feature_importance_l1-flow-l4-rgb.png} \includegraphics[scale=0.3]{feature_importance_l1-l4-flow.png} \includegraphics[scale=0.3]{feature_importance_l1-l4-rgb.png} \caption{Feature importance for the three contrasts: $L1^{flow} - L4^{rgb}$ (top), $L4^{flow} - L2^{flow}$ (middle) and $L4^{rgb} - L2^{rgb}$ (bottom)} \label{fig:feature_importance} \end{figure} \end{comment} \section{Discussion} Reproducing the results of \cite{gucclu2015deep} we have shown that lateral areas are best predicted by the last layers of both streams whereas occipital areas are best predicted by first layers of both streams. We have also shown that foveal areas are best predicted by last layers of the raw frames stream and that peripheric areas are best predicted by the first layers of the flow fields stream. We have introduced a compression procedure for video representation that does not alter too much the channel structure of the network, yielding tremendous gains in performance compared to PCA. The linear prediction from deep video features yields predictions scores that are far better than chance. However the TVL1 algorithm \cite{zach2007duality} used in the TSN network does not produce high quality flow fields. Using more recent algorithms to compute optical flow such as Flownet 2 \cite{ilg2017flownet}, our performance could be further improved. The TSN Network would have to be retrained though. In contrast to \cite{gucclu2017increasingly}, the data used to train the network are not the same as the data presented to the subjects. We rely in fact on transfer between computer vision datasets and the visual content used for visual stimulation. This transfer is imperfect: the Berkeley video dataset contains videos of landscapes and animated pictures that are not present in the Kinetic dataset, which introduces some noise. In conclusion, our study provides key insights that areas have a role linked to their retinotopic representation when performing action recognition. Future studies should focus on finessing this result by using a network tuned for other tasks. \section{Acknowledgments} This project has received funding from the European Union’s Horizon 2020 Research and Innovation Programme under Grant Agreement No. 720270 (HBP SGA1). \bibliographystyle{apacite} \setlength{\bibleftmargin}{.125in} \setlength{\bibindent}{-\bibleftmargin}

a8f8de1c25a236daba8c00d8be74033549e29f7f

\section{Introduction} We prove ABP (Alexandroff-Bakelmann-Pucci) inequalities for a class \textit{singular submanifolds of arbitrary codimension} in the Euclidean space which includes the class of varifolds without boundary and with bounded mean curvature. More specifically, we consider the following class. \begin{Definition}\label{definition of (m,h)} Suppose $ 1 \leq m \leq n $ are integers, $ \Omega $ is an open subset of $ \Real{n+1} $, $ \Gamma $ is relatively closed in $ \Omega $ and $ h \geq 0 $. We say that $ \Gamma $ is a $ (m,h) $ subset of $ \Omega $ provided it has the following property: if $ x \in \Gamma $ and $ f $ is a $ \mathscr{C}^{2} $ function in a neighbourhood of $ x $ such that $ f|\Gamma $ has a local maximum at $ x $ and $ \nabla f(x) \neq 0 $, then \begin{equation*} \trace_{m}\Der^{2}f(x) \leq h |\nabla f(x)|, \end{equation*} where $ \trace_{m} \Der^{2}f(x) $ is the sum of the lowest $ m $ eigenvalues of $ \Der^{2}f(x) $. \end{Definition} This class has been recently introduced\footnote{This definition is equivalent to \cite[2.1]{MR3466806}, by \cite[8.1]{MR3466806}.} in \cite[2.1]{MR3466806} in the analysis of the area blow-up of sequences of smooth submanifolds (or varifolds) with a uniform bound on the mean curvature. In fact, as proved in \cite[2.6]{MR3466806}, if $ M_{i} $ is a sequence of $ m $ dimensional varifolds in an open subset $ \Omega $ of $ \Real{n+1} $ with mean curvatures uniformly bounded in $ \Lp{\infty} $ by a non-negative number $ h $ and such that \begin{equation*} \limsup_{i \to \infty}|\partial M_{i}|(U) < \infty \quad \textrm{whenever $ U \subset \subset \Omega $}, \end{equation*} where $ |\partial M_{i}| $ is the measure associated to the generalized boundary of $ M_{i} $, then the smallest closed subset $ Z $ of $ \Omega $ such that the areas of $ M_{i} $ are uniformly bounded as $ i \to \infty $ on compact subsets of $ \Omega \sim Z $ is an $ (m,h) $ subset of $ \Omega $. As a simple consequence of this general fact, it follows that the support of every $ m $ dimensional varifold in $ \Omega $ with no boundary and mean curvature bounded in $ \Lp{\infty} $ by $ h $ is an $(m,h)$ subset of $ \Omega $, see \cite[2.8]{MR3466806}. Special cases of \ref{definition of (m,h)}, under the name of viscosity minimal sets, have been studied in \cite{2017arXiv170507948S} and \cite{2017arXiv170801549S}. We describe now the results of the present paper. Firstly we introduce some notation\footnote{For basic notation, see the end of the introduction.} that we use through the rest of this work. Suppose $ \mathscr{C}_{1}(0)=\mathbf{U}^{n}(0,1) \times \mathbf{R} \subseteq \mathbf{R}^{n+1} $, $ \Gamma \subseteq \mathscr{C}_{1}(0) $ is relatively closed and $ a > 0 $. For every $ x \in \mathbf{R}^{n} $ we define $ P_{a,x} $ to be the paraboloid of opening $ a $ and center $ x $ touching $ \overline{\Gamma} $ from above. For every closed set $ C \subseteq \mathbf{B}^{n}(0,1) $ we define the touching set \begin{equation*} A_{a}(\Gamma; C) \end{equation*} as the set of $ (z,\eta) $ such that $z \in \Sigma(P_{a,x}) \cap \overline{\Gamma}$ for some $ x \in C $ and $ \eta \in \Nor(\Sigma(P_{a,x}), z) $ with $ |\eta |=1 $ and $ \eta_{n+1} > 0 $. Moreover, we let \begin{equation*} A'_{a}(\Gamma; C) = \{ z' : \textrm{$(z, \eta) \in A_{a}(\Gamma; C)$ for some $ \eta \in \mathbf{S}^{n} $}\}. \end{equation*} When $ C = \mathbf{B}^{n}(0,1) $ we let $ A_{a}(\Gamma; C) = A_{a}(\Gamma) $ and $ A'_{a}(\Gamma;C) = A'_{a}(\Gamma) $. For submanifolds of arbitrary codimension we obtain the following ABP inequality: \begin{Theorem}[ABP estimate, arbitrary codimension]\label{ABP arbitrary cod} Suppose $ 1 \leq m \leq n $ are integers, $ h \geq 0 $, $ a > 0 $ and $ \Gamma $ is an $ (m,h) $ subset of $ \mathscr{C}_{1}(0) $ such that $ \Haus{m}(\Gamma) < \infty $. For every closed set $ C $ in $ \mathbf{B}^{n}(0,1) $, if $ \varnothing \neq A_{a}(\Gamma; C) \subseteq \Gamma \times \mathbf{S}^{n} $, then \begin{equation*} \Haus{n}(C) \leq \gamma \Big(1+a+\frac{1}{a}\Big)^{n-m}\Big(1 + a + \frac{h}{a}\Big)^{m} \int_{\Gamma^{(m)}}\Haus{n-m}\{ \eta : (z,\eta) \in A_{a}(\Gamma; C) \} d\Haus{m}z, \end{equation*} where $ \Gamma^{(m)} $ is the set of $ z \in \Gamma $ where $ \Gamma $ can be touched from $ n+1-m $ linearly independent directions by an open ball of $ \Real{n+1} $ and $ \gamma = 4^{n-m}(2m + 4h)^{m} $. \end{Theorem} In case of hypersurfaces, the previous result can be refined as follows: \begin{Corollary}[ABP estimate, codimension 1]\label{ABP} Suppose $ m = n $ in \ref{ABP arbitrary cod}. For every closed set $ C $ in $ \mathbf{B}^{n}(0,1) $, if $ \varnothing \neq A_{a}(\Gamma; C) \subseteq \Gamma \times \mathbf{S}^{n} $, then \begin{equation}\label{ABP:2} \Haus{n}(C) \leq \gamma (1+a+ha^{-1})^{n}(1+4a^{2})^{1/2}\Haus{n}(A'_{a}(\Gamma; C)), \end{equation} where $ \gamma = (2n + 4h)^{n} $. \end{Corollary} As a special case of \ref{ABP} we obtain the following ABP estimate proved in \cite[2.1]{MR2334822} for viscosity subsolutions of the minimal surface equation\footnote{The results in \cite{MR2334822} are actually proved for a class of elliptic operators $ F $ that are uniformly elliptic in a neighbourhood of the origin and $ F(0,p,z,x) =0 $ whenever $ p $, $ z $ and $ x $ are close to $ 0 $.} . \begin{Corollary}[Savin]\label{Savin ABP} Let $ u $ be a continuous viscosity subsolution in $ \mathbf{U}^{n}(0,1) $ of the minimal surface equation, let $ \Gamma \subseteq \Real{n+1} $ be the graph of $ u $ and let $ 0 < a \leq 1 $. Then there exists a small universal constant $ \gamma $ (depending only on $ n $) such that for every closed set $ C \subseteq \mathbf{B}^{n}(0,1) $ such that $ A'_{a}(\Gamma; C) \subseteq \mathbf{U}^{n}(0,1) $, \begin{equation*} \Haus{n}(A'_{a}(\Gamma;C)) \geq \gamma \Haus{n}(C). \end{equation*} \end{Corollary} Similar estimates have been proved in several other works since then; see, for example, \cite[3.5]{MR3500837} \cite[6.5]{MR3695374}, \cite[3.5]{MR3265174} and, in a Riemannian setting, \cite[Theorem 1.2]{MR2989991}. It is well known, starting with the prioneering work of Krilov and Safanov, that the ABP estimate plays a fundamental role in the derivation of Harnack-type estimates. In this regard \ref{ABP} can be employed to extend to $ (n,h) $ sets in $ \mathbf{R}^{n+1} $ the Measure-Estimate in \cite[6.1]{MR2757359} (the same estimate is implicitly contained in the first part of the proof of the Harnack inequality in \cite[1.1]{MR2334822}). This type of estimate is also known as weak Harnack inequality, see \cite[6.2]{MR3695374}. Having the ABP estimate at our disposal the proof goes along the line of \cite[2.2]{MR2334822}. \begin{Theorem}[Measure-Estimate]\label{Extrinsic weak Harnack inequality} For every $ n \geq 1 $ there exist $ \alpha > 1 $ and $ 0 < \beta < 1 $ such that the following statement holds. For every $ \mu > 0 $ there exists an integer $ k \geq 1 $ so that if $ \Gamma $ is an $(n,h)$ subset of $ \mathcal{C}_{1}(0) $ with $ \Haus{n}(\Gamma) < \infty $, $ 0 \leq h < \alpha^{-k-1} $, \begin{equation*} \Gamma \subseteq \{ x : x_{n+1} \leq 0 \} \quad \textrm{and} \quad \Gamma \cap \{ x : |x'|\leq 1/4, \; -\alpha^{-k-1}/48 \leq x_{n+1} \leq 0 \} \neq \varnothing, \end{equation*} then \begin{equation*} \Leb{n}(\mathbf{B}^{n}(0, 1/3) \sim A'_{\alpha^{-1}}(\Gamma)) \leq \mu \Leb{n}(\mathbf{B}^{n}(0, 1/3)). \end{equation*} \end{Theorem} We finally remark that the proof of \ref{ABP arbitrary cod} and \ref{ABP}, though based on the same general idea of \cite[2.1]{MR2334822}, use a substantially different method to treat the arbitrary-codimension setting of \ref{ABP arbitrary cod}. This method is based on the theory of curvature for arbitrary closed sets recently developed in \cite{2017arXiv170801549S} (see also \cite{MR534512}, \cite{MR2031455} for earlier contributions in this direction). The proof of \cite[2.1]{MR2334822} goes in the following way: firstly the problem is reduced to semi-concave functions by inf-convolution approximation; since such functions are differentiable on the contact set, it is possible to introduce an auxiliary function (\emph{vertex map}), mapping each contact point into the vertex of the corresponding touching paraboloid; the vertex map is Lipschitz and its derivative can be suitably estimated using the assumptions on the elliptic operator $ F $; henceforth, an application of the classical area formula for Lipschitz maps gives the conclusion. In the present paper we also use a vertex map, but now its domain is a subset of the normal bundle of the $ (m,h) $ set $ \Gamma $, namely $ A_{a}(\Gamma; C) $. Then, in order to complete the proof following the idea of \cite[2.1]{MR2334822}, we need a suitable area formula for functions whose domain is a subset of the normal bundle, see \ref{general facts}\eqref{general facts: Coarea}, and suitable estimates for the differential of the vertex map, that can be deduced from the upper bound on the trace of the second fundamental form in \ref{trace and viscosity minimal sets}. \paragraph*{Notation.}The distance function from a closed set $ \Gamma $ in $ \Real{n+1} $ is $ \bm{\delta}_{\Gamma} $ and the nearest point projection onto $ \Gamma $ is $ \bm{\xi}_{\Gamma} $. If $T$ is a linear subspace of $ \Real{n+1} $ then $ T_{\natural} $ is the orthogonal projection onto $ T $. If $ z \in \Real{n} \times \Real{} \cong \Real{n+1} $ we define $ z' \in \Real{n} $ so that \mbox{$(z', z_{n+1}) = z $.} If $ \rho > 0 $ and $ x \in \Real{n} $ we define \begin{equation*} \mathcal{C}_{\rho}(x) = \Real{n+1} \cap \{ z : |z'-x| < \rho \}, \end{equation*} \begin{equation*} \mathbf{B}^{n}(x,\rho) = \Real{n} \cap \{y : | y-x| \leq \rho \}, \quad \mathbf{U}^{n}(x,\rho) = \Real{n} \cap \{y : | y-x| < \rho \}. \end{equation*} For every function $ f : D \rightarrow \Real{} $, where $ D \subseteq \Real{n} $, we let \begin{equation*} \Sigma(f) = \{ (x,f(x)) : x \in D \} \quad \textrm{and} \quad \Sigma^{+}(f) = \{ (x,y) : y > f(x) \}. \end{equation*} If $ a > 0 $ and $ x \in \Real{n} $ then a function $ P : \Real{n} \rightarrow \Real{} $ is a \emph{paraboloid of center $ x $ and opening $ a $} if there exists $ t \in \Real{} $ such that \begin{equation*} P(y) = (a/2)|y-x|^{2} + t \quad \textrm{whenever $ y \in \Real{n} $}. \end{equation*} If $ \Gamma \subseteq \Real{n+1} $ is a closed set and $ f : \Real{n} \rightarrow \Real{} $ is continuous, we say that \emph{$ f $ touches $ \Gamma $ from above} if and only if \begin{equation*} \Sigma(f) \cap \Gamma \neq \varnothing \quad \textrm{and} \quad \Sigma^{+}(f)\cap \Gamma = \varnothing. \end{equation*} \section{Curvature of closed sets} \label{Curvature of minimal varieties} Suppose $ n \geq 1 $ is an integer and $ \Gamma \subseteq \Real{n+1} $ is closed. For every $ r > 0 $ we define \begin{equation*} N_{r}(\Gamma)= (\Gamma \times \mathbf{S}^{n}) \cap \{ (z,\eta): \bm{\delta}_{\Gamma}(z+r\eta) = r \}; \end{equation*} the \emph{generalized normal bundle of $ \Gamma $} is then given by \begin{equation*} N(\Gamma) = \bigcup_{r > 0}N_{r}(\Gamma). \end{equation*} If $ 0 \leq m \leq n+1 $ the \emph{$ m $-th stratum of $ \Gamma $} is \begin{equation*} \Gamma^{(m)} = \Gamma \cap \{ z : \dim \bm{\xi}_{\Gamma}^{-1}\{z\} = n+1-m \}. \end{equation*} It is not difficult to see that $ \Gamma^{(m)} $ is the set of $ z \in \Gamma $ where $ \Gamma $ can be touched from $ n+1-m $ linearly independent directions by an open ball of $ \Real{n+1} $; see \cite[5.1-5.3]{2017arXiv170801549S}. We prove in \cite[3.10(1)]{2017arXiv170801549S} that $ N_{r}(\Gamma) $ is a locally $ \rect{n} $ rectifiable closed subset of $ \Real{n+1} \times \mathbf{S}^{n} $ and we deduce that $ N(\Gamma) $ is a countably $ \rect{n} $ rectifiable Borel subset of $ \Real{n+1} \times \mathbf{S}^{n} $ since $ N_{r}(\Gamma) \subseteq N_{s}(\Gamma) $ if $ s < r $. We remark that $ N(\Gamma) $ coincides with the normal bundle introduced in \cite[\S 2.1]{MR2031455} and it is well known that if $ \Gamma $ is a set of positive reach then $ N(\Gamma) $ coincides with the classical unit normal bundle, as defined in \cite[3.1.21]{MR0257325}, and it has locally finite $ \Haus{n} $ measure (see \cite[4.4]{2017arXiv170801549S}). In \cite[4.5-4.12]{2017arXiv170801549S} we prove that if $ \Gamma $ is an arbitrary closed subset of $ \Real{n+1} $ then for $ \Haus{n} $ a.e.\ $(z,\eta) \in N(\Gamma) $ there exist a linear subspace $ T_{\Gamma}(z,\eta) $ of $ \Real{n+1} $ with $ \dim T_{\Gamma}(z,\eta) \geq 0 $ and a bilinear form $ Q_{\Gamma}(z,\eta) :T_{\Gamma}(z,\eta) \times T_{\Gamma}(z,\eta) \rightarrow \Real{} $ that coincides with the second fundamental form introduced in \cite[4.5]{MR1021369} when $ \Gamma $ is a set of positive reach. Henceforth we refer to this bilinear form as \emph{the second fundamental form of $ \Gamma $.} The principal curvatures $ - \infty < \kappa_{1}(z,\eta) \leq \ldots \leq \kappa_{n-1}(z,\eta) \leq \infty $ of $ \Gamma $ at $ (z,\eta) $ can be defined as \begin{equation*} \kappa_{m+1}(z,\eta) = \infty, \quad \textrm{$ \kappa_{1}(z,\eta), \ldots , \kappa_{m}(z,\eta) $ are the eigenvalues of $ Q_{\Gamma}(z,\eta) $} \end{equation*} where $ m = \dim T_{\Gamma}(z,\eta) $. They coincide $ \Haus{n} $ almost everywhere in $ N(\Gamma) $ with the principal curvatures introduced in \cite[p.\ 244]{MR2031455} by different methods, see \cite[4.16]{2017arXiv170801549S}. We remark that the functions mapping $ \Haus{n} $ almost all $ (z,\eta) \in N(\Gamma) $ into $ T_{\Gamma}(z,\eta)_{\natural} \in \Hom(\Real{n+1}, \Real{n+1}) $ and $ Q_{\Gamma}(z, \eta) \circ \bigodot_{2}T_{\Gamma}(z,\eta)_{\natural} \in \Hom(\bigodot_{2}\Real{n+1}) $ are $ \Haus{n} $ measurable. For reader's convenience in the next theorem we summarize two facts from the general theory developed in \cite{2017arXiv170801549S} that we employ in the proof of the ABP estimate. \begin{Theorem}\label{general facts} Let $ \Gamma $ be a closed subset of $ \Real{n+1} $. Then the following statements hold. \begin{enumerate} \item\label{general facts:basis} For every $ r > 0 $ and for $ \Haus{n} $ a.e.\ $ (z,\eta) \in N_{r}(\Gamma) $ the approximate tangent cone $ \Tan^{n}(\Haus{n}\restrict N_{r}(\Gamma), (z,\eta)) $ is an $ n $-dimensional plane and there exist $ u_{1}, \ldots , u_{n} \in \Real{n+1} $ such that $ \{ u_{1}, \ldots , u_{n}, \eta \} $ is an orthonormal basis of $\Real{n+1}$ and the orthonormal vectors \begin{equation*} \bigg(\frac{1}{(1+\kappa_{i}(z,\eta)^{2})^{1/2}} u_{i}, \frac{\kappa_{i}(z,\eta)}{(1+\kappa_{i}(z,\eta)^{2})^{1/2}} u_{i} \bigg) \end{equation*} form a basis of $ \Tan^{n}(\Haus{n}\restrict N_{r}(\Gamma), (z,\eta)) $ (if $ \kappa_{i}(z, \eta) = \infty $ then the corresponding vector equals $ (0, u_{i}) $). \item\label{general facts: Coarea} If $f$ is a $ (\Haus{n}\restrict N(\Gamma)) $ integrable $ \overline{\Real{}} $ valued function, then \begin{flalign*} & \int_{N(\Gamma)|\Gamma^{(m)}} f(a,u)\, \prod_{i=1}^{m}\frac{1}{(1+\kappa_{i}(a,u)^{2})^{1/2}} \, d\Haus{n}(a,u) \\ & \quad = \int_{\Gamma^{(m)}}\int_{\{z\} \times N(\Gamma,z)}f \, d\Haus{n-m}\, d\Haus{m}z. \end{flalign*} \end{enumerate} \end{Theorem} \begin{proof} See \cite[4.14, 5.6]{2017arXiv170801549S}. \end{proof} In order to combine this measure-theoretic theory with techniques developed in PDE's theory, we have to analize the behaviour of the principal curvatures introduced above in order to produce useful estimates. At this point it is worth to recall that if $ \kappa_{1} $ is the principal curvature of the graph $ \Gamma $ of the primitive of the (ternary) Cantor function, which is a convex $ \mathscr{C}^{1} $ function from $ \Real{} $ to $ \Real{} $, then, letting $ L = \bigcap_{r > 0}N_{r}(\Gamma) $, there exists $ M \subseteq L $ such that $ \Haus{1}(M)> 0 $ and \begin{equation*} \kappa_{1}(z,\eta) = \infty \quad \textrm{for every $(z,\eta) \in M $.} \end{equation*} The reader may consult \cite[5.11]{2017arXiv170801549S} for details on this example. In order to prove the main result of this section, firstly we need a generalization of the barrier principle in \cite[7.1]{MR3466806}. \begin{Lemma}\label{weak maximum principle} Suppose $ 1 \leq m \leq n $ are integers, $ f : \Real{n} \rightarrow \Real{} $ is function pointwise differentiable of order $ 2 $ at $ 0 $ such that $ f(0) =0 $ and $ \Der f(0) =0 $, $ h \geq 0 $, $ \Omega $ is an open subset of $ \Real{n+1} $, $ \Gamma $ is an $ (m,h) $ subset of $ \Omega $ such that $ 0 \in \Gamma $ and \begin{equation*} \Gamma \cap V \subseteq \{ z : z_{n+1}\leq f(z') \} \end{equation*} for some open neighbourhood $ V $ of $ 0 $. Then, denoting by $ \chi_{1} \geq \ldots \geq \chi_{n} $ the eigenvalues of $ \Der^{2}f(0) $, it follows that \begin{equation*} \chi_{1} + \ldots + \chi_{m} \geq -h. \end{equation*} \end{Lemma} \begin{proof} Let $ \epsilon > 0 $ and $ \psi(x) = \frac{1}{2}\Der^{2}f(0)(x,x) + \epsilon |x|^{2} $ for $ x \in \mathbf{R}^{n} $. There exists $ r > 0 $ such that $ f(x)\leq \psi(x) $ for every $ x \in \mathbf{U}^{n}(0,r) $. By \cite[7.1]{MR3466806}, if $ \kappa_{1} \leq \ldots \leq \kappa_{n} $ are the principal curvatures of $ M = \{(x, \psi(x)) : x \in \mathbf{R}^{n} \} $ at $ (0,0) $ with respect to the unit normal that points into $ \{ z : z_{n+1} \leq \psi(z') \} $, then \begin{equation*} \kappa_{1} + \ldots + \kappa_{m} \leq h. \end{equation*} Since a standard and straightforward computation shows that $ \kappa_{i} = -\chi_{i} -\epsilon $ for $ i = 1, \ldots , n $, we obtain the conclusion letting $ \epsilon \to 0 $. \end{proof} The main result of this section provides the crucial estimate for the trace of the second fundamental form of an $ (m,h) $ set. As the example of the primitive of the Cantor function clearly shows, this estimate cannot be expected to hold without a specific geometric hypothesis (as the weak maximum principle for $ (m,h) $ sets). The proof of the following theorem is adapted from \cite[7.5]{2017arXiv170801549S}, where certain stationary submanifolds (viscosity minimal sets) are treated. \begin{Theorem}\label{trace and viscosity minimal sets} If $ \Omega $ is an open subset in $ \Real{n+1} $, $\Gamma$ is an $ (m,h) $ subset of $ \Omega $ that is a countable union of sets with finite $ \Haus{m} $ measure and if $ \kappa_{1} \leq \ldots \leq \kappa_{n} $ are the principal curvatures \mbox{of $ \overline{\Gamma} $,} then $ \kappa_{m+1}(z, \eta) = \infty $ and \begin{equation*} \textstyle \sum_{i=1}^{m}\kappa_{i}(z,\eta) = \trace Q_{\overline{\Gamma}}(z, \eta) \leq h \end{equation*} for $ \Haus{n} $ a.e.\ $ (z, \eta) \in N(\overline{\Gamma}) \cap \{ (z, \nu) : z \in \Gamma \} $. In particular, if $ \Gamma $ is the support of an $ m $-dimensional integral varifold in $ \Omega $ without boundary and with bounded variational mean curvature $ H $, then \begin{equation*} \textstyle \sum_{i=1}^{m}\kappa_{i}(z,\eta) = \trace Q_{\overline{\Gamma}}(z, \eta) = -H(z) \bullet \eta \end{equation*} for $ \Haus{n} $ a.e.\ $ (z, \eta) \in N(\overline{\Gamma}) \cap \{ (z, \nu) : z \in \Gamma \} $. \end{Theorem} \begin{proof} Suppose $ A = \overline{\Gamma} $, $ \lambda > (2m-1) + 1 $ and the functions $ \bm{\nu}_{A} $ and $ \bm{\psi}_{A} $ are defined as in \cite[3.1]{2017arXiv170801549S}. Given such $ \lambda $ we define for $ r > 0 $ the set $ N_{r} $ as in \cite[3.10]{2017arXiv170801549S}. We recall from the proof of \cite[4.14]{2017arXiv170801549S} that for $ \Leb{1} $ a.e.\ $ r > 0 $ and for $ \Haus{n} $ a.e.\ $ x \in N_{r} $, if $ \chi_{1} \leq \ldots \leq \chi_{n} $ are the eigenvalues of $ \ap \Der \bm{\nu}_{A}(x)| \Tan^{n}(\Haus{n}\restrict N_{r}, x) $ then \begin{equation}\label{trace and viscosity minimal sets:1} \kappa_{i}(\bm{\psi}_{A}(x)) = \chi_{i}(1-r\chi_{i})^{-1} \quad \textrm{for $ i = 1, \ldots , \dim T_{A}(\bm{\psi}_{A}(x)) $.} \end{equation} We select $ 0 < r < (2h)^{-1} $ so that \cite[3.10(3),(4)]{2017arXiv170801549S} hold for $ \Haus{n} $ a.e.\ $ x \in N_{r} $; then we fix $ x \in N_{r} \cap \bm{\xi}_{A}^{-1}(\Gamma) $ satisfying the conclusions of \cite[3.10(2),(3),(4)]{2017arXiv170801549S}. We assume $ \bm{\xi}_{A}(x) =0 $ and let $ T = \{ v : v \bullet \bm{\nu}_{A} (x) =0 \} $. It follows that there exists a Lipschitzian function $ f : T \rightarrow T^{\perp} $ such that $ f $ is pointwise differentiable of order $ 2 $ at $ 0 $, $ \Der f(0) =0 $, \begin{equation*} \Der^{2}f(0)(u,v) \bullet \bm{\nu}_{A}(x) = -\ap\Der \bm{\nu}_{A}(x)(u) \bullet v \quad \textrm{whenever $ u , v \in T $,} \end{equation*} \begin{equation*} W \cap \bm{\delta}_{A}^{-1}\{r\} = W \cap \{ \chi + f(\chi) : \chi \in T \} \quad \textrm{for some neighbourhood $ W $ of $ x $.} \end{equation*} Choose $ 0 < s < r/2 $ such that $ \mathbf{U}(x,s) \subseteq W $ and let $ g(\zeta) = f(\zeta) - x $ for $ \zeta \in T $, \begin{equation*} U= T_{\natural}\big(\mathbf{U}(x,s) \cap \{ \chi + f(\chi) : \chi \in T \}\big), \quad V = \{ y-x: y \in T_{\natural}^{-1}(U) \cap \mathbf{U}(x,s) \}. \end{equation*} Then $ V $ is an open neighbourhood of $ 0 $ and \begin{equation*} V \cap A \subseteq \{ z : z \bullet \bm{\nu}_{A}(z) \leq g(T_{\natural}(z)) \bullet \bm{\nu}_{A}(z) \}. \end{equation*} This inclusion can be proved as follows: if there was $y \in \mathbf{U}(x,s) \cap T_{\natural}^{-1}[U]$ such that $y-x \in A $ and $ y \bullet \bm{\nu}_{A}(x) > f(T_{\natural}(y))\bullet \bm{\nu}_{A}(x) $ then, noting that $ \bm{\nu}_{A}(x) = r^{-1}x $, \begin{equation*} T_{\natural}(y) + f(T_{\natural}(y)) \in \mathbf{U}(x,s) \cap \bm{\delta}_{A}^{-1}\{r\}, \quad |T_{\natural}(y) + f(T_{\natural}(y))-y| < r, \end{equation*} we would conclude \begin{equation*} |T_{\natural}(y) + f(T_{\natural}(y))-(y-x)| = r - (y-f(T_{\natural}(y)))\bullet \bm{\nu}_{A}(x)< r \end{equation*} which is a contradiction. Let $ \chi_{1} \leq \ldots \leq \chi_{n} $ be the eigenvalues of $\ap \Der \bm{\nu}_{A}(x)|T$. Then $ 1 - \chi_{1}r, \ldots , 1 - \chi_{n}r $ are the eigenvalues of $ \ap \Der \bm{\xi}_{A}(x)|T$ by \cite[3.5]{2017arXiv170801549S} and $ -\chi_{1}, \ldots , - \chi_{n} $ are the eigenvalues of $ \pt\Der^{2} g(0) \bullet \bm{\nu}_{A}(x) $. Therefore we apply a rotated version of \ref{weak maximum principle} to conclude that \begin{equation}\label{trace and viscosity minimal sets:2} \chi_{1} + \ldots + \chi_{m} \leq h. \end{equation} Since $ \chi_{j} \geq -(\lambda -1)^{-1}r^{-1}$ whenever $ j = 1 , \ldots , n $ by \cite[3.10(2)]{2017arXiv170801549S}, we get \begin{equation*} \chi_{j} - (m-1)(\lambda-1)^{-1}r^{-1} \leq \sum_{i=1}^{m} \chi_{i} \leq h, \end{equation*} whence we conclude $ \chi_{j} < r^{-1} $ whenever $ j = 1 , \ldots , m $. Therefore, \begin{equation*} \| \textstyle \bigwedge_{m}\big((\Haus{n}\restrict N_{r},n)\ap \Der \bm{\xi}_{A}(x)\big) \| > 0. \end{equation*} We remind that the assertions proved so far hold for $\Haus{n}$ a.e.\ $ x \in N_{r} \cap \bm{\xi}_{A}^{-1}(\Gamma) $ and for $ \Leb{1} $ a.e.\ $ 0 < r < (2h)^{-1} $. We prove now the following \emph{Lusin $(N)$ condition}: if $ S \subseteq \Gamma $ and $ \Haus{m}(S \cap A^{(m)}) =0 $ then $ \Haus{n}(N(A) \cap \{ (x,\nu) : x \in S\} ) =0 $. Using \cite[3.10(1), 4.3, 5.3]{2017arXiv170801549S} we get that \begin{equation*} A \cap \{ x : \Haus{n-m}(\bm{\xi}_{A}^{-1}\{x\} \cap N_{r} ) >0 \} \subseteq \bigcup_{i=0}^{m}A^{(i)} \end{equation*} for every $ r > 0 $ and, noting that $ \Haus{m}(A^{(i)}) =0 $ for $ i = 0, \ldots , m-1 $ by \cite[5.2]{2017arXiv170801549S}, we deduce \begin{equation*} \Haus{m}\big( S \cap \{ x : \Haus{n-m}(\bm{\xi}_{A}^{-1}\{x\} \cap N_{r} ) >0 \} \big) =0. \end{equation*} Noting \cite[3.10(1)(2)]{2017arXiv170801549S}, we apply \cite[7.4]{2017arXiv170801549S} with $ W $ and $ f $ replaced by $ N_{r} \cap \bm{\xi}_{A}^{-1}(\Gamma) $ and $ \bm{\xi}_{A} $ to get \begin{equation*} \Haus{n}(\bm{\xi}_{A}^{-1}(S) \cap N_{r} ) =0 \quad \textrm{for $ \Leb{1} $ a.e.\ $ 0 < r < (2h)^{-1} $,} \end{equation*} whence the desired Lusin condition follows from \cite[4.3]{2017arXiv170801549S}. By \cite[5.2, 5.9]{2017arXiv170801549S} it follows that $ \dim T_{A}(z,\eta) = m $ and $ \kappa_{m+1}(z,\eta) = \infty $ for $ \Haus{n} $ a.e.\ $ (z, \eta) \in N(A) \cap \{ (z, \nu) : z \in \Gamma \} $. Moreover we combine \eqref{trace and viscosity minimal sets:1} and \eqref{trace and viscosity minimal sets:2} to infer that \begin{equation}\label{trace and viscosity minimal sets:3} \sum_{i=1}^{m}\frac{\kappa_{i}(\bm{\psi}_{A}(x))}{1 + r \kappa_{i}(\bm{\psi}_{A}(x))} \leq h \end{equation} for $ \Haus{n} $ a.e.\ $ x \in N_{r} \cap \bm{\xi}_{A}^{-1}(\Gamma) $ and for $ \Leb{1} $ a.e.\ $ 0 < r < (2h)^{-1} $. We choose a sequence $ r_{i} > 0 $ converging to $ 0 $ such that if $ M_{i} $ is the set of $ x \in N_{r_{i}} \cap \bm{\xi}_{A}^{-1}(\Gamma) $ such that \eqref{trace and viscosity minimal sets:3} is satisfied, then $ \Haus{n}(N_{r} \cap \bm{\xi}_{A}^{-1}(\Gamma) \sim M_{i}) =0 $ for every $ i \geq 1 $. Then we readily infer that \begin{equation*} \trace Q_{A}(z,\eta) \leq h \end{equation*} for every $ (z,\eta) \in \bigcap_{i=1}^{\infty}\bigcup_{j=i}^{\infty}\bm{\psi}_{A}(M_{i}) $ and, noting that $ \bm{\psi}_{A}(N_{r}) \subseteq \bm{\psi}_{A}(N_{s}) $ if $ s < r $, we additionally get that \begin{equation*} \textstyle \Haus{n}\Big( N(A) \cap \{ (z, \nu) : z \in \Gamma \} \sim \bigcap_{i=1}^{\infty}\bigcup_{j=i}^{\infty}\bm{\psi}_{A}(M_{i}) \Big) =0. \end{equation*} If $ \Gamma $ is the support of an $ m $ dimensional integral varifold in $ \Omega $ without boundary and with mean curvature bounded in $ \Lp{\infty} $ by $ h $, then, noting that $ \Gamma $ is an $ (m,h) $ subset of $ \Omega $ by \cite[2.8]{MR3466806}, we may achieve the conclusion using the Lusin $(N)$ condition above in combination with \cite[5.2, 5.9]{2017arXiv170801549S} and \cite[4.2]{MR2472179}. \end{proof} \begin{Remark} We may conjecture that \ref{trace and viscosity minimal sets} could also be proved for classes of \emph{singular submanifolds of possibly unbounded mean curvature}. For example, if $ \Gamma $ is the support of an $ m $ dimensional integral varifold in $ \Real{n+1} $ without boundary and with mean curvature $ H $ in $ \Lp{m} $, is it true that $ \kappa_{m+1}(z,\eta) = \infty $ and \begin{equation*} \textstyle \sum_{i=1}^{m}\kappa_{i}(z,\eta) = - H(z) \bullet \eta \end{equation*} for $ \Haus{n} $ a.e.\ $ (z,\eta) \in N(\Gamma) $? More generally, what can be concluded in case of rectifiable varifolds with a uniform lower bound on the density? To treat these cases the method of \ref{trace and viscosity minimal sets} is not powerful enough and new refined tools seem to be necessary. \end{Remark} \section{Proof of \ref{ABP arbitrary cod} and \ref{ABP}} We notice that $ A_{a}(\Gamma; C) $, $ A'_{a}(\Gamma; C) $ and \begin{equation*} \{ z : \textrm{$(z, \eta) \in A_{a}(\Gamma; C)$ for some $ \eta \in \mathbf{S}^{n} $} \} \end{equation*} are closed subset of $ \overline{\Gamma} \times \mathbf{S}^{n} $, $ \mathbf{B}^{n}(0,1) $ and $ \overline{\Gamma} $ respectively. \begin{proof*}[Proof of \ref{ABP arbitrary cod}] Clearly, for every $ (w,\eta) \in A_{a}(\Gamma; C) $ there exists a unique $ x \in C $ such that $ w \in \Sigma(P_{a,x}) $ and $ \eta \in \Nor(\Sigma(P_{a,x}),w) $; in fact \begin{equation*} x = w' + a^{-1}\eta_{n+1}^{-1}\eta'. \end{equation*} Henceforth, we define the map $ \Psi : A_{a}(\Gamma; C) \rightarrow C $ so that $ \Psi(w,\eta)= w' + \frac{1}{a\eta_{n+1}}\eta' $ and we notice that $ \Psi[A_{a}(\Gamma; C)]= C $. It is not difficult to check that\footnote{In fact, a straightforward computation shows that \emph{if $ n \geq 1 $ is an integer, $ P $ is a paraboloid of center $ 0 $ and opening $ a > 0 $, $ x \in \Real{n} $ and $ \eta \in \Nor(\Sigma(P), (x, P(x)) ) $ with $ |\eta |= 1 $ and $ \eta_{n+1} > 0 $, then \begin{equation*} \sup\{s : \bm{\delta}_{\Sigma(P)}((x,P(x)) + s \eta ) = s \} = (a\eta_{n+1})^{-1}. \end{equation*}}} \begin{equation*} \bm{\delta}_{\overline{\Gamma}}\big(w + (a\eta_{n+1})^{-1}\eta\big) = (a\eta_{n+1})^{-1} \end{equation*} whenever $ (w, \eta) \in A_{a}(\Gamma; C) $. Henceforth $ A_{a}(\Gamma; C) \subseteq N_{a^{-1}}(\overline{\Gamma}) $ and, if $ \kappa_{1}, \ldots , \kappa_{n} $ are the principal curvatures of $ \overline{\Gamma} $, it follows from \cite[4.12]{2017arXiv170801549S} that \begin{equation}\label{ABP arbitrary cod:1} \kappa_{i}(w, \eta) \geq - a\eta_{n+1} \end{equation} for $ \Haus{n} $ a.e.\ $ (w, \eta) \in A_{a}(\Gamma; C) $ and $ i = 1, \ldots , n $. Employing \ref{trace and viscosity minimal sets} we deduce that \begin{equation}\label{ABP arbitrary cod:2} \kappa_{i}(w, \eta) \leq (m-1) a \eta_{n+1} + h, \quad \kappa_{m+1}(w,\eta) = \infty, \end{equation} for $ \Haus{n} $ a.e.\ $ (w,\eta) \in A_{a}(\Gamma; C) $. We compute now the jacobian of $ \Psi $ using the orthonormal basis provided by \ref{general facts}\eqref{general facts:basis}. For $ \Haus{n} $ a.e.\ $ (w,\eta) \in A_{a}(\Gamma; C) $, if $ u_{1}, \ldots , u_{n} $ are the orthonormal vectors provided by \ref{general facts}\eqref{general facts:basis}, noting that $ \eta_{n+1}^{-1} \leq (1+4a^{2})^{1/2}\leq 1+ 2a $ and using \eqref{ABP arbitrary cod:1}-\eqref{ABP arbitrary cod:2}, we can easily estimate \begin{equation*} | \Der \Psi(w,\eta)(0,u_{i})| \leq \frac{1}{a}\Big( \frac{1}{\eta_{n+1}} + \frac{|\eta'|}{\eta_{n+1}^{2}}\Big) \leq \frac{1}{a\eta_{n+1}} + \frac{2}{\eta_{n+1}}\leq 4 + 4a + \frac{1}{a} \end{equation*} for $ i = m+1, \ldots , n $, and \begin{flalign*} & |\Der \Psi(w,\eta)(u_{i}, \kappa_{i}(w,\eta) u_{i})| \leq 1 +\frac{|\kappa_{i}(w,\eta)|}{a\eta_{n+1}} + \frac{|\kappa_{i}(w,\eta)||\eta'|}{a\eta_{n+1}^{2}} \\ & \quad \quad \leq 1 +\frac{|\kappa_{i}(w,\eta)|}{a\eta_{n+1}} + 2\frac{|\kappa_{i}(w,\eta)|}{\eta_{n+1}} \leq \gamma \Big(1+a+\frac{h}{a}\Big), \end{flalign*} for $ i = 1, \ldots m $, where $ \gamma = 2m + 4h $. Therefore for $ \Haus{n} $ a.e.\ $ (w,\eta) \in A_{a}(\Gamma; C) $, we get \begin{flalign*} & \| \textstyle{\bigwedge_{n}} [\Der \Psi(w,\eta)|\Tan^{n}(\Haus{n}\restrict A_{a}(\Gamma; C) , (w,\eta))] \| \\ & \quad \quad \leq 4^{n-m}\gamma^{m} \Big(1+a+\frac{1}{a}\Big)^{n-m}\Big(1 + a + \frac{h}{a}\Big)^{m} \prod_{i=1}^{m}\frac{1}{(1+\kappa_{i}(w,\eta)^{2})^{1/2}}. \end{flalign*} Now we employ \cite[3.2.22(3)]{MR0257325}, \ref{general facts}\eqref{general facts: Coarea} and the Lusin $(N)$ condition obtained in the proof of \ref{trace and viscosity minimal sets} to conclude \begin{flalign*} & \Haus{n}(C) \leq \int_{C}\Haus{0}\{ (w,\eta) : \Psi(w,\eta)= x \} \,d \Haus{n}x \\ & \quad = \int_{A_{a}(\Gamma; C)| \Gamma^{(m)}} \| \textstyle{\bigwedge_{n}} [\Der \Psi(w,\eta)|\Tan^{n}(\Haus{n}\restrict A_{a}(\Gamma; C), (w,\eta))] \| \,d\Haus{n}(w,\eta) \\ & \quad \leq \gamma_{1}\int_{A_{a}(\Gamma; C)| \Gamma^{(m)}} \prod_{i=1}^{m}\frac{1}{(1+\kappa_{i}^{2})^{1/2}} \, d\Haus{n} \\ & \quad \leq \gamma_{1} \int_{\Gamma^{(m)}}\Haus{n-m}\{ \eta : (z,\eta) \in A_{a}(\Gamma; C) \} d\Haus{m}z, \end{flalign*} where $ \gamma_{1} = 4^{n-m}\gamma^{m} \Big(1+a+\frac{1}{a}\Big)^{n-m}\Big(1 + a + \frac{h}{a}\Big)^{m} $. \end{proof*} \begin{Lemma}\label{ABP : aux lemma} Suppose $ \Gamma \subseteq \mathscr{C}_{1}(0) $ is relatively closed such that $ \Haus{n}(\Gamma) < \infty $, $ a > 0 $, $ C \subseteq \mathbf{B}^{n}(0,1) $ is closed and \begin{equation*} B= \{ z : \textrm{$(z, \eta) \in A_{a}(\Gamma; C)$ for some $ \eta \in \mathbf{S}^{n} $} \}. \end{equation*} Then \begin{equation*} \Haus{n}(B) \leq (1+4a^{2})^{1/2}\Haus{n}(A'_{a}(\Gamma;C)). \end{equation*} \end{Lemma} \begin{proof} If $ z \in B $, $ x \in C $ and $ \nu \in \Real{n+1} $ such that $ z \in \Sigma(P_{a,x}) $ and $ \nu \in \Nor(\Sigma(P_{a,x}),z) $ with $ \nu_{n+1}> 0 $, then elementary considerations imply that for every $ v \in \Real{n+1} $ with $ v \bullet \nu > 0 $ there are $ \epsilon > 0 $ and $ r > 0 $ such that \begin{equation*} \{ y : | s(y-z)-v|< \epsilon \; \textrm{for some $ s > 0 $} \} \cap \mathbf{U}(z,r) \subseteq \Sigma^{+}(P_{a,x}), \end{equation*} whence we deduce that $ \Tan^{n}(\Haus{n}\restrict B,z) \subseteq \Real{n+1} \cap \{ w : w \bullet \nu \leq 0 \} $ by \cite[3.2.16]{MR0257325}. Since $ B $ is $ n $ rectifiable by \cite[4.12]{2017arXiv170309561M}, we deduce from \cite[3.2.19]{MR0257325} that for $ \Haus{n} $ a.e.\ $ z \in B $ there exists $ x \in C $ such that $ z \in \Sigma(P_{a,x}) $ and \begin{equation*} \Tan^{n}(\Haus{n}\restrict B,z) = \Tan(\Sigma(P_{a,x}),z). \end{equation*} Therefore for $ \Haus{n} $ a.e.\ $ z \in B $ there exists $ x \in C $ and an orthonormal basis $ e_{1}, \ldots , e_{n} $ of $ \Real{n} $ such that \begin{equation*} e_{n} = \nabla P_{a,x}(z')/|\nabla P_{a,x}(z')| \quad \textrm{if $ \nabla P_{a,x}(z') \neq 0 $}, \end{equation*} and \begin{equation*} \Big\{(e_{1}, 0),\; \ldots , \; (e_{n-1},0),\; \frac{(e_{n}, |\nabla P_{a,x}(z')|)}{(1 + |\nabla P_{a,x}(z')|^{2})^{1/2}}\Big\} \end{equation*} is an orthonormal basis of $ \Tan^{n}(\Haus{n}\restrict B,z) $. If $ p : B \rightarrow \Real{n} $ is given by $ p(z)= z' $ for $ z \in B $, noting that $ p $ is injective, the conclusion can be easily obtained applying \cite[3.2.20]{MR0257325} with $ W $ and $ f $ replaced by $ B $ and $ p $. \end{proof} \begin{proof*}[Proof of \ref{ABP}] Combine \ref{ABP arbitrary cod} and \ref{ABP : aux lemma}. \end{proof*} \section{Proof of \ref{Extrinsic weak Harnack inequality}}\label{section: A measure estimate} \begin{Lemma}\label{measure to point estimate} For every $ n \geq 1 $ there exist $ \alpha > 1 $ and $ 0 < \beta < 1 $ such that if $ 0 \leq h < a \leq \alpha^{-1} $, $ \Gamma $ is an $ (n,h) $ subset of $ \mathcal{C}_{1}(0) $ with $ \Haus{n}(\Gamma) < \infty $, $ x_{0} \in \mathbf{U}^{n}(0,1) $, $ r > 0 $ such that \begin{equation*} \mathbf{B}^{n}(x_{0},r) \subseteq \mathbf{U}^{n}(0,1) \quad \textrm{and} \quad A'_{a}(\Gamma) \cap \mathbf{U}^{n}(x_{0},r) \neq \varnothing, \end{equation*} then \begin{equation*} \Leb{n}(A'_{\alpha a}(\Gamma) \cap \mathbf{U}^{n}(x_{0}, r/8)) \geq \beta r^{n}. \end{equation*} \end{Lemma} \begin{proof} Let $ x_{1} \in \mathbf{B}^{n}(0,1) $ such that $ \Sigma(P_{a,x_{1}}) \cap \Gamma \cap \mathcal{C}_{r} (x_{0}) \neq \varnothing $. For every $ \gamma > 1 $ we let \begin{equation*} \phi(t) = \begin{cases} - \gamma^{-1}(16^{\gamma}-1) & \textrm{if $ 0 \leq t \leq 1/16 $}\\ -\gamma^{-1}(t^{-\gamma} - 1) & \textrm{if $ 1/16 \leq t \leq 1 $}\\ 0 & \textrm{if $ t \geq 1 $} \\ \end{cases} \end{equation*} and $ \psi(x) = P_{a,x_{1}}(x) + a r^{2} \phi(|x-x_{0}|/r) $ for every $ x \in \Real{n} $. Firstly, we prove the existence of a number $ 0 < \gamma < \infty $ depending only on $ n $ such that if $ h < a \leq (16^{\gamma + 1} + 2)^{-1} $ then\footnote{$Q_{\Sigma(\psi)}$ is the second fundamental of the $ \Sigma(\psi)$.} \begin{equation}\label{measure to point estimate:1} \trace Q_{\Sigma(\psi)}(z) \bullet (-\nabla \psi(z'),1) < -h(1 + |\nabla \psi(z')|^{2})^{1/2} \end{equation} for every $ z \in \Sigma(\psi) $ with $ r/16 < | z'-x_{0}| < r $. We fix $ x \in \Real{n} $ with $ r/16 < | x-x_{0}| < r $ and let $ z = (x, \psi(x)) $ and $ t = |x-x_{0}|/r $. We notice that \begin{equation*} |\nabla \psi(x)| \leq a(2+16^{\gamma+ 1}) \leq 1 \end{equation*} and if $ e_{1}, \ldots , e_{n} $ is an orthonormal basis of $ \Real{n} $ such that $ e_{n} = \nabla \psi(x)/ |\nabla \psi(x) | $ if $\nabla \psi(x) \neq 0 $, then \begin{equation*} \Big\{(e_{1}, 0), \ldots , (e_{n-1}, 0), \frac{(e_{n}, |\nabla \psi(x)|)}{(1+|\nabla \psi(x)|^{2})^{1/2}} \Big\} \end{equation*} is an orthonormal basis of $ \Tan(\Sigma(\psi),z) $. Henceforth, letting $ \zeta = (1+|\nabla \psi(x)|^{2})^{-1/2} $ and $ \nu = \zeta (-\nabla \psi(z'),1) $ and noting that $ 1/2 \leq \zeta \leq 1 $, we compute \begin{flalign*} & \trace Q_{\Sigma(\psi)}(z)\bullet \nu = \zeta a (n-1) (1 + t^{-\gamma -2}) \\ & \qquad\qquad - \zeta a (\gamma + 2)t^{-\gamma - 2} \sum_{i=1}^{n-1}\bigg|\frac{x-x_{0}}{|x-x_{0}|} \bullet e_{i} \bigg|^{2} + \zeta^{3}a (1 + t^{-\gamma -2}) \\ & \qquad\qquad - \zeta^{3}a (\gamma + 2)t^{-\gamma -2}\bigg|\frac{x-x_{0}}{|x-x_{0}|} \bullet e_{n} \bigg|^{2} \\ & \qquad \leq \zeta a(n-1)(1 + t^{-\gamma - 2}) - a \zeta^{3}(\gamma + 2)t^{-\gamma - 2} + a \zeta^{3}(1 + t^{-\gamma - 2}) \\ & \qquad \leq an(1 + t^{-\gamma - 2}) - \frac{a}{8}(\gamma + 2)t^{-\gamma - 2}. \end{flalign*} Since $ 0 \leq h < a $, we can evidently choose $ \gamma> 0 $ depending only on $ n $ so that the last line in the previous estimate is less than $ -h $ and \eqref{measure to point estimate:1} is proved. Next we prove that if $ h <a \leq (16^{\gamma + 1}+2)^{-1} $ then there exists $ z \in \Gamma \cap \overline{\mathcal{C}_{r/16}(x_{0})} $ such that \begin{equation}\label{measure to point estimate:2} 0 \leq P_{a,x_{1}}(z') - z_{n+1} \leq \gamma^{-1}(16^{\gamma} - 1) a r^{2}. \end{equation} In fact, since $ \Sigma^{+}(\psi) \cap \Gamma \cap \mathcal{C}_{r}(x_{0}) \neq \varnothing $ and $ \Sigma^{+}(\psi + ar^{2}\gamma^{-1}(16^{\gamma} - 1)) \cap \overline{\Gamma} = \varnothing $, we infer that there exists $ 0 < t \leq ar^{2}\gamma^{-1}(16^{\gamma} - 1) $ such that $ \psi + t $ touches $ \overline{\Gamma} $ from above. Since $ \psi(x) + t > P_{a,x_{1}}(x) $ if $ |x-x_{0}| \geq r $, we infer that \begin{equation*} \varnothing \neq \Sigma(\psi + t) \cap \Gamma \subseteq \mathcal{C}_{r}(x_{0}); \end{equation*} moreover, it follows from \ref{weak maximum principle} and \eqref{measure to point estimate:1} that \begin{equation*} \Sigma(\psi + t) \cap \Gamma \cap \big(\mathcal{C}_{r}(x_{0}) \sim \overline{\mathcal{C}_{r/16}(x_{0})} \big) = \varnothing. \end{equation*} If $ z \in \Sigma(\psi + t) \cap \Gamma \cap \overline{\mathcal{C}_{r/16}(x_{0})} $ then \begin{equation*} z_{n+1}=\psi(z')+t=P_{a,x_{1}}(z') - ar^{2}\gamma^{-1}(16^{\gamma}-1) + t \leq P_{a,x_{1}}(z') \end{equation*} and \eqref{measure to point estimate:2} follows. We fix now $ a $, $ h $ and $ z $ as in \eqref{measure to point estimate:2} and let $ \theta > 0 $ to be chosen later depending only on $ \gamma $. For every $ y \in \mathbf{B}^{n}(z', r/64) $ we select $ c_{y} \in \Real{} $ such that the paraboloid \begin{equation*} Q_{y}(x) = P_{a,x_{1}}(x) + \theta\frac{a}{2}|x-y|^{2}+c_{y} \end{equation*} touches $ \overline{\Gamma} $ from above. Noting that $ Q_{y}(z') \geq z_{n+1} $ and using \eqref{measure to point estimate:2}, we choose $ \theta $ large enough depending on $ \gamma $ so that $ Q_{y}(x) > P_{a,x_{1}}(x) $ whenever $ | x - z'| \geq r/16 $, which implies that \begin{equation*} \varnothing \neq \Sigma(Q_{y}) \cap \Gamma \subseteq \mathcal{C}_{r/16}(z') \subseteq \mathcal{C}_{r/8}(x_{0}). \end{equation*} Since $ Q_{y} $ is a paraboloid of center $ (\theta/(1+\theta))y + (1/(1+\theta))x_{1} $ and opening $ (\theta+1)a $ we can apply \ref{ABP} with $ C $ replaced by \begin{equation*} \bigg\{ \frac{\theta}{1 + \theta}y + \frac{1}{\theta + 1}x_{1} : y \in \mathbf{B}^{n}(z', r/64) \bigg\} \subseteq \mathbf{B}^{n}(0,1) \end{equation*} to obtain the conclusion. \end{proof} \begin{Remark} The proof of \ref{measure to point estimate} closely follows \cite[2.2]{MR2334822}, where continuous viscosity super-solutions of certain elliptic equations (including the minimal surface equation) are treated. \end{Remark} \begin{proof*}[Proof of \ref{Extrinsic weak Harnack inequality}] This is a standard argument that combines the Vitali-type covering lemma in \cite[2.3]{MR2334822} with the measure-estimate \ref{measure to point estimate}. Let $ \alpha $ and $ \beta $ be as in \ref{measure to point estimate}, $ \epsilon = (48\alpha^{k+1})^{-1} $, $ \mu > 0 $ and let $ k $ be a positive integer to be chosen later depending only on $ n $ and $ \mu $. If $ P $ is the paraboloid of center $ 0 $ and opening $ 48\epsilon $ touching $ \overline{\Gamma} $ from above, we observe that the set of touching points of $ P $ with $\overline{\Gamma}$ is contained in $ \mathcal{C}_{1/3}(0) $. For every integer $ j \geq 0 $ we let $ F_{j} = \mathbf{B}^{n}(0,1/3) \cap A'_{48\epsilon\alpha^{j}}(\Gamma) $ and, noting that $ F_{0} \neq \varnothing $, we combine \cite[2.3]{MR2334822} (or \cite[6.4]{MR2757359}) and \ref{measure to point estimate} to conclude \begin{equation*} \Leb{n}(\mathbf{B}^{n}(0,1/3) \sim A'_{\alpha^{-1}}(\Gamma)) \leq (1-\beta_{1})^{k}\Leb{n}(\mathbf{B}^{n}(0,1/3)), \end{equation*} where $ \beta_{1} $ depends only on $ \beta $. Now we choose $ k $ so that $ (1-\beta_{1})^{k} \leq \mu $. \end{proof*}

1562a00876353b909e7fc58d84a5521ca539702d

\section{Introduction} The 45-m radio telescope in the Nobeyama Radio Observatory (NRO) is one of the largest millimeter-wave telescopes in the world (e.g., see \cite{mor81}; \cite{aka83}). Many observational studies have been conducted using this telescope, including investigations of star formation, galaxy evolution, and interstellar chemistry. The NRO 45-m telescope has been employed in a wide range of research fields and it is clear that it is an important instrument at present. The stable operation of receivers and the development of receivers with new capabilities using advanced technology are important to continue making new scientific discoveries. Several receivers have been developed and operated at the NRO in the millimeter band, especially at $\lambda\sim$3 mm, which is the highest frequency range in the NRO 45-m telescope. For example, \citet{ina87} developed a superconductor-insulator-superconductor (SIS) mixer in 80--120 GHz and constructed a single-beam receiver designated as S100. This SIS mixer was fabricated as an Nb/Al-AlOx/Nb junction and it was constructed with the four junctions connected in series. The mixer mount had two mechanical tuners, which comprised a backshort plunger and a stub-tuner for adjusting the source impedance. Subsequently, \citet{sun93} reported a tunerless mixer mount for the same frequency band. They calculated the embedding impedance as viewed from the SIS junctions and the expected double sideband (DSB) receiver noise temperature based on the theoretical design of the mixer mount. Thus, they successfully developed a low noise SIS receiver ($T_{\rm RX}$(DSB) = 35--70 K) at $f_{\rm LO}$ = 80--120 GHz. Moreover, \citet{sun95} developed a focal plane array receiver designated as S115Q. This receiver was installed with fixed tuned single sideband (SSB) mixer mounts and it had four (2$\times$2) beams. The SSB receiver noise temperature was measured as about 150--200 K at $f_{\rm LO}$ = 102--110 GHz. In 2000, a new multi-beam receiver was developed in the 100-GHz band (25 BEam Array Receiver System; BEARS) and installed in the 45-m telescope \citep{sun00}. This receiver was a focal plane array and it was constructed with 25 (5$\times$5) DSB mixers and feed horns \citep{yam00}. Over the next 10 years, several new receivers were developed with novel waveguide-type sideband-separating (2SB) mixer technology and they entered open-use operation in the 45-m telescope. This mixer can separate the radio frequency (RF) signal into an USB (upper sideband) and a LSB (lower sideband) by using a quadrature hybrid in either the signal or local oscillator (LO) path (\cite{ker98}; S. M. X. Claude et al. 2000\footnote{Claude, S. M. X., et al. ALMA Memo 316, 2000\\ $\langle$http://library.nrao.edu/public/memos/alma/main/memo316.pdf$\rangle$.}). Operation of the 2SB mode is required for the receivers on the Atacama Large Millimeter/sub-millimeter Array (ALMA), and many 2SB mixers and receiver systems were developed during this period for millimeter and sub-millimeter telescopes throughout the world. For example, \citet{asa03} and S. Asayama et al. (2004)\footnote{Asayama, S., et al. ALMA Memo 481, 2004\\ $\langle$http://library.nrao.edu/public/memos/alma/main/memo481.pdf$\rangle$.} developed an integrated 2SB mixer in the 100-GHz band, and conducted the first astronomical observations with the waveguide-type 2SB mixer on the prototype ALMA antenna. Moreover, E. F. Lauria et al. (2006)\footnote{Lauria, E. F., et al. ALMA Memo 553, 2006\\ $\langle$http://library.nrao.edu/public/memos/alma/main/memo553.pdf$\rangle$.} and \citet{nak07} reported the first astronomical observations with the SMT (Submillimeter Telescope) and the first large-scale multi-line mapping observations with the 60-cm survey telescope, respectively, with the 2SB mixer in the 200 GHz band. At the NRO, a new single-beam 2SB receiver was developed in the 100-GHz band and installed in the 45-m telescope during 2007. This receiver designated as T100 comprises an OMT (ortho-mode transducer) and two 2SB mixers, which are both based on a waveguide technique. \citet{nak08} reported the first astronomical observation using the waveguide-type dual-polarization 2SB receiver system in the 100-GHz band. At that time, the SIS mixer device in the 100-GHz band in the 45-m telescope was a parallel-connected twin-junction (PCTJ) type device. The PCTJ circuit comprises two identical SIS junctions and a stripline inductor to tune out the capacitance of the junctions. The junctions are connected in parallel through the stripline inductor (\cite{nog95}; \cite{shi97}). \citet{asa03} developed a DSB mixer using a PCTJ circuit device and reported that the receiver noise temperature was less than 25 K in the LO frequency range of 95--120 GHz, with an intermediate frequency (IF) range of 4.0--7.5 GHz. Moreover, the SSB receiver noise temperature was measured for the 2SB mixer using this PCTJ device as lower than approximately 50 K over an RF frequency range of 85--120 GHz with 4.0-8.0 GHz IF in the new 2SB receiver system on the 45-m telescope \citep{nak13}. The mixer for the PCTJ device has high performance, as mentioned above, so this design was employed in the mixer for the ALMA band-4 ($f_{\rm RF}$ = 125--163 GHz) and band-8 ($f_{\rm RF}$ = 385--500 GHz) receivers (\cite{sat08}; \cite{asa14}). However, we found that the mixer with the PCTJ device in the 100-GHz band on the 45-m telescope exhibited severe gain saturation when connected to a load of 273 K during chopper-wheel calibration. In fact, we confirmed that the molecular line peak intensity of the standard calibration source for the 45-m telescope, such as the $^{12}$CO ($J$ = 1--0) line in Orion KL, increased by more than 10 \% compared with that for the old array junction device. It is well known that the dynamic range of an SIS mixer depends on $N$ and $f$, where $N$ is the number of series-connected junctions and $f$ is the frequency (\cite{tuc85}), and a theoretical explanation of the gain compression was provided by A. R. Kerr (2002)\footnote{Kerr, A. R. ALMA Memo 401, 2002\\ $\langle$http://library.nrao.edu/public/memos/alma/main/memo401.pdf$\rangle$.}. Therefore, our design of the PCTJ circuit device, with $N$ = 2, exhibited excessively large gain compression at least in the 100-GHz band, and thus it is important to develop a device with a series array of SIS junctions with $N >$2 to improve the linear performance. In this study, we developed a novel series array of SIS junction device in the 100-GHz band for a new multi-beam receiver system called FOREST (FOur beam REceiver System on the 45-m Telescope). We developed this SIS mixer to meet the following aims. \begin{enumerate} \item To design a tuning circuit for a series array of SIS junctions ($N \geq$ 3) in the 100-GHz band ($f_{\rm RF}$ = 80--116 GHz) in order to reduce the gain compression (at least $<$ 10\%) in the black body radiation at room temperature for chopper-wheel calibration. \item To expand the IF frequency range from 4.0--8.0 GHz to 4.0--12.0 GHz because the IF circuit of FOREST has a wide frequency range compared with the previous receivers in the 45-m telescope. An IF band width of 8 GHz is needed for the simultaneous observation of three carbon monoxide (CO) isotopic molecular lines, which are the most basic and important lines for studies of interstellar molecular gas properties. \item To develop DSB mixers with the lowest possible receiver noise temperature for the PCTJ circuit device ($\sim$20 K) for the 2SB mixers on FOREST. \end{enumerate} In Section 2, we describe the design of each section of the mixer, i.e., the waveguide probe, RF impedance matching circuit, and RF band filter for the IF signal. In Section 3, we present the results measured for the receiver noise temperature and the gain compression for the DSB mixers in the laboratory, as well as test observations using the newly developed mixer in FOREST on the 45-m telescope. \section{Design of the SIS Mixer} A SIS junction and RF matching circuit were designed considering a wide RF range (80--116 GHz; fractional bandwidth is 35\%), wide IF range (4--12 GHz), low noise performance ($T_{\rm rx}$(DSB) $\lesssim$20 K), and low gain compression ($<$10\% to 2SB mode with room temperature noise). These goals were determined based on the requirements for open-use of the 45-m telescope. The SIS mixer design comprised two independent planar circuits. One circuit was constructed with the probe antenna and RF filter to convert from the waveguide mode to the microstrip line (MSL) mode and to reject the RF band for the IF signal, respectively, on a 300 $\mu$m-thick quartz substrate. The other circuit was the impedance matching line between the feed point and the SIS junction on a 300 nm-thick SiO$_{2}$ thin film. The transmission lines in both circuits were fabricated with superconducting niobium (Nb) thin film. The structures of these planar circuits were designed using electromagnetic analysis with HFSS software\footnote{ANSYS HFSS $\langle$https://www.ansys.com/products/electronics/ansys-hfss$\rangle$} and an electrical circuit simulator Genesys\footnote{Keysight Genesys $\langle$https://www.keysight.com/en/pc-1297125/genesys-rf-and-microwave-design-softwares$\rangle$}. \subsection{Waveguide Probe} \begin{figure} \begin{center} \includegraphics[width=8cm]{fig1.eps} \end{center} \caption{(a) Pattern of the RF and IF circuit on the SIS mixer chip. The feed point is located in the center of the double-sided bow-tie type probe. (b) Feed point impedance of the waveguide probe on the Smith chart normalized by 70 $\Omega$ and calculated with the HFSS. The frequency range is from 80 to 118 GHz.}\label{fig:chip} \end{figure} The construction of the split-block waveguide unit for a mixer, which comprised RF and LO waveguides, a LO coupler, and a channel for the SIS mixer chip, was same as the design employed in a previous study (\cite{asa04}). The RF and LO signals passed through a tapered impedance transformer, which changed the standard W band waveguide (2.54 $\times$ 1.27 mm) into a 1/5 reduced-height waveguide (2.54 $\times$ 0.254 mm) in dimensions, and they were fed to the feed point on the SIS mixer chip. The SIS mixer chip (0.4 $\times$ 7 $\times$ 0.3$^{t}$ mm$^{3}$) was placed into a chip channel in the mixer block and one end of the planar circuit is connected to a 50 $ \Omega$ IF line by a 25 $\mu$m-diameter Al wire to extract the IF output. The channel slot on the other side was filled with indium, which held the SIS chip and connected to electrical ground. The waveguide probe was a double-sided bow-tie antenna, which was designed by \citet{asa03}, as shown in Figure~\ref{fig:chip}(a). The feed point was located in the center of the bow-tie antenna and it was treated as a lumped-gap-source port in HFSS. The feed point impedance of this waveguide probe is shown in Figure~\ref{fig:chip}(b) and it was designed as 70 $\Omega$. The calculated return loss from the input waveguide to the feed point on the SIS mixer chip was lower than 25 dB in the frequency range of 80--118 GHz. \subsection{Series-Connected Array of SIS Junctions} \begin{figure} \begin{center} \includegraphics[width=8cm]{fig2.eps} \end{center} \caption{(a) Design of the impedance transformer from the feed point to the SIS junction array on the bow-tie probe. (b) Model of the RF matching circuit. The circuit comprises four transmission lines. TR1 and TR3 are high-impedance coplanar waveguides, and TR2 and TR4 are low-impedance microstrip lines. The impedances as viewed from the SIS junction (A) and from the feed point (B) are shown in Figure 3.}\label{fig:circuit} \end{figure} \begin{figure*} \begin{center} \includegraphics[width=16cm]{fig3.eps} \end{center} \caption{(a) Embedding impedances as viewed from the SIS junction (at position A in Figure 2(b)) and (b) the impedance through the circuit from the feed point (at position B in Figure 2(b)). The Smith charts for (a) and (b) are normalized by $R_{\rm opt}$ = 25 $\Omega$ and the feed impedance 70 $\Omega$, respectively. Both of the analytical frequency ranges are 70--130 GHz. The dashed circle and the square in (a) represent the magnitude of the reflection coefficient $|\rho|$ $\leq$0.4 and the calculated region of dynamic resistance ($R_{\rm dyn}$), respectively. The gray shaded region at the top right indicates an unstable region where the mixer has negative dynamic resistance. The frequency characteristics of solid and dashed lines are the CLTL circuit in this work and the PCTJ circuit device in the previous work \citep{asa03}, respectively.}\label{fig:smith} \end{figure*} \begin{figure*} \begin{center} \includegraphics[width=16cm]{fig4.eps} \end{center} \caption{(a) Scanning electron microscope image of the tuning circuit for the mixer device. The boundary plane between the microstrip line (MSL) and coplanar waveguide (CPW) is the region within the dotted circle. (b) Structure of the capacitively loaded transmission line, which comprises the MSL and CPW where the center conductors and planes are the same, but the center conductor is different at each level.}\label{fig:cpwmsl} \end{figure*} According to the theoretical and experimental previous works (e.g. \cite{tuc85} and \cite{ker90}), the RF bandwidth ($\Delta\omega$) was determined from the $\omega R_{\rm n}C_{\rm j}$ product, where $R_{\rm n}$ is the junction normal state resistance and $C_{\rm j}$ is the junction specific capacitance. 1/$\omega R_{\rm n}C_{\rm j} \propto \Delta\omega/\omega_{0} \propto J_{\rm c}/\omega$, so this value was selected 2.8 at the center frequency ($\omega_{0}$) of 97.5 GHz, and thus the junction critical current density ($J_{\rm c}$) was 2.9 kA cm$^{-2}$. The sizes of a single junction and the target $R_{\rm n}$ were to be assumed 1.1 $\times$ 1.1 $\mu$m $^{2}$ and 62.7 $\Omega$, respectively. In astronomical observations with radio telescopes, the SIS mixers must exhibit good linear performance, i.e., a large dynamic range, between the input noise to the mixer and the output voltage from a detector attached to the receiver. During the calibration of the output signal intensity from the mixer receiver, such as a molecular line spectrum and continuum wave, a black body noise source at room temperature is typically used, which is well known as the chopper-wheel method \citep{uli76}. However, it is known that a single-junction SIS mixer readily saturates in response to a signal with very low ($\sim$pW) input power \citep{fel83}. Thus, a series SIS junction array is important for increasing the dynamic range of the SIS mixer. \citet{tuc85} reported that the saturation power level (P$_{\rm sat}$) can be calculated as follows: \begin{equation} P_{\rm sat} \propto \left(\frac{Nh\omega}{e}\right)^{2}\frac{L}{2R_{\rm L}}, \end{equation} where $N$ is the number of junctions in a series-connected array. Therefore, an array-junction mixer provides an increased dynamic range. We designed $N$ = 3, 4, 5, and 6 array junction on a coplanar waveguide (CPW) devices for testing under the assumption of an equivalent single junction \citep{tuc85}. Therefore, the size of each junction was multiplied by the size of a single junction and by $\sqrt[]{\mathstrut N}$. \citet{ton13} noted that it is important to consider the effect of the interconnecting lines between the individual junctions for mixers at a high frequency. The frequency band of our mixer was about 100 GHz and the lengths of the interconnecting lines were 8.5 $\mu$m of CPW, which are smaller than 1/200 of the wavelength ($\sim$1.9 mm), so we considered that the effects of these lines would have been small on the electrical circuit. Therefore, we treated part of the series array junctions as an equivalent single junction. \subsection{Impedance Matching Line} \begin{figure*} \begin{center} \includegraphics[width=16cm]{fig5.eps} \end{center} \caption{Simulated coupling efficiency (a) and expected receiver noise temperature (b) based on the designed impedance matching line between the feed point and the SIS junction. Solid and dashed lines represent the array junction device and the PCTJ circuit device, respectively. Dotted lines represent the changes in the noise temperatures due to the effective length extension of the coplanar waveguides for +3.4 $\mu$m (short circuit) and +6.8 $\mu$m (open circuit). The dashed-dotted line in (b) is the quantum limit.}\label{fig:couple} \end{figure*} We designed an end-loaded stub type matching circuit for the impedance transformer from the feed point to the SIS junction array. This circuit comprised two sections with high ($\sim$90 $\Omega$) and low ($\sim$10 $\Omega$) characteristic impedance transmission lines, and it was located on the bow-tie probe (Figure~\ref{fig:circuit}(a)). The high and low impedance lines were implemented using a CPW and a superconducting MSL, respectively. Therefore, this circuit can be viewed as a capacitively loaded transmission line (CLTL). This structure and its equivalent circuit are shown in Figure~\ref{fig:circuit}(a) and (b), respectively. As shown in this figure, the structures of SIS junction and TR1 are likely to be a capacitively loaded coplanar waveguide (CLCPW; \cite{ker96}). The impedances as viewed from the SIS junction (position A in Figure~\ref{fig:circuit}(b)) and from the feed point (position B in Figure~\ref{fig:circuit}(b)) are shown in Figure~\ref{fig:smith}. In Figure~\ref{fig:smith}(a), the center of impedance on the Smith chart is where the SIS junction had the lowest noise temperature, but we designed the circuit so the impedance would not enter the unstable region ($R_{\rm dyn} <$ 0 or =$\infty$), as shown by the negative resistance on the I-V curve in the simulation \citep{ino12}. No negative resistance is visible in the bottom left region ($R_{\rm dyn} >$ 0) and the mixer could operate in a stable manner. In addition, for the impedance located in the circle with a radius of 0.4, as represented by the dashed line in Figure~\ref{fig:smith}(a), the SIS mixer exhibited acceptable performance in terms of both the mixer noise temperature and gain \citep{ker95}. In Figure~\ref{fig:smith}(b), the center of impedance on the Smith chart is the feed impedance 70 $\Omega$. Since the CLTL of multi-stage tuning circuit is adopted, the distance from the center is almost similar value over the full RF range on the Smith chart. On the other hand, the PCTJ circuit device show a peaky resonace frequency because it has only one stage impedance transformer \citep{asa03}. This tuning line comprised two different structure lines, i.e., MSL and CPW, and they were connected inline. Therefore, the boundary plane between the MSL and CPW was a discontinuous structure (Figure~\ref{fig:cpwmsl}). \citet{bei93} suggested that an effective length extension could occur at the edge face of the planar circuit due to a parasitic component. They calculated the effective length extensions at the edge of the CPW as 3.4 $\mu$m for the short circuit and 6.8 $\mu$m for the open circuit. In order to estimate the extension in the electromagnetic analysis, we modeled a two-terminal-pair circuit, where we assumed no physical length and no loss between the MSL and CPW with HFSS. As a result, the Z-matrix for this circuit was inductive and the effective length was longer than the combined physical length of MSL and CPW. The extension was estimated as 4.8 $\mu$m at 100 GHz according to our calculations, where this value was between the lengths estimated for the short and open circuits by \citet{bei93}. Therefore, the lengths of the CPWs in the CLTL circuit used in our mixer device for testing were designed as 3.4 $\mu$m and 6.8 $\mu$m short on both sides of the CPW. However, the expected degradation of the receiver noise temperature due to this effect was approximately as only 1--2 K at both edges of the frequency range covered (Figure~\ref{fig:cpwmsl}(b) because these extension lengths were 4\% of the corresponding wavelength of the CPW at most. We simulated the performance of the SIS mixer with the CLTL circuit described above. The coupling efficiency of the RF/LO signal through the impedance matching line to the SIS junction is shown in Figure~\ref{fig:couple}(a), and the expected receiver noise temperature based on the calculations is shown in Figure~\ref{fig:couple}(b). When the coupling efficiency is 100\%, all of the input signal is fed to the SIS junction without reflection. We calculated the reflection coefficient ($S_{11}$) between the feed point and the SIS junction based on the calculation of electrical circuit using Genesys software, and estimated the coupling efficiency ($1-|S_{11}|^{2}$). The receiver noise temperature was simulated with the SIS mixer analyzer (SISMA; Shan \& Shi 2007, private communication), which is based on quantum mixing theory \citep{tuc79} with 3-port and 5-port approximations \citep{ton90}. In this calculation, we applied the embedding impedance of the RF circuit calculated with Genesys (Figure~\ref{fig:smith}(a)) and the noise temperature of the IF amplifier was assumed to be 10 K. This calculation excluded the loss of optics, such as the RF window, feed horn, and LO coupler, as well as the IF signal characteristic compared with the actual receiver system. Under these conditions, the expected receiver noise temperature was approximately 10 K, which is twice the quantum limit ($h\nu/k$). Moreover, we confirmed that the array junction receiver covered a wider frequency range than the PCTJ device, i.e., more than 92\% compared with the range of 75--125 GHz. \subsection{RF filter design for IF characteristic} \begin{figure} \begin{center} \includegraphics[width=8cm]{fig6.eps} \end{center} \caption{Simulated return and insertion losses for the choke filter versus frequency. The solid and dashed lines are the insertion and return losses, respectively. The frequency range is 0--120 GHz, which includes the IF and RF frequencies. The shaded region shows the actual IF and RF bandwidth of FOREST.}\label{fig:filter} \end{figure} The IF range for FOREST is 4--12 GHz (8 GHz bandwidth) because the molecular lines of $^{12}$CO ($J$ = 1--0; $f_{\rm rest}$ = 115.271 GHz), $^{13}$CO ($J$ = 1--0; $f_{\rm rest}$ = 110.201 GHz), and C$^{18}$O ($J$ = 1--0; $f_{\rm rest}$ = 109.782 GHz) can be obtained simultaneously in a single-sideband. The differences in the remaining rest frequencies between $^{12}$CO and C$^{18}$O is $\sim$5.5 GHz, and the IF bandwidth must be more than this difference. Therefore, it is important to design an RF rejection filter with low insertion loss in this IF range. The design of the RF filter is shown in Figure~\ref{fig:chip}(a). In order to extract the IF signal from the mixing device, a low-pass filter was needed to transmit the signal for the IF range and to block the RF range from the probe. We designed an RF choke filter based on a previous study by \citet{asa03}. This RF choke filter comprised high and low characteristic impedance 1/4$\lambda$ transmission lines (stepped-impedance low-pass filter), and the RF impedance at the input of the choke filter as viewed from the mixing device was larger by a factor of the squared impedance ratio of the pair. In addition, the impedance section was much shorter than the wavelength of the IF frequencies. A mixer chip with this IF circuit has been used in several receiver systems (e.g., S. Asayama et al. 2004$^{2}$, \cite{nak13}). The simulated insertion and return losses of this choke filter calculated with HFSS are shown in Figure~\ref{fig:filter}. The insertion loss of the filter at 12 GHz was approximately --0.8 dB. When we considered a cascade arrangement for the SIS junction, the IF output section including the RF choke filter, and a low noise amplifier, the increase in the receiver noise temperature at 12 GHz was 2 K at most relative to that at 4 GHz under the assumption of 20 K and 0 dB for the noise temperature and conversion gain of the SIS mixer, respectively, and 10 K for the noise temperature of the amplifier. This degradation value did not have a large impact relative to the noise temperature of this receiver system. However, the performance of the RF choke filter at a higher IF frequency ($\gtrsim$12 GHz) was more than -1 dB. Therefore, we designed and tested a new low-pass filter as a hammer type, as suggested by A. Navarrini \& B. Lazareff (2001)\footnote{ Navarrini, A., \& Lazareff, B. ALMA Memo 351, 2001\\ $\langle$http://library.nrao.edu/public/memos/alma/main/memo351.pdf$\rangle$.}, for the ALMA receiver in the 275--370 GHz band. We optimized the size of the hammer pattern to reject the RF frequency band (80--120 GHz) and determined a number for the pattern between the probe and IF output port. We successfully designed large rejection in the RF range ($\lesssim$-30 dB) and a small insertion loss in the IF range from 4 to 20 GHz ($\lesssim$0.5 dB) \citep{koz15}. The details of this filter design will be described in another paper. \section{Results} \subsection{Receiver Noise Temperature} \begin{figure} \begin{center} \includegraphics[width=8cm]{fig7.eps} \end{center} \caption{Noise temperatures measured for the newly developed array junction devices (4J and 6J) and existing PCTJ devices in the DSB mode corresponding to the LO frequency. The measurement system include the SIS mixer as well as the feed horn and IF components. PCTJ devices were measured in 10 mixers independently, and the open triangles and shaded region represent the average values and minimum-to-maximum range, respectively.}\label{fig:noise} \end{figure} We measured the noise temperatures for the newly developed array junction devices in the DSB mode using the standard Y-factor method for hot (room temperature) and cold (liquid nitrogen cooled) loads with a power meter. Figure~\ref{fig:noise} shows the results, where we compared the array junction devices (4J and 6J) with existing PCTJ devices. It should be noted that these noise temperatures included the contributions from the SIS mixer but also from the losses due to the feed horn, cooled isolator, and HEMT (high electron mobility transistor) amplifier in the range of 4--8 GHz IF. We successfully developed suitable array junction devices with $\sim$15--30 K for the receiver noise temperature $T_{\rm rx}$, which was almost the same as the performance of the PCTJ device. Moreover, we confirmed that the array junction devices covered a wider frequency range than PCTJ device. The trends in the frequency characteristics of both the array junctions and PCTJ devices were consistent with the expected features based on the circuit design (Figure~\ref{fig:couple}). However, $T_{\rm rx}$ at 12 GHz in the IF frequency range was typically $\sim$1.5 times higher than that at 4 GHz, and this degradation was large compared with the expected loss due to the RF filter as mentioned in 2.4. This increase in the noise temperature at the high IF range was probably due to the conversion gain characteristics of the SIS mixer as well as the loss characteristics of the coaxial cable and the cooled isolator before the HEMT amplifier in this measurement system, rather than the loss due to the RF filter in the mixer chip. It has been reported that the mixer noise temperature depends on the number of junctions $N$ and it generally increases with $\sqrt[]{\mathstrut N}$ \citep{tuc85}. However, our results for the 4J ($N$ = 4) and 6J ($N$ = 6) devices did not follow this trend as shown in Figure~\ref{fig:noise}. The results were probably affected by the device fabrication quality such as the implementation of $J_{\rm c}$ and the sizes of the junctions, rather than the effect of the number of junctions, because these devices were fabricated in different processes. Furthermore, \citet{cre87} reported that the measured mixer noise temperature increased as $N$ increased and decreased with the value of $\omega R_{\rm n}C_{\rm j}$. According to their results, the effect of $N$ on the increase in the noise temperature in our device ($\omega R_{\rm n}C_{\rm j}$ = 2.8) was probably small. The extension length effect in the planar circuit was not confirmed adequately in our devices. Our results indicated that this effect possibly had a small impact on the device in the millimeter wavelength range. According to our calculations, the extension lengths were about 4\% at most for the corresponding wavelength of the CPW. Moreover, the expected degradation of the receiver noise temperature due to this effect was approximately only 1--2 K at both edges of the frequency range covered (Figure 4(b)), and it could not be measured by our measurement system. \subsection{Gain compression} As mentioned in the introduction, the gain compression was more than 10\% with PCTJ devices based on the previous 2SB receiver in astronomical observations with the 45-m telescope. Based on the calculations reported by A. R. Kerr (2002)$^{4}$, 10--20 \% gain compression for the $N$ = 2 mixer under operation in the 2SB mode in the 100-GHz band corresponds to a mixer gain of 0--4 dB. This value is fairly consistent with the maximum mixer gain calculated for the PCTJ type receiver at around 110 GHz according to SISMA (Figure~\ref{fig:couple}). The gain compression for the array junction receiver was expected to improve and it was estimated at $\sim\frac{1}{N^{2}}$ times smaller than that with the PCTJ type receiver if they have the same conversion gain. We measured the maximum gain compression values for the $N$ = 3 and $N$ = 5 mixers at a load of 273 K in the laboratory and the values obtained were 21.2$\pm$1.2\% and 10.6$\pm$0.7\%, respectively. Therefore, the equivalent gain compression values in the 2SB mode were calculated as about 12\% for $N$ = 3 and 6\% for $N$ = 5 mixers. The value was larger than the target performance for $N$ = 3, but that for $N$ =5 was suitable to meet the requirements for open-use with FOREST on the 45-m telescope. The detailed measurements of the gain compression for our mixer, including the measurement system and method, will be presented in a separate study (T. Nakajima et al. in prep.). \subsection{First Light of Molecular Line on the Telescope} \begin{figure} \begin{center} \includegraphics[width=8cm]{fig8.eps} \end{center} \caption{First $^{12}$CO ($J$ = 1--0) molecular spectrum toward a carbon star obtained with the newly developed array junction mixer on the FOREST receiver.}\label{fig:12co} \end{figure} After evaluating the receiver noise temperature and its frequency characteristics with the array junction devices in the laboratory, we developed eight 2SB mixers for the FOREST receiver with four beams, where they could simultaneously obtain the dual polarization signals as well as the upper and lower sidebands in each polarization. Details of the receiver system for FOREST were described previously by \citet{nak12}, \citet{mur13}, and \citet{min16}. The first astronomical signal was obtained from the $^{12}$CO ($J$ = 1--0) spectrum in the USB toward the well-known carbon star IRC+10216 on May 19, 2011 with FOREST using the array junction mixer. Figure~\ref{fig:12co} shows the spectrum received on beam-1 at the center point of this object. Test observations of simultaneous $^{12}$CO ($J$ = 1--0), $^{13}$CO ($J$ = 1--0), and C$^{18}$O ($J$ = 1--0) spectra obtained with the four-beam 2SB mixer receiver toward the Orion KL molecular cloud were presented by \citet{min16}. At present, FOREST is using one of these receivers for open-use on the 45-m telescope and the system noise temperature including the atmosphere is typically 150--300 K. \section{Summary} In this study, we developed a series-connected array of SIS junction device in the 100-GHz band DSB mixer and it was installed in the multi-beam receiver FOREST on the NRO 45-m telescope. An array junction mixer ($N \geq$ 3) was designed in order to reduce the gain compression under hot load temperatures compared with the existing PCTJ type mixer. The multi-stage impedance transformer from the feed point to the junction array employed a CLTL, which comprised two sections with high ($\sim$90 $\Omega$) and low ($\sim$10 $\Omega$) characteristic impedance transmission lines. The structure of this tuning line is simple and easy to fabricate, and the RF impedance matched in a wide frequency range. The RF/LO coupling efficiency was more than 92\% and the expected receiver noise temperature was approximately two times the quantum limit, at least from 75 GHz to 125 GHz. We fabricated series-connected array devices where $N$ = 3--6 and measured their performance in the laboratory. As a result, we successfully developed suitable array junction devices with $\sim$15--30 K of $T_{\rm rx}$ (DSB) in the range of 80--120 GHz. The minimum value was fairly similar to the performance of the PCTJ device, but the RF frequency range was significantly wider than that for the existing device. We considered the effective length extension at the edge of planar circuit, but the impact of this effect on the noise performance was not clearly confirmed. Our results indicated that this effect possibly had a small impact on the device in the millimeter wavelength range, or it may have been too small to measure with our measurement system. The gain compression values were measured as 21.2$\pm$1.2\% and 10.6$\pm$0.7\% for the $N$ = 3 and $N$ = 5 mixers, respectively, at a load of 273 K in the DSB mode. Thus, the equivalent gain compression with the $N$ = 5 device using a 2SB mixer was expected to be approximately 6\%, which is sufficient to meet the requirements for open-use with FOREST on the 45-m telescope. Finally, the newly developed array junction mixer was installed in the FOREST receiver and we successfully obtained the $^{12}$CO ($J$ = 1--0) molecular line toward IRC+10216 with the NRO 45-m telescope. \begin{ack} The authors would like to thank Tetsuya Katase and Makoto Koyano at Osaka Prefecture University, and Makio Ito, Naoki Akiyama, Kouta Zengyo, and Masahiro Suzuki at Nagoya University for their contributions to measuring the performance of the SIS mixer. We also wish to thank Tomonori Tamura and Shohei Ezaki for the fabricating the SIS device at the Advanced Technology Center (ATC), National Astronomical Observatory of Japan, and Chihaya Kato for provide the SEM image of the SIS device. We are also grateful to Shin'ichiro Asayama, Jun Maekawa, Nario Kuno, Ryohei Kawabe, and the entire staff of the Nobeyama Radio Observatory for their useful discussions and support. \end{ack}

213efa3a08f017ea4ee76a18cf3b46c4f43e4df9

\section{\label{sec:intro}Introduction} The act of passing information between brain regions produces waves of neural activity. These waves are readily observed using non-invasive techniques such as electroencephalography (EEG) and magnetoencephalography (MEG) \citep{Alexander2015}, as well as in brain slices \citep{Golomb97}. Both experimental and theoretical work has shown that EEG/MEG recordings and evoked potentials can exhibit travelling and standing waves \cite{Nunez2006}. In particular travelling waves are seen in EEG sleep recordings propagating across the cortex at a speed of about 1.2-7.0 m/s \cite{Massimini2004}. Standing waves are often associated with idle brain states. For example, standing waves at $\alpha$ frequency (8-13 Hz) are observed in the vicinity of the visual cortex when the subject has their eyes closed \cite{Berger1929}. Another commonly observed spatial pattern is the so called \emph{bump attractor}. This spatially localised increase in population firing is produced in working memory tasks and the location of the bump can be linked to memory location \cite{Wimmer2014}. Traditionally, neural field models are used to describe wave and bump states in the brain. Although inspired by biology, these models are entirely phenomenological in nature. Even so, they have been particularly successful in describing neurophysiological phenomena, such as EEG/MEG rhythms \citep{Zhang1996}, working memory \citep{Laing2002}, binocular rivalry \citep{Bressloff2012} and orientation tuning in the visual cortex \citep{Ben-Yishai1995}. They are typically cast as a system of non-local differential equations which describe the spatiotemporal evolution of coarse grained population variables, such as the firing rate of a neuronal population, the average synaptic current, or the mean membrane potential \citep{Coombes2014}. The first attempt at a neural field mode is attributed to Beurle \cite{Beurle1956}, who built a model to describe the propagation of activation in a given volume of neural tissue. This model was purely excitatory, but even so allowed him to examine the propagation of large scale brain activity. In the 1970s Wilson and Cowan \cite{Wilson1972, Wilson1973} extended this model to include a second inhibitory layer. Unlike Beurle, they were interested in spatially localised bump solutions. In his seminal paper, Amari \cite{Amari1975,Amari1977} created what is now known as the standard neural field equation. By introducing a Mexican hat type coupling function (local excitation and long range inhibition), he reduced the model to a single equation with a mixture of excitatory and inhibitory connections. This allowed him to construct explicit solutions for spatially localised patterns, and assess their stability (at least for a Heaviside firing rate function). For a review of the Amari model and \emph{bumps} in one spatial dimension see \cite{Coombes2005} and for a discussion of \emph{bubbles} in two spatial dimensions see \cite{Bressloff2013}. One of the main assumptions in many types of neural field model, and especially those for describing EEG, is that point-wise they describe a density of neurons operating in a near synchronous regime \cite{Nunez2005}. This wholly reasonable assumption can be traced back to the observation that an EEG scalp electrode, which typically experiences the activity of roughly $10^9$ cortical pyramidal cells, can only detect an electric field if all the individual cell dipoles add coherently \cite{daSilva05}. However, a near synchrony assumption in any neural mass model precludes its use in describing the increase and decrease of power commonly seen in given EEG/MEG frequency bands. These \textit{temporal} variations are believed to be the result of changes in synchrony within the neural tissue. The former phenomenon is called event-related synchronisation (ERS), and the latter event-related desynchronisation (ERD) \cite{Pfurtscheller1999}. Consequently, there is a pressing need to develop the next generation of neural field models, which include this notion of a dynamic within-population synchrony (not fixed to be near synchronous), to more accurately describe the evolution of large scale spatio-temporal brain rhythms. When looking at within population synchronisation one typically uses a spiking neural network model. However, these high dimensional models are almost impossible to gain insight from. In an ideal world there would be a mathematical procedure for linking microscopic dynamics to macroscopic dynamics. This link has proved elusive for the majority of spiking models. However, Luke \emph{et al.} \citep{Luke2013} showed that the $\theta$-neuron model is amenable to such a reduction for pulsatile coupling. Montbri\'o \emph{et al.} \citep{Montbrio2015} have used a similar approach to reduce a network of quadratic integrate-and-fire neurons. Laing \cite{Laing2015} has also shown that the same approach can be applied to a network of spatially extended $\theta$-neurons, in the presence of gap junction coupling. In previous papers \citep{Byrne2017, Coombes2017}, we have shown that the approach of Luke \emph{et al.} can be extended to incorporate a biologically realistic form of synaptic coupling. Here we build on this work to construct a neural field model that incorporates within population synchrony and a realistic form of synaptic coupling. We shall refer to this model as a \emph{next generation neural field model}. In this paper we show, using a mixture of analysis and simulation, that this new neural field model can support exotic patterned states more reminiscent of high dimensional spiking networks, with spatio-temporal patterns showing the evolution of synchrony. We begin with an overview of the model formulation in \S\ref{sec:model} and outline the necessary steps for reduction to a neural field. A Turing instability analysis of the model is covered in \S\ref{sec:turing}. Here, we show that the system can be unstable to both static and dynamic Turing patterns for a wide window of parameter space. More interestingly, we show that when the Turing bifurcation collides with a Hopf bifurcation patterned states emerge in which there exists an oscillating structure within a spatially localised bump. The Turing analysis is complimented with a numerical bifurcation analysis for the full nonlinear model in \S\ref{sec:numeric}. Here we further examine the emergent patterns away from bifurcation, as well as consider localised travelling waves. Finally, in \S\ref{sec:discussion} we discuss our main results as well as natural extensions of the work presented. \section{The Model} \label{sec:model} \noindent We first consider a network of $N$ coupled quadratic integrate-and-fire (QIF) neurons, uniformly distributed along a line of length $\Lambda$ such that the $j$th neuron is at position $x_j = -\Lambda/2 +(j-1)\Delta x$, where $j=1,\ldots,N$ and $\Delta x = \Lambda/(N-!)$ is the spacing between neurons. The coupling between neuron $i$ and neuron $j$ depends only upon the distance between the two neurons, $w_{ij}^m=w_m(|x_i-x_j|)$, where $m$ is a label used to keep track of neural subpopulations. We shall focus on the choice $\Lambda \gg 1$ (so that we deal with large spatial scale systems). We write the network dynamics for the voltage, $v_i \equiv v(x_i,t)$, in the form \begin{equation} \dot{v}_i = v_i^2 +\eta_i + I_i , \qquad i=1,\ldots, N, \end{equation} subject to \textit{reset}, $v_i \rightarrow v_r$, whenever a firing threshold $v_{th}$ is reached by $v_i$. The time at which the $i$th cell reaches threshold from below for the $s$th time will be denoted by $T_i^s$, $s\in \NSet$. Here the background drives $\eta_i$ will be assumed to be heterogeneous and chosen from a Lorentzian distribution, \begin{align} L(\eta) = \frac{1}{\pi}\frac{\Delta}{(\eta-\eta_0)^2+\Delta^2}, \end{align} where $\eta_0$ is the centre of the distribution and $\Delta$ is the half width. We assume a synaptic input current of the form \begin{equation} I_i(t) = \sum_m g^i_m (t)(v_\text{syn}^m-v_i(t)), \end{equation} for a global conductance $g^i_m $, synaptic reversal potential $v_\text{syn}^m$ and local voltage $v_i$. In what follow we will assume $m=\{1,2\}$ or $m = 1$, but one should note that this framework can be extended to include multiple types of synapses. The interplay of excitation and inhibition, on different spatial scales, is known to play an important role in the generation of global spatially patterned states \citep{Ermentrout1979,Bressloff2001,Coombes2010}. Hence, we separate our synaptic conductance into two parts, such that we have one excitatory synaptic current ($v_\text{syn}>0$) and one inhibitory ($v_\text{syn}<0$), each with different spatial ranges. Each of the synaptic conductances will be taken to mimic that of a synapse with a finite rise and fall time that evolves according to \begin{equation} Q_m g^i_m = \frac{\kappa_m}{N} \sum_{s\in\NSet} \sum_{j=1}^N w_{ij}^m \delta(t-T_j^s), \end{equation} for some coupling strength $\kappa_m$, where $Q$ is the linear 2\textsuperscript{nd} order differential operator \begin{equation} Q_m = \left(1 + \tau_m\FD{}{t}\right)^2. \label{eq:Q} \end{equation} Here $\tau_m$ is the synaptic time scale, and $Q_m$ has a response (Green's function) $s(t)=\tau_m^{-2} t {\rm e}^{-t/\tau_m}$ for $t\geq 0$ (and is zero otherwise), which is a popular choice for many synapse models in computational neuroscience \citep{Gabbiani2010}. It is well known that the QIF model is formally equivalent to the $\theta$-neuron model \cite{Ermentrout1986} under the transformation $v_i = \tan(\theta_i/2)$, for $\theta_i \in [0, 2 \pi)$ (when the threshold $v_{th}$ and reset $v_r$ are set to $+\infty$ and $-\infty$, respectively). This relationship allows us to construct a $\theta$-neuron network dynamics as \begin{align} \dot{\theta_i} &= 1-\cos\theta_i + (1+\cos\theta_i)\eta_i \nonumber\\ &\phantom{.}\hspace{1.7em}+ \sum_m g^i_m [(1+\cos\theta_i)v_\text{syn}^m - \sin\theta_i], \label{eq:theta_discrete} \\ Q_m g^i_m &= \frac{2 \kappa_m}{N}\sum_{j=1}^N w_{ij}^m\delta(\theta_j-\pi). \label{eq:g_discrete} \end{align} To obtain \eqref{eq:g_discrete} we have made use of the fact that $\delta(t-T_j^s) = \delta(\theta_j- \pi) |\dot{\theta}_j(T_j^m)|$. As such, we say that neuron $j$ ``spikes" whenever $\theta_j$ increases through $\pi$. The network formulation in terms of dynamics on a circle is particularly useful since we no longer have to worry about handling the discontinuous reset process as we would have to do for a QIF network. \subsection{Mean field limit} We take the large $N$ limit, $N \rightarrow \infty$, which allows us to describe the system in terms of a continuous probability distribution function $\rho(x,\eta,\theta,t)$, with $x,\eta \in \RSet$, $\theta \in [0,2 \pi)$, and $t\in \RSet^+$, which satisfies the continuity equation: \begin{equation} \PD{}{t}\rho +\PD{}{\theta}(\rho v_{\theta}) = 0, \label{eq:continuity} \end{equation} where $v_{\theta}$ is the following realisation of \eqref{eq:theta_discrete}, \begin{align} v_{\theta}= 1-\cos\theta + (&1+\cos\theta)\eta \nonumber \\ &+ \sum_m g_m [(1+\cos\theta)v_\text{syn}^m - \sin\theta] . \label{eq:theta_continuous} \end{align} The mean field representation of the synaptic inputs $g_m$ are written as follows, \begin{align} Q_m g_m (x,t) = \frac{\kappa_m}{\pi}\sum_{l \in \ZSet}\int_{-\infty}^\infty\d y & \int_0^{2\pi}\d\theta \int_{-\infty}^\infty\d\eta \rho(y,\eta,\theta,t) \nonumber \\ &\times w_m(x-y){\rm e}^{il(\theta-\pi)}, \label{eq:g_continuous} \end{align} where we have used the result $2 \pi\delta(\theta-\pi) = \sum_{l \in \ZSet} {\rm e}^{il (\theta-\pi)}$ to write the right hand side in terms of exponentials. The formula for $v_{\theta}$ \eqref{eq:theta_continuous} may be conveniently written in terms of ${\rm e}^{\pm i\theta}$ as \begin{equation} v_{\theta} = \phi ({\rm e}^{i\theta} +{\rm e}^{-i\theta}) + \chi, \label{eq:c_exponential_form} \end{equation} where $\phi = (\eta+I-1)/2$ and $\chi = \eta+I+1$. Note that a similar approach has previously been considered by Laing \citep{Laing2015}. However, his focus was on smooth (non-pulsatile) interactions with a first order model of the synapse, and he did not consider reversal potentials.\\ To reduce the system we make use of the Ott-Antonsen (OA) ansatz \citep{Ott2008}. This decomposes $\rho$ in the form $\rho(x,\eta,\theta,t) = L(\eta) F(x,\eta,\theta,t)/(2 \pi)$, where $F$ is $2 \pi$-periodic in $\theta$ with a Fourier series representation $F(x,\eta,\theta,t) = \sum_n F_n(x,\eta,t) {\rm e}^{in \theta}$. The OA ansatz restricts the choice of $F_n$ such that $F_n(x,\eta,t) = \alpha(x,\eta,t)^n$, with $|\alpha(x,\eta,t)|<1$ to ensure convergence. Hence, $\rho$ can be written as \begin{equation} \rho(x,\eta,\theta,t) = \frac{L(\eta)}{2 \pi} \left\{ 1 + \left[\sum_{n=1}^\infty \alpha(x,\eta,t)^n {\rm e}^{i n \theta} + \text{cc} \right] \right\}, \label{eq:rho_OA} \end{equation} where $\text{cc}$ denotes the complex conjugate. Substituting \eqref{eq:c_exponential_form} into the continuity equation \eqref{eq:continuity} and balancing terms in ${\rm e}^{i\theta}$ gives an evolution equation for $\alpha$: \begin{equation} \PD{}{t} \alpha - i \alpha^{2} \phi - i\alpha \chi - i \phi = 0 . \label{eq:cont_a} \end{equation} We define the Kuramoto order parameter as follows, \begin{equation} z(x,t) = \int_0^{2 \pi} \d \theta \int_{-\infty}^\infty \d \eta \rho(x,\eta,\theta,t) {\rm e}^{i \theta}, \label{eq:kuramoto_order1} \end{equation} where $|z|\leq 1$. The Kuramoto order parameter is a complex number $z=R{\rm e}^{i\Psi}$, whose magnitude $R$ represents the degree of within population synchrony and angle $\Psi$ represents the average phase of the population. Substituting \eqref{eq:rho_OA} into \eqref{eq:kuramoto_order1} we find that \begin{equation} \bar{z}(x,t) = \int_{-\infty}^\infty \d \eta L(\eta) \alpha(x,\eta,t), \label{eq:kuramoto_orderr2} \end{equation} where $\bar{z}$ denotes the complex conjugate of $z$. As the Lorentzian has two simple poles $\eta_\pm=\eta_0\pm i\Delta$, the above integral may be performed by choosing a large semi-circular contour in the lower half $\eta$-plane and using the residue theorem, to yield $\bar{z}(x,t) = \alpha(x,\eta_-,t)$.\\ We use \eqref{eq:kuramoto_order1} to write \eqref{eq:g_continuous} as \begin{equation} \begin{aligned} Q_m g_m (x,t) &= \kappa_m\int_{-\infty}^{\infty} \d y w_m(x-y) f(z(y,t)),\\ & \equiv \kappa_m w_m\otimes f(z), \label{eq:evolution_g} \end{aligned} \end{equation} where $\otimes$ represents a spatial convolution and $f$ is the population firing rate \begin{equation} \begin{aligned} f(z) &= \frac{1}{\pi}\left\{1+\left[\sum_{l=1}^\infty (-1)^l z^l + \text{cc} \right]\right\} \nonumber\\ &= \frac{1}{\pi}\frac{1-\left| z\right|^2}{1+z+\overline{z}+\left|z\right|^2}, \hspace{1em} |z|<1. \label{eq:firing_rate} \end{aligned} \end{equation} Note that \eqref{eq:evolution_g} takes the form of a generalised neural field equation, where the firing rate function $f$ is a derived quantity that depends on the within population synchrony. Also noteworthy is the fact that this firing rate function is not a sigmoid. It is a highly non-linear function that depends on the intrinsic population dynamics. The firing rate can be plotted as a function of the Kuramoto order parameter (Fig.~\ref{fig:firing_rate}). As expected, the firing rate is highest at $z={\rm e}^{i\pi}$ (where a singe neuron fires as $\theta$ increases through $\pi$). \begin{figure \centering \includegraphics[width=0.9\linewidth]{firing_rate} \caption{\textbf{Firing rate dynamics:} Density plot showing the firing rate $f$ as a function of the complex Kuramoto order parameter $z=R{\rm e}^{i\Psi}$. Firing is highest near $z={\rm e}^{i\pi}$. This corresponds to highly synchronous behaviour where all of the phases of the neurons go through $\pi$ simultaneously. \label{fig:firing_rate}} \end{figure} The dynamics of $z$ are obtained by evaluating \eqref{eq:cont_a} at $\eta_- = \eta_0 - i\Delta$, and taking the complex conjugate, which gives $\partial z /\partial t = \mathcal{F}(z;\eta_0,\Delta)+\sum_m\mathcal{G}(z,g_m;v_{\text{syn}}^m)$, where \begin{align} \mathcal{F}(z;\eta_0,\Delta)&= -i\frac{(z-1)^2}{2}+\frac{(z+1)^2}{2}\left[i \eta_0 -\Delta \right], \label{eq:F_internal} \\ \mathcal{G}(z,g;v_{\text{syn}})&= g \left[i\frac{(z+1)^2}{2}v_{\text{syn}} - \frac{(z^2-1)}{2} \right] . \label{eq:G_external} \end{align} As in \cite{Coombes2017}, we interpret (\ref{eq:F_internal}) as describing the intrinsic population dynamics and (\ref{eq:G_external}) as the dynamics generated by synaptic coupling. In summary, we have the following evolution equations \begin{align} &\partial_t z = \mathcal{F}(z) + \sum_m\mathcal{G}_m(z,g_m) , \label{eq:ZEvol} \\ &(1+\tau_m\partial_t)^2 g_m = \kappa_m w_m \otimes f(z), \label{eq:GEvol} \end{align} where $\mathcal{G}_m(z,g_m)\equiv \mathcal{G}(z,g_m;v_{\text{syn}}^m)$ and we have omitted the dependence on control parameters. The dynamical system~\eqref{eq:ZEvol}--\eqref{eq:GEvol} has therefore one complex ($z$) and $|m|$ real ($g_m$) state variables. In what follows, we choose normalised exponentially-decaying synaptic kernels of the form $w_m(x) = \exp(-\beta_m |x|)/(2\beta_m)$, with Fourier transform $\widehat w_m(k) =1/(1+(k/\beta_m)^2)$. As the Fourier transform is a rational function, the evolution equations~\eqref{eq:ZEvol}--\eqref{eq:GEvol} are equivalent to the PDEs \citep{Coombes2003} \begin{align} &\partial_t z = \mathcal{F}(z) + \sum_m \mathcal{G}_m(z,g_m) , \label{eq:ZPDE} \\ &(1-\partial_{xx}/\beta_m^2) (1+\tau_m\partial_t)^2 g_m = \kappa_m f(z). \label{eq:GPDE} \end{align} In what follows, we will show that the model can support Turing patterns, travelling waves, and other complex global spatio-temporal patterns, for $m=\{1,2\}$ and travelling fronts for $m=1$. In the former case, this is achieved by choosing $\beta_1>\beta_2$ such that $w_1$ has a shorter spatial scale than $w_2$. For $v_\text{syn}^1>0$ and $v_\text{syn}^2<0$ this has the overall effect of short range excitation and long range inhibition, and for $v_\text{syn}^1<0$ and $v_\text{syn}^2>0$ inhibition dominates at short distances and excitation at longer distances. \section{ Instability of the homogeneous steady state} \label{sec:turing} We require excitation and inhibition to generate Turing patterns, hence we let $m=\{1,2\}$ in this section. We will also assume $\beta_1=1$ (without loss of generality) and define $\beta=\beta_2<1$ as the parameter which measures the difference in the spatial scales of $w_1$ and $w_2$. We first recast the system \eqref{eq:ZPDE}--\eqref{eq:GPDE} as a first-order evolution equation in the state variables $u = (a,b, K_1, g_1, K_2, g_2)$, where $a = {\rm Re}(z)$, $b = {\rm Im}(z)$ and $K_m = (1+\tau_m \partial /\partial t)g_m$, and seek stationary and spatially homogeneous states $u(x,t) = u^*$ for all $x$ and $t$. We apply a small perturbation of the form $\widetilde u(x,t) = u^* + \widetilde u(x,t)$, where $ \widetilde u(x,t) = A {\rm e}^{\lambda t} {\rm e}^{ikx}$, $\lambda \in \mathbb{C}$, $k\in \mathbb{R}$ and $A\in \mathbb{C}^6$. Using the following identity \[ w_m(x)\otimes{\rm e}^{ikx} = \widehat{w}_m(k){\rm e}^{ikx} , \] we obtain, to leading order, $ \partial_t \widetilde u(x,t) = \mathcal{J}(k)\widetilde u(x,t), $ where $\mathcal{J}$ is the following ($k$-dependent) Jacobian, evaluated at $u^*$, \begin{equation*} \mathcal{J}(k) = \left(\begin{matrix} \mathcal{J}_{11} & \mathcal{J}_{12} \\ \mathcal{J}_{21}(k) & \mathcal{J}_{22} \end{matrix}\right), \label{jacobian1} \end{equation*} where we highlight the $k$-dependence \begin{equation*} \mathcal{J}_{21}(k) = \left(\begin{matrix} \tau_1^{-1}\kappa_1 \widehat{w}_1(k)\partial_a f \hspace{0.5em} \phantom{.}& \tau_1^{-1}\kappa_1 \widehat{w}_1(k)\partial_b f \vspace{0.5em}\\ 0 & 0 \vspace{0.5em}\\ \tau_2^{-1}\kappa_2 \widehat{w}_2(k)\partial_a f \hspace{0.5em} \phantom{.}& \tau_2^{-1}\kappa_2 \widehat{w}_2(k)\partial_b f\vspace{0.5em}\\ 0 & 0 \end{matrix}\right). \label{jacobian2} \end{equation*} and show the other elements in Appendix \ref{app:jacobian}. The complex eigenvalues $\lambda = \nu + i\omega$ satisfies the characteristic equation \begin{equation} \mathcal{E}(\lambda,k) = \det{\left| \mathcal{J}(k)-\lambda I_4\right|} = 0, \label{eq:characteristic_formula} \end{equation} where $I_4$ is the $4\times4$ identity matrix. A homogeneous steady state $u^*$ is linearly stable to perturbations ${\rm e}^{ikx}$ if $\nu(k) < 0$ for all $k$. By the implicit function theorem, a branch of solutions $\lambda(k)$ to~\eqref{eq:characteristic_formula} touches the imaginary axis when \begin{equation} \partial_k \mathcal{M} \partial_\omega \mathcal{N} - \partial_\omega \mathcal{M} \partial_k \mathcal{N} = 0, \label{eq:nonDegenracy} \end{equation} where $\mathcal{M}={\rm Re}(\mathcal{E})$ and $\mathcal{N}={\rm Im}(\mathcal{E})$. We observe a Hopf bifurcation of the spatially uniform state $u^*$ if \eqref{eq:characteristic_formula} holds for $\nu(0)=0$ and $k=0$, i.e. if there is a non-zero solution of: \begin{align} \omega^6 + p_4\omega^4 + p_2 \omega^2 + p_0 &= 0, \label{eq:Hopf1}\\ p_5 \omega^4 + p_3 \omega^2 + p_1 &= 0. \label{eq:Hopf2} \end{align} Here, $p_i$ are scalars which depend on $u_*$ and the control parameters of the problem (see Appendix \ref{app:coeff}). Solving \eqref{eq:Hopf1}--\eqref{eq:Hopf2}, for fixed width of the Lorentzian $\Delta$, synaptic coupling strengths $\kappa_1$, $\kappa_2$, synaptic time constants $\tau_1$, $\tau_2$, and relative width of the synaptic kernel $\beta$, gives the locus of Hopf bifurcations as a function of the synaptic reversal potential $v_\text{syn}$ and background drive $\eta_0$ (Fig \ref{fig:turing_curves}(a) green curve). We fix the synaptic reversal potentials to be equal and opposite, and define $v_\text{syn} = v_\text{syn}^1 = -v_\text{syn}^2$. For simplicity we set $\kappa_1=\kappa_2$ and $\tau_1=\tau_2$. For this choice of parameters, excitation dominates for short range interactions for $v_\text{syn}>0$ and inhibition dominates for short range interactions for $v_\text{syn}<0$. Note, also, that the Hopf bifurcation occurs for the same value of $\eta_0$ for all $v_\text{syn}$. This is a consequence of the choice of equal coupling strengths and time constants. If this balance is disrupted the Hopf bifurcation will depend upon $v_\text{syn}$ and for some parameter choices we see several Hopf curves in the $(v_{\text{syn}},\eta_0)$-plane. The homogeneous steady state $u^*$ undergoes a static Turing bifurcation if there exists a non-zero critical wavenumber $k_c$ such that \eqref{eq:characteristic_formula} and \eqref{eq:nonDegenracy} hold for $\nu(k_c) =0$, $\omega(k_c)=0$. This leads to the conditions \begin{align} &p_0 + q^1_0 \widehat w_1 + q^2_0 \widehat w_1 = 0, \label{eq:Turing1}\\ &\left[q_0^1\FD{\widehat w_1}{k_c} +q_0^2\FD{\widehat w_2}{k_c}\right]\left(p_1+q^1_1 \widehat w_1 + q^2_1 \widehat w_2 \right)= 0, \label{eq:Turing2} \end{align} where $\widehat{w}_i$ and its derivative depend on $k_c$ and $\beta$, and the scalars $p_i$, $q_i^j$ depend on $u_*$ and the control parameters of the problem (see Appendix \ref{app:coeff}). As with the Hopf bifurcation, the locus of Turing bifurcations can be plotted as a function of the synaptic reversal potential $v_\text{syn}$ and background drive $\eta_0$ (Fig.~\ref{fig:turing_curves} red curve). The curve exhibits a turning point, therefore the system has two Turing bifurcations for sufficiently large values of $v_\text{syn}$. If we fix $v_\text{syn}$ at a value in this region, and ascend the bifurcation diagram by increasing $\eta_0$, the spectrum $\lambda(k)$ touches the imaginary axis of the $(\nu,\omega)$-plane with wave number $k_c$ (lower Turing bifurcation) (Fig. \ref{fig:turing_curves}(c)) and continues to move to the right, implying that $u^*$ is unstable to a range of perturbations with wavenumbers $k \in (k_1,k_2)$. As $\eta_0$ is further increased, the spectrum returns to the left-hand plane (upper Turing bifurcation), which restores the stability of $u^*$. Note the value of $k_c$ is not necessarily the same on the upper and lower branches. The scenario described above is robust to perturbations in other parameters: we found that changes in $\tau_1$, $\tau_1$ and $\Delta$ do not significantly affect the location of the Turing bifurcations. Increasing $\kappa_2$ results in an increase of the gap between the two critical values of $\eta_0$, and hence a larger window of instability, while increasing $\kappa_1$ decreases this gap and at a certain point the shape is inverted, such that the unstable region lies predominantly in the $\eta_0<0$ region of the plot. As $\kappa_1$ corresponds to the synaptic strength for the excitatory current for $v_\text{syn}>0$, when it becomes significantly larger than $\kappa_2$ excitation dominants even for long range interactions. As patterns arise from the interplay of excitation and inhibition, an inhibitory external drive ($\eta_0<0$) is needed to observe Turing patterns in this regime. \begin{figure*} \centering \includegraphics{two-par-bif-diag-compound} \caption{Instabilities of the homogeneous steady state. (a): Two-parameter bifurcation diagram of Hopf and Turing instabilities of the homogeneous steady state in the $(v_{\text{syn}},\eta_0)$-plane, where $v_{\text{syn}}=v_{\text{syn}}^1=-v_{\text{syn}}^2$. The Hopf (green), Turing (red) and Turing-Hopf (blue) curves partition the plane into five sectors. The Hopf and Turing curves cross at a codimension-2 point where both instabilities occur simultaneously. The homogeneous steady state $u^*$ is stable in I. (b): $u^*$ is unstable to global periodic oscillations in II, and direct numerical simulations close to the instability show a stationary Turing pattern. We plot $R(x,t) = |z(x,t)|$ obtained via direct numerical simulation of~\eqref{eq:ZPDE}--\eqref{eq:GPDE}, for $m=\{1,2\}$ with exponentially-decaying kernels (top panel). The eigenvalues of the spatially clamped system at the bifurcation point are shown in the bottom panel. (c): $u^*$ is Turing unstable in III, hence, we observe stationary periodic patterns. The spectrum of the linearised operator around $u^*$ at bifurcation is also shown (bottom panel). (d): $u^*$ is Turing-Hopf unstable in IV and the instability (bottom panel), gives rise to a periodic wavetrain, as expected. (e): spatio-temporal pattern obtained in V, where $u^*$ is Turing and Turing-Hopf unstable. The spectrum in the bottom panel is at the codimension-2 point, marked in (a) with a red circle. Parameters: $\Delta=0.5$, $\kappa_1=\kappa_2=5$, $\tau_1=\tau_2=0.2$. } \label{fig:turing_curves} \end{figure*} A Turing-Hopf instability of the homogeneous steady state occurs if there exists a non-zero wave number $k_c$ such that \eqref{eq:characteristic_formula} and \eqref{eq:nonDegenracy} hold with $\nu(k_c)=0$ and $\omega(k_c) = \pm \omega_c \neq 0$. This instability, which we will also refer to as \textit{dynamic} Turing bifurcation, elicits wavetrains with wavenumber $k_c$ and phase velocity $\omega_c/k_c$ (near bifurcation). From \eqref{eq:characteristic_formula} and \eqref{eq:nonDegenracy} we obtain a system of the form \begin{align} & \omega^6 +p_4 \omega^4 + P_2\omega^2 + P_0 = 0, \label{eq:dynTuring1}\\ & p_5\omega^4 + P_3 \omega^2 + P_1 = 0, \label{eq:dynTuring2}\\ & \left[ Q_2^1\FD{\widehat w_1}{k} +Q_2^2\FD{\widehat w_2}{k}\right] (-6\omega^5 +4 p_4 \omega^3 - 2 P_2\omega) \nonumber \\ &\phantom{.}\hspace{0.5em}-\left[ Q_3^1\FD{\widehat w_1}{k} +Q_3^2\FD{\widehat w_2}{k}\right](5p_5\omega^4 -3 P_3 \omega^2 + P_1) = 0, \label{eq:dynTuring3} \end{align} where $P_i = p_i +q^1_i \widehat w_1(k) + q^2_i \widehat w_1(k) $, $Q_i^j =q_i^j\omega^i +q_{i-2}^j\omega^{i-2}$ and $p_i$, $q_i^j$ are scalars which depend on the control parameters (see Appendix \ref{app:coeff}). In passing, we note that the characteristic equation now has an imaginary part, resulting in the two conditions~\eqref{eq:dynTuring1}, \eqref{eq:dynTuring2}. Solving \eqref{eq:dynTuring1}--\eqref{eq:dynTuring3} allows us to plot the locus of the Turing-Hopf bifurcations in the $(v_{\text{syn}},\eta_0)$-plane (Fig.~\ref{fig:turing_curves} blue curve), together with the Hopf and Turing bifurcation curves. Note that the Turing-Hopf curve intersects with the Hopf curve at $v_{\text{syn}}\approx 8$. The value of $k_c$ deceases along the Turing-Hopf curve as $v_{\text{syn}}$ is increased, such that at the point where the two curves collide $k_c=0$ on the Turing-Hopf curve. \begin{figure*} \centering \includegraphics{further-spatiotemporal-patterns} \caption{Further spatio-temporal patterns supported by the neural field model. (a): Pattern found in sector V of Fig.~\ref{fig:turing_curves}(a), for $\eta_0=60$, $v_\text{syn}^1=-v_\text{syn}^2=10$. The pattern displays a time-periodic structure within each bump of a Turing pattern. (b): \emph{Wandering} bump observed in region IV of Fig.~\ref{fig:turing_curves}, for $\eta_0=2.5$, $v_\text{syn}^1=-v_\text{syn}^2=-20$. Similar patterns are also found in sector II (not shown). (c): Weakly unstable standing wave obtained for $\eta_0=15$, $v_\text{syn}=-10$. After undergoing a breathing instability for a long transient, the standing wave evolves towards a stable wavetrain (not shown). Other parameter values as in Fig \ref{fig:turing_curves}.} \label{fig:patten_zoo} \end{figure*} The curves in Fig.~\ref{fig:turing_curves}(a) were computed using {\sc XPPAUT} \cite{Ermentrout02}, by continuing a suitable algebraic problem in one of the parameters of the system ($v_\text{syn}$). The equations defining an equilibrium are solved simultaneously with \eqref{eq:Hopf1}--\eqref{eq:Hopf2}, \eqref{eq:Turing1}--\eqref{eq:Turing2} or \eqref{eq:dynTuring1}--\eqref{eq:dynTuring3} and continued in parameter space. The curves partition the $(v_{\text{syn}},\eta_0)$-space into five sectors (labelled I--V). In addition to sectors where $u^*$ is stable (I), $u^*$ is unstable to bulk oscillations (II), Turing instabilities (III), and Turing-Hopf instabilities (IV), a fifth sector (V) is generated by the crossing of the Turing and Hopf curves. At the intersection (codimension-2) point, the spectrum of the linearised operator around $u^*$ has one zero eigenvalue and two complex conjugate eigenvalues, where the critical wavenumber $k_c$ is non zero for the zero eigenvalue and equal to zero for the complex conjugate pair (Fig \ref{fig:turing_curves}(e)). Direct numerical simulations close to the instability confirm the predictions of the linear stability analysis: bulk oscillations are observed in in section I (Fig \ref{fig:turing_curves}(b)) stationary Turing patterns are observed in sector III (Fig \ref{fig:turing_curves}(c)) and both wavetrains and standing waves are seen in region IV (Fig \ref{fig:turing_curves}(d)). In sector V, the simultaneous Turing and Hopf unstable modes compete, resulting in a characteristic complex spatio-temporal pattern, where temporal oscillations develop within each bump of a Turing pattern (Fig ~\ref{fig:turing_curves}(e)). The existence of the intersection point of the Turing and Hopf curves and the observation of the exotic spatio-temporal patterns are robust to changes in parameters. In passing we note that, close to onset, we could find only large amplitude patterns of the type shown in Figs~\ref{fig:turing_curves}(c)-(e), indicating that the Turing and Turing-Hopf bifurcations are subcritical (as confirmed below by numerical continuation). Interestingly, we also observe dynamic Turing patterns in region II, implying that the Hopf bifurcation does not stabilize the system, instead it creates bistability in this region, where the system supports both bulk oscillations and dynamic global patterns. We also found other complex spatio-temporal patterns away from bifurcation onset, using direct numerical simulation (Fig.~\ref{fig:patten_zoo}). The system supports time-periodic patterns containing structures within bumps (Fig.~\ref{fig:patten_zoo}(a)). This pattern is observed as we move away from the Turing bifurcation but stay close to the Hopf curve. Structures of this type, in which the bumps of a Turing pattern are periodically modulated in time, were observed initially in sector IV of Fig. \ref{fig:turing_curves}(a). Spatio-temporal patterns of this form, with modulation at the core, were found only in parameter sets where the Turing and Hopf codimension-2 point is present, as expected. Structures within bump solutions are not seen in standard neural field models. They are however commonly observed in spiking neuron models \citep{Laing01a, Chow2006}, which emphasises that this next generation neural mass model retains information about the underlying spiking model. In region IV, we also observe an number of interesting patterns, such as the \emph{wandering} bump (Fig. \ref{fig:patten_zoo}(b)) and a form of standing wave, where both the width and the height of the bumps is periodically modulated (Fig. \ref{fig:patten_zoo}(c)). Numerical simulations indicate that the wandering bump is stable for a large area of parameter space and can coexist with a periodic wavetrain. The standing wave, however, persists for a long time but ultimately is unstable and transitions to a periodic travelling wave (not shown). \section{Numerical bifurcation analysis}\label{sec:numeric} In order to study patterns away from bifurcations, we employ numerical bifurcation techniques, which allows us to compute coherent structures, determine their stability, and track their dependence on control parameters. Here, we employ the numerical tool kit developed by Avitabile \cite{Avitabile2016} and employed in the context of standard neural fields in \cite{Rankin:2014bz,Avitabile:2014wg}. This tool kit can be used to compute waves and patterns, and their stability. We refer the reader to recent reviews on numerical bifurcation analysis for coherent structures \cite{Laing2014, Champneys2007}. \begin{figure} \centering \includegraphics{turing-continuation} \caption{Continuation of spatially periodic patterns. A bifurcation diagram in $\eta_0$ of two stationary patterned states, with different wavenumbers. Each branch of patterned states are connected to the branch of homogeneous states via a subcritical Turing bifurcations and reconnect to the steady state within the region of instability. Only the rightmost and left most Turing points correspond to a change in stability of the system. The homogeneous steady state undergoes a Hopf bifurcation at $\eta_0= 3.298$ as expected from the instability analysis in \S\ref{sec:turing}. More interestingly, we also see Hopf bifurcations along the branch of periodic solutions. Solid (dashed) lines represent stable (unstable) solutions. Parameter values: $v_\text{syn}=15$, $\Lambda=12\pi$, other parameters as in Fig.~\ref{fig:turing_curves}(a).} \label{fig:bump_continuaton1} \end{figure} \begin{figure*} \centering \includegraphics{wavetrain-continuation} \caption{ Numerical bifurcation analysis of wavetrains. (a): Bifurcation diagram of wavetrains in the parameter $\eta_0$, for various values of the domain size, where $k_c=0.739$. Here we show only the maximum of the synaptic conductance $g_1$ as our solution measure. The green solution bifurcates off the steady state solution (grey) at a Turing-Hopf bifurcation, and the other solutions go unstable at slightly larger values of $\eta_0$. (b): the branches in (a) are plotted using the phase velocity $c$ as solution measure. The diagrams in (a) and (b) confirm that the Turing-Hopf bifurcation is subcritical. At the critical points, the wavetrain emerges with a non-zero phase velocity ($\omega_c/k_c$). (c): Branches of wavetrains in the continuation parameter $\Lambda$, for $\eta_0=0$. This curve is effectively a dispersion relation, showing the wave speed as function of the spatial period. In this region of parameter space the branch is an isola, therefore the wavetrain exists only for a finite range of periods. Parameter values: $v_{\text{syn}}=-30$, other parameters as in Fig. \ref{fig:turing_curves}.} \label{fig:periodic_waves_domain_cont} \end{figure*} First, we rescaled space such that $x\in [-0.5, 0.5]$, to highlight the dependence on the scale of the domain size $\Lambda$. The equivalent PDE formulation~\eqref{eq:ZPDE}--\eqref{eq:GPDE} becomes \begin{equation} \label{eq:WH-PDE} \begin{aligned} \partial_t z & = \big( \mathcal{F}(z) + \mathcal{G}(z,I) \big) ,\\ (\Lambda^2 - \partial_{xx})(1+\tau_1\partial_t)^2 g_1 & = \Lambda^2\kappa_1 f(z) , \\ (\Lambda^2 - \partial_{xx}/\beta^2)(1+\tau_2\partial_t)^2 g_2 & = \Lambda^2\kappa_2 f(z). \end{aligned} \end{equation} We perform numerical bifurcation analysis for heterogeneous spatially patterned steady states and periodic travelling waves of the system above. We construct the stationary patterns by solving for $(z,g_1,g_2)$ the boundary value problem \begin{equation} \label{eq:WH-PDE-SS} \begin{aligned} \mathcal{F}(z) + \mathcal{G}_1(z,g_1) + + \mathcal{G}_2(z,g_2) & = 0, \\ (\Lambda^2 - \partial_{xx}) g_1 - \Lambda^2\kappa_1 f(z) & = 0, \\ (\Lambda^2 - \partial_{xx}/\beta^2) g_2 - \Lambda^2\kappa_2 f(z) & = 0, \end{aligned} \end{equation} with Neumann boundary conditions. Wavetrains are solutions to~\eqref{eq:WH-PDE} with $z(x,t) = Z(x-ct)$, $g_1(x,t) = G_1(x-ct)$ and $g_2(x,t) = G_2(x-ct)$ for $c \in \mathbb{R}$, where $Z(\xi)$, $G_1(\xi)$ and $G_2(\xi)$ are $\Lambda$-periodic. To compute wavetrains, we solve for $(Z,G_1,G_2,c)$ the boundary value problem \begin{equation} \label{eq:WH-PDE-TW} \begin{aligned} c \partial_\xi z + \Lambda \big( \mathcal{F}(Z) + \mathcal{G}_1(Z,G_1)+ \mathcal{G}_2(Z,G_2) \big) &= 0, \\ (\Lambda^2 - \partial_{\xi\xi})(1-c\tau\partial_\xi)^2 G_1 - \Lambda^2\kappa f(Z) &= 0, \\ (\Lambda^2 - \partial_{\xi\xi}/\beta^2)(1-c\tau\partial_\xi)^2 G_2 - \Lambda^2\kappa f(Z) &= 0, \\ \psi(Z,G_1,G_2) &= 0, \end{aligned} \end{equation} posed on $\xi \in [-1/2,1/2]$ with periodic boundary conditions. The last equation in~\eqref{eq:WH-PDE-SS} and \eqref{eq:WH-PDE-TW} is a standard phase condition \cite{doedel1981auto} \begin{equation} \label{eq:phase_cond} \begin{aligned} \psi(Z,G_1,G_2) & = \int_{-1/2}^{1/2} \d\xi \; \frac{\d}{\d \xi} \widetilde Z(\xi) \big( Z(\xi) - \widetilde Z(\xi) \big)\\ & + \int_{-1/2}^{1/2} \d\xi \; \frac{\d}{\d \xi} \widetilde G_1(\xi) \big( G(\xi) - \widetilde G_1(\xi) \big) \\ & + \int_{-1/2}^{1/2} \d\xi \; \frac{\d}{\d \xi} \widetilde G_2(\xi) \big( G(\xi) - \widetilde G_2(\xi) \big), \end{aligned} \end{equation} where $(\widetilde Z, \widetilde G_1, \widetilde G_2)$ is a reference template solution, such as one of the solutions obtained via direct simulation. We discretised the differential operators \eqref{eq:WH-PDE-TW} using standard differentiation matrices, which are also used to compute linear stability of the coherent structures~\cite{Avitabile2016}. \subsection{Turing patterns}\label{sub:bumps} We first analyse the stationary patterns seen in \S\ref{sec:turing}, using numerical continuation to verify the analytical results, determine the criticality of the Turing bifurcation, and examine the behaviour of these solutions away from the onset of the instability. We continued solutions to the boundary value problem~\eqref{eq:WH-PDE-SS} in the parameter $\eta_0$ with $\Lambda=12\pi$, $v_{\text{syn}}=15$ and all other parameters as in Fig.~\ref{fig:turing_curves}(a). This corresponds to making a vertical excursion through the 2-parameter bifurcation diagram Fig.~\ref{fig:turing_curves}(a). Solving \eqref{eq:Turing1}--\eqref{eq:Turing2} we found two Turing bifurcations at $\eta_0=-0.648$, $k_c=0.738$ and $\eta_0=12.67$, $k_c=0.969$. Hence, we numerically continued patterns states with $k=0.738$ (blue) and $k=0.969$ (red) (Fig. \ref{fig:bump_continuaton1}). As expected, the homogeneous steady state bifurcates to patterns at $\eta_0=-0.648$ (blue) and $\eta_0=12.67$ (red), corresponding to the Turing bifurcations found in \S\ref{sec:turing}. The stability of the patterned states, as well as the homogeneous state, was numerically calculated along the continuation branch. Along with the Turing bifurcations, the homogeneous steady state undergoes a Hopf bifurcation at $\eta_0=3.298$, which matches the value found analytically. The patterned Turing solutions were found to go unstable to a globally oscillating periodic pattern, through a Hopf bifurcation at $\eta_0=2.8380$ (blue) and $\eta_0=5.8546$ (red). The patterned solutions are also unstable (through a symmetry breaking bifurcation) to broader patterns at low values of $\eta_0$ As anticipated, the bifurcations from the homogeneous state steady state are subcritical, hence we have bistability between a spatial pattern and the homogeneous state and the bistability region occurs in a wide region of parameter space. Further numerical continuation results (not shown) indicate that the region of bi-stability increases as the reversal potential $v_\text{syn}$ is increased. As $v_\text{syn}$ decreases towards $0$, the static Turing bifurcation points collide, and the patterned states cease to exist, as predicted from the Turing analysis in \S\ref{sec:turing}. We found the scenario presented above to be robust to changes in the other parameters. \subsection{Wavetrains}\label{subsub:moiving_bumps} We now shift our focus to wavetrain solutions originating at a Turing-Hopf bifurcation of the homogeneous steady state. We recall that we find these states as solutions to~\eqref{eq:WH-PDE-TW} with periodic boundary conditions, hence $\Lambda$ corresponds to the spatial period of the wavetrain profile. In addition, the phase velocity $c$ of a wavetrain is accessible from the boundary-value problem solution. As in \S\ref{sub:bumps}, the spatial frequency $k_c$ at bifurcation was found by solving \eqref{eq:dynTuring1}--\eqref{eq:dynTuring3}, and we continued patterns with this wave number by setting $\Lambda=2\pi/k_c$, where $k_c=0.739$ (Fig. \ref{fig:periodic_waves_domain_cont}(a)--(b) green curve). We also continued spatial patterns with $k=2 k_c$ (red curve) and $k=k_c/2$ (blue curve), this was achieved by posing the system on domains $\Lambda=\pi/k_c$ and $\Lambda=4\pi/k_c$, respectively. The wavetrains bifurcate from the homogeneous steady state with a non-zero phase velocity, at different values of $\eta_0$ (Fig.~\ref{fig:periodic_waves_domain_cont}(b)). The value for which periodic waves emerge for $\Lambda=2\pi/k_c$ corresponds to the Turing-Hopf bifurcation (green dot) found in \S\ref{sec:turing}, and the speed is equal to $\omega_c/k_c$. The wavetrains are unstable for low values of $\eta_0$. For the smaller domain size they are unstable to the stable steady state, whereas for the larger domain size they transition to finer patterns, with a smaller spatial wavelength $k$. To fully explore the relationship between the wave speed and the spatial period, we fix the value of $\eta_0=0$ and use $\Lambda$ as a continuation parameter. This opens up the possibility to trace branches of solutions in the $(c,\Lambda)$-plane, and hence, approximate the dispersion curve of the waves (Fig. \ref{fig:periodic_waves_domain_cont}(c)). We found that wavetrains occur in isolas, and therefore only exist for a finite range of spatial periods. For $5.2<\Lambda<13.1$, a fast (stable) wavetrain coexist with slower (unstable) one. Whereas, when $ 3.2<\Lambda<5.2 $ and $13.1<\Lambda<28.0 $ the two waves are unstable. For all other values of $\Lambda$ we do not see wavetrain solutions. If $\eta_0$ is increased, the dispersion curve is no longer an isola, but rather a monotonically increasing function. Thus, in this regime, the system supports periodic travelling waves for all values of the spatial period, above a threshold value. \subsection{Fronts}\label{sub:waves} Standard neural field models are known to support travelling fronts (for exponentially decaying kernels), which are travelling waves whose profile connects a uniform high-activity state to another low-activity state \cite{Ermentrout1998, Bressloff2012a}. It is therefore natural to search for these coherent structures in our new neural field model~\eqref{eq:ZEvol}--\eqref{eq:GEvol}. As in any other nonlocal neural field, the existence of the high- and low-activity states depends on the choice of the synaptic kernel. As inhibition is not necessary for the generation of localised patterns, such as travelling fronts, in this section we consider only a single synaptic conductance, $m=1$ and suppress the $m$ label, leading to the equivalent PDE formulation \begin{equation} \label{eq:EK-PDE} \begin{aligned} \partial_t z & = \mathcal{F}(z) + \mathcal{G}(z,g), \\ (1 - \partial_x^2)(1+\tau\partial_t)^2 g & = \kappa f(z) , \end{aligned} \end{equation} which we pose on $\xi \in \mathbb{R}$. We now set $z(x,t) = Z(x-ct)$, $g(x,t) = G(x-ct)$, and we seek travelling fronts as bounded solutions $U(\xi) = (U_1(\xi), \ldots, U_6(\xi))$ to the boundary-value problem \begin{equation}\label{eq:spatialDynamic} \partial_\xi U = \mathcal{N}(U), \quad \xi \in \mathbb{R}, \qquad \lim_{\xi \to \pm \infty} U(\xi) = U^{\mp}, \end{equation} where $\mathcal{N} \colon \mathbb{R}^6 \to \mathbb{R}^6$ is the real-valued nonlinear function \[ \mathcal{N}(U)= \begin{pmatrix} -\real [ \mathcal{F}( U_1 + i U_2) + \mathcal{G}( U_1 + i U_2, U_3) ]/c \\ -\imag [ \mathcal{F}( U_1 + i U_2) + \mathcal{G}( U_1 + i U_2, U_3) ]/c \\ (U_3 - U_4)/(\tau c) \\ U_5 \\ U_6 \\ U_5 + \big[ U_6 + \kappa f(U_1 + i U_2) - U_4\big]/(\tau c) \end{pmatrix}. \] In this spatial-dynamical system formulation of the problem, the first three components of $U$ have a direct interpretation in terms of the state variables $(z,g)$ of \eqref{eq:EK-PDE}, \[ U_1 = A \equiv \real Z, \qquad U_2 = B \equiv \imag Z, \qquad U_3 = G, \] whereas $U_4$, $U_5$ and $U_6$ are auxiliary variables, necessary to cast the problem as a system of first-order differential equations in $\xi$. The equilibria $ U^\pm = (U_1^\pm,U_2^\pm,U_3^\pm,U_3^\pm,0,0) $ of the spatial-dynamical system~\eqref{eq:spatialDynamic} correspond to high- and low-activity homogeneous steady states of~\eqref{eq:EK-PDE} and are completely determined by solving for $(U_1,U_2,U_3)$ the algebraic problem \begin{equation}\label{eq:clampedEq} \begin{aligned} \real [ \mathcal{F}( U_1 + i U_2) + \mathcal{G}( U_1 + i U_2, U_3) ] & = 0,\\ \imag [ \mathcal{F}( U_1 + i U_2) + \mathcal{G}( U_1 + i U_2, U_3) ] & = 0,\\ \kappa f(U_1 + i U_2) - U_3 & = 0. \end{aligned} \end{equation} We define $U^+$ as the high activity state, which displays high synaptic conductance and $U^-$ as the low activity state, displaying lower synaptic conductance. Note that we also seek the symmetric counterpart solution where $\lim_{\xi \to \pm \infty} U(\xi) = U^{\pm}$. We have continued solutions to~\eqref{eq:clampedEq} in $\eta_0$ and $v_\text{syn}$ using XPPAUT~\cite{Ermentrout02} (Fig. \ref{fig:front_continuation}(a)). The system has three fixed points in the region enclosed by the saddle-node curves, two of which are stable ($U^\pm$). Therefore, we look for travelling waves in this shaded region of parameter space. We note that, as $\tau$ is decreased, a Hopf bifurcation of the $U^-$ state arises, opening up the possibility of creating heteroclinic connections to periodic orbits, which we will not consider further here. \begin{figure*} \centering \includegraphics{front-continuation-compound} \caption{Numerical continuation results for travelling fronts. (a): two-parameter bifurcation diagram of stationary homogeneous states of~\eqref{eq:EK-PDE} (equilibria of~\eqref{eq:spatialDynamic}) in the $(\eta_0,v_{\text{syn}})$-plane. We plot the locus of saddle-node bifurcations of the steady state, colliding at a cusp bifurcation. To the left of the cusp we identify a low-activity state, $U^-$, and a high-activity one, $U^+$, which coexist and are stable in the shaded area. Parameter values: $\Delta=0.5$, $\kappa=5$, $\tau=1$. (b) Travelling front profile, computed by solving~\eqref{eq:spatialDynamic} on a truncated domain $[-30,30]$, for $\eta_0=-3$, $v_{\text{syn}}=4$. We show the profiles in the synaptic conductance variable $G(\xi)$ and the synchrony variable $R(\xi) = |Z(\xi)|$. (c) The travelling front in (b) is a heteroclinic orbit connecting the equilibria $U^+$ and $U^-$ of the spatial-dynamical system~\eqref{eq:spatialDynamic}. We show a projection of the (approximate) heteroclinic orbit in the $(A, B, G)$-space. (d) Numerical continuation of the travelling front $U(\xi)$ found in (b) (blue) and its symmetric counter $U(-\xi)$ (red), using $\eta_0$ as continuation parameter and $c$ as solution measure. The fronts live on an isolated branch and destabilise at saddle-node bifurcations. The bifurcation curve is symmetric, with respect to the axis $c=0$. Solutions with $c>0$ and $c<0$ are related via the transformation $U(\xi)\mapsto U(-\xi)$. } \label{fig:front_continuation} \end{figure*} The state at $\xi\to-\infty$ ($\xi \to \infty$) displays a high-conductance (low-conductance) $G$, hence a high (low) firing, and it is therefore referred to as the high (low) activity state (Fig. \ref{fig:front_continuation}(b)). We computed travelling waves using the routines from~\cite{Avitabile2016}, which allows us to study the spectral stability of the waves. Computations are performed on a large truncated domain $\Lambda = 60$, and plotted on $[-10,10]$ for convenience. Interestingly, we see ripples in the wake of the front, which indicates that the solution connects a node to a focus. The solution in the phase space $(A,B,G)$ illustrates that the high-activity state ($U^-$) is a focus, and the low activity one ($U^+$) a node. We continued the travelling wave shown in Fig.~\ref{fig:front_continuation}(b) (blue), and its symmetric counterpart (red), in the mean background drive $\eta_0$, using $c$ as a solution measure (Fig.~\ref{fig:front_continuation}(d)). Fronts live on an isolated branch and destabilise at saddle-node bifurcations. The bifurcation diagram is symmetric with respect to the axis $c=0$, as solutions on the branch with $c>0$ and $c<0$ are related via the transformation $U(\xi) \mapsto U(-\xi)$. Up to $6$ coexisting waves exist in a wide region of $\eta_0$ parameter space, albeit only two of them are stable. In addition, we found stable waves in which the high activity state $U^+$ is moving across the tissue, invading the low activity state $U^-$, and vice versa. The inversion of velocity occurs where the blue and red stable branches overlap. As the isola of travelling fronts is traced, these solutions gain or lose oscillations in the wake of the wave. To investigate further this aspect, we monitor the spectrum of $U^\pm$ (as equilibria of the spatial-dynamical system~\eqref{eq:spatialDynamic}) as we move along the branch. Taking a vertical excursion in Fig.~\ref{fig:front_continuation}(d), we find three independent solutions, one of which is stable and two of which unstable. We examine the stability of the fixed points $U^\pm$ in the travelling wave frame for each of these solutions. As expected, oscillations in the wake of the wave are absent where the unstable spectrum of $U^-$ is purely real (Fig.~\ref{fig:wave_connections}(a)), and they develop when a complex-conjugate pair of eigenvalues crosses the imaginary axis (Fig.~\ref{fig:wave_connections}(b)). We also find that the amplitude of the oscillations is small when the real part of the complex eigenpair is small (Fig.~\ref{fig:wave_connections}(c)). \begin{figure*} \centering \includegraphics{wave-connections} \caption{Spectrum of the high- and low-activity homogeneous states connected by selected travelling wave profiles, in Fig. \ref{fig:front_continuation}. (a) $c=-0.8769$, (b) $c=0.3594$, (c) $c=0.9470$. Parameter values: $\eta_0=-3$, $v_\text{syn}=4$, $\Delta=0.5$, $\kappa=5$, $\tau=1$.} \label{fig:wave_connections} \end{figure*} \section{\label{sec:discussion}Discussion} We have presented the derivation of an atypical neural field model from a network of spatially distributed $\theta$-neurons. In the reduced model, which we dub a \emph{next generation neural field model}, within population synchrony drives the population firing rate. The new model supports a range of patterns, such as bumps, waves and breathers. Noteworthy is the state characterised by structures within bumps, as these states are not seen in standard neural mass models. These structures instead typify patterns seen in networks of spiking neurons, which signifies that by maintaining the notion of within population synchrony this neural field model can retain information about the underlying spiking network. Exotic states, with within bump oscillations, were found in the region where the Hopf and Turing bifurcations collided. A Turing instability analysis provided us with an understanding of how the system behaved close to bifurcation points, allowing us to determine when the system transitioned from the homogeneous steady state. However, unlike the Amari model we cannot use the Heaviside approximation to make further analytical progress since the firing rate is now a fixed real valued function of the Kuramoto order parameter. As such, we have moved to numerical continuation techniques to analyse the behaviour of the system away from these bifurcation points. Numerical techniques were also used to examine the existence and stability of travelling fronts. Previous work \cite{Byrne2017} illustrated that the point version of this model with an external time-dependent drive (without spatial extent) could support $\beta$-rebound, an event-related modulation of the beta rhythm, as seen in MEG. An interesting extension of this work would be to include an additional drive in the neural field model to examine how the inclusion of space affects $\beta$ band modulations, and perhaps explain why $\beta$-rebound is seen in both the contralateral and the ipsilateral hemispheres during movement. More generally the model parameters can be altered so that the population oscillates at other frequencies, and hence, used to explain other event-related desynchronisation/synchronisation phenomena in the brain. In \S\ref{sec:numeric}, we pointed to the existence of a front which connects periodic orbits to nodes/focuses, for the model with an exponential coupling kernel. We observed such fronts by decreasing the synaptic time constant $\tau$ (Fig. \ref{fig:front_orbits}). The numerical machinery used here doesn't allow for the continuation of these solutions, known as defects. The analysis of defects is still an open problem. Close examination of Fig. \ref{fig:front_orbits} reveals that there are two fronts in the connection between the node and the limit cycle, which appear to be moving at different speeds. Even more interesting, would be the analysis of fronts which connect two periodic orbits of different amplitudes. The spreading of such a wave across the cortex could be viewed as the spreading of an epileptic seizure. \begin{figure*} \centering \includegraphics{front-oscillations} \caption{Travelling wave connecting a periodic orbit to a node: Surface plots showing the evolution of (a) the synaptic conductance $g$ and (b) the synchrony $R$ for a front which connects an oscillatory state to a fixed point state. Simulations for the system defined by \eqref{eq:EK-PDE}, with $\eta_0=-5$, $v_\text{syn}=10$, $\Delta=0.5$, $\kappa=5$, $\tau=0.2$ } \label{fig:front_orbits} \end{figure*} Another numerical challenge would be to continue the exotic patterned states seen in \S\ref{sec:turing} when the Turing and Hopf bifurcations collide. These patterns have both a spatial and a temporal period, which would require extending the numerical machineries \citep{Avitabile2016} to continue both a spatial and a temporal pattern. This has been achieved in \citep{Avitabile2008} for the Brusselator model. A natural extension to the work presented in both \S\ref{sec:turing} and \S\ref{sec:numeric} would be to include a second spatial dimension. It is more natural to view the cortex as a two dimensional sheet and examine the propagation of waves across it. We would expect that the 2D system supports the two dimensional versions of the patterned states presented here, but also potentially some more exotic states. Extending both the Turing analysis and the numerical machinery to include a second spatial extension is worthy of further exploration. \section{Acknowledgements} DA was supported by the Engineering and Physical Sciences Research Council under grant EP/P510993/1. SC was supported by the European Commission through the FP7 Marie Curie Initial Training Network 289146, NETT: Neural Engineering Transformative Technologies.

0070911b1f68035e35a75643d8238a84c3682976

\section{Introduction} Tidal friction becomes orders of magnitude larger as stars leave the main sequence (MS) and ascend the red giant branch (RGB). Binaries that suffered relatively weak tidal effects on the MS may suffer catastrophic orbital decay during the sub-giant branch (SGB) \citep{2013ApJ...772..143S}, RGB and asymptotic giant branch, resulting in the binary components coming into contact. Sufficiently massive secondaries, $M_2 \ga 10^{-2}\, M_\odot$, may then initiate a common envelope spiral-in and ejection of the primary's envelope \citep{1976IAUS...73...75P}, forming a close binary containing the secondary and a helium-core white dwarf. Smaller secondaries may give rise to mergers with the helium core of the primary, or destruction in the envelope of the primary \citep{1998A&A...335L..85N}. The engulfment of companions proceeds from smaller to larger orbital separation. Most previous studies (e.g. \citealt{2008MNRAS.386..155S,2011ApJ...737...66K,2012ApJ...761..121M,2014ApJ...794....3V}) have focused on the end result, namely the critical separation outside of which systems may survive as the primary transitions into a white dwarf. This paper focuses on earlier stages, when the primary has left the MS but is well below the tip of the RGB, as these are more commonly found in spectroscopic surveys. This paper was motivated by the close binaries found in the APOGEE survey. Employing three years of APOGEE \citep{2017AJ....154...94M} observations from the twelfth data release of the Sloan Digital Sky Survey \citep{2015ApJS..219...12A}, \cite{2016AJ....151...85T} compiled a catalog of 376 newly-found close binary systems. These are single-lined binaries with radial velocity fits for the orbit and secondary minimum mass, and a large range of Galactocentric radius, metallicity and evolutionary state of the primary. This sample is unique, as it contains both dwarf and giant primaries, and secondaries ranging from planetary to stellar masses. The present paper focuses on the 180 primaries which are post-MS, as theory predicts these systems have much stronger tidal friction. Interestingly, brown dwarf (BD) secondaries are nearly as common as stellar-mass secondaries in the APOGEE-1 sample. The putative ``brown-dwarf desert", a lack of close binaries with solar-type primaries and BD secondaries, is {\it not} found in this sample, in seeming contradiction to decades of previous surveys of FGK dwarfs \citep{2006ApJ...640.1051G}. The presence of secondaries, from planetary to brown dwarf to stellar mass, allows a test of synchronization and orbital decay over the entire range, from small-secondary Darwin unstable systems, which will come into contact due to orbital decay, as well as large-secondary systems, which will quickly synchronize. The latter then evolve on the RGB evolutionary time of the primary, with the spin frequency nearly equal to the orbital frequency as the star evolves. Both equilibrium and dynamical tides are considered in this paper. For the equilibrium tide, shearing of the tidally forced fluid motion is dissipated as heat by turbulent viscosity in the convective envelope. As gravity waves are evanescent there, and the frequencies are far below acoustic waves, the fluid motion is not wavelike. Rather, fluid nearly follows equipotential surfaces, and large scale shearing motions are present. This fluid motion is sometimes approximated with the analytic ``equilibrium tide" solution out of convenience, although that derivation is only formally valid in radiative zones \citep{1998ApJ...502..788T}. However, since the surface of the star is nearly an equipotential, this analytic solution performs well, and is convenient. It is used in this paper. The fluid motions in the convective zone are damped by turbulent viscosity, in which resonant turbulent eddies due to thermal convection transport momentum and damp the tidal shear flow. The dissipation rate for this process depends on the uncertain details of the interaction of small-scale turbulence with a mean flow, but the two theories discussed in this paper have scalings $P_{\rm orb}^{-4}$ and $P_{\rm orb}^{-5}$ for the dissipation rate, and hence decrease much more slowly than the dynamical tides' $P_{\rm orb}^{-7.7}$. Hence it is expected that the dynamical tide dominates at small separation and the equilibrium tide at larger separation. One of the aims of this paper is to estimate the critical separation at which dynamical and equilibrium tide dissipation are comparable. The dissipation of the equilibrium tide by turbulent viscosity in convection zones was first developed by \citet{1977A&A....57..383Z}. In that paper, a turbulent viscosity $\nu_\ell = v_\ell \ell / 3$ was proposed, where $\ell \sim H$ and $v_\ell \sim (F/\rho)^{1/3}$ are the size and velocity of the large, energy-bearing eddies. Here $H$ is the pressure scale height, $F$ is the heat flux, and $\rho$ is the mass density. In many situations where this theory is applied, the large eddy turn-over time, $\tau_\ell = \ell / v_\ell$ is much longer than the forcing period, $P_{\rm f}$, and the eddies cannot efficiently transport momentum and damp the shear flow. \citet{1989A&A...220..112Z} proposed that large, non-resonant eddies move a small fraction $P_{\rm f}/2\tau_\ell$ of an overturn, and so the eddy turnover time should be reduced by this linear-in-period factor. \citeauthor{1977Icar...30..301G} (1977; hereafter GN) argued that resonant eddies on smaller scales damp the shear. There was later support for dissipation by resonant eddies in both analytic calculations \citep{1997ApJ...486..403G} and numerical simulations \citep{2011ApJ...731...67P, 2011ApJ...734..118P, 2012MNRAS.422.1975O}. Assuming Kolmogorov scalings for the turbulent eddies, GN argue for a quadratic reduction factor $(P_{\rm f}/2\pi \tau_\ell)^2$. For $P_{\rm f} \ll \tau_\ell$, this leads to a very large suppression of the viscosity. By contrast, even for resonant eddies, some numerical simulations \citep{2011ApJ...731...67P, 2011ApJ...734..118P} find a scaling closer to linear, perhaps due to the fact that the largest eddies do not follow inertial-range, Kolmogorov scalings. Other simulations \citep{2012MNRAS.422.1975O} find a quadratic scalings, with a negative value for the viscosity at high frequencies. As the RGB stars studied here may have turnover times $\ga 10^2\, \rm days$ while the orbital periods are of order days to weeks, there may be several orders of magnitude difference in the predictions given by un-reduced (here called ``standard" or ``std"), linear and quadratic scalings. The treatment of the equilibrium tide in the present paper was influenced by \citet{1992RSPTA.341...39P} and \citet{1995A&A...296..709V}, who studied circularization of binaries containing an RGB star. The latter presented analytic formulae for circularization which take into account the evolution of the RGB star to larger radius and luminosity. They applied this analytic circularization formula to 28 binaries, finding that all systems could be well explained with a non-reduced kinematic viscosity. Equivalent formulae are derived here for the orbital decay problem and for reduced viscosity. Binary orbital frequencies are below the frequencies of acoustic and fundamental waves, and hence only low-frequency gravity waves can be resonantly excited by the tide. During the SGB and RGB phase, the core is radiative and internal gravity waves are excited by the tide at the boundary between the radiative zone and the convective envelope \citep{1998ApJ...507..938G, 2016CeMDA.126..275B, 2013MNRAS.434.1079C, 2013MNRAS.432.2339I, 2017A&A...604A.112G, 2017ApJ...849L..11W, 2017MNRAS.467.2146K}. The waves then travel inward toward the center. Two damping mechanisms may prevent the formation of a standing wave. Radiative diffusion is dominant on the SGB and RGB, and can easily damp the dynamical tide from any orbiting companion, independent of its mass. Even if radiative diffusion were not present, companions larger than roughly 1 Jupiter mass will give rise to such large wave amplitudes near the center that the stratification is overturned and the wave breaks, depositing it's energy as heat and torquing the gas in the wave-breaking layer \citep{2010MNRAS.404.1849B,2011MNRAS.414.1365B}. In this paper it is assumed that the angular momentum is efficiently redistributed to the rest of the star. Numerous studies (e.g., \citealt{2008MNRAS.386..155S,2011ApJ...737...66K,2012ApJ...761..121M,2014ApJ...794....3V}) have considered the effect of post-main sequence stellar evolution and tidal evolution on planetary orbits, with the goal of predicting the properties of planetary systems around white dwarfs. For massive planets in close orbits, orbital decay can lead to engulfment of the planet. By contrast, smaller mass planets that do not suffer orbital decay must wait for the star's radius to expand out to their orbit. \citet{2012ApJ...761..121M} employed turbulent viscosity in the star's convective envelope as the tidal friction, and experimented with different prescriptions for reduction of the viscosity when the eddy turnover time is longer than the forcing period. However, since the semi-major axes explored were all large ($\ga 1$ AU), eddies in the convection zone turn over fast compared to the forcing period, and the full turbulent viscosity ends up being used. For those planetary systems that do not undergo a merger, high mass loss rates during the post-main sequence evolution will cause orbits to expand as the star's mass decreases. The paper is organized as follows. The prescriptions for the equilibrium and dynamical tide dissipation rates are discussed in Sections \ref{sec:eqtide} and \ref{Dynamical Tide Dissipation Rate}, respectively. Section \ref{sec:examples} shows examples of orbital evolution. Theory is compared with data from the APOGEE survey and the exoplanets in Section \ref{sec:acrit}. Conclusions are given in Section \ref{sec:conclusion}. Analytic approximations to the equilibrium tide dissipation rate and critical separation are given in Appendices \ref{sec:analytic_eq_tide} and \ref{sec:analytic_acrit}, respectively. \section{Equilibrium Tide Dissipation Rate} \label{sec:eqtide} Consider a primary star of mass $M_1$ and radius $R_1$ in a circular orbit of separation $a$ with a secondary star of mass $M_2$. The tide raised by the secondary in the convective envelope of the primary creates a time-dependent fluid shear, which is damped by the turbulent viscosity of convective eddies. This process transfers energy and angular momentum from the orbit to the stellar convection zone. For simplicity, efficient angular momentum redistribution is assumed in the primary so that the rotation rate, $\Omega$, is uniform over the star. The orbital torque, $N$, is related to the energy dissipation rate in the rotating frame, $\dot{E}$, through the pattern speed $n-\Omega$ as $N=\dot{E}/(\Omega-n)$, where $n=\sqrt{ G(M_1+M_2)/a^3}$ is the orbital frequency. This relation is valid even as $\Omega \rightarrow n$, i.e. synchronous rotation. The torque $N \rightarrow 0$, since $\dot{E}\sim(\Omega-n)^2 \rightarrow 0$ as $\Omega \rightarrow n$. This torque changes the orbital angular momentum, $L=\mu n a^2$, where $\mu=M_1M_2/(M_1+M_2)$ is the reduced mass of the system. The semi-major axis then changes at a rate \begin{equation} \label{dota_dotE} \dot{a} = - \frac{2\dot{E}}{\mu n (n - \Omega)a}. \end{equation} Since $\dot{E}>0$, the orbit decays for a slowly rotating star ($\Omega < n$) and expands for a rapidly rotating star ($\Omega > n$). Equation~\ref{dota_dotE} agrees with the end result of the derivation in \cite{1981A&A....99..126H}, and is also valid even when the moment of inertia of the primary star is changing with time due to stellar evolution. For massive secondaries, the spin of the primary may be tidally synchronized to the orbit, with subsequent tidal evolution occurring on the stellar evolution timescale of the primary (e.g., \citealt{2015A&A...574A..39D}). The conserved total angular momentum of the system, assumed aligned, is $J=L + S_1 + S_2$, where $S_1=I_1\Omega$ is the spin angular momentum of the primary, and $S_2$ is that of the secondary. Here $I_1$ is the moment of inertia of the primary, and $\Omega$ is the primary's angular velocity. Since the moment of inertia of the secondary $I_2$ is much smaller than $I_1$, we ignore $S_2$. In that case, \begin{eqnarray} L + S_1 & \simeq & \rm constant. \label{eq:LplusS} \end{eqnarray} At each time $t$, $\Omega(t)$ is determined from the initial values $L(0)$ and $S_1(0)$ as $\Omega(t) = (L(0)+S_1(0)-L(t))/I_1(t)$. In this way a separate differential equation is not needed for $\Omega$. One technical point is that even a tiny amount of mass loss would cause an artificial spinup of the primary according to Equation \ref{eq:LplusS}, since $L$ would decrease even at fixed $a$. To eliminate this issue, the two masses are fixed during the evolution. This is a good approximation for the stars near the base of the giant branch considered here, since little mass has been lost, and e.g., the expansion of the orbit would be tiny at this stage. The viscous dissipation rate for incompressible flow is given by \citep{1959flme.book.....L} \begin{equation} \label{eq:Edotvisc} \dot{E} = \frac{1}{2} \sum_{i=x,y,z} \sum_{k=x,y,z} \int \rho \nu \bigg( \frac{\partial v_i}{\partial x_k} + \frac{\partial v_k}{\partial x_i} \bigg)^2 d^3x \end{equation} where the sums are over the three spatial directions, $v_i$ is the velocity of the tidal flow, $x_i$ are Cartesian coordinates, $\nu$ is the (isotropic) kinematic viscosity, and $\rho$ is the mass density. Isotropic turbulent viscosity is assumed for simplicity. Numerical simulations find modest deviations from isotropy for Boussinesq convection \citep{2009ApJ...705..285P,2011ApJ...734..118P}. The velocity of the tidal flow may be represented as a spherical harmonic expansion \begin{eqnarray} \label{eq:vexpansion} \vect{v} & = & \sum_{\ell m} (-i \omega_m) \left[ \xi_{r, \ell m}(r)Y_{\ell m}(\theta,\phi)\vect{e}_r \right. \nonumber \\ & + & \left. \xi_{h, lm}(r) r \vect{\nabla} Y_{\ell m}(\theta,\phi) \right] e^{-i \omega_m t} \end{eqnarray} where $\xi_{r, \ell m}$ and $\xi_{h,\ell m}$ are the radial and horizontal component of the Lagrangian displacement vector, and $\omega_m = m(n-\Omega)$ is the forcing frequency in the rotating frame. Plugging Equation \ref{eq:vexpansion} into Equation \ref{eq:Edotvisc} and performing the angular integral gives \begin{equation} \begin{split} \dot{E} & = \frac{1}{2} \sum_{\ell m}\omega_{m}^2 \int^{R_1}_{r_{\rm{bcz}}} dr r^2\rho \nu \bigg[4\bigg( \frac{d\xi_{r, \ell m}(r)}{dr} \bigg)^2 \\ & + 2\ell(\ell+1)\bigg( \frac{d\xi_{h, \ell m}(r)}{dr} + \frac{\xi_{r, \ell m}(r)}{r} - \frac{\xi_{h, \ell m}(r)}{r} \bigg)^2 \\ & + 2\bigg(\ell(\ell+1)\frac{\xi_{h, \ell m}(r)}{r} - 2\frac{\xi_{r, \ell m}(r)}{r}\bigg)^2 \bigg] \end{split} \label{Edot_general} \end{equation} where $r_{\rm bcz}$ is the radius at the base of surface convective zone. Since the $m=0$ term has zero frequency it does not contribute. The $\pm m$ terms give equal contributions. The tidal potential in the primary due to a secondary orbiting at co-latitude $\pi/2$ and and orbit angle $(n-\Omega)t$ is \begin{eqnarray} \begin{split} U = & \sum_{\ell=2}^\infty \sum_{m=-\ell}^\ell U_{\ell m}(r) Y_{\ell m}(\theta,\phi) e^{-i\omega_{m}t} \\ = & -GM_2 \sum_{\ell=2}^\infty \sum_{m=-\ell}^\ell \frac{4\pi}{2\ell+1} \frac{r^{\ell}}{a^{\ell+1}}\\ & \times Y_{\ell m}\left( \frac{\pi}{2},0 \right) Y_{\ell m}(\theta,\phi) e^{-i\omega_{m}t}, \end{split} \end{eqnarray} which is smaller than the potential $GM_1/R_1$ at the surface of the primary by a small factor $\epsilon = (M_2/M_1)(R_1/a)^3$ for $\ell=2$. In the equilibrium tide approximation (e.g., \citealt{1977Icar...30..301G}), the radial and horizontal displacements are \begin{equation*} \xi_{r, \ell m}(r) = -\frac{U_{\ell m}(r)}{g} \end{equation*} and \begin{equation*} \xi_{h, \ell m}(r) = -\frac{1}{\ell(\ell+1)} \frac{ U_{\ell m}(r) }{ g} \bigg( 2\ell - \frac{d\ln g}{d\ln r} \bigg) \end{equation*} Given the run of $\rho$, $\nu$, $g$, $\xi_{r,\ell m}$ and $\xi_{h,\ell m}$ versus $r$ for a stellar model, $\dot{E}$ is computed by numerical integration of Equation \ref{Edot_general} for a given stellar model. Analytic approximations for this integral are discussed in the Appendix. When the turnover time $\tau_{\rm ed}$ of the large eddies becomes longer than the forcing period $P_f=P_{\rm orb}/2$ (for $m=2$), it is expected than turbulent viscosity is reduced, since the large eddies cannot transport momentum efficiently. Three models of turbulent viscosity will be investigated: un-reduced (``standard") viscosity \begin{eqnarray} \nu_{\rm std} & = & \frac{1}{3} v_{\rm ed} \alpha H, \label{eq:nustd} \end{eqnarray} Zahn's formula with a linear reduction \begin{eqnarray} \nu_{\rm Z} & = & \nu_{\rm std} \times {\rm min} \left( 1 , \frac{P_f}{2\tau_{\rm ed}} \right), \label{eq:nuZ} \end{eqnarray} and Goldreich and Nicholson's formula with a quadratic reduction \begin{eqnarray} \nu_{\rm GN} & = & \nu_{\rm std} \times {\rm min} \left( 1 , \left( \frac{P_f}{2\pi \tau_{\rm ed}} \right)^2 \right), \label{eq:nuGN} \end{eqnarray} The numerical factors in each expression are somewhat arbitrary \citep{2009ApJ...705..285P}. Due to the large eddy velocity near the surface, the eddy turnover time will generally be shorter than forcing periods of interest there. Reduced viscosity is important deep in the convection zone where eddy velocities are small, due to increasing density. If the viscosity scales with orbital period $P_{\rm orb}=2\pi/n$ as $\nu \propto P_{\rm orb}^\alpha$, then for $\Omega \ll n$, $\dot{E} \propto M_2^2 a^{-9+3\alpha/2}=M_2^2 a^{-(\beta+2)}$ and $\dot{a} \propto M_2 a^{-7+3\alpha/2}=M_2 a^{-\beta}$ with $\beta=7-3\alpha/2$. The relevant values of $\beta$ are then 7, 5.5 and 4 for un-reduced, linear and quadratic scalings. The exponent $\beta$ is a crucial parameter that directly determines the relative number of systems at different orbital separation. Appendix \ref{sec:analytic_eq_tide} discusses analytic approximations to the dissipation rate for the different turbulent viscosity assumptions. The convective envelope is treated as an $n=3/2$ polytrope, with the entropy increasing as the star ascends the RGB. These approximations will be used in Section \ref{sec:acrit} to understand the critical orbital separation out to which orbital decay is expected to have caused systems to merge. \subsection{A Numerical Example} For close-in orbits and deep stellar convection zones, the viscosity is expected to be significantly reduced. This section shows a numerical example to illustrate the reduction factor. Figure \ref{fig:viscosity_energy_dissipation_logP} shows the depth dependence of turbulent viscosity (top panel), eddy turnover time and interior mass (middle panel), and $d\dot{E}/d\ln P$ from Equation \ref{Edot_general} (bottom panel); here $P$ is pressure. Representative values have been used, with a $M_1=1M_{\odot}$ RGB star with radius $R_1=10R_{\odot}$, companion mass $M_2=0.01\, M_{\odot}$, orbital separation $a=0.1\, \rm AU$ and orbital period $P_{\rm orb}=11.4$ days. The middle panel of Figure \ref{fig:viscosity_energy_dissipation_logP} shows that this RGB star has eddy turnover times $\tau_{\rm ed} = \alpha_{\rm mlt} H/v_{\rm ed} \ga 100\, \rm days$ near the peak of the integrand. Hence tidal forcing for any orbit shorter than $P_{\rm orb}\simeq 200\, \rm days$ is expected to have reduced viscosity. The peak in eddy turnover time is due to the scale height becoming small both near the surface, due to lower temperature, and toward the center, due to higher gravity. Near the peak of the integrand the eddies are turning over 10 times slower than the tidal forcing period. The top panel shows that this leads to a reduction factor of $10^{-1}$ and $10^{-4}$ relative to un-reduced viscosity, for the Zahn and GN viscosities, respectively. The bottom panel shows that $d\dot{E}/d\ln P$ has a strong peak for un-reduced viscosity, due to decreasing mass near the surface and decreasing tidal shear toward the interior. The integrands for Zahn and GN viscosities are reduced in size and with flattened peaks closer to the surface. The Zahn case still has a peak in the integrand well below the depth where $P_f = \tau_{\rm ed}$, however the GN case has a flat integrand over 3-4 decades in pressure, ending at $P_f = \tau_{\rm ed}$. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{10Msun_10Rsun_nu_Edot.pdf} \caption{Depth dependence of quantities needed for the turbulent viscosity dissipation rate in the convective envelope. The top panel shows the run of the three prescriptions for turbulent viscosity as a function of pressure in the convective envelope. The middle panel shows the eddy turnover time (green solid line), forcing period ($P_{\rm orb}/2=5.7\, \rm days$, the black horizontal line) and interior mass $m(r)$ in terms of ${\rm log}\,P$. The bottom plot gives the integrand of the energy dissipation integral. The top and bottom panels show the un-reduced, Zahn and GN prescriptions as black, green and cyan lines, respectively. The base of the convective zone is on the right of the figure, where $\log\,P = 13.5$. The parameters used are a $M_1=1\, M_{\odot}$ RGB star with radius $R_1 = 10\, R_{\odot}$, companion mass $M_2=0.01\, M_{\odot}$, and separation $a = 0.1$ AU (orbital period $P_{\rm orb}=11.4$ days).} \label{fig:viscosity_energy_dissipation_logP} \end{figure} \section{ Dynamical Tide Dissipation Rate} \label{Dynamical Tide Dissipation Rate} The dynamical tide involves the excitation of internal gravity waves near the radiative-convective boundary (RCB). If damping is weak, waves will reflect in the core and form standing waves. When the wave can be damped in less than one group velocity travel time the result is a traveling wave. In the traveling wave regime, the dissipation rate is given by the inward-going wave luminosity, $L_{\rm dyn}$. During the SGB and RGB phases the core is radiative and the envelope convective. The dynamical tide is excited at the RCB, with the waves propagating inward toward the center (e.g., \citealt{1998ApJ...507..938G}). There are two damping mechanisms that may cause the wave to damp at or before it reaches the center. First, radiative diffusion damping becomes progressively more important as the star evolves off the MS, due to the large number of internal gravity wave nodes in the core \citep{1977AcA....27...95D}. This mechanism depends only on the orbital period, and is independent of the companion mass. Second, waves may ``break" nonlinearly at the center, and not reflect back \citep{2010MNRAS.404.1849B,2011MNRAS.414.1365B}. Here wave breaking may mean overturning the local stratification, or strong wave-wave interactions which transfer energy from the tidally-excited fluid motion to daughter waves \citep{2012ApJ...751..136W,2010MNRAS.404.1849B,2011MNRAS.414.1365B, 2016ApJ...816...18E}. Nonlinear wave breaking depends on the companion mass, as well as the orbital period. The traveling wave luminosity is computed from stellar models by solving the linearized radial momentum and continuity equations and applying the appropriate boundary conditions for an inward-going traveling wave in the radiative zone. Define the radial Lagrangian displacement $\xi_r$, the potential $\psi = \delta p/\rho + U$, the Eulerian pressure perturbation $\delta p$, and the tidal potential $U$. The radial momentum equation and continuity equations, in the Cowling approximation, are then (e.g., \citealt{1989nos..book.....U}) \begin{eqnarray} \frac{d\psi}{dr} & = & \frac{N^2}{g} \left( \psi - U \right) - (N^2-\sigma^2) \xi_r \end{eqnarray} and \begin{eqnarray} \frac{d\xi_r}{dr} & = & \xi_r \left( \frac{g}{c^2} - \frac{2}{r} \right) + \psi \left( \frac{k_h^2}{\sigma^2} - \frac{1}{c^2} \right) + \frac{U}{c^2}, \end{eqnarray} where the horizontal wavenumber is $k_h^2=\ell(\ell+1)/r^2 $. At the surface, the evanescent wave boundary condition is \citep{1989nos..book.....U} $\delta p/\rho = \psi - U = g\xi_r$. The inward-going traveling wave boundary condition is \begin{eqnarray} \frac{d\left( \psi - \psi_0 \right)}{dr} & = & i k_r \left( \psi - \psi_0 \right), \end{eqnarray} where $\psi_0$ is an approximate long-wavelength, particular solution and $k_r=k_hN/\omega$ is the radial wavenumber. The value of $\psi_0$ can be computed from the equilibrium tide as \citep{2010ApJ...714....1A,2012ApJ...751..136W, 2017ApJ...849L..11W} \begin{eqnarray} \psi_0 & = & \frac{\omega^2}{\ell(\ell+1)} \frac{d(r^2\xi_{\rm r,eq})}{dr}. \end{eqnarray} The dynamical tide pieces of $\xi_r$ and $\psi$ are denoted $\xi_{\rm r,dyn} = \xi_r - \xi_{\rm r,eq}$ and $\psi_{\rm dyn} = \psi - \psi_0$. In the gravity wave propagation zone, the traveling wave luminosity is given by \begin{equation} L_{\rm dyn} = r^2 \int d\Omega \rho \psi_{\rm dyn} \dot{\xi}_{\rm r,dyn}, \label{L_dyn_numerical} \end{equation} which involves an integral over two spherical harmonics. The purpose of subtracting the long-wavelength response is to decrease the size of the oscillatory part of $L_{\rm dyn}(r)$, making it easier to isolate the value in the propagation zone far away from the RCB. The decrease of gravity wave energy in the core due to radiative diffusion can be parametrized as \begin{eqnarray} \alpha & \equiv & \frac{\dot{E}_{\rm diff}}{L_{\rm dyn}} \simeq 2 \int_0^{r_{\rm rcb}} dr \frac{k_h^3 N}{\omega^4} \nabla_{\rm ad}(\nabla_{\rm ad}-\nabla)\frac{4\sigma T^4 g^2}{3\kappa P^2} \nonumber \\ & \equiv & \left( \frac{P_{\rm orb}}{P_{\rm orb, diff}} \right)^4. \label{eq:alpha} \end{eqnarray} Here $\dot{E}_{\rm diff}$ is the dissipation rate including both the inward and outward going waves, giving the factor of 2 in Equation \ref{eq:alpha}. The adiabatic and stellar temperature gradients are denoted $\nabla_{\rm ad}$ and $\nabla$, respectively, and $\kappa$ is the opacity. This formula was derived for the dynamical tide component using the low frequency, quasi-adiabatic approximation discussed in \citet{1989nos..book.....U}, and is valid even in the degenerate core and non-degenerate burning shell, where most of the contribution arises. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{Porbcrit_vs_Radius.pdf} \caption{ Critical orbital period during the evolution for stars of mass $M_1=1\,\,(\rm top\,\,panel),2\,\,(\rm middle\,\,panel),3\, M_\odot$ (bottom panel). The solid line shows the critical orbital period, $P_{\rm orb,\,diff}$ outside of which the radiative diffusion damping timescale is shorter than twice the group travel time across the core, and the traveling wave regime obtains. The dashed lines show the minimum possible orbital period $P_{\rm orb,\,min}$ for a star of that mass and radius, where the companion is orbiting at the surface of the star. At a given stellar radius, the standing wave limit can only occur if the dashed line is below the solid line.} \label{fig:Porbcritdiff} \end{figure} Outside $P_{\rm orb, diff}$, defined in Equation \ref{eq:alpha}, radiative diffusion damping is strong, and the traveling wave limit occurs, and vice versa for $P_{\rm orb} < P_{\rm orb, diff}$. Figure \ref{fig:Porbcritdiff} shows $P_{\rm orb,\,diff}$ versus stellar radius, $R$, which acts as a proxy for age. In all three cases, radiative diffusion will lead to the traveling wave limit for orbital periods $P_{\rm orb} \ga 1\, \rm day$ over part of the SGB and all the RGB. This enhanced radiative diffusion is caused by the short vertical wavelength in the core. Radiative diffusion may also have an important effect on the driving of waves at the RCB. For very thin surface convection zones in MS stars of mass $M_1 \ga 1.3\, M_\odot$, the thermal time can become shorter than the forcing period, and radiative diffusion will rapidly damp out temperature differences induced by the wave, effectively eliminating the buoyancy force. The wave luminosity is expected to be suppressed when $\omega \la \omega_{\rm diff}$, where the thermal diffusion frequency at the RCB is approximated as \begin{eqnarray} \omega_{\rm diff} & \simeq & \frac{F}{\rho g \lambda^2}. \label{omega_diff} \end{eqnarray} Here $F$ is the flux and $\lambda \equiv|\ell(\ell+1)(dN^2/dr) / (\omega^2r^2)|^{-1/3} \simeq H (\omega/c_sk_h)^{2/3}$ is the Airy wavelength of the gravity wave \citep{1998ApJ...507..938G}. In this paper it is assumed that rapid thermal diffusion will greatly reduce the wave luminosity, so that it is set to zero when the thermal time $t_{\rm th}=PC_pT/gF$ is shorter than the forcing period $P_{\rm f}$. As the convective envelope rapidly deepens on the SGB, where $\omega_{\rm diff}$ decreases fast and $t_{\rm th}$ increases fast, $\omega = \omega_{\rm diff}$ and $t_{\rm th}=P_{\rm f}$ occur at almost the same stage during the evolution. Therefore, turning on the dynamical tides at $t_{\rm th}>P_{\rm f}$ is a good approximation. This effectively sets the inward-going wave luminosity to zero on the MS for the $M_1=2$ and $3\, M_\odot$ cases. As the star leaves the MS, the convection zone will deepen rapidly and hence our assumption allows the dynamical tide dissipation to turn on suddenly at the end of the MS. This effect is clearly evident in the results for the $M_1=2$ and $3\, M_\odot$ stars. In the calculations of orbital decay in later sections, it is convenient to have tabulated formulas for the wave luminosity that can be rapidly evaluated, as opposed to solving the above boundary value problem. If $\lambda \simeq \left( \omega / ck_h \right)^{2/3} H$ is much smaller than the local scale height $H$, which is valid when the forcing frequency is much smaller than the Lamb frequency at the RCB, the Airy approximation is good and the wave luminosity can be written in the form \citep{1998ApJ...507..938G} \begin{equation} \begin{split} L_{\rm dyn} = & \bigg(\frac{3^{2/3}\Gamma^2(1/3)}{2\pi}\bigg)[\ell(\ell+1)]^{-4/3}\omega^{11/3}\\ & \times \bigg(\rho r^5 \bigg|\frac{dN^2}{d{\rm ln}r}\bigg|^{-1/3}\frac{\zeta^2\xi^2_{r,{\rm eq}}}{r^2}\bigg), \end{split} \label{L_dyn_analytical} \end{equation} where the dimensionless parameter $\zeta$ is defined by \begin{eqnarray} \frac{d\xi_{\rm r,dyn}}{dr} & \equiv & \zeta \frac{\xi_{\rm r,eq}}{r}, \end{eqnarray} and $\zeta$ is set by matching the solution in the convection zone to that in the radiative zone. All the quantities in Equation \ref{L_dyn_analytical} are evaluated at the RCB. MESA models are used to compute $\zeta$ as a function of age for each stellar model. Our numerical calculations find that $\zeta$ grows strongly during the RGB. Equating the analytical formula in Equation \ref{L_dyn_analytical} to the numerical results generated from Equation \ref{L_dyn_numerical} for each stellar model gives the parameter $\zeta$. When both dynamical and equilibrium tides are included, the orbital decay rate becomes $\dot{a} = - 2(\dot{E}_{\rm eq} + |L_{\rm dyn}|)/ \mu n (n - \Omega)a$. When applying Equation \ref{L_dyn_analytical} in the calculations of orbital decay, the use of the Airy approximation on the radiative side of the RCB requires that the wavelength is always much shorter than a scale height. For closer orbits, the larger forcing frequency implies larger wavelengths, and the luminosity can be larger than implied by Equation \ref{L_dyn_analytical}, and this approach may underestimate the orbital decay rate. Given $L_{\rm dyn}$, the nonlinearity of the wave must be checked by evaluating \begin{equation} k_r \xi_{r,{\rm dyn}}=\sqrt{\frac{[\ell(\ell+1)]^{3/2} N L_{\rm dyn}}{4\pi\rho r^5\omega^4}} \end{equation} in the radiative zone. If $\alpha > 1$, or if $\alpha < 1$ but $k_r \xi_{r,{\rm dyn}}> 1$, the dissipation rate is given by the full $|L_{\rm dyn}|$. On the other hand, if $\alpha < 1$ and $k_r \xi_{r,{\rm dyn}}<1$, the dissipation rate is set to zero as the wave reflects back and forms a standing wave, with much smaller dissipation rate. Lastly, we have ignored the Coriolis force in computing the dynamical or quasi-static tidal response. This is not a good approximation when the rotation is nearly synchronized, as occurs for sufficiently massive companions. As the forcing frequency $\sim 2(n-\Omega)$ becomes small, inertial waves may be excited, possibly giving much larger dissipation rates than included here (e.g. \citealt{2007ApJ...661.1180O}). For SGB and RGB stars, the large expansion after the MS implies these stars will be slowly rotating before tidal torques become important. In this case, the bottleneck is first spinning the stars up to synchronous rotation. At that point, the orbital decay rate becomes small and further evolution is on the nuclear timescale of the primary. Increasing the dissipation rate at fixed $n-\Omega$ will tend to make $n-\Omega$ smaller, until the synchronization rate becomes comparable to the nuclear timescale of the primary. Hence we believe including the Coriolis force and inertial waves will not significantly change our results for orbital decay. \section{ Examples of Orbital Decay} \label{sec:examples} This section presents calculations of orbital decay for a range of primary mass, secondary mass and initial separation. Each integration of the equation $\dot{a} = - 2(\dot{E}_{\rm eq} + |L_{\rm dyn}|)/ \mu n (n - \Omega)a$ includes the dynamical tide, as well as a prescription for the turbulent viscosity used to damp the equilibrium tide. Calculations using Zahn and Goldreich-Nicholson viscosity are compared to assess if they result in potentially detectable differences in the critical separation for rapid orbital decay. The range of substellar companion masses is chosen to span the range of synchronized and non-synchronized cases. The range of primary masses and post-MS evolutionary state represent the bulk of binaries with SGB and RGB primaries and a substellar secondary. Modules for Experiments in Stellar Astrophysics (MESA, {version 8845}, \citealt{2011ApJS..192....3P,2013ApJS..208....4P,2015ApJS..220...15P}) is used to provide the stellar structure for three stars of mass $M_1=1, 2$ and $3\, M_{\odot}$ during the MS, and ending on the RGB. The initial metallicity is $Z=0.02$. The type 2 opacity table is used. The nuclear burning network used is ``o18\_and\_ne22.net". The mixing length factor is 2. The Schwarzschild criterion for the definition of the convective zone is applied. In the following three sections, results are presented for the three different primary masses. \subsection{$1\, M_\odot$ Model} Sun-like stars have a radiative core and a relatively thick convective envelope on the MS, allowing the ingoing-wave dynamical tide to operate. The convection zone deepens significantly on the RGB in both mass and radius, giving rise to equilibrium tide dissipation rates many orders of magnitude larger than on the MS. At fixed semi-major axis, the dynamical tide increases strongly during the SGB, and is relatively constant on the RGB. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{1M_10MJ_ZN.pdf} \caption{Orbital decay for a $M_1=1\, M_\odot$ primary, with companion mass $M_2= 0.01\, M_{\odot}$ and initial separations $a=$ 0.05 (black), 0.15 (green) and 0.25 (cyan) AU. Zahn's turbulent viscosity prescription ($\nu_{\rm Z}$) has been used for the equilibrium tide. Here $\log g$ of the primary star shows the evolution of the star, from left to right. (Top panel) Semi-major axis vs $\log g$. The black dotted-dash line shows the stellar radius, $R_1$. (Middle panel) Primary rotation rate ($\Omega$, solid lines) and binary orbital frequency ($n$, dashed lines). (Bottom panel) Dynamical tide ($|L_{\rm dyn}|$, dashed lines) and equilibrium tide ($\dot{E}_{\rm eq}$, solid lines) dissipation rates. } \label{1M_10MJ_ZN} \end{figure} The top panel of Figure~\ref{1M_10MJ_ZN} shows the evolution of $a$ for a $M_1=1M_{\odot}$ primary and a $M_2=0.01M_{\odot}$ secondary, including dynamical tides and equilibrium tides with Zahn's prescription ($\nu_{\rm Z}$) for viscosity. The surface gravity, $\log\,g\equiv \log_{10}(g/(\rm cm\,s^{-2}))$ is a proxy for the evolutionary state of the primary star from the MS phase ($\log\,g =4.5 $) to the RGB ($\log\,g \ll 4.5$). When the orbital decay rate is small, $a$ is constant. Since the orbital decay rate increases so strongly after the MS, the orbit will decay rapidly compared to the stellar evolution timescale, and the line will become nearly vertical. The system merges at the end of the MS for $a_{\rm init}=0.05$ AU, and during the RGB phase for $a_{\rm init}=0.15$ and $0.25$ AU. From the middle panel, none of the examples synchronize for this relatively low companion mass (starting from $\Omega=0$). The bottom panel shows the dynamical tide wave luminosity $L_{\rm dyn}$ and the equilibrium tide energy dissipation rate $\dot{E}_{\rm eq}$ during the evolution. For the case of $a_{\rm init}=0.05$ AU, the dynamical tide dominates. For the case of $a_{\rm init}=0.15$, the dynamical and equilibrium tides are comparable before the RGB phase, then the equilibrium tide increases much faster than the dynamical tide. In the case of $a_{\rm init}=0.25\, $AU, the orbit shrinks mainly due to the equilibrium tide. There are two trends that favor the equilibrium tide over the dynamical tide for wider orbits. First, they have different dependence on orbital period, with $\dot{E}_{\rm eq} \propto P_{\rm orb}^{-5.5}$ and $L_{\rm dyn} \propto P_{\rm orb}^{-7.67}$, so the former decreases outward more slowly. Second, the dynamical tide luminosity has an initial increase by several orders of magnitude during the SGB, but then becomes relatively constant during the RGB (at fixed semi-major axis). This is in contrast to the equilibrium tide, which shows a continuous increase up the RGB. Hence for decay of wider orbits, which occurs for a more evolved primary star, the equilibrium tidal friction is more important. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{1M_10MJ_GN.pdf} \caption{ Same as Fig.~\ref{1M_10MJ_ZN} but using GN's viscosity ($\nu_{\rm GN}$) for the equilibrium tide dissipation. } \label{1M_10MJ_GN} \end{figure} Figure \ref{1M_10MJ_GN} is the same as Figure \ref{1M_10MJ_ZN} but uses GN's viscosity prescription instead of Zahn's. The equilibrium tide dissipation rate is then much smaller since $\nu_{\rm GN} \ll \nu_{\rm Z}$. The $a=0.05\, \rm AU$ orbit decays at a similar time as in Figure \ref{1M_10MJ_ZN} since the dynamical tide dominates. It also dominates early in the evolution for the $a=0.15\, \rm AU$ case, although there is little orbital decay during that time. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{1M_015_big_companion_ZN.pdf} \caption{The effect of synchronous rotation on orbital decay for $M_1=1\, M_\odot$, $M_2=0.001$ (black lines), $0.01$(green lines) and $0.2\, M_\odot$ (cyan lines). Zahn's viscosity is used. (top panel) Semi-major axis ($a$) versus evolutionary state of the primary ($\log g$), and stellar radius. (bottom panel) Primary rotation rate ($\Omega$, solid lines), orbital frequency ($n$, dashed lines) and orbital frequency at which the Darwin instability begins ($n_{\rm max}$, dotted line), evaluated for secondary mass $M_2=0.2\, M_\odot$.} \label{1M_015_big_companion_ZN.pdf} \end{figure} Next, Figure \ref{1M_015_big_companion_ZN.pdf} compares the orbital decay for a $M_1=1\, M_\odot$ primary with three different companion masses with Zahn's reduced viscosity. For the low mass companion, $M_2=10^{-3}\, M_{\odot} \simeq 1\, M_{\rm Jup}$, $\dot{a} \propto M_2$ leads to a small orbital decay rate, and the spin is far from synchronous. By comparison, the $M_2=0.01M_{\odot}$ case is also not synchronized, but the orbital decay occurs faster due to the larger mass. The higher mass, $M_2=0.2\, M_\odot$ case would have had even stronger orbital decay if synchronization did not occur. However, this system synchronizes on the SGB at $\log\,g=3.5$ by the equilibrium tide, after which point the orbit evolves on the much slower stellar evolution timescale, so that this case actually lives {\it longer} than the two lower mass cases. The lower panel of Figure \ref{1M_015_big_companion_ZN.pdf} shows that destruction occurs due to the Darwin instability, at separation $a_{\rm D} \simeq \left( 3I_1/\mu \right)^{1/2} \simeq \left(3I_1/M_2 \right)^{1/2}$ and orbital frequency $ n_{\rm max} \simeq \sqrt{G(M_1+M_2)/a_{\rm D}^3}\simeq \sqrt{GM_1/a_{\rm D}^3}$. \footnote{$a_{\rm D}$ can be derived by solving for the separation where $\partial J/\partial n =0$ assuming $\Omega=n$.} Beyond this point, the spin has larger angular momentum than the orbit, and the orbital frequency will tend to increase rapidly, leaving the spin frequency behind. Hence the critical semi-major axis out to which orbits are destroyed is not a monotonic function of $M_2$; for small $M_2$ it increases (see Appendix \ref{sec:analytic_acrit}), while for large $M_2$ it decreases. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{1M_015_big_companion_GN.pdf} \caption{Same as Figure \ref{1M_015_big_companion_ZN.pdf} but with GN's reduced viscosity. } \label{1M_015_big_companion_GN.pdf} \end{figure} Figure \ref{1M_015_big_companion_GN.pdf} is the same as Figure \ref{1M_015_big_companion_ZN.pdf} but uses GN's reduced viscosity prescription. For the $M_2=0.001M_{\odot}$ and $M_2=0.01M_{\odot}$ cases, the systems merge later than for $\nu_{\rm Z}$ since $\nu_{\rm GN} \ll \nu_{\rm Z}$. As the primary is not synchronized in both cases, the system merges earlier for the higher mass companion. For the $M_2=0.2M_{\odot}$ high mass companion, the orbit is synchronized later than in Figure \ref{1M_015_big_companion_ZN.pdf}, but the merger occurs at a similar time due to the Darwin instability. \subsection{2 $M_{\odot}$ Model} For $M_1=2\, M_{\odot}$, the star has a convective core on the MS which inhibits nonlinear wave breaking at the center.\footnote{ An outward traveling wave flux excited at the radiative-convective boundary of the central convection zone is ignored in this paper.} More importantly, the surface convection zone is so thin that radiative diffusion damping suppresses the driving of the wave (see Equation \ref{omega_diff} and the surrounding discussion). On the SGB, the core becomes radiative and the surface convection zone deepens, at which point both efficient excitation and nonlinear wave breaking can occur. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{2M_10MJ_ZN.pdf} \caption{Same as Figure \ref{1M_10MJ_ZN} but with $M_1=2\, M_\odot$, $M_2=0.01\, M_\odot$ and $a_{\rm init}=0.05$ (black lines), 0.10 (green lines) and 0.15 (cyan lines) AU. Zahn's turbulent viscosity prescription ($\nu_{\rm Z}$) has been used for the equilibrium tide. } \label{2M_10MJ_ZN.pdf} \end{figure} Figure \ref{2M_10MJ_ZN.pdf} shows the evolution for $M_1=2\, M_{\odot}$ and $M_2=0.01M_{\odot}$ starting from $a_{\rm init}=$ 0.05, 0.10 and 0.15 AU. Zahn's viscosity is used, and the dynamical tide is included. The bottom panel shows that $L_{\rm dyn}=0$ until $\log g=3.5$, where the surface convection zone deepens and $t_{\rm th}>P_{\rm f}$. Further $\dot{E}_{\rm eq}$ is much smaller than the $M_1=1\, M_\odot$ case on the MS, due to the smaller convection zone. The end result is that tidal friction in this model is much weaker than the $M_1=1\, M_\odot$ model on the MS, but comparable on the SGB and RGB. The middle panel shows that none of the cases synchronize for this companion mass. In the bottom panel, $L_{\rm dyn}$ dominates by orders of magnitude in the $a=0.05\, \rm AU$ and $a=0.1$ AU cases once it turns on. This causes the immediate decay of the orbit, on a shorter timescale than the radius expansion of the primary, seen as vertical lines in the top panel. For the $0.15\, \rm AU$ case, dynamical tides are important only briefly after they turn on, and then the equilibrium tide dominates until the merger. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{2M_10MJ_GN.pdf} \caption{Same as Figure \ref{2M_10MJ_ZN.pdf} but with GN's viscosity prescription. } \label{2M_10MJ_GN.pdf} \end{figure} Figure \ref{2M_10MJ_GN.pdf} is the same as Figure \ref{2M_10MJ_ZN.pdf} but uses GN's viscosity prescription. There is almost no difference for the merger time for the case of $a_{\rm init}=$ 0.05 and 0.1 AU since dynamical tides dominate. However, since the $a_{\rm init}=$ 0.15 AU case is dominated by the equilibrium tide, it merges later than for Zahn's prescription, since $\dot{E}_{\rm eq}$ is much smaller. For the $a_{\rm init}=$ 0.15 AU case, the system merges at $\log\,g=2.4$, which is near dredge-up, where the convective envelope reaches the hydrogen burning shell, and the radius first contracts and then expands. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{2M_010_big_companion_ZN.pdf} \caption{Same as Figure \ref{1M_015_big_companion_ZN.pdf} but with $M_1=2\, M_\odot$, and $M_2=$ 0.001 (black), 0.01 (green) and 0.5 (cyan), starting separation $a_{\rm init}=0.1\, \rm AU$, and uses Zahn's prescription for viscosity. In the lower panel, the green dotted line displays the orbital frequency above which the Darwin instability occurs ($n_{\rm max}$), evaluated with $M_2=0.5M_{\odot}$.} \label{2M_010_big_companion_ZN.pdf} \end{figure} Figure \ref{2M_010_big_companion_ZN.pdf} again uses $M_1=2\, M_\odot$ and compares tracks with companion masses $M_2=0.001, 0.01$ and $0.5\,M_{\odot}$ for $a_{\rm init}=0.1\, \rm AU$ with Zahn's prescription for viscosity. In the second panel, the z-shape in $n_{\rm max}$ near $\log\,g=3.5 \sim 3.4$ occurs between central hydrogen exhaustion and shell ignition, where the star first shrinks and then expands. Similar to the $M_1=1\, M_\odot$ case, the two low mass cases are not synchronized while the higher mass case is synchronized until the Darwin instability. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{2M_010_big_companion_GN.pdf} \caption{Same as Figure \ref{2M_010_big_companion_ZN.pdf} but with GN's viscosity prescription. In the lower panel, the green dotted line displays the orbital frequencies above which the Darwin instability occurs ($n_{\rm max}$), evaluated with $M_2=0.5M_{\odot}$.} \label{2M_010_big_companion_GN.pdf} \end{figure} Figure \ref{2M_010_big_companion_GN.pdf} is the same as Figure \ref{2M_010_big_companion_ZN.pdf}, but uses GN's viscosity prescription. For the $M_2=0.001M_{\odot}$ case, dynamical tides are more important than for Zahn's viscosity, and the merger occurs slightly later compared to Figure \ref{2M_010_big_companion_ZN.pdf}. For the $M_2=0.01M_{\odot}$ case, the spin is not synchronized and dynamical tides dominate, and so the merger is near that in Figure \ref{2M_010_big_companion_ZN.pdf}. For the $M_2=0.5M_{\odot}$ case, the spin synchronizes leading to a merger at a similar time as in Figure \ref{2M_010_big_companion_ZN.pdf}. \subsection{3 $M_{\odot}$ Model} The $M_1=3\, M_{\odot}$ case is qualitatively similar to that of $M_1=2\, M_{\odot}$. Tidal dissipation is suppressed during the MS, effectively turning on when the convective envelope thickens near $\log g \simeq 3.0$ on the SGB. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{3M_10MJ_ZN.pdf} \caption{Same as Figure \ref{1M_10MJ_ZN} but with $M_1=3\, M_\odot$, $M_2=0.01\, M_\odot$ and $a_{\rm init}=0.05$ (black lines), 0.15 (green lines) and 0.2 (cyan lines) AU. Zahn's prescription is used.} \label{3M_10MJ_ZN.pdf} \end{figure} The bottom panel of Figure \ref{3M_10MJ_ZN.pdf} shows that the dynamical tide is much bigger than equilibrium tide when it turns on, and immediately causes both the $a=0.05$ and $0.15\, \rm AU$ orbits to decay. Only the $a=0.2\, \rm AU$ orbit is sufficiently wide that the star has time to move up the RGB and the equilibrium tide can dominate. None of the cases have synchronous spin for this companion mass (middle panel). \begin{figure} \centering \includegraphics[width=0.5\textwidth]{3M_10MJ_GN.pdf} \caption{Same as Figure \ref{3M_10MJ_ZN.pdf} but with GN's viscosity prescription.} \label{3M_10MJ_GN.pdf} \end{figure} Figure \ref{3M_10MJ_GN.pdf} shows the same experiment as Figure \ref{3M_10MJ_ZN.pdf} but with GN's viscosity. The merger in the $a_{\rm init}=0.05$ and $0.15$ AU cases is due to the dynamical tide and agrees closely with Figure \ref{3M_10MJ_ZN.pdf}. The $a_{\rm init}=0.2$ AU case ends up dominated by the equilibrium tide, and is not synchronized, and hence merges much later than that in Figure \ref{3M_10MJ_ZN.pdf}. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{3M_10MJ_ZN_close.pdf} \caption{Close-up of the bottom panel of Figure \ref{3M_10MJ_ZN.pdf}, near where the convective envelope deepens and tidal friction increases dramatically.} \label{3M_10MJ_ZN_close.pdf} \end{figure} Figure \ref{3M_10MJ_ZN_close.pdf} shows a close-up of the dissipation rates in the bottom panel of Figure \ref{3M_10MJ_ZN.pdf}. For all three cases, the dynamical tide (dashed lines) dominates the equilibrium tide (solid lines) as soon as it turns on. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{3M_015_big_companion_ZN.pdf} \caption{Same as Figure \ref{1M_015_big_companion_ZN.pdf} but with $M_1=3\, M_\odot$, and $M_2=$ 0.001 (green), 0.01 (blue) and 0.5 (red), and starting separation $a_{\rm init}=0.15\, \rm AU$. In the lower panel, the green dotted line displays the orbital frequencies above which the Darwin instability occurs ($n_{\rm max}$), evaluated with $M_2=0.5M_{\odot}$.} \label{3M_015_big_companion_ZN.pdf} \end{figure} Figure \ref{3M_015_big_companion_ZN.pdf} again compares the evolution with Zahn's viscosity and including the dynamical tide for three different companion masses. All three cases merge over a small range of $\log g$ soon after the dynamical tide turns on. For the most massive $M_2=0.5\, M_\odot$ companion, in spite of the fact that the spin becomes synchronous, the binary is still short lived since the Darwin instability turns on at roughly the same time. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{3M_015_big_companion_GN.pdf} \caption{Same as Figure \ref{3M_015_big_companion_ZN.pdf} but with GN's viscosity prescription.} \label{3M_015_big_companion_GN.pdf} \end{figure} The result for GN's viscosity in Figure \ref{3M_015_big_companion_GN.pdf} is very similar to Figure \ref{3M_015_big_companion_ZN.pdf}. As the dynamical tide for the $3M_{\odot}$ case is strong for all the three companion masses, the different viscous dissipation reduction factors make almost no change in the orbital evolution. \section{Critical Semi-major Axis for Mergers} \label{sec:acrit} Examples of orbital decay were shown in Section \ref{sec:examples} to understand the importance of the turbulent viscosity prescription, the strength of dynamical versus equilibrium tide, and synchronous spin at each stage in a star's evolution. In this section, calculations of the ``critical radius", $a_{\rm crit}(t)$, which depends on the age of the system, are presented. The critical radius is defined as the {\it initial} separation out to which orbits would have decayed down to the surface of the primary by the age $t$. Few binaries are expected to be found with (present day) $a \la a_{\rm crit}(t)$ since those orbits should have decayed and the binary already merged, while binaries with $a \ga a_{\rm crit}$ are relatively unaffected by orbital decay. Hence if a lack of systems is observed for some range of semi-major axis, the plot of $a_{\rm crit}$ versus $\log g$ shows which range might be absent of binaries due to tides, and for which range the lack of systems must have some other explanation as tides become ineffective. Figures \ref{acrit_1Msun.pdf}, \ref{acrit_2Msun.pdf} and \ref{acrit_3Msun.pdf} show our calculations of $a_{\rm crit}$ versus $\log g$ for a primary mass of $M_1=1,2,3\, M_\odot$, respectively, with the open circles showing data for binary systems from the APOGEE survey. The different lines in each plot are for different secondary masses, $M_2$. Zahn's prescription for reduced viscosity will be used in all plots. This reduces to standard (un-reduced) viscosity when eddy turnover times are short. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{acrit_1Msunv2.pdf} \caption{The critical semi-major axis $a_{\rm crit}$ versus $\log\,g$ for $M_1=1M_{\odot}$ with three companion masses: $M_2 = 0.001M_{\odot}$ (black), $M_2 = 0.01M_{\odot}$ (green) and $M_2 = 0.1M_{\odot}$ (cyan). The black dots are the APOGEE data \citep{2016AJ....151...85T}, with $M_1$ between $0.5M_{\odot}$ and $1.5M_{\odot}$, and $M_2$ between $1M_{\rm Jup}$ - $100M_{\rm Jup}$. } \label{acrit_1Msun.pdf} \end{figure} Figure \ref{acrit_1Msun.pdf} compares calculations of $a_{\rm crit}$ versus $\log\,g$ for $M_1=1M_{\odot}$. These calculations are compared to APOGEE binaries \citep{2016AJ....151...85T} for SGB and RGB primary stars in the mass range $M_1=0.8$ - $1.5\, M_\odot$. All the observed systems have $a>a_{\rm crit}$, meaning that the orbital decay rate is small compared to the stellar evolution timescale. The lack of systems at $a<a_{\rm crit}$ may be interpreted as either the population of closer systems with $a < a_{\rm crit}$ have already merged due to orbital decay, or such close binary systems are rare or never formed in the first place. Note that for extremely small $a_{\rm init}$, companions of mass $M_2 = 0.001$ - $0.1\, M_{\odot}$ would suffer orbital decay before the end of the MS, and well before the RGB. This corresponds to the starting point of $a_{\rm init}$ lines in the bottom left of the figure. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{acrit_2Msunv2.pdf} \caption{Same as Figure \ref{acrit_1Msun.pdf} for $M_1=2M_{\odot}$ and $M_2 = 0.001$ (black), $0.01$ (green) and $0.5M_{\odot}$ (cyan). The black dots are the APOGEE data \citep{2016AJ....151...85T}, with $M_1=1.5$ - $2.5M_{\odot}$ and $M_2=1$ - $500M_{\rm Jup}$.} \label{acrit_2Msun.pdf} \end{figure} Figure \ref{acrit_2Msun.pdf} shows the $M_1=2M_{\odot}$ case. Due to the thin surface convection zone, and a central convection zone, dynamical tides are assumed ineffective on the MS. The surface convection zone deepens near $\log g \simeq 3.5$, at which point both dynamical and equilibrium tides increase dramatically. This causes rapid orbital decay over a range of small orbital separation. At $a_{\rm crit}=0.05$ AU, dynamical tides rapidly shrink the orbit, and even the high mass companions can't synchronize the orbit within the short orbital decay timescale. Therefore the $a_{\rm crit}$ lines for the three $M_2$ are close. At $a_{\rm crit} = 0.1$ AU for the $M_2 = 0.01M_{\odot}$ secondary, dynamical tides are still strong and the orbit shrinks at $\log\,g=3.5$. Slow orbital decay occurs both for the high mass companion, due to synchronous spin, and low mass companions due to the weak tidal force. All but one of the APOGEE systems have $a \geq a_{\rm crit}$, again showing that the $a \leq a_{\rm crit}$ systems, especially with $\log g \la 3.5$, may have already been destroyed. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{acrit_3Msunv2.pdf} \caption{Same as Figure \ref{acrit_1Msun.pdf} for $M_1=3M_{\odot}$ with three companion masses $M_2 = 0.001M_{\odot}$ (black), $M_2 = 0.01M_{\odot}$ (green) and $M_2 = 0.5M_{\odot}$ (cyan). The black dots are the APOGEE data \citep{2016AJ....151...85T}, with $M_1=2.5$ - $3.5M_{\odot}$ and $M_2=1$ - $500M_{\rm Jup}$.} \label{acrit_3Msun.pdf} \end{figure} Figure \ref{acrit_3Msun.pdf} shows the $M_1=3M_{\odot}$ case. Similar to the $2M_{\odot}$, $a_{\rm crit}=0.05$ AU case, dynamical tides are strong and synchronization does not occur. The system separation decreases quickly at $\log\,g=3.0$ for the range $M_2=0.001$ - $0.5M_{\odot}$. For the $M_2 = 0.5M_{\odot}$ secondary, the distance of 0.2 AU is far enough away for the high mass companion to synchronize the orbit, which makes the system survive longer compared with using a $M_2 = 0.01M_{\odot}$ or $M_2 = 0.001M_{\odot}$ secondary. The observed systems are again shown in black circles. In this case there are three observed binaries near the $a=a_{\rm crit}$ lines, that may be undergoing more rapid orbital decay. Wide orbits with with $a_{\rm obs}>0.8$ AU are not shown in Figure \ref{acrit_1Msun.pdf}, \ref{acrit_2Msun.pdf} and \ref{acrit_3Msun.pdf}. Analytic estimates for $a_{\rm crit}$ are given in Appendix \ref{sec:analytic_acrit} in Equations \ref{eq:aD_vs_R}, \ref{eq:acritstd}, \ref{eq:acritz}, \ref{eq:acritgn} and \ref{acrit_dyn_R}. Since $\dot{E}$ scales as a high power of $a$, $a_{\rm crit}$ is a very weak function of $M_2$ for equilibrium and dynamical tides with non-synchronized primary. The dependence on $R_1$ is nearly linear for the equilibrium tide, but is dominated by the SGB and base of the RGB for the dynamical tide. Next, the calculations of $a_{\rm crit}$ are compared to binaries containing an exoplanet and host star from \citet{2014PASP..126..827H} (downloaded from exoplanets.org). Figure \ref{acrit_krxir_1015Msun_exo.pdf} shows stars in the mass range $M_1=1-1.3\, M_{\odot}$ with planetary mass companions. These stars are in the MS or early SGB. \footnote{The reason why the upper limit is set at $1.3M_\odot$ is because stars with $M>1.3M_\odot$ have a convective core and thin convective envelope, hence both the dynamical and equilibrium tide dissipation rates are expected to be smaller by comparison, giving small $a_{\rm crit}$ during the MS stage.} Immediately apparent in Figure \ref{acrit_krxir_1015Msun_exo.pdf} is the strong dependence of $a_{\rm crit}$ on stellar mass over the gravity range $\log g = 4.0-4.5$ and mass range $M_1=1.0-1.5\, M_\odot$. The sharp decrease in $a_{\rm crit}$ at fixed $\log g$ for slightly higher $M_1$ is due to the dynamical tide shutting off for stars with convective cores. The weaker equilibrium tide dissipation rate leads to much smaller $a_{\rm crit}$ when convective cores are present. Stars with a convective core on the MS initially have small $a_{\rm crit}$ while the convective core is present, and then $a_{\rm crit}$ suddenly increases at the end of the MS when the core becomes radiative and the convective envelope deepens. In addition, for the case of $M_1=1.2M_{\odot}$ - $1.5M_{\odot}$, at nearly the same time that the central convective core ceases, $k_r\xi_r$ becomes close to 1 and the dynamical tides become effective. The reason of the increase in $k_r\xi_r$ is that the density near the star center increases and the inner turning point of the wave moves inward as composition gradients also increase during the evolution. This result agrees with \cite{2010MNRAS.404.1849B}. This strong dependence on stellar mass implies that care must be taken when comparing an observed orbital separation, $a$, with a calculated critical separation, $a_{\rm crit}$. Uncertainties in spectroscopic determination of stellar mass and radius may not be small enough to accurately decide which theory curve is appropriate. The error bars on primary star gravity in Figure \ref{acrit_krxir_1015Msun_exo.pdf} are $\Delta(\log g) \simeq 0.1-0.2$, and uncertainties in fitting mass are at the level $\Delta(M_1) \simeq 0.1-0.2\, M_\odot$. If, for example, all data points in the range $\log g = 4.2 - 4.5$ were compared with the $M_1=1.0\, M_\odot$ line, a large number of systems would have $a \la a_{\rm crit}$ since dynamical tide dissipation is present for that stellar mass. However, if the stellar mass was slightly higher, with $M_1=1.1-1.2\, M_\odot$, then most of the points have $a>a_{\rm crit}$. A detailed comparison of each system with the measured $M_1$ indeed shows that nearly all exoplanets plus host star have $a \ga a_{\rm crit}$. However, given the size of the error bars as compared to the rapid change in the theory curves this result should be taken with some caution. For non-transiting binaries detected by the radial velocity method, only the minimum mass $M_2 \sin(i)$ is measured, where $i$ is the binary inclination. If the primary is not synchronized, using $M_2\sin(i)$ as the secondary mass instead of $M_2$ will lead to smaller $a_{\rm crit}$, although this dependence is very weak (see Appendix \ref{sec:analytic_acrit}). This would tend to make points have larger $a/a_{\rm crit}$. Next, if the primary star is synchronized, $a_{\rm D}$ varies inversely with $M_2$, tending to make the $a_{\rm crit}$ larger, and would tend to make smaller $a/a_{\rm crit}$. In practice systems with $\sin(i) \ll 1$ are rare. For the APOGEE systems and the exoplanet host stars, our results are consistent with nearly all systems having $a \ga a_{\rm crit}$. A natural question is what the distribution of orbital separation would look like if a large number of systems did indeed have $a \ll a_{\rm crit}$. This would imply that a large number of binaries have orbital decay times short compared to their stellar evolution timescale, and are being observed in a short-lived phase just before merger. Moreover, the tail of the semi-major axis distribution at $a<a_{\rm crit}$ should be accompanied by a much larger number $\propto \dot{a}^{-1} \propto a^{9.5}$ at larger separation, where $\beta=9.5$ for dynamical tides. For example, if $\sim 10$ systems were found at $a/a_{\rm crit} = 0.5$, this should be accompanied by $\sim 10 \times 2^{9.5} \simeq 7000$ systems at $a \simeq a_{\rm crit}$. The large reservoir of systems with slow orbital decay would then feed a small tail of systems with rapid orbital decay at smaller separation. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{acrit_krxir_1015Msun_exo.pdf} \caption{Critical semi-major axis $a_{\rm crit}$ versus $\log\,g$ for companion mass $M_2=10^{-3}\, M_\odot$ and primary masses $M_1=1.0M_\odot$ (blue), $1.1M_\odot$ (black), $1.2M_\odot$ (magenta), $1.3M_\odot$ (yellow), $1.4M_\odot$ (orange), $1.5M_\odot$ (cyan). The exoplanet plus host star data from \citet{2014PASP..126..827H} is given by open circles with error bars for $\log\,g$ in blue. } \label{acrit_krxir_1015Msun_exo.pdf} \end{figure} \section{Conclusion} \label{sec:conclusion} Motivated by current and future surveys that find binaries with SGB or RGB primaries and stellar or substellar secondaries, MESA models for primary stars of mass $M_1=1, 2$ and $3\, M_\odot$ have been used to compute dynamical and equilibrium tidal dissipation rates. The resultant orbital decay rate was used to compare merger times for different primary and secondary masses and orbital separations, as well as different prescriptions for turbulent viscosity. The role of synchronization of the primary's spin and the Darwin instability have been taken into account. The dynamical tide dominates for close-in systems, with less evolved primaries. The equilibrium tide dominates for wider systems and more evolved primaries. The dividing line between the two depends on primary and secondary masses, as well as the prescription for reduced viscosity. The tidal evolution depends sensitively on the primary star's mass. For stars of mass $M_1 \ga 1.3\, M_\odot$, equilibrium and dynamical tidal friction is strongly suppressed on the MS, and turns on suddenly during the sub-giant branch phase as the convective envelope deepens. For close-in systems this may result in orbital decay that proceeds rapidly compared to the stellar evolutionary timescale. The reason why the equilibrium tide is suppressed at the MS stage is because the energy dissipation rate depends on the convective envelope mass and the eddy velocity. These two physical quantities are much larger in the RGB phase. In the MS phase, the dynamical tide is suppressed because the wave can't propagate inward, and the convective envelope for $M_1 \ga 1.3\, M_\odot$ is very thin. For $M_2=1$ - $10\, M_{\rm Jup}$, the low mass secondary cannot provide enough angular momentum to synchronize the spin. For synchronization to occur, $M_2$ must be on the order of $0.1M_\odot$. After it reaches the Darwin instability, the system merges soon after. Neither small ($M_2\lesssim 10 M_{\rm Jup}$) nor large ($M_2\gtrsim 100 M_{\rm Jup}$) companions can give rise to fast orbital decay. Small mass companions exert weak tidal forces, thereby causing low energy dissipation rates. High mass companions synchronize their orbits quickly, resulting in a small forcing frequency and consequently low energy dissipation rate. Only intermediate mass ($10\lesssim M_2 \lesssim 100 M_{\rm Jup}$) secondaries can sustain a large energy dissipation rate. By accounting for the Darwin instability, dynamical tide, and Zahn's prescription for the reduced viscosity for the equilibrium tide, we define a critical separation $a_{\rm crit}$ below which the system merges rapidly compared to the stellar evolution timescale. Figures \ref{acrit_1Msun.pdf}, \ref{acrit_2Msun.pdf} and \ref{acrit_3Msun.pdf} show that the majority of APOGEE binaries show separations larger than $a_{\rm crit}$ for the observed $\log g$, indicating that these systems likely have very slow orbital decay. A small number of systems have observed separations $a<a_{\rm crit}$, implying rapid orbital decay. To compare to exoplanets and their host stars from the compilation in \citet{2014PASP..126..827H}, $a_{\rm crit}$ was computed over a finer grid of primary mass $M_1=1.0-1.5\, M_\odot$. The size of the convective core during the MS changes rapidly over this stellar mass range. As a result, the dynamical tide may turn on suddenly when the convective core disappears, and $a_{\rm crit}$ rapidly moves outward. Figure \ref{acrit_krxir_1015Msun_exo.pdf} compares our calculations to the exoplanet data. In a detailed comparison of the measured mass to the appropriate theory curve, we find that only a few systems have $a \la a_{\rm crit}$. However, this result must be taken with caution given the uncertainties in measuring primary gravity and mass. The comparison between the calculations of $a_{\rm crit}$ and semi-major axis data in this paper is an extension of \citet{2017MNRAS.470.2054C}, who studied hot Jupiters undergoing orbital decay due to the dynamical and equilibrium tides. \citet{2017MNRAS.470.2054C} computed orbital decay rates for five systems with detections or upper limits. These authors also found that the dynamical tide dominated for short-period orbits, and that radiative diffusion damping is important during the post main sequence phase. The present paper is focused on systems with no measured orbital decay rate, for which the only comparison is with the distribution of semi-major axis. Our study extends \citet{2017MNRAS.470.2054C} by carrying out a parameter study for a wide range of stellar mass, evolutionary state of the star, and companion mass. Our study also includes a criterion for the traveling wave limit to apply due to either strong radiative diffusion damping or nonlinear wave breaking when $k_r \xi_r \ga 1$ at the center. \citet{2013ApJ...772..143S} discussed observational evidence that SGBs with planetary-mass companions show a strong deficit of systems with $a \la 0.67\, \rm AU$ ($P_{\rm orb}=200$ days). Further, the closest systems at $0.67\, \rm AU$ had fairly circular orbits. They proposed a scenario to explain this with tides, in which tidal friction is weak on the MS but increases dramatically on the sub-giant branch. Their scenario requires $R_1 \sim 3$ - $4\, R_\odot$ primary stars to cause orbital decay out to $0.67\, \rm AU$, and circularization of the orbits just outside this. Our results show that, with the tidal friction mechanisms included here, that orbital decay rate can only affect systems out to $a=0.05$ - $0.15\, \rm AU$ for this range of radii (and depending on the stellar mass), and that tidal friction is many orders of magnitude too weak to affect systems at $a=0.67\, \rm AU$. \section*{Acknowledgements} MS thanks Chenliang Huang, Dom Pesce and Scott Suriano for the very useful suggestions in improving the paper. The authors thank the referee for comments which significantly improved the paper. This work was supported by NASA ATP grant NNX14AB40G.

2e448c07c712bc7a82c21f400e41f10c298f3568

\section{Introduction} The high-luminosity phase of the Large Hadron Collider (HL-LHC) is foreseen to start in 2026 and will deliver an integrated luminosity of up to 4000\,fb$^{-1}$ in about 10 years. The HL-LHC will operate at an instantaneous luminosity up to $\mathcal{L}=7.5\times10^{34}$\,cm$^{-2}$s$^{-1}$ which corresponds to an average $<\mu>=200$ inelastic proton-proton collisions per bunch-crossing (pile-up). The HL-LHC will offer us the opportunities to precisely measure the Higgs coupling using as many Higgs production and decay channels as possible. These precision measurements will provide constraints on potential non-Standard Model physics. With the 4000\,fb$^{-1}$ of data, it allows the exploration of Higgs potential by studying Higgs-boson pair production. The high luminosity will also extend the reach of many new physics searches. The HL-LHC will present an extremely challenging environment to the ATLAS experiment, well beyond that for which it was designed. An upgrade of the ATLAS detector~\cite{ATLAS-detector} will be needed before the start of this new phase to cope with the high-radiation environment and the large increase in pile-up. \section{Upgrades of ATLAS detector for HL-LHC} To reach the physics goals at HL-LHC, maintaining or improving the performance of the ATLAS detector is essential. The ATLAS detector will be upgraded to cope with the increased occupancies and data rates. The inner tracking detector will be completely replaced with a new all-silicon tracker. The Transition Radiation Tracker (TRT) will not be used any more. For the Muon detector, new inner barrel chambers will be installed for trigger. The Calorimeters themselves are expected to operate reliably during the HL-LHC period, but the readout system of all calorimeters will be upgraded. Actually all readout has to be upgraded because the trigger rate will be increased by an order of magnitude. In addition, the high-granularity timing detector has been proposed and approved. A possible high-$\eta$ muon tagger is under consideration. This proceeding focuses on few aspects. More details are in Technical Design Reports (TDRs). TDRs for Pixel~\cite{pixel-TDR}, Strip~\cite{strip-TDR}, Muon~\cite{Muon-TDR}, Liquid Argon Calorimeter~\cite{LAr-TDR}, Tile Calorimeter~\cite{Tile-TDR} and TDAQ~\cite{TDAQ-TDR} upgrade have been approved at ATLAS experiment. The HGTD has produced a technical proposal~\cite{HGTD-TP} and the TDR is planned for early 2019. \subsection{New Inner Tracker} An all-silicon Inner Tracker (ITk) has been designed to completely replace the current Inner Detector (ID). As shown in Fig.~\ref{fig:ITk-layout}, the ITk comprises two subsystems: a Strip Detector surrounding a Pixel Detector. The Strip Detector has four barrel layers and six end-cap petal-design disks on each side, both having double modules each with a small stereo angle to add $z$ $(R)$ resolution in the barrel (end-caps). The Strip Detector, covering $|\eta|<2.7$, is complemented by a five layer Pixel Detector extending the coverage to $|\eta|<4$. Radiation-hard sensors and readout chips that can withstand the particle fluence have been designed. High-granularity pixels are necessary to reduce the occupancy and handle the high pile-up environment. The inclined sensors in barrel reduce the amount of silicon needed by maximizing angular coverage per sensor. End-cap modules are supported in rings instead of traditional disks to simplify mechanics and services routing while optimizing the distribution of hits along high rapidity tracks. The silicon surface is about 200\,m$^2$ with $\sim$5\,G pixels and $\sim$80\,M strips. \begin{figure}[h] \centering \includegraphics[width=0.7\textwidth]{ITk-layout.png} \caption{A schematic layout of the ITk Inclined Duals layout for the HL-LHC phase of ATLAS~\cite{pixel-TDR}. Here only one quadrant and only active detector elements are shown. The horizontal axis is the axis along the beam line with zero being the interaction point. The vertical axis is the radius measured from the interaction region. The outer radius is set by the inner radius of the barrel cryostat that houses the solenoid and the electromagnetic calorimeter.} \label{fig:ITk-layout} \end{figure} \subsection{Muon Spectrometer} The upgrades to the Muon Spectrometer will be done in two phases. All the upgrades for phase-I and II are shown in Fig.~\ref{fig:muon-upgrade}. In the Phase-I upgrade starting in 2019, the present small wheels of the end-cap layers (between toroid and end-cap Calorimeter) will be completely replaced. New Small Wheel (NSW) are required to maintain low $p_{\rm{T}}$ lepton triggers in high-rate environment after Run 2. It could reject up to 90\% of fake triggers at high rates. In the NSW region, the Micromegas (MM) and small-strip Thin Gap Chambers (sTGC) will take the dual role of trigger and precision chambers. Upgrades will also be done to the inner barrel resistive plate chambers (RPC). The phase-II upgrade will be mainly about trigger. New inner RPC stations will be installed. And the monitored drift tubes (MDTs) information will be added at the hardware trigger. In addition, a high-$\eta$ tagger is under investigation. \begin{figure}[h] \centering \includegraphics[width=0.7\textwidth]{Muon-upgrades.png} \caption{$R$-$z$ view of the Phase-II ATLAS muon spectrometer layout~\cite{Muon-TDR}. The drawings show the new detectors to be added in the Phase-II upgrade (red text), those to be installed during Phase-I upgrade (green text), and those that will remain unchanged from the Run 1 layout (black text). } \label{fig:muon-upgrade} \end{figure} \subsection{High-Granularity Timing Detector} The pile-up will be one of the main challenges of the HL-LHC, the interaction region will spread over about 50\,mm (RMS) along the beam axis and produce an average of 1.6 collisions/mm for $<\mu>=200$. This is to be compared with the 0.24 collisions/mm in Run 2. The precise assignment of tracks to hard-scatter vertices is crucial to mitigate the pile-up effects. This depends on the space separation of vertices in the beam direction. It will be very challenging at HL-LHC because of the high density of pile-up. The longitudinal impact parameter ($z_0$) resolution in the forward region is limited by multiple scattering. A powerful new way to address this challenge is to exploit the time spread of the interactions to distinguish between collisions occurring very close in space but well separated in time. A High-Granularity Timing Detector (HGTD), based on low gain avalanche detector technology, is proposed for the ATLAS Phase-II upgrade. As shown in Fig.~\ref{fig:HGTD}, the HGTD will be located just outside of ITk, covering the pseudorapidity region between 2.4 and 4.0. In order to ensure a timing resolution of 30\,ps per track over the whole HL-LHC period, four layers of active sensors will be built. At the smallest radius the occupancy is 10\% for a pixel size of $1.3\times1.3$\,mm$^2$. \begin{figure}[h] \centering \includegraphics[width=0.7\textwidth]{HGTD.png} \caption{The HGTD installed outside of ITk~\cite{HGTD-TP}. } \label{fig:HGTD} \end{figure} \subsection{Trigger and DAQ Upgrade} The high instantaneous luminosity means higher data rates. It poses significant challenges to the Trigger and to the DAQ system to fully exploit the physics potential of the HL-LHC. The baseline architecture is based on a single-level hardware trigger plus event filter. The trigger rate will be 1\,MHz instead of 100\,kHz of Run 2. This will be a big challenge for the detector readout. Up to 10\,kHz event data are sent into storage. This is ten times of the Run 2 output data rate. To cope with the increased occupancies and high data rates, the readout electronics for all systems will be upgraded. \section{Expected performance of physics objects} With the upgrades to the ATLAS detector and the new designed TDAQ system, the physics objects are expected to have a comparable or better performance than Run 2. The NSW phase-I upgrade and the phase-II upgrades of Muon Spectrometer are not included in the simulation. The new ITK has been included. \subsection{Tracking Performance} The track reconstruction efficiency in the ITk is found to be stable to better than 1\% for $t\bar{t}$ events with 40 to 250 pile-up, as shown on the left side of Fig.~\ref{fig:tracking} for different regions in $\eta$. For a fixed tracking efficiency, it is important to control the rate of fake or mis-reconstructed tracks. An inclusive measure for the rate of such fake or mis-reconstructed tracks is the ratio of the number of reconstructed tracks to the number of generated true particles. Because of the limited momentum resolution, the dominant contribution to the additional reconstructed tracks in the forward region is due to mis-measured low-$p_{\rm{T}}$ tracks that have a reconstructed transverse momentum above 1\,GeV. As can be seen on the right side of Fig.~\ref{fig:tracking}, the ratio does not vary over the full range of pile-up studied between 40 and 250 by more than 1\% for all $\eta$ regions. \begin{figure}[h] \centering \includegraphics[width=0.46\textwidth]{efficiency_vs_mu.png} \hspace{0.5cm} \includegraphics[width=0.46\textwidth]{fakeRate_vs_mu.png} \caption{Left: Tracking efficiency~\cite{pixel-TDR}. Right: Ratio of reconstructed tracks to generated charged particles~\cite{pixel-TDR}. In both cases, the results are shown as a function of pile-up for $t\bar{t}$ events for pile-up $\mu$ between 40 and 250, for different regions in $\mu$.} \label{fig:tracking} \end{figure} Better track parameter resolution is expected with ITk due to the smaller pixels. The resolution of momentum is shown in the left plot of Fig.~\ref{fig:trkRes} as an illustration. The momentum resolution is nearly a factor 2 better in the ITk than in the Run 2 ID, primarily due to the higher precision of the strip tracker compared to the TRT at high $p_{\rm{T}}$, and due to the reduced material at low $p_{\rm{T}}$. It degrades towards the forward region, mostly due to the fact that the solenoid field in this region is weaker and the transverse path length of the tracks gets shorter. Better momentum resolution leads to better mass resolution as shown in the right plot of Fig.~\ref{fig:trkRes}. The width of the $HH\rightarrow ZZ\rightarrow 4\mu$ and $H\rightarrow \mu\mu$ invariant mass signal have been studied. The resolution is visibly improved with ITk. The resolution of transverse impact parameter ($d_0$) of ITk is expected to be comparable with Run 2, while the resolution of $z_0$ will be significantly better than Run 2 benefitting from the smaller pixel pitch in the $z$ direction. \begin{figure}[h] \centering \includegraphics[width=0.46\textwidth]{pT.png} \hspace{0.5cm} \includegraphics[width=0.46\textwidth]{width-dimu-4l.png} \caption{(a) Track parameter resolution of $p_{\rm{T}}$ as a function of $\eta$~\cite{pixel-TDR}. (b) The width of the $H\rightarrow ZZ\rightarrow 4\mu$ and $H\rightarrow \mu\mu$ invariant mass signal as a function of the $|\eta|$ of the muon with the largest $|\eta|$ value~\cite{pixel-TDR}.} \label{fig:trkRes} \end{figure} The resolutions of these parameters directly determine the performance of the detector in terms of its capability for $b$-tagging and lepton or jet reconstruction. \subsection{$b$-tagging performance} The $b$-tagging algorithms used in ATLAS are based on multivariate techniques combing the impact parameters of associated tracks and the properties of reconstructed secondary vertex. The $b$-tagging performance is characterized by the probability to identify jets containing a $b$-hadron decay ($b$-jet efficiency) and by the rejection of jets not containing a $b$- or $c$-hadron as a $b$-jet (light-jet rejection). The $b$-tagging algorithm called MV2 has been fully re-optimized for the new ITk layout. Fig.~\ref{fig:btaggin-ITk-vs-run2} shows the light-jet rejection versus the $b$-tagging efficiency obtained by $t\bar{t}$ events with an average of 200 pile-up. For comparison purposes, the performance of the Run 2 detector for $|\eta|<2.5$ for events with an average of $\sim$30 pile-up events is also shown. The ITk performs better than the current ID, even at pile-up levels of HL-LHC. Even in the very forward region, the ITk provides significant discrimination power between $b$- and light-jets. \begin{figure}[h] \centering \includegraphics[width=0.46\textwidth]{rejection-vs-btagging-run2VSITk.png} \caption{Performance of the MV2 $b$-tagging algorithms in $t\bar{t}$ events with 200 pile-up for the ITk layout~\cite{pixel-TDR}. The rejection of light jets for different $\eta$ regions is shown as a function of $b$-jet efficiency. For comparison purposes, the performance for ATLAS during Run 2 with an average of 30 pile-up events is shown as crosses.} \label{fig:btaggin-ITk-vs-run2} \end{figure} The $b$-tagging is particulary sensitive to the contamination of pile-up tracks. It considers tracks with large impact parameters. So tracks from nearby pile-up are likely to be selected. It is essential to mitigate effects from pile-up. \subsection{Pile-up Jets Suppression} Pile-up jets are tagged with the discriminant $R_{p_{\rm{T}}}$. It is defined as the scalar $p_{\rm{T}}$ sum of the tracks that are associated with a jet and originate from the hard-scatter vertex PV$_0$, divided by the fully calibrated jet $p_{\rm{T}}$, i.e. \vspace{-0.5cm} \begin{center} \begin{equation} \label{eq:RpT} R_{\rm{p_{T}}}=\frac{\sum_k p_{\rm{T}}^{\rm{trk_k}}(\rm{PV}_0)}{p_{\rm{T}}^{\rm{jet}}} \end{equation} \end{center} Pile-up jets typically have a small value of $R_{\rm{p_{\rm{T}}}}$ corresponding to jets with small charged particle $p_{\rm{T}}$ fraction originating from the hard scatter vertex PV$_0$, as shown in the left plot of Fig.~\ref{fig:rejection-pileup-vs-eff-HS}. The right plot of Fig.~\ref{fig:rejection-pileup-vs-eff-HS} shows the rejection (defined as the inverse of the mis-tag efficiency) of pile-up jets as a function of the efficiency for selecting hard-scatter jets using the $R_{p_{\rm{T}}}$ discriminant for jets with low and high $p_{\rm{T}}$ in $t\bar{t}$ events with $<\mu>$ = 200 with and without the HGTD, for different time resolution ($\sigma(t)$) values. The extended coverage of ITk makes the track-based pile-up suppression possible in the forward region. A significant improvement in performance of up to a factor of 4 higher pile-up jet rejection at constant efficiency is achieved with the use of timing information. \begin{figure}[h] \centering \includegraphics[width=0.46\textwidth]{R-pT.png} \hspace{0.5cm} \includegraphics[width=0.46\textwidth]{rejection-pileup-vs-eff-HS-lowpT.png} \caption{(a) $R_{\rm{p_{\rm{T}}}}$ distribution for hard-scatter and pile-up jets with $40<p_{\rm{T}}<50$\,GeV in different $\eta$ regions~\cite{phaseII-scoping}. (b) Pile-up jet rejection as a function of hard-scatter jet efficiency in the $2.4<|\eta|<4.0$ region, for the ITk-only and ITk + HGTD scenarios with different time resolutions~\cite{HGTD-TP}.} \label{fig:rejection-pileup-vs-eff-HS} \end{figure} \section{Bench mark analyses to evaluate the physics performance at HL-LHC} The effects of an upgraded ATLAS detector are taken into account by applying energy smearing, object efficiencies and fake rates to truth level quantities, and following parameterizations based on detector performance studies with full simulation and HL-LHC conditions. An integrated luminosity of 3000\,fb$^{-1}$ and an average pile-up of $<\mu> = 200$ are assumed. \subsection{Higgs-boson pair production} Measurement of the Higgs-boson pair production is a major goal of the HL-LHC. The small cross section of $\sigma(pp\rightarrow HH)\sim 40$\,fb at $\sqrt{s}=14$\,TeV motivates the use of the dominant gluon fusion production. The channels in which at least one Higgs decaying to $b\bar{b}$ are promising. Thus the high-performance $b$-tagging capability is of critical importance for these analyses. The SM non-resonant $HH$ production process is dominated by gluon-gluon fusion, leading to centrally produced Higgs bosons, hence the extended forward tracking capability of the ITk is not expected to lead to large improvements in sensitivity. The sensitivity of $HH\rightarrow 4b$ channel is estimated as an illustration. The cross section of this channel is predicted to be 13.2\,fb in the SM. This channel has a high sensitivity to $b$-jet trigger threshold. If there is no upgrade to the trigger system, the $p_{\rm{T}}$ threshold for the $b$-jets would be 100\,GeV. The left plot of Fig.~\ref{fig:HH-4b} shows the 95\% C.L. upper limit on the cross section as a function of the jet $p_{\rm{T}}$ threshold without considering the systematic uncertainty. Relative to 100\,GeV, the threshold of 65\,GeV which could be reached with the trigger upgrade, will improve the sensitivity by a factor of 2. So the Trigger system upgrade is critical. In the right plot, the upper limit on the Higgs trilinear self-coupling $\lambda_{HHH}$ has been set which is also dependent on the trigger threshold. Based on these results, it is not possible to get a good constraint on SM. More channels are need to be combined to get enough statistics. The combination of the different sub-channels may provide a confirmation of double Higgs production, and perhaps shed some light on the size of $\lambda_{HHH}$. \begin{figure}[h] \centering \includegraphics[width=0.46\textwidth]{cs-limit-4b.png} \hspace{0.5cm} \includegraphics[width=0.46\textwidth]{par-limit-4b.png} \caption{Left: Non-resonant $HH\rightarrow 4b$ $\sigma/\sigma_{SM}$ 95\% exclusion sensitivity as a function of the minimum offline jet $p_{\rm{T}}$~\cite{TDAQ-TDR}. Right: Allowed intervals for the $\lambda_{HHH}$ parameter assuming the standard model as function of the minimum offline jet $p_{\rm{T}}$~\cite{TDAQ-TDR}.} \label{fig:HH-4b} \end{figure} \subsection{$t\bar{t}$ resonance search} BSM theories predict new particles with masses at the TeV scale. Specifically a Leptophobic $Z^{'}$ arising from topcolour-assisted technicolour is considered which decays primarily to a pair of top quarks. Thus a $t\bar{t}$ resonance search is a benchmark analysis for evaluating BSM physics prospects at the HL-LHC. The most recent ATLAS search~\cite{Run2:ttbar}, using 36.1\,fb$^{-1}$ of data taken at $\sqrt{s}=13$\,TeV, excludes such $Z^{'}$ bosons for masses less than 3.2\,TeV. At higher masses of the new heavy particles, the produced top quarks will be highly boosted. The top candidates will tend to produce $b$-jets with $p_{\rm{T}} > 600$\,GeV leading to a very dense tracking environment. Thus to be sensitive in this mass region, it is critical to maintain stable track reconstruction efficiency with increasing $p_{\rm{T}}$. Fig.~\ref{fig-Zp} (a) shows the track reconstruction efficiency as a function of jet $p_{\rm{T}}$ for tracks in jets for which $Z^{'}$(5\,TeV)$\rightarrow t\bar{t}$ events with an average of 200 pile-up events have been used. While the current detector shows a significant loss of efficiency for high-$p_{\rm{T}}$ jets is this boosted topology, a similar degradation is not observed for the ITk, thanks to the improved sensor granularity and the enlarged lever arm of the 5 pixel layers. A topcolour model containing a spin-1 $Z^{'}$ boson is used as the signal assuming a signal width of 1.2\%. Samples at $m_{Z^{'}}=1-7$\,TeV were generated. Leading order cross sections are used for each signal sample, multiplied by a factor of 1.3 to account for NLO effects. The expected upper limits set on the signal cross section $\times$ branching ratio as a function of the signal mass are shown in the right plot of Fig.~\ref{fig-Zp}. If there is no signals observed, the mass reach of the search is estimated to be 4\,TeV for the leptophobic topcolour-assisted technicolour $Z^{'}$~\cite{ttbar-HL}. The mass reach of $Z^{'}$ is estimated to be 3\,TeV with 300\,fb$^{-1}$. Apparently these estimations are too conservative. The limit on the mass reach is expected to be significantly improved. \begin{figure}[h] \centering \includegraphics[width=0.46\textwidth]{TrackEff_vs_jetPt.png} \hspace{0.5cm} \includegraphics[width=0.46\textwidth]{uplimit-Zp.png} \caption{Left: Reconstruction efficiency as a function of jet $p_{\rm{T}}$, for tracks in jets from $Z^{'}$(5\,TeV)$\rightarrow t\bar{t}$ events with an average of 200 pile-up events~\cite{pixel-TDR}. Right: The expected upper limit on the cross section $\times$ branching ratio of a leptophobic topcolour-assisted technicolour $Z^{'}$ boson for masses in the range $1-7$\,TeV, assuming an integrated luminosity of 3000\,fb$^{-1}$ at $\sqrt{s}=14$\,TeV~\cite{strip-TDR}.} \label{fig-Zp} \end{figure} \section{Conclusions} The ATLAS detector upgrade for HL-LHC manages to maintain or improve the performance in very dense environment with pile-up up to 200. The all-silicon ITk with extended coverage will improve the tracking performance. As a complement of ITk, the new designed timing detector plays important role in the mitigation of pile-up effects. The upgrade to the trigger system is critical to keep lower trigger threshold. With the efforts we have made and the potential improvements in future, the performance of the physics objects reconstruction is expected to be better than the current detector. \begin{flushleft}

c57f26b88547dd60cb5b28ad25f866cd8e347e88

\section{Introduction} In this paper, we study the problem of the following non-convex composition minimization {\small \begin{align}\label{VRNonCS-SCSG:ProblemMainComposition} \mathop {\min }\limits_{x \in {\mathbb{R}^N}} \left\{ {f( x )\mathop = \limits^{{\rm{def}}} F( {G( x )} ) \mathop = \limits^{{\rm{def}}} \frac{1}{n}\sum\limits_{i = 1}^n {{F_i}\left( {\frac{1}{n}\sum\limits_{j = 1}^n {{G_j}( x )} } \right)} } \right\}, \end{align}} where $f$: $\mathbb{R}^N$ $\to$ $\mathbb{R}$ is a non-convex function, each $F_i$: $\mathbb{R}^M \to \mathbb{R}$ is a smooth function, each $G_i$: $\mathbb{R}^N$$\to$ $\mathbb{R}^M$ is a mapping function, $n$ is the number of $F_i$'s and $G_j$'s. We call $G(x)$:$=\frac{1}{n}\sum\nolimits_{j=1}^nG_j(x)$ the inner function, and $F( {w} )$:$ = \frac{1}{n}\sum\nolimits_{i = 1}^n {{F_i}( { w} )} $ the outer function. There are many machine learning application such as such as reinforcement learning \cite{sutton1998reinforcement,wang2017stochastic,wang2016accelerating} and nonlinear embedding \cite{hinton2003stochastic,dikmen2015learning}, that can be formed to the composition problem with two finite-sum structure $\frac{1}{n}\sum\nolimits_{i = 1}^n {{F_i}( {\frac{1}{n}\sum\nolimits_{j = 1}^n {{G_j}( x )} } )} $. For example, { \begin{align*} \mathop {\min }\limits_x {\| {\mathbb{E}[ B ]x - \mathbb{E}[ b ]} \|^2}, \end{align*} where $\mathbb{E}[ B ] = I - \gamma {P^{\pi} }$, $\gamma\in(0,1)$ is a discount factor, $P^{\pi}$ is the transition probability, $\mathbb{E}[ b ] = {r^{\pi} }$, and $r^{\pi}$ is the expected state transition reward.} Another example is the mean-variance in risk-averse learning: \begin{align*} \mathrm{min}_x\, \mathbb{E}_{a,b}[ h( {x;a,b} ) ] + \lambda \mathrm{Var}_{{a,b}}[ {h( {x;a,b} )} ], \end{align*} where $h(x;a,b)$ is the loss function with random variables $a$ and $b$. $\lambda>0$ is a regularization parameter. Stochastic neighbour embedding (SNE) \cite{hinton2003stochastic} is the non-convex problem that map data from a high dimensional space to a low dimensional space. \begin{align* \mathop {\min }\nolimits_x \sum_t^{} {\sum_i^{} {{p_{i|t}}\log \frac{{{p_{i|t}}}}{{{q_{i|t}}}}} } , \end{align*} where \begin{align*} {p_{i|t}} = \frac{{\exp ( - {{\left\| {{z_t} - {z_i}} \right\|}^2}/2\sigma _i^2)}}{{\sum\nolimits_{j \ne t} {\exp ( - {{\left\| {{z_t} - {z_j}} \right\|}^2}/2\sigma _i^2)} }}, \quad {q_{i|t}} = \frac{{\exp ( - {{\left\| {{x_t} - {x_i}} \right\|}^2})}}{{\sum\nolimits_{j \ne t} {\exp ( - {{\left\| {{x_t} - {x_j}} \right\|}^2})} }}, \end{align*} and $\sigma_i$ is {the predefined parameter to control the sensitivity to the distance.} $\{z_i\}_{i=1}^n$ and $\{x_i\}_{i=1}^n$ denote the representation of $n$ data points in the high dimensional space and the low dimensional space, respectively. Recently, there many stochastic optimization methods solving the composition problem, such as stochastic gradient method \cite{wang2017stochastic,wang2016accelerating} and the variance-reduction based method \cite{lian2016finite,liu2017duality,liu2017variance}. However, there are two main problems encountered in the composition function: 1) the inner function $G(x)$ is the finite-sum structure. When the number of $G_i(x)$ is large, it will need more computation cost; 2) if the inner function $G(x)$ is estimated, the expectation of the stochastic gradient $f(x)$ with respect to $i_k,j_k\in [n]$ is not equal to the $\nabla f(x)$. That is \begin{center} $\mathbb{E}_{i_k,j_k}[ ( \partial {G_{j_k}} (x))^\mathsf{T}\partial {F_{i_k}}( \tilde G(x)) ] \ne \nabla f( x )$, \end{center} where $\tilde G(x)) $is the estimation of $G(x)$, $\partial {G_{j_k}}$ is the partial gradient of $G_{j_k}(x)$. Furthermore, we use the query complexity to evaluate the algorithm, that is the number of {component function} queries used to compute the gradient. Stochastic gradient method, such as Stochastic compositional gradient descent (SCSG) \cite{wang2017stochastic} estimates the inner function $G(x)$ by an iterative weighted average of the past values of the $G(x)$, then perform the stochastic quasi-gradient iteration. The advantage of this method is that it does not depend on $n$ but with poor query complexity to the desired point. Variance-reduction method such as Compositional-SVRG \cite{lian2016finite} estimates the inner function $G(x)$ and the gradient of function $f(x)$ by using the finite-sum structures, which deriving the linear convergence rate with the relationship of $n$. Table \ref{VRNonCS-SCSG:Table} present the query complexity result with different algorithms. \begin{table*}[t] \centering \begin{tabular}{|l|c|c|} \hline {Algorithm} &{Strongly Convex} &Non-convex \\ \hline SCGD \cite{wang2017stochastic} &{$\mathcal{O}(1/\varepsilon^{3/2})$} &{$\mathcal{O}(1/\varepsilon^{4})$} \\ \hline Acc-SCGD \cite{wang2017stochastic} &{$\mathcal{O}(1/\varepsilon^{5/4})$} &{$\mathcal{O}(1/\varepsilon^{7/2})$} \\ \hline ASC-PG \cite{wang2016accelerating} &{$\mathcal{O}(1/\varepsilon^{5/4})$} &{$\mathcal{O}(1/\varepsilon^{9/4})$} \\ \hline SC-SVRG\cite{lian2016finite}\cite{liu2017variance} & ${\cal O}\left( {\left( {n + L_f^2/{\mu ^4}} \right)\log \left( {1/\varepsilon } \right)} \right).$ &{$\mathcal{O}( n^{4/5}/\varepsilon )$} \\ \hline SC-SCSG & ${\cal O}\left( {\left( {\min \left\{ {n,\frac{1}{{\varepsilon {\mu ^2}}}} \right\} + \frac{{L_f^2}}{{{\mu ^2}}}\min \left\{ {n,\frac{1}{{{\mu ^2}}}} \right\}} \right)\log \left( {1/\varepsilon } \right)} \right)$ &{$\mathcal{O}( \min \{ {{1}/{{{\varepsilon ^{9/5}}}},{{{n^{4/5}}}}/{\varepsilon }} \} )$} \\ \hline \end{tabular}% \vspace{5pt} \caption{Comparison of the query complexity with different algorithms } \label{VRNonCS-SCSG:Table}% \end{table*}% Motivated by the recent work \cite{lei2017less,lei2017non,AllenZhu2017-natasha2} that the convergence rate of the finite-sum structure function has the general result under the relationship between $n$ and $\varepsilon$. Here, we use $\varepsilon$ to evaluate the terminal of the convex and non-convex function by $f\left( x \right) - f\left( {{x^*}} \right) \le \varepsilon$ and ${\left\| {\nabla f\left( x \right)} \right\|^2} \le \varepsilon$, respectively, where $x^*$ is the optimal point in the convex function. The core aspect of these kinds of algorithms is similar to the stochastic variance-reduced gradient (SVRG) that using a snapshot vector to compute the ``gradient" of the function. The difference lies that the gradient is no longer computed directly but rather using the random subset, called stochastically controlled stochastic gradient (SCSG). We explore the SCSG based method to the composition problem with both convex and non-convex function and analyze the corresponding the convergence and query complexity. In this paper, we develop a novel stochastic composition optimization through stochastically controlled stochastic gradient (SC-SCSG) method to two finite-sum structure. The main contributions are summarized below: \begin{itemize} \item We provide the variance reduction based method to estimate the inner function $G(x)$. Similar to the SCSG that estimate the gradient, the function $G(x)$ can also be estimated by a snapshot $\tilde{x}_s$, in which $G(\tilde{x}_s)$ is not computed directly, but rather based on the random subset from $[n]$. We also analyze the size of the subset such that can lead to the desired precision for both convex and non-convex function. \item After obtaining the estimated inner function, we consider the gradient of the function $f(x)$. Here, we can also apply the SCSG based method to estimate the gradient. However, there are two situations encountered in the estimate process. 1) the expectation of the gradient is no longer the unbiased estimation. 2) the gradient of $f(\tilde{x}_s)$ at the snapshot is formed by two random subsets, which are used for the function $F_i$ and $G_j$ respectively. Nevertheless, we also provide the bound of the subset size that we can use the estimated gradient to update the iteration. The details analysis can be referred to Section \ref{VRNonCS-SCSG:Section:SCSG-convex}. \item The mini-batch version of the proposed algorithm is also provided for both the convex and non-convex function. The corresponding query complexities are improved based on the size of the mini-batch. More information can be referred to Section \ref{VRNonCS-SCSG:Section:SCSG-minibatch}. \end{itemize} \subsection{Results} We give the general query complexity of the composition problem based on SCSG based method. The results present us an intuitive explanation for comparing with other algorithms. Note that the Algorithm \ref{VRNonCS-SCSG:AlgorithmI} can be used to both convex and non-convex problems that deriving the corresponding query complexities. Furthermore, Algorithm \ref{VRNonCS-SCSG:AlgorithmII} present the mini-batch version of the proposed method. \textbf{{Convex function}} The query complexity for the convex function is \begin{center} $\mathcal{O}\left( {\left( {\min \left\{ {n,\frac{1}{{{\varepsilon\mu ^2}}}} \right\} + \frac{{L_f^2}}{{{\mu ^2}}}\min \left\{ {n,\frac{1}{{{\mu ^2}}}} \right\}} \right)\log \left( {1/\varepsilon } \right)} \right)$, \end{center} where $\mu$ is the constant of strongly convex of $f(x)$. The result is the same as that of \cite{lian2016finite} if $n\le 1/(\varepsilon\mu^2)$ \textbf{Non-convex function} The query complexity is {$\mathcal{O}( \min \{ {{1}/{{{\varepsilon ^{9/5}}}},{{{n^{4/5}}}}/{\varepsilon }} \} )$}, which can be better than that of \cite{wang2016accelerating} and comparable to that of \cite{liu2017variance}. \textbf{{Mini-batch}} For the mini-batch version, the query complexity can be improved to some extent comparing with above results, that is {$\mathcal{O}( \min \{ {{1}/{{{\varepsilon ^{9/5}}}},{{{n^{4/5}}}}/{\varepsilon }} \} /b^{1/5})$} and \begin{center} $\mathcal{O}\left( {\left( {\min \left\{ {n,\frac{1}{{{\varepsilon\mu ^2}}}} \right\} + \frac{{L_f^2}}{{{b\mu ^2}}}\min \left\{ {n,\frac{1}{{{\mu ^2}}}} \right\}} \right)\log \left( {1/\varepsilon } \right)} \right) $, \end{center} for convex and non-convex function. \subsection{Related work} As the data increase, stochastic optimization has been the popular method in machine learning and deep learning, especially for the finite-sum function. The typical algorithm include (stochastic gradient descent) SGD \cite{ghadimi2016accelerated}, stochastic variance reduction gradient (SVRG)\cite{johnson2013accelerating,reddi2016stochastic}, stochastic dual coordinate ascent (SDCA) \cite{shalev2014accelerated,shalev2013stochastic} and the accelerated method Nesterov's method \cite {nesterov2013introductory}, accelerated randomized proximal coordinate (APCG) \cite{lin2014accelerated,lin2014acceleratedSIAM} and Katyusha method \cite{allen2016katyusha}. As the function is finite-sum structure, the general process for optimization is randomly selected one or a block component function to estimate the gradient. Thus the estimated gradient leads to the large variance of the gradient. Variance reduction method estimates the gradient by using a snapshot in which the gradient of the function is computed at this point, which can appropriately reduce the variance. The composition function can also be solved by using above algorithms, however, two finite-sum structures prevent implementation directly due to the fact that the computation of the inner function may increase the query complexity. Recently, Wang \textit{et al.} \cite{wang2017stochastic} first proposed the first-order stochastic compositional gradient methods (SCGD) to solve such problems, which used two steps to alternately update the variable and inner function. The SCGD method has the query complexity $\mathcal{O}(\varepsilon^{-7/2})$ for the general function and $\mathcal{O}(\varepsilon^{-5/4})$ for the strongly convex function. Liu \textit{et al.} \cite{wang2016accelerating} employed Nesterov's method to accelerate the composition problem with $\mathcal{O}(\varepsilon^{-5/4})$ and $\mathcal{O}(\varepsilon^{-9/4})$ for strongly convex and non-convex function. However, these methods estimate the inner function by an iterative weighted average of the past function. Such estimation did not take advantage of the finite-sum structure. Based on the variance reduction technology, Lian \textit{et al.} \cite{lian2016finite} first applied the SVRG-based method to estimate the inner function $G(x)$ and the gradient of the function $f(x)$ as well. The linear convergence rate is obtained. In the following, Liu \textit{et. al} \cite{liu2017duality} apply the duality-free method to the composition problem and derive the linear convergence rate as well. Yu and Huang \cite{yu2017fast} applied the ADMM-based \cite{ boyd2011alternating} method and provide an analysis of the convex function without requiring Lipschitz smoothness. Moreover, Liu \textit{et. al} \cite{liu2017variance} considered the non-convex function and analyzed the query complexity with different sizes of the inner function and outer function. The details of the query complexity are provided. There are many recent papers considering the variance reduced method that estimates the gradient using the random subset rather than computing directly. Lei and Jordan \cite{lei2017less} proposed an SCSG method to the convex finite-sum function, and then applied to the non-convex problem in \cite{lei2017non} that using less than a single pass to compute the gradient at the snapshot point. In the following, Allen-Zhu \cite{AllenZhu2017-natasha2} also proposed Natasha1.5 algorithm, in which the gradient for each epoch is based on the random subset. Moreover, the objective function has the regularization term. Liu \textit{et. al} \cite{liu2018stochastic} applied the SCSG based method to the zeroth-order optimization with the finite-sum function. The rest of paper is organized as follows: In section \ref{VRNonCS-SCSG:Section:Preliminaries}, we give preliminaries used for analyzing the proposed algorithm. Section \ref{VRNonCS-SCSG:Section:SCSG} presents the SCSG-based method for the composition problem. we give the convergence and query complexity for the convex and non-convex function in Section \ref{VRNonCS-SCSG:Section:SCSG-convex} and Section \ref{VRNonCS-SCSG:Section:SCSG-Non-convex}, respectively. Section \ref{VRNonCS-SCSG:Section:SCSG-minibatch} gives the mini-batch version. We conclude our paper in Section \ref{VRNonCS-SCSG:Section:Conclusion}. \section{Preliminaries}\label{VRNonCS-SCSG:Section:Preliminaries} Throughout this paper, we use the Euclidean norm denoted by $\|\cdot\|$. We use $i \in [ n ]$ and $j \in [ m ]$ to denote that $i$ and $j$ are generated from $[ n ] = \{ {1,2,...,n} \}$, and $[ m ] = \{ {1,2,...,m} \}$. We denote by ${( {\partial G( x )} )^\mathsf{T}}\nabla F( {G( x )} )$ the full gradient of the function $f$, $\partial G( x )$ the partial gradient of $G$, and ${( {\partial G_{j_k}( x )} )^\mathsf{T}}\nabla F_{i_k}( {G( x )} )$ as the stochastic gradient of the function $f$, where $i_k$ and $j_k$ are randomly selected from $[n]$ and $[m]$. We use $\mathbb{E}$ to denote the expectation. Note that all the variable such as subset $\cal A$ and $\cal B$, element $i_k$ and $j_k$ are independently selected from $[n]$ or $[m]$, in particular, the element in $\cal A$ and $\cal B$ are independent. So we use $\mathbb{E}$ in instead of $\mathbb{E}_{i_k}$,$\mathbb{E}_{j_k}$ ,$\mathbb{E}_{\cal A}$ and $\mathbb{E}_{\cal B}$ except particular stated. We use $A=|\cal A|$ to denote the number of the elements in the set $\cal D$ and define ${G_\mathcal{A}(x)} = \frac{1}{A}\sum_{1 \le j \le A} {{G_{\mathcal{A}\left[ j \right]}(x)}} $. Recall two definitions on Lipschitz function and smooth function. \begin{definition} \label{VRNonCS-SCSG:DefinitionLipschitzFunction} A function $p$ is called a Lipschitz function on $ \mathcal{X}$ if there is a constant $B_p$ such that $ \| {p( x ) - p( y )} \| \le {B_p}\| {x - y} \|$, $\forall x,y\in \mathcal{X}$. \end{definition} \begin{definition} A function $p$ is called a $L_p$-smooth function on $ \mathcal{X}$ if there is a constant $L_p$ such that $\| {\nabla p( x ) - \nabla p( y )} \| \le {L_p}\| {x - y} \|$, and equal to $p( y ) \le p( x ) + \langle {\nabla p( x ),y - x} \rangle + {L_p}/2{\| {y - x} \|^2}$, $\forall x,y\in \mathcal{X}$. \end{definition} We make the following assumptions used for the discussion of the convergence rate and complexity analysis. \begin{assumption}\label{VRNonCS-SCSG:Assumption:f-strong} For function $f$: ${\mathbb{R}^M} \to {\mathbb{R}}$, all $i\in [n]$, \begin{itemize} \item $F$ is $\mu$-strongly convex satisfying $f\left( y \right) \ge f\left( x \right) + \langle f\left( x \right),y - x\rangle + \frac{\mu }{2}E{\left\| {x - y} \right\|^2}$. \item $f$ has the optimal point $x^*$, then $\langle f\left( {{x_k}} \right),{x^*} - {x_k}\rangle \le - \mu {\left\| {{x_k} - {x^*}} \right\|^2}$. \end{itemize} \end{assumption} \begin{assumption}\label{VRNonCS-SCSG:Assumption:G} For function $G_j$: ${\mathbb{R}^N} \to {\mathbb{R}^M}$, all $j\in [m]$, \begin{itemize} \item $G_j$ has the bounded Jacobian with a constant ${B_G}$, that is $\| {\partial G_j( x )} \| \le {B_G}$, $\forall x \in {\mathbb{R}^N}$, then $ G_j( x )$ is also a Lipschitz function that satisfying $ \| {G_j( x ) - G_j( y )} \| \le {B_G}\| {x - y} \|$, $\forall x,y \in {\mathbb{R}^N}$. \item $G_j$ is $L_G$-smooth satisfying $ \| {\partial G_j( x ) - \partial G_j( y )} \| \le {L_G}\| {x - y} \|$, $\forall x,y \in {\mathbb{R}^N}$. \end{itemize} \end{assumption} \begin{assumption}\label{VRNonCS-SCSG:Assumption:F} For function $F_i$: ${\mathbb{R}^M} \to {\mathbb{R}}$, all $i\in [n]$, \begin{itemize} \item $F_i$ has the bounded gradient with a constant ${B_F}$, that is $\|{\nabla F_i( y )} \| \le {B_F}$, $\forall y \in {\mathbb{R}^M}$. \item $F_i$ is $L_F$-smooth satisfying $ \| {\nabla F_i( x ) - \nabla F_i( y )} \| \le {L_F}\| {x - y} \|$, $\forall x,y \in {\mathbb{R}^M}$. \end{itemize} \end{assumption} \begin{assumption} \label{VRNonCS-SCSG:Assumption:GF} For function $F_i(G(x))$: ${\mathbb{R}^N} \to {\mathbb{R}}$, all $i\in [n]$, there exist a constant $L_f$ satisfying \begin{align* \| {{( {\partial {G_j}( x )} )^\mathsf{T}}\nabla {F_i}( {G( x )} ) - {( {\partial {G_j}( y )} )^\mathsf{T}}\nabla {F_i}( {G( y )} )} \|\nonumber \le {L_f}\| {x - y} \|,\forall j \in [m], \forall x,y \in {\mathbb{R}^N}. \end{align*} \end{assumption} \begin{assumption} \label{VRNonCS-SCSG:AssumptionIndependent} We assume that $i_k$ and $j_k$ are independently and randomly selected from $[n]$ and $[m]$, $z\in \mathbb{R} ^M, x\in \mathbb{R}^N$, \begin{align*} \mathbb{E}[ {{( {\partial {G_{j_k}}( x )} )^\mathsf{T}}\nabla {F_{i_k}}( z )} ] = {( {\partial G( x )} )^\mathsf{T}}\nabla F( z ), \end{align*} \end{assumption} \begin{assumption}\label{VRNonCS-SCSG:Assumption:Middle-Bound} We assume that $H_1$ and $H_2$ are the upper bounds on the variance of the functions $G(x)$ and $(\partial G(x))^\mathsf{T}\nabla F(y)$, respectively, that is, \begin{align*} \frac{1}{n}\sum\limits_{i = 1}^n& {{{\left\| {G(x) - {G_i}(x)} \right\|}^2}} \le {H_1}.\\ \frac{1}{{n^2}}\sum\limits_{j = 1}^{n} \sum\limits_{i = 1}^n& {{{\left\| {{(\partial G(x))^\mathsf{T}}\nabla F(y) - {(\partial {G_j}(x))^\mathsf{T}}\nabla {F_i}(y)} \right\|}^2}} \le {H_2}. \end{align*} } \end{assumption} In the paper, we denote by $x_k^s$ the $k$-th inner iteration at $s$-th epoch. But in each epoch analysis, we drop the superscript $s$ and denote by $x_k$ for $x_k^s$ . We let $x^*$ be the optimal solution of $f(x)$. Throughout the convergence analysis, we use $\mathcal{O}( \cdot )$ notation to avoid many constants, such as $B_F$, $B_G$, $L_F$, $L_G$ and $L_f$,... that are irrelevant with the convergence rate and provide insights to analyze the iteration and query complexity. \section{Stochastic Composition via SCSG for the composition problem} \label{VRNonCS-SCSG:Section:SCSG} \begin{algorithm*}[t] \caption{SC-SCSG for the composition problem} \label{VRNonCS-SCSG:AlgorithmI} \begin{algorithmic} \Require $K$, $S$, $\eta$ (learning rate), $\tilde{x}_0$ and $\mathcal{D} = \left[ {{\mathcal{D}_1},{\mathcal{D}_2}} \right]$ \For{$s =0,1, 2,\cdots,S-1$} \State Sample from $[n]$ for D times to form mini-batch $\mathcal{D}_1$ \State Sample from $[n]$ for D times to form mini-batch $\mathcal{D}_2$ \State $\nabla {\hat f_{ \mathcal{D}}}({{\tilde x}_s}) = {(\partial G_{ \mathcal{D}_1}({{\tilde x}_s}))^\mathsf{T}}{\nabla {F_{{ \mathcal{D}_2}}}(G_{ \mathcal{D}_1}({{\tilde x}_s}))} $\Comment{D Queries} \State $x_0=\tilde{x}_s$ \For{$k =0,1,2,\cdots,K-1$} \State Sample from $[m]$ to form mini-batch $\mathcal{A}$ \State ${{\hat G}_k} = {G_{{{\cal A}}}}({x_k}) - {G_{{{\cal A}}}}({{\tilde x}_s}) + {G_{\mathcal{D}_1}}({{\tilde x}_s})$\Comment{A Queries} \State Uniformly and randomly pick $i_k$ and $j_k$ from $[n]$ \State Compute the estimated gradient $\nabla {{\tilde f}_k}$ from (\ref{VRNonCS-SCSG:Definition:Estimate-f}) \Comment{4 Queries} \State ${x_{k+1}}= {x_k} - {\eta} \nabla {{\hat f}_k} $ \EndFor \State Update $\tilde{x}_{s+1}=x_K $ \EndFor \\ \textbf{Output:} $\hat x_k^s$ is uniformly and randomly chosen from $s\in\{0,...,S-1\}$ and $k\in \{0,..,K-1\}$. \end{algorithmic} \end{algorithm*} In this section, we present the variance-reduction based method for the composition problem, which can be used for both the convex and non-convex function. Before describing the proposed algorithm, we recall the original SVRG \cite{johnson2013accelerating}. The general process of the SVRG works as follows. The update process is divided into $S$ epochs, each of the epoch consists of K iterations. At the beginning of each epoch, SVRG define a snapshot vector $\tilde{x}_s$, and then compute the full gradient $\nabla f(\tilde{x}_s)$. In the inner iteration of the current epoch, SVRG defines the estimated gradient by randomly selecting $i_k$ from [n] at the $k$-th iteration, \begin{align}\label{VRNonCS-SCSG:Definition:Estimate-f-svrg} {(\partial G({x_k}))^\mathsf{T}}\nabla {F_{{i_k}}}({G_k}) - {(\partial G({{\tilde x}_s}))^\mathsf{T}}\nabla {F_{{i_k}}}(G({{\tilde x}_s})) + \nabla f({{\tilde x}_s}). \end{align} However, for the composition problem, there are also variance-reduction based methods in \cite{lian2016finite}, \cite{liu2017duality} and \cite{liu2017variance}. The difference with SVRG is that there is another estimated function for $G(x)$, as $G(x)$ is also the finite-sum structure. There methods defined the estimate function as \begin{align}\label{VRNonCS-SCSG:Definition:Estimate-G-svrg} {{\tilde G}_k} = {G_{{{\cal A}}}}({x_k}) - {G_{{{\cal A}}}}({{\tilde x}_s}) + {G}({{\tilde x}_s}), \end{align} where $\mathcal{A}$ is the mini-batch formed by randomly sampling from $[n]$. Whereas, as the number of the inner function $G_j$ and the outer function $F_i$ increase, it is not reasonable to compute the full gradient of $f(x)$ and the full function $G(x)$ directly for each epoch. Extended from the SCSG \cite{lei2017non}\cite{lei2017less} and Natasha1.5 \cite{AllenZhu2017-natasha2}, we present a new algorithm for the composition problem as shown in Algorithm \ref{VRNonCS-SCSG:AlgorithmI}. First of all, we introduce the two subset $\mathcal{D}_1$ and $\mathcal{D}_2$, which are independent with each other and formed by randomly selecting from $[n]$, respectively. We define $\mathcal{D} = \left[ {{\mathcal{D}_1},{\mathcal{D}_2}} \right]$ for a new variable. $\mathcal{D}_1$ is used for estimating the inner function. Based on the variance reduction technology, the estimated inner function at $k$-th iteration of $s$-th epoch is \begin{align}\label{VRNonCS-SCSG:Definition:Estimate-G} {{\hat G}_k} = {G_{{{\cal A}}}}({x_k}) - {G_{{{\cal A}}}}({{\tilde x}_s}) + {G_{\mathcal{D}_1}}({{\tilde x}_s}), \end{align} where the subset of $\mathcal{A}$ is the same as in (\ref{VRNonCS-SCSG:Definition:Estimate-G-svrg}). Note that $\cal A$ and $\cal D$ are independent with each other. The difference with (\ref{VRNonCS-SCSG:Definition:Estimate-G-svrg}) is computing the third them that is under the subset $\mathcal{D}_1$ rather than $[n]$ as in (\ref{VRNonCS-SCSG:Definition:Estimate-G-svrg}). Throughout the paper, we assume that $|\cal A|\le |\cal D|$. $\mathcal{D}_2$ is use to estimate the outer function $F$. The key distinguish with \cite{lei2017non, lei2017less,AllenZhu2017-natasha2} is the biased full gradient of $f(\tilde{x}_s)$. We define this estimated full gradient of $f(\tilde{x}_s)$ for each epoch as $\nabla {\hat f_{ \mathcal{D}}}({{\tilde x}_s}) = {(\partial G_{\mathcal{D}_1}({{\tilde x}_s}))^\mathsf{T}}\nabla {F_{{ \mathcal{D}_2}}}(G_{ \mathcal{D}_1}({\tilde x}_s))$. However, ${\mathbb{E}_{\mathcal{A},\mathcal{D}}}[ {\nabla {{\hat f}_{\cal D}}({{\tilde x}_s})} ] \ne \nabla f({{\tilde x}_s})$. Then, we estimate the gradient of the $f(x_k)$ by \begin{align}\label{VRNonCS-SCSG:Definition:Estimate-f} \nabla {{\tilde f}_k} = {(\partial {G_{{j_k}}}({x_k}))^\mathsf{T}}\nabla {F_{{i_k}}}({{\hat G}_k}) - {(\partial {G_{{j_k}}}({{\tilde x}_s}))^\mathsf{T}}\nabla {F_{{i_k}}}({G_{\mathcal{D}_1}}({{\tilde x}_s})) + \nabla {{\hat f}_\mathcal{D}}({{\tilde x}_s}), \end{align} where $i_k$ and $j_k$ are randomly selected from [n] at the $k$-th iteration for function $F$ and $G$, respectively. Furthermore, ${\mathbb{E}_{i_j,j_k\mathcal{A},\mathcal{D}}}[ {\nabla {{\tilde f}_k}} ] \ne \nabla f({x_k})$ as well. This gives us more discussion about the upper bound with respect to the estimated function and the gradient under the new random subset $\cal D$. \subsection{Technical Tool} For the subset $\mathcal{A} \subseteq [n]$, we present the following lemma that the variance of a random variable decreases by a factor $|\cal A|$ if we choose $|\cal A|$ independent element from $[n]$ and average them. The proof process is trivial and can be referred to Appendix. However, it present an important tool for analyzing the query complexity under the different size of the subset. \begin{lemma}\label{VRNonCS-SCSG:Appendix:Lemma:Inequation-indepent} If $v_1,...,v_m\in \mathbb{R}^d$ satisfy $\sum\nolimits_{i = 1}^m {{v_i}} = \vec 0$, and $\cal A$ is a non-empty, uniform random subset of $[m]$, $A=|\cal A|$, then \begin{center} ${\mathbb{E}_{\cal A}} {{{\left\| {\frac{1}{A}\sum\nolimits_{b \in {\cal A}} {{v_b}} } \right\|}^2}} \le \frac{{\mathbb{I}\left( {A < m} \right)}}{A}\frac{1}{m}\sum\limits_{i = 1}^m {v_i^2}.$ \end{center} Furthermore, if the elements in $\cal A$ are independent, then \begin{center} $ {\mathbb{E}_\mathcal{A}}{{{\left\| {\frac{1}{A}\sum\nolimits_{b \in \mathcal{A}} {{v_b}} } \right\|}^2}} = \frac{1}{{An}}\sum\limits_{i = 1}^n {v_i^2}.$ \end{center} \end{lemma} Based on Lemma \ref{VRNonCS-SCSG:Appendix:Lemma:Inequation-indepent}, we can obtain the inequality with two-variables ${\mathcal{D}_1}$ and ${\mathcal{D}_2}$, which are used for the gradient of $f(x)$ with the partial gradient $\partial G(x)$. \begin{lemma}\label{VRNonCS-SCSG:Appendix:Lemma:Inequation-indepent-double} If $w_1,...,w_n\in \mathbb{R}^{M\times N}$ and $v_1,...,v_n\in \mathbb{R}^M$ satisfy $(\frac{1}{{{n}}} \sum\nolimits_{i \in [n]} {{w_{i}}} )^\mathsf{T}( \frac{1}{{{n}}} \sum\nolimits_{j \in [n]} {{v_{j}}} ) = \bar w^\mathsf{T}\bar v$, and $\mathcal{D} = \left[ {{\mathcal{D}_1},{\mathcal{D}_2}} \right]$ is a non-empty, uniform random subset consist of ${{\mathcal{D}_1}} $ and $ {{\mathcal{D}_2}} $, which are independently and uniformly selected from $[n]$, $D=|\mathcal{D}_1|=|\mathcal{D}_2|$, then \begin{align*} {\mathbb{E}_\mathcal{D}}{\left\| {\frac{1}{{\left| {{\mathcal{D}_1}} \right|\left| {{\mathcal{D}_2}} \right|}}\left( {\sum\nolimits_{d_1 \in {\mathcal{D}_1}} {{w_{d_1}}} } \right)^\mathsf{T}\left( {\sum\nolimits_{d_2 \in {\mathcal{D}_2}} {{v_{d_2}}} } \right) - \bar w\bar v} \right\|^2} =& {\mathbb{E}_\mathcal{D}}{\left\| {\frac{1}{D^2}\left( {\sum\nolimits_{\left[ {{d_1},{d_2}} \right] \in {\cal D}} {\left( {({w_{{d_1}}})^\mathsf{T}{v_{{d_2}}} - \bar w^\mathsf{T}\bar v} \right)} } \right)} \right\|^2}\\ \le&\frac{{\mathbb{I}\left({D^2<n^2} \right)}}{D^2}\frac{1}{{{n^2}}}\sum\limits_{i,j = 1}^n {{{\left\| {({w_i})^\mathsf{T}{v_j} - \bar w^\mathsf{T}\bar v} \right\|}^2}}. \end{align*}} \end{lemma} \subsection{Bounds analysis of the estimated function and the gradient} Here, we mainly give different kinds of bounds for the proposed algorithm, such as ${\mathbb{E}_{\mathcal{A},\mathcal{D}_1}}\| {\hat G}_k - G(x_k) \|^2$, $\mathbb{E}_{\mathcal{A,D}}\| E_{i_k,j_k}[ \nabla {\tilde f}_k ] - \nabla f(x_k) \|^2$ and $\mathbb{E}_{i_k,j_k,\cal A, \cal D}\| \nabla {\tilde f}_k - \nabla f( x_k ) \|^2$. These bounds will be used to analyze the convergence rate and query complexity. We assume that these bound are all base on Assumption \ref{VRNonCS-SCSG:Assumption:G}-\ref{VRNonCS-SCSG:Assumption:Middle-Bound}. Parameters such as $B_G$, $B_F$, $L_G$, $L_F$ and $L_f$ in the bound are all from these Assumptions. We do not define the exact value of parameters such as $h$, $A$ and $D$, which have great influence on the convergence and will be clearly defined in the query analysis. Our proposed bound are similar to that of \cite{lian2016finite}, \cite{liu2017duality} and \cite{liu2017variance}, but, the difference lies that there is an extra subset $\cal D$, which shows an interesting phenomenon. That is when the subset $\cal D$ is equal to the $[n]$, the corresponding bounds are the same as in \cite{lian2016finite}, \cite{liu2017duality} and \cite{liu2017variance}. However, it is the independent subset $\cal D$ that gives more general query complexity result for the problem (\ref{VRNonCS-SCSG:ProblemMainComposition}). The following bounds are all used for the composition problem for both convex and non-convex problem based on the Lemma \ref{VRNonCS-SCSG:Appendix:Lemma:Inequation-indepent} and Lemma \ref{VRNonCS-SCSG:Appendix:Lemma:Inequation-indepent-double}. The more details of the proof can be referred to Appendix. For simplicity, we drop the superscript $i_k$, $j_k$, $\cal A$ and $\cal D$ for the expectation with $\mathbb{E}$ in the proof. \begin{lemma}\label{VRNonCS-SCSG:Lemma:Bound-estimate-G-2} Suppose Assumption \ref{VRNonCS-SCSG:Assumption:G} and \ref{VRNonCS-SCSG:Assumption:Middle-Bound} holds, for ${{\hat G}_k}$ defined in (\ref{VRNonCS-SCSG:Definition:Estimate-G}) with $D =\left|{\cal D}_1\right|$ and $A =\left|{\cal A}\right|$, we have \begin{align*} {\mathbb{E}_{\mathcal{A},\mathcal{D}_1}}{\| {{{\hat G}_k} - G({x_k})} \|^2} \le 4\left( {\frac{{\mathbb{I}\left( {A < n} \right)}}{A} + \frac{{\mathbb{I}\left( {D < n} \right)}}{D}} \right)B_G^2\mathbb{E}{\left\| {{x_k} - {{\tilde x}_s}} \right\|^2} + 2\frac{{\mathbb{I}\left( {D < n} \right)}}{D}{H_1}. \end{align*}} \end{lemma} \begin{lemma}\label{VRNonCS-SCSG:Lemma:Bound-estimate-G-3} Suppose Assumption \ref{VRNonCS-SCSG:Assumption:G}, \ref{VRNonCS-SCSG:Assumption:F}, \ref{VRNonCS-SCSG:AssumptionIndependent} and \ref{VRNonCS-SCSG:Assumption:Middle-Bound} holds, for ${{\hat G}_k}$ defined in (\ref{VRNonCS-SCSG:Definition:Estimate-G}) and ${\nabla {{\tilde f}_k}}$ defined in (\ref{VRNonCS-SCSG:Definition:Estimate-f}) with $\mathcal{D} = \left[ {{\mathcal{D}_1},{\mathcal{D}_2}} \right]$ and $D =\left|{\cal D}_1\right|=\left|{\cal D}_2\right|$, we have \begin{align*} {\mathbb{E}_{\mathcal{A,D}}}{\| {{\mathbb{E}_{i_k,j_k}}[ {\nabla {{\tilde f}_k}} ] - \nabla f({x_k})} \|^2} \le& 4B_G^4L_F^2\left( {4\frac{{\mathbb{I}\left( {A < n} \right)}}{A} + 4\frac{{\mathbb{I}\left( {D < n} \right)}}{D}} \right)\mathbb{E}{\left\| {{x_k} - {{\tilde x}_s}} \right\|^2}\\& + 16B_G^2L_F^2\frac{{\mathbb{I}(D < n)}}{D}{H_1} + 4\frac{{\mathbb{I}(D^2 < n^2)}}{{D^2}}{H_2}. \end{align*}} \end{lemma} \begin{lemma}\label{VRNonCS-SCSG:Lemma:Bound-estimate-G-4} Suppose Assumption \ref{VRNonCS-SCSG:Assumption:G}-\ref{VRNonCS-SCSG:Assumption:Middle-Bound} holds, for ${{\hat G}_k}$ defined in (\ref{VRNonCS-SCSG:Definition:Estimate-G}) and ${\nabla {{\tilde f}_k}}$ defined in (\ref{VRNonCS-SCSG:Definition:Estimate-f}) with $\mathcal{D} = \left[ {{\mathcal{D}_1},{\mathcal{D}_2}} \right]$ and $D =\left|{\cal D}_1\right|=\left|{\cal D}_2\right|$, we have \begin{align*} \mathbb{E}_{i_k,j_k,\cal A, \cal D}{\| {\nabla {{\tilde f}_k} - \nabla f\left( {{x_k}} \right)} \|^2} \le& 5B_G^4L_F^2\left( {\frac{L_f^2}{B_G^4L_F^2} + 4\frac{{\mathbb{I}\left( {A < n} \right)}}{A} + 4\frac{{\mathbb{I}\left( {D < n} \right)}}{D}} \right)\mathbb{E}{\left\| {{x_k} - {{\tilde x}_s}} \right\|^2}\\& + 20B_G^2L_F^2\frac{{\mathbb{I}(D < n)}}{D}{H_1} + 5\frac{{\mathbb{I}(D^2 < n^2)}}{{D^2}}{H_2}. \end{align*}} \end{lemma} As can be seen from the above results directly, when $A$ and $D$ increase, the upper bounds are more approximating the related bounds as in \cite{lian2016finite,liu2017duality,liu2017variance}. Though there are extra terms with respect to $A$ and $D$, it gives us another direction for analyzing the convergence rate and query complexity. As the convergence rate not only depends on the convergence sequence, but also the terms including the event function $\mathbb{I}$. Thus, we can obtain the lower bound range of $A$ and $D$ that is related to $\varepsilon$. The result in Lemma \ref{VRNonCS-SCSG:Lemma:Bound-estimate-G-3} and \ref{VRNonCS-SCSG:Lemma:Bound-estimate-G-4} are similar except the extra term $L_f^2\mathbb{E}||x_k-\tilde{x}_s||^2$. This is due to the fact that the order of the expectation is different. This difference derives from the proof process by using the smoothness of the function $f(x)$ and the update of $x$ in Algorithm \ref{VRNonCS-SCSG:AlgorithmI}. Furthermore, these two lemmas can be both applied to analyze the convergence rate and query complexity of the convex and non-convex composition problem. \section{Stochastic Composition via SCSG method for the Convex Composition problem} \label{VRNonCS-SCSG:Section:SCSG-convex} In this section, we analyze the proposed method for the convex composition problem. We first present the convergence of the proposed algorithm, and then give the query complexity. Thought the proof is similar to that of \cite{lian2016finite} and \cite{xiao2014proximal}, we present a more clear and simple process as there is an extra term deriving from the subset $\cal D$. In order to ensure the convergence of the proposed algorithm, we obtain the desired parameters' setting, such as $A$, $D$, $K$, $\eta$ and $h$. Based on the setting, we can obtain the corresponding query complexity, which is better than or equal to the SVRG-based method in \cite{lian2016finite} and \cite{liu2017duality}. This is in fact that the event function $\mathbb{I}$ has the influence on the size of $A$ and $D$. \subsection{Convergence analysis for the convex problem} Based on the strong convex and smoothness of the function of $f(x)$, we provide the convergence sequence, in which the parameters are not defined. But the sequences motivate us to consider the parameters' setting such that lead to the desired convergence rate. \begin{theorem}\label{VRNonCS-SCSG:Lemma:NewSequence} Suppose Assumption \ref{VRNonCS-SCSG:Assumption:f-strong}- \ref{VRNonCS-SCSG:Assumption:Middle-Bound} holds, in Algorithm \ref{VRNonCS-SCSG:AlgorithmI}, let $h>0, \eta>0$, $A=|\cal A|$, $D=|\mathcal{D}_1|=|\mathcal{D}_2|$, $K$ is the number of the inner iteration, $x^*$ is the optimal point, we have \begin{center} ${\rho _1}\mathbb{E}{\left\| {{{\tilde x}_{s + 1}} - {x^*}} \right\|^2} \le \left( {\frac{1}{K} + {\rho _2}} \right)\mathbb{E}{\left\| {{{\tilde x}_s} - {x^*}} \right\|^2} +\rho_3,$ \end{center} where \begin{align} \label{VRNonCS-SCSG:Lemma:NewSequence-V} V =& B_G^4L_F^2\left( {4\frac{{\mathbb{I}\left( {A < n} \right)}}{A} + 4\frac{{\mathbb{I}\left( {D < n} \right)}}{D}} \right),\\ \label{VRNonCS-SCSG:Lemma:NewSequence-V1} {V_1} =& 20B_G^2L_F^2\frac{{\mathbb{I}(D < n)}}{D}{H_1} + 5\frac{{\mathbb{I}({D^2} < {n^2})}}{{{D^2}}}{H_2},\\ \label{VRNonCS-SCSG:Lemma:NewSequence-r1} {\rho _1} =& \left( {2\mu - h - 4V\frac{1}{h} - \left( {12L_f^2 + 10V} \right)\eta } \right)\eta, \\ \label{VRNonCS-SCSG:Lemma:NewSequence-r2} {\rho _2} =& 2\left( {2V\frac{1}{h} + 5\left( {L_f^2 + V} \right)\eta } \right)\eta,\\ \label{VRNonCS-SCSG:Lemma:NewSequence-r3} {\rho _3} =& \frac{1}{h}\eta {\frac{4}{5}{V_1}} + 2{\eta ^2}{V_1}. \end{align}} \end{theorem} We do not give the convergence form for the update of iteration as the we do no sure whether $\rho_1$ is positive or not. Based on above equality in Lemma \ref{VRNonCS-SCSG:Lemma:NewSequence}, we assume that $\rho_1>0$ in (\ref{VRNonCS-SCSG:Lemma:NewSequence-r1}), then we can obtain \begin{align}\label{VRNonCS-SCSG:Lemma:NewSequence-convergence} \mathbb{E}{\left\| {{{\tilde x}_S} - {x^*}} \right\|^2} \le& {\rho ^S}\mathbb{E}{\left\| {{{\tilde x}_0} - {x^*}} \right\|^2} + \frac{{{\rho _3}}}{{{\rho _1}}}\sum\limits_{s = 0}^S {{\rho ^s}} \nonumber\\ \le & {\rho ^S}\mathbb{E}{\left\| {{{\tilde x}_0} - {x^*}} \right\|^2} + \frac{{{\rho _3}}}{{{\rho _1}}}\frac{1-\rho^S}{{1 - \rho }}, \end{align}} where $\rho=(\frac{1}{K}+\rho_2)/\rho_1$, $\rho_2$ and $\rho_3$ defined in (\ref{VRNonCS-SCSG:Lemma:NewSequence-r2}) and (\ref{VRNonCS-SCSG:Lemma:NewSequence-r3}), the last inequality is based on the formula of geometric progression. Thus, if the $\tilde{x}_S$ converge to the optimal point $x^*$, we need to require that $\rho<1$ and the second term ${\rho _3}( {1 - {\rho ^S}} )/( {{\rho _1}( {1 - \rho } )} )$ is less than $\varepsilon/2$. Actually, if $D=n$, the second term is equal to zero satisfying the requirement directly, which is similar to the convergence results in \cite{lian2016finite} and \cite{liu2017duality}. \subsection{Query complexity analysis for the convex problem} Based on the above result in (\ref{VRNonCS-SCSG:Lemma:NewSequence-convergence}), we analyze the query complexity. Furthermore, we also present the parameters' setting, and then derived the query complexity, in which the details information can be referred to the Appendix. \begin{corollary}\label{VRNonCS-SCSG:Lemma:paraeters setting} Suppose Assumption \ref{VRNonCS-SCSG:Assumption:f-strong}- \ref{VRNonCS-SCSG:Assumption:Middle-Bound} holds, in Algorithm \ref{VRNonCS-SCSG:AlgorithmI}, let $h=\mu$, the step size is $\eta \le {\mu }/({{135L_f^2}})$, the subset size of $\cal A$ is $ A = \min \{ {n,128B_G^4L_F^2/{{{\mu ^2}}}} \}$, the subset size of $\mathcal{D}_1$ and $\mathcal{D}_2$ are both $D = \min \left\{ {n,5\left( {16B_G^4L_F^2{H_1} + 4{H_2}} \right)/({{4\varepsilon {\mu ^2}}}}) \right\}$, the number of the inner iteration is $K \ge 540L_f^2/\mu ^2$, the number of outer iteration is $S \ge 1/( \log ( 1/\rho ) )\log ( 2E\| {\tilde x}_0 - x^* \|^2/\varepsilon )$. the query complexity is {\small \begin{align*} {\left( {D + KA} \right)S} = \mathcal{O}\left( {\left( {\min \left\{ {n,\frac{1}{{{\varepsilon\mu ^2}}}} \right\} + \frac{{L_f^2}}{{{\mu ^2}}}\min \left\{ {n,\frac{1}{{{\mu ^2}}}} \right\}} \right)\log \left( {1/\varepsilon } \right)} \right). \end{align*} } \end{corollary} As can be seen from the above result, Corollary \ref{VRNonCS-SCSG:Lemma:paraeters setting} present the general query complexity under different parameters. Comparing $n$ with corresponding parameters, we analyze the query complexity separately. We remove the parameters such as $B_G^2$, $L_F^2$, $H_1$ and $H_2$, and analyze the size with the order of $1/\mu$. Though the comparison is not exactly correct, we present the results to illustrate the corresponding different algorithms. We can directly obtain that $1/\mu^2<1/(\varepsilon\mu^2)$. We consider three situations comparing with $n$, that is to present the value of the $min$ function, \begin{itemize} \item $\frac{1}{{{\mu ^2}}} \le \frac{1}{{\varepsilon {\mu ^2}}} \le n$. When n is large enough such that we can obtain the query complexity is ${\cal O}( {( {1/( {\varepsilon {\mu ^2}} ) + {}L_f^2/{\mu ^4}} )\log ( {1/\varepsilon } )} )$. This result avoids the situation that computing the full gradient of $f(x)$ and the full function $G(x)$ for the large-scale number of n. What's more, this result is better than Compositional-SVRG \cite{lian2016finite} and \cite{liu2017duality}. \item $\frac{1}{{{\mu ^2}}} \le n \le \frac{1}{{\varepsilon {\mu ^2}}}$. When n is smaller than $1/(\varepsilon\mu^2)$, the query complexity becomes ${\cal O}( {( {n + {}L_f^2/{\mu ^4}} )\log ( {1/\varepsilon } )} )$, which is the same as Compositional-SVRG \cite{lian2016finite}. That is we need to compute the full gradient of $\nabla f(\tilde{x}_s)$ as in (\ref{VRNonCS-SCSG:Definition:Estimate-f-svrg}). The estimation of inner function $G(x)$ is the same as in \cite{lian2016finite}. \item $n \le \frac{1}{{{\mu ^2}}} \le \frac{1}{{\varepsilon {\mu ^2}}}$. When $n$ is small, the query complexity becomes ${\cal O}( {( {n + L_f^2n/{\mu ^2}} )\log ( {1/\varepsilon } )} )$. The result has the similar form to SVRG \cite{johnson2013accelerating}. This also gives us the intuition that the inner function should be computed directly rather than estimated. \end{itemize} \section{Stochastic Composition via SCSG method for the Non-convex composition problem} \label{VRNonCS-SCSG:Section:SCSG-Non-convex} In this section, we give the analysis of the convergence analysis and the query complexity under the proposed algorithm for the non-convex composition. We first present the new reformed sequence with respect to $E[f({x_k})] + {c_k}E{\left\| {{x_k} - {{\tilde x}_s}} \right\|^2}$, in which the parameters are not well defined. Then, we sum-up these sequence based on the SVRG-based on the framework, in which there is a snapshot point ${\tilde x}_s$ in each epoch. The last not least, we present the query complexity analysis and derive the optimal parameters' setting such that improve the query complexity. \subsection{Convergence analysis for the non-convex problem} We first present the new form sequence under the Lyapunov function based on the smoothness of $f(x)$ and the update of $x$. The new parameters such as $c_k$, $u_k$ and $J_k$ will be used to form sequence such that we can obtain the convergence sequence. \begin{lemma}\label{VRNonCS-SCSG:Non:Lemma:NewSequence} Suppose Assumption \ref{VRNonCS-SCSG:Assumption:G}- \ref{VRNonCS-SCSG:Assumption:Middle-Bound} holds, in Algorithm \ref{VRNonCS-SCSG:AlgorithmI}, we can obtain the following new sequence with respect to $f(x_k)$ and $||x_k-\tilde{x}_s||^2$, let $h>0,\eta>0$, $A =\left|{\cal A}\right|$ and $D =\left|{\cal D}_1\right|=\left|{\cal D}_2\right|$, we have \begin{align*} \mathbb{E}[f({x_{k + 1}})] + {c_{k + 1}}\mathbb{E}{\left\| {{x_{k + 1}} - {{\tilde x}_s}} \right\|^2} \le \mathbb{E}[f({x_k})] + {c_k}\mathbb{E}{\left\| {{x_k} - {{\tilde x}_s}} \right\|^2} - {u_k}{\left\| {\nabla f({x_k})} \right\|^2} +J_k, \end{align*} where \begin{align} \label{VRNonCS-SCSG:Non:Lemma:NewSequence-W} W =& B_G^4L_F^2\left( {4\frac{{\mathbb{I}\left( {A < n} \right)}}{A} + 4\frac{{\mathbb{I}\left( {D < n} \right)}}{D}} \right),\\ \label{VRNonCS-SCSG:Non:Lemma:NewSequence-c_k} {c_k} =& {c_{k + 1}}\left( {1 + \left( {\frac{2}{h} + 4hW} \right)\eta + 10\left( {L_f^2 + W} \right){\eta ^2}} \right) + 2W\eta + 5(L_f^2+W){L_f}{\eta ^2},\\ \label{VRNonCS-SCSG:Non:Lemma:NewSequence-u_k} {u_k} =& \left( {\left( {\frac{1}{2} - h{c_{k + 1}}} \right)\eta - \left( {{L_f} + 2{c_{k + 1}}} \right){\eta ^2}} \right),\\ \label{VRNonCS-SCSG:Non:Lemma:NewSequence-W_1} {W_1} =& 20B_G^2L_F^2\frac{{\mathbb{I}(D < n)}}{D}{H_1} + 5\frac{{\mathbb{I}(D^2 < n^2)}}{{D^2}}{H_2},\\ \label{VRNonCS-SCSG:Non:Lemma:NewSequence-J_k} J_k=& \left( {\frac{1}{2} + h{c_{k + 1}}} \right){\frac{4}{5}W_1}\eta + \left( {{L_f} + 2{c_{k + 1}}} \right){W_1}{\eta ^2}. \end{align} } \end{lemma} Based on the above inequality with respect to the sequence $\mathbb{E}[f({x_k})] + {c_k}\mathbb{E}{\left\| {{x_k} - {{\tilde x}_s}} \right\|^2}$ and Algorithm \ref{VRNonCS-SCSG:AlgorithmI}, we can obtain the convergence form in which the parameters are not clearly defined. \begin{theorem}\label{VRNonCS-SCSG:Non:Lemma:Convergence form} Suppose Assumption \ref{VRNonCS-SCSG:Assumption:G}- \ref{VRNonCS-SCSG:Assumption:Middle-Bound} holds, in Algorithm \ref{VRNonCS-SCSG:AlgorithmI}, we can obtain the following new sequence with respect to $f(x_k)$ and $||x_k-\tilde{x}_s||^2$. $K$ is the number of inner iterations, $S$ is the number of inner iterations, we have \begin{align*} u_0\mathbb{E}[{\| {\nabla f({\hat x_k^s })} \|^2}] \le \frac{{f({x_0}) - f({x^*})}}{{KS}}+J_0, \end{align*} where $\hat x_k^s$ is the output point, $J_0$ and $u_0$ are defined in (\ref{VRNonCS-SCSG:Non:Lemma:NewSequence-J_k}) and (\ref{VRNonCS-SCSG:Non:Lemma:NewSequence-u_k}). \end{theorem} \subsection{Query complexity analysis for the non-convex problem} Consider the convergence form above, we actually can't obtain the convergence rate if the parameter $u_k$ in (\ref{VRNonCS-SCSG:Non:Lemma:NewSequence-u_k}) is negative. Furthermore, there is extra term $J_0$ derived from the subset $\cal D$. We need to consider the size of the subset $\cal D$ such that we can keep the $J_0$ under our desired degree of accuracy $\epsilon$. What's more, the parameter $c_k$ in (\ref{VRNonCS-SCSG:Non:Lemma:NewSequence-c_k}) is not a constant, which has a relationship with $K$ and $\eta$. Based on these influence element, we consider the parameters' setting and give the query complexity. \begin{corollary}\label{VRNonCS-SCSG:Non:Lemma:paraeters setting} Suppose Assumption \ref{VRNonCS-SCSG:Assumption:G}- \ref{VRNonCS-SCSG:Assumption:Middle-Bound} holds, in Algorithm \ref{VRNonCS-SCSG:AlgorithmI}, for the step $\eta>0$, by setting $h =\sqrt {1/\eta}$, the set-size of the subset $\mathcal{D}_1$ and $\mathcal{D}_2$ are $ D = \min \left\{ {n,\mathcal{O}(1/\varepsilon )} \right\}$, the set-size of $\cal A$ is $A = \min \left\{ {n,\mathcal{O}\left( {1/\eta } \right)} \right\}$, the number of inner iteration is $K \le \mathcal{O}\left( {1/{\eta ^{3/2}}} \right)$, the total number of iteration is $T = \mathcal{O}\left( {1/\left( {\varepsilon \eta } \right)} \right)$, then we can obtain $\mathbb{E}[\| \nabla f(\hat x_k^s ) \|^2] \le \varepsilon$. \end{corollary} The above corollary gives the parameters' setting except the step $\eta$, note that the outer number of iteration $S$ has the relationship with $T$ and $K$, that is $T=SK$. Here, we present the optimal step $\eta$ such that we reach the improved the query complexity. \begin{corollary} Suppose Assumption \ref{VRNonCS-SCSG:Assumption:G}- \ref{VRNonCS-SCSG:Assumption:Middle-Bound} holds, in Algorithm \ref{VRNonCS-SCSG:AlgorithmI}, the step size is $\eta = \min \{ 1/n^{2/5},\varepsilon ^{2/5} \}$, then the query complexity is \begin{center} $\mathcal{O}\left( \min \left\{ {\frac{1}{{{\varepsilon ^{9/5}}}},\frac{{{n^{4/5}}}}{\varepsilon }} \right\} \right),$ \end{center} \end{corollary} \begin{proof} Based on the parameters' setting, that is $ D = \min \left\{ {n,\mathcal{O}(1/\varepsilon )} \right\}$, $A = \min \left\{ {n,\mathcal{O}\left( {1/\eta } \right)} \right\}$, $K \le \mathcal{O}\left( {1/{\eta ^{3/2}}} \right)$, and $T = \mathcal{O}\left( {1/\left( {\varepsilon \eta } \right)} \right)$, we have, { \begin{align*} \mathcal{O}\left( {\frac{T}{K}\left( {D + KA} \right)} \right) =& \mathcal{O}\left( {\frac{1}{{\varepsilon \eta }}\left( {\frac{D }{K} + A} \right)} \right)\\ =& \mathcal{O}\left( {\frac{1}{{\varepsilon \eta }}\left( {\min \left\{ {n,\frac{1}{\varepsilon }} \right\} {{\eta ^{3/2}}} + \frac{1}{\eta }} \right)} \right)\\ =& \mathcal{O}\left( {\frac{1}{\varepsilon }\left( {\min \left\{ {n,\frac{1}{\varepsilon }} \right\}{{\eta ^{1/2}}} + \frac{1}{{{\eta ^2}}}} \right)} \right)\\ \ge& \mathcal{O}\left( \min \left\{ {\frac{1}{{{\varepsilon ^{9/5}}}},\frac{{{n^{4/5}}}}{\varepsilon }} \right\} \right), \end{align*} } where the optimal $\eta = \min \left\{ {1/{n^{2/5}},{\varepsilon ^{2/5}}} \right\}$. \end{proof} As can be sen from the above result, we can see that when $n$ is large enough the query complexity become $ \mathcal{O}(1/\varepsilon^{9/5})$, that is the gradient and the inner function are estimated rather than computed the full value directly. The corresponding is better than the accelerated method in \cite{wang2016accelerating}, in which the query complexity does not depend on $n$. Furthermore, when $n\le 1/\varepsilon$, the query complexity is $\mathcal{O}(n^{4/5}/\varepsilon)$, which is the same as in \cite{liu2017variance} for the case of the problem in (\ref{VRNonCS-SCSG:ProblemMainComposition}). \section{Mini-batch version of SC-SCSG for the composition problem} \label{VRNonCS-SCSG:Section:SCSG-minibatch} \begin{algorithm}[t] \caption{Mini-batch version of SC-SCSG for the composition problem} \label{VRNonCS-SCSG:AlgorithmII} \begin{algorithmic} \Require $K$, $S$, $\eta$ (learning rate), $\tilde{x}_0$ and $\mathcal{D} = \left[ {{\mathcal{D}_1},{\mathcal{D}_2}} \right]$ \For{$s =0,1, 2,\cdots,S-1$} \State Sample from $[n]$ for D times to form mini-batch $\mathcal{D}_1$ \State Sample from $[n]$ for D times to form mini-batch $\mathcal{D}_2$ \State $\nabla {\hat f_{ \mathcal{D}}}({{\tilde x}_s}) = {(\partial G_{ \mathcal{D}_1}({{\tilde x}_s}))^\mathsf{T}}{\nabla {F_{{ \mathcal{D}_2}}}(G_{ \mathcal{D}_1}({{\tilde x}_s}))} $\Comment{D Queries} \State $x_0=\tilde{x}_s$ \For{$k =0,1,2,\cdots,K-1$} \State Sample from $[m]$ to form mini-batch $\mathcal{A}$ \State ${{\hat G}_k} = {G_{{{\cal A}}}}({x_k}) - {G_{{{\cal A}}}}({{\tilde x}_s}) + {G_{\mathcal{D}_1}}({{\tilde x}_s})$\Comment{A Queries} \State ${\Lambda_0}=0$ \For{t=1,...,b} \State Uniformly and randomly pick $i_k$ and $j_k$ from $[n]$ \State Compute the estimated gradient $\nabla {{\tilde f}_k}$ from (\ref{VRNonCS-SCSG:Definition:Estimate-f}) \Comment{4 Queries} \State ${\Lambda _{t+1}}={\Lambda _t}+\nabla {{\tilde f}_k}$ \EndFor \State ${\Lambda}={\Lambda _b}/b$ \State ${x_{k+1}}= {x_k} - {\eta} {\Lambda} $ \EndFor \State Update $\tilde{x}_{s+1}=x_K $ \EndFor \\ \textbf{Output:} $\hat x_k^s$ is uniformly and randomly chosen from $s\in\{0,...,S-1\}$ and $k\in \{0,..,K-1\}$. \end{algorithmic} \end{algorithm} In this section, we present the mini-batch version of the proposed method in Algorithm \ref{VRNonCS-SCSG:AlgorithmII}. The difference with Algorithm \ref{VRNonCS-SCSG:AlgorithmI} is the computation of the gradient of the $f(x)$. Furthermore, the convergence proof with the related upper bounds are almost the same except the following lemma. By using the Lemma \ref{VRNonCS-SCSG:Appendix:Lemma:Inequation-indepent}, we derive the similar bound but the first term is reduced by a factor of $b$, where $b$ is the number of the mini-batch. Note that here the element in the mini-batch are independent, we can obtain the result directly. The details can be referred to Appendix. \begin{lemma}\label{VRNonCS-SCSG-miniBatch:Lemma:Bound-estimate-G-4} Suppose Assumption \ref{VRNonCS-SCSG:Assumption:G}-\ref{VRNonCS-SCSG:Assumption:Middle-Bound} holds, for ${{\hat G}_k}$ defined in (\ref{VRNonCS-SCSG:Definition:Estimate-G}) and $\Lambda $ defined in Algorithm \ref{VRNonCS-SCSG:AlgorithmII} with $\mathcal{D} = \left[ {{\mathcal{D}_1},{\mathcal{D}_2}} \right]$ and $D =\left|{\cal D}_1\right|=\left|{\cal D}_2\right|$, we have \begin{align*} &\mathbb{E}_{i_k,j_k,\cal A, \cal D}{\| {\Lambda - \nabla f\left( {{x_k}} \right)} \|^2}\\ \le&5B_G^4L_F^2\left( {\frac{L_f^2}{bB_G^4L_F^2} + 4\frac{{\mathbb{I}\left( {A < n} \right)}}{A} + 4\frac{{\mathbb{I}\left( {D < n} \right)}}{D}} \right)\mathbb{E}{\left\| {{x_k} - {{\tilde x}_s}} \right\|^2} + 20B_G^2L_F^2\frac{{\mathbb{I}(D < n)}}{D}{H_1} + 5\frac{{\mathbb{I}(D^2 < n^2)}}{{D^2}}{H_2}, \end{align*} \end{lemma} Based on the above lemma, we can obtain the query complexity for both convex and non-convex problem. As the process of the proof are similar to that of Corollary \ref{VRNonCS-SCSG:Lemma:paraeters setting} and Corollary \ref{VRNonCS-SCSG:Non:Lemma:paraeters setting}, we give the following result directly. The difference of the parameters' setting are $K$, and $\eta$ due to the fact of the mini-batch. \begin{corollary}\label{VRNonCS-SCSG-miniBatch:Lemma:paraeters setting} Suppose Assumption \ref{VRNonCS-SCSG:Assumption:f-strong}- \ref{VRNonCS-SCSG:Assumption:Middle-Bound} holds, in Algorithm \ref{VRNonCS-SCSG:AlgorithmII}, for convex problem, let $h=\mu$, the step size is $\eta \le {b\mu }/({{135L_f^2}})$, the subset size of $\cal A$ is $ A = \min \{ {n,128B_G^4L_F^2/{{{\mu ^2}}}} \}$, the subset size of $\mathcal{D}_1$ and $\mathcal{D}_2$ are both $D = \min \{ n,5( 16B_G^4L_F^2{H_1} + 4{H_2} )/(4\varepsilon \mu ^2) \}$, the number of the inner iteration is $K \ge 540L_f^2/(b\mu ^2)$, the number of outer iteration is $S \ge 1/( \log ( 1/\rho ) )\log ( 2E\| {\tilde x}_0 - x^* \|^2/\varepsilon )$. The query complexity is \begin{center} ${\left( {D + KA} \right)S} = \mathcal{O}\left( {\left( {\min \left\{ {n,\frac{1}{{{\varepsilon\mu ^2}}}} \right\} + \frac{{L_f^2}}{{{b\mu ^2}}}\min \left\{ {n,\frac{1}{{{\mu ^2}}}} \right\}} \right)\log \left( {1/\varepsilon } \right)} \right).$ \end{center} \end{corollary} \begin{corollary}\label{VRNonCS-SCSG-miniBatch-non:query complexity} Suppose Assumption \ref{VRNonCS-SCSG:Assumption:G}- \ref{VRNonCS-SCSG:Assumption:Middle-Bound} holds, in Algorithm \ref{VRNonCS-SCSG:AlgorithmII}, Let $h =\sqrt {b/\eta}$, the step size is $\eta = {b^{3/5}}\min \{ 1/n^{2/5},\varepsilon ^{2/5} \}$, the set-size of $\cal A$ is $A = \min \left\{ {n,\mathcal{O}\left( {b/\eta } \right)} \right\}$, the set-size of the subset $\mathcal{D}_1$ and $\mathcal{D}_2$ are $ D = \min \left\{ {n,\mathcal{O}(1/\varepsilon )} \right\}$, the number of inner iteration is $K \le \mathcal{O}\left( {b^{1/2}/({\eta ^{3/2}})} \right)$, the total number of iteration is $T = \mathcal{O}\left( {1/\left( {\varepsilon \eta } \right)} \right)$, in order to obtain obtain $\mathbb{E}[\| \nabla f(\hat x_k^s ) \|^2] \le \varepsilon$.The query complexity is \begin{center} $\frac{1}{{{b^{1/5}}}}\mathcal{O}\left( {\min \left\{ {\frac{1}{{{\varepsilon ^{9/5}}}},\frac{{{n^{4/5}}}}{\varepsilon }} \right\}} \right)$ \end{center} \end{corollary} From the above result of the query complexity of the convex and non-convex problem, we can see that both of their step size $\eta$ and the number of inner iteration $K$ increase. These two key parameters lead to the improved the query complexity of both convex and non-convex problem. \section{Conclusion}\label{VRNonCS-SCSG:Section:Conclusion} In this paper, we propose the variance reduction based method for the convex and non-convex composition problem. We apply the stochastically controlled stochastic gradient to estimate inner function $G(x)$ and the gradient of $f(x)$. The query complexity of our proposed algorithm is better than or equal to the current methods on both convex and non-convex function. Furthermore, we also present the corresponding mini-batch version of the proposed method, in which the query complexities are improved as well. In the future, we can consider the non-smooth function of the composition problem with the method of the stochastically controlled stochastic gradient. {\small \bibliographystyle{unsrt

a3b48265638bb8fbc5ed36e819091a8624e5231f

\section{Introduction} By $X \gg Y$ or $Y \ll X$, we mean that $X \geq cY$, for some absolute constant $c > 0$. The expression $X \approx Y$ means that both $X \gg Y$ and $X \ll Y$ hold. The expression $X \gtrsim Y$ or $Y \lesssim X$ means that $X \gg Y/(\log X)^c$, for some absolute constant $c > 0$. When dependence on a certain parameter needs to be emphasized, we write the parameter in the subscript, e.g., $Y \ll_\epsilon X$ means that that the hidden constant depends on $\epsilon$. Given finite subsets $A, B$ of a ring, the {\em sum set} and {\em product set} are defined respectively as \[ A + B := \{a + b: a \in A, b \in A \},\] and \[AB := \{ab: a \in A, b \in A \}. \] Erd\H{o}s and Szemer\'edi~\cite{es83} conjectured that, for finite sets of integers, one of these must be nearly as large as possible. \begin{conjecture}[Erd\H{o}s, Szemer\'edi]\label{conj:sumProduct} Let $A$ be a finite set of integers. Then, for any $\delta < 1$, \begin{equation}\label{eq:sumProduct} |A+A| + |AA| \gg_\delta |A|^{1+\delta}. \end{equation} \end{conjecture} In their initial work on the problem, Erd\H{o}s and Szmer\'edi showed that Conjecture \ref{conj:sumProduct} holds for some fixed $\delta>0$. Subsequent works by Nathanson~\cite{nathanson97}, Ford~\cite{ford98}, Elekes~\cite{elekes97}, and Solymosi~\cite{solymosi08} gave increasing values for $\delta$. Solymosi used a beautiful and simple geometric argument to show that Conjecture~\ref{conj:sumProduct} holds with $\delta \leq 1/3 - \varepsilon$ for any $\varepsilon > 0$ whenever $A$ is a set of real numbers. This bound stood until 2015, when Konyagin and Shkredov~\cite{ks15} combined Solymosi's geometric insight with Shkredov's work in additive combinatorics to get a slight improvement. Further incremental progress was made by Konyagin and Shkredov~\cite{ks16}, by Rudnev, Shkredov and Stevens~\cite{rss16}, and by Shakan~\cite{shakan18}. Currently, Shakan's result gives the best bound for $\delta$, showing that Conjecture~\ref{conj:sumProduct} holds with $\delta \leq 1/3 + 5/5277$, whenever $A$ is a set of real numbers. The conjecture has also been studied for other fields and rings. Konyagin and Rudnev~\cite{kr13} generalized Solymosi's geometric argument to finite sets of complex numbers, showing that Conjecture~\ref{conj:sumProduct} holds with $\delta \leq 1/3 - \varepsilon$ for any $\varepsilon > 0$ whenever $A \subset \mathbb{C}$. For quaternions, Chang~\cite{chang05} proved the bound $\delta \leq 1/54$. This was improved upon by Solymosi and Wong~\cite{sw17}, who showed that, for finite subsets of the quaternions, Conjecture~\ref{conj:sumProduct} holds with $\delta \leq 1/3 - \epsilon$, for any $\epsilon > 0$. For more detail on the sum-product conjecture and its variants, see the recent survey of Granville and Solymosi~\cite{gs16}. Our contribution is to generalize Konyagin and Shkredov's proof to quaternions, hence passing the $\delta=1/3$ barrier. \begin{theorem}\label{th:main} Let $A$ be a finite set of quaternions. Then, there is a constant $c > 0$ such that \begin{equation}\label{eq:main} |A+A| + |AA| \gg |A|^{4/3 + c}. \end{equation} \end{theorem} We make no attempt to prove Theorem~\ref{th:main} for the largest possible value of $c$, instead preferring to keep the exposition relatively simple and self contained. \paragraph{Overview} Our proof follows the general outline of Konyagin-Shkredov in \cite{ks15} and \cite{ks16}. Since our aim is to keep this paper self contained rather than obtain the best value for $c$, at various points we make do with weaker estimates than the ones used in these papers. We split the problem into two cases, depending on the additive energy of $A$ (see Section~\ref{sec:prelims} for the definitions). In the case that this additive energy is small, we prove an appropriate generalization of Solymosi's argument, in the spirit of Konyagin-Rudnev~\cite{kr13} and Solymosi-Wong~\cite{sw17} (see Section~\ref{sec:large-k}). In the case that this additive energy is large, we adapt the arguments of Konyagin and Shrkedov~\cite{ks16} to work for quaternions (see Section~\ref{sec:small-k}). This requires us to replace an application of the Szmer\'edi-Trotter theorem by a generalization proved by Solymosi and Tao~\cite{st12}, and to adapt the definitions and arguments of Konyagin and Shrkedov to work when multiplication is not commutative. This is done in Section~\ref{sec:energies}. \section{Preliminaries} \label{sec:prelims} Given finite subsets $A, B$ of the quaternions, $\ensuremath{\mathbb H}$, the {\em sum set} and {\em product set} are defined respectively as \[ A + B := \{a + b: a \in A, b \in A \},\] and \[AB := \{ab: a \in A, b \in A \}. \] We define the {\em negation} of $A$ to be \[ -A := \{-a : a \in A \}, \] and the {\em inverse} of $A$ to be \[ A^{-1} := \{a^{-1} : a \in A \}. \] The {\em difference set} is defined to be $A - B$, and the {\em ratio set} is defined as $A/B := AB^{-1} \cap B^{-1}A$. Denote by $\delta_{A + B}(x)$ and by $\delta_{AB}(x)$ the number of representations of $x$ of the form $a + b$ and $ab$ with $a \in A$, $b \in B$, respectively. Let \[ E^+_k (A, B) := \sum_{x} \delta_{A-B}(x)^k, \quad \mbox{and } \quad E^*_k (A, B) := \sum_{x} \delta_{AB}(x)^k. \] The {\em additive energy of $A$ and $B$}, denoted by $E^+(A, B)$, is defined to be: \begin{align*} E^+(A, B) & := E^+_2(A, B) = \sum_{x} \delta_{A - B}(x)^2 = |\{(a, b, c, d) \in (A \times B)^2: a - b = c - d\}| \\ & = |\{(a, b, c, d) \in (A \times B)^2: a - c = b - d\}| = \sum_{x} \delta_{A - A}(x)\delta_{B-B}(x)\\ & = |\{(a, b, c, d) \in (A \times B)^2: a + d = c + b\}| = \sum_{x} \delta_{A + B}(x)^2. \end{align*} Similarly, the {\em multiplicative energy of $A$ and $B$}, denoted by $E^*(A, B)$, is defined to be \begin{align*} E^*(A, B) := E^*_2(A, B) & = \sum_{x} \delta_{AB}(x)^2 = |\{(a, b, c, d) \in (A \times B)^2: ab = cd\}|.\\ & = |\{(a, b, c, d) \in (A \times B)^2: c^{-1}a = db^{-1} \}| = \sum_{x} \delta_{A^{-1}A}(x) \delta_{BB^{-1}}(x). \end{align*} When $A = B$, we write $E^+_k(A)$, $E^+(A)$, $E^*_k(A)$, and $E^*(A)$ to simplify notation. We note that, since multiplication of quaternions is not commutative, some care is necessary when dealing with product sets, ratio sets, and the multiplicative energy. The Cauchy-Schwarz inequality implies the following lower bounds on the additive and multiplicative energies: \begin{equation} \label{eq:additivecs} E^+(A, B) \geq \frac{|A|^2|B|^2}{|A + B|}, \quad \mbox{ and }\quad E^+(A, B) \geq \frac{|A|^2|B|^2}{|A - B|}. \end{equation} \begin{equation} \label{eq:multiplicativecs} E^*(A, B) \geq \frac{|A|^2|B|^2}{|AB|}. \end{equation} \subsection{Results from incidence geometry} In this section, we set up some notation and results about {\em quaternionic lines}. In Section~\ref{sec:energies}, these results will be used to connect various energies with the sizes of the sum and product sets. Define a {\em left line} to be any set in $\ensuremath{\mathbb H}^2$ of the form $\{(a,b)+t(c,d) : t \in \ensuremath{\mathbb H}\}$, for some $a,b,c,d \in \ensuremath{\mathbb H}$ with $(c, d) \neq (0, 0)$. Similarly, define {\em right}, {\em lr-mixed} and {\em rl-mixed lines} to be sets of the form $\{(a,b)+(c,d)t : t \in \ensuremath{\mathbb H}\}$, $\{(a,b)+(tc,dt) : t \in \ensuremath{\mathbb H}\}$, and $\{(a,b)+(ct,td) : t \in \ensuremath{\mathbb H}\}$ respectively. We say two lines are of the {\em same type} if both are sets of the same form. It is straightforward to check that any two distinct quaternionic lines {\em of the same type} intersect in at most one point. Throughout this writeup, a set of quaternionic lines is restricted to have all lines of the same type. Given a point set $\mathcal P$ and a set of quaternionic lines $\mathcal L$, both in $\ensuremath{\mathbb H}^2$, we say that an incidence is a pair $(p,l) \in \mathcal P \times \mathcal L$ with the point $p$ lying on the line $l$. We denote by $\mathcal I(\mathcal P, \mathcal L)$ the set of incidences in $\mathcal P \times \mathcal L$. The following theorem is a special case of a result of Solymosi and Tao~\cite{st12}: \begin{theorem} \label{th:solymosi-tao} Let $\epsilon > 0$. Then there exists a constant $A = A_\epsilon > 0$ such that \[ |\mathcal I(\mathcal P, \mathcal L)| \leq A|\mathcal P|^{2/3 + \epsilon}|\mathcal L|^{2/3} + \frac{3}{2}|\mathcal P| + \frac{3}{2}|\mathcal L| \] whenever $\mathcal P$ is a finite set of points in $\ensuremath{\mathbb H}^2$, and $\mathcal L$ is a finite set of quaternionic lines in $\ensuremath{\mathbb H}^2$. \end{theorem} As a consequence, we have the following upper bound on the number of {\em $k$-rich lines} determined by a finite set in $\ensuremath{\mathbb H}^2$: \begin{corollary} \label{co:solymosi-tao} Let $\epsilon > 0$. Then for any set $\mathcal P$ of points in $\ensuremath{\mathbb H}^2$ and, for any integer $k \geq 2$, the number of quaternionic lines containing at least $k$ points of $\mathcal P$, denoted by $n_k$, satisfies \[ n_k \ll_\epsilon \frac{|\mathcal P|^{2+\epsilon}}{k^3} + \frac{|\mathcal P|}{k} . \] \end{corollary} Next, we use Corollary~\ref{co:solymosi-tao} to obtain the following upper bound on the number of collinear triples: \begin{lemma} \label{le:collinear-triples} Let $\epsilon > 0$. For a finite set $A$ of quaternions, let $T$ be the number of collinear triples in $A \times A \subset \ensuremath{\mathbb H}^2$. Then \[ T \ll_\epsilon |A|^{4+\epsilon} \log |A|. \] \end{lemma} \begin{proof} Consider a partitioning of the set of lines into dyadic blocks, where the $i^{th}$ block consists of lines containing between $2^i$ and $2^{i+1}$ points, with $1 \leq i \leq \lceil \log |A| \rceil$. Let $n_i$ be the number of lines in the $i^{th}$ set. Then the number of collinear triples is \begin{equation} \label{eq:triples} T \leq \sum_{i = 1}^{\lceil \log |A| \rceil} {2^{i+1} \choose 3} n_{i}. \end{equation} This allows us to use Corollary~\ref{co:solymosi-tao} with $|P| = |A|^2$ to bound $n_i$, giving \begin{align*} T & \leq \sum_{i = 1}^{\lceil \log |A| \rceil} {2^{i+1} \choose 3} n_{i} \ll_\epsilon \sum_{i = 1}^{\lceil \log |A| \rceil} {2^{i+1} \choose 3} \left(\frac{|A|^{4+\epsilon}}{(2^{i})^3} + \frac{|A|^2}{2^{i}} \right)\\ & \ll_\epsilon \sum_{i = 1}^{\lceil \log |A| \rceil} \left(|A|^{4+\epsilon} + |A|^2 2^{2i} \right) \ll_\epsilon |A|^{4+\epsilon} \log |A|. \end{align*} \end{proof} \section{Bounding energies} \label{sec:energies} In this section we collect various bounds on the sizes of sum and product in terms of energies. All the results and proofs presented in this section have appeared in various papers. We simply present the proofs and adapt them to our setting. First, we give some basic estimates on additive energies using the Cauchy-Schwarz inequality. The following appears as Lemma 2.4 and 2.5 in~\cite{li11}. \begin{lemma} \label{le:e1.5e3} Let $A, B$ be a finite sets of quaternions. Then \[ |A|^2E^+_{1.5}(A)^2 \leq E_{3}^+(A)^{2/3} E_{3}^+(B)^{1/3} E^+(A, A+B). \] \end{lemma} \begin{proof} Let $A_x = A \cap (A + x)$, and note that $|A_x| = \delta_{A-A}(x)$. Equation~\eqref{eq:additivecs} implies that \begin{equation*} |A_x|^{1.5}|B| \leq E^+(A_x, B)^{1/2}|A_x + B|^{1/2} |A_x|^{1/2}. \end{equation*} This gives \begin{align} \label{eq:b1} \nonumber E^+_{1.5}(A)^2|B|^2 & = \left(\sum_{x \in A-A} \left(|A_x|^{1.5}|B|\right)\right)^2\\ \nonumber &\leq \left( \sum_{x \in A-A} \left(E^+(A_x, B)^{1/2}|A_x + B|^{1/2}|A_x|^{1/2} \right)\right)^2\\ \nonumber &\leq \sum_{x \in A-A} E^+(A_x, B) \sum_{x \in A-A} |A_x + B| |A_x|\\ \nonumber &\leq \sum_{x \in A-A} E^+(A_x, B) \sum_{x \in A-A} |(A + B)_x| |A_x|\\ &= \left(\sum_{x \in A-A} E^+(A_x, B)\right) E^+(A, A+B). \end{align} where the second inequality follows from the Cauchy-Schwarz inequality, and the third inequality is implied by the set inclusion $A_x + B \subseteq (A + B)_x$. Next, note that \begin{align} \label{eq:b2} \nonumber \sum_{x} E^+(A_x, B) &= \sum_{x} \sum_{y} \delta_{A_x - A_x}(y) \delta_{B-B}(y) = \sum_{x} \sum_{y} \delta_{A_y - A_y}(x) \delta_{B-B}(y)\\ \nonumber &= \sum_{y} \sum_{x} \delta_{A_y - A_y}(x) \delta_{B-B}(y) = \sum_{y} \delta_{A - A}(y)^2 \delta_{B-B}(y)\\ &\leq \left( \sum_{y} \delta_{A - A}(y)^3\right)^{2/3} \left(\sum_{y} \delta_{B-B}(y)^3 \right)^{1/3} = E_3^+(A)^{2/3}E_3^+(B)^{1/3}, \end{align} where the second to last step is a standard application of H\"older's inequality. The statement of the lemma now follows by combining~\eqref{eq:b1}~and~\eqref{eq:b2}. \end{proof} The following lemma establishes a connection between the additive energy and the size of the product set, and appears as Theorem~9 in~\cite{ks15}. The technique used here was introduced by Elekes and Ruza in~\cite{er03}. \begin{lemma} \label{le:additive} Let $\epsilon > 0$, and $A$ be a finite set of quaternions. Then \[ E^+(A)^4 \lesssim_{\epsilon} |AA| |A|^{10+\epsilon}. \] \end{lemma} \begin{proof} In order to bound $E^+(A)$, we may restrict our attention to elements that have many realizations in $A+A$. Let $F$ be the set of such elements, i.e., \[ F = \left\{ x \in A+A: \delta_{A+A}(x) > \frac{E^+(A)}{2|A|^2} \right\}. \] Now \[ \sum_{x \notin F} \delta_{A+A}(x)^2 \leq \frac{E^+(A)}{2|A|^2} \sum_{x \notin F} \delta_{A+A}(x) \leq \frac{E^+(A)}{2}. \] In particular, we have \[ E^+(A) \leq 2 \sum_{x \in F} \delta_{A+A}(x)^2. \] Let $\mathcal{P} \subset (A \cup F) \times (A \cup F)$. We will double count the number of collinear triples in $\mathcal{P}$, denoted by~$T$. Denote by $\delta_F$ the quantity $\sum_{x \in F} \delta_{A+A}(x)$. Since, for all $x$, $\delta_{A+A}(x) \leq |A|$, we have \[ \delta_F = \sum_{x \in F} \delta_{A+A}(x) \geq \sum_{x \in F} \frac{\delta_{A+A}(x)^2}{|A|} \geq \frac{E^+(A)}{2|A|}. \] This gives \[ |F| + |A| \leq \sum_{x \in F} \left( \delta_{A+A}(x) \cdot \frac{2|A|^2}{E^+(A)}\right) + |A|\left( \delta_F\cdot \frac{2|A|}{E^+(A)} \right) = \frac{4|A|^2\delta_F}{E^+(A)}. \] Combining the above with Lemma~\ref{le:collinear-triples} implies \begin{equation} \label{eq:triplesupperbound} T \ll_\epsilon |A \cup F|^{4+\epsilon} \log |A \cup F| \lesssim_\epsilon \left( \frac{|A|^2 \delta_F}{E^+(A)} \right)^{4+\epsilon}. \end{equation} We now obtain a lower bound on $T$. For each $a \in A$, let $F_a = \{b \in A: a + b \in F\}$. Fix $(a, b) \in A^2$, and consider a quadruple of the form $(c, d, e, f) \in (F_a \times F_b)^2$ with $cd = ef$, or equivalently $df^{-1} = c^{-1}e$. Note that the line $\{(a, b) + (ct, tf) : t \in \ensuremath{\mathbb H} \}$ contains the points \[ (a, b), (a + c, b + f), (a + e, b + d) \in \mathcal P, \] with $t = 0, 1$, and $t = df^{-1} = c^{-1}e $ respectively. Hence, each quadruple corresponds to a collinear triple in $\mathcal P$. By Equation~\eqref{eq:multiplicativecs}, the number of such quadruples is at least \[ \frac{|F_a|^2|F_b|^2}{|F_a F_b|} \geq \frac{|F_a|^2|F_b|^2}{|AA|}. \] It follows that the number of collinear triples in $P$ is at least: \begin{align} \label{eq:tripleslowerbound} \nonumber T &\geq \sum_{a, b \in A} \frac{|F_a|^2|F_b|^2}{|AA|} = \frac{1}{|AA|} \left(\sum_{a \in A } |F_a|^2\right)^2\\ &\geq \frac{1}{|AA|} \left(\frac{1}{|A|} \left(\sum_{a \in A} |F_a|\right)^2 \right)^2 \gg \frac{\delta_F^4}{|AA||A|^2}. \end{align} Combining Equations~\eqref{eq:triplesupperbound}~and~\eqref{eq:tripleslowerbound} gives: \[ E^+(A)^{4 + \epsilon} \lesssim_\epsilon |A|^{10 + 2\epsilon} \delta_F^{\epsilon} |AA|. \] Finally, since $E(A) \geq |A|^2$ and $\delta_F = \sum_{x \in F} \delta_{A+A}(x) \leq |A|^2$, we get \[ E^+(A)^{4} \lesssim_\epsilon |AA||A|^{10 + \epsilon}. \] \end{proof} An important tool in connecting incidence results and sum-product estimates is the quantity $d_*(A)$ defined below, which captures the multiplicative structure of $A$. We define \[ d_*(A) := \min_{t > 0} \min_{\varnothing \neq Q, R \subset \ensuremath{\mathbb H} \setminus \{0\}} \frac{|Q|^2|R|^2}{|A|t^3}, \] where the second minimum is taken over all sets $Q$ and $R$ such that $\max\{|Q|, |R|\} \geq |A|$, and, for every $x \in A$, the bound $|Q \cap R x| \geq t$ holds. It is not hard to check that $d_*(A) \leq |A|$. Roughly speaking, the closer $d_*(A)$ is to $|A|$, the more multiplicative structure $A$ has. The following is a key lemma concerning $d_*(A)$ and will be used to connect additive energies with the product set. The version for real numbers appeared as Lemma 13 in~\cite{ks16}. \begin{lemma} \label{le:stsets} Let $\epsilon > 0$, $A$ and $B$ be finite sets of quaternions, and $\tau \geq 1$ an integer. Then \[ |\{x : \delta_{A-B}(x) \geq \tau\}| \ll_\epsilon \frac{|A||B|^{2 + \epsilon}}{\tau^3} d_*(A)^{1+\epsilon}. \] \end{lemma} \begin{proof} Note that we may assume $A$ and $B$ are large (with respect to the hidden constant), since otherwise the statement is trivially true. Let $Q, R$ be finite sets of quaternions, and $t > 0$ an integer such that, for all $x \in A$, $|Q \cap R x| \geq t$. For any integer $\tau \geq 1$, let \[D_\tau = \{x \in A - B: \delta_{A - B}(x) \geq \tau \}. \] To prove the lemma, it suffices to show that \begin{equation} \label{eq:required} |D_\tau| \ll_\epsilon \frac{|Q|^{2+\epsilon}|R|^{2+\epsilon}|B|^{2+\epsilon}}{t^3\tau^3}. \end{equation} Let $\sigma = |\{ a - b = x : a \in A, b \in B, x \in D_\tau \}|$, and note that, by definition, \begin{equation} \label{eq:sigmalb} \sigma = \sum_{x \in D_\tau} \delta_{A-B}(x) \geq \tau|D_\tau|. \end{equation} On the other hand, for each $x \in A$, $|Q \cap Rx| \geq t$ implies that $\delta_{R^{-1}Q}(x) \geq t$. This gives \begin{equation} \label{eq:sigmaub} \sigma \leq |\{r^{-1}q - b = x: q \in Q, r \in R, b \in B, x \in D_\tau\}|\cdot t^{-1}. \end{equation} By ignoring the condition $x \in D_\tau$, Equation~\eqref{eq:sigmaub} immediately implies the bound \begin{equation} \label{eq:sigmaub-easy} \sigma \leq \frac{|Q||R||B|}{t}. \end{equation} When $t^2\tau^2 \ll |Q||R||B|$, Equations~\eqref{eq:sigmalb}~and~\eqref{eq:sigmaub-easy} imply \[ |D_\tau| \leq \frac{|Q||R||B|}{t\tau} \ll \frac{|Q|^2|R|^2|B|^2}{t^3\tau^3}, \] allowing us to restrict our attention to the case when $t^2\tau^2 \gg |Q||R||B|$. In this scenario, we reformulate Equation~\eqref{eq:sigmaub} as an incidence problem. For $r \in R$ and $d \in D_\tau$, let $l_{r,d} = \{ (x, y) : r^{-1}y - x = d \}$ be a quaternionic line. Consider the family $\mathcal L = \{ l_{r,d}: r \in R, d \in D_\tau \}$ of $|R||D_\tau|$ lines. Let $\mathcal P$ be the point set $Q \times B$. Now, by Theorem~\ref{th:solymosi-tao}, for $\epsilon' = \epsilon/3$, we have \begin{align} \label{eq:sigmaub1} \nonumber \sigma & \leq |\mathcal I(\mathcal P, \mathcal L)| \cdot t^{-1}\\ & \ll \left(A_{\epsilon'} |\mathcal P|^{2/3 + \epsilon'}|\mathcal L|^{2/3} + |\mathcal P| + |\mathcal L|\right) \cdot t^{-1}. \end{align} If the first term in Equation~\eqref{eq:sigmaub1} dominates, we have \begin{equation*} t \tau |D_\tau| \ll A_{\epsilon'} |\mathcal P|^{2/3 + \epsilon'}|\mathcal L|^{2/3} = A_{\epsilon'}(|Q||B|)^{2/3 + \epsilon'}(|R||D_\tau|)^{2/3}. \end{equation*} Rearranging gives \begin{align*} |D_\tau| & \ll_\epsilon \frac{|Q|^{2 + \epsilon}|B|^{2 + \epsilon}|R|^2}{t^3\tau^3}. \end{align*} Now suppose that the bound \eqref{eq:required} does not hold, and assume, without loss of generality, that $|Q| = \max\{|Q|, |R|\} \geq |A|$. If the second term dominates, then \begin{equation} \label{eq:stcontradiction1} \frac{|Q|^2|R|^2|B|^2}{t^2\tau^2} \ll t \tau |D_\tau| \ll |P| = |Q||B|. \end{equation} Note that we must have $t \leq \min\{|Q|, |R|\} = |R|$, $\tau \leq |B|$, and $\tau \leq |A| \leq |Q|$. Together with Equation~\eqref{eq:stcontradiction1}, this gives a contradiction. Finally, if the third term dominates, then \[ t\tau |D_\tau| \ll |L| = |R||D_\tau|. \] Combined with the assumption that $t^2\tau^2 \geq |Q||R||B|$, this implies \[ |Q||B||R| \ll |R|^2. \] But this is a contradiction, since $|Q| \geq |R|$, and $B$ is large enough. \end{proof} We now give some consequences of Lemma~\ref{le:stsets}. The proof of all three corollaries follows the same basic outline. \begin{corollary} \label{co:e1.5} Let $\epsilon > 0$, and $A, B$ be a finite sets of quaternions. Then \[ E^+(A) \lesssim_{\epsilon} |A|^{1 + \epsilon} d_*(A)^{1/3 + \epsilon} E_{1.5}^+(A)^{2/3}. \] \begin{proof} Let $\Delta$ be a parameter to be specified later, and recall that \[ E^+(A) = \sum_{x} \delta_{A-A}(x)^2 = \sum_{x\,:\,\delta_{A-A}(x) < \Delta} \delta_{A-A}(x)^2 + \sum_{x\,:\,\delta_{A-A}(x) \geq \Delta} \delta_{A-A}(x)^2. \] We first consider sums with fewer than $\Delta$ realizations. \begin{align} \label{eq:a1} \nonumber \sum_{x\,:\,\delta_{A-A}(x) < \Delta} \delta_{A-A}(x)^2 &\leq \max_{x\,:\,\delta_{A-A}(x) < \Delta}\{\delta_{A-A}(x)^{1/2}\}\sum_{x\,:\,\delta_{A-A}(x) < \Delta} \delta_{A-A}(x)^{1.5}\\ &\leq \Delta^{1/2}E^+_{1.5}(A). \end{align} To bound the contribution of sums with more than $\Delta$ realizations, we use a dyadic decomposition along with Lemma~\ref{le:stsets} (with $B = A$ and $\epsilon' = 3\epsilon/2$). \begin{align} \label{eq:a2} \nonumber \sum_{x\,:\,\delta_{A-A}(x) \geq \Delta} \delta_{A-A}(x)^2 & = \sum_{j = 1}^{\lceil \log |A| \rceil + 1} \sum_{x\,:\, \Delta2^{j-1} \leq \delta_{A-A}(x) < \Delta 2^{j}} \delta_{A-A}(x)^2\\ \nonumber& \ll_{\epsilon'} \sum_{j = 1}^{\lceil \log |A| \rceil + 1} \frac{|A|^{3+\epsilon'}d_*(A)^{1 + \epsilon'}}{(\Delta2^{j-1})^3} (\Delta 2^{j})^2\\ & \lesssim_{\epsilon'} \frac{|A|^{3+\epsilon'}d_*(A)^{1 + \epsilon'}}{\Delta^3}. \end{align} Combining~\eqref{eq:a1},~\eqref{eq:a2}, and setting $\Delta = \left(\frac{|A|^{3+\epsilon'}d_*(A)^{1 + \epsilon'}}{E^+_{1.5}(A)}\right)^{2/3}$ finishes the proof. \end{proof} \end{corollary} \begin{corollary} \label{co:eds} Let $\epsilon > 0$, and $A, B$ be a finite sets of quaternions. Then \[ E^+(A, B) \lesssim_\epsilon |A||B|^{3/2 + \epsilon}d_*(A)^{1/2 + \epsilon} . \] \end{corollary} \begin{proof} Let $\Delta$ be a parameter to be specified later. Then \begin{align*} E^+(A, B) & = \sum_{x} \delta_{A-B}(x)^2\\ & = \sum_{x\,:\,\delta_{A-B}(x) < \Delta} \delta_{A-B}(x)^2 + \sum_{j = 1}^{\lceil \log |A| \rceil + 1} \sum_{x\,:\, \Delta2^{j-1} \leq \delta_{A-B}(x) < \Delta 2^{j}} \delta_{A-B}(x)^2\\ & \ll_{\epsilon'} \Delta |A||B| + \sum_{j = 1}^{\lceil \log |A| \rceil + 1} \frac{|A||B|^{2 + \epsilon'}d_*(A)^{1 + \epsilon'}}{(\Delta2^{j-1})^3}(\Delta2^j)^2\\ &\lesssim_{\epsilon'} \Delta |A||B| + \frac{|A||B|^{2 + \epsilon'}d_*(A)^{1 + \epsilon'}}{\Delta}. \end{align*} Setting $\Delta = \left(|B|^{1 + \epsilon'}d_*(A)^{1 + \epsilon'}\right)^{1/2}$ completes the proof. \end{proof} \begin{corollary} \label{co:e3ds} Let $\epsilon > 0$, and $A, B$ be a finite sets of quaternions. Then \[ E^+_3(A) \lesssim_\epsilon |A|^{3 + \epsilon} d_*(A)^{1 + \epsilon}. \] \end{corollary} \begin{proof} \begin{align*} E_3^+(A) & = \sum_{x} \delta_{A-A}(x)^3 = \sum_{j = 1}^{\lceil \log |A| \rceil + 1} \sum_{x\,:\, 2^{j-1} \leq \delta_{A-A}(x) < 2^{j}} \delta_{A-A}(x)^3\\ & \ll_\epsilon \sum_{j = 1}^{\lceil \log |A| \rceil + 1} \frac{|A|^{3 + \epsilon}d_*(A)^{1 + \epsilon}}{(2^{j-1})^3}(2^j)^2 \lesssim_\epsilon |A|^{3 + \epsilon}d_*(A)^{1 + \epsilon}. \end{align*} \end{proof} We now give the main result of this section. The following theorem will be used in Section~\ref{sec:small-k} when the multiplicative energy is small. \begin{theorem} \label{th:smallk} Let $\epsilon > 0$, and $A$ be a finite set of quaternions. Then \[ |A+A| \gtrsim_{\epsilon} \frac{|A|^{14/9 - \epsilon}}{d_*(A)^{5/9}}. \] \end{theorem} \begin{proof} The proof combines Equation~\eqref{eq:additivecs}, Corollary~\ref{co:e1.5}, Lemma~\ref{le:e1.5e3}, Corollary~\ref{co:eds}, and Corollary~\ref{co:e3ds} (in the specified order). This gives, for $\epsilon' = \epsilon/15$, \begin{align*} \frac{|A|^{12}}{|A+A|^3} &\leq E^+(A)^3 \\ &\lesssim_{\epsilon'} |A|^{3 + 3\epsilon'}d_*(A)^{1 + 3\epsilon'}E^+_{1.5}(A)^2\\ &\lesssim_{\epsilon'} |A|^{1 + 3\epsilon'}d_*(A)^{1 + 3\epsilon'}E_3^+(A)E^+(A, A+A)\\ &\lesssim_{\epsilon'} |A|^{2 + 3\epsilon'}d_*(A)^{3/2 + 4\epsilon'}E_3^+(A)|A+A|^{3/2 + \epsilon'}\\ &\lesssim_{\epsilon'} |A|^{5 + 4\epsilon'}d_*(A)^{5/2 + 5\epsilon'}|A+A|^{3/2 + \epsilon'}. \end{align*} This implies \[ |A+A|^{9/2 + \epsilon'} \gtrsim_{\epsilon'} \frac{|A|^{7 - 4\epsilon'}}{d_*(A)^{5/2 + 5\epsilon'}}. \] The theorem now follows by noting that $|A+A| \leq |A|^2$ and $d_*(A) \leq |A|$. \end{proof} \section{Main Theorem} \label{sec:proof} We are finally ready to prove Theorem~\ref{th:main}. The section is organized as follows: In Section~\ref{sec:initial}, we use standard arguments to find a large subset of $A$ with properties that will be convenient to work with. Section~\ref{sec:small-k} will deal with the case when the additive energy of this subset is large, and Section~\ref{sec:large-k} will deal with the case when the additive energy is small. \subsection{Initial setup and pigeonholing} \label{sec:initial} Let $A$ be a finite set of quaternions. Since we are not concerned about constants, it suffices to prove~\eqref{eq:main} for a subset $A' \subseteq A$ containing a positive proportion of elements of $A$. We view a quaternion $a = w + xi + yj + zk$ as a vector $(w, x, y, z) \in \ensuremath{\mathbb R}^4$, and say that $w$ and $(x, y, z)$ are, respectively, the {\em real} and {\em imaginary} parts of $a$. By a simple pigeonholing argument, there exists a set $A'$ of size at least $|A|/16$ such that all elements of $A'$ lie in the same orthant of $\ensuremath{\mathbb R}^4$. We assume that $0 \notin A'$, and that each element of $A'$ has non-negative real part. The latter is without loss of generality since multiplication by reals is commutative and, hence, doesn't affect the sizes of the sum and product sets. To simplify notation, we identify $A'$ with $A$, and assume that $A$ satisfies the properties that we need. To estimate the multiplicative energy, we will consider a subset of $A/A$, which we build incrementally. Let $R_0 = \{\lambda \in A/A: \delta_{A^{-1}A}(\lambda) \geq \delta_{AA^{-1}}(\lambda)\}$, and assume, without loss of generality, that \[ \sum_{\lambda \in R_0} \delta_{A^{-1}A}(\lambda)\delta_{AA^{-1}}(\lambda) \geq \sum_{\lambda \notin R_0} \delta_{A^{-1}A}(\lambda)\delta_{AA^{-1}}(\lambda). \] This gives \[ E_*(A) = \sum_{\lambda \in A/A} \delta_{A^{-1}A}(\lambda)\delta_{AA^{-1}}(\lambda) \leq 2 \sum_{\lambda \in R_0} \delta_{A^{-1}A}(\lambda)\delta_{AA^{-1}}(\lambda) \leq 2 \sum_{\lambda \in R_0} \delta_{A^{-1}A}(\lambda)^2. \] Next, we restrict our attention to $\lambda \in R_0$ with $\delta_{A^{-1}A}(\lambda) \geq E_*(A)/4|A|^2$. Let $R_1 = \{\lambda \in R_0: \delta_{A^{-1}A}(\lambda) \geq E_*(A)/4|A|^2\}$. This is possible because ratios in~$R_0 \setminus R_1$ can not account for too much the multiplicative energy, i.e., \[ \sum_{\lambda \in R_0 \setminus R_1} \delta_{A^{-1}A}(\lambda)^2 < \frac{E_*(A)}{4|A|^2} \sum_{\lambda \in R_0\setminus R_1} \delta_{A^{-1}A}(\lambda) \leq \frac{E_*(A)}{4} \leq \frac{1}{2}\sum_{\lambda \in R_0} \delta_{A^{-1}A}(\lambda)^2. \] This implies that ratios in $R_1$ give the following bound on the multiplicative energy: \begin{equation}\label{eq:multiplicative-1} E_*(A) \ll \sum_{\lambda \in R_1} \delta_{A^{-1}A}(\lambda)^2. \end{equation} For a positive integer $\tau$, let $S'_\tau$ be the set $ S'_\tau = \{\lambda \in R_1: \tau \leq \delta_{A^{-1}A}(\lambda) < 2\tau \}. $ Since $1 \leq \delta_{A^{-1}A}(\lambda) \leq |A|$, a dyadic decomposition of the summation in~\eqref{eq:multiplicative-1} gives \[ \sum_{\lambda \in R_1} \delta_{A^{-1}A}(\lambda)^2 = \sum_{i = 0}^{\lceil \log |A| \rceil} \sum_{\lambda \in S'_{2^i}} \delta_{A^{-1}A}(\lambda)^2. \] It follows, by the pigeonhole principle, that there exists a $\tau \geq E_*(A)/4|A|^2$ such that elements of $S_\tau$ contribute at least $1/\lceil \log |A|\rceil$ to the sum. This implies \[ E_*(A) \lesssim \sum_{\lambda \in S'_\tau} \delta_{A^{-1}A}(\lambda)^2 \ll |S'_\tau|\tau^2. \] Finally, for each $\lambda \in S'_\tau$, consider the set $A_\lambda := A \cap A \lambda$. Then $|A_\lambda| = \delta_{A^{-1}A}(\lambda)$, which implies that $\tau^2 \leq E_+(A_\lambda) \leq 8\tau^3$. A dyadic decomposition of this interval, along with the pigeonhole principle, implies that there exists $K \geq 1$ such that the set $S_\tau = \{ \lambda \in S'_\tau: E_{+}(A_\lambda) \approx \tau^3/K\} $ has cardinality $|S_\tau| \gtrsim |S'_\tau|$. To sum up, for an integer \begin{equation} \label{eq:taubound} \tau \gg E^*(A)/|A|^2, \end{equation} we have found a set $S_\tau \subseteq A/A$ such that \begin{equation} \label{eq:energybound} E_*(A) \lesssim |S_\tau|\tau^2, \end{equation} and, for every $\lambda \in S_\tau$, we have \[ \delta_{A^{-1}A}(\lambda) \geq \delta_{AA^{-1}}(\lambda),\,\, |A_\lambda| = \delta_{A^{-1}A}(\lambda) \approx \tau, \mbox{ and } E_+(A_\lambda) \approx \tau^3/K \mbox{ for some } K \geq 1. \] \subsection{The case when $K$ is small} \label{sec:small-k} We will show that there exists a {\em large} set $A' \subseteq A$ such that $d_*(A')$ is {\em small}. The result will then follow by using Theorem~\ref{th:smallk}. Recall that for $\lambda \in A/A$, $A_\lambda = A \cap A \lambda \subseteq A$. We will require the {\em Katz-Koester} inclusion, i.e., for any $\lambda \in A/A$, $A_\lambda A_\lambda \subseteq (AA) \cap (AA) \lambda$. Clearly $A_\lambda A_\lambda \subseteq (AA)$, so it suffices to show $A_\lambda A_\lambda \subseteq (AA) \lambda $. To see this, consider $a = b \cdot c \in A_\lambda A_\lambda$. Since $c \in A_\lambda$, there exists $c' \in A$ such that $c = c' \lambda$, that is, $c' = c \lambda^{-1} \in A$. Now we may write $a = b \cdot c = b \cdot c \cdot \lambda^{-1} \cdot \lambda = b \cdot c' \lambda \in (AA)\lambda$. Note that \[ \sum_{a \in A} |A \cap a S_\tau| = \sum_{\lambda \in S_\tau} |A \cap A \lambda| \geq |S_\tau|\tau. \] Then, by the pigeonhole principle, there exists $a \in A$ such that $A' := A \cap a S_\tau $ has size at least $|S_\tau|\tau/|A|$. Observe that Lemma~\ref{le:additive} implies that, for all $\lambda \in S_\tau$, \[ |A_\lambda A_\lambda| \gtrsim_{\epsilon'} \frac{E_+(A_\lambda)^4}{|A_\lambda|^{10 + \epsilon'}} \approx_{\epsilon'} \frac{\tau^{2 - \epsilon'}}{K^4}. \] Now, for $b \in A'$ we can write $b = a\lambda $ for some $\lambda \in S_\tau$. This implies \[ | (AA) \cap (AAa^{-1})b | = | (AA) \cap (AA) \lambda | \geq |A_\lambda A_\lambda| \gtrsim_{\epsilon'} \frac{\tau^{2 - \epsilon'}}{K^4}. \] Since this holds for an any $b \in A'$, we get \[ d_*(A') \lesssim_{\epsilon'} \frac{|AA|^4}{|A'|(\tau^{2 - \epsilon'}/K^4)^3} \approx_{\epsilon} \frac{|AA|^4 K^{12}}{|A'| \tau^{6 - 3\epsilon'}}, \] by letting $Q = AA$, $R = AAa^{-1}$, and $t = \tau^{2 - \epsilon'}/K^4$ in the definition of $d_*(A')$. Combining this with Theorem~\ref{th:smallk} gives \[ |A+A|^9 \gtrsim_{\epsilon'} \frac{|A'|^{14 - \epsilon'}}{d_*(A')^{5}} \gtrsim_{\epsilon'} \frac{|A'|^{19 - \epsilon'} \tau^{30 - 15\epsilon'}}{K^{60} |AA|^{20}}. \] We then apply Equations~\eqref{eq:energybound},~\eqref{eq:taubound},~and~\eqref{eq:multiplicativecs}, in the specified order, to get \begin{align*} K^{60} |AA|^{20} |A+A|^9 &\gtrsim_{\epsilon'} |A'|^{19 - \epsilon'} \tau^{30 - 15\epsilon'}\gtrsim_{\epsilon'} \left(\frac{|S_\tau| \tau}{|A|} \right)^{19 - \epsilon'} \cdot \tau^{30 - 15\epsilon'} \\ & \approx_{\epsilon'} \left(|S_\tau| \tau^{2}\right)^{19 - \epsilon'} \cdot \frac{ \tau^{11 - 14\epsilon' }}{|A|^{19 - \epsilon'}} \gtrsim_{\epsilon'} E^*(A)^{19 - \epsilon'} \cdot \frac{\tau^{11 - 14\epsilon' }}{|A|^{19 - \epsilon'}} \\ & \approx_{\epsilon'} \left(\frac{E^*(A)}{|A|} \right)^{19 - \epsilon'}\cdot \tau^{11 - 14\epsilon'} \gg_{\epsilon'} \left(\frac{E^*(A)}{|A|} \right)^{19 - \epsilon'} \cdot \left(\frac{E^*(A)}{|A|^2}\right)^{11 - 14\epsilon'}\\ & \approx_{\epsilon'} \frac{E^*(A)^{30- 15\epsilon'} }{|A|^{41 -29\epsilon'}} \gtrsim_{\epsilon'} \frac{1}{|A|^{41 -29\epsilon'}} \cdot \left(\frac{|A|^4}{|AA|}\right)^{30- 15\epsilon'}\\ & \approx_{\epsilon'} \frac{|A|^{79 - 31\epsilon'}}{|AA|^{30- 15\epsilon'}}. \end{align*} That is, we have \[ |AA|^{50} |A+A|^{9} \gtrsim_\epsilon \frac{|A|^{79 - \epsilon}}{K^{60}}. \] By choosing $K$ sufficiently small depending on $|A|$, this implies that $\min(|AA|, |A+A|) \geq |A|^{79/59 - \delta} > |A|^{4/3 }$, where $\delta$ is a small constant depending on $K$. \subsection{The case when $K$ is large} \label{sec:large-k} Now we handle the case that $K$ is larger than the small function of $|A|$ chosen at the end of the previous section. For convenience of notation, we label the slopes in $S_\tau$ by distinct integers, and let $A_i = A_{\lambda_i} = A \cap A{\lambda_i}$ for each $\lambda_i \in S_\tau$. A key observation, first exploited by Solymosi in the real case \cite{solymosi08} and by Solymosi and Wong \cite{sw17} for quaternions, is that, for distinct $\lambda_i, \lambda_j \in S_\tau$ and $a_1,a_2 \in A_i, b_1,b_2 \in A_j$ with $(a_1,b_1) \neq (a_2,b_2)$, we have $(a_1, a_1\lambda_i) + (b_1, b_1 \lambda_j) \neq (a_2, a_2 \lambda_i) + (b_2, b_2 \lambda_j)$. Indeed, since $\lambda_i \neq \lambda_j$, we have that $(\lambda_i - \lambda_j)$ is invertible. Hence, given $x,y \in \mathbb{H}$, we can solve the equations \[a + b = x, \qquad a \lambda_i + b \lambda_j = y \] uniquely for $a,b$. For pairs of distinct elements $\{\lambda_i, \lambda_j\}, \{\lambda_k, \lambda_\ell\} \in S_\tau^2$, we say that $\{\lambda_i, \lambda_j\}$ {\em conflicts} with $\{\lambda_k, \lambda_\ell\}$ if $\{\lambda_i, \lambda_j\} \neq \{\lambda_k, \lambda_\ell\}$ and $$((A_{i} + A_{j}) \times ( A_{i}\lambda_i + A_{j} \lambda_j)) \cap ((A_{i} + A_{j}) \times ( A_{i} \lambda_i + A_{j} \lambda_j)) \neq \emptyset.$$ In order to find a large set of distinct elements of $(A+A)\times(A+A)$, we will consider sums $((a, a\lambda_i) + (b, b\lambda_j)) \in ((A_i, A_i \lambda_i) + (A_j, A_j \lambda_j))$ for edges $\{\lambda_i, \lambda_j\}$ in a carefully selected graph $G \subset S_\tau^2$. We construct $G$ so that few pairs of edges conflict. For each pair of edges that does conflict, we apply an observation of Konyagin and Shkredov \cite{ks15} (adapted from the real to quaternionic case) to bound the number of elements of $(A+A) \times (A+A)$ that we count more than once. First we give the observation of Konyagin and Shkredov. Using the assumption on the additive energy of each $A_i$, we can bound the number of repeated sums coming from each pair of conflicting edges. \begin{lemma}\label{lem:intersectionSize} Let $\{\lambda_i, \lambda_j\}$ , $\{\lambda_k, \lambda_\ell\}$ be a pair of conflicting edges. Then, \begin{equation}\label{eq:conflictIntersection} |((A_{i} + A_{j}) \times ( A_{i} \lambda_i + A_{j} \lambda_j)) \cap ((A_{k} + A_{\ell}) \times ( A_{k}\lambda_k + A_{\ell} \lambda_\ell ))| \lesssim \tau^2 K^{-1/2}.\end{equation} \end{lemma} \begin{proof} Suppose, without loss of generality, that $\lambda_\ell$ is distinct from $\lambda_i$ and $\lambda_j$. Note that quadruples $(a_i, a_j, a_k, a_\ell) \in A_i \times A_j \times A_k \times A_\ell$ that contribute to (\ref{eq:conflictIntersection}) satisfy \begin{align} \label{eq:equalXCoord}a_i + a_j &= a_k + a_\ell, \\ \label{eq:equalYCoord} a_i \lambda_i + a_j \lambda_j &= a_k \lambda_k + a_\ell \lambda_\ell. \end{align} Subtracting equation $\ref{eq:equalXCoord}$ multiplied by $\lambda_\ell$ on the right from equation $\ref{eq:equalYCoord}$, we obtain \begin{equation}\label{eq:combinedSum} a_i(\lambda_i - \lambda_\ell) + a_j(\lambda_j - \lambda_\ell) = a_k (\lambda_k - \lambda_\ell). \end{equation} Hence, $|(\mathcal{A}_{\lambda_1} + \mathcal{A}_{\lambda_2}) \cap (\mathcal{A}_{\lambda_3} + \mathcal{A}_{\lambda_4})|$ is bounded above by the number $T$ of $(a_1,a_2,a_3) \in A_1 \times A_2 \times A_3$ that satisfy equation \ref{eq:combinedSum}. Applying Cauchy-Schwarz twice, we obtain \begin{align*} T &\leq |A_{3}|^{1/2} E_+^{1/2}(A_{1}(\lambda_1 - \lambda_4), A_{2}(\lambda_2-\lambda_4)),\\ &\leq |A_{3}|^{1/2} E_+^{1/4}(A_{1}) E_+^{1/4}(A_{2}). \end{align*} By construction, $|A_i| = \tau$ and $E^+(A_i) \approx \tau^3 K^{-1}$ for each $i$. Hence, \[T \lesssim \tau^2 K^{-1/2},\] as claimed. \end{proof} We also need to limit the number of pairs $(\lambda_i,\lambda_j),(\lambda_k, \lambda_\ell) \in G$ for which this intersection is non-empty. This involves controlling when such intersections can occur, and choosing $G$ accordingly. First, we establish that any ratio $(a_i + a_j)^{-1}(a_i \lambda_i + a_j \lambda_j)$ with $a_i \in A_i, a_j \in A_j$ must be close to $\lambda_i$. This is Lemma 7 in Solymosi and Wong \cite{sw17}. \begin{lemma}\label{lem:conflictBall} For any distinct $\lambda_i, \lambda_j \in S_\tau$ and $a_i \in A_i$, $a_j \in A_j$, we have $$\|(a_i + a_j)^{-1}(a_i \lambda_i + a_j \lambda_j) - \lambda_i\| \leq \|\lambda_j - \lambda_i\|.$$ \end{lemma} \begin{proof} Recall that $a_i$ and $a_j$ are in the same orthant, so $\|a_i+a_j\| \geq \|a_j\|$. Using the identity $(a_i+a_j)^{-1} = (1- (a_i+a_j)^{-1}a_j)a_i^{-1}$, we have \begin{align*} \|(a_i+a_j)^{-1}(a_i \lambda_i + a_j \lambda_j) - \lambda_i\| &= \| (1-(a_i+a_j)^{-1}a_j)\lambda_i + (a_i+a_j)^{-1}a_j \lambda_j - \lambda_i \|,\\ &= \|(a_i + a_j)^{-1} a_j (\lambda_j - \lambda_i) \|, \\ &= \|\lambda_j - \lambda_i\|\,\|a_j\|\,\|a_i+a_j\|^{-1}, \\ &\leq \| \lambda_j-\lambda_i \|. \end{align*} \end{proof} Now we define $G$. Let $M$ be a positive integer parameter that we will fix later. Let $G$ be the graph on vertex set $S_\tau$ formed by joining each $\lambda \in S_\tau$ to its $M$ closest neighbors, breaking ties arbitrarily. Since the degree of each vertex of $G$ is at least $M$, the number of edges in $G$ is at least $M |S_\tau|/2$. We also need to bound the number of conflicting pairs of edges in $G$. Let $B_i$ be the smallest closed ball centered at $\lambda_i$ that contains at least $M+1$ points of $S_\tau$ (including $\lambda_i$ itself), and let $R_i$ be the radius of $B_i$. While a given ball $B_i$ may contain an arbitrary number of points of $S_\tau$, the interior of each ball contains at most $M$ points. Note that, if $\{\lambda_i, \lambda_j\} \in G$, then at least one of $\lambda_i \in B_j$ or $\lambda_j \in B_i$. In addition, from Lemma \ref{lem:conflictBall}, we have that, for any distinct $\lambda_i, \lambda_k$, if there exist $\lambda_j, \lambda_\ell$ such that $\{\lambda_i,\lambda_j\}$ conflicts with $\{\lambda_k, \lambda_\ell\}$, then $B_i \cap B_k \neq \emptyset$. For each $\lambda_i$, let $L_i$ be the set of $\lambda_j$ such that $R_j \geq R_i$ and $B_i \cap B_j \neq \emptyset$. Given any $\lambda_i$, and $\lambda_j \in L_i$, the number pairs of edges $\{\lambda_i, \lambda_k\}, \{\lambda_j, \lambda_\ell\}$ that conflict is bounded above by the product of the degrees of $\lambda_i$ and $\lambda_j$. Hence, to bound the number of conflicting pairs of edges, it will suffice to place upper bounds on the degree of each vertex of $G$, and the size of each set $L_i$. This is done in Lemmas \ref{lem:degreeBound} and \ref{lem:conflictBound}, respectively. The proofs of both Lemma \ref{lem:degreeBound} and \ref{lem:conflictBound} both rely on similar geometric averaging arguments. In each case, for a fixed $\lambda_i$, we bound the number of interesting $\lambda_j$ in an arbitrary cone of constant size with vertex $\lambda_i$. By linearity of expectation, this gives an upper bound on the number of interesting $\lambda_j$ in all directions from $\lambda_i$. The following simple geometric lemma used in both proofs. \begin{lemma}\label{lem:vectorDiff} Let $\mathbf{x}, \mathbf{y}$ be arbitrary vectors in $\mathbb{R}^4$, and let $\mathbf{u}$ be a unit vector. Let $\mathbf{x}^\perp$ and $\mathbf{y}^\perp$ be the projections of $\mathbf{x}$ and $\mathbf{y}$ onto the subspace orthogonal to $\mathbf{u}$. Then, $$\| \mathbf{x} - \mathbf{y} \|^2 \leq \|\mathbf{x}\|^2 + \|\mathbf{y}\|^2 - 2(\mathbf{x}\cdot \mathbf{u})(\mathbf{y}\cdot \mathbf{u})+2\|\mathbf{x}^\perp\|\|\mathbf{y}^\perp\|.$$ \end{lemma} \begin{proof} \begin{align*} \|\mathbf{x} - \mathbf{y}\|^2 &= \|(\mathbf{x}\cdot \mathbf{u} - \mathbf{y} \cdot \mathbf{u})\mathbf{u} + \mathbf{x}^\perp - \mathbf{y}^\perp \|^2, \\ &= (\mathbf{x}\cdot \mathbf{u} - \mathbf{y} \cdot \mathbf{u})^2 + \|\mathbf{x}^\perp - \mathbf{y}^\perp\|^2, \\ &\leq (\mathbf{x}\cdot \mathbf{u} - \mathbf{y} \cdot \mathbf{u})^2 + (\|\mathbf{x}^\perp\| + \|\mathbf{y}^\perp\|)^2, \\ &= \|\mathbf{x}\|^2 + \|\mathbf{y}\|^2 - 2(\mathbf{x}\cdot \mathbf{u})(\mathbf{y}\cdot \mathbf{u})+2\|\mathbf{x}^\perp\|\|\mathbf{y}^\perp\|. \end{align*} \end{proof} \begin{lemma}\label{lem:degreeBound} For each $i$, the degree $\deg(\lambda_i)$ of $\lambda_i$ in $G$ is bounded by \[\deg(\lambda_i) \ll M.\] \end{lemma} \begin{proof} Take $\lambda_i$ to be at the origin. Let $\mathbf{u}$ be an arbitrary unit vector, and let $C$ be the cone $$C = \{\mathbf{x} : \mathbf{x} \cdot \mathbf{u} > (\sqrt{3}/2)\|\mathbf{x}\|\}.$$ We will use the decomposition $\mathbf{x} = (\mathbf{x} \cdot \mathbf{u})\mathbf{u} + \mathbf{x}^\perp$. Note that, if $\mathbf{x} \in C$, then $$\|\mathbf{x}^\perp\|^2 = \|\mathbf{x}\|^2 - (\mathbf{x} \cdot \mathbf{u})^2 < (1/4)\|\mathbf{x}\|^2.$$ Let $\mathbf{x}, \mathbf{y}$ be vectors in $C$ with $\|\mathbf{x}\|\geq\|\mathbf{y}\|$. By Lemma \ref{lem:vectorDiff}, \begin{align*} \| \mathbf{x} - \mathbf{y} \|^2 &\leq \|\mathbf{x}\|^2 + \|\mathbf{y}\|^2 - 2(\mathbf{x}\cdot \mathbf{u})(\mathbf{y}\cdot \mathbf{u})+2\|\mathbf{x}^\perp\|\|\mathbf{y}^\perp\|, \\ &< \|\mathbf{x}\|^2 + \|\mathbf{y}\|^2 - \|\mathbf{x}\|\|\mathbf{y}\|,\\ &\leq \|\mathbf{x}\|^2. \end{align*} To put this another way, the distance between any pair of points in $C$ is less than the distance from the more distant point to $\lambda_i$. Consequently, $\lambda_i$ has at most $M$ neighbors in $C$. Indeed, suppose for contradiction that there are $M+1$ neighbors in $C$, and let $\lambda$ be a neighbor at the largest distance from $\lambda_i$. Then, all of the remaining $M$ of the neighbors are closer to $\lambda$ than $\lambda_i$ is. This is a contradiction, since only the $M$ closest points to $\lambda$ are selected as neighbors. The preceding argument applies for an arbitrary unit vector $\mathbf{u}$. Suppose, for contradiction, that $\lambda_i$ has more than $cM$ neighbors for a sufficiently large constant $c$. Then, the expected number of neighbors in the cone corresponding to a uniformly random $\mathbf{u}$ will be greater than $M$, but we have shown that there cannot be any $\mathbf{u}$ for which this number is greater than $M$. \end{proof} Next we bound $|L_i|$ for an arbitrary $\lambda_i$. \begin{lemma}\label{lem:conflictBound} For each $i$, $$|L_i| \ll M.$$ \end{lemma} \begin{proof} Take $\lambda_i$ to be at the origin. Let $\mathbf{u}$ be an arbitrary unit vector, and let $C$ be the cone $$C = \{\mathbf{x} : \mathbf{x} \cdot \mathbf{u} > 0.99 \|\mathbf{x}\|\}.$$ Note that, if $\mathbf{x} \in C$, then $$\|\mathbf{x}^\perp\|^2 = \|\mathbf{x}\|^2 - (\mathbf{x} \cdot \mathbf{u})^2 < 0.02 \|\mathbf{x}\|^2.$$ If $\mathbf{x}, \mathbf{y}$ are vectors in $C$, then by Lemma \ref{lem:vectorDiff}, \begin{align} \| \mathbf{x} - \mathbf{y} \|^2 &\leq \|\mathbf{x}\|^2 + \|\mathbf{y}\|^2 - 2(\mathbf{x}\cdot \mathbf{u})(\mathbf{y}\cdot \mathbf{u})+2\|\mathbf{x}^\perp\|\|\mathbf{y}^\perp\|, \nonumber \\ &< \|\mathbf{x}\|^2 + \|\mathbf{y}\|^2 - 1.92\|\mathbf{x}\|\|\mathbf{y}\|. \label{eq:L_iDistance} \end{align} Define the following subsets of $C$: \begin{align*} C_1 &= \{\mathbf{x} \in C : \|\mathbf{x}\| < R_i\}, \\ C_2 &= \{\mathbf{x} \in C : R_i \leq \|\mathbf{x}\| < 1.5 R_i \}, \\ C_3 &= \{\mathbf{x} \in C : 1.5 R_i \leq \|\mathbf{x}\| \}. \end{align*} We will show that each of $C_1, C_2,$ and $C_3$ contains at most $M$ points of $L_i$. Since every point of $S_\tau$ that is at distance less than $R_i$ from $\lambda_i$ is a neighbor of $\lambda_i$, it is immediate that $C_1$ contains fewer than $M$ points of $S_\tau$. Let $\mathbf{x}, \mathbf{y}$ be vectors in $C_2$ with $R_i < \|\mathbf{y}\| \leq \|\mathbf{x}\| \leq 1.5 R_i$. By Equation~\eqref{eq:L_iDistance}, \begin{align*} \| \mathbf{x} - \mathbf{y} \|^2 &< \|\mathbf{x}\|^2 + \|\mathbf{y}\|^2 - 1.92\|\mathbf{x}\|\|\mathbf{y}\|, \\ &\leq \|\mathbf{x}\|^2 - 0.92 \|\mathbf{x}\| R_i, \\ &\leq 0.87 R_i. \end{align*} In other words, each pair of vectors in $C_2$ is at distance less than $R_i$. By assumption, if $R$ is the radius of the ball associated to any $\lambda \in L_i$, we have $R \geq R_i$. If there are $M+1$ points of $L_i$ in $C_2$, then the ball associated to each of these points contains all $M+1$ of the points in its interior, which contradicts the construction of the balls. Let $\mathbf{x}, \mathbf{y}$ be vectors in $C_3$ with $1.5 R_i < \|\mathbf{y}\| \leq \|\mathbf{x}\|$. We claim that \begin{equation}\label{eq:C3claim} \|\mathbf{x} - \mathbf{y} \| < (\|\mathbf{x}\| - R_i)^2. \end{equation} From (\ref{eq:L_iDistance}), in order to prove (\ref{eq:C3claim}), it is enough to show that \begin{align} \|\mathbf{y}\|^2 + 2\|\mathbf{x}\|R_i - 1.92 \|\mathbf{x}\| \|\mathbf{y}\| - R^2 < 0,& \label{eq:C3functToMaximize}\\ \text{under the assumption that} \hspace{1cm} \|\mathbf{x}\| \geq \|\mathbf{y}\| \geq 1.5 R_i.& \label{eq:C3constraint} \end{align} Considering the partial derivatives of the left hand side of (\ref{eq:C3functToMaximize}), we can see that this quadric attains its unique unconstrained maximum at $\|\mathbf{y}\| = (2/1.92)$, $\|\mathbf{x}\| = (2/1.92)^2$. Since this point is outside of the region defined by (\ref{eq:C3constraint}), it is enough to consider the boundaries of this region. When (\ref{eq:C3functToMaximize}) is constrained to the line $\|\mathbf{x}\|=\|\mathbf{y}\|$, we have $$-0.98\|\mathbf{x}\|^2 + 2\|\mathbf{x}\|R_i + R_i^2.$$ The derivative of this is negative for $\|\mathbf{x}\| \geq 1.5 R_i$. When (\ref{eq:C3functToMaximize}) is constrained to the line $\|\mathbf{y}\|=1.5R$, we have $$1.25R_i^2 - 0.88\|\mathbf{x}\|R_i.$$ Here, the derivative is always negative. Hence, it suffices to consider the point $\|\mathbf{x}\| = \|\mathbf{y}\| = 1.5 R_i$. At this point, the left hand side of (\ref{eq:C3functToMaximize}) is $-.07 R_i < 0$, as claimed. The meaning of (\ref{eq:C3claim}) is that, given any two points in $C_3$, the distance from the point that is further from $\lambda_i$ to the nearest point of $B_i$ is greater than the distance between the two points. Consequently, if $C_3$ contains $M+1$ points of $L_i$, the ball around the furthest point (which intersects $B_i$) must contain all $M+1$ points strictly in its interior, which is a contradiction. \end{proof} Let $T \subset G^2$ be the set of pairs of edges in $G$ that conflict. Applying Lemmas \ref{lem:degreeBound} and \ref{lem:conflictBound}, \begin{equation}\label{eq:conflictingPairsBound} |T| \leq \sum_i |L_i| \deg(\lambda_i) \ll |S_\tau| M^2.\end{equation} We have \[(A+A) \times (A+A) \supset \bigcup_{\{\lambda_i, \lambda_j\} \in G} (A_i + A_j)\times ( A_i \lambda_i + A_j \lambda_j).\] Using inclusion-exclusion and Lemma \ref{lem:intersectionSize}, this implies that \[|(A+A) \times (A+A)| \geq |G|\tau^2 - c|T|\tau^2 K^{-1}\] for some $c>0$. Since the degree of each $\lambda_i \in G$ is at least $M$, we have $|G| \geq |S_\tau|M/2$. Together with equation \ref{eq:conflictingPairsBound}, we get \[|(A+A)\times (A+A)| \geq |S_\tau|M \tau^2/2 - c |S_\tau| M^2 \tau^2 K^{-1/2}.\] Taking $M = c'K^{1/4}$ for a sufficiently small constant $c'$, we obtain \[|A+A|^2 \gg |S_\tau| \tau^2 K^{1/4} \gtrsim E^*(A) K^{1/4} \geq |A|^4 |AA|^{-1} K^{1/4} .\] Assuming that $K \gg |A|^{\delta}$ for some fixed $\delta > 0$, this shows that $|A+A|+|AA| \gtrsim |A|^{4/3 + \delta/12}$, as claimed. \section*{Acknowledgments} We would like to thank Oliver Roche-Newton for sharing his unpublished notes on Konyagin and Shkredov's sum-product bound. We would like to thank Adam Sheffer for extensive discussions on an earlier version of this paper. Work on this project by Ben Lund was supported by NSF grant DMS-1344994 (RTG in Algebra, Algebraic Geometry, and Number Theory, at the University of Georgia). \bibliographystyle{plain}

6660963acec6304e9f2891691c7d252ff1f76d41

\section{Introduction} \label{sec:intro} Typical domain adaptation methods are designed to transfer supervision from a single source domain. However, in many practical applications, we have access to multiple sources. For instance, in sentiment analysis of product reviews, we can often transfer from a wide range of product domains, rather than one. This can be particularly promising for target domains which do not match any one available source well. For example, the \textit{Kitchen} product domain may include reviews on \textit{pans}, \textit{cookbooks} or \textit{electronic devices}, which cannot be perfectly aligned to a single source such as \textit{Cookware}, \textit{Books} or \textit{Electronics}. By intelligently aggregating distinct and complementary information from multiple sources, we may be able to better fit the target distribution. A straightforward approach to utilizing data from multiple sources is to combine them into a single domain. This strategy, however, does not account for distinct relations between individual sources and the target example. Constructing a common feature space for this heterogeneous collection may wash out informative characteristics of individual domains and also lead to \textit{negative transfer} \cite{rosenstein2005transfer}. Therefore, we propose to explicitly model the relationship between different source domains and target examples. We hypothesize that different source domains are aligned to different sub-spaces of the target domain. Specifically, in this paper, we model the domain relationship with a mixture-of-experts (MoE) approach \cite{jacobs1991adaptive}. For each target example, the predicted posterior is a weighted combination of all the experts' predictions. The weights reflect the proximity of the example to each source domain. Our model learns this \textit{point-to-set} metric automatically, without additional supervision. We define the \textit{point-to-set} metric using Mahalanobis distance \cite{weinberger2009distance} between individual examples and a set (i.e. domain), which are computed within the hidden representation space of our model. The main challenge is to learn this metric in an unsupervised setting. We address it through a \textit{meta-training} procedure, in which we create multiple \textit{meta}-tasks of domain adaptation from the source domains. In each \textit{meta}-task, we pick one of the source domains as \textit{meta-target}, and the rest source domains as \textit{meta-sources}. By minimizing the loss using the MoE predictions on \textit{meta-target}, we are able to learn both the model and the metric simultaneously. To further improve transfer quality, we align the encoding space of our target and source domains via adversarial learning. We evaluate our approach on sentiment analysis using the benchmark multi-domain Amazon reviews dataset \cite{Chen2012msda,ziser2017nscl} as well as on part-of-speech (POS) tagging using the SANCL dataset \cite{petrov2012sancl}. Experiments show that our approach consistently improves the adaptation results over the best single-source model and a unified multi-source model. On average, we achieve a 7\% relative error reduction on the Amazon reviews dataset, and a 13\% on the SANCL dataset. Importantly, the POS tagging experiments on the SANCL dataset demonstrate that our method is able to robustly handle negative transfer from unrelated sources (e.g., \textit{Twitter}) and utilize it effectively to consistently improve performance. \section{Related Work} \label{sec:relwork} \paragraph{Unsupervised domain adaptation} Most existing domain adaptation methods focus on aligning the feature space between source and target domains to reduce the domain shift \cite{ben2007analysis,blitzer2007da,blitzer2006domain,pan2010cross}. Our approach is close to the representation learning approaches, such as the denoising autoencoder \cite{glorot2011domain}, the marginalized stacked denoising autoencoders \cite{Chen2012msda}, and domain adversarial networks \cite{tzeng2014deep,ganin2016domain,zhang2017aspect,Shen2018WassersteinDG}. In contrast to these previous approaches, however, our approach not only learns a shared representation space that generalizes well to the target domain, but also captures informative characteristics of individual source domains. \paragraph{Multi-Source domain adaptation} The main challenge in using multiple sources for domain adaptation is in learning domain relations. Some approaches assume that all source domains are equally important to the target domain \cite{li2008multi, luo2008transfer, crammer2008learning}. Others learn a global domain similarity metric using labeled data in a supervised fashion \cite{yang2007cross,duan2009domain,yu2018meta} or use predefined similarity measures \cite{ruder2017knowledge}. Alternatively, \newcite{mansour2009domain} and \newcite{bhatt2016cross} utilize unlabeled data of the target domain to find a distribution weighted combination of the source domains or to construct an auxiliary training set of the source domain instances close to the target domain instances. Recent adversarial methods on multi-source domain adaptation \cite{zhao2018multiple,chen2018multinomial} align source domains to the target domains globally, without accounting for the distinct importance of each source with respect to a specific target example. Beyond the global \textit{domain-to-domain} relations, \newcite{ruder2017transfer} use \textit{example-to-domain} similarity measures to select data from multiple sources for a specific target domain, which are then combined to train a single classifier. The work most related to ours is by \newcite{kim2017domain}. They also model the \textit{example-to-domain} relations, but use an attention mechanism. The attention module is learned using limited training data from the target domain in a supervised fashion. Our method, however, works in an unsupervised setting without utilizing any labeled data from the target domain. \section{Methodology} \label{sec:method} \paragraph{Problem definition} We follow the unsupervised multi-source domain adaptation setup, assuming access to labeled training data from $K$ source domains: $\{\mathcal{S}_i\}_{i=1}^{K}$ where $\mathcal{S}_i\triangleq\{(x_t^{\mathcal{S}_i}, y_t^{\mathcal{S}_i})\}_{t=1}^{\left|\mathcal{S}_i\right|}$, and (optionally) unlabeled data from a target domain: $\mathcal{T}\triangleq\{x_t^{\mathcal{T}}\}_{t=1}^{\left|\mathcal{T}\right|}$. The goal is to learn a model using the source domain data, that generalizes well to the target domain. \paragraph{Notations} For the rest of the paper, we denote an individual example as $x$, and a batch of examples as $\mathbf{x}$. We use superscript to denote the domain from which an example is sampled, and use subscript to denote the index of an example. \subsection{Overview of Our Approach} \input{figures/moe.tex} We model the multiple source domains as a mixture of experts, and learn a \textit{point-to-set} metric $\alpha$ to weight the experts for different target examples. The metric is learned in an unsupervised manner. Our model consists of four key components as shown in Figure \ref{fig:moe}, namely the \textit{encoder} ($E$), \textit{classifier} ($F$), \textit{metric} ($M$) and \textit{adversary} ($D$). We use a typical neural multi-task learning architecture \cite{caruana1997multitask}, with a shared encoder across all sources, and domain-specific classifiers ($\{F^{\mathcal{S}_i}\}_{i=1}^K$). Each input is first encoded with $E$, and then fed to each classifier to obtain the domain-specific predictions (i.e. posteriors). The final predictions are then weighted based on the \textit{metric} (see Equation \ref{eq:1}). We start by describing the representation learning component. \subsection{Representation} \label{sec:repr} Our goal is to design an encoder that supports transfer, while maintaining source domain-specific information. Depending on different tasks and datasets, we select appropriate encoders --- \textit{multilayer perceptron} (MLP), \textit{convolutional neural network} (CNN) or \textit{long short-term memory networks} (LSTM) (see Section \ref{subsec:implement} for details). We further add an adversarial module ($D$) on top of the encoder, in order to align the target domain with the sources. $D$ is typically designed as a parameterized classifier in domain adversarial networks \cite{ganin2016domain,zhang2017aspect}, which is trained jointly with the encoder and the classifiers through a minimax game. Here, we instead use Maximum Mean Discrepancy (MMD) \cite{gretton2012kernel} as our adversary. This distance metric measures the discrepancy between two distributions explicitly in a non-parametric manner, greatly simplifying the training procedure compared to domain adversarial networks which use an additional domain classifier module. \subsection{Mixture of Experts} \label{sec:moe} Given an example $x$ from the target domain, we model its posterior distribution as a mixture of posteriors produced by models trained on different source domain data: \begin{equation} \begin{split} p_{moe}(y|x) & = \sum_{i=1}^{K} \alpha (x, \mathcal{S}_i)\cdot p^{\mathcal{S}_i}(y|x) \\ & = \sum_{i=1}^{K} \alpha (x, \mathcal{S}_i)\cdot \mathtt{softmax}\big(\mathbf{W}^{\mathcal{S}_i} E(x)\big) \end{split} \label{eq:1} \end{equation} $p^{\mathcal{S}_i}$ is the posterior distribution produced by the $i^{th}$ source classifier $F^{\mathcal{S}_i}$ (the $i^{th}$ expert). $\mathbf{W}^{\mathcal{S}_i}$ is the output layer weights of $F^{\mathcal{S}_i}$, $\alpha$ is a parameterized metric function that measures how much confidence we put in the specific source expert for a given example $x$.\footnote{In typical MoE frameworks \cite{jacobs1991task,jacobs1991adaptive,shazeer2017outrageously}, $\alpha$ is commonly realized as a ``gating network", which produces a normalized weight vector that determines the combination of experts depending solely on the input example. Such gating networks, however, do not yield promising results in our scenario. We hypothesize that both the input example and the underlying domain distribution should be captured for determining the credit assignment.} To derive $\alpha$, we first define a \textit{point-to-set} Mahalanobis distance metric between an example $x$ and a set $\mathcal{S}$: \begin{equation*} d(x, \mathcal{S}) = \Big( \big(E(x) - \mu^\mathcal{S}\big)^\top \mathbf{M}^{\mathcal{S}}\big(E(x) - \mu^\mathcal{S}\big) \Big)^{\frac{1}{2}} \end{equation*} where $\mu^\mathcal{S}$ is the mean encoding of $\mathcal{S}$. In its original form, the matrix $\mathbf{M}^\mathcal{S}$ played the role of the inverse covariance matrix. However, computing the inverse of the covariance matrix is both time consuming and numerically unstable in practice. Here we allow $\mathbf{M}$ to denote any positive semi-definite matrix which is to be estimated during training \cite{weinberger2009distance}. To guarantee the positive semi-definiteness of $\mathbf{M}$, we approximate $\mathbf{M}$ with $\mathbf{M} = \mathbf{U} \mathbf{U}^\top$, where $\mathbf{U}\in \mathbb{R}^{h\times r}$, $h$ is the dimension of hidden representations and $r$ is a hyper-parameter controlling the rank of $\mathbf{M}$. Based on the distance metric, we further derive a confidence score $e(x, \mathcal{S}_i)=f\big(d(x, \mathcal{S}_i)\big)$ for each specific expert. The final metric values $\alpha(x, \mathcal{S}_i)$ are then obtained by normalizing these confidence scores: \begin{equation} \alpha (x, \mathcal{S}_i) = \frac{\exp\big(e(x, \mathcal{S}_i)\big)}{\sum_{j=1}^{K}\exp\big(e(x, \mathcal{S}_j)\big)} \end{equation} Here, we explain our design of $e(x, \mathcal{S})$ on two tasks, respectively \textit{binary classification} and \textit{sequence tagging}, which are also used for evaluation in this paper (Section \ref{sec:exp-setup}). \paragraph{Binary classification} The \textit{point-to-set} Mahalanobis distance metric measures the distance between an example $x$ and the mean encoding of $\mathcal{S}$, i.e. $\mu^\mathcal{S}$, while taking into account the (pseudo) covariance of $\mathcal{S}$. In binary classification, however, the mean vector $\mu^\mathcal{S}$ is likely to be located near the decision boundary, particularly under a balanced setting. Therefore, a small $d(x, \mathcal{S})$ actually implies lower confidence of the corresponding classifier, which is counter-intuitive. To this end, we instead define the confidence $e(x, \mathcal{S})$ as the difference between the distances from $x$ to each category of $\mathcal{S}$, referred to as \textit{Maximum Cluster Difference} in \newcite{ruder2017knowledge}: \begin{equation*} e(x, \mathcal{S}) = \left|d(x, \mathcal{S}^{+}) - d(x, \mathcal{S}^{-})\right| \end{equation*} Here $\mathcal{S}^+$ and $\mathcal{S}^-$ stand for the positive space and negative space of $\mathcal{S}$ respectively. Consequently, if $x$ is either far away from $\mathcal{S}$ (i.e., $x$ is not in the manifold of $\mathcal{S}$) or near the classification boundary, we will get a small $e(x, \mathcal{S})$ indicating a low confidence to the corresponding prediction. On the contrary, if $x$ is much closer to a specific category of $\mathcal{S}$ than other categories, the classifier will get a higher confidence. \paragraph{Sequence tagging} For sequence tagging tasks (e.g., POS tagging), we compute the distance metric at the \textit{token level}.\footnote{This actually makes it a multi-class classification problem with respect to every token of a sequence.} Unlike in \textit{binary classification}, the decision boundary here is more complicated, and the label distribution is typically imbalanced. The mean vector $\mu^\mathcal{S}$ is unlikely to be located at the decision boundary. So we directly use the (reverse) distance as the confidence value for each token $x$: \begin{equation*} e(x, \mathcal{S}) = -d(x, \mathcal{S}) \end{equation*} \subsection{Training} \label{subsec:train} Since we do not have annotated data in the target domain, we have to learn our model in an unsupervised fashion. Inspired by the recent progress on few-shot learning with metric-based models such as matching network \cite{vinyals2016matching,yu2018meta} and prototypical network \cite{snell2017prototypical}, we propose the following \textit{meta-training} approach. Given $K$ source domains, each source domain will be considered as a target, referred to as \textbf{\textit{meta-target}}, with the rest of the source domains as \textbf{\textit{meta-sources}}. This way, we obtain $K$ (\textit{meta-sources}, \textit{meta-target}) training pairs for domain adaptation. Then, we apply our MoE formulation over these \textit{meta-training} pairs to learn the metric. At testing time, the metric will be applied to all the $K$ source domains for each example in the target domain. We optimize two main objectives: the MoE objective and the multi-task learning (MTL) objective. \paragraph{MoE objective} For each example in each \textit{meta-target} domain, we compute its MoE posterior using the corresponding \textit{meta-sources}. Therefore, we get the following MoE loss over the entire multi-source training data: \begin{equation} \begin{split} \mathcal{L}_{moe} & = -\sum_{i=1}^{K}\sum_{j=1}^{\left|\mathcal{S}_i\right|}\log p_{moe}(y^{\mathcal{S}_i}_{j}|x^{\mathcal{S}_i}_{j}) \\ & = -\sum_{i=1}^{K}\sum_{j=1}^{\left|\mathcal{S}_i\right|}\log \sum_{l=1, l\ne i}^{K}\alpha(x, \mathcal{S}_l)\cdot p^{\mathcal{S}_l}(y^{\mathcal{S}_i}_{j}|x^{\mathcal{S}_i}_{j}) \end{split} \end{equation} Note that $\alpha$ is normalized over the \textit{meta-sources} for each \textit{meta-target}, rather than over all the $K$ sources. \paragraph{MTL objective} For each \textit{meta-target}, we further optimize a supervised cross-entropy loss using the corresponding labels. All supervised objectives are optimized jointly with the encoder being shared, resulting in the following multi-task learning objective: \begin{equation} \mathcal{L}_{mtl}=-\sum_{i=1}^{K}\sum_{j=1}^{\left|\mathcal{S}_i\right|}\log p^{\mathcal{S}_i}(y^{\mathcal{S}_i}_{j}|x^{\mathcal{S}_i}_{j}) \end{equation} \input{algorithms/training.tex} \paragraph{Adversary-augmented MoE} We use MMD \cite{gretton2012kernel} as the adversary to minimize the divergence between the marginal distribution of target domain and source domains. Specifically, at each training epoch, given the $K$ batches $\{\mathbf{x}^{\mathcal{S}_1},\mathbf{x}^{\mathcal{S}_2},...,\mathbf{x}^{\mathcal{S}_{K}}\}$ from all the source domains, we sample a batch (unlabeled) $\mathbf{x}^\mathcal{T}$ from our target domain, and minimize the MMD: \begin{equation} \mathcal{L}_{adv} = \mathtt{MMD}^2( \mathbf{x}^{\mathcal{S}_1} \cup ... \cup \mathbf{x}^{\mathcal{S}_{K}},\mathbf{x}^\mathcal{T}) \end{equation} where \begin{equation*} \begin{split} &\mathtt{MMD}(\mathcal{D}^\mathcal{S}, \mathcal{D}^\mathcal{T}) = \\ &\left\Vert \frac{1}{|\mathcal{D}^\mathcal{S}|}\sum_{x_s \in \mathcal{D}^\mathcal{S}}\phi\big(E(x_s)\big) - \frac{1}{|\mathcal{D}^\mathcal{T}|}\sum_{x_t \in \mathcal{D}^\mathcal{T}}\phi\big(E(x_t)\big)\right\Vert_{\mathcal{H}} \end{split} \end{equation*} measures the discrepancy between $\mathcal{D}^\mathcal{S}$ and $\mathcal{D}^\mathcal{T}$ based on Reproducing Kernel Hilbert Space (RKHS). $\phi(\cdot)$ is the feature map induced by a universal kernel. We follow \newcite{bousmalis2016domain} and use a linear combination of multiple RBF kernels: $\kappa(\bm{h}_i, \bm{h}_j) = \Sigma_n \exp(-\frac{1}{2\sigma_n}\left\Vert \bm{h}_i - \bm{h}_j\right\Vert^2)$. \paragraph{Entropy regularization} In the \textit{meta-training} process, for each example $x$ in \textit{meta-target}, we know exactly from which source $x$ is sampled. This provides additional insight that the $\alpha$ distribution is skewed, which can be utilized as a soft constraint. Therefore, we propose to regularize the entropy of the $\alpha$ distribution over all the sources, rather than \textit{meta-sources}:\footnote{Alternatively, we can directly exploit this supervision and minimize the KL divergence of the $\alpha$ distribution and its ground truth one-hot distribution. In practice, however, we found it beneficial to allow examples from one domain to be attended to different sources. This observation may be attributed to the fact that each domain indeed consists of multiple latent sub-domains.} \begin{equation*} H\big(\bm{\alpha}(x, \cdot)\big) = -\sum_{l=1}^{K}\mathbf{\alpha}(x, \mathcal{S}_l) \cdot \log \mathbf{\alpha}(x, \mathcal{S}_l) \end{equation*} \begin{equation} \mathcal{R}_h = \sum_{i=1}^{K}\sum_{j=1}^{\left|\mathcal{S}_i\right|}H\big(\bm{\alpha}(x^{\mathcal{S}_i}_{j}, \cdot)\big) \end{equation} \paragraph{Joint learning} Our final objective is the weighted combination of each individual component loss: \begin{equation} \begin{split} \mathcal{L} = &~ \lambda \cdot \mathcal{L}_{moe} + (1 - \lambda) \cdot \mathcal{L}_{mtl} \\ & + \gamma \cdot \mathcal{L}_{adv} \\ & + \eta \cdot \mathcal{R}_{h} \end{split} \end{equation} where $\lambda$ controls the balance of the MoE loss and MTL loss. $\gamma$ is set to 0 in \textit{non-adversarial} setting when unlabeled data from the target domain is not provided. Additionally, it would be straightforward to add an MoE loss for labeled data in the target domain if they are available, thus extending our framework to a setting where we have few-shot target annotations. The training process is shown in Algorithm 1. \section{Experimental Setup} \label{sec:exp-setup} \subsection{Task and Dataset} \paragraph{Sentiment classification} We use the multi-domain Amazon reviews dataset \cite{blitzer2007da}, one of the standard benchmark datasets for domain adaptation. It contains reviews on four domains: \textit{Books} (B), \textit{DVDs} (D), \textit{Electronics} (E), and \textit{Kitchen appliances} (K). We follow the specific experiment settings proposed by \newcite{Chen2012msda} (\textsc{Chen12}) and \newcite{ziser2017nscl} (\textsc{Ziser17}). \begin{enumerate} \item In \textsc{Chen12}, each domain has 2,000 labeled examples for training (1,000 positive and 1,000 negative), and the target test set has 3,000 to 6,000 examples.\footnote{This dataset has been processed by the author to TF-IDF representations, using the 5,000 most frequent unigram and bigram features, thus word order information is not available.} \item In \textsc{Ziser17}, each domain also has 2,000 labeled examples (1,000 positive and 1,000 negative), sampled differently from \textsc{Chen12} \end{enumerate} For each dataset, we conduct experiments by selecting the target domain in a round-robin fashion. Following the protocol in previous work, we use cross-validation over source domains for hyper-parameters selection for each adaptation task \cite{zhao2018multiple}. When training with an adversary, we use the 2,000 examples training set of the target domain as the unlabeled data in both the settings. In \textsc{Ziser17}, the same data is also used for test, resulting in a transductive setting. \paragraph{Part-of-Speech tagging} We further consider a sequence tagging task, where the metric is computed over the token-level encodings and multi-class predictions are made at the token (word) level. We use the SANCL dataset \cite{petrov2012sancl} which contains part-of-speech (POS) tagging annotations in 5 web domains: \textit{Emails}, \textit{Weblogs}, \textit{Answers}, \textit{Newsgroups}, and \textit{Reviews}. Among these, \textit{Newsgroups}, \textit{Reviews}, and \textit{Answers} have both a validation and a test set, and are used as target domains. The test set from \textit{Weblogs} and \textit{Emails} are used as individual source domains. The tagging is performed using the Universal POS tagset \cite{petrov2011universal}. We also use \textit{Twitter} \cite{liu2018parsing} as an additional training source. Since it differs substantially from other sources and the target domain, we can assess our model's ability to handle negative transfer. We consider 750 sentences from each SANCL source domain for training, and up to 2,250 sentences from the \textit{Twitter} dataset to magnify the negative transfer. The validation set in the standard split of each target domain is used for hyper-parameters selection and early-stopping in our experiments \subsection{Baselines} We verify the efficacy of our approach (MoE) in \textit{non-adversarial} and \textit{adversarial} settings respectively. In both settings, we compare our approach against the following two baselines: \begin{itemize} \item \textbf{best-SS}: the best single-source adaptation model among all the sources. \item \textbf{uni-MS}: the unified multi-source adaptation model, which is trained using the combination of all the source domain data with single-source transfer methods. uni-MS is a common and strong baseline for multi-source domain adaptation \cite{zhao2018multiple}. \end{itemize} For the rest of the paper, we name the adversarial counterpart of the models as \textbf{$*$-A}. In the adversarial setting on \textsc{Chen12}, in addition to best-SS and uni-MS with adversarial loss, we further compare with the following two systems that also utilize unlabeled data from target domain. \begin{itemize} \item \textbf{mSDA}: the marginalized stacked denoising autoencoder \cite{Chen2012msda}. mSDA outperforms prior deep learning and shallow learning approaches such as structural correspondence learning \cite{blitzer2007da} and denoising autoencoder \cite{glorot2011domain}. \item \textbf{MDAN}: the multi-source domain adversarial network \cite{zhao2018multiple}. MDAN gives the state-of-the-art performance for multi-source domain adaptation on \textsc{Chen12}. It generalizes the domain adversarial network to multiple source domain adaptation by selectively backpropagating the domain discrimination loss according to domain classification error. \end{itemize} \subsection{Implementation Details} \label{subsec:implement} For \textsc{Chen12}, since the dataset is in TF-IDF format and the word ordering information is not available, we use a multilayer perceptron with an input layer of 5,000 dimensions and one hidden layer of 500 dimensions as our encoder. For \textsc{Ziser17}, we instead use a convolutional neural network encoder with a combination of kernel widths 3 and 5 \cite{kim2014cnn}, each with one hidden layer of size 150, which are then concatenated to a 300 dimension representation.\footnote{Note that with a more extensive architecture search, we are likely to achieve better results. This, however, is not the main focus of this work.} For the POS tagging encoder, we use a hierarchical bidirectional LSTM (BiLSTM) network, which contains a character-level BiLSTM for generating individual word representations, followed by a word-level BiLSTM that generates contextualized word representations. For MMD, we follow \newcite{bousmalis2016domain} and use 19 RBF kernels with the standard deviation parameters ranging from $10^{-6}$ to $10^6$.\footnote{Detailed values are presented in the supplementary material in \newcite{bousmalis2016domain}.} \input{tables/amazon-results.tex} \input{figures/metric.tex} All the models were trained using Adam with weight decay. Learning rate is set to $10^{-4}$ for \textsc{Chen12} and $10^{-3}$ for \textsc{Ziser17} and POS tagging. We use mini-batches of 32 samples from each domain. We tune the coefficients $\lambda$, $\eta$ for each adaptation task. $\gamma$ is set to 1 for all experiments. \section{Results} \subsection{Sentiment Analysis on Amazon Reviews} We report our results on the Amazon reviews datasets in Table \ref{tab:result-chen12} (\textsc{Chen12}) and Table \ref{tab:result-ziser17} (\textsc{Ziser17}). Our approach (MoE) consistently achieves the best performance across different settings and tasks. The results clearly demonstrate the value of using multiple sources. In most cases, even a unified model performs better than the oracle best single source. By smartly combining all the sources, our model outperforms the unified model significantly. One exception is the task of \textquotedblleft B,D,K-E\textquotedblright~in \textsc{Chen12}, where the unified multi-source model doesn't improve over the best single source model, constituting a \textit{negative transfer} scenario. However, even in this scenario, our approach still performs significantly better, demonstrating its robustness in handling \textit{negative transfer}. \paragraph{Impact of adversarial adaptation} We achieve consistent improvements over the baseline systems with the addition of the adversarial loss. In most cases, MoE also achieves additional improvement (e.g., 79.42\% \textit{vs.} 80.87\% in ``D,E,K-B"). We notice that in some cases, e.g., ``B,D,K-E" in \textsc{Chen12} and ``B,E,K-D" in \textsc{Ziser17}, the adversarial loss doesn't help MoE. This might be attributed to the fact that by aligning the target distribution with the source domains, the representation space becomes more compact, thus making it more difficult to capture source domain-specific characteristics and increasing the difficulty of metric learning in MoE \paragraph{Analysis on the metric ($\alpha$)} Figure~\ref{fig:alpha-amazon} visualizes the distribution of $\alpha$ values, learned by our model in different tasks, across the source domains. The visualization is based on 200 examples for each domain randomly sampled from the corresponding test set. From the heatmap we can see that for a specific target domain, different examples may have different $\alpha$ distributions. Moreover, for most examples, the $\alpha$ distribution is skewed, indicating that our model draws on a few most informative source domains. Figure~\ref{fig:alpha-example} exemplifies the above point. For instance, the first review about ``charger" and ``battery" is closer to the \textit{Electronics} source domain. This relation is successfully captured by the $\alpha$ distribution produced by our model. \input{tables/amazon-entropy-reg.tex} We further investigate the impact of entropy regularization over $\alpha$. Table \ref{tab:ablation-entropy} summarizes the ablation test results of entropy regularization ($\eta = 0$) on \textsc{Chen12} and \textsc{Ziser17}. It shows that entropy regularization benefits our model under both non-adversarial and adversarial settings. \input{tables/sancl-results.tex} \subsection{Part-of-Speech Tagging} \label{subsec:pos-results} Table \ref{tab:pos-tagging} summarizes our results on POS tagging. Again, our approach consistently achieves the best performance across different settings and tasks. Adding \textit{Twitter} as a source leads to a drop in performance for the unified model, as a result of negative transfer. Our method, however, robustly handles negative transfer and manages to even benefit from this additional source. \paragraph{Impact of negative transfer} Table \ref{tab:pos-tagging-alphas} presents the $\alpha$ distribution learned by the metric, on average for all tokens of the target domain. As we can see, our model (MoE-A) effectively learns to decrease the weights on \textit{Twitter}, demonstrating again its ability to alleviate negative transfer. \input{tables/sancl-alphas.tex} We further study the impact of this outlier source by varying the amount of \textit{Twitter} data used during training. We gradually increase the number of \textit{Twitter} instances by 750. As shown in Table~\ref{tab:pos-varying-twitter}, the increase of the \textit{Twitter} data does not benefit the unified multi-source model (uni-MS-A), and even amplifies negative transfer for the \textit{Answers} and \textit{Reviews} domains. However, the performance of our MoE (MoE-A) model stays stable, consistently increasing with more \textit{Twitter}, showing robustness in handling negative transfer. \input{tables/sancl-varying-twitter.tex} \section{Conclusion} In this paper, we propose a novel mixture-of-experts (MoE) approach for unsupervised domain adaptation from multiple diverse source domains. We model the domain relations through a \textit{point-to-set} distance metric, and introduce a \textit{meta-training} mechanism to learn this metric. Experimental results on sentiment classification and part-of-speech tagging demonstrate that our approach consistently outperforms various baselines and can robustly handle negative transfer. The effectiveness of our approach suggests its potential application to a broader range of domain adaptation tasks in NLP and other areas. \section*{Acknowledgments} We thank MIT NLP group and the anonymous reviewers for their helpful comments. We also thank Shiyu Chang and Mo Yu for insightful discussions on metric learning. This work is supported by the MIT-IBM Watson AI Lab. Any opinions, findings, conclusions, or recommendations expressed in this paper are those of the authors, and do not necessarily reflect the views of the funding organizations.

5ef8443fd472eb741565ca287fdac012c2ea9769

\section{Introduction} Many problems plague big cities, including transportation, pollution, security, and public safety. With the proliferation of mobile devices and availability of more datasets, there is great potential to revisit these problems and introduce more efficient solutions. As a start, in this study, we shall focus on shared transportation as a solution to congestion problems in big cities. Recent advances in mobile technology enabled the creation and rise of transportation network companies, like Uber and Lyft. Although shared transportation is not new, these companies were able to realize them at scale. Unfortunately, these companies, in order to maintain their competitive edge, do not disclose much of their data or analysis hence we investigate the matter with open datasets. The approach to the study of human mobility depends on the spatial scale of the target varying from individual mobility to higher levels (buildings, city blocks, etc.). As vehicular transportation is the quintessence of urban mobility we focus on the study of trips happening in a city. A common representation of trips is through their origins and destination. In this paper, we use origin-destination (OD) representation to study the trips and investigate various scenarios of shared transportation. A comparison of similarity/distance scores is then presented using a more detailed representation of trips based on a sample of 50 waypoints (spatio-temporal triples of x,y,t) for each trip (instead of OD - endpoints only). Although similarity of trajectories has long been a subject of interest, the concept of time and its effects are not well-studied with respect to the trajectories. We approach this problem using spatio-temporal similarity of waypoints of the trips. Public transportation is a big part of large cities around the world, but fine-grained shared mobility has received less attention. Recently, it has been seen for its potentials in the reducing travel times and traffic leading to increased productivity, less fuel consumption, and less pollution. In this paper, we use our proposed similarity measure to formulate and investigate three scenarios of shared transportation. This study provides the following main contributions: \begin{enumerate*} \item Analysis of trips data from a spatio-temporal perspective, \item Proposing a measure of spatio-temporal similarity of trajectories (trips) with flexibility across spatial or temporal elements. A comparison is conducted of various measures available in the literature, \item Study of various shared transportation scenarios of Catch-a-Ride and Car Pooling. In addition, a novel graph formulation of a free float Car-Sharing scenario is presented. \end{enumerate*} The study finds that trip distance and duration follow lognormal and gamma distributions (with gamma providing a better fit), and demonstrates the utility of the score and spectral clustering in finding spatio-temporal clusters. Similarity measures available in the literature have a quadratic computation complexity (in the number of waypoints), while our score has linear complexity with the flexibility to prefer time over space or vise versa. Shared transportation achieves roughly 25\% decrease in distances traveled by cars. The Car-Sharing scenario is optimal in the number of cars (minimum) and also the consecutive pick-ups in terms of combined temporal and spatial distances with the algorithm being bounded by the complexity of the matching in bipartite graphs ($O(n^3)$ in the number of nodes - in our case number of trips). The rest of the paper is structured as follows: Related work in Sec. \ref{sec:relatedwork} followed by exploratory data analysis in Sec. \ref{sec:analysis}, and definition of similarity score and its potential in finding spatio-temporal clusters in Sec. \ref{sec:similarity} and \ref{sec:clustering}. The score is used for dynamic matching of trips for Catch-a-Ride and Car Pooling scenarios of shared transportation in Sec. \ref{sec:matching}. Sec. \ref{sec:compare} presents a comparison of various measures of similarity. Formulation of the Car Sharing problem and the algorithm is presented in Sec. \ref{sec:carsharing}. Lastly, Sec. \ref{sec:conclusion} concludes the study. \vspace{-5pt} \section{Related Work} \label{sec:relatedwork} We explore the literature from three perspectives: \begin{enumerate*} \item The study of trips for their characteristics in time and space. \item The similarity of trips using the trajectory of movements. \item The concept of shared transportation and its variations. \end{enumerate*} Several studies use publicly available taxicab datasets to characterize trips. \citet{Ferreira2013} analyzed a large dataset of taxicab trips (170M trips a year, for 2 years) of NYC for visualization of multiple use cases such as activity in different regions. Users trips have been extracted from geo-tagged images and studied for patterns of visitations \cite{arase2010mining}. A recent study focused on the travel patterns arising from Internet-based ride-sharing platforms using data from DiDi, China's biggest ride-hailing company \cite{Dong2018}. They found two types of individual behavior: commuters and taxi-like roaming. In a study of spatial patterns based on the origin-destination representation of trips, \citet{Guo2012} proposed a distance measure based on the number of shared nearest neighbors and used it to find spatial clusters of the trips. We follow a similar approach but we define the similarity score and clustering based on temporal aspect in addition to space. \citet{Veloso2011} studied a taxi dataset for Lisbon, Portugal. They visualized and explored various spatio-temporal features including distributions of trip durations, distances, pickups, and drop-offs. Uber and Lyft trips in San Francisco were studied and visualized by SF County Transportation Authority with respect to location and time \cite{sfTNC}. Trip trajectories have been studied for storage and performance optimizations, as well as traffic insights. The similarity of trips bounded by traffic network was studied in \cite{Abraham2012}, using spatial and temporal points of interest and the distance of the trip to them. \citet{tiakas2006trajectory} defined spatial similarity between two trajectories as the average point-wise distance of points in a trajectory and temporal similarity as the sum of differences of consecutive points relative to the maximum of both. \citet{Toohey2015} performed a comparative study of various similarity measures. They consider measures of longest common subsequence (LCSS), Frechet distance, dynamic time warping (DTW) and edit distance for trajectories of delivery drivers in the UK. These metrics are detailed in section \ref{sec:compare} for comparison. In a similar study, \citet{Wang2013} investigated the effectiveness of various similarity scores on multiple transformations of trajectories, with parameters controlling the amplitude of change in the trips. \citet{VanKreveld2007} studied the similarity of trajectories and subtrajectories considering time shifts, to find most similar trajectories. They propose various polynomial time algorithms; $O(n^4)$ for the most sophisticated case with time shifts and subtrajectories. \citet{yan2017itdl} proposed a method of generating embeddings of Points of Interest (POI) with a spatial context that can be used to measure the similarity of place types. This has potential applications in the matching of rides and profiling trajectories. Concerning rapidly expanding ride-hailing networks, \citet{Tian2013} presented a system of dynamic ride-sharing called Noah where three algorithms of matching are considered. Minimizing traveled distances or times along with maximizing number of matches are common non-commercial objectives of dynamic ride-sharing systems \cite{AGATZ2012295,DiFebbraro2013}. \citet{Gidofalvi2008} studied the social barriers of shared mobility and proposed a system of ride sharing that incorporates social relationships in the matching criteria. \citet{Masoud2017} studied another variant of ride-sharing, assuming a p2p (similar to carpooling) and flexible (multi-hop with transfers) scenario and provides an optimal solution. \citet{Ma2013} propose a framework of large-scale dynamic ride-sharing for taxis and evaluate it based on Beijing trips. \citet{Zhang2014carpool} present a scalable framework of ride sharing with a focus for carpool capabilities (i.e. a taxicab with a passenger and a route, picking up another passenger for a similar trip) and evaluate it on trips in the city of Shenzhen. They show that their system, \textit{CallCab}, can reduce the total traveled distances by 60\% and wait times by 41\%. \citet{khan2017ride} posit that ride-sharing is more effective when users can go out of their way slightly and also agree on common destinations. A less studied scenario of shared transportation is car sharing. \citet{Katzev2003} studied effects of car sharing on mobility and environmental impacts. Several similar, recent studies have investigated mobility and impacts of car sharing \cite{Baptista2014,Nijland2017,Musso2012}. \citet{HANNA2016254} formulated the car sharing problem as a bipartite graph and solved the matching for three criteria: min cost, min makespan (limit of distance for matches), and strategic manipulability (no user has the incentive to misreport its location and hence distance). In \cite{agatz2016autonomous}, authors discuss several challenges of realizing car and ride sharing, including optimization of multiple trade-offs (number of vehicles, convenience, total travel time, operating costs). This new concept of car sharing and ownership, coupled with the rise of autonomous and automated transportation systems (e.g. self-driving cars) might change the face of mobility in urban environments. In this paper, we present a novel formulation for scheduling a car sharing/automated fleet of vehicles with the minimum number of cars. \begin{figure}[!t] \subcaptionbox{8-9am distance}{% \includegraphics[width=0.49\columnwidth]{images/trip_distance_dist_8}% } \subcaptionbox{4-5pm distance}{% \includegraphics[width=0.49\columnwidth]{images/trip_distance_dist_16}% } \subcaptionbox{8-9am duration}{% \includegraphics[width=0.49\columnwidth]{images/trip_duration_dist_8}% } \subcaptionbox{4-5pm duration}{% \includegraphics[width=0.49\columnwidth]{images/trip_duration_dist_16}% }% \caption{Distribution of duration and traveled distance of trips in different times of day.} \label{fig:distr_dist_dur} \end{figure} \section{Dataset and Analysis}\label{sec:analysis} The dataset used for this study is mobility traces for the city of Cologne \cite{kolntmc}. The data was collected as part of an initiative to realistically reproduce vehicular mobility in the city\footnote{http://kolntrace.project.citi-lab.fr/}. We choose this dataset as it provides us with fine-grained traces of trips in a long timespan of 24hours in a major city. This dataset contains traces consisting of location, time and speed of vehicles on a trip (700K trips, 350M records in a 400 $km^2$ area). To analyze different times of the day and understand temporal dynamics, we select four 1-hour time windows throughout the day (Morning 8-9am, Noon 12-1pm, Evening 4-5pm, Night 8-9pm). For the OD representation of the trips, the first and last appearances of each ID are extracted. \subsection{Exploratory Analysis of Spatial and Temporal Characteristics} We start with an exploratory analysis of the dataset in order to better understand the vehicular mobility through the characteristics that pertain to the trips. We comparatively investigate (different times of the day) the trip duration and distance distributions in \ref{sssec:distr} (more accurately the distance between origin and destination represents the displacement and we use them interchangeably), number of trip waypoints (records in data) in \ref{sssec:wp}, spatial spread of the trips in \ref{sssec:spatial}, temporal distributions (in \ref{sssec:temporal}) and the relation of trip duration to trip start location in \ref{sssec:relation}. \subsubsection{Duration and Distance Distribution}\label{sssec:distr} Figure \ref{fig:distr_dist_dur} shows the distribution of trip duration and distance in each hour of the day. As observable, the distribution in all cases is right-skewed. Since one common distribution to model the trip duration is log-normal \cite{enroute,dandy1984variability,rakha2010trip,arroyo2004modeling}. The fit of lognormal is plotted on the histograms and compared to Gamma distribution which provides a better fit (visually) as it is a more flexible (general) distribution. The duration has a lighter tail comparing to lognormal. Figure \ref{fig:cdf_dist_dur} compares duration and traveled distance in different times of day. Trips that happen around evening rush hour are longer than rest of the day (e.g., $\approx90\%$ of trips are at most $\approx 1000$ seconds vs $\approx 700$ seconds for noon). As for the distance, similar trait holds except the probability of shorter (3KM or less) trips are almost the same. Also, morning trips tend to be lengthier than other times of the day. \begin{figure} \setlength{\belowcaptionskip}{-13pt} \subcaptionbox{} {\includegraphics[width=0.49\columnwidth]{images/trip_distance_cdf_comparison}} \subcaptionbox{} {\includegraphics[width=0.49\columnwidth]{images/trip_duration_cdf_comparison}} \caption{Comparison of trip (a) distance (b) duration through out the day.} \label{fig:cdf_dist_dur} \end{figure} \begin{figure}[b] \includegraphics[width=0.66\columnwidth,height=0.4\columnwidth]{images/trip_waypoint_count_cdf_comparison} \caption{CDF of number of waypoints per trip in different times of day.} \label{fig:cdf_waypoint} \end{figure} \begin{figure*}[!ht] \subcaptionbox{8-9am\label{fig:distance_distr:a}}{% \includegraphics[width=0.49\columnwidth]{images/reappcount_vs_distance_jointplot_8}% } \subcaptionbox{12-13pm\label{fig:distance_distr:b}}{% \includegraphics[width=0.49\columnwidth]{images/reappcount_vs_distance_jointplot_12}% } \subcaptionbox{4-5pm\label{fig:distance_distr:c}}{% \includegraphics[width=0.49\columnwidth]{images/reappcount_vs_distance_jointplot_16}% } \subcaptionbox{8-9pm\label{fig:distance_distr:d}}{% \includegraphics[width=0.49\columnwidth]{images/reappcount_vs_distance_jointplot_20}% } \caption{Joint distribution of count of waypoints for a trip vs distance in different times of day.} \label{fig:joint_plot_waypoint_distance} \end{figure*} \begin{figure*}[!ht] \subcaptionbox{8-9am}{% \includegraphics[width=0.48\columnwidth]{images/heatmap_uniq_trip_appearance_8}% } \subcaptionbox{12-13pm}{% \includegraphics[width=0.48\columnwidth]{images/heatmap_uniq_trip_appearance_12}% } \subcaptionbox{4-5pm}{% \includegraphics[width=0.48\columnwidth]{images/heatmap_uniq_trip_appearance_16}% } \subcaptionbox{8-9pm}{% \includegraphics[width=0.48\columnwidth]{images/heatmap_uniq_trip_appearance_20}% } \caption{Heatmap of unique number of trips spanning the city (on a 300x300 grid). It resembles the city map with its traffic load in different times of the day.} \label{fig:heatmaps_uniq} \end{figure*} \subsubsection{Number of waypoints}\label{sssec:wp} Waypoints are simply put, reappearances of the same trip\_id in traces. Although this measure is not an intrinsic characteristic of the trips, rather the result of data generation method, it can provide us with insights for understanding the dataset and later for feature engineering and representations for the trips. Figure \ref{fig:cdf_waypoint} shows the cumulative distribution of the number of waypoints per trip throughout the day. The waypoint count is higher for 4 pm trips ($\approx90\%$ of having roughly 1000 waypoints or less, while at noon the number becomes $\approx$750). Morning and night periods (8 am and 8 pm) follow similar patterns. To further address the implications of such data we compare waypoint count to distance. Figure \ref{fig:joint_plot_waypoint_distance} presents joint-distribution plots of waypoint count and traveled distances. The Pearson correlation coefficient is 0.59, 0.89, 0.71 and 0.81 for the hours of the day in the study respectively which suggests the existence of a relatively strong positive linear correlation between the two peaking at noon time. This, in turn, suggests that the number of waypoints can be a good representative of the distance traveled in this data (possibly due to equidistant time intervals of trip logging in data generation). \subsubsection{Spatial Spread}\label{sssec:spatial} Figure \ref{fig:heatmaps_uniq} shows 300x300 heatmaps, depicting the structure of the city based on the volume of traffic (unique count of trips in a grid cell in the hour). At noon, the traffic seems to be more focused on the middle and south-east side of the city with less load on the highways/beltways. There is a significant drop in traffic towards the night. The spatial spread of the trips represents the map of city roads, this suggests the existence of a subspace in the spatial space of trips (so a more compact representation might be more useful comparing to lat-lon. coordinate space). \begin{figure}[t] \subcaptionbox{} {\includegraphics[width=0.49\columnwidth]{images/trip_start_time_cdf_comparison}} \subcaptionbox{} {\includegraphics[width=0.49\columnwidth]{images/trip_end_time_cdf_comparison}} \caption{Comparison of trip (a) start time (b) end time through out the day.} \label{fig:cdf_temporal} \end{figure} \subsubsection{Temporal Distribution}\label{sssec:temporal} CDFs of start and end times of the trips are shown in figure \ref{fig:cdf_temporal}. More morning and night trips start in the first 10 minutes of the hour than the first 10 minutes of noon or evening rush hour although in the case of 4 pm, towards the end of the hour, more trips are started (there is a jump in the number of trips that start after 2200 seconds). At half hour marker, almost 10\% more trips have already finished in the morning and nighttime comparing to noon or 4 pm (close to 50\% vs 40\%). \subsubsection{Relation of space and time}\label{sssec:relation} We observe the relationship of trip start location and its duration using box plots in Fig. \ref{fig:box_plots} throughout the day. The repeating patterns each corresponds to same longitude grid cells (left to right, going from west to east) and the overall trend moving from left to right in the figures corresponds to higher latitudes. The patterns suggest that trips happening in the east of the city are longer in general and trips happening on the north side are slightly longer as well. This may be explained by highways (German Autobahn) around the city to the east and north-west. \begin{figure}[t] \setlength{\belowcaptionskip}{-13pt} \includegraphics[width=0.99\columnwidth]{images/8am_boxplot_10by10grid_vs_duration} \caption{Box plot of duration of 8am trips versus their starting cell in a 10x10 grid of the city area. Repeating patterns are observed on west to east traversal of the grid. Overall trend shows a slight increase in duration towards the north.} \label{fig:box_plots} \end{figure} \subsection{Similarity}\label{sec:similarity} In order to better understand the trips, we propose a similarity measure as the arithmetic mean of point-wise (e.g. origin vs origin, destination vs destination) distances each achieved through the weighted geometric mean of Euclidean similarity (spatial) and their temporal similarity. Both location and time features are scaled to range $[0,1]$. The intuition behind such a choice is that the Geometric mean adjusts the spatial component of two points based on their temporal similarity and arithmetic mean provides a measure based on all the point-wise similarities. Mathematically: $$ psim(p_1,p_2) = e^{\frac{w_1 \ln(\frac{1}{1+dist(p_1,p_2)}) + w_2 \ln(\frac{1}{1+time(p_1,p_2)})}{w_1+w_2}}$$ $$ sim(t_1,t_2) = \frac{\sum_{i=0}^n psim(t_1^i,t_2^i)}{n} \label{math:sim_measure}$$ Where $dist(p1,p2)$ is the spatial distance of two points (Euclidean in this case), $time(p_1,p_2)$ is the absolute difference of the points in time and $w_1$ and $w_2$ are the weight given to them respectively. We use $1/(1+dist)$ to convert a distance into similarity. The similarity of trips $sim(t_1,t_2)$ is the average of all the $n$ points (waypoints, or just endpoints in case of OD representation). In this study, for clustering and matching of trips, higher weight ($w_1 = [0.6 , 0.7]$ vs $w_2 = [0.3 , 0.4]$) is given to the spatial component so that time difference acts as the adjusting factor in determining the similarity measure for a pair of trips. Effect of the choice of weights on the results of the matching is available in Sec. \ref{sec:compare}. \begin{figure}[t] \setlength{\belowcaptionskip}{-13pt} \includegraphics[width=0.55\columnwidth]{images/8am_sim_pdf_both} \label{fig:similarity measure distribution} \caption{Distribution of the values of similarity. Laplacian kernel is applied to create more distinction between values. } \end{figure} \subsection{Clustering and Visualization}\label{sec:clustering} With the help of similarity defined above in section \ref{sec:similarity}, we are able to generate spatio-temporal clusters by using Spectral Clustering (SC). Because of the spatial constraints enforced by the map of the city, spatial features do not form a convex space and hence we need clustering algorithms that work with connectivity instead of shape. We tried DBSCAN and SC algorithms; Only SC was able to find meaningful clusters. SC make use of the spectrum (eigenvalues) of the similarity matrix of the data to project the data into fewer dimensions before clustering. In order to visualize the data, Multi-Dimensional Scaling (MDS) and Principal Component Analysis (PCA) are used. Figure \ref{fig:clustering_mds_pca} shows the resulting clusters for morning trips on 2D scaling and 2D projection of OD representation into first two principal component's space. First two PCs capture over 82\% of the variance of the data. To test the quality of this clustering, we attempt to interpret the resulting clusters. Clusters are shown on a scaled lat. lon. plane with mean and median of start and end points in Figure \ref{fig:clustering_spatial}. The ellipsoid axes correspond to the std. of lat. and lon. of the cluster. These clusters cover the area of the city with some of them spanning same areas especially near the city's center (downtown). Interestingly, as Figure \ref{fig:clustering_time} shows, there is a clear separation in time for those clusters that are spatially overlapping. This suggests the clustering scheme is successful in finding clusters that show different spatio-temporal traits. \begin{figure}[t] \setlength{\belowcaptionskip}{-5pt} \subcaptionbox{} {\includegraphics[width=0.49\columnwidth]{images/8am_2mds_projection_spectral_clusters}} \subcaptionbox{} {\includegraphics[width=0.49\columnwidth]{images/8am_2pc_projection_spectral_clusters}} \caption{Results of Spectral Clustering for 8-9am visualized (each color represents a cluster) on (a) 2D MDS plot (b) First two Principal Components. Other times of the day exhibit similar patterns.} \label{fig:clustering_mds_pca} \end{figure} \begin{figure}[t] \setlength{\belowcaptionskip}{-5pt} \includegraphics[width=0.55\columnwidth]{images/cluster_distribution_in_space} \caption{Spread of clusters in space} \label{fig:clustering_spatial} \end{figure} \begin{figure}[!h] \setlength{\belowcaptionskip}{-13pt} \subcaptionbox{} {\includegraphics[width=0.49\columnwidth]{images/8am_kde_trip_startime}} \subcaptionbox{} {\includegraphics[width=0.49\columnwidth]{images/8am_kde_trip_endtime}} \caption{Kernel Density Estimates of 8am trips (a) start times (b) end times. There is a clear separation in time for the clusters that are spatially intermixed.} \label{fig:clustering_time} \end{figure} \section{Matching Trips with OD Representation}\label{sec:matching} The existence of clusters and a spectrum of similarity suggest the availability of trips nearby. Knowing this we can use variations of the similarity score for applications such as shared transportation. Thus, we refine the similarity to reflect more on each scenario as well as a dynamic matching of trips for each case. We focus on morning time window (8 am) and for all cases, the trips are divided into two categories: Riders and Rides (2000 riders vs 10000 rides). The goal is to find matches for the rider to ride with. Trip end times are used as the expected time of arrival (latest possible). \subsection{Catch-a-Ride} In this scenario, riders have a time margin to reach the pickup points (ride's origin) and a margin for them to arrive at their destination after drop-off (ride's destination). We redefine the similarity score as Catch-a-Ride (CaR) score between $t_1$ and $t_2$ as if $t_1$ can move to $t_2$ and ride with it. Therefore, $t_2$ has to start after $t_1$ and finish before it to have a high score (and be close nearby). We can do this by modifying the similarity score so that the time component is the time of $t_2 - t_1$ for the origin points and time of $t_1 - t_2$ for destinations rather than the absolute values. This will create an asymmetric affinity matrix $A$ where $A_{i,j}$ is the $CaR\_score(t_i,t_j)$. \subsubsection{Clustering and Visualization} We decompose the matrix into one symmetric (average of the matrix and its transpose) and one anti-symmetric (skew-symmetric) component (half of the difference of matrix and its transpose) and then use the symmetric part for clustering and visualization. Fortunately, the ratio sum of squares of the symmetric matrix to the original is 98\% which suggests the matrix is almost symmetric in nature and the symmetric component can be a good representative of it. This is expected as the CaR\_score is a subset of the similarity score where the time difference is not the absolute value. Clusters (using spectral clustering) are visualized in Figure \ref{fig:clustering_CaR} on MDS and 2PC plots. \begin{figure}[!t] \setlength{\belowcaptionskip}{-13pt} \subcaptionbox{} {\includegraphics[width=0.49\columnwidth]{images/8am_2d_mds_WtR_2k_sample}} \subcaptionbox{} {\includegraphics[width=0.49\columnwidth]{images/8am_2pc_pca_WtR_2k_sample}} \caption[Visualization of Catch-a-ride clusters]{Visualization of the clusters on (a) 2D DMS (b) First 2 PCs. CaR score results in nested ellipses in MDS. By definition, $CaR\_score$ and $carpool\_score$ are complements of each other with respect to original similarity score and hence the clusters (and MDS and PCA) would be the same (with some rotation).} \label{fig:clustering_CaR} \end{figure} \subsubsection{Dynamic Matching} For each rider, the potential matches are filtered out by using distance and time thresholds. The trend of the number of matches per threshold is presented in Figure \ref{fig:CaR_count_per_dist_and_time} (the number of riders where they have at least L matches). As observable, the number of matches is more sensitive to distance thresholds compared to time. After 250 seconds, the number of riders that get a match at any value of L scantly changes. At any distance threshold, the number of riders with at least L matches significantly drop with an increase of L ($\approx$900 riders get at least 5 matches while $\approx$1500 get at least a match). Next, we choose the thresholds of 1800 meters and 900 seconds (which enables us to provide a match for almost 3/4 of the riders) and use the CaR\_score as the matching criteria. Note that the choice of one is independent of others and hence the greedy approach is optimal. The results of such a matching scheme are presented in Table \ref{tab:dynamic_match_car_carpool}. The total travel required by both the requests (riders) and the matches (rides) are $12990.199$ KM where 33.5\% of it is traveled by the matches. Since roughly 75\% of the requests have a match (1496 out of 2000), the ratio of matched to total for those is 45.4\% which means in cases where at least a potential match exists, \textbf{$\approx$55\%} of the traveled kilometers can be saved. Since 5235.3 km of the 8633.8 are taken by those who have matches, the total travel done by unmatched requests is 3398.5 KM. Then, the traveled distances with ride-sharing amount to 3398.5 + 4356.3, the ratio of which to total (12990.2) is 59.7\% which yields \textbf{a saving of 40.3\%}, compared to having no shared transportation at the cost of commuting to pick up and drop off locations ($\approx$1000 KM each, average of 0.66 KM for each rider). \begin{figure}[t] \centering \setlength{\belowcaptionskip}{-10pt} \subcaptionbox{} {\includegraphics[width=0.49\columnwidth]{images/CaR_match_per_dist_thresh}} \subcaptionbox{} {\includegraphics[width=0.49\columnwidth]{images/CaR_match_per_time_thresh}} \caption{Number of potential matches for CaR (similarly for CarPool) (a) per distance (time threshold fixed at 900 sec) (b) per time (distance threshold fixed at 1800 meters).} \label{fig:CaR_count_per_dist_and_time} \end{figure} \newcolumntype{b}{>{\footnotesize}X} \newcolumntype{s}{>{\hsize=.33\hsize}X} \begin{table*}[ht] \centering \caption{Results of matching trips for ride sharing Catch-a-Ride (CaR) and carpool (CP) scenarios.} \label{tab:dynamic_match_car_carpool} \begin{tabularx}{\textwidth}{bss} \hline & CaR & CarPool \\ \hline match travels (km) & 4356.368 & 5486.785 \\ req travels (km) & 8633.831 & 8633.831 \\ match to total travel ratio & 33.5\% & 38.8\%\\ origin-origin distance (km) & 1017.665 & 948.438 \\ dest-dest distance (km) & 1045.912 & 1019.560\\ origin-origin times (sec) & 70185 & 83171\\ dest-dest times (sec) & 88873 & 88064 \\ \# req with at least a match & 1496 & 1458\\ req travels for least a match (km) & 5235.319 & 4938.073 \\ match to total travel ratio (at least a match) & 45.4\% & 52.6\% \\ \hline \end{tabularx} \end{table*} \subsection{Car Pooling} Similar to the previous scenario, but instead the ride is willing to go out of their way (within a threshold) to pick the rider up and drop them off before arriving at its own destination on time. The clustering is omitted as the constructed affinity matrix is the transpose of CaR scenario and hence would result in the same symmetric component. The dynamic matching is discussed next. The threshold filters are applied to find potential matches. The trends of the number of potential matches over time and distance thresholds follow a very similar curve to CaR (Fig. \ref{fig:CaR_count_per_dist_and_time}). There is a slightly lower number of requests with a match at each threshold, but the overall trends are alike with more sensitivity to distance thresholds. Distance and time thresholds of 1800 meters and 900 seconds are again applied, and the CP\_score is used as the matching criteria. Results are presented in Table \ref{tab:dynamic_match_car_carpool}. The total KMs traveled would be the request travels plus the pickup and drop off ($8633.8 + 948.4 + 1019.5 = 10601.7$). Without any carpooling, the total travels by matches and requests would be $8633.8 + 5486.8 = 14120.6$KM and hence in existence of the ride-sharing there is 25\% decrease in total KMs traveled at the cost of drivers going a total of 948.4 KMs out of their way (average of 0.65 KM per driver) for pick up and 1019.5 KMs for drop off (0.69 KM average). CDFs of resulting pick-up and drop-off distances and times are presented in Figure \ref{fig:CP_result_dist_time} \begin{figure} \centering \subcaptionbox{}{\includegraphics[width=0.49\columnwidth]{images/CP_result_distance}} \subcaptionbox{}{\includegraphics[width=0.49\columnwidth]{images/CP_result_time}} \caption{CDF of CP results of pick-up and drop-off (a) distance (b) time.} \label{fig:CP_result_dist_time} \vspace{-10pt} \end{figure} \section{Comparison with Other Similarity Scores} \label{sec:compare} The explained scenarios only utilize the OD representation for the trips (and hence only the endpoints) but the similarity score is not limited to OD representation. As mentioned in the \ref{sec:relatedwork} section, various measures of similarity or distance have been proposed for comparison of trajectories. Longest Common Sub-Sequence (LCSS) is a measure of closeness based on the number of points shared by two trajectories (points within thresholds of each other). Frechet distance is the maximum distance between two curves over all parameterization of the curves. This can be imagined as the minimum length of the leash required to take a dog for a walk when the human and dog have their own path to travel with all the possible combinations of speed for each (no backward movement allowed). Dynamic Time Warping (DTW) works by finding the minimum cost warp path where a warp path is the sequence of changes that warps a trajectory into the other. In this section, we investigate the performance of Catch-a-Ride scenario using DTW, Frechet Distance and LCSS and our score (referred to as WGM for Weighted Geometric Means, weights of 0.6 and 0.4 for distance and time resp.). In all cases, we use a 50-waypoint (sampled) representation of the trips instead of only OD and the same filtering is applied (and hence measures related to the request set are kept constant). The results are presented in table \ref{tab:comparison}. In case of WGM with time, the weights of the geometric mean are given as 0.1 for spatial and 0.9 for temporal parts. In DTW with time, the cost is the product of distance and time difference instead of distance only. As the table shows, Frechet distance (which is an upper-bound on the distance of the trajectories by definition) provides closest matches in distance (but not in time). On the other hand, DTW with time, results in least orig-orig and dest-dest time but \textbf{not} distance. WGM achieves the results with both time and space in mind so at a tiny cost in orig-orig and dest-dest distance it beats others (original versions) in temporal aspect . Also with the flexibility of weights, we can tune it for more focus on time or space. From a runtime point of view, our score is linear in the number of waypoints for the trips whereas LCSS and DTW have runtime and space complexity of $O(N^2)$ and Frechet takes even longer with $O(N^2 log(N))$. To better understand the sensitivity of results to the choice of the weight of the temporal component in WGM, trends of change in pick-up and drop-off distance and time is plotted against temporal weight (w\_t) in Fig. \ref{fig:oodd_dist_time_per_wt}. With an increase in w\_t, more preference is given to time element of the similarity and hence the matches are closer in time (total pick-up and drop-off times decreases) while they are not necessarily closer in space. This can be used to tune the score for different applications with levels of dependence on time versus location. \begin{figure} \centering \setlength{\belowcaptionskip}{-10pt} \subcaptionbox{}{\includegraphics[width=0.49\columnwidth]{images/oodd_dist_per_wt}} \subcaptionbox{}{\includegraphics[width=0.49\columnwidth]{images/oodd_time_per_wt}} \caption{pick-up and drop-off (a) distance (b) time.} \label{fig:oodd_dist_time_per_wt} \end{figure} \begin{table*}[ht] \centering \caption{Comparison on different similarity/distance scores on Catch-a-Ride scenario.} \label{tab:comparison} \begin{tabularx}{\textwidth}{bssssss} \hline & WGM & LCSS & Frechet & DTW & DTW w/ time & WGM w/ time\\ \hline match travels (km) & 5655.997 & 5393.861 & 5671.631 & 5652.613 & 6038.372 & 5945.291 \\ req travels (km) & 11706.334 & 11706.334 & 11706.334 & 11706.334 & 11706.334 & 11706.334 \\ match to total travel ratio & 32.57\% & 31.54\% & 32.63\% & 32.56\% & 34.02\% & 33.68\% \\ origin-origin distance (km) & 1149.131 & 1442.782 & 1104.907 & 1127.623 & 1202.733 & 1203.724 \\ dest-dest distance (km) & 1188.705 & 1421.046 & 1144.947 & 1165.916 & 1222.000 & 1255.597 \\ origin-origin times (sec) & 71365 & 86071 & 74544 & 72567 & 52866 & 57896\\ dest-dest times (sec) & 80860 & 101652 & 89570 & 86655 & 57546 & 62424 \\ \# req with at least a match & 1357 & 1357 & 1357 & 1357 & 1357 & 1357\\ req travels for least a match (km) & 7244.179 & 7244.179 & 7244.179 & 7244.179 & 7244.179 & 7244.179 \\ match to total travel ratio (at least a match) & 43.84\% & 42.67\% & 43.91\% & 43.82\% & 45.46\% & 45.07\% \\ \hline \end{tabularx} \end{table*} \section{Car Sharing}\label{sec:carsharing} The similarity score, in addition to clustering or dynamic matching of trips, can be used to find a scheduling for car sharing applications. This is particularly useful in case of the existence of autonomous driving systems. Our aim here is to service all our requests (trips) with the minimum number of cars shared while in all the possible assignment, we focus on the ones that minimize the distance and time different (based on the similarity score) between consecutive trips assigned to the same car. The same set of 2000 trips is used as the request set. The problem is formulated as a graph: \begin{itemize} \item Each node represents a trip (N nodes). Each directed edge between any 2 given nodes represents the possibility of the destination node to use the car/get serviced after the source node of that edge (i.e. an edge $(a,b)$ exist only if the end point of $a$ is within time and distance threshold of start point of $b$ and $b$ starts after $a$ ends). The similarity score is used as edge weights. \item The problem as stated above (i.e. finding the minimum number of cars) is translated into min-path partitioning of graphs (a path in our graph represents the chain of trips being serviced by the same car). This problem is NP-hard in general graphs (proof by reduction from Hamiltonian Cycle (HC): HC exists if and only if the graph can be covered with 1 path, hence if we find the min number of paths to be 1, the graph has an HC). \item Since an edge only exists if the trip corresponding to the destination node happens after the source node's, cycles are not possible and a partial ordering is maintained (i.e. the graph is a DAG - directed acyclic graph). Fortunately, there exists a polynomial time algorithm that can solve the path partitioning problem for DAGs. \end{itemize} The algorithm works by converting the DAG G into a bipartite graph B where matching in B translates into path partitioning in G: \begin{itemize} \item Given G, construct B by creating nodes $i$ in the left part and $i'$ in the right part of B for every node $i$ in G. An \textit{undirected} edge exists between $i$ and $j'$ in B if the directed edge (i,j) exists in G. \item Running a maximal maximum matching (max-cardinality max-weight) on B would yield the minimum number of cars and their chain of the trip. This can be achieved by modifying weights on the edges by adding N times the maximum score to all edges (so that no augmenting path exists that has fewer edges and higher weight) and then running a maximum matching algorithm (e.g. Hungarian algorithm or Edmond-Karp max-flow equivalent problem). Assume $T = max(score)$ this means every original edge has a value of $[0,T]$, so a matching of size K has a maximum total of $(K * N +1 )* T$. If we take off a match of size $K$ total cost is reduced by a maximum of $(K * N +1 )* T = K N T + T$ and adding a new matching of bigger size K+1 even at the minimum possible weight would be $(K+1) * N * T = K N T + NT$. Since the minimum with $K+1$ has definitely more total weight than maximum with $K$ the algorithm would then always choose the longest path. \item Each chain of trips (path in DAG) has exactly one start and one end (they can be the same), hence the number of chains would be the number of either start or end nodes. Any right part node of B not in the matching is the start of a chain of trips so the number of such nodes is the required number of cars (N - matching\_cardinality). To retrieve the chains for each $i'$ in the right part that is not in the matching, find $i$ in the left part and follow its match chain recursively (i.e. start with $j'$ that $i$ is matched to). \end{itemize} The graph constructed by thresholds of 900 seconds and 1.8 KM, results in 38730 edges (2K nodes). The matching cardinality is 1370 which means 630 cars are needed to service 2000 trips with 149 of them not belonging to any chains (singleton trips), leaving 481 chains with the average length of $\approx$3.88. The distribution of chain lengths, chain travels, consecutive pick-up travels and times are presented in Figure \ref{fig:CSH_results} (CDFs are omitted for brevity). Majority of non-singleton trips have chains of size $\leq$4. Approximately 90\% of chains travel <30 KMs, and their total consecutive pick-up distance and time are at most $\approx3.5$ KMs and 1300 seconds. \begin{figure}[] \centering \setlength{\belowcaptionskip}{-1pt} \subcaptionbox{} {\includegraphics[width=0.49\columnwidth]{images/chain_len_hist}} \subcaptionbox{} {\includegraphics[width=0.49\columnwidth]{images/chain_travel_hist}} \subcaptionbox{} {\includegraphics[width=0.49\columnwidth]{images/pickup_travels_hist}} \subcaptionbox{} {\includegraphics[width=0.49\columnwidth]{images/pickup_time_hist}} \caption{Carsharing results (a) chain length (b) total km traveled (c) total km traveled to pickup next rider (d) total time spent to pickup next rider.} \label{fig:CSH_results} \vspace{-6pt} \end{figure} \section{Concluding Remarks}\label{sec:conclusion} In this paper, driven by data, we explored various spatio-temporal characteristics of the trips (of the city of Cologne). The systematic analysis can be applied to other datasets for understanding such characteristics and for comparison purposes. Next, We have proposed a similarity score, \textbf{W}eighted \textbf{G}eometric \textbf{M}ean (\textbf{WGM}), that reflects both spatial and temporal similarities with knobs that can be used to favor the spatial or temporal similarity (through weight adjustments). It is also efficient to calculate. This score is used in clustering of trips, successfully identifying clusters separated by distance or time. Next, scenarios of shared transportation including Catch-a-Ride and CarPool are discussed and WGM is used as the criteria for dynamic matching of trips for this purpose. A comparison of the performance of various similarity scores/distances on the Catch-a-Ride scenario is also given. WGM is able to capture similar trips in both distance and time dimensions. The weights (of the geometric mean) can be used to capture more similarity in one dimension or the other which gives our score more flexibility while it remains computationally simple compared to other metrics. Lastly, we formulated the problem of car sharing (autonomous fleet of vehicles) using graph theory and algorithms (with our score as the weights for the graph) which achieves the optimal number of cars required with the minimum spatio-temporal difference between consecutive pick-ups. To follow up, we plan to confirm the reproducibility of the results on different datasets. A potential dataset is the simulator-generated traces of the trips based on traffic cameras around the globe \cite{enroute,Thakur2012hotplanet}. Additionally, we plan to find embedding spaces that reflect the similarity of trips. In general, trips are represented through thousands (and not necessarily the same number) of waypoints. This makes detailed comparisons computationally intensive. Encompassing trips in an embedding space (with a fixed number of dimensions) can ease the problem. Another potential application is in a real-time matching of trips where the details of paths are not shared (e.g. for privacy concerns, or compression). The notion of similarity and its variations can be used for the design of systems with pollution reduction and public safety in mind potentially through participatory sensing and crowdsourcing. These steps would lead to the concept of profiles of vehicular mobility as a way to describe human mobility behavior. \begin{acks} This project was partially supported by NSF grant 1320694 and UF Informatics Institute. \end{acks}

ce6bf593601336015c6da5338e697f1f3509197d

\section{Introduction} Making optimal and adaptive intervention decisions in the face of uncertainty is a central task in precision medicine, computational social science, and artificial intelligence. In healthcare, the problem of learning optimal policies is studied under the heading of \emph{dynamic treatment regimes} \cite{Moodie13DTR}. The same problem is called \emph{reinforcement learning} in artificial intelligence \cite{sutton98reinforcement}, and \emph{optimal stochastic control} \cite{bertsekas96neuro} in engineering and signal processing. In all of these cases, a policy (a function of historical data to some space of possible actions, or a sequence of such functions) is chosen to maximize some pre-specified outcome quantity, which might be abstractly considered a \emph{utility} (or \emph{reward} in reinforcement learning). Increasingly, ideas from optimal policy learning are being applied in new contexts. In some areas, particularly socially-impactful settings like criminal justice, social welfare policy, hiring, and personal finance, it is essential that automated decisions respect principles of fairness since the relevant data sets include potentially sensitive attributes (e.g., race, gender, age, disability status) and/or features highly correlated with such attributes, so ignoring fairness considerations may have socially unacceptable consequences. A particular worry in the context of automated sequential decision making is ``perpetuating injustice,'' i.e., when maximizing utility maintains, reinforces, or even introduces unfair dependence between sensitive features, decisions, and outcomes. Though there has been growing interest in the issues of fairness in machine learning \cite{pedreshi08discrimination,feldman15certifying,hardt16equality,kamiran13quantifying,corbett2017algorithmic,jabbari16fair,kusner17counterfactual,zhang18fairness,mitchell18speak,Zhang17causal}, so far methods for optimal policy learning subject to fairness constraints have not been well-explored. As a motivating example, we consider a simplified model for a children's welfare screening program, recently discussed in \cite{chouldechova2018case, Hurley18child}. A hotline for child abuse and neglect receives many thousands of calls a year, and call screeners must decide on the basis of calculated risk estimates what action to take in response to any given call, e.g., whether or not to follow up with an in-person visit from a caseworker. The idea is that only cases with substantial potential risk to the child's welfare should be prioritized. The information used to determine the calculated risk level and thereby the agency's action includes potentially sensitive features, such as race and gender, as well as a myriad of other factors such as perhaps whether family members receive public assistance, have an incarceration history, record of drug use, and so on. Though many of these factors may be predictive of subsequent negative outcomes for the children, there is a legitimate worry that both risk calculations and policy choices based on them may depend on sensitive features in inappropriate ways, and thereby lead to unfair racial disparities in the distribution of families investigated, and perhaps separated, by child protective services. Learning high-quality policies that satisfy fairness constraints is difficult due to the fact that multiple sources of bias may occur in the problem simultaneously. One kind of bias, which we call \emph{retrospective bias}, has its origin in the historical data used as input to the policy learning procedure. This data may reflect various systematic disparities and discriminatory historical practices in our society, including prior decisions themselves based on poor data. Algorithms trained on such data can maintain these inequities. Furthermore, decision making algorithms may suffer from what we call \emph{prospective} sources of bias. For instance, suppose the functional form of the chosen decision rule explicitly depends on sensitive features in inappropriate ways. In that case, making decisions based on the new decision rule may perpetuate existing disparities or even introduce disparities that were previously absent. Avoiding this sort of bias may involve imposing non-trivial restrictions on the policy learning procedure. Finally, learning high-quality policies from observational data requires dealing with \emph{confounding bias}, where associations between decision and reward cannot be used directly to assess decision quality due to the presence of confounding variables, as well as \emph{statistical bias} due to the reliance on misspecified statistical models. Policy learning algorithms that respect fairness constraints must address all of these sources of bias. In this paper, we use tools from mediation analysis and causal inference to formalize fairness criteria as constraints on certain impermissible causal pathways from sensitive features to actions or outcomes \cite{NabiShpitser18Fair}. Moreover, we describe how all the aforementioned biases can be addressed by a novel combination of methods from causal inference, constrained optimization, and semiparametric statistics. Our main theoretical result illustrates in what sense enacting fair policies can ``break the cycle of injustice'': we show how to learn policies such that the joint distribution induced by these policies (in conjunction with reward/utility mechanisms outside the policy-maker's control) will satisfy specified fairness constraints while remaining ``close'' to the generating distribution. To our knowledge, this paper constitutes the first attempt to integrate algorithmic fairness and policy learning with the possible exception of \citet{jabbari16fair}, which addressed what we call prospective bias in the context of Markov Decision Processes. To precisely describe our approach, we must introduce some necessary concepts and tools from causal inference and policy learning. Then, we summarize the perspective on algorithmic fairness in prediction problems from \citet{NabiShpitser18Fair}, and adapt this framework to learning optimal fair policies. We illustrate our proposal via experiments on synthetic and real data. \section{Notation and preliminaries} \label{sec:prelim} Consider a multi-stage decision problem with $K$ pre-specified decision points, indexed by $k = 1, \ldots, K$. Let $Y$ denote the final outcome of interest and $A_k$ denote the action made (treatment administered) at decision point $k$ with the finite state space of ${\cal A}_k$. Let $X$ denote the available information prior to the first decision, and $Y_k$ denote the information collected between decisions $k$ and $k+1$, ($Y \equiv Y_{K}$). $\overline{A}_k$ represents all treatments administered from time $1$ to $k$; likewise for $\overline{Y}_k$. We combine the treatment and covariate history up to treatment decision $A_k$ into a history vector $H_k$. The state space of $H_k$ is denoted by ${{\cal H}_k}$. Note that while our proposal in this paper applies to arbitrary state spaces, we present examples with continuous outcomes and binary decisions for simplicity. The goal of policy learning is to find policies that map vectors in ${\cal H}_k$ to values in ${\cal A}_k$ (for all $k$) that maximize the expected value of outcome $Y$. In offline settings, where exploration by direct experimentation is impossible, finding such policies requires reasoning counterfactually, as is common in causal inference. The value of $Y$ under an assignment of value $a$ to variable $A$ is called a \emph{potential outcome} variable, denoted $Y(a)$. In causal inference, quantities of interest are defined as functions of potential outcomes (also called counterfactuals). Estimating these functions from observational data is a challenging task, and requires assumptions linking potential outcomes to the data actually observed. Our assumptions can be formally represented using \textit{causal graphs}. In a directed acyclic graph (DAG), nodes correspond to random variables, and directed edges represent direct causal relationships. As an example, consider the single treatment causal graph of Fig. \ref{fig:triangle}(a). $X$ is a direct cause of $A$, and $A$ is both a direct cause of $Y$ as well as an indirect cause of $Y$ through $M$. A variable like $M$ which lies on a causal pathway from $A$ to $Y$ is called a \textit{mediator}. For more details on causal graphical models see, e.g., \citet{spirtes01causation} and \citet{pearl09causality}. In what follows let $Z$ denote the full vector of observed variables in the causal model, e.g., $Z = (Y,M,A,X)$ in our Fig. \ref{fig:triangle}(a). A causal parameter is said to be \emph{identified} in a causal model if it is a function of the observed data distribution $p(Z)$. In causal DAGs, distributions of potential outcomes are identified by the \emph{g-formula}. For background on general identification theory, see \citet{shpitser18identification}. As an example, the distribution of $Y(a)$ in the DAG in Fig.~\ref{fig:triangle}(a) is identified by $\sum_{X,M} p(Y | a,M,X) p(M | a,X) p(X)$. Note that some causal parameters may be identified even in causal models with hidden (``latent'') variables, typically represented by acyclic directed mixed graphs (ADMGs) \citep{shpitser18identification}. Though we do not apply our methods to hidden variable models here, the general approach and many of the specific learning strategies we propose are applicable in contexts with hidden variables, so long as the relevant parameters are identified. In our sequential setting, $Y(\overline{a}_K)$ represents the response $Y$ had the fixed treatment assignment strategy $\overline{A}_K = \overline{a}_K$ been followed, possibly contrary to fact. The contrast $\mathbb{E}[Y(\overline{a}_K)] - \mathbb{E}[Y(\overline{a}'_K)]$, where $\overline{a}_K$ is the treatment history of interest and $\overline{a}'_K$ is the reference treatment history, quantifies the \emph{average causal effect} of $\overline{a}_K$ on the outcome $Y$. \subsection*{Mediation and path-specific analysis} One way to understand the mechanisms by which treatments influence outcomes is via mediation analysis. The simplest type of mediation analysis decomposes the causal effect of $A$ on $Y$ into a \textit{direct effect} and an \textit{indirect effect} mediated by a third variable. Consider the graph in Fig \ref{fig:triangle}(a): the direct effect corresponds to the path $A \rightarrow Y$, and indirect effect corresponds to the path through $M$: $A \rightarrow M \rightarrow Y$. In the potential outcome notation, the direct and indirect effects can be defined using nested counterfactuals such as $Y(a, M(a'))$ for $a, a' \in \cal A$, which denotes the value of $Y$ when $A$ is set to $a$ while $M$ is set to whatever value it would have attained had $A$ been set to $a'$. Under certain identification assumptions discussed by \citet{pearl01direct}, the distribution of $Y(a, M(a'))$ (and thereby direct and indirect effects) can be nonparametrically identified from observed data by the following formula: $p(Y(a, M(a')) = \sum_{X, M} p(Y \mid a, X, M) p(M \mid a', X) p(X)$. More generally, when there are multiple pathways from $A$ to $Y$ one may define various \textit{path-specific effects} (PSEs), which under some assumptions may be nonparametrically identified by means of the \textit{edge g-formula} provided in \citet{shpitser15hierarchy}. We define a number of PSEs relevant for our examples below. For a general definition, see \citet{shpitser13cogsci}. \subsection*{Policy counterfactuals and policy learning} \label{sec:policy} Let $f_A = \{f_{A_1}, \ldots, f_{A_K} \}$ be a sequence of decision rules. At the $k$th decision point, the $k$th rule $f_{A_k}$ maps the available information prior to the $k$th treatment decision $H_k$ to treatment decision $a_k$, i.e.\ $f_{A_k}: {{\cal H}_k} \mapsto {{\cal A}_k}$. Given $f_A$ we define the counterfactual response of $Y$ had $A$ been assigned according to $f_A$, or $Y(f_A)$, by the following recursive definition \citep[cf.][]{robins04SNMM,thomas13swig}: {\small \begin{align*} Y\big( \big\{ f_{A_k}\big(H_k(f_{A})\big): A_k \in \pa_{\cal G}(Y)\cap A \big\}, \big\{ \pa_{\cal G}(Y) \setminus A \big\}(f_A) \big). \end{align*} In words: the potential outcome $Y$ had any parent of $Y$ that is in $A$ been set to $f_A$ in response to counterfactual history $H_k$ up to $k$, where this history behaves as if $A$ were set to $f_A$ \textit{and} any parent of $Y$ that is not in $A$, behaves as if $A$ were set to $f_A$. Under a causal model associated with the DAG $\cal G$, the distribution $p(Y(f_A))$, is identified by the following generalization of the g-formula: {\scriptsize \begin{align*} \sum_{Z \setminus \{Y, A \}} \prod_{V \in Z \setminus A} p\big(V | \{ f_{A_k}(H_k): A_k \in \pa_{\cal G}(V) \cap A \}, \pa_{\cal G}(V) \setminus A \big). \end{align*} }% As an example, $Y(a = f_A(X))$ in Fig.~\ref{fig:triangle}(a) is defined as $Y(a=f_A(X), M(a=f_A(X),X), X)$, and its distribution is identified as $\sum_{x,m} p(Y | a=f_A(x),M=m,X=x) p(M | a=f_A(x),X=x) p(X=x)$. Given an identified response to a fixed set of policies $f_A$, we consider search for the optimal policy set $f^*_A$, defined to be one that maximizes $\mathbb{E}[Y(f_A)]$. Since $Y(f_A)$ is a counterfactual quantity, validating the found set of policies is difficult given only retrospective data, with statistical bias due to model misspecification being a particular worry. This stands in contrast with online policy learning problems in reinforcement learning, where new data under any policy may be generated and validation is therefore automatic. Partly in response to this issue, a set of orthogonal methods for policy learning have been developed that model different parts of the observed data likelihood function. Q-learning, value search, and g-estimation are common methods used in dynamic treatment regimes literature for learning optimal policies \cite{Moodie13DTR}. We defer detailed descriptions to later in the paper and the supplement. \begin{figure*} \begin{center} \begin{tikzpicture}[>=stealth, node distance=1.0cm] \tikzstyle{format} = [draw, very thick, circle, minimum size=5.0mm, inner sep=0pt] \begin{scope} \path[->, very thick] node[format] (c) {$X$} node[format, below right of=c] (a) {$A$} node[format, above right of=a] (m) {$M$} node[format, below right of=m] (y) {$Y$} (c) edge[blue] (a) (a) edge[blue] (y) (a) edge[blue] (m) (m) edge[blue] (y) (c) edge[blue] (m) (c) edge[blue] (y) node[below of=a, yshift=0.1cm, xshift=0.3cm] (l) {$(a)$} ; \end{scope} \begin{scope}[xshift=3.0cm] \path[->, very thick] node[format] (x) {$X$} node[format, below right of=x] (s) {$S$} node[format, above right of=s] (m) {$M$} node[format, below right of=m] (a) {$A$} node[format, above right of=a] (y) {$Y_1$} node[format, right of=y] (y2) {$Y_2$} (s) edge[blue] (a) (s) edge[blue] (m) (m) edge[blue] (a) (x) edge[blue] (m) (x) edge[blue] (a) (a) edge[blue] (y) (m) edge[blue] (y) (s) edge[blue] (y) (x) edge[blue, bend left=25] (y) (x) edge[blue, bend left] (y2) (a) edge[blue] (y2) (y) edge[blue] (y2) (m) edge[blue, bend left] (y2) node[below of=a, yshift=0.1cm, xshift=0.0cm] (l) {$(b)$} ; \end{scope} \begin{scope}[xshift=8.0cm] \path[->, very thick] node[format] (x) {$X$} node[format, below right of=x] (s) {$S$} node[format, above right of=x] (m) {$M$} node[format, right of=s] (a1) {$A_1$} node[format, right of=m] (y1) {$Y_1$} node[ right of=a1] (dots1) {$\ldots$} node[ right of=y1] (dots2) {$\ldots$} node[format, right of=dots1] (aK) {$A_K$} node[format, right of=dots2] (yK) {$Y_K$} (y1) edge[blue] (dots2) (a1) edge[blue] (dots2) (y1) edge[blue] (dots1) (a1) edge[blue] (dots1) (dots1) edge[blue, bend left=0] (aK) (dots2) edge[blue, bend left=0] (aK) (m) edge[blue, bend left=0] (aK) (y1) edge[blue] (aK) (a1) edge[blue, bend right=15] (aK) (s) edge[blue, , bend right=25] (aK) (x) edge[blue, bend left=0] (aK) (dots1) edge[blue] (yK) (dots2) edge[blue] (yK) (m) edge[blue, bend left=25] (yK) (x) edge[blue, bend left=0] (yK) (y1) edge[blue, bend left=15] (yK) (s) edge[blue, bend left=0] (yK) (a1) edge[blue] (yK) (aK) edge[blue] (yK) (x) edge[blue] (s) (s) edge[blue] (a1) (m) edge[blue] (a1) (x) edge[blue] (a1) (s) edge[blue] (m) (x) edge[blue] (m) (a1) edge[blue] (y1) (m) edge[blue] (y1) (x) edge[blue, bend left=0] (y1) (s) edge[blue] (y1) node[below of=a1, yshift=0.1cm, xshift=0.5cm] (l) {$(c)$} ; \end{scope} \end{tikzpicture} \end{center} \caption{(a) A simple causal DAG, with a single treatment $A$, a single outcome $Y$, a vector $X$ of baseline variables, and a single mediator $M$. (b) A causal DAG corresponding to our (simplified) child welfare example with baseline factors $X$, sensitive feature $S$, action $A$, vector of mediators (including e.g.\ socioeconomic variables, histories of drug treatment) $M$, an indicator $Y_1$ of whether a child is separated from their parents, and an indicator of child hospitalization $Y_2$. (d) A multistage decision problem, which corresponds to a complete DAG over vertices $X,S,M,A_1,Y_1,\cdots,A_K,Y_K$. } \label{fig:triangle} \end{figure*} \section{From fair prediction to fair policies} \label{sec:fairpolicy} \citet{NabiShpitser18Fair} argue that fair inference for prediction requires imposing hard constraints on the prediction problem, in the form of restricting certain path-specific effects. We adapt this approach to optimal sequential decision-making. A feature of this approach is that the relevant restrictions are user-specified and context-specific; thus we will generally require input from policymakers, legal experts, bioethicists, or the general public in applied settings. Which pathways may be considered impermissible depends on the domain and the semantics of the variables involved. We do not defend this perspective on fairness here for lack of space; please see \citet{NabiShpitser18Fair} for more details. We summarize the proposal from \citet{NabiShpitser18Fair} with a brief example, inspired by the aforementioned child welfare case. Consider a simple causal model for this scenario, shown in Fig.\ 1(b). Hotline operators receive thousands of calls per year, and must decide on an action $A$ for each call, e.g., whether or not to send a caseworker. These decisions are made on the basis of a (high-dimensional) vectors of covariates $X$ and $M$, as well as possibly sensitive features $S$, such as race. $M$ consists of mediators of the effect of $S$ on $A$. $Y_1$ corresponds to an indicator for whether the child is separated from their family by child protective services, and $Y_2$ corresponds to child hospitalization (presumably attributed to domestic abuse or neglect). The observed joint distribution generated by this causal model would be $p(Y_1,Y_2,A,M,S,X)$. The proposal from \citet{NabiShpitser18Fair} is that fairness corresponds to the impermissibility of certain path-specific effects, and so fair inference requires decisions to be made from a counterfactual distribution $p^*(Y_1,Y_2,A,M,S,X)$ which is ``nearby'' to $p$ (in the sense of minimal Kullback-Leibler divergence) but where these PSEs are constrained to be zero. They call $p^*$ the distribution generated by a ``fair world.'' Multiple fairness concerns have been raised by experts and advocates in discussions of the child protection decision-making process \cite{chouldechova2018case, Hurley18child}. For example, it is clearly impermissible that race has any direct effect on the decision made by the hotline screener, i.e., that all else being held fixed, members from one group have a higher probability of being surveilled by the agency. However, it is perhaps permissible that race has an indirect effect via some mediated pathway, e.g., if race is associated with some behaviors or features which themselves ought to be taken into consideration by hotline staffers, because they are predictive of abuse. If that's true, then $S \rightarrow A$ would be labeled an impermissible pathway whereas $S \rightarrow M \rightarrow A$ (for some $M$) would be permissible. Similarly, it would be unacceptable if race had an effect on whether children are separated from their families; arguably both the direct pathway $S \rightarrow Y_1$ and indirect pathway though hotline decisions $S \rightarrow A \rightarrow Y_1$ should be considered impermissible. Rather than defend any particular choice of path-specific constraints, we note that the framework outlined in \citet{NabiShpitser18Fair} can flexibly accommodate any set of given constraints, as long as the PSEs are identifiable from the observed distribution. \subsection{Inference in a nearby ``fair world''} We now describe the specifics of the proposal. We assume the data is generated according to some (known) causal model, with observed data distribution $p(\cdot)$, and that we can characterize the fair world by a fair distribution $p^*(\cdot)$ where some set of pre-specified PSEs are constrained to be zero, or within a tolerance range. Without loss of generality we can assume the utility variable $Y$ is some deterministic function of $Y_1$ and $Y_2$ (i.e., $Y \equiv u(Y_1,Y_2)$) and thus use $Y$ in place of $Y_1$ and $Y_2$ in what follows. Then $Z = (Y,X,S,M,A)$ in our child welfare example. For the purposes of illustration, assume the following two PSEs are impermissible: PSE$^{sa}$, corresponding to the direct effect of $S$ on $A$ and defined as $\mathbb{E}[A(s,M(s'))] - \mathbb{E}[A(s')]$, and PSE$^{sy}$, corresponding to the effect of $S$ on $Y$ along the edge $S \to Y$, and the path $S \to A \to Y$ and defined as $\mathbb{E}[Y(s, A(s,M(s')), M(s'))] - \mathbb{E}[Y(s')]$. If the PSEs are identified under the considered causal model, they can be written as functions of the observed distribution. For example, the unfair PSE of the sensitive feature $S$ on outcome $Y$ in our child welfare example may be written as a functional PSE$^{sy} = g_1(Z) \equiv g_1(p(Y, X, S, M, A))$. Similarly the unfair PSE of $S$ on $A$ is PSE$^{sa}$ $= g_2(Z) \equiv g_2(p(Y, X, S, M, A))$. Generally, given a set of identified PSEs $g_j(Z)$ $\forall j \in \{ 1,..., J \}$ and corresponding tolerated lower/upper bounds $\epsilon_j^-, \epsilon_j^+$, the fair distribution $p^*(Z)$ is defined: {\small \begin{align} p^*(Z) &\equiv \arg \min_{q} \hspace{0.1cm} D_{KL}(p||q) \nonumber \\ \text{ subject to} \hspace{0.2cm} & \epsilon_j^- \leq g_j(Z) \leq \epsilon_j^+, \hspace{0.2cm} \forall j \in \{ 1,...,J \}, \label{eqn:p-star} \end{align} }% where $D_{KL}$ is the KL-divergence and $J$ is the number of constraints.\footnote{Note that in our examples $J$ will typically be $K+1$, i.e., one constraint for the $S$ to $Y$ paths and one constraint for each set of paths from $S$ to $A_k$. We allow for $J$ constraints in general to accommodate more complex settings (e.g., where there are multiple sensitive features, multiple outcomes, or a different set of pathways are constrained).} In finite sample settings, \citet{NabiShpitser18Fair} propose solving the following constrained maximum likelihood problem: {\small \begin{align} \widehat{\alpha} &= \arg \max_{\alpha} \hspace{0.1cm} {\cal L}(Z; {\alpha}) \nonumber \\ \text{ subject to} \hspace{0.2cm} & \epsilon_j^- \leq \widehat{g}_j(Z) \leq \epsilon_j^+, \hspace{0.2cm} \forall j \in \{ 1,...,J \}, \label{eqn:c-mle} \end{align} }% where $\widehat{g}_j(Z)$ are estimators for the chosen PSEs and ${\cal L}(Z; \alpha)$ is the likelihood function. The most relevant bounds in practice are $\epsilon_j^- = \epsilon_j^+ = 0$. Given an approximation of $p^*$ learned in this way, \citet{NabiShpitser18Fair} transform regression problems originally defined on $p$ into regression problems on $p^*$. In other words, instead of learning a regression function $\mathbb{E}[Y \mid M, A, S, X]$ on the observed data distribution $p(Z)$, they approximate the fair distribution $p^*(Z)$ by constrained maximum likelihood and classify new instances using the constrained model. As we discuss in more detail later, \citet{NabiShpitser18Fair} choose to average over certain variables to accomodate the fact that new instances are generated from $p$ rather than $p^*$. \subsection{Fair decision-making} In the sequential decision setting, there are multiple complications. In particular, we aim to learn high-quality policies while simultaneously making sure that the joint distribution induced by the policy satisfies our fairness criteria, potentially involving constraints on multiple causal pathways. This problem must be solved in settings where distributions of some variables, such as outcomes, are not under the policy-maker's control. Finally, we must show that if the learned policy is adapted to new instances (drawn from the original observed distribution) in the right way, then these new instances combined with the learned policy, constrained variables, and variables outside our control, together form a joint distribution where our fairness criteria remain satisfied. Consider a $K$-stage decision problem given by a DAG where every vertex pair is connected, and with vertices in a topological order $X, S, M, A_1, Y_1, \ldots, A_K, Y_K$. See Fig.\ 1(c). Note that the setting where $S$ can be assumed exogenous is a special case of this model with missing edge between $X$ and $S$. Though we only assume a single set of permissible mediators $M$ here, at the expense of some added cumbersome notation all of the following can be extended to the case where there are distinct sets of mediators $M_1, \ldots, M_K$ preceding every decision point. (We extend the results below to that setting in the Supplement.) We will consider the following PSEs as inadmissible: PSE$^{sy}$, representing the effect of $S$ on $Y$ along all paths \emph{other than} the paths of the form $S \to M \to \ldots \to Y$; and PSE$^{sa_k}$, representing the effect of $S$ on $A_k$ along all paths \emph{other than} the paths of the form $S \to M \to \ldots \to A_k$. That is, we consider \emph{only} pathways connecting $S$ and $A_k$ or $Y$ through the allowed mediators $M$ to be fair. In this model, these PSEs are identified by \cite{shpitser13cogsci}: {\small \begin{align*} \text{PSE}^{sy} &= \mathbb{E}[Y(s,M(s'))] - \mathbb{E}[Y(s')] \\ = \sum_{X,M} &\{ \mathbb{E}[Y | s,M,X] - \mathbb{E}[Y | s',M,X] \} p(M | s',X) p(X)\\ \\ \text{PSE}^{sa_k} &= \mathbb{E}[A_k(s,M(s'))] - \mathbb{E}[A_k(s')] \\ = \sum_{X,M} &\{ \mathbb{E}[ A_k | s,M,X] - \mathbb{E}[ A_k | s',M,X] \} p(M | s',X) p(X) \end{align*} }% Numerous approaches for estimating and constraining these identified PSEs are possible. In this paper, we restrict our attention to semiparametric estimators, which model only a part of the likelihood function while leaving the rest completely unrestricted. Estimators of this sort share some advantages with parametric methods (e.g., often being uniformly consistent at favorable rates), but do not require specification of the full probability model. Specifically, we use estimators based on the following result: \begin{thm} Assume $S$ is binary. Under the causal model above, the following are consistent estimators of PSE$^{sy}$ and PSE$^{sa_k}$, assuming all models are correctly specified: {\small \begin{align} &\widehat{g}^{sy}(Z) = \\ \nonumber &\frac{1}{N} \sum_{n=1}^{N} \Big\{ \frac{\mathbb{I}(S_n = s)}{p(S_n | X_n)} \ \frac{p(M_{n} | s', X_n)}{p(M_{n} | s, X_n)} - \frac{\mathbb{I}(S_n = s')}{p(S_n | X_n)} \Big\} \ Y_n \\ \nonumber \\ &\widehat{g}^{sa_k}(Z) =\\ \nonumber &\frac{1}{N} \sum_{n=1}^{N} \Big\{ \frac{\mathbb{I}(S_n = s)}{p(S_n | X_n)} \ \frac{p(M_{n} | s', X_n)}{p(M_{n} | s, X_n)} - \frac{\mathbb{I}(S_n = s')}{p(S_n | X_n)} \Big\} \ A_{kn} \end{align} } \label{thm:est} \end{thm} These inverse probability weighted (IPW) estimators use models for $M$ and $S$. Thus, we can approximate $p^*$ by constraining only the $M$ and $S$ models, i.e., obtaining estimates $\hat{\alpha}_{m}$ and $\hat{\alpha}_s$ of the parameters $\alpha_{m}$ and $\alpha_s$ in $p^*(M|S,X;\alpha_{m})$ and $p^*(S|X;\alpha_s)$ by solving (\ref{eqn:c-mle}). The outcomes $Y_k$ and decisions $A_k$ are left unconstrained. This is subtle and important, since it enables us to choose our optimal decision rules $f^*_A$ without restriction of the policy space and allows the mechanism determining outcomes $Y_k$ (based on decisions $A_k$ and history $H_k$) to remain outside the control of the policy-maker. Consequently, we can show that implementing this procedure guarantees that the joint distribution over all variables $Z$ induced by 1) the constrained $M$ and $S$ models, 2) the conditional distributions for $A_k$ given $H_k$ implied by the optimal policy choice, and 3) \emph{any} choice of $p(Y_k | A_k,H_k)$ will (at the population-level) satisfy the specified fairness constraints. We prove the following result in the Supplement: \begin{thm} Consider the K-stage decision problem described by the DAG in Fig.~\ref{fig:triangle}(c). Let $p^*(M|S,X; \alpha_{m})$ and $p^*(S|X; \alpha_s)$ be the constrained models chosen to satisfy PSE$^{sy} = 0$ and PSE$^{sa_k} = 0$. Let $\tilde{p}(Z)$ be the joint distribution induced by $p^*(M|S,X; \alpha_{m})$ and $p^*(S|X;\alpha_s)$, and where all other distributions in the factorization are unrestricted. That is, {\small \begin{align*} \tilde{p}(Z) \equiv p(X)p^*(S|X;\alpha_s) &p^*(M|S,X;\alpha_m)\\ &\times \prod_{k=1}^K p(A_k|H_k) p(Y_k|A_k,H_k).\\ \end{align*} }% Then the functionals PSE$^{sy}$ and PSE$^{sa_i}$ taken w.r.t.\ $\tilde{p}(Z)$ are also zero. \end{thm} This theorem implies that any approach for learning policies based on $\tilde{p}(Z)$ addresses both retrospective bias (since the fairness criterion violation present in $p(Z)$ is absent in $\tilde{p}(Z)$) and prospective bias (since the criterion holds in $\tilde{p}(Z)$ for any choice of policy on $A_k$ inducing $p(A_k | H_k)$). As we discuss in detail in the next section, modified policy learning based on $\tilde{p}(Z)$ requires special treatment of the constrained variables $S$ and $M$. New instances (e.g., new calls to the child protection hotline) will be drawn from the unfair distribution $p$, not $\tilde{p}$. So, the enacted policy cannot use empirically observed values of $S$ or $M$. In what follows, our approach is to either average over $S$ and $M$ (following \citet{NabiShpitser18Fair}), or resample observations of $S$ and $M$ from the constrained models. \section{Estimation of optimal policies in the fair world} \label{sec:estfairpolicy} In the following, we describe several strategies for learning optimal policies, and our modifications to these strategies based on the above fairness considerations. \subsection{Q-learning} In Q-learning, the optimal policy is chosen to optimize a sequence of counterfactual expectations called Q-functions. These are defined recursively in terms of value functions $V_k(\cdot)$ as follows: {\small \begin{align} Q_K(H_K, A_K) &= \mathbb{E}[Y_K(A_K) \mid H_K], \nonumber \\ V_K(H_K) &= \max_{a_K} Q_K(H_K, a_K), \label{eq:QK} \end{align} } and for $k = K-1, \ldots, 1$ {\small \begin{align} Q_k(H_k, A_k) &= \mathbb{E}[V_{k+1}(H_{k+1},A_k) \mid H_k], \nonumber \\ V_k(H_k) &= \max_{a_k} Q_k(H_k, a_k). \label{eq:Qi} \end{align} }% Assuming $Q_k(H_k, A_k)$ is parameterized by $\beta_k$, the optimal policy at each stage may be easily derived from Q-functions as $f^*_{A_k}(H_k) = \argmax_{a_k} {Q}_k(H_k, a_k; \widehat{\beta}_k)$. Q-functions are recursively defined regression models where outcomes are value functions, and features are histories up to the current decision point. Thus, parameters $\beta_k$ $(k=1,\ldots,K)$ of all Q-functions may be learned recursively by maximum likelihood methods applied to regression at stage $k$, given that the value function at stage $k+1$ was already computed for every row. See \citet{Moodie13DTR} for more details. Note that at each stage $k$, the identity $Q_k(H_k, A_k) = \mathbb{E}[V_{k+1}(H_{k+1},A_k) \mid H_k] = \mathbb{E}[V_{k+1}(H_{k+1}) \mid A_k, H_k]$ only holds under our causal model if the \emph{entire past} $H_k$ is conditioned on. In particular, $\mathbb{E}[V_{k+1}(H_{k+1},A_k) \mid H_k \setminus \{ M, S \} ] \neq \mathbb{E}[V_{k+1}(H_{k+1}) \mid A_k, H_k \setminus \{M, S \} ]$. To see a simple example of this, note that $Y_K(a_1)$ is not independent of $A_1$ conditional on just $X$ in Fig.~\ref{fig:triangle}(c), due to the presence of the path $Y_K \gets M \to A_1$; however the indepenence does hold conditional on the entire $H_1 = \{X, S, M\}$ \cite{thomas13swig}. In a fair policy learning setting, though $\{M, S\}$ may be in $H_k$, we cannot condition on values of ${M, S}$ to learn fair policies since these values were drawn from $p$ rather than $p^*$. There are multiple ways of addressing this issue. One approach is to modify the procedure to obtain optimal policies that condition on all history \emph{other than} $\{M, S\}$. We first learn $Q_k$s using (\ref{eq:QK}) and (\ref{eq:Qi}). We then provide the following modified definition of Q-functions defined directly on $p^*$: {\small \begin{align*} & Q^{*}_k(H_k \setminus \{M, S\}, A_k; \beta_k) = \\ & \frac{1}{Z} \ \sum_{m, s} Q_k(H_k, A_k; \beta_k) \ \prod_{i = 1}^{k} p(A_i | H_i \setminus \{M, S\}, m, s) . \\ & \prod_{i = 2}^{k - 1} p(M_i | A_i, H_i \setminus \{M, S\}, m, s) \ p^*(m, s | X), \end{align*} }% for $k = K, \ldots, 1$, {\small \begin{align*} Z &= \sum_{m, s} p^*(m, s | X) \ \prod_{i = 1}^{k} p(A_i | H_i \setminus \{M, S\}, m, s) \\ & \prod_{i = 2}^{k - 1} p(M_i | A_i, H_i \setminus \{M, S\}, m, s). \end{align*} }% The optimal fair policy at each stage is then derived from $Q^{*}$-functions as $f^*_{A_k}(H_k) = \argmax_{a_k} {Q^{*}}_k(H_k \setminus \{M, S\}, a_k; \widehat{\beta}_k)$. As an alternative approach, we can compute the original $Q$-functions defined in (\ref{eq:QK}) and (\ref{eq:Qi}) with respect to $p^{*}(Z)$ by ignoring the observed values $M_n$ and $S_n$ for the $n$th individual and replacing them with samples drawn from $p^{*}(M|S,X;\alpha_m)$ and $p^{*}(S|X;\alpha_s)$. Then, in (\ref{eq:QK}) and (\ref{eq:Qi}), the history at the $k$th stage, $H_k$, gets replaced with $H^{*}_k = \{H_k \setminus \{M, S\}, M^{*}, S^{*} \}$. \subsection{Value search} It may be of interest to estimate the optimal policy within a restricted class ${\cal F}$. One approach to learning the optimal policy within ${\cal F}$ is to directly search for the optimal ${ f}^{*,{\cal F}}_{ A} \equiv \argmax_{{f}_{A} \in {\cal F}} \mathbb{E}[Y({f}_{A})]$, which is known as \textit{value search}. The expected response to an arbitrary policy $\phi = \mathbb{E}[Y({f}_{A})]$, for ${f_A} \in {\cal F}$ can be estimated in a number of ways. Often $\widehat{\phi}$ takes the form of a solution to some estimating equation $\mathbb{E}[ h(\phi) ] = 0$ solved empirically given samples from $p(Z)$. A simple estimator for $\phi$ that uses only the treatment assignment model $\pi(H_k; \psi) \equiv p(A_k=1 | H_k)$ is the IPW estimator that solves the following estimating equation: {\small \begin{align} \label{eqn:ipw} \mathbb{E}\left[ \prod_{k = 1}^{K} \left\{ C_{f_{A_k}} / \pi_{f_{A_k}}(H_k; \widehat{\psi}) \right\} \times Y - \phi \right] = 0, \end{align} }% where $\small C_{f_{A_k}} \equiv \mathbb{I}(A_k = f_{A_k}(H_k))$, $\pi_{f_{A_k}}(H_k;\psi) \equiv \pi(H_k;\psi) f_{A_k}(H_k) + (1-\pi(H_k;\psi)) (1 - f_{A_k}(H_k))$, the expectation is evaluated empirically, and $\widehat{\psi}$ is fit by maximum likelihood. Finding fair policies via value search involves solving the same problem with respect to $p^*(Z)$ instead. Given known models $p^*(M | S,X; \alpha_m)$ and $p^*(S | X; \alpha_s)$, we may consider two approaches. The first one involves solving a modified estimating equation of the form {\small \begin{align*} &\mathbb{E}^*[ h(\phi) ] \equiv\\ &\mathbb{E}\left[ \sum_{m, s} \mathbb{E}[ h(\phi) | M,S,X ] p^*(M | S,X; \alpha_m) p^*(S | X; \alpha_s) \right] = 0 \end{align*} }% with respect to $p^*(Z \setminus \{M, S \})$. The alternative is to solve the original estimating equation $\mathbb{E}[ h(\phi) ] = 0$ with respect to $p^*(Z)$ by replacing observed values $M_n$ and $S_n$ for the $n$th individual with sampled values $M^*_n$ and $S^*_n$ drawn from $p^{*}(M|S,X;\alpha_m)$ and $p^{*}(S|X;\alpha_s)$. In both approaches, the optimal fair policy at each stage is then derived by replacing the history at the $k$th stage, $H_k$, with $H^*_k = \left\{H_k \setminus \{M, S\}, M^*, S^* \right\}$. Given constrained models $p^*(M | S,X; \alpha_m)$, and $p^*(S | X; \alpha_s)$ representing $p^*(Z)$, we can perform value search by solving the given estimating equation empirically on a dataset where every row $x_n, s_n, m_n$ in the data is replaced with $I$ rows $x_n, s^*_{ni}, m^*_{ni}$ for $i = 1, \ldots I$, with $m^*_{ni}$ and $s^*_{ni}$ drawn from $p^*(M | S, x_n; \alpha_m)$ and $p^*(S | x_n; \alpha_s)$, respectively. \subsection{G-estimation} A third method for estimating policies is to directly model the counterfactual contrasts known as \emph{optimal blip-to-zero functions} and then learn these functions by a method called g-estimation \citep{robins04SNMM}. In the interest of space, we defer a full description of blip-to-zero functions and g-estimation to the Supplement, where we also present some results for our implementation of fair g-estimation. \subsection{Tradeoffs and treatment of constrained variables} We've proposed to constrain the $M$ and $S$ models to satisfy given fairness constraints. Since empirically observed values of $M$ and $S$ are sampled from $p$ rather than $p^*$ (or $\tilde{p}$), our approach requires resampling or averaging over these features. The choice of models to constrain involves a tradeoff. The more models are constrained, the closer the KL distance between $p$ and $p^*$, but the more features have to be resampled or averaged out; that is, some information on new instances is ``lost.'' Alternative approaches may constrain fewer or different models in the likelihood (for example, we could have elected to constrain the $Y$ model instead of $S$). However, the benefit of our approach here is that we can guarantee, with outcomes $Y$ outside the policy-maker's control, that the induced joint distribution will satisfy the given fairness constraints (by Theorem 2), whereas alternative procedures which aim to avoid averaging or resampling will typically have no such guarantees. Another alternative that avoids averaging over variables altogether is to consider likelihood parameterizations where the absence of a given PSE directly corresponds to setting some variation-independent likelihood parameter for the $Y$ model to zero (cf.\ \citet{chiappa2018path}). While such a parameterization is possible for linear structural equation models, it is an open problem in general for arbitrary PSEs and nonlinear settings. Developing novel, general-purpose alternatives that transfer observed distributions to their ``fair versions,'' while avoiding resampling and averaging, is an open problem left to future work. \section{Experiments} \label{sec:exp} \subsection*{Synthetic data} We generated synthetic data for a two-stage decision problem according to the causal model shown in Fig.~\ref{fig:triangle}(c) ($K = 2$), where all variables are binary except for the continuous response utility $Y \equiv Y_2$. Details on the specific models used are reported in the Supplement. We generated a dataset of size $5,000$, with $100$ bootstrap replications, where the sensitive variable $S$ is randomly assigned and where $S$ is chosen to be an informative covariate in estimating $Y$. We use estimators in Theorem \ref{thm:est} to compute PSE$^{sy}$, PSE$^{sa_1}$, and PSE$^{sa_2}$ which entail using $M$ and $S$ models. In this setting, the PSE$^{sy}$ is $1.918$ (on the mean scale) and is restricted to lie between $-0.1$ and $0.1$. The PSE$^{sa_1}$ is $0.718$, and PSE$^{sa_2}$ is $0.921$ (on the odds ratio scale) and both are restricted to lie between $0.95$ and $1.05$. We only constrain $M$ and $S$ models to approximate $p^*$ and fit these two models by maximizing the constrained likelihood using the R package \texttt{nloptr}. The parameters in all other models were estimated by maximizing the likelihood. Optimal fair polices along with optimal (unfair) policies were estimated using the two techniques described in Section \ref{sec:estfairpolicy} (where we used the ``averaging'' approach in both cases). We evaluated the performance of both techniques by comparing the population-level response under fair policies versus unfair policies. One would expect the unfair policies to lead to higher expected outcomes compared to fair policies since satisfying fairness constraints requires sacrificing some policy effectiveness. The expected outcomes under unfair polices obtained from Q-learning and value search were $7.219 \!\pm\! 0.005$ and $7.622 \!\pm\! 0.265$, respectively. The values dropped to $6.104 \!\pm\! 0.006$ and $6.272 \!\pm\! 0.133$ under fair polices, as expected. In addition, both fair and unfair optimal polices had higher expected outcomes than the observed population-level outcome, using both methods. In our simulations, the population outcome under observed policies was $4.82 \!\pm\! 0.007$. Some additional results are reported in the Supplement. \subsection*{Application to the COMPAS dataset} COMPAS is a criminal justice risk assessment tool created by the company Northpointe that has been used across the US to determine whether to release or detain a defendant before their trial. Each pretrial defendant receives several COMPAS scores based on factors including but not limited to demographics, criminal history, family history, and social status. Among these scores, we are primarily interested in the ``risk of recidivism." We use the data made available by Propublica and described in \citet{angwin16machine}. The COMPAS risk score for each defendant ranges from 1 to 10, with 10 being the highest risk. In addition to this score ($A$), the data also includes records on defendant’s age ($X_1 \in X$), gender ($X_2 \in X$), race ($S$), prior convictions ($M$), and whether or not recidivism occurred in a span of two years ($R$). We limited our attention to the cohort consisting of African-Americans and Caucasians, and to individuals who either had not been arrested for a new offense or who had recidivated within two years. Our sample size is 5278. All variables were binarized including the COMPAS score, which we treat as an indicator of a binary decision to incarcerate versus release (pretrial) ``high risk'' individuals, i.e., we assume those with score $\geq 7$ were incarcerated. In this data, 28.9\% of individuals had scores $\geq 7$. Since the data does not include any variable that corresponds to utility, and there is no uncontroversial definition of what function one should optimize, we define a heuristic utility function from the data as follows. We assume there is some (social, economic, and human) cost, i.e., negative utility, associated with incarceration (deciding $A=1$), and that there is some cost to releasing individuals who go on to reoffend (i.e., for whom $A=0$ and $R=1$). Also, there is positive utility associated with releasing individuals who do not go on to recidivate (i.e., for whom $A=0$ and $R=0$). A crucial feature of any realistic utility function is how to balance these relative costs, e.g., how much (if any) ``worse'' it is to release an individual who goes on to reoffend than to incarcerate them. To model these considerations we define utility $Y \equiv (1-A) \times \{ -\theta R + (1-R) \} - A$. The utility function is thus parameterized by $\theta$, which quantifies how much ``worse'' is the case where individuals are released and reoffend as compared with the other two possibilities which are treated symmetrically. We emphasize that this utility function is a heuristic we use to illustrate our optimal policy learning method, and that a realistic utility function would be much more complicated (possibly depending also on factors not recorded in the available data). We apply our proposed Q-learning procedure to optimize $\mathbb{E}[Y]$, assuming $K=1$ and exogenous $S$. The fair policy constrains the $S \rightarrow A$ and $S \rightarrow Y$ pathways. We describe details of our implementation as well as additional results in the Supplement. The proportion of individuals incarcerated ($A=1$) is a function of $\theta$, which we plot in Fig.\ 2 stratified by racial group. See the Supplement for results on \emph{overall} incerceration rates, which also vary among the policies. The region of particular interest is between $\theta=2$ and $3$, where fair and unrestricted optimal policies differ and both recommend lower-than-observed overall incarceration rates (see Supplement). For most $\theta$ values, the fair policy recommends a decision rule which narrows the racial gap in incarceration rates as compared with the unrestricted policy, though does not eliminate this gap entirely. (Constraining the causal effects of race through mediator $M$ would go further in eliminating this gap.) In regions where $\theta >3$, both optimal policies in fact recommend higher-than-observed overall incarceration rates but a narrower racial gap, particularly for the fair policy. Comparing fair and unconstrained policy learning on this data serves to simultaneously illustrate how the proposed methods can be applied to real problems and how the choice of utility function is not innocuous. \begin{figure}[t!] \begin{center} \includegraphics[scale=.32]{GroupRates.pdf \end{center} \caption{Group-level incarceration rates for the COMPAS data as a function of the utility parameter $\theta$.} \vspace{-4.7mm} \label{sim} \end{figure} \section{Conclusion} \label{sec:conc} We have extended a formalization of algorithmic fairness from \citet{NabiShpitser18Fair} to the setting of learning optimal policies under fairness constraints. We show how to constrain a set of statistical models and learn a policy such that subsequent decision making given new observations from the ``unfair world'' induces high-quality outcomes while satisfying the specified fairness constraints in the induced joint distribution. In this sense, our approach can be said to ``break the cycle of injustice'' in decision-making. We investigated the performance of our proposals on synthetic and real data, where in the latter case we have supplemented the data with a heuristic utility function. In future work, we hope to develop and implement more sophisticated constrained optimization methods, to use information as efficiently as possible while satisfying the desired theoretical guarantee, and to explore nonparametric techniques for complex settings where the likelihood is not known. \clearpage \section*{Acknowledgments} This project is sponsored in part by the National Institutes of Health grant R01 AI127271-01 A1 and the Office of Naval Research grant N00014-18-1-2760. {\small \section*{Appendix A: G-estimation} G-estimation applies to structural nested models, which directly model the counterfactual deviations in outcome from a reference treatment value (which we take to be $A=0$) conditional on history, assuming all future decisions are already optimal. Specifically, for each decision point $k$ we posit a \emph{structural nested mean model (SNMM)} parameterized by $\psi$ as follows: {\small \begin{align*} \gamma_k(H_k, &a_k; \psi) =\\ &\mathbb{E}[ Y(\bar{a}_{k-1}, a_k, f^*_{\underline{A}_{k+1}}) - Y(\bar{a}_{k-1}, a_k=0, f^*_{\underline{A}_{k+1}}) \mid H_k ], \end{align*} }% where $\underline{A}_{k+1}$ represents all treatments administered from time $k+1$ onwards. In words, $\gamma_k$ is the contrast of the counterfactual mean (conditional on observed history $H_k$) where the past decisions are set to their observed values, the present decision is either $a_k$ or a reference decision $a_k = 0$, and all future decisions are made optimally, $f^*_{\underline{A}_{k+1}}$. Note that if the true $\gamma_k(H_k,a_k;\psi)$ were known, the optimal treatment policies are those that maximize this ``blip'' function at each stage: $f^*_{A_k} = \argmax_{a_k} \gamma_k(H_k, a_k; \psi)$. In order to estimate $\psi$ using data, let {\small \begin{align} U({\bf \psi, \zeta(\psi), \alpha}) = \sum_{k=1}^{K} &\left\{ G_k({\bf \psi}) - \mathbb{E}\left[G_k({\bf \psi}) \mid H_k; {\bf \zeta} \right] \right\} \nonumber \\ \times \ &\left\{ d_k(H_k, A_k) - E\left[d_k(H_k, A_k) \mid H_k; {\bf \alpha}\right] \right\}, \label{eqn:G-estim} \end{align} }% where $d_k(H_k, A_k)$ is any function of $H_k$ and $A_k$ and $G_k({\bf \psi})$ is defined as {\small \begin{align*} Y - \gamma_k(H_k, a_k; {\bf \psi}) + \sum_{i = k + 1}^{K} \left[ \gamma_i(H_i, a^*_i; {\bf \psi}) - \gamma_i(H_i, a_i; {\bf \psi}) \right], \end{align*} }% ($a^*_i$ is the optimal decision at $i$th stage). Consistent estimators of ${\bf \psi}$ can be obtained solving the estimating equations $\mathbb{E}[U({\bf \psi, \zeta(\psi), \alpha})] = 0$, as shown in \citet{robins04SNMM}. Both of the modifications discussed for Q-learning and value search must be applied when learning fair optimal policies by g-estimation. Specifically, we determine optimal polities not from the SNMM contrast $\gamma_k(H_k, a_k; \psi) = \mathbb{E}[ Y(\bar{a}_{k-1}, a_k, f^*_{\underline{A}_{k+1}}) - Y(\bar{a}_{k-1}, a_k = 0, f^*_{\underline{A}_{k+1}})\mid H_k ]$ itself, but rather from a modified contrast $\gamma^*_k(H_k \setminus M, a_k; \psi) = \sum_{m, s} \gamma_k(H_k, a_k; \psi) p^*(M | S,X) p^*(S | X) = \mathbb{E}[ Y(\bar{a}_{k-1},a_k, f^*_{\underline{A}_{k+1}}) - Y(\bar{a}_{k-1}, a_k = 0, f^*_{\underline{A}_{k+1}})\mid H_k \setminus \{M, S\} ]$ which does not use $M$ and $S$. This is analogous to removing $M$ and $S$ from the Q-functions defined in Section 4 and is done for the same reason: $M, S$ are drawn from $p(Z)$, not $p^*(Z)$. Second, the estimating equations for $\psi$ must use constrained models (in particular for $M$ and $S$), and must be empirically solved using observations only from $p^*(Z)$. As was done with value search, we solve equation (\ref{eqn:G-estim}) empirically using a dataset where each row $x_n,s_n,m_n$ is replaced by $I$ rows of the form $x_n,s^*_{ni},m^*_{ni}$, $i = 1, \ldots, I$, with $s^*_{ni}$ and $m^*_{ni}$ drawn from $p^*(S | x_n; \alpha_s)$ and $p^*(M | x_n, S; \alpha_m)$, respectively. \section*{Appendix B: Simulation details and additional results on synthetic data} Here we report the precise parameter settings used in our simulation studies. The following regression models were used in our simulation study of the two-stage decision problem: {\small \begin{align*} X_1 &\sim | \mathcal{N}(0, 1) | \nonumber \\ (X_2, X_3) &\sim \mathcal{N}(0, \text{diag}(2)) \nonumber \\ S &\sim \text{Bernoulli}(p = 0.5) \nonumber \\ \text{logit}(p(M = 1)) &\sim -1 + X_1 + X_2 + X_3 + S\\ &+ 3SX_1 + SX_2 + SX_3 \nonumber \\ \text{logit}(p(A_1 = 1)) &\sim 1 - X_1 + X_2 + S + M - SX_1 + SX_2\\ &+ MS - 3MX_1 + 0.5MX_2 \nonumber \\ \text{logit}(p(Y_1 = 1)) &\sim -2 + X_1 + X_2 + S + M + A\\ &+ SX_2 + MS + AS + AM \nonumber \\ \text{logit}(p(A_2 = 1)) &\sim 1 - X_1 + X_2 + M + A + W\\ &+ S(1 - X_1 + X_2 + M - A) \\ &- 3MX_1 + 0.5MX_2 - AX_1 - AX_2 \nonumber \\ Y& = 2.5 + X_1 + X_2 + M + W + B\\ &+ S(1+ X_1+ X_2 + M + A + W) \\ &+ A(1 + M - 2W) + MW\\ &+ B(-X_1 + 2X_2 - M) + WX_1 + \mathcal{N}(0, 1) \end{align*} }% For this two-stage setting we estimated the optimal policies using Q-learning and value search. In value search, we considered restricted class of polices of the form $p(A_1 = 1 | X, S, M) = -1 + \alpha_x X + \alpha_s S + \alpha_m M + \alpha_{sx} SX + \alpha_{sm} SM + \alpha_{mx} MX $, and $p(A_2 = 1 | X, S, M, A_1, Y_1) = -1 + \alpha_x X + \alpha_s S + \alpha_m M + \alpha_a A + \alpha_{y_1} Y_1 + \alpha_{sx} SX + \alpha_{sm} SM + \alpha_{mx} MX + \alpha_{as} AS + \alpha_{ax} AX$ where all $\alpha$s range from $-3$ to $3$ by $0.5$ increments and estimated the value of policies for each combination of $\alpha$s using equation (\ref{eqn:ipw}). A third method for estimating policies is to directly model the counterfactual contrasts known as \emph{optimal blip-to-zero functions} and then learn these functions by g-estimation \citep{robins04SNMM}; see Appendix A. We implemented our modified fair g-estimation for a single-stage decision problem and compared the results with Q-learning and value search. The results are provided in Table 1. The data generating process for the single-stage decision problem matches the causal model shown in Fig.~\ref{fig:triangle}(a) where $X, S, M,$ and $A$ were generated the same way as described above. The outcome $Y$ was generated from a standard normal distribution with mean $-2 + X + S + M + A -3SX_2 + MS + AS + AM + AX_2 + AX_3$. We used estimators in Theorem \ref{thm:est} to compute PSE$^{sy}$ and PSE$^{sa}$ which require using $M$ and $S$ models. In this synthetic data, the PSE$^{sy}$ was $1.618$ (on the mean scale) and was restricted to lie between $-0.1$ and $0.1$. The PSE$^{sa}$ was $0.685$ (on the odds ratio scale) and was restricted to lie between $0.95$ and $1.05$. \begin{table}[t!] \label{tab:outcome} \centering \begin{tabular}{ c | c | c } & \small \textbf{Unfair Policy} & \small \textbf{Fair Policy} \\ [0.1cm] \hline \small \textbf{Q-learning } & $ 1.414 \!\pm\! 0.0056 $ & $ 1.189 \!\pm\! 0.0059 $ \\ [0.1cm] \small \textbf{value search } & $ 1.134 \!\pm\! 0.0245 $ & $ 1.056 \!\pm\! 0.0299 $ \\ [0.1cm] \small \textbf{g-estimation } & $ 1.375 \!\pm\! 0.0099 $ & $ 1.312 \!\pm\! 0.0102 $ \end{tabular} \caption{{\small Comparison of population outcomes $\mathbb{E}[Y]$ under policies learned by different methods. The value under the observed policy was $ 0.24 \!\pm\! 0.006 $. }} \end{table} \section*{Appendix C: Details and additional results on the COMPAS data experiment} The regression models we used in the COMPAS data analysis were specified as follows: {\small \begin{align*} \text{logit}(p(M=1)) &\sim X_1 + X_2 + S + SX_1 + SX_2\\ \text{logit}(p(A=1)) &\sim X_1 + X_2 + S + M + SX_1\\ & \hspace{0.1cm} + SX_2 + MS + MX_1 + MX_2\\ Y &\sim X_1 + X_2 + S + M + A + SX_1 + SX_2 \\ & \hspace{0.1cm} + AS + AM + MS + MX_1 + MX_2\\ & \hspace{0.1cm} + AX_1 + AX_2 \end{align*} }% For estimating the PSEs which we constrain, we used the same IPW estimators described in the main paper and reproduced in the theorem below. We constrained the PSEs to lie between $-0.05$ and $0.05$ and $0.95$ and $1.05$, respectively. In Fig.\ 3, we compare the overall incarceration rates recommended by the optimal fair and unconstrained policies on the COMPAS data, as a function of the utility parameter $\theta$. For low values of $\theta$ the incarceration rate is zero, and becomes higher as $\theta$ increases, but differentially for the fair and unconstrained optimal policies. The difference between the policies depends crucially on the utility function. For some values of the utility parameter, the unfair and fair policies coincide, but for other values we would expect significantly different overall incarceration rates as well as different disparities between racial groups (see result in the main paper). In Fig.\ 4, we show the relative utility achieved by the optimal fair and unconstrained policies, as well as the utility of the observed decision pattern, as a function of $\theta$. As expected, choosing an optimal policy improves on the observed policy, with the unfair (unconstrained) choice being higher utility than the fair (constrained) choice; we sacrifice some optimality to satisfy the fairness constraints. However, the difference depends on the utility parameter and for a range of parameter values the fair and unfair policies are nearly the same in terms of optimality (even when they may disagree on the resulting incarceration rate, around $\theta=2.6$). The fair and unfair policies drift far apart in terms of utility around $\theta=3$, when the policies recommend an incarceration rate comparable to or higher than the observed rate. \addtocounter{figure}{2} \begin{figure}[t!] \begin{center} \includegraphics[scale=.32]{OverallRate.pdf} \end{center} \caption{Overall incarceration rates for the COMPAS data as a function of the utility parameter $\theta$.} \end{figure} \begin{figure}[t!] \begin{center} \includegraphics[scale=.32]{ExpectedUtility.pdf} \end{center} \caption{The relative utility of policies for the COMPAS data as a function of the utility parameter $\theta$.} \end{figure} \newpage \section*{Appendix D: Proofs} \begin{thm Assume $S$ is binary. Under the causal model above, the following are consistent estimators of PSE$^{sy}$ and PSE$^{sa_k}$, assuming all models are correctly specified: {\small \begin{align} &\widehat{g}^{sy}(Z) = \\ \nonumber &\frac{1}{N} \sum_{n=1}^{N} \Big\{ \frac{\mathbb{I}(S_n = s)}{p(S_n | X_n)} \ \frac{p(M_{n} | s', X_n)}{p(M_{n} | s, X_n)} - \frac{\mathbb{I}(S_n = s')}{p(S_n | X_n)} \Big\} \ Y_n \\ \nonumber \\ &\widehat{g}^{s{a}_k}(Z) =\\ \nonumber &\frac{1}{N} \sum_{n=1}^{N} \Big\{ \frac{\mathbb{I}(S_n = s)}{p(S_n | X_n)} \ \frac{p(M_{n} | s', X_n)}{p(M_{n} | s, X_n)} - \frac{\mathbb{I}(S_n = s')}{p(S_n | X_n)} \Big\} \ A_{kn} \end{align} }% \end{thm} \begin{prf} The latent projection \cite{verma90equiv} of any $K$ stage DAG onto $X,S,M,A,Y$ suffices to identify and estimate the two path-specific effects in question, and this latent projection is the complete DAG with topological ordering $X,S,M,A,Y$. The consistency of the estimators above then follows directly from derivations in \cite{tchetgen12semi}. As an example, we have the following derivation for the first term of $g^{sy}(Z)$: {\small \begin{align*} & \sum_{X,M} \mathbb{E}\left[ Y| s, M, X \right]p(M|s',X)p(X)\\ &= \sum_{X,M,A,Y} Y p(Y | s,M,A,X) p(A | s,M,X) p(M | s',X) p(X)\\ &= \sum_{X,S,M,A,Y} \frac{\mathbb{I}(S = s)p(M | s', X)}{p(S | X) p(M | s, X)} Y dp(Y,S,M,A,X)\\ &= \mathbb{E}\left[ \frac{\mathbb{I}(S = s)p(M | s', X)}{p(S | X) p(M | s, X)} Y \right]\\ \end{align*} }% which is precisely the identifying functional for the first term of the PSE we are interested in. That the above estimator is consistent for this functional is a standard result. \end{prf} \begin{thm} Consider the K-stage decision problem described by the DAG in Fig 1c. Let $p^*(M|S,X; \alpha_{m})$ and $p^*(S|X; \alpha_s)$ be the constrained models chosen to satisfy PSE$^{sy} = 0$ and PSE$^{sa_k} = 0$. Let $\tilde{p}(Z)$ be the joint distribution induced by $p^*(M|S,X; \alpha_{m})$ and $p^*(S|X;\alpha_s)$, and where all other distributions in the factorization are unrestricted. That is, {\small \begin{align*} \tilde{p}(Z) \equiv p(X)p^*(S|X;\alpha_s) &p^*(M|S,X;\alpha_m)\\ &\times \prod_{k=1}^K p(A_k|H_k) p(Y_k|A_k,H_k).\\ \end{align*} }% Then the functionals PSE$^{sy}$ and PSE$^{sa_k}$ taken w.r.t.\ $\tilde{p}(Z)$ are also zero. \end{thm} \begin{prf} Let $Y \equiv Y_K$. Because $M$ preceeds all $A_k,Y_k$ for $k = 1, \ldots K$, it suffices to consider the latent projection with only variables $X,S,M,A,Y$ without affecting identifiability considerations. Then we have the following: {\small \begin{align*} &\widetilde{\text{PSE}}^{sy} \\ &= \tilde{\mathbb{E}}[Y(s,M(s'))] - \tilde{\mathbb{E}}[Y(s')] \\ &= \sum_{X,M} \{ \tilde{\mathbb{E}}[Y | s,M,X] - \tilde{\mathbb{E}}[Y | s',M,X] \} p^*(M | s',X;{\alpha}_m) p(X) \\ &= \sum_{X,M} \{ {\mathbb{E}}[Y | s,M,X] - {\mathbb{E}}[Y | s',M,X] \} p^*(M | s',X;{\alpha}_m) p(X) \\ &= \sum_{X,M,Y} Y \{ p(Y | s,M,X) - p(Y | s',M,X) \} p^*(M | s',X;{\alpha}_m) p(X) \\ &= \sum_{X,S,M,Y} Y \Big\{ \frac{\mathbb{I}(S = s)}{p^*(S | X;\alpha_s)} \ \frac{p^*(M | s', X;\alpha_m)}{p^*(M | s, X;\alpha_m)} - \frac{\mathbb{I}(S = s')}{p^*(S | X;\alpha_s)} \Big\} \\ & \times p(Y | M,S,X) p^*(M | S,X; \alpha_m) p^*(S | X; \alpha_s) p(X)\\ &=0 \end{align*} }% by choice of $p^*(M |S, X;\alpha_m)$ and $p^*(S | X;\alpha_s)$. The proof is structurally the same for $\widetilde{\text{PSE}}^{sa_k}$. \end{prf} \section*{Appendix E: Modified results with multiple sets of mediators} In the main paper, we discussed a $K$-stage decision problem with one set of permissible mediators, $M$. Here, we extend those results to the setting where we have multiple sets of mediators $M_1, \ldots, M_K$, i.e., a DAG with topological ordering $X, S, M_1, A_1, Y_1, \ldots, M_K, A_K, Y_K$. In this case, we consider the following paths impermissible: PSE$^{sy}$, representing the effect of $S$ on $Y$ along all paths \emph{other than} the paths of the form $S \to M_k \to \ldots \to Y$ ($\forall k$); and PSE$^{sa_k}$, representing the effect of $S$ on $A_k$ along all paths \emph{other than} the paths of the form $S \to M_j \to \ldots \to A_k$ ($\forall j \leq k$). That is, we consider \emph{only} pathways connecting $S$ and $A_k$ or $Y$ through the allowed mediators $M_1, \ldots, M_K$ to be fair. In this case, the PSEs are identified by a modification of the previous formula given in Section 3.2. {\small \begin{align*} &\text{PSE}^{sy} = \mathbb{E}[Y(s,M_1(s'),\ldots,M_K(s'))] - \mathbb{E}[Y(s')] \\ &= \sum_{x,\overline{m}_K,\overline{a}_{K-1},\overline{y}_{K-1},} \{ \mathbb{E}[Y | s,\overline{M}_K, \overline{A}_{K-1},\overline{Y}_{K-1},X]\\ &- \mathbb{E}[Y | s', \overline{M}_K, \overline{A}_{K-1},\overline{Y}_{K-1},X] \} \prod_{k=1}^K p(M_k | s',\overline{A}_{k-1},\overline{Y}_{k-1},X)\\ &\times \prod_{k=1}^{K-1} p(A_k | s, \overline{M}_k,\overline{A}_{k-1},\overline{Y}_{k},X) p(Y_k | s,\overline{M}_k,\overline{A}_{k},\overline{Y}_{k-1},X) p(X) \\ &\mbox{} \end{align*} } \vspace{-1.5cm} {\small \begin{align*} &\text{PSE}^{sa_k} = \mathbb{E}[A_k(s,M_1(s'),\ldots,M_K(s'))] - \mathbb{E}[A_k(s')] \\ &= \sum_{x,\overline{m}_k,\overline{a}_{k-1},\overline{y}_{k-1},} \{ \mathbb{E}[A_k | s,\overline{M}_k, \overline{A}_{k-1},\overline{Y}_{k-1},X]\\ &- \mathbb{E}[A_k | s', \overline{M}_k, \overline{A}_{k-1},\overline{Y}_{k-1},X] \} \prod_{k=1}^K p(M_k | s',\overline{A}_{k-1},\overline{Y}_{k-1},X)\\ &\times \prod_{j=1}^{k-1} p(A_j | s, \overline{M}_j,\overline{A}_{j-1},\overline{Y}_{j},X) p(Y_j | s,\overline{M}_j,\overline{A}_{j},\overline{Y}_{j-1},X) p(X) \end{align*} }% With these definitions, we can replace the estimators in Theorem 1 with: {\small \begin{align*} &\widehat{g}^{sy}(Z) = \\ \nonumber &\frac{1}{N} \sum_{n=1}^{N} \Big\{ \frac{\mathbb{I}(S_n = s)}{p(S_n | X_n)} \ \prod_{k=1}^K \frac{p(M_{k,n} | s', \overline{A}_{k-1,n},\overline{Y}_{k-1,n}, X_n)}{p(M_{k,n} | s, \overline{A}_{k-1,n},\overline{Y}_{k-1,n}, X_n)}\\ &- \frac{\mathbb{I}(S_n = s')}{p(S_n | X_n)} \Big\} \ Y_n \\ \\ &\widehat{g}^{sa_k}(Z) =\\ \nonumber &\frac{1}{N} \sum_{n=1}^{N} \Big\{ \frac{\mathbb{I}(S_n = s)}{p(S_n | X_n)} \ \prod_{k=1}^K \frac{p(M_{k,n} | s', \overline{A}_{k-1,n},\overline{Y}_{k-1,n}, X_n)}{p(M_{k,n} | s, \overline{A}_{k-1,n},\overline{Y}_{k-1,n}, X_n)} \\ &- \frac{\mathbb{I}(S_n = s')}{p(S_n | X_n)} \Big\} \ A_{kn} \nonumber \end{align*} }% Then, in Theorem 2 we analogously define $\tilde{p}(Z)$ as follows: {\small \begin{align*} \tilde{p}(Z) \equiv p(X)p^*(S|X;\alpha_s) \prod_{k=1}^K \Big\{ &p^*(M_k|S,\overline{A}_{k-1},\overline{Y}_{k-1},X;\alpha_{m})\\ &\times p(A_k|H_k) p(Y_k|A_k,H_k) \Big\}.\\ \end{align*} }% In this case we constrain the $S$ and $M_k$ models $\forall k$, the rest of the procedure remaining the same. Aside from the form of the identifying functional, the proofs of modified versions of Theorem 1 and Theorem 2 are analogous. {\small \bibliographystyle{plainnat}

72b36681f64f627beb5dc52c2d1ab645480a9262

\section{Conclusion} \label{conclusion} In this work we have explored the possibility of reproducing the measured relic DM abundance from (p)reheating. This scenario is interesting because it does not rely on interaction of DM particles with the SM particles, as required for standard freeze-out mechanism, of which no hint has been observed in direct and indirect searches. While the production of DM during reheating (perturbative decay) is rather simple to estimate and the abundances can directly be obtained from the branching ratios of SM and DM from inflaton, the dynamics of preheating dynamics can not be solved analytically as the equation of evolution becomes non-linear after the initial stage of preheating. We have used publicly available code {\tt LATTICEEASY} to simulate the dynamics of preheating for quadratic inflation and find that DM self-interactions can suppress the production of DM from preheating. However, this suppression remains inadequate to reproduce the observed DM relic abundance, respecting the Bullet cluster bound. We show that, in order to achieve further depletion in the DM number density, a non-relativistic phase of the inflaton before it decays completely to SM and a cannibalization mechanism needs to be invoked. Note that the same DM quartic parameter which suppresses the preheating production, naturally generates cannibalization processes through 4-to-2 annihilations which can deplete the DM relic abundance. In more general models, DM annihilation into other dark sector particles could help depleting the DM abundance, relaxing the length of the non-relativistic phase of the inflaton. \section{Constraints}\label{constraints} In this section, we briefly discuss the relevant constraints, especially focusing on a light DM candidate. $\bullet$ While only active neutrinos lead to $N_\text{eff}=3.046$~\cite{Mangano:2005cc}, the presence of any additional relativistic species enhances $N_\text{eff}$. Primordial nucleosynthesis yields, independently of measurement of the baryon density from CMB observations, $\Delta N_\text{eff}<1$ at 95\%~CL~\cite{Mangano:2011ar}. Using CMB measurements along with the BBN yields data can improve the constraint to $\Delta N_\text{eff}<0.38$ at 95\% CL~\cite{Cyburt:2015mya} (see also Refs.~\cite{Cooke:2017cwo, Tanabashi:2018oca}). This is a constraint we keep in mind when having a DM of the mass range keV-MeV (i.e. relativistic during BBN). To satisfy it, the ratio of dark sector energy density to energy density of relativistic SM particles is required to be $\lesssim 0.051$ at the onset of BBN \cite{Cyburt:2015mya}. $\bullet$ We consider DM particles with mass range of minimum $m_\chi\gtrsim \mathcal{O}(10)$~keV as constrained by Lyman-$\alpha$ observations~\cite{Baur:2015jsy, Irsic:2017ixq}. $\bullet$ The presence of a light scalar field, the DM, during inflation can leave imprints in the primordial power spectrum. In the context of curvaton scenarios~\cite{PhysRevD.42.313, Linde:1996gt,Enqvist:2001zp,Lyth:2001nq,Moroi:2001ct} these possibilities have been considered. The generic signatures include significant non-Gaussianity and isocurvature modes in the primordial power spectrum.% \footnote{Note that, in the context of multi-field inflation scenarios, where more than one fields contribute to the exponential expansion with comparable energy densities, isocurvature perturbation can also arise~\cite{POLARSKI1992623,Langlois:1999dw,Gordon:2000hv,Lalak:2007vi}.} Recent results from Planck~\cite{Akrami:2018odb} find no evidence for isocurvature perturbation, imposing stringent constraints. The success of single field slow-roll inflation paradigm in the light of present CMB data implies that such an additional light field, if present during inflation, can only contribute subdominantly to the energy density of the Universe during inflation. For a light field during inflation, $m^\text{eff}_{\chi} \ll H_{\rm inf}$, where $m^\text{eff}_{\chi}$ and $H_{\rm inf}$ denote the effective mass of the light scalar and Hubble parameter during inflation, respectively. The light field will acquire quantum fluctuations proportional to $H_{\rm inf}$~\cite{Linde:1982uu, Starobinsky:1982ee, Vilenkin:1982wt,Starobinsky:1994bd} with its root mean squared value given by $\frac{\mathcal{O}(0.1)}{\lambda_\chi^{1/4}}\,H_{\rm inf}$~\cite{Starobinsky:1994bd}. The condensate, thus formed, remains frozen until the end of inflation while the Hubble parameter becomes small compared to the (possibly thermal) mass of the light field. At this point the field starts to oscillate and behaves like matter before finally decaying into the $\chi$ particles (in the quartic region~\cite{Kainulainen:2016vzv}). Note that, since we assume very small coupling with the SM fields (i.e. negligible portal coupling with the Higgs boson), $\chi$ particles never thermalize with the SM sector. Similar ideas have been studied in Refs.~\cite{Enqvist:2014zqa,Nurmi:2015ema,Kainulainen:2016vzv}. In Ref.~\cite{Kainulainen:2016vzv} it has been shown that, such a scenario would be quite constrained from non-observation of isocurvature perturbation, in particular, for large field inflation models (with large $H_{\rm inf} \simeq 10^{13}-10^{14}$ GeV). In order to evade this issue, as we will elaborate in the subsequent section, we consider sizable coupling of the $\chi$ field with the inflaton $\phi$. This enhances the effective mass (i.e. the second derivative of the relevant potential) of the $\chi$ field during inflation and ensures that the field does not receive large fluctuations during the same period. $\bullet$ We demand a successful inflation, able to reproduce the observed scalar spectral index $n_s$, the tensor-to-scalar ratio $r$ and the amplitude of the scalar perturbation $A_s$. $\bullet$ We finally demand to reproduce the observed DM relic abundance. \subsection{Dark Matter Cannibalization} In this scenario DM particles never reach thermal equilibrium with the SM, due to the very small coupling between the two sectors, and hence they can not be produced via the standard WIMP scenario. First DM particles could be produced by reheating and preheating mechanisms, but also by the FIMP mechanism. However here we focus on a scenario where the bulk of the original DM particles were produced by the (p)reheating dynamics. In fact, FIMP mechanism typically requires a portal coupling $\lambda_{\chi H}\sim 10^{-10}$ in order to reproduce the observed DM abundance. In our case however, we have $\lambda_{\chi H}\ll 10^{-10}$, so that the FIMP production via SM particle annihilation and Higgs decays is subdominant with respect to the (p)reheating production. Moreover, due to the dynamics within the dark sector, characterized by sizable couplings between DM particles, DM self-interactions could play a crucial role in the generation of the DM relic abundance. In fact, one can consider a framework where the freeze-out proceeds via $N$-to-$n$ number-changing processes, where $N$ DM particles annihilate into $n$ of them (with $N>n\geq 2$). This possibility was first studied in Ref.~\cite{Carlson:1992fn} and recently named the `SIMP paradigm'~\cite{Hochberg:2014dra}. This has been entertained recently in strongly self-interacting DM models that annihilate in number-depleting 3-to-2~\cite{Dolgov:1980uu,Carlson:1992fn,Hochberg:2014dra,Hochberg:2014kqa,Bernal:2015bla,Bernal:2015lbl,Lee:2015gsa,Choi:2015bya,Hansen:2015yaa,Bernal:2015ova,Kuflik:2015isi,Hochberg:2015vrg,Choi:2016hid,Pappadopulo:2016pkp,Farina:2016llk,Choi:2016tkj,Dey:2016qgf,Choi:2017mkk,Ho:2017fte,Dolgov:2017ujf,Garcia-Cely:2017qpx,Choi:2017zww,Chu:2017msm,Duch:2017khv,Heikinheimo:2018esa} or 4-to-2 interactions~\cite{Bernal:2015xba,Bernal:2017mqb,Heikinheimo:2017ofk,Herms:2018ajr,Bernal:2018ins}. In the case where the Higgs portal is very suppressed, the Boltzmann equation which describes the evolution of the DM number density $n(T')$ is: \begin{equation} \frac{dn}{dt}+3\,H(T)\,n=-\langle\sigma v^2\rangle_{3\to 2}\left[n^3-n\,n_\text{eq}^2\right]-\langle\sigma v^3\rangle_{4\to 2}\left[n^4-n^2\,n_\text{eq}^2\right], \end{equation} where $H(T)$ is the Hubble expansion rate as a function of the temperature $T$ of the visible sector and $n_\text{eq}(T')$ represents the equilibrium DM number density at a dark temperature $T'$. The factors $\langle\sigma v^2\rangle_{3\to 2}$ and $\langle\sigma v^3\rangle_{4\to 2}$ correspond to the generalized annihilation cross sections of the 3-to-2 and 4-to-2 DM annihilations, respectively. In the case where the DM stability is guaranteed by a $\mathbb{Z}_2$ symmetry, 3-to-2 processes are forbidden and the DM annihilations are driven by the 4-to-2 processes, with a cross section that in the non-relativistic limit is given by~\cite{Bernal:2015xba} \begin{equation} \langle\sigma v^3\rangle_{4\to 2}\sim\frac{27\sqrt{3}}{8\pi}\frac{\lambda_\chi^4}{m_\chi^8}. \end{equation} Let us recall that the entropies of the dark and the visible sectors are separately conserved, because they were always kinetically decoupled from each other. It is then useful to compute the entropy ratio $\xi$ between them, which is also a conserved quantity. In particular, if the freeze-out happens non-relativistically, one has \begin{equation} \label{entropy:cor} \xi \equiv \frac{s}{s'} \sim \frac{T'_\text{FO}}{3.6~\text{eV}\,(1+ 2.5\,T'_\text{FO}/m_\chi) \times \Omega_\chi h^2}\,, \end{equation} where $s$ and $s'$ are the entropy densities of the SM sector and the dark sector, respectively, and $T'_\text{FO}$ is the temperature of the dark sector when the DM freeze-out takes place~\cite{Carlson:1992fn,Bernal:2015xba}. Substituting $\Omega_\chi h^2 \sim 0.12$~\cite{Aghanim:2018eyx} and the solution of $T'_\text{FO}$~\cite{Bernal:2015xba}, the proper DM abundance requires \begin{equation} \label{entropy:observed} \xi \sim 7\times 10^9 \, \left(\frac{m_\chi}{\text{GeV}}\right) \frac{1}{1 + 0.03 \left(\lambda_\chi\,\frac{\text{GeV}}{m_\chi}\right)^{4/7}}\,. \end{equation} The entropy ratio is nearly proportional to the DM mass, up to a weak dependence on the self-coupling $\lambda_\chi$. Moreover, to match the observed DM relic abundance there is always a $\xi\gg 10^2$ for each DM mass. The model therefore never leads to observable extra radiation~\cite{Carlson:1992fn}, consistent with BBN/CMB bounds. Fig.~\ref{fig:cannibal} shows the entropy ratio $\xi$ (left panel) and the energy density ratio $f$ (right panel) between the two sectors needed in order to reproduce the observed DM abundance, in the ($m_\chi$, $\lambda_\chi$) plane. Let us emphasize that even if the entropy ratio is a constant, the energy density ratio in general is not. For a fixed number of relativistic degrees of freedom, $f$ does not vary while the DM stays relativistic. During that period, the two quantities are related by \begin{equation} f=\frac{g_{\star S}^{4/3}}{g_\star}\,\xi^{-4/3}. \end{equation} In Fig.~\ref{fig:cannibal} the upper left corner (light blue) shows the region where DM elastic scatterings are too strong, and in tension with the Bullet Cluster constraints: $\sigma/m_\chi < 1.25$~cm$^2$/g at 68\% CL~\cite{Markevitch:2003at,Clowe:2003tk,Randall:2007ph}. Additionally, the upper band corresponding to non-perturbative quartic couplings ($\lambda_\chi>4\pi$) is also discarded. \begin{figure}[t!] \begin{center} \includegraphics[width=.49\textwidth]{Entropy.pdf} \includegraphics[width=.49\textwidth]{EnergyDensity.pdf} \caption{Values of $\xi$ and $f$ that yield the observed DM relic density, in the ($m_\chi$, $\lambda_\chi$) plane. The light blue region corresponds to $\sigma/m_\chi > 1.25$~cm$^2$/g and it is excluded by cluster observations. The area where $\lambda_\chi > 4\pi$ is shown in light red. }\label{fig:cannibal} \end{center} \end{figure} \section{Introduction} The existence of non-luminous Dark Matter (DM) has been well established thanks to several evidences in cosmology and astrophysics at different scales~\cite{Bertone:2010zza}. While these signatures only require DM to posses gravitational interactions and to be non-relativistic sufficiently early, the possibility that it can interact weakly with the Standard Model (SM) particles have been widely considered in various well motivated extensions of the SM~\cite{Jungman:1995df,Bertone:2004pz,Bertone:2010zza}. However, very little is known about the particle properties of DM and its interactions with the SM particles. It is generally assumed that DM consists of Weakly Interacting Massive Particles (WIMPs), which, as the name suggests, weakly interact with SM particles~\cite{Arcadi:2017kky}. This assumption opens up the the possibility that DM particles were in thermal equilibrium with the SM in the early Universe; and thus, were produced via the standard freeze-out mechanism. Further, such an assumption provided several avenues to search for DM. However, the various direct~\cite{Akerib:2016vxi, Cui:2017nnn, Aprile:2018dbl} and indirect~\cite{Aguilar:2016kjl,Fermi-LAT:2016uux} DM detectors, as well as the collider experiments~\cite{Aaboud:2016obm, Sirunyan:2017hnk}, have not measured any of such interactions so far. The most constraining limits for WIMP DM come from the direct detection experiments, especially for DM masses in the GeV ballpark. Moreover, while the large scale structure and Lyman-$\alpha$ data~\cite{Baur:2015jsy, Irsic:2017ixq} constraint the mass of the DM particle to be greater than $\mathcal{O}(10)$~keV, the success of Big Bang Nucleosynthesis (BBN) prevents any non-SM relativistic degree of freedom to be in thermal equilibrium with the SM particles at the onset of BBN ($\simeq 1$~MeV)~\cite{Cyburt:2015mya}. This clearly disfavors a DM candidate with mass $\lesssim 1$~MeV which thermalizes with the SM particles \footnote{However, in certain scenarios it is possible to accommodate such a light thermal DM invoking additional interactions~\cite{Berlin:2017ftj,Dutra:2018gmv,Berlin:2018ztp}.} While, in no way the current searches exclude the WIMP scenarios in its entirety, it may be worth exploring different kind of DM candidates which feebly interact with the SM particles. Generically such DM candidates do not thermalize with the SM particles in the early Universe, due to the feeble interaction strength. Several such scenarios have been proposed~\cite{McDonald:2001vt,Choi:2005vq,Kusenko:2006rh,Petraki:2007gq,Hall:2009bx}, both with a simple DM candidate, or even with a more complex Dark Sector (DS). Since in all these scenarios, DM is initially not in thermal equilibrium with the SM particles, the production mechanisms usually involve assumptions on the initial abundance of the DM or DS particles in the very early Universe. For example, in the simplest Feebly Interacting Massive Particle (FIMP) scenario, the initial abundance of the FIMP DM is usually assumed to be zero (see Ref.~\cite{Bernal:2017kxu} for a recent review). Therefore, an estimation of the initial abundance of DM or DS particles can be important in establishing the viability of a particular DM scenario. In the very early Universe, the paradigm of cosmological inflation~\cite{Starobinsky:1980te,Linde:1981mu,PhysRevLett.48.1220} has been well-established, single field slow-roll inflation scenario have been very successful in the light of present data~\cite{Akrami:2018odb}. In this framework, post-inflationary particle production takes place during the (p)reheating epoch~\cite{PhysRevD.42.2491, PhysRevLett.73.3195, PhysRevD.51.5438, PhysRevD.56.3258}. While during reheating particles are produced due to perturbative decay of the inflaton, preheating is a fast and efficient process, when particles are produced due to parametric resonance. This takes place even before the reheating epoch. Thanks to the Bose enhancement, this process is especially efficient in producing bosons which couple to the inflation. In the present article a simple extension of the SM, accommodating an inflaton and a scalar DM (stabilized using an effective $\mathbb{Z}_2$ symmetry), has been considered. Further, DM is assumed to posses negligible interaction with SM particles. Within this framework, we have studied the production of DM after inflation, during the epoch of preheating and reheating. The role of various interaction terms between the inflaton and the DM, as well as the DM self-interactions, have been analyzed in details. This article is organized as follows. In section~\ref{model} the model is presented, making especial emphasis on requirements for driving the inflation. In the following section~\ref{constraints} existing constraints on the relevant model parameters various cosmological and astrophysical considerations have been sketched. Subsequently, in section~\ref{results} we discuss the production of the scalar DM during (p)reheating and estimate the relic abundance of the DM. Finally, in section~\ref{conclusion} we summarize our findings. \section{Dark Matter Production}\label{results} \begin{figure}[t!] \centering \includegraphics[width=0.475\linewidth]{bottomtop.pdf} \caption{Energy fraction $f(T)\equiv\frac{\rho_\chi(T')}{\rho_\text{SM}(T)}$ required to get the observed DM abundance as a function of $1/T$ (in MeV$^{-1}$), for $m_\chi=0.01,~0.05,~0.3,~2,~10,~50$ and $300$~MeV (from top to bottom). The dashed line at the top-right corner corresponds to the observed $f(T_0)\sim 5.3$, as measured by Planck. } \label{bottomtop} \end{figure} The ratio of the energy density of the DM to that of the SM at a temperature $T$, $f(T)\equiv\frac{\rho_\chi(T')}{\rho_\text{SM}(T)}$, measured at the present day CMB temperature $T_0$ is $f(T_0)\sim 5.3$~\cite{Aghanim:2018eyx} (and a corresponding DM temperature $T'$). In the early Universe, however, this ratio is expected to be much smaller since SM radiation cools down faster than the DM component. Assuming the two sectors were always completely disconnected, with corresponding entropies separately conserved, the evolution of $f$ is shows in Fig.~\ref{bottomtop}, for different DM masses: $m_\chi=0.01,~0.05,~0.3,~2,~10,~50$ and $300$~MeV (from top to bottom). The initial values (at high temperatures) for all curves are tuned so that at low temperatures they reproduce the observed DM abundance, as measured by Planck, and showed with a dashed line. When $T'\gtrsim m_\chi$, $f$ tends to be constant showing that both components scale in the same way, as radiation. An analog behavior appears for $T<0.7$~eV, near the matter-radiation equality, when both the SM and the DM behave like matter. The kinks at $T\sim 0.1$~MeV and $T\sim 100$~MeV correspond to the electron-positron annihilation and the QCD phase transition, respectively. In the following we will focus on how (p)reheating could generate the DM to SM energy density ratio in order to reproduce the observed DM abundance, firstly assuming that there are no sizable self-interactions for the DM, and secondly taking them into account. \subsection{Dark Matter Production during (P)reheating} In this subsection we consider the DM production after inflation during the (p)reheating epoch. The process of reheating involves perturbative decay of the inflaton and the production of a particular species during this epoch can be estimated by the branching fractions of the inflaton into it. However, during preheating, which precedes the reheating epoch, bosonic particles coupled to the inflaton field are produced via parametric or tachyonic resonance, and it is a rapid non-perturbative process. The initial stage of preheating can be described in terms of a set of Mathieu equations in the Fourier space and can be studied analytically~\cite{PhysRevLett.73.3195}. The Fourier modes, which are present in the well-known instability bands, grow exponentially at the onset of preheating and are interpreted as exponential particle production% ~\cite{PhysRevLett.73.3195}. However, after the initial stage of preheating, non-linear effects including back-reaction and re-scatterings become relevant~\cite{PhysRevD.56.3258}. Thus, during this phase, accurate analytical estimations are hard to obtain. Therefore, numerical codes have been developed and used to estimate the particle production during this epoch. We have used the publicly available code {\tt LATTICEEASY}~\cite{Felder:2000hq} in order to the study the DM production during preheating. The inflationary production of scalar DM with intermediate masses has been studied by various authors~\cite{Dev:2013yza,Nurmi:2015ema,Kainulainen:2016vzv,Bertolami:2016ywc,Heurtier:2017nwl,Cosme:2017cxk,Cosme:2018nly,Alonso-Alvarez:2018tus}. \begin{figure}[t!] \centering \includegraphics[width=0.475\linewidth]{upto1leasy.pdf} \caption{Evolution of the energy fraction $\rho_\chi/(\rho_\phi+\rho_H)$ as a function of time for $m_{\phi}=10^{-6}M_\text{Pl}$, $\lambda_{\phi}=10^{-14}$, $\lambda_{\phi\chi}=10^{-7}$, $\lambda_{\phi H}=10^{-7}$, $\lambda_{H}=10^{-7}$ and $\sigma_{\phi H}=10^{-10}M_\text{Pl}$. $\lambda_{\chi}=10^i$, for $i=-7$ (uppermost line) to $i=0$ (lowermost line), increasing $i$ in one unit each time. Higher values of $\lambda_\chi$ give rise to more suppression of DM energy density. } \label{ratioplot} \end{figure} In Fig.~\ref{ratioplot}, we plot the evolution during preheating of the ratio of energy density of $\chi$ ($\rho_{\chi}$) to the combined energy density of $\phi$ ($\rho_{\phi}$) and Higgs ($\rho_{H}$) together, i.e. $\dfrac{\rho_{\chi}}{\rho_{\phi}+\rho_{H}}$. We vary $\lambda_{\chi}$ while keeping the other parameters fixed as follows: In order to ensure a successful inflationary phase, $m_\phi=10^{-6}~M_{\rm Pl}$ and $\lambda_\phi=10^{-14}$. Further, $\lambda_{\phi\chi}=\lambda_{\phi H}=10^{-7}$ such that the flatness of the inflationary potential is not ruined, as discussed in section~\ref{model}. A small $\lambda_H=10^{-7}$, as we will elaborate in the subsequent discussion, avoids the suppression of the energy flow into Higgs during preheating. Finally, we have assumed $\sigma_{\phi H}=10^{-10}M_\text{Pl}$ respecting the relation $\lambda_{H}>\frac{\sigma_{\phi H}^2}{2\beta^2m_{\phi}^2}$, again as discussed in section~\ref{model}, while $\sigma_{\chi H} = 0$ in order to avoid thermalization of $\chi$ and Higgs. In the figure, $m_\phi t=0$ denotes the onset of preheating at the end of inflation where we set $\phi_{ini} \simeq 0.2\,M_{\rm Pl}$ at which \texttt{LATTICEEASY} starts to evolve the preheating dynamics. While for different values of $\lambda_{\chi}$ while the ratio changes significantly, as shown in Fig.~\ref{ratioplot}, the generic features of the curves resemble each other, since these are dependent on the dynamics of preheating in general. As emphasized already, production of $\chi$ and $H$ quanta during the initial stage of preheating is governed by the relevant Mathieu equation and the energy transfer from the oscillating inflaton condensate to the $\chi$ and $H$ sector is significant. The energy transfer is most efficient when the inflaton is near the minimum of the potential, i.e. $\phi(t) \simeq 0$, since during this phase the adiabatic conditions are strongly violated (due to large $\dot{\phi}$) and also the contribution of $\phi(t)(\simeq 0)$ to the effective mass of $\chi$ and $H$ becomes small. While the production of $\chi$ is dominated by the parametric resonance due to the absence of the trilinear term $\sigma_{\phi \chi}$, the production of $H$ can receive contributions due to both parametric and tachyonic resonances (for small $k$ modes in particular, in our context)~\cite{Dufaux:2006ee}. The oscillatory features in the energy densities (and in their ratio, as shown in Fig.~\ref{ratioplot}) during the initial stage of preheating can be attributed to the characteristic solution of Mathieu equation. However, as the $\chi$ and $H$ fields are populated, as discussed in the previous paragraph, the back-reaction and re-scattering become significant, gradually suppressing the energy transfer to the $\chi$ and $H$ fields, and the ratio $\dfrac{\rho_{\chi}}{\rho_{\phi}+\rho_{H}}$ eventually stabilizes \footnote{In particular, once in the initial stage the resonant modes $k_*$ of $\chi$ ($H$) get excited, annihilation of those quanta generate inflaton quanta, re-scattering of $\chi$ ($H$) quanta against inflaton zero mode produces $\chi$ ($H$) and inflaton quanta with the momentum $k_*/2$. In the next stage, violent phase of non-linear dynamics, the $k_*/2$ modes grow in amplitude and shift towards $k_*$. In the third stage, the distribution smooths out and spreads towards higher momentas. For an elaborate discussion, see Ref.~\cite{Podolsky:2005bw}.} For example, as shown in the figure, with $\lambda_\chi = 10^{-7}$ (the highest line) the increment continues until $m_\phi t\sim80$, at which point $\rho_{\chi}$ becomes comparable to that of the inflaton and does not grow any further. Note that, with $\lambda_{\phi H} \simeq \lambda_{\phi \chi}$ (both set to $\mathcal{O}(10^{-7})$), as the self-coupling of $\chi$ ($\lambda_{\chi}$) is enhanced compared to that of $H$, the energy flow in the $\chi$ sector decreases substantially. In particular, as shown in the figure, for higher quartic couplings: $\lambda_{\chi}=10^0$ (lowermost line), $10^{-1}$, ... $10^{-6}$ and $10^{-7}$ (uppermost line) the ratio $\dfrac{\rho_{\chi}}{\rho_{\phi}+\rho_{H}}$ varies from $\mathcal{O}(10^{-6})$ to $\mathcal{O} (10^{-1})$, respectively. A similar suppression due to quartic self-interaction has also been observed in Refs.~\cite{Prokopec:1996rr, Hyde:2015gwa}. This feature can be attributed to the fact that the quartic coupling, after initial exponential production of DM, contributes to the effective mass of $\chi$ as $\sqrt{\frac{\lambda_{\chi}}{2}\langle \chi^2\rangle}$. This large contribution makes subsequent production of $\chi$ energetically expensive, effectively blocking the same~\cite{Dufaux:2006ee}. It is worth mentioning here that the production of a boson during preheating can be suppressed by reducing its interaction strength with the inflaton $\lambda_{\phi\chi}\lesssim 4 \times 10^{-8}$, which leads to an inefficient and short period of preheating, due to the smallness of the relevant parameter $\lambda_{\phi\chi}\,\phi_{ini}^2/m_{\phi}^2 \lesssim 1$~\cite{Hardy:2017wkr}.% \footnote{However, this does not seem to lead to adequate suppression for different DM masses we have considered. Also, if the interaction is significantly reduced, the effective mass of the boson may fall below $H_{\rm inf}$. We will not consider this possibility.} Finally, although not explicitly shown in Fig.~\ref{ratioplot} at the end of preheating, the remaining energy density in the inflaton condensate continues to oscillate before it eventually decays, in our context, mostly producing the SM Higgs, thanks to the trilinear $\sigma_{\phi H}$ term \footnote{Note that during the oscillation of the inflaton, an effective trilinear coupling proportional to $\lambda_{\phi \chi} \phi(t)$ can be generated. Thus, setting $\sigma_{\phi \chi}=0$ does not completely prohibit the perturbative decay of the inflaton into $\chi$. However with time, as the amplitude of inflaton oscillation $\phi(t)$ decreases below $\mathcal{O}(10^{-3}M_{\rm Pl})$, $\sigma_{\phi H}\simeq \mathcal{O}(10^{-10} M_{\rm Pl})$ becomes dominant. For simplicity we have ignored this contribution in our estimation.} While we are interested in the initial value of $f = \rho_\chi/\rho_\text{SM}$ at the end of (p)reheating, the decay mentioned above ensures that $\dfrac{\rho_{\chi}} {\rho_{\phi}+\rho_{H}}$, as plotted in the figure above, gives a good estimate of $f$ at the end of (p)reheating assuming that the inflaton does not go through a non-relativistic phase before it decays. However, as shown in Fig.~\ref{bottomtop}, only a small initial DM energy density (compared to that of the SM) at high temperature suffices to produce the right DM relic abundance. Even with a rather large $\lambda_{\chi}$, and assuming that the perturbative decay of the inflaton $\phi$ only contributes to $\rho_\text{SM}$, we find it difficult to achieve such a small ratio, and produce substantially large initial abundance of $\chi$ at the end of preheating, especially for $m_{\chi} \gtrsim 0.02$ MeV. However, as will be discussed subsequently in this section, the mass range $m_{\chi} \lesssim 0.02$~MeV, giving rise to the right initial abundance with large $\lambda_{\chi}$, as discussed above, have been disfavored due to rather large self-interaction from the Bullet cluster bounds. This makes it apparent that the scenario, as described so far, requires mechanisms to suppress the initial abundance of $\chi$ in order to be viable. It is well-known that the energy density associated to the oscillations of the inflaton, in the context of quadratic potential, behaves as matter. However, it has been shown that, although substantial energy density remains in the inflaton (condensate and the excitation combined) at the end of preheating, in presence of the trilinear $\sigma$ terms the equation of state right after preheating is rather close to that of the radiation $w \simeq \frac{1}{4}$ \cite{Dufaux:2006ee}. This is in contrast to the case of quartic inflation, where the equation of state, under similar circumstances, resembles that of radiation \footnote{In Ref.~\cite{Dufaux:2006ee}, the equation of state during the preheating period has been studied in detail. It has been shown that for a preheating scenario (with quadratic inflaton potential) where there is no trilinear coupling of inflaton to other species, the equation of state at first increases to $1/4$, then it decreases slowly towards $0$. On the other hand, in a scenario with trilinear term, the equation of state increases to a value above $1/4$ and does not decrease thereafter. This phenomenon has been explained by studying the fraction of number density stored in the relativistic modes of the fields. It has been observed that, for the first scenario, it decreases after initial increment for the inflaton field. Whereas for the second scenario, this fraction remains stable for the entire simulation time scale. Therefore, for simplicity, we can assume the inflaton to be a relativistic species after preheating for our case with trilinear coupling.} However, if inflaton becomes non-relativistic before it decays, its energy density after the relativistic to non-relativistic transition, scales as $a^{-3}$, while for the relativistic species (SM and DM particles) the energy density scales as $a^{-4}$, where $a$ is the scale factor. This can lead to an enhancement in the relative energy fraction stored in the inflaton field at the end of preheating, and the subsequent decay process into SM can suppress the DM energy density relative to SM. The dynamics of the Universe during the phase between the end of preheating (defined as the initial exponential particle production) and the decay of inflaton (governed by the decay width $\Gamma_\phi$ of inflaton, which is generally much larger than the preheating time scale) is complicated and contains processes like turbulence and thermalization~% \cite{Aarts:2000mg, Allahverdi:2000ss, Felder:2000hr, Salle:2000hd, Allahverdi:2002pu, Micha:2003ws, Micha:2004bv, Desroche:2005yt, Mukaida:2015ria}. The enhancement of the relative energy density of inflaton from the end of preheating to its decay depends on details of this stage. Although the parameter $a_d/a_t$, written as a ratio of scale factors during the decay of inflaton $a_d$ and during the transition of inflaton energy density from relativistic to non-relativistic $a_t$ (where we assume that both the inflaton decay and its transition from relativistic to non-relativistic phase are instantaneous), enhances the inflaton energy density relative to the other (relativistic) energy densities, the full details of the phase between end of preheating and decay of inflaton may add additional information to get the correct enhancement factor. We use $a_d/a_t$ as an effective enhancement factor of inflaton energy density with respect to other energy densities. The factor $a_d/a_t$ is by construction $\geq1$, as even if the inflaton decays when it is relativistic, i.e. $a_d< a_t$, the inflaton energy density does not get enhanced. In Fig.~\ref{ratioplot2}, we plot $f\equiv\frac{\rho_\chi}{\rho_\text{SM}}=\frac{\rho_\chi^{ini}}{\rho_\phi^{ini}\times\frac{a_d}{a_t}+\rho_H^{ini}}$ (where $\rho_i^{ini}$ is the initial energy density of species $i$ just after preheating) as a function of $\lambda_{\chi}$ at the end of preheating for different values of $a_d/a_t$. For the top-most line of this plot, it is clear from Fig.~\ref{ratioplot} why the ratio does not change with $\lambda_{\chi}$, when $\lambda_{\chi}$ is not significantly larger than $\lambda_{\phi\chi}$, and after that the ratio falls off with increasing $\lambda_{\chi}$. The other lines in this plot correspond to the cases where the inflaton energy density gets a boost by a factor $a_d/a_t>1$. \begin{figure}[t!] \centering \includegraphics[width=0.475\linewidth,height=0.340\linewidth]{fvslambda6plots.pdf} \caption{Energy fraction $f\equiv\rho_\chi/\rho_\text{SM}$ at the end of reheating as a function of $\lambda_{\chi}$, for the same benchmark point of Fig.~\ref{ratioplot}. The lines correspond to $a_d/a_t=10^0$ (uppermost) to $10^6$ (lowermost), increasing by an order of magnitude each time. } \label{ratioplot2} \end{figure} \begin{figure}[t!] \centering \includegraphics[width=0.475\linewidth]{masl6.pdf} \caption{$\lambda_{\chi}$ vs $m_{\chi}$ (in MeV) plot for allowed relic abundance (without any depletion mechanism in DM sector). The lines from top-left corresponds to $a_d/a_t=10^0$ to $10^6$. The shaded region is excluded from Bullet cluster bounds. } \label{masl} \end{figure} Fig.~\ref{masl} combines the information of Figs.~\ref{bottomtop} and~\ref{ratioplot2}. It depicts the values of the quartic coupling $\lambda_\chi$ required in order to reproduce the observed DM relic abundance as a function of the DM mass, for different ratios $a_d/a_t$: from $10^0$ (uppermost) until $10^6$ (lowermost), increasing by an order of magnitude each time. The shaded gray region is in tension with the Bullet cluster constraint on DM self-interactions. The figure shows that the case where $a_d/a_t=1$ is excluded by observations. However, having $a_d/a_t\gg 1$ gives additional suppression of the DM to SM ratio if the inflaton fully decays into SM only. An additional suppression in the ratio $f$ by this factor can reproduce the right DM abundance of a specific $m_\chi$ with a smaller value of $\lambda_{\chi}$. Thus, as shown in Fig.~\ref{masl}, an appropriate $a_d/a_t$ may lead to the right DM abundance, evading Bullet cluster bounds.\footnote{Quartic inflation does not give a better result in terms of getting right DM relic abundance. This is because, the energy density of a quartic potential evolves as radiation while oscillating around its minima. For the same trilinear coupling strength, the inflaton may decay while still relativistic (because inflation may be rather light in this case), thereby possibly decreasing the duration of a non-relativistic phase.} Let us emphasize that in this figure we are assuming that there are no interactions within the dark sector able to modify the DM number density. However, when $\lambda_\chi$ becomes sizable, that may not be a suitable assumption. In fact, in this case DM cannibalization will inevitably deplete the DM sector. This is especially efficient in the presence of large self-interaction parameter $\lambda_{\chi}$ which, incidentally, also helps to suppress the DM production during preheating. We will discuss this possibility in detail in the next subsection. \section{Conclusion} \section*{Acknowledgments} The authors would like to thank Juan Pablo Beltrán Almeida and Supratik Pal for fruitful discussions. We acknowledge partial support from the European Union Horizon 2020 research and innovation programme under the Marie Sk{\l}odowska-Curie: RISE InvisiblesPlus (grant agreement No 690575) and the ITN Elusives (grant agreement No 674896). NB is also supported by the Universidad Antonio Nariño grants 2017239 and 2018204, and by the Spanish MINECO under Grant FPA2017-84543-P. AC acknowledges support from Department of Science and Technology, India, through INSPIRE faculty fellowship, (grant no: IFA 15 PH-130, DST/INSPIRE/04/2015/000110). AP thanks the computational facilities of Indian Statistical Institute, Kolkata. AP is supported by Council of Scientific and Industrial Research (CSIR), India, File no. 09/093(0169)/2015 EMR-I. In addition to the software packages cited above, this research made use of IPython~\cite{Perez:2007emg}, Matplotlib~\cite{Hunter:2007ouj} and SciPy~\cite{SciPy}. \section{The Model}\label{model} The production of a scalar DM and SM particles during (p)reheating epoch generally depends on the inflaton potential at the end of inflation, and also on the interaction terms between the inflaton and the respective sectors. In this section, we sketch the model of our interest, elaborating on the inflationary scenario and the relevant interaction terms. On top of the SM field content, our model contains two real scalar fields, the inflaton $\phi$ and the DM $\chi$.\footnote{Note that a stable inflaton could also play the role of DM~\cite{Liddle:2006qz,Cardenas:2007xh,Panotopoulos:2007ri,Liddle:2008bm,Bose:2009kc,Lerner:2009xg,DeSantiago:2011qb,Khoze:2013uia,Mukaida:2014kpa,Fairbairn:2014zta,Bastero-Gil:2015lga,Kahlhoefer:2015jma,Tenkanen:2016twd,Daido:2017wwb,Choubey:2017hsq,Daido:2017tbr,Hooper:2018buz,Borah:2018rca,Manso:2018cba,Rosa:2018iff,Almeida:2018oid}.} In order to ensure the stability of the DM, a $\mathbb{Z}_2$ symmetry is imposed under which only the DM is odd. The scalar potential is, then, given by \begin{eqnarray}\label{potential0} V &=& \frac{m_{\phi}^2}{2}\phi^2+\frac{m_\chi^2}{2}\chi^2-\frac{\mu_H^2}{2}|H|^2+\frac{\sigma_3}{3} \phi^3+\frac{\lambda_{ \phi}}{4}\phi^4 + \frac{\lambda_{\chi}}{4}\chi^4 + \frac{\lambda_H}{4} |H|^4\nonumber\\ &&+ \frac{\sigma_{\phi \chi}}{2} \phi \chi ^2 + \frac{\sigma_{\phi H}}{2} \phi |H|^2 + \frac{\lambda_{\phi \chi}}{2} \phi^2 \chi ^2+ \frac{\lambda_{\phi H}}{2} \phi^2 |H|^2 + \frac{\lambda_{\chi H}}{2} \chi^2 |H|^2+V_\text{gravity}\,, \end{eqnarray} where $H$ is the SM Higgs field. Note that $\mu_H$ and $\lambda_H$, at the electroweak scale, are constrained in order to reproduce the observed Higgs mass and vacuum expectation value ({\it vev}). In the present context, we will consider large field inflationary models. Two such scenarios have been widely studied, one with the quartic potential~\cite{Linde:1983gd} and another with quadratic potential~\cite{Mukhanov:1981xt} for the inflaton field $\phi$, see also Refs.~\cite{Linde:2007fr,Martin:2013tda} for reviews. In their simplest incarnations, both of these scenarios have been in tension with the present constraints from the CMB data~\cite{Akrami:2018odb}, the quartic one being worse. However, it has been pointed out that introduction of a non-minimal coupling with gravity can render both these scenarios viable~\cite{Fakir:1990eg,Linde:2011nh,Kallosh:2013tua}. While we will consider quadratic inflation in the presence of a non-minimal gravitational coupling, we will also comment on the consequences of considering a quartic potential, especially in the context of DM production during (p)reheating. The non-minimal couplings to gravity are expressed as \begin{equation} V_\text{gravity}=\frac12\left(\xi_\phi\phi^2+\xi_\chi\chi^2+2\xi_H|H|^2\right)\mathcal{R}, \end{equation} where $\mathcal{R}$ is the Ricci scalar~\cite{Birrell:1982ix}. For simplicity, we will set $\xi_\chi = 0 = \xi_{H}$. Before discussing more on the inflationary aspects, which include terms depending only on the inflaton $\phi$, a discussion on various other relevant parameters are in order. As we will see in section~\ref{results}, they play an important role during the (p)reheating. The trilinear terms $\frac{\sigma_{\phi \chi}}{2} \phi \chi ^2 $, $\frac{\sigma_{\phi H}}{2} \phi |H|^2$ and $\frac{\sigma_3}{3} \phi^3$ can potentially give rise to a non-zero {\it vev} for the inflaton field at the end of inflation.% \footnote{These terms are allowed by symmetries and required for the draining of the excess energy stored in inflaton after preheating, as 2-to-2 scatterings of inflatons into DM or SM through terms $\frac{\lambda_{\phi \chi}}{2} \phi^2 \chi ^2$ and $ \frac{\lambda_{\phi H}}{2} \phi^2 |H|^2$ can not transfer the energy stored in inflaton completely.} This will generate an effective mass term, presumably large, for both the scalars coupled to the inflaton. In order to keep the (p)reheating dynamics simple, and the DM mass as a free parameter of the theory (independent of $\lambda_{\phi\chi}$), we discard this possibility. Further, we also ensure that the minima of the scalar potential is at zero, and is achieved when all the scalar fields assume zero {\it vev}.% \footnote{Since we are interested in the the epoch after inflation, where all relevant scales are much larger compared to the electroweak scale, we ignore the zero-temperature electroweak {\it vev} of the Higgs field.} These lead to the conditions $\lambda_{\chi}>\frac{\sigma_{\phi \chi}^2} {2\alpha^2m_{\phi}^2}$, $\lambda_{H}>\frac{\sigma_{\phi H}^2}{2\beta^2m_{\phi}^2}$ and $\lambda_{\phi}>\frac{2\sigma_3^2}{9\gamma^2 m_{\phi}^2}$ given arbitrary real constants $\alpha$, $\beta$ and $\gamma$, such that $\alpha^2+\beta^2+\gamma^2=1$. Also the quartic couplings $\lambda_{\phi \chi}$, $\lambda_{ \phi H}$ and $\lambda_{\chi H}$ need to be positive. Note that at high scales ($\sim 10^{-5}M_\text{Pl}$) the quartic coupling $\lambda_H$ becomes negative, due to the large quantum corrections mostly from the top quark Yukawa coupling \footnote{Since in our context both $\lambda_{\phi \chi}$ and $\lambda_{\phi H}$ remain small, as we will discuss latter, the quantum contribution from the scalar fields $\phi$ and $\chi$ do not improve the situation.} This has been well studied in the literature, see e.g. Refs.~\cite{Sher:1988mj,EliasMiro:2011aa,Degrassi:2012ry,Buttazzo:2013uya}. Its consequence during inflation and possible remedies are also known, see e.g. Refs.~\cite{Espinosa:2007qp,Enqvist:2014bua,Shkerin:2015exa,Kearney:2015vba,Espinosa:2015qea} (also Ref.~\cite{Espinosa:2007qp,Herranen:2014cua,Herranen:2015ima,Ema:2017loe} for $\xi_H \neq 0$). Since our main focus is on the post inflationary production of DM particle during (p)reheating, we will assume that the quartic coupling $\lambda_H$ remains positive during (p)reheating epoch. This can be simply achieved by introducing another scalar field, which couples to the Higgs boson~\cite{Lebedev:2012zw,EliasMiro:2012ay}. We will assume that such a scalar, if exists, thermalizes with the SM particles, and if coupled to the inflaton, its production and dynamics during (p)reheating would be similar as that of the Higgs boson itself; but we will not explicitly include it in our discussion. For the rest of this article, we will simply assume that during (p)reheating $\lambda_H$ remains positive. For the inflationary paradigm we consider the quadratic potential for the inflaton $\phi$. In this case, in Eq.~\eqref{potential0} the quadratic term for the inflaton field $\phi$ dominates over the quartic term, which is assumed to be small. Introducing a non-minimal coupling $\xi_{\phi} \simeq \mathcal{O}(10^{-3})$ with $m_\phi \simeq \mathcal{O}(10^{-6})~M_\text{Pl}$~\cite{Linde:2011nh} the quadratic potential can produce the scalar spectral index $n_s$, the tensor-to-scalar ratio $r$ and the amplitude of the scalar perturbation $A_s$, consistently with the current Planck data~\cite{Akrami:2018odb}. In particular, a benchmark with $\xi_\phi=2\times 10^{-3}$, $m_{\phi}=5.3\times10^{-6}~M_\text{Pl}$ (with $\phi_{\rm pivot} \simeq 14\,M_\text{Pl}$ and $\phi_{\rm end} \simeq M_\text{Pl}$), leads to the following values of the observable parameters: the scalar spectral index $n_s \simeq 0.963$, the amplitude of the scalar perturbation $A_s \simeq 2.22\times 10^{-9} $ and the tensor-to-scalar ratio $r \simeq 0.043$, while the inflation lasts for $\sim 62~e$-foldings. It has also been shown in literatures \cite{Markkanen:2015xuw,Fairbairn:2018bsw} that a non-minimal coupling of $\chi$ to gravity can produce DM gravitationally, even if the DM has no other coupling with any other sector. Note that, the success of inflation critically depends on the flatness of the inflaton potential, such that the slow-roll parameters remains small during the entire inflation. In case of the quadratic inflation, the contribution from the quartic term, therefore, must remain insignificant. Note that, simply forbidding $\lambda_{\phi}$ at the tree-level, would not suffice, since it can be generated due to quantum corrections. The relevant renormalization group equations (RGE) are given by~\cite{DeSimone:2008ei,Lerner:2011ge,Ema:2017ckf}, \begin{eqnarray} 16\pi^2 \frac{d \lambda_{\phi}} {d\ln{\mu}} &=& 8 \lambda_{\phi H}^2 + 2 \lambda_{\phi \chi}^2+ 18 c_\phi^2\lambda_{\phi} ^2\,,\label{eq:rge1}\\ 16\pi^2 \frac{d \lambda_{\phi H}}{d\ln{\mu}} &=&\lambda_{ \phi H}\left[8c_\phi\lambda_{ \phi H} + 12 \lambda_H+6c_\phi^2\lambda_\phi-\frac32(3g_L^2+g_Y^2)+6y_t^2\right],\label{eq:rge2}\\ 16\pi^2 \frac{d \xi_\phi}{d\ln{\mu}} &=&6\left(\xi_\phi-\frac16\right)c_\phi\lambda_\phi,\label{eq:rge3} \end{eqnarray} where \begin{equation} c_\phi=\frac{1+\xi_\phi \phi_i^2/M_\text{Pl}^2}{1+(6\xi_\phi-1)\,\xi_\phi \phi_i^2/M_\text{Pl}^2}, \end{equation} with $g_L$ and $g_Y$ corresponding to the $SU(2)$ and $U(1)$ gauge couplings and $y_t$ being the Yukawa coupling of the top quark. As shown in the equations above, the radiative contribution to $\lambda_{\phi}$ involves both $\lambda_{\phi \chi}$ and $\lambda_{\phi H}$. Thus, for requiring $\lambda_{\phi} \lesssim 10^{-14}$ (to ensure the smallness of this term during inflation) without invoking any tuning, we assume that $\lambda_{\phi \chi}$, $\lambda_{\phi H} \lesssim \mathcal{O}(10^{-7})$ during inflation and right after the inflation at the onset of (p)reheating epoch. Similar argument prevents $\sigma_3$ from being large. Further, $\sigma_{\phi \chi}$ and $\sigma_{\phi H}$ can also contribute to the RGE of $m_{\phi}$ and $\sigma_3$, and we ensure that these parameters remain small during (p)reheating after inflation in our discussion in section~\ref{results}. Finally, we comment on the possibility of considering the quartic inflation where the quadratic term $\frac{m_{\phi}^2}{2} \phi^2$ is negligible in Eq.~\eqref{potential0} compared to $\frac{\lambda_{\phi}}{4} \phi^4$ during inflation~\cite{Fakir:1990eg,Makino:1991sg,Komatsu:1997hv,Komatsu:1999mt,Tenkanen:2017jih}. However, to ensure that $\phi$ remains massive after inflation (without the necessity of a phase transition), and to ensure that the minima of the scalar potential remains at zero, when all the scalar fields are set to zero, a mass term for $m_{\phi}\ll 10^{-6} M_{\rm Pl}$ would be desired. It has been shown that, in this case, the current CMB data can be matched with $N\sim 60~e$-folds and $\xi_\phi\sim 790N\sqrt{\lambda_\phi}$. Note that the nature of the inflaton potential, whether quartic or quadratic, affects the post inflation dynamics of the inflaton, and hence the preheating process. Furthermore, a heavy inflaton can possibly behave as matter after the rapid preheating process, before it decays injecting significant entropy, and diluting the abundance of any light particle produced during preheating. In section~\ref{results}, a quadratic potential will be assumed for $\phi$. However, we will also comment on the impact of the choice of quartic inflation in the same context.

47224efdb0acdeaf24522a260d462eae0ca30282

\section{Integral modeling} \subsection{Integrated modeling of the electric double layer}\label{ss:intgrated_dl} In the previous sections we discussed the interaction of a plasma with a solid surface in terms of elementary surface collision processes. Similar to collisions in the bulk of the discharge they are characterized by collision probabilities (viz: sticking, reflection, and secondary electron emission coefficients) which have to be determined either by independent measurements or by separate quantum-mechanical calculations. In contrast to bulk processes, however, where the surrounding plasma does not affect the collision partners, in a surface collision the plasma affects the target. The parameters used to characterize a surface collision depend thus on the plasma. Taking the plasma-induced modifications of the surface into account requires a selfconsistent modeling of the mutual influence between the plasma and the solid. Such an integral modeling approach is of course very demanding involving, in general, processes and species acting on different scales. In this section we will thus exemplify the approach only for an idealistic situation, where the only physical consequence of the plasma-surface interaction is an electric response leading to the build-up of an electric double layer at the plasma-solid interface, as discussed in Sec.~\ref{ss:psi-processes}, cf. Fig.~\ref{fig:interface-physics}. First steps towards a selfconsistent description of this double layer, for a floating dielectric surface, have been made by Bronold and Fehske~[\onlinecite{BF17}]. In this subsection we summarize and comment on their work. The formation of an electron-depleted region on the plasma side of the interface and an electron-rich region inside or on top of the solid, depending on its electronic structure, is the most fundamental manifestation of the interaction of a solid surface with an ionized gas. It arises because electrons in the plasma outrun heavy species leading to a more efficient electron deposition due to electron absorption than electron extraction by neutralization/de-excitation of ions/radicals. That the electric response of the plasma-solid interface leads to an electric double layer, having a negative part inside the solid and a positive part inside the plasma, is known since the beginnings of gaseous electronics~[\onlinecite{LM24}]. Ever since, however, the focus of interest has been on the plasma-based electron-depleted part of the double layer--the plasma sheath--how it merges with the quasi-neutral bulk plasma~[\onlinecite{Robertson13,Brinkmann09,Franklin03,Riemann91,SB90}] and how it is affected by the emissive properties of the surface~[\onlinecite{CU16,LW15,SHK13,SKR09,TLC04,HZ66}]. The negative part of the double layer found essentially no attention. Yet it is an integral part of the electric response of the solid to the plasma. Usually, the theoretical descriptions of the electric response of the plasma-solid interface assume that the processes inside the solid occur on time scales too fast and length scales too small to affect the plasma~[\onlinecite{Franklin76}]. Based on such a view, it is thus sufficient to replace the plasma-facing solid by an object with a geometrical boundary and probabilities for electron sticking/reflection, ion neutralization, and secondary electron emission~[\onlinecite{LL05}]. Within such an approach~[\onlinecite{BGL10,Kushner04,GMB02}] it is of course impossible to investigate the plasma-induced modifications of the electronic structure of the solid, which in turn however may strongly affect the probabilities for charge transfer. We consider this to be a particularly severe drawback for the modeling of microdischarges on semiconducting substrates~[\onlinecite{DOL10,KSO12}]. Due to the continuing miniaturization of these structures~[\onlinecite{EPC13}] the electron transit times through the plasma sheath may become comparable to the electron relaxation time inside the solid. Between subsequent electron encounters the electrons inside the solid can thus no longer equilibrate. It is hence no longer viable to describe charge transfer across the plasma-solid interface by a number of fixed surface parameters in such situations. Instead charge-transfer has to be considered as the linking part of a selfconsistent modeling of the charge dynamics on both sides of the interface. As illustrated in Fig.~\ref{IntModelProcesses}, the selfconsistent integral modeling of the electric response of the plasma-solid interface tracks electrons ($e^-$) and ions ($i^+$) generated by impact ionization inside the plasma to the inside of the solid, where they recombine--after energy- and momentum-relaxation--either radiatively by band-to-band transitions or non-radiatively via trap states as conduction band electrons ($e_*^-$) and valence band holes($h^+$). The plasma source is thus linked to the plasma sink inside the solid. We expect an understanding of this link to be central for the progress of future efforts combining gaseous and semiconductor electronics~[\onlinecite{TWH11,OE05}]. \begin{figure}[t] \includegraphics[width=\linewidth]{ProjectCartoon} \caption{(Color online) Illustration of the electric response of a floating dielectric plasma-solid interface. On the right is shown the traditional modeling treating the solid as a black box, characterized by surface parameters such as the electron sticking coefficient and the ion wall recombination probability. The left depicts the processes actually taking place inside the solid when plasma is destroyed. Electrons ($e^-$) and ions ($i^+$), generated by impact ionization in the plasma, hit the dielectric solid, thereby injecting conduction band electrons ($e_*^-$) and valence band holes ($h^+$), which after relaxation may either recombine non-radiatively or radiatively. As a result, electron depletion occurs in front of the surface leading to a positive space charge which in turn is balanced by a negative space charge inside the solid (electric double layer). The selfconsistent kinetic modeling~[\onlinecite{BF17}] treats impact ionization in the plasma and recombination in the solid on an equal footing by tracing the ambipolar charge transport across the interface, allowing for quantum-mechanical reflection/transmission and charge transport/relaxation on both sides of the interface.} \label{IntModelProcesses} \end{figure} To get an idea about how to organize the self-consistent integral modeling of the electric response of the plasma-solid interface, let us consider a dielectric solid facing a plasma~[\onlinecite{BF17}]. The modeling then is based on two sets of spatially separated Boltzmann equations, one for the electrons and ions inside the plasma and one for the conduction band electrons and valence band holes inside the solid (using an electron-hole picture inside the solid simplifies the calculations). Defining a species index $s=e,i,*,h$ to denote electrons, ions, conduction band electrons, and valence band holes, the Boltzmann equations for the quasi-stationary distribution functions $F_s^\gtrless(z,E,\vec{K})$ for left and right moving particles can be generally written as (measuring length in Bohr radii and energy in Rydbergs)~[\onlinecite{BF17}] \begin{widetext} \begin{equation} \left[\pm v_s(z,E,\vec{K})\frac{\partial}{\partial z}+\gamma^\gtrless_s[F_{s^\prime}^>,F_{s^\prime}^<]\right] F_s^{\gtrless}(z,E,\vec{K})=\Phi^\gtrless_s[F_s^>,F_s^<,F_{s^\prime}^>,F_{s^\prime}^<] \label{BTE} \end{equation} \end{widetext} with \begin{equation} v_s(z,E,\vec{K}) = 2\bigg(\frac{m_e}{m_s}[E-U_s(z)] - (\frac{m_e}{m_s}\vec{K})^2 \bigg)^{1/2}~ \end{equation} the velocity of the particles normal to the (planar) interface at $z=0$. Here, $z$, $E$, and $\vec{K}$ denote the distance from the interface, the total energy, and the lateral momentum of the particles. The functions $U_s(z)$ are shown in Fig.~\ref{IntModelNotation} (together with other quantities relevant for the description of the electric double layer). They define the regions in $z\!\!-\!\!E$ space where the respective species move freely~[\onlinecite{BF17}]. The alignment of the electronic energies of the solid with the ones of the plasma is controlled by the electron affinity $\chi$ and the energy gap $E_g$. The functions $\gamma^\gtrless_s$ and $\Phi^\gtrless_s$ denoting, respectively, the rates for out-scattering and the in-scattering collision integrals depend on the scattering process. \begin{figure}[b] \includegraphics[width=\linewidth]{IntModelingPro} \caption{ Notation used for the description of an electric double layer at a floating dielectric plasma-solid interface with negative space charge inside the solid and positive space charge in front of it. Shown are the edges of the conduction ($U_*$) and valence ($U_{vb}$) bands, the edge for the motion of valence band holes ($U_{h}$), the potential energies for electrons ($U_e$) and ions ($U_i$) on the plasma side, and the energetic range, specified by the ion's ionization energy $I$ and its broadening $\Gamma$, in which hole injection occurs due to the neutralization of ions at the interface. Source, reservoir, and quasi-neutral regions are indicated as they will arise in the course of the calculation (adopted from Ref.~\onlinecite{BF17}).} \label{IntModelNotation} \end{figure} The Boltzmann equations themselves are not sufficient. They have to be augmented by matching and boundary conditions for the distribution functions, the former applying to $z=0$ and the latter to the asymptotic regions of the interface. For a dielectric plasma-solid interface without interface/surface states the matching conditions for the electron distribution functions become ($E>0$) \begin{align} F_{e,*}^{>,<}(0,E,\vec{K}) &= R(E,\vec{K}) F_{e,*}^{<,>}(0,E,\vec{K})\nonumber\\ &+ [1-R(E,\vec{K})]F_{*,e}^{>,<}(0,E,\vec{K})~, \label{Fe_gtrless_match} \end{align} while the hole and ion distribution functions are connected by ($E>E_g+\chi$) \begin{align} F_h^<(0,E,\vec{K}) &= F_h^>(0,E,\vec{K}) + \alpha S_h^<(E,\vec{K})~~, \nonumber\\ F_i^>(0,E,\vec{K}) &= (1-\alpha)F_i^<(0,E,\vec{K})~. \label{Fi_gtr_match} \end{align} The function $R(E,\vec{K})$ is the quantum-mechanical reflection coefficient for electrons due to the surface potential which can be, for instance, modelled by a three-dimensional potential step with a mismatch in the electron mass arising from the difference between the effective electron mass $m_e^*$ inside the solid and the electron mass $m_e$ in the plasma, $\alpha$ is the neutralization probability for an ion at the interface, and $S_h^<(E,\vec{K})$ is a function specifying the accompanying hole injection into the valence band of the dielectric. An illustration of the matching conditions, coupling the distribution functions of the conduction band electrons ($e_*^-$) and valence band holes ($h^+$) in the solid with the distribution functions of the electrons ($e^-$) and ions ($i^+$) in the plasma, is included in Fig.~\ref{IntModelProcesses}. The particular form of $S_h^<(E,\vec{K})$, into which $F_i^<$ enters~[\onlinecite{BF17}], depends on the neutralization process. In case the neutralization induces also secondary electron emission, the matching condition for the electron distribution functions has to be augmented by a function $S_e^>(E,\vec{K})$. With the Poisson equation for the electric potential energy $U_c(z)$ (again given in atomic units), \begin{align} \frac{d}{dz}\varepsilon(z)\frac{d}{dz} U_c(z) = 8\pi\big[\rho_w(z)\theta(-z)-\rho_p(z)\theta(z)\big]~, \label{PE} \end{align} where $\rho_w(z)=n_*(z)-n_h(z)-n_D+n_A>0$ and $\rho_p(z)=n_i(z)-n_e(z)>0$ are, respectively, the charge densities inside the surface (assumed to contain donors and acceptors with concentrations $n_D$ and $n_A$) and the plasma, to be obtained from the distribution functions by integration, \begin{eqnarray} n_s(z)=\int\frac{dE d\,^2K}{(2\pi)^3} \frac{F^>_s(z,E,\vec{K})+F^<_s(z,E,\vec{K})}{v_s(z,E,\vec{K})}~, \label{Density} \end{eqnarray} and the matching conditions for $U_c(z)$ at $z=0$, Eqs.~\ref{BTE}--\ref{Fi_gtr_match} form a closed set of equations for the distribution functions and the electric potential energy provided they are augmented by boundary conditions far away from the interface, at $z=z_b\!<\!0$ and $z=z_w\!>\!0$ (see Fig.~\ref{IntModelNotation}). \begin{figure}[t] \includegraphics[width=\linewidth]{iTiO2Edges} \caption{Upper panel: Edges for the conduction and valence bands (solid red and blue lines), the free hole motion (dashed blue lines), and the potential energies for electrons and ions on the plasma side (also indicated by solid red and blue lines) for an intrinsic ${\rm TiO}_2$ surface in contact with an hydrogen plasma. Shaded regions indicate respectively the reservoir and source which have been set up to provide the correct physical boundary conditions for the double layer. Lower panel: Density profiles $\rho_w(z)=[n_*(z)-n_h(z)]\theta(-z)$ and $\rho_p(z)=[n_i(z)-n_e(z)]\theta(z)$. By definition they are both positive outside the shaded regions, that is, in the regions which are physically relevant. The interface is collisionless on both sides and perfectly absorbing. Material and plasma parameters are $\chi=4.8\,{\rm eV}$, $E_g=3.3\,{\rm eV}$, $m_e^*=m_e$, $m_h=0.8\,m_e$, $\varepsilon=6$, $I=13.6\,{\rm eV}$, $\Gamma=2\,{\rm eV}$, $k_BT_*=k_BT_h=0.2\,{\rm eV}$, $k_BT_e=10\,k_BT_i=2\,{\rm eV}$. Due to the somewhat unrealistic temperatures of the charges on both sides of the interface (required to stabilize the numerics) the Debye screening length on the solid side $\lambda_D^w=2.2\cdot 10^{-6}\,{\rm cm}$ is comparable to the Debye screening length on the plasma side $\lambda_D^p=1.6\cdot 10^{-6}\,{\rm cm}$. Due to the absence of collisions the numerical values, however, should not be taken literally. A collisional theory would produce different density and potential profiles. } \label{IntModelData} \end{figure} Essential for the selfconsistent description of the electric response is the implementation of the boundary conditions far away from the interface. It depends on how the interface, and hence the electric double layer, is electrically connected to the outside. For an electrically floating interface the double layer is embedded between two quasi-neutral, field-free regions. The distribution functions in these regions are of course unknown. They are themselves the result of the electric response and can thus not be used as boundary conditions for the Boltzmann equations. A possibility to overcome this complication is to enforce two inflection points in the profile of $U_C(z)$ mimicking the quasi-neutral, field-free surroundings of the double layer, see Fig.~\ref{IntModelNotation}. The price to be paid comes in the form of eight parameters: Four boundary densities ($n_{b*}, n_{bh}, n_{se}$, and $n_{si}$), for each of the charged species considered, and four spatial coordinates ($z_b, z_1, z_p$, and $z_w$), two of them denote the positions of the inflection points ($z_1, z_p$), one the position of the particle reservoir in the solid ($z_b$), and one the position of the source ($z_w$) in the plasma. The adjustment of the distribution functions to the interface can then be emulated without simulating the bulk of the solid and the plasma. The framework outlined indicates what is needed for an integral modeling of the electric response. Besides specifying collision processes on both sides of the interface, matching and boundary conditions for the distribution functions have to be worked out--based on, respectively, the emissive properties of the plasma-solid interface and its asymptotics on both sides. To gain first insights into the selfconsistent electric response of a plasma-solid interface Bronold and Fehske~[\onlinecite{BF17}] applied the equations listed above to a collisionless, perfectly absorbing interface. By ignoring the collision integrals, the lateral momentum $\vec{K}$ could be eliminated from all equations, including the matching conditions. In addition, the solutions of the Boltzmann equations turned out to be no longer explicit functions of $z$ but only functions of $E$ and $U_C(z)$. The first integral of the Poisson equation could thus be obtained analytically, greatly simplifying the further numerical treatment. The selfconsistent embedding finally led to four nonlinear algebraic equations for the potential energies $U_C(z_b), U_C(z_1), U_C(z_p)$, and $U_C(z_w)$ which could be solved numerically with moderate effort. Results for the perfectly absorbing, collisionless interface--for parameters applicable to intrinsic ${\rm TiO}_2$ in contact with an hydrogen plasma--are shown in Fig.~\ref{IntModelData}. In a collisionless theory recombination and generation of electron-hole pairs inside the solid are excluded. Nothing balances thus--strictly speaking--the permanent influx of electrons and ions onto the interface. For a quasi-stationary regime to develop it is necessary to introduce by hand a recombination condition. The condition employed in~Ref.~[\onlinecite{BF17}] utilized the fact that intraband energy and momentum relaxation is much faster than interband recombination. The electron and hole densities inside the solid can thus be split into thermalized/trapped [$n^t_{*,h}(z)$] and free parts [$n_{*,h}^j(z)$], with only the former acting as a source in the Poisson equation, and the latter assumed to cancel once integrated over a spatial strip extending from the inflection point $z_1<0$ inside the solid to the interface at $z=0$, \begin{eqnarray} \int_{z_1}^0 \!\! dz \, \rho_w^j(z) = \int_{z_1}^0 \!\!dz \, [\,n_*^j(z)-n_h^j(z)\,] = 0~. \label{ReCombCond} \end{eqnarray} With this postulate the collisionless theory provided in total eight equations for the eight parameters mentioned above. A selfconsistent description of the electric response, indicated by a selfconsistent embedding of the electric double layer between two quasi-neutral, field-free regions, could thus be realized without taking collisions into account. The absence of collisions can be at most justified on the plasma side, where electrons are strongly depleted and hence nearly collisionless, and ions are only subject to ion-neutral collisions which however have to be considered only in particular situations~[\onlinecite{Sternovsky05,Riemann03,SG91}] or when a very precise description of the plasma sheath is required~[\onlinecite{TC17}]. Hence, on the plasma side ignoring or including collisions is essentially a question of how accurate one wants to describe the plasma sheath. On the solid side of the interface, however, neglecting collisions makes the theory conceptually incomplete. Without recombination, either radiatively or non-radiatively, both involving collisions between conduction band electrons and valence band holes, with the latter mediated by trap states (see Fig.~\ref{IntModelProcesses}), no quasi-stationarity can be achieved because nothing balances the injection of conduction band electrons and valence band holes from the plasma. The carrier concentration inside the solid would just grow indefinitely. Future work has thus to include collisions, at least on the solid side of the interface. The ad-hoc condition~\eqref{ReCombCond} could then be replaced by a consistent truncation of the perturbation theory treatment of charge relaxation and recombination. Work in this direction is in progress. In this subsection we discussed a general framework extending the kinetic modeling of the charge dynamics of the plasma to the inside of the solid. Within such an approach the electric modification of the plasma-solid interface--the build-up of the electric double layer--can be described selfconsistently, treating the plasma-based positive and the solid-based negative part of the double layer on an equal footing. In particular for hybrid electronics, using arrays of microdischarges integrated on semiconducting substrates, it may be necessary to adopt this type of integral modeling. In discharges used for materials processing, chemical and structural responses of the interface, involving the formation of adsorption layers and/or the sputtering of the outer layers of the interface, are of course intimately coupled to the electric response. To include them into the framework we just described is in principle possible. Staying at the level of Boltzmann equations is rather advantageous at this point. For each species of interest one has to add a Boltzmann equation with appropriate collision terms and matching/boundary conditions. Ways to treat rough and disordered interfaces are also conceivable. The numerical effort however will be rather high. An integral modeling, taking the electric, chemical, and structural response of the plasma-solid interface selfconsistently into account will be possible using high performance computing. The limitations of the present approach are mostly due to the simplified treatment of the solid. The related processes in the surface material require \textit{ab initio} modeling. In the next section such an approach that is based on nonequilibrium Green functions will be outlined. At the same time, there the treatment of the plasma particles will be substantially simplified compared to the present treatment. So both methods have complementary features. \section{Introduction}\label{s:intro} Modern plasma physics has three main research topics [\onlinecite{doe-report-17}]: high-temperature plasmas, in particular magnetic fusion; high-density plasmas (``warm dense matter'', laser plasmas, inertial confinement fusion); and low-temperature plasmas. The location of these areas in the density-temperature plain is sketched in Fig.~\ref{fig:nt-plane}. In each of these fields, the processes at the plasma wall play a crucial role, both, for fundamental understanding and for technological applications. Therefore, progress in the simulation of plasma-solid interaction is of crucial importance in each of these fields. \begin{figure}[h] \begin{center} \hspace{-0.cm}\includegraphics[width=0.5\textwidth]{n-t-ebene-sfb.png} \end{center} \caption{Low-temperature plasmas (LTP) are one of three main current research topics in plasma physics, aside from magnetic fusion and dense plasma (warm dense matter, WDM) \cite{doe-report-17}. These systems cover a huge parameter range in the density-temperature plane. LTP (the blue box) range from low (electron) density to atmospheric pressure (right edge). Representatives of solids facing the plasma are metals and semiconductors (electron-hole plasmas) sketched by the red areas. Relevant dimensionless parameters are the classical coupling parameter, $\Gamma=e^2/{\bar r} k_BT$, the quantum coupling parameter, $r_s={\bar r}/a_B$ and the degeneracy parameter of the electrons, $\Theta=k_BT/E_F$, with ${\bar r}, a_B$ and $E_F$ denoting the mean interparticle distance, the Bohr radius and the Fermi energy, respectively.} \label{fig:nt-plane} \end{figure} Here we concentrate on plasma-solid processes in low-temperature plasma, although some of our results are expected to be of interest also for the high-temperature plasmas. This field has experienced impressive progress over the last two decades, both, in experiments and applications. Aside from traditional applications also new materials, in particular \textit{nanomaterials}, are coming into the focus [\onlinecite{meyyappan_jpd_11,ostrikov_aip_13}]. This includes carbon based materials such as carbon nanotubes and graphene nanoribbons that have a size-dependent bandgap and promise exciting electronic and optical properties, e.g. [\onlinecite{son_prl_06, prezzi_prb_08}]. The role of plasmas, in the context of these novel materials, is only poorly explored yet. There are impressive first results on plasma synthesis of these systems. On the other hand, it will also be interesting to use such materials inside a plasma and to utilize their properties in a discharge environment. These applications are only emerging, and a brief discussion of some aspects will be given in this paper. Yet for their success, and for solids embedded in plasmas in general, it will be crucial to have available accurate simulations of the plasma-solid interface, as has been pointed out in many places, e.g. [\onlinecite{adamovich_2017_plasma, doe-report-17}]. The interest in low-temperature plasmas arises from the peculiar properties of these systems. These plasmas typically have a low degree of ionization and cover a broad pressure range, from below one Pa to atmospheric pressure, see Fig.~\ref{fig:ltp-properties}. \begin{figure}[h] \begin{center} \hspace{-0.cm}\includegraphics[width=0.49\textwidth]{plasma_properties.pdf} \end{center} \caption{Low-temperature plasmas being composed of neutrals (atoms, molecules), ions and electrons comprise a number of very unusual properties: they are non-isothermal ($T_i\ne T_e$), far from thermal equilibrium (non-Maxwellian velocity distributions), and they contain electrons and ions of a very broad range of kinetic energies. In the plasma boundary region (``sheath'') ions may reach high energies that can be exploited for materials modification, sputtering and ion implantation. At the same time, electrons with energies in the eV range are able to excite and ionize neutrals and trigger chemical reactions.} \label{fig:ltp-properties} \end{figure} These plasmas are non-thermal, i.e. the electron and ion temperatures may differ by several orders of magnitude. Moreover, electrons and ions may be far from thermal equilibrium, being described by non-Maxwellian velocity distributions. The existence of energetic electrons with temperatures in the eV-range, which is sufficient to excite or ionize atoms and molecules, is of high interest for applications in surface chemistry, biology and even medicine. On the other hand, these plasmas may contain highly energetic ions that are accelerated by strong electric fields, in particular in the surface near region (the ``plasma sheath''). These energies are sufficient for mechanical modification of the solid surface such as defect creation, ion implantation or sputtering. Even though there have been remarkable recent advances, both, in plasma modeling and surface science simulations, the combination of the two is still at a very early stage. Current simulations in low-temperature plasma physics that are based on kinetic theory or fluid theory have achieved a high quality description of the dynamics of electrons, ions and neutrals in the plasma bulk, including elastic and inelastic scattering processes, cf. top part of Fig.~\ref{fig:psi_today}. At the same time, these simulations often omit plasma-surface processes or treat them phenomenologically. \begin{figure}[h] \begin{center} \hspace{-0.cm}\includegraphics[width=0.5\textwidth]{left_plasma_half.png}\\ \hspace{-0.cm}\includegraphics[width=0.5\textwidth]{right_solid_half.png} \end{center} \caption{\textbf{Top}: Sketch of the current approach to include surface properties in plasma simulations via phenomenological parameters such as sputter rates $R^{\rm sput}$, electron and ion sticking coefficients, $S_e, S_i$ and the secondary electron emission coefficient, $\gamma_e$. While the plasma is treated by advanced approaches, the atomic structure of the solid and the surface is not resolved. \textbf{Bottom:} In contrast, in surface science, atomic level information of the surface is taken into account, whereas plasma effects are approximated via independent impacts of neutrals or ions.} \label{fig:psi_today} \end{figure} For example, in many advanced kinetic simulations based on the Boltzmann equation, e.g. [\onlinecite{hagelaar_2005,donko_2016}] or particle in cell (PIC-MCC) simulations, e.g. [\onlinecite{ebert_2016,becker_psst17}] neutrals are treated as a homogeneous background, and their interaction with surfaces is not included in the description. However, the effect of energetic neutrals maybe crucial for secondary electron emission (SEE), as was demonstrated in PIC simulations of Derszi \textit{et al.} [\onlinecite{derzi_2015_effects}]. Another effect that can be important for the behavior of the plasma are the properties of the surface, such as surface roughness, oxidation or coverage by an adsorbate. For example, Phelps and Petrovic [\onlinecite{phelps_1999_cold}] convincingly demonstrated that the plasma modification of a metal surface may change SEE by several orders of magnitude, see Sec.~\ref{s:influences}. The conclusion is that an accurate theoretical treatment of the solid surface may be very important for low-temperature plasma simulations. This suggests to resort to surface science methods where a very accurate atomic level treatment of solid surfaces has been achieved by \textit{ab initio} methods such as density functional theory (DFT). Surface simulations have also incorporated the impact of energetic projectiles to simulate sputtering, e.g. [\onlinecite{ brault_frontiers_18}] or energy loss (stopping power), e.g. [\onlinecite{zhao_2015_comparison,balzer_prb16}]. However, these are typically simulations where single ions or neutrals are treated, but the effect of a plasma and of its nonequilibrium properties [cf. Fig.~\ref{fig:ltp-properties}] and the plasma-induced modifications of the surface have not been taken into account so far. From the examples presented above that demonstrated the mutual influences between plasma and surface, it is clear that further progress in an accurate modeling of the plasma-solid interface requires to go beyond an independent treatment of both sides. Instead it is necessary to develop a combined theory and simulation of the entire system. This is a challenging project that requires strong input from plasma physics and surface science, simultaneously. In fact, such a research effort is under way at Kiel University, in collaboration with Greifswald University and the INP Greifswald. The concept of this project was first presented in 2015 [\onlinecite{interface}] and continuously developed since then. It is the goal of this article to present these ideas and first simulation results and to outline further directions and perspectives of development. We note that similar concepts have been developed by Graves, Brault and Neyts and others in the frame of MD simulations, e.g. [\onlinecite{graves_2009_molecular, neyts2017molecular}], see Sec.~\ref{s:mesoscopic}. The main difference is that those simulations usually neglect the electronic degrees of freedom, in particular, quantum effects and internal relaxation processes in the solid. This paper is organized as follows. In Sec.~\ref{s:influences} we discuss the mutual influences between plasma and solid that motivates the development of a novel approach combining plasma and surface science methods. In Sec.~\ref{s:concepts} we discuss more in detail the relevant physical processes and effects at the plasma-solid interface. This sets the basis for the required theoretical approaches that are capable to accurately simulating plasma-surface processes and discuss their respective advantages and problems. We identify four different methods that are discussed in detail with increasing degree of complexity. In Sec.~\ref{s:mesoscopic} we discuss the first one -- mesoscopic approaches to the plasma-solid interface -- in particular molecular dynamics and discuss acceleration approaches that are of potential relevance for plasma-surface interaction. In Sec.~\ref{s:qkin} we discuss the second class of methods that is based on the quantum Boltzmann equation. In Secs.~\ref{s:abinit} we consider the third approach that is based DFT and TDDFT. Finally, in Sec.~\ref{s:negf} the fourth approach is discussed that is based on nonequilibrium Green functions (NEGF) and leads to generalized quantum kinetic equations. The analysis concludes with Sec.~\ref{s:integrated} where we outline first steps towards an integrated plasma-surface modeling, and we present our conclusions in Sec.~\ref{s:conclusion1}. \section{Mutual influences of plasma and solid}\label{s:influences} \begin{figure} \begin{center} \hspace{-0.cm}\includegraphics[width=0.4\textwidth]{phelps_see.png}\\ \hspace{-0.cm}\includegraphics[width=0.4\textwidth]{see_donko2.png} \end{center} \caption{\textbf{Top}: Secondary electron emission yield per argon atom or ion for a broad variety of metals (symbols). ``Clean metals'' refers to beam experiments where the surface was cleaned via ion sputtering. ``Dirty metals'' denotes measurements following varying degrees of surface exposure to oxygen, to water or ambient gas. Figure from Ref.~\onlinecite{phelps_1999_cold} where additional details are given. \textbf{Bottom:} Effect of SEE on ion density in the bulk of a AC discharge (13.56 MHz, p=5Pa, for varying electrode voltage), for different SEE models: no SEE (black) to a full treatment of ion and atom induced SEE (green). From Ref.~\onlinecite{derzi_2015_effects}.} \label{fig:see} \end{figure} As we discussed in the Introduction there are many processes that couple the plasma and the surface. Here we discuss a few examples. The first is the effect of energetic neutrals impacting the surface. These neutrals are efficiently produced in the case of a strong sheath electric field, by means of charge exchange collisions. Energetic neutrals maybe crucial for secondary electron emission from metal surfaces, as was shown in Ref.~\onlinecite{phelps_1999_cold}, see also Fig.~\ref{fig:see}. This was confirmed by PIC simulations where neutrals above a threshold energy of 23eV were traced individually [\onlinecite{derzi_2015_effects}]. An example where the surface properties affect the plasma is related to plasma electrons hitting a solid surface. The standard assumption in plasma simulations is that these electrons are lost without reflection, e.g. [\onlinecite{sheehan_2013_kinetic}]. Only recently a microscopic calculation of the electron sticking coefficient was performed by Bronold and Fehske [\onlinecite{bronold_2015_absorption}] that demonstrated that this picture has to be revised. Recent PIC simulations by Sun \textit{et al.} confirmed that finite electron reflectivity, in combination with SEE, may have a significant influence on the plasma parameters of a CCRF discharge for pressures of several tens of Pa [\onlinecite{sun_psst_2018}]. An even more striking example of the effect of the properties of the surface on the plasma properties is secondary electron emission. Phelps and Petrovic compiled extensive experimental data for the SEE yield from different metals over a broad range of impact energies of argon atoms and ions [\onlinecite{phelps_1999_cold}]. They compared SEE from surfaces that were cleaned by ion sputtering (cf. curves labeled ``clean metals'' in the top part of Fig.~\ref{fig:see}) to the SEE yield obtained from surfaces that have been in contact with a plasma (or to oxygen or ambient gas, for the original references see Ref.~\onlinecite{phelps_1999_cold}). The authors note a dramatic difference of the SEE yield from a ``dirty'' surface compared to a ``clean'' one which may exceed two orders of magnitude for SEE due to neutral atoms. For energetic atoms (energies above 200 eV) this difference is much larger than the difference between different metals. The crucial importance of SEE has been confirmed in many simulations. For example, Derszi \textit{et al.} performed PIC simulations where they included SEE according to various models via modified cross sections ~[\onlinecite{derzi_2015_effects}]. In that work it was found that a realistic (``dirty'' [\onlinecite{phelps_1999_cold}]) surface gives rise to a significant increase (up to a factor of two) of the ion density, even far away from the electrode compared to simulations where SEE is neglected. This trend is seen in the bottom part of Fig.~\ref{fig:see}, compare the green and black curves. At the same time, the experimental data shown in the top part of Fig.~\ref{fig:see} suggest that there remain substantial uncertainties in the values of the SEE coefficient. It is also not clear how long the surface was treated. In a real plasma treatment experiment a ``clean'' surface may correspond to the initial state of an electrode which, ultimately, turns into a ``dirty'' metal that is covered by adsorbates or an oxide layer. Thus, more accurate experimental and theoretical knowledge of the SEE for different materials and varying degrees of surface coverage will be very important for applications. Strong surface effects have also been observed in the field emission. For example, Li \textit{et al.} studied the effect of surface roughness on the field emission by including a phenomenological geometric enhancement factor [\onlinecite{li_2013}]. These examples show that the simple concept of fixed phenomenological surface parameters, such as the SEE coefficient $\gamma_e$, for a given surface material, that was sketched in Fig.~\ref{fig:psi_today} has to be questioned. Instead the SEE coefficient has to take into account the properties of the surface and also the exposure to the plasma. \begin{figure}[h] \begin{center} \hspace{-0.cm}\includegraphics[width=0.5\textwidth]{new_picture_without_arrows.png} \end{center} \caption{Sketch of the plasma-solid interface which comprises the plasma sheath and the plasma facing activated layers of the solid [\onlinecite{interface}]. Among the relevant processes are diffusion, adsorption (``sticking'') and desorption of neutrals, penetration (stopping) of ions and electron transfer between solid and plasma. The influence of the plasma on the solid and vice versa is a major challenge for a predictive theoretical treatment and require a combination of various theoretical approaches, see Fig.~\ref{fig:theory}.} \label{fig:psi} \end{figure} In order to make reliable predictions, novel theoretical concepts are needed that include the whole scope of complex physical and chemical processes that occur at the plasma-solid interface which include secondary electron emission, sputtering, neutralization and stopping of ions, adsorption and desorption of neutral particles as well as chemical reactions. Therefore, the plasma-solid system should be treated as single entity which we call ``plasma-solid interface'' [\onlinecite{interface}]. It comprises the plasma sheath and the plasma facing atomic layers of the surface that are influenced (excited or ``activated'') by the plasma. This new theoretical concept is sketched in Fig.~\ref{fig:psi}. \section{The plasma-solid interface: Physical processes and theoretical approaches}\label{s:concepts} \subsection{Overview on the physical processes at the plasma-solid interface}\label{ss:psi-processes} Let us now look more in detail at the properties of the plasma-solid interface and the relevant processes. We already discussed in the beginning, cf. Fig.~\ref{fig:nt-plane}, that plasma and solid are characterized by a huge density gap leading, in many cases, to an enormous difference in length and time scales. First of all, the plasma-solid system is in a stationary state. This state differs from thermal equilibrium due to the nonequilibrium character of the plasma [cf. Fig.~\ref{fig:ltp-properties}] and due to fluxes of particles (electrons, ions, neutrals) that cross the boundary in both directions. On the largest scale (the scale of the Debye length, Fig.~\ref{fig:interface-physics}.a) the interface is characterized by the density profiles of electrons and ions. Electron depletion in the plasma sheath near a surface [cf. Fig.~\ref{fig:psi_today}] gives rise to an excess positive charge in front of the surface. The missing negative charge has to accumulate inside the surface giving rise to an \textit{electric double layer}. Charged double layers are a common phenomenon in liquids and were originally studied in electrolytes by Helmholtz~[\onlinecite{helmholtz}]. The corresponding effect for the plasma-solid interface was predicted by Bronold and co-workers [\onlinecite{heinisch_2012_electron}] and turns out to be very different, by the different composition of the system, and more complex, due to its nonequilibrium nature. The peculiarities of the electric double layer that are caused by the plasma properties--the non-Maxwellian velocity distributions and the time variation, in case of an rf field, on the scale of nanoseconds--are qualitatively understood. On the other hand, what is far less understood, is the impact of the solid on this charge distribution: the influence of the nanoscale surface structure [Fig.~\ref{fig:psi_today}.b] as well as of the atomic scale structure [Fig.~\ref{fig:psi_today}.c] and processes between solid and plasma. \begin{figure*}[h] \begin{center} \hspace{-0.cm}\includegraphics[width=0.88\textwidth]{the_plasma_interface_uebersicht.png} \end{center} \caption{Sketch of the physical processes at the plasma-solid interface--from the largest to the smallest length scale. \textbf{Top}: The electric double layer (on the scale of the Debye length, on the plasma side, and a few nanometers, in the solid) resulting from electron depletion in the plasma sheath [cf. Fig.~\ref{fig:psi_today}] is characterized by the local difference of the nonequilibrium ion and electron densities and is accompanied by electron accumulation in the solid which is influenced by the processes in figure parts b) and c). \textbf{Bottom left}: on the scale of the surface roughness (typically nanometers) the surface exhibits local variations of the morphology and chemical composition. \textbf{Bottom right}: atomic scale modification of the surface and the plasma sheath caused by individual particle impacts, charge transfer, chemical reactions etc. The relevant processes are indicated inside the figure parts.} \label{fig:interface-physics} \end{figure*} Thus, an analysis of the surface and near-surface structure of the solid in the presence of the plasma and of surface processes under these conditions is required. This includes the structural and chemical response of the material. Surface science experiments are required to investigate the atomic-scale structure [Fig.~\ref{fig:psi_today}.c], e.g., surface relaxation and reconstruction, adsorbate species, and surface defects, as well as the evolution of the nanoscale morphology, e.g., the formation of islands, pits, steps, and ultrathin films on the surface [Fig.~\ref{fig:psi_today}.b]. Also the modification of the material in the near-surface region (e.g., crystallinity, porosity, composition, defect density) has to be studied as function, e.g., of ion and neutral particle energy and plasma density. It would be highly desirable that surface science and plasma physics experiments obtain the concentration of charged and neutral species at and near the surface as well as chemical binding energies of various species. The information that should be obtained from theory includes the relevant cross sections and rate coefficients, neutral sticking coefficients, SEE coefficients, diffusion coefficients, information on energy dissipation channels and time scales etc. Another key topic is the charge transfer dynamics across the interface in the presence of the plasma. Here, plasma physics experiments are needed that provide key information, including the electron and ion fluxes to and from the surface [Fig.~\ref{fig:psi_today}.a]. This should be complemented by surface science experiments measuring the secondary electron current from the surface and the plasma-induced modification of the band structure of the surface material. These quantities are the combined result of a multitude of physical processes. Theory and simulations should attempt to resolve the individual contributions due to the neutralization of ions in front of the surface or electron impact ionization and ion stopping inside the solid. Furthermore, it would be desirable if surface science experiments could provide insight into the complex energy landscape of the solid surface, its modification due to the plasma environment, as well as the plasma-induced space charge region inside the solid. The experimental information should be complemented by novel theoretical approaches that will lead to accurate results for electron and ion velocity distributions in the plasma sheath, on secondary electron emission, on the sticking coefficients of electrons, ions, and neutrals and on the charge of nanoparticles embedded in the plasma. All these processes evolve in time (on the scale of seconds), as a result of the surface modification by the plasma. Thus, there is a direct coupling between effects on the atomic scale and the macroscopic plasma behavior that needs to be explored. \subsection{Theoretical approaches for the plasma-solid interface}\label{ss:psi-theory} Let us discuss, in the following the theoretical strategy needed to tackle these problems. Aside from the different pressure, length and time scales, the main difficulty is that both sides of the plasma-solid interface are governed by completely different physics: low-density gas-like behavior, in the plasma, versus quantum dynamics of electrons, in the solid, coupled to the lattice dynamics; this situation, the relevant processes and scales are sketched in ~Fig.\ref{fig:theory}. An accurate simulation of plasma-surface processes, first of all, requires a reliable description of the plasma and the solid. Standard tools in plasma simulations are fluid simulations and kinetic theory (cf. blue box in Fig.~\ref{fig:theory}). Here two main approaches are in use: direct solution of the Boltzmann equation or particle in cell simulations with Monte Carlo collisions (PIC-MCC). The particle dynamics have to be coupled to the dynamics of the electromagnetic field on the basis of Maxwell's equations or the Poisson equation for the electrostatic potential. Finally, these simulations require surface parameters as an input---the fluxes $\textbf{J}^s$---and deliver the corresponding fluxes $\textbf{J}^p$, as an output, cf. Fig.~\ref{fig:theory}. To obtain the necessary surface information, one first of all needs to obtain the ground state properties of the solid--the energy spectrum (band structure) and the Kohn-Sham orbitals--which is done by density functional theory (DFT) simulations [Sec.~\ref{s:abinit}], cf. right part of Fig.~\ref{fig:theory}. However, DFT is known to have problems, in particular, in treating materials with strong electronic correlations including various oxides. Here, many-body approaches are being used that include the Bethe-Salpeter equation (BSE), e.g. [\onlinecite{onida_rmp_02}], dynamical mean field theory (DMFT), e.g. [\onlinecite{kotliar_rmp_06}], or quantum Monte Carlo (QMC) for the ground state or finite temperature [\onlinecite{foulkes_rmp_01,DORNHEIM_physrep18}]. \begin{figure*} \begin{center} \hspace{-0.cm}\includegraphics[width=0.78\textwidth]{theory_fcse.pdf} \end{center} \vspace{-2.0cm} \caption{Theoretical methods for the description of the plasma-solid interface \cite{interface}, as sketched in Fig.~\ref{fig:psi}. Some of the processes of interest are listed in the figure. Note the dramatically different length scales and the very different properties of the plasma and the solid, requiring fundamentally different methods to be applied on the plasma and the solid side. Standard methods for the bulk solid are Density functional theory (DFT), Bethe-Salpeter equation (BSE), Dynamical Mean Field Theory (DMFT), and Quantum Monte Carlo (QMC). To simulate surface processes (central box), additional non-adiabatic (time-dependent) approaches are required: molecular dynamics (MD), Kinetic Monte Carlo (KMC), Quantum Boltzmann eqution, Born-Oppenheimer MD (BO-MD), time-dependent DFT (TDDFT), Nonequilibrium Green functions (NEGF) and \textit{ab initio} NEGF (AI-NEGF). To account for the complex interactions between plasma and solid, the corresponding methods have to be properly linked: plasma simulations should provide the momentum dependent fluxes $\textbf{J}^p_a$ of all species ``a'' to the surface whereas surface simulations deliver the corresponding fluxes $\textbf{J}^s_a$ that leave the surface. Bulk solid simulations provide the band structure $\epsilon_\lambda$ and reactive force fields (FF), whereas surface simulations return the updated surface morphology, chemical modifications etc. For details see text.} \label{fig:theory} \end{figure*} \begin{figure*} \begin{center} \hspace{-2.cm}\includegraphics[width=1.1\textwidth]{methods_space_time_nokmc.pdf} \end{center} \vspace{-2.0cm} \caption{ Approximate length and time scales accessible with different simulation methods for plasma-surface applications that are listed in Fig.~\ref{fig:theory}. The shortcuts are the same as in Fig.~\ref{fig:theory} except for DFT-MD, which is equivalent to BO-MD, and QBE which stands for Quantum Boltzmann equation. Note that the comparison is only qualitative as different methods may apply to different processes. Also, the physical time resolution is often much less (larger minimum scale) than the required time step that is dictated by numerical stability. The \textit{ab initio} approaches \textbf{TDDFT} and \textbf{NEGF} resolve the electronic length and time scales and apply to ultrafast processes. \textbf{DFT-MD} has a more limited resolution of electronic relaxation processes. The upper limits of DFT, TDDFT and NEGF are set by the required basis dimension and accessible number of time steps. \textbf{QBE} resolves spatial details on the level of 100 inter-atomic distances and the relaxation time of the electrons in the solid. The upper length limit is determined by the imposed accuracy (level of coarse graining, pink arrow). \textbf{MD} propagates only the heavy particles neglecting electronic degrees of freedom. The accessible simulation dimensions can be increased via parallelization. Simulation times are restricted by the number of time steps and can, in some cases, be extended by additional ``acceleration'' methods (blue arrows). On the plasma side, \textbf{PIC-MCC} simulations resolve approximately one tenth of the electron Debye length and one plasma period and may extend to centimeters and milliseconds. \textbf{Fluid simulations} contain an additional coarse graining with respect to the particle velocities which limits their lower scale limits, compared to PIC, but extends their upper limits. KMC may, in principle, extend plasma-surface simulations to minutes and millimeters but is not considered here due to its largely uncontrolled character for the present applications, see text. Figure adapted from Ref.~\onlinecite{abraham_phd_18}. } \label{fig:scales} \end{figure*} Next, if the solid comes in contact with a low-temperature plasma, energetic electrons, ions or neutrals may excite the electrons of the solid and the lattice. This is already not captured by ground state DFT but requires time-dependent extensions, cf. the approaches listed in the central box of Fig.~\ref{fig:theory}. \textit{The top row in the central box} lists mesoscopic approaches--molecular dynamics (MD) and kinetic Monte Carlo (KMC)--that do not treat the quantum dynamics of the electrons explicitly. This simplification allows one to access comparatively large time and length scales, see also Fig.~\ref{fig:scales}. In particular with KMC, in principle, one can reach minutes of simulation time and length scales of centimeters. KMC has been successfully used in many plasma simulations, e.g. [\onlinecite{marinov_2016, guerra, abraham_jap_15, fujioka_phd_15, polonsky_epjd18}], and it is an integral part of many multiscale simulation concepts because it is very flexible with respect to the inclusion of new processes. However, all these processes are treated in a very simplified manner using process rates that often do not include all relevant parameter dependencies. In cases where the complete set of relevant processes is well known, KMC may, nevertheless, be a powerful tool. In contrast, for new problems--such as the plasma-solid interface--where the complete set of events is not known \textit{a priori}, the accuracy and predictive power of KMC is rather limited. We will, therefore, not discuss KMC in detail here--the interested reader is referred to Refs.~\onlinecite{guerra, marinov_2016,rosenthal_phd_13, abraham_phd_18}. Instead, here we will concentrate on MD because it has a much stronger foundation and can retain a first-principle character once the information on the interaction potentials is derived from \textit{ab initio} approaches in cases when electronic and quantum effects are not important. We will give a more detailed discussion in Sec.~\ref{s:mesoscopic}. For an overview on the length and time scales that can be reached by MD and the other methods, see Fig.~\ref{fig:scales}. The \textit{second row in the central box} of Fig.~\ref{fig:theory} lists models based on the quantum Boltzmann equation -- the quantum generalization of classical kinetic equations. Here extensive recent work is due to Bronold and co-workers who developed simulations of the charge transfer, electron sticking and other processes. The corresponding approach and some results will be summarized in Sec.~\ref{s:qkin}. The \textit{third row in the central box} of Fig.~\ref{fig:theory} is devoted to time-dependent simulations that are based on DFT. The first approach is Born-Oppenheimer MD where the ions are moved with classical molecular dynamics whereas the electrons are assumed to follow the ion dynamics adiabatically, thereby remaining in the (time evolving) ground state that is obtained quantum-mechanically, by ground state DFT. However, for strong excitation and/or fast processes the adiabatic approximation fails. Even though in plasma-surface interaction the mean excitation of the surface (per atom) may be small, local excitations may be strong, e.g. due to the impact of plasma particles. The corresponding non-adiabatic extension of DFT is time-dependent DFT (TDDFT) [\onlinecite{runge_gross}]. TDDFT is successfully being used for many surface processes. Applications to ion stopping have been performed e.g. in Ref. \onlinecite{zhao_2015_comparison}. An overview on this method is presented in Sec.~\ref{s:abinit}. Finally, the \textit{fourth row in the central box} contains another \textit{ab initio} method: nonequilibrium Green functions (NEGF), e.g. [\onlinecite{balzer-book,schluenzen_cpp16}]. This method generalizes the quantum Boltzmann equation (second row, Sec.~\ref{s:qkin}) to fast processes and is, in particular, well suited to accurately treat electronic correlation effects in the surface material. First applications to ion stopping were performed recently [\onlinecite{balzer_prb16}]. Finally, we also list \textit{ab initio} NEGF simulations -- a recently developed combination of ground state DFT and NEGF [\onlinecite{marini_2009_yambo}]. An overview on NEGF methods will be presented in Sec.~\ref{s:negf}. Even though the \textit{ab initio} methods are, obviously, the most accurate theoretical approaches, they are extremely CPU-time demanding which strongly limits the accessible length and time scales. For example, Born-Oppenheimer MD simulation require a time step around $0.1\dots 1$ femtoseconds, which allows one to treat on the order $100\dots 1000$ atoms for $1\dots 100$ picoseconds, during a week of simulations on massively parallel hardware, e.g. Ref.~\onlinecite{hutter_wires_12}. The demand for TDDFT and NEGF is several orders of magnitude larger. Typical length and time scales are summarized in Fig.~\ref{fig:scales}. Thus, despite their accuracy, it is prohibitive to apply \textit{ab initio} methods to all problems of the plasma-solid interface. They should be applied to those processes where such an involved treatment is without an alternative, in particular, when important processes would be lost otherwise, e.g. via an averaging or coarse graining procedure. Even though there is an impressive list of applications of all four plasma-surface simulation approaches, until now these have been developed essentially in isolation from each other. At the same time there is a high need for smart combinations of the different methods to cover the length and time scales of the plasma-solid interface at a sufficient accuracy and to capture all relevant physical and chemical processes. We hope that the present analysis of each of theses methods will highlight their strengths and weaknesses and stimulate such combinations. \section{Mesoscopic simulation approaches: molecular dynamics. Acceleration and extension concepts}\label{s:mesoscopic} We start with the method that extends to the largest time and length scales of those that are listed in the center of Fig.~\ref{fig:theory}: molecular dynamics [for the discussion of KMC, see Sec.~\ref{s:concepts}]. This method is based on empirical interaction potentials and does not include quantum effects in the dynamics of the particles, in particular no explicit electronic effects. This approach has to be clearly distinguished from \textit{ab initio} MD (or Born-Oppenheimer) MD where the electrons are time-propagated as well using density functional theory simulations (see Sec.~\ref{s:abinit}). Nevertheless, such a semiclassical modeling is often sufficient for the treatment of the dynamics of neutral particles on a surface: diffusion, adsorption and desorption or many chemical reactions--a technique that is well developed in surface science and in theoretical chemistry, e.g. Ref.~ \onlinecite{gross-book}. Similarly, MD simulations are well established in low-temperature plasmas, e.g. to compute first principle structural properties of dust particles [\onlinecite{com-plasma_springer_14}] or the diffusion coefficient in a strongly correlated magnetized plasma [\onlinecite{ott_prl_11}]. MD simulations are also actively used in plasma-surface simulations, e.g. [\onlinecite{graves_2009_molecular,neyts2017molecular}], and recent applications include cluster growth [\onlinecite{abraham_jap16}] and sputtering [\onlinecite{brault_frontiers_18}]. The semiclassical MD simulations solve Newton's equations for all particles exactly. The quality of the results, obviously, crucially depends on the accuracy of the input data, most importantly, the effective pair potentials or force fields. These quantities are usually derived from microscopic quantum simulations or are adjusted to reproduce experimental data. Typically MD simulations for atoms require a time step of the order of $1$ fs and can, in principle, treat huge systems by using massively parallel hardware. For example, Ref.~\onlinecite{nakano_08} reported simulations of a system containing $10^{11}$ atoms that reach times of the order of several milliseconds. However, this is presently only possible on the largest supercomputers or on dedicated hardware, e.g.~[\onlinecite{piana_13}]. At the same time, even though parallelization allows to reach larger system sizes, it does not help to extend the simulation duration. Despite these impressive records, it is clear that in the near future MD simulations for plasma-surface processes will remain many orders of magnitude short of system sizes and length scales needed to compare with experiments. In plasma physics, these are minutes and (at least) micrometers, respectively. Therefore, additional strategies are needed. One way is of course the use of additional approximations leading to simplified models at the expense of accuracy and reliability. Here, we are concentrating on other methods that avoid simplifications of the equations of motion and to retain the first principles character of the MD simulations. The idea is to invoke additional information on the system properties that allow one to effectively \textit{accelerate} the simulations and/or to extend them to larger scales \textit{without loosing accuracy}. There exists a variety of acceleration strategies including hyperdynamics [\onlinecite{voter_97}], metadynamics [\onlinecite{LaioParrinello2002}] or temperature accelerated dynamics [\onlinecite{Sorensen2000}], for additional comments see Sec.~\ref{s:abinit}. A more recent concept is collective variable driven hyperdynamics [\onlinecite{Bal2015}] that was reported to achieve, for some applications, speed-ups of about nine orders of magnitude. These methods have been successfully applied in surface physics and chemistry, and a more detailed discussion of these very diverse acceleration/extension developments, of their respective strengths and limitations was presented recently in Ref.~\onlinecite{bonitz_psst18}. The above methods are not easily applied to the heterogeneous plasma-solid system. Here recently a multi-scale simulation concept has been proposed that overcomes the low-density problem of the gas phase, for details see Ref.~[\onlinecite{brault_frontiers_18}]. Another approach developed by two of the present authors [\onlinecite{abraham_jap16, abraham_cpp18}] uses a \textit{selective acceleration of some relevant processes} and, thereby, achieved speed-ups exceeding a factor $10^9$, see Sec.~\ref{ss:spa}. The third direction of developments we underline here does not aim at accelerating the \textit{ab initio} simulations but to extend them to longer times by a suitable combination with analytical models [\onlinecite{franke_2010_diffusion, filinov_psst18_1}]. These methods will be called below \textit{Dynamical freeze out of dominant modes} (DFDM) and are briefly discussed in Sec.~\ref{ss:freezout} \subsection{MD simulations employing selective process acceleration (SPA)}\label{ss:spa} The authors of reference \onlinecite{abraham_jap16} considered, as an example for plasma-surface interaction, the deposition of gold atoms onto a polymer surface. The MD simulations tracked each individual atom, its diffusion on the surface, the emergence and growth of clusters and, eventually, the coalescence of the latter. The influence of the plasma environment was treated statistically by taking into account the impact of energetic ions that leads to the formation of surface defects that trap incoming atoms and prevent their diffusion. Varying the fraction of trapped atoms allows to mimick the flux of ions to the surface. In the MD simulations the isotropic Langevin equation of motion for all gold particles with the mass $m$ and spatial coordinates $\mathbf{r}=(\mathbf{r}_1, \mathbf{r}_2, \ldots)$ were solved: \begin{equation} m \ddot{\mathbf{r}} = - \nabla U(\mathbf{r}) - \frac{m}{t_\mathrm{damp}} \dot{\mathbf{r}} + \sqrt{\frac{2m k_\mathrm{B}T}{t_\mathrm{damp}}} \,\mathbf{R}(t)\,, \label{eq:eqmotion_langevin} \end{equation} where the potential $U$ describes the interaction between gold particles. For this potential \textit{ab initio} force field data are being used (the MD simulations used the LAMMPS package). Further, $t_\mathrm{damp}$ has the role of a damping parameter, and $\mathbf{R}$ is a delta-correlated Gaussian random process. This random force and the viscous damping simulate the effect of the polymer on the heavy gold particles. This is motivated by the fact that the interaction between gold atoms by far exceeds the one between gold and polymer, so details of the latter are of minor importance. While the first term on the r.h.s. of Eq.~(\ref{eq:eqmotion_langevin}) favors cluster formation (gold atoms settle in the minima of the total potential), the last two terms induce a diffusive motion with the diffusion coefficient \begin{equation} \label{eq:diffusioncoeff} D=\frac{1}{m} {k_\mathrm{B}T t_\mathrm{damp}}\,. \end{equation} Thus, it is clear that the utilization of the Langevin dynamics allows one to control the speed of the surface diffusion and bulk by choosing a specific combination of the temperature $T$ and the damping parameter $t_\mathrm{damp}$. Beyond that, it is possible to add a spatial (or directional) dependence to the diffusion coefficient if one lets the damping parameters depend on the position of the particle. Based on the above considerations, Abraham \textit{et al}.~[\onlinecite{abraham_jap16}] developed a procedure to simulate the growth of nanogranular gold structures on a thin polymer film. By choosing appropriate ratios of the damping parameters, one can make sure that the atoms spend most of the time in the surface layer where they perform a random walk. The use of Langevin dynamics is restricted to the polymer surface whereas in the plasma, the dynamics of the gold atoms are purely microscopic. This allows one to add particles to the system by creating particles at the top of the simulation box and assigning them an initial velocity towards the substrate. Therefore, it is possible to perform the simulation with values of the deposition rates $J_\mathrm{sim}$ and diffusion coefficients $D_\mathrm{sim}$ that are much higher than the values in typical experiments. In Ref.~[\onlinecite{abraham_jap16}], it was argued that the simulations yield an adequate description of a real experimental deposition process if the ratio $J_\mathrm{sim}/D_\mathrm{sim}$ is equal to the ratio $J_\mathrm{exp}/D_\mathrm{exp}$ of the corresponding quantities in the experiment. The idea behind that is that -- at least at the early stage of the deposition process -- the growth should be essentially determined by the average distance an atom travels on the surface between successive depositions of atoms. Hence, the absolute time of the process is assumed to be irrelevant. The results presented in Ref.~[\onlinecite{abraham_jap16}] were obtained with a time step of $\SI{1}{fs}$, and a damping parameter for the diffusion in $x$- and $y$-directions of $\SI{1}{ps}$. The temperature and the deposition rate were set to match the conditions of the experimental results in Ref.~\onlinecite{schwartzkopf_2015_real} for the sputter deposition of gold on polystyrene. Using these parameters, the direct MD simulation time for the growth of a thin gold film is roughly $10^9$ times shorter than the corresponding time in the experiment. Or in other words, the duration of the MD simulations could be extended by nine orders of magnitude. To verify the validity of such a dramatic shift of the time scales and to obtain the applicability limits, comprehensive tests of the method against experimental results were performed, see also Refs.~\onlinecite{abraham_phd_18,abraham_epjd18} for a discussion. In particular, as one accelerates only selected processes, i.e., the deposition of atoms and the diffusion of atoms on the surface, one has to make sure that the neglect of other processes, e.\,g., the relaxation of a cluster structure, does not lead to artifacts in the simulation results. In Ref.~\onlinecite{abraham_jap16}, the method was tested by comparing several quantities describing the evolution of the gold film morphology with the results of time-resolved in situ grazing incidence x-ray scattering (GISAXS) experiments of Schwartzkopf \textit{et al.} [\onlinecite{schwartzkopf_2015_real}]. Indeed, many of the observed features could be reproduced by the simulations, for film thicknesses up to \SI{3}{nm}. This thickness corresponds to an impressive effective simulation time of \SI{367}{s} which is directly suited for comparison with measurements. The present approach of selective acceleration of dominant processes can be generalized to other systems as well. A recent application concerned the deposition and growth of bi-metallic clusters on a polymer surface [\onlinecite{abraham_cpp18}] where the acceleration allowed one to study the very slow process of demixing of the two metals. Applications of this approach to various plasma processes should be possible as well. In addition to the deposition of neutral atoms, the method also allows one to describe the impact of ions and the growth of charged clusters. \subsection{Dynamical freeze out of dominant modes (DFDM)}\label{ss:freezout} We now discuss an approach that does not accelarate MD, as the ones mentioned above, but takes advantage of the intrinsic hierarchy of relaxation processes existing in any many-body system. When a system is excited, typically, small scale processes and correlation effects will tend to equilibrate fast whereas large scale effects such as particle transport will occur on longer time scales. In between these scales one expects the establishment of equilibrium (or stationary) velocity distributions, for details and more examples, see Refs.~[\onlinecite{bonitz_qkt,bonitz_psst18}]. This means that during the course of the evolution the information and detail required to describe the entire process is systematically reduced. Finally, in thermodynamic equilibrium the system would be completely described by a few macroscopic variables such as temperature and density. This hierarchical character of the evolution is known for a long time and also called ``coarse graining''. It means that a full microscopic description is only needed for the early period of the evolution whereas, for later times, it is sufficient to capture the dynamics of the relevant ``modes'' or degrees of freedom. This has lead to simplified models such as kinetic equations, fluid equations or rate equations where the full N-particle information is mapped onto a limited number of quantities. Even though each of these equations is an approximation to the full many-body equations, these equations are accurate within their respective relaxation stages and time scales. Of course, kinetic equations, fluid models or rate equations contain input parameters (such as collision cross sections or collision integrals, transport coefficients or reaction rates, respectively) that are usually derived by applying suitable approximations to the many-body problem. It is the accuracy of these approximations that governs the accuracy of the model. Consider, as an example, a master (or rate) equation \begin{eqnarray} \frac{dp_i(t)}{dt} &=& \sum\limits_{j\ne i} \left\{ \Gamma_{j\to i}\, p_j(t) - \Gamma_{i\to j}\, p_i(t) \right\},\; \label{eq:master} \\ & &0 \le p_i(t) \le 1, \quad \sum_i p_i(t)=1\,, \nonumber \end{eqnarray} where $i$ is a multi-index numbering the configurations the system can have at some stage of the relaxation which occur with a probability $p_i(t)$. $\Gamma_{i\to j}$ are the transition rates (probability per unit time) from state $i$ to $j$. The first term on the right hand side of Eq.~(\ref{eq:master}) describes processes which increase the probability to realize state $i$ (``gain''), whereas the second term describes the analogous loss processes. Let us now return to our general question about the accuracy of the model (\ref{eq:master}). Such an equation can often be rigorously derived from the underlying classical or quantum equations of motion, and its accuracy is only limited by the accuracy of the transition rates $\Gamma_{i\to j}$ that are typically calculated using various approximation schemes. If, on the other hand, no approximations would be made and a general form be permitted where $\Gamma_{i\to j}\to \Gamma_{i\to j}(t;\{p_k(t)\})$ may depend on time as well as on all $p_i$ for the present and earlier times (memory) this would, in general, result in an exact equation (\ref{eq:master}). In fact, in Ref.~[\onlinecite{bonitz_psst18}] it was proposed to obtain the exact expressions for $\Gamma_{a\to b}$ from first principle MD simulations. Imagine that in the course of the evolution the system reaches a state where only a small number of configurations can be realized the description would much simpler compare to the full N-body dynamics, even if the involved rates are complicated. Moreover, it can be expected that, in many cases, the functional form of the rates $\Gamma_{i\to j}(t;\{p_k(t)\})$ will simplify in the course of the evolution, and they even may converge to stationary values. This procedure was demonstrated in Refs.~[\onlinecite{filinov_psst18_1, filinov_psst18_2}] for a simple example: the adsorption kinetics of argon atoms on a platinum (111) surface. There it was shown that the sticking probability of the atoms can be derived from an equation of the type (\ref{eq:master}) with just three different states. The transition probabilities between these states were computed by first principle MD, and their time dependence was analyzed. It turned out that, in fact, these probabilities converge to constant values, $\Gamma_{i\to j}(t;\{p_k(t)\}) \to \Gamma_{i\to j}^{\rm EQ}$, within approximately $t=t^{\rm EQ}\approx 20$ ps, i.e. as soon as the adsorbate atoms have equilibrated with the surface. Thus, using the first-principle transition probabilities obtained from MD in the master equation (or rate equation), its solution will be essentially exact, for times $t\ge t^{\rm EQ}$. This allows one to extend first principle-type simulations to times long enough to compare with experiments [\onlinecite{bonitz_psst18}] without actually performing MD simulations. The reason is that one was able to identify the dominant collective modes that fully describe the system at long times and emerge dynamically during the evolution. A similar approach was developed by Franke and Pehlke [\onlinecite{franke_2010_diffusion}]. They performed DFT simulations of the diffusion of a 1,4-butanedithiol molecule on a gold surface and also mapped this on a master equation. Even though they did not consider in detail \textit{ab initio} results for the transition probabilities, their results confirm that the type of extension of accurate simulations can also be applied with quantum simulations, instead of semiclassical MD. The idea of mapping the microscopic dynamics onto a small set of relevant degrees of freedom is also utilized in the computation of the secondary electron emission coefficient. This also leads to a system of coupled rate equations as is explained in section~\ref{s:qkin}. \input{QKsee_bpf} \input{plasma-solid-interface-dft-tddft.tex} \section{Nonequilibrium Green functions-based \textit{ab initio} simulations}\label{s:negf} We now return to the statistical approach to quantum many-body systems. We have already seen in Sec.~\ref{s:qkin} that a quantum generalization of kinetic theory allows for an efficient description of plasma-surface interaction processes such as secondary electron emission. The analysis of that section, however, used a simplified treatment of the solid, assuming spatial homogeneity and effective mass models and did not resolve the electronic dynamics in the material. Here we present a generalized quantum kinetic approach where these effects in the solid can be included straightforwardly. This will be demonstrated on the example of ion stopping focusing on recent simulation results of Balzer, Bonitz and co-workers [\onlinecite{balzer_prb16, balzer_prl_18}]. At the same time, the NEGF approach is computationally extremely expensive and presently does not allow for a full quantum-mechanical treatment of the electron dynamics inside the projectile as well. So here an Ehrenfest-type dynamics will be employed, as was discussed already in Sec.~\ref{ss:tddft}. This means that the present model is--in terms of effects included and neglected--complementary to both, the quantum Boltzmann approach presented in Sec.~\ref{s:qkin} and to the DFT concept of Sec.~\ref{s:abinit}. \subsection{Definitions and basic concept of NEGF}\label{sss:negf-defs} The method of nonequilibrium (real-time) Green functions is a successfull approach to quantum many-body systems out of equilibrium, cf. Refs.~\onlinecite{keldysh64, kadanoff-baym-book}. The method is a straightforward generalization of classical kinetic theory (e.g. Boltzmann equation) and of the quantum Boltzmann equation approach overcoming the limitations of the latter. These limitations include the restriction to times larger than the correlation time and fundamental problems such as incorrect conservation laws (e.g. conservation of kinetic energy instead of total energy) and relaxation toward an equilibrium state of an ideal gas (Fermi, Bose or Maxwell distribution) instead of the one of an interacting system, for a detailed discussion, see Refs.~\onlinecite{bonitz_qkt, bonitz-etal.96pla, bonitz-etal.96jpb, bonitz96pla}. Generalized quantum kinetic equations that are based on nonequilibrium Green functions (or, alternatively, on density operator theory [\onlinecite{bonitz_qkt, kremp-etal.97ap}]) overcome these problems. The NEGF approach has been successfully applied to an impressively diverse array of systems, including nuclear matter, by Danielewicz, Köhler and others, e.g. Refs.~\onlinecite{DANIELEWICZ_84_ap2,koehler_prc_95}, to optically excited semiconductors and quantum dots by Schäfer, Haug, Banyai, Bonitz and others, e.g. Refs.~\onlinecite{banyai_prl_95,bonitz-etal.96jpb, kwong-etal.98pss, binder-etal.97prb, bonitz_prb_7, balzer_prb_9}, to dense laser plasmas by Kremp and Bonitz, e.g. Refs.~\onlinecite{kremp_99_pre, bonitz_99_cpp}, to few electron atoms, Refs.~\onlinecite{stefanucci_cambridge_2013, balzer_pra_10, balzer_pra_10_2}, and correlated fermions in lattice systems, Refs.~\onlinecite{verdozzi_jpcs16,schluenzen_prb16} and many other problems. In the present context of plasma-surface interaction we expect that the NEGF approach will allow one to study non-adiabatic effects, in particular, relaxation processes in the surface that are initiated by the impact of plasma particles. The NEGF-method is formulated in second quantization (for textbook or review discussions, see Refs.~\onlinecite{stefanucci_cambridge_2013,kadanoff-baym-book, schluenzen_cpp16}), in terms of creation (annihilation) operators $c_{i\sigma}$ ($c^{\dagger}_{i\sigma}$) for electrons in a single-particle orbital $|i\rangle$ with spin projection $\sigma$. Below we will consider a spatially inhomogeneous lattice configuration where $i$ labels the spatial coordinates of individual lattice points. The creation and annihilation operators are time-dependent via the Heisenberg representation of quantum mechanics. The central quantity that determines all time-dependent observables is the one-particle NEGF (we use $\hbar = 1$), \begin{equation} G_{ij\sigma}(t,t') = -i \langle T_{\cal C}c_{i\sigma}(t)c^{\dagger}_{j\sigma}(t')\rangle\,, \label{eq:negf} \end{equation} where the expectation value is computed with the equilibrium density operator of the system. For completeness we mention that times are running along the Keldysh contour $\cal C$, and $T_{\cal C}$ denotes ordering of operators on $\cal C$ (this is merely a formal trick for the theory development, all practical calculations are done for real-time quantitites, for details see Ref.~ \onlinecite{balzer-book}). For example, the time-dependent electron density on site $i$ follows from $G$ via $n_i(t)= -i G_{ii\sigma}(t, t^+)$, where $t^+ \equiv t+\epsilon$, with $\epsilon>0$ and $\epsilon \to 0$. If the site indices are taken different, $i\ne j$, the Green function describes time-dependent transitions of electrons between two lattice sites. In similar manner one computes the density matrix, currents, mean energies, optical absorption or electrical conductivity from $G$. The NEGF obeys the two-time Keldysh-Kadanoff-Baym equations (KBE)~[\onlinecite{kadanoff-baym-book}] \begin{eqnarray} \label{eq.kbe} \sum_k&\left[i\partial_t\delta_{ik} -h_{ik\sigma}(t)\right]G_{kj\sigma}(t,t')\\ &=\delta_{\cal C}(t-t')\delta_{ij}+\sum_{k}\int_{\cal C} ds\,\Sigma_{ik\sigma}(t,s)G_{kj\sigma}(s,t')\,,\nonumber \end{eqnarray} where $h$ contains kinetic, potential and mean field energy contributions whereas correlation effects are included in the selfenergy $\Sigma$ [here we do not consider spin changes and we omit the second equation which is the adjoint of (\ref{eq.kbe})]. Without the right hand side, Eq.~(\ref{eq.kbe}) would be equivalent to a Vlasov equation or its quantum generalization (time-dependent Hartree-Fock, TDHF). The r.h.s. contains correlation effects that are responsible for relaxation, dissipation and include scattering of electrons with electrons, ions or lattice vibrations (phonons). Notice the time integral on the r.h.s. which incorporates memory effects that are important to correctly treat correlations. The standard Boltzmann equation [Sec.~\ref{s:qkin}] is recovered by evaluating this time integral approximately via a retardation expansion [\onlinecite{bonitz_qkt, bonitz_cpp18}] or by a saddle point method as done in Sec.~\ref{s:qkin}. In this Section we will not consider this long-time limit but concentrate on the fast ion stopping dynamics in a solid. The NEGF formalism is formally exact if the selfenergy would be known exactly. The approach is internally consistent, obeys conservation laws and is applicable to arbitrary length and time scales. Its accuracy is determined by the proper choice for a single function -- the selfenergy. For an overview on the treatment of weak and strong correlations in solids an optical lattices, see Ref.~\onlinecite{schluenzen_cpp16}. In the following we consider, as an example of relevance to plasma-surface interaction, the NEGF approach to the energy loss of energetic ions in a solid. \subsection{Nonequilibrium Green functions approach to ion stopping in nanomaterials}\label{ss:negf-stopping} The energy loss of energetic ions in a solid (stopping power) is an old problem that has been studied in great detail. There exist extensive references and successful code packages such as SRIM [\onlinecite{trim}]. However, standard methods assume linear response, i.e. the material is weakly perturbed by the projectile and the response is computed in perturbation theory. While this maybe correct when considering an extended area of the surface where the effect of the projectile is small, on average, locally the excitation maybe strong. In the following we, therefore, attempt to perform a space and time resolved analysis of the projectile-solid interaction including electronic correlation effects in the solid. This will be particularly important for strongly correlated materials. We consider a surface with strong electronic correlations that is modeled by a Hubbard hamiltonian (\ref{eq:ham1}) with hopping amplitude $J$ [$\langle i,j\rangle$ denotes nearest neighbors] and onsite interaction strength $U$. \begin{eqnarray} \label{eq:ham1} H_\textup{e}&=- J\sum_{\langle i,j\rangle,\sigma} c_{i\sigma}^\dagger c_{j\sigma} + U\sum_{i}\left(n_{i\uparrow}-\frac{1}{2}\right)\left(n_{i\downarrow}-\frac{1}{2}\right)&\nonumber\\ & -\frac{Z_\textup{p}e^2}{4\pi\epsilon_0}\sum_{i,\sigma}\frac{c_{i\sigma}^\dagger c_{i\sigma}}{|\vec{r}_\textup{p}(t)-\vec{R}_i|} + \sum_{\langle i,j\rangle,\sigma}W_{ij}(t)c_{i\sigma}^\dagger c_{j\sigma} \,. \end{eqnarray} \begin{figure}[h] \begin{center} \hspace{-0.cm}\includegraphics[width=0.485\textwidth]{stopping1.pdf} \includegraphics[width=0.485\textwidth]{stopping2.pdf} \end{center} \vspace{-.40cm} \caption{Energy loss of doubly charged helium ions in a graphene sheet~[\onlinecite{balzer_prb16}]. \textbf{(a)} Simulation results for $U/J=1.6$, $J=3.15$\,eV and $\gamma=0.55$ and for different cluster size $L$ are compared with SRIM and TDDFT simulations of Ojanper\"a \textit{et al.} [\onlinecite{ojanpera_prb_14}] for an infinite system. The NEGF data (lines: Hartree approximation, red symbols: second Born selfenergy) show good agreement and, in addition, extend to lower projectile energies. \textbf{(b)} A honeycomb cluster with $L=54$ sites, $U/J=4$ $J=2.8$\,eV, $\gamma=0$, is studied with different selfenergy approximations: second Born, third-order and T-matrix approximation showing a clear impact of correlations.} \label{fig:stopping_vs_impact} \end{figure} The strength of correlations is measured by the ratio $U/J$ and is typically in the range from $0$ to $10$. The second line of Eq.~(\ref{eq:ham1}) contains the coupling of the lattice electrons located at coordinate ${\vec R}_i$ with a positively charged projectile of charge $Z_p$ that is treated classically (Ehrenfest dynamics) by solving Newton's equation for the trajectory $\textbf{r}_p(t)$ under the influence of all Coulomb forces with the lattice electrons. The final term allows to improve the model by accounting for modification of the hopping rates due to the projectile according to $W_{ij}(t)=\gamma [W_{ii}(t)+W_{jj}(t)]/2$, where $W_{ii}$ is the magnitude of the Coulomb potential of the projectile at lattice site ``i'' and $\gamma$ is a phenomenological parameter of the order unity [\onlinecite{balzer_prb16}]. The KBE (\ref{eq.kbe}) with the hamiltonian (\ref{eq:ham1}) have been solved numerically for a two-dimensional hexagonal lattice as is known e.g. for graphene [cf. top part of Fig.~\ref{fig:doublons}] of finite size (the number of lattice sites $L$ varies from $24$ to $384$) for a broad range of projectile velocities~[\onlinecite{balzer_prb16}], see Fig.~\ref{fig:stopping_vs_impact}. For the correlation selfenergy $\Sigma$ several approximations were used: no selfenergy (this corresponds to the mean field or Hartree approximation), second Born (second order in the electron-electron interaction), the third order and the T-matrix (strong coupling) approximation. The results demonstrate good agreement with available data from SRIM [\onlinecite{trim}] and TDDFT [\onlinecite{ojanpera_prb_14}] simulations for high and intermedeate impact energies, down to 10 keV which confirms the choice of the selfenergy. The strength of the NEGF approach is that it also provides data for lower energy, as are typical for low-temperature plasmas. Moreover, the simulations are not restricted to spatially uniform systems but are also directly applicable to finite systems such as graphene nanoclusters, nanostructured surfaces. A particular interesting example are small graphene ``nanoribbons'' that have a size-dependent band gap, e.g. [\onlinecite{son_prl_06, prezzi_prb_08}], giving rise to promising electronic and optical properties. For an overview on plasma synthesis of nanomaterials, see Refs. \onlinecite{meyyappan_jpd_11, ostrikov_aip_13}. In these systems, finite size effects play an important role which is in particular true for the stopping power, as was demonstrated in the top part of Fig.~\ref{fig:stopping_vs_impact}. Finally, the NEGF approach also applies to strongly correlated materials the potential of which for plasma applications has not been explored yet. So far we have not clarified what is the energy loss mechanism for the projectile -- except for the fact that the projectile produces, within the above model, purely electronic excitations (lattice vibrations can be included straightforwardly, but this will not be of interest here, assuming that only times shorter than about 100 fs are considered). In the following we use our time-dependent and space resolved simulations to investigate the electronic correlations effects that are excited by the projectile in a graphene-type nanostructure more in detail. Of particular interest is the local excitation of double occupancies (``doublons''), $d_i(t) =\langle \hat{n}_{i\uparrow}(t)\hat{n}_{i\downarrow}(t)\rangle\,$. These are pairs of electrons (with opposite spin projection) that occupy the same lattice site ``i'' and can be considered bound states, despite their repulsive Coulomb interaction. \begin{figure}[h] \begin{center} \hspace{-0.cm}\includegraphics[width=0.2\textwidth]{lattice.pdf} \hspace{-0.cm}\includegraphics[width=0.485\textwidth]{lattice1.pdf} \hspace{-0.cm}\includegraphics[width=0.485\textwidth]{lattice2.pdf} \end{center} \vspace{-.40cm} \caption{Time-dependent response of a strongly correlated finite honeycomb cluster ($L=12$ sites, $U/J=10$) to a charged projectile penetrating through the center (point C in top figure). \textbf{(a)} site-averaged double occupation, $d_\textup{av}(t)=\frac{1}{L}\sum_i d_i(t)$, for charge $Z=1$ (dashed lines), and $Z=2$ (full lines). \textbf{(b)} The densities on sites A (full line) and B (dashes) closest to the projectile, for the case $Z=2$, after Ref.~[\onlinecite{balzer_prl_18}]. } \label{fig:doublons} \end{figure} If the initial state of the lattice is ``half filled'', i.e. there are as many electrons as lattice sites, this system will behave as an insulator, because any additional electron will have to sit on an already occupied site. This, however, costs an energy $U$, cf. the hamiltonian (\ref{eq:ham1}). An analysis of the energy spectrum reveals that this electron occupies an excited energy band ($U$ above the lower band). If an external electric field would applied, this electron would be mobile, once it has overcome this energy gap. If the system is now hit by an energetic ion, part of this energy could be transformed into excitation of electrons to the upper Hubbard band, i.e. to the formation of doublons and of a conducting state. This hypothesis is tested on an example in Fig.~\ref{fig:doublons} where a projectile penetrates a honeycomb cluster in the center. As a result, in the cluster the electrons are forced towards the center, which is seen by a density increase on the nearest site (B) and a corresponding decrease at the next nearest site (A). Quickly after the projectile has passed through the densities return to their initial values. At the same time, the double occupation (a correlation effect) of the site A, $d_A$, is increased as well, which is due to a transition of the whole electronic system to a higher energy eigenstate. Most importantly, this excitation is stored in the system even after the projectile has left, it only redistributes equally among the sites A and B (cf. the average that is shown by the green curve). The largest effect is observed for energies of 120 eV, in the case of protons (480 eV, for alpha particles) [\onlinecite{balzer_prl_18}]. This way part of the energy of the projectile from the plasma is stored in non-trivial electronic excitations of the solid which might find interesting applications in the future. We note that these simulations are very CPU time consuming. This allows to do simulations only for selected impact parameters (no averaging over trajectories), as in TDDFT simulations [\onlinecite{ojanpera_prb_14}]. Also, long time scales (that describe, e.g. the coupling of electronic and phononic degrees of freedom) are not easily accessible. Also, a full quantum treatment of projectiles with internal degrees of freedom (atoms, larger ions, molecules) will require further developments. In this respect, these NEGF simulations are complementary to the quantum kinetic models of Sec.~\ref{s:qkin} or TDDFT simulations, cf. Sec.~\ref{ss:tddft}. \subsection{\textit{Ab initio} NEGF (AI-NEGF)}\label{ss:ainegf} The NEGF simulations of ion stopping were, so far, restricted to lattice model systems, in order to systematically explore the role of electronic correlations and to investigate finite size effects in nanostructures. In order to extend this approach to a broad range of realistic materials one can couple NEGF simulations to a Kohn-Sham basis that is precomputed by a ground state DFT simulation, as was discussed in Sec.~\ref{s:abinit}. In fact, such an approach was developed a few years ago by Marini and co-workers within their Yambo code [\onlinecite{marini_2009_yambo}] and provides a suitable basis for future improvements of the NEGF simulations towards realistic material properties. The public equilibrium part of the program is well suited to compute the dielectric function, optical properties of materials and was successfully tested in Kiel for graphene [\onlinecite{heese_bsc_17}]. In the following we plan to extend the nonequilibrium part by including stopping simulations as explained in Sec.~\ref{ss:negf-stopping}. \section{Synergies of the surface simulation approaches}\label{s:synergies} Let us now discuss what kind of synergies exist between the various surface simulation methods that were listed in the central box of Fig.~\ref{fig:theory} and were discussed in Secs.~\ref{s:mesoscopic}, \ref{s:qkin}, \ref{s:abinit} and \ref{s:negf}. These methods are very different in terms of accessible length and time scales and the nature of physical approximations, therefore, an efficient and reliable combination into a single single computer code (multi-scale approach) does not seem realistic, at the moment. Nevertheless, it is very interesting to explore what synergies these concepts may possess and how they may be combined for a better understanding of selected relevant surface simulation processes. Of particular interest are, of course, such processes that can be treated with more than one method which offers the potential of cross checks, benchmarks and improvements of approximations. In table~\ref{tab:processes-methods} we list such processes and discuss the potential of the surface simulation approaches. \begin{table*}[] \centering \begin{tabular}{|l|l|l|l|} \hline \textbf{Process} & \textbf{Methods} & \textbf{Pros} & \textbf{Cons}\\\hline Ion stopping & MD &large system & el. adiabatic, unknown accuracy of empirical forces \\ & DFT & accounts for electronic structure & el. adiabatic, small system, limited accuracy of approx.XC\\ &&&kinetic energy limits due to pseudopotentials\\ & TDDFT & electronically non-adiabatic& unknown accuracy of approx. XC\\ &&& only electronic stopping from selected trajectories\\ & NEGF & electronic correlations& finite system, Hubbard-type model \\ && Ref.~\onlinecite{balzer_prb16}, Sec.~\ref{ss:negf-stopping}& classical ion (Ehrenfest dynamics) \\ & AI-NEGF & accounts for electronic structure, & small systems, approximate electronic correlations\\ && possible with Yambo, Sec.~\ref{ss:ainegf} [\onlinecite{marini_2009_yambo}] & not yet tested for plasma-surface applications\\ \hline Ion neutralization & QBE & correlations in projectile, Ref.~\onlinecite{pamperin_2015_many}& model surface\\ & TDDFT & accounts for electronic structure & unknown/limited accuracy of adiabatic approx to XC\\\hline Electron sticking/ & QBE &Ref.~\onlinecite{bronold_2015_absorption}& model surface\\ absorption & && \\\hline SEE & QBE & correlations in projectile, Sec.~\ref{s:qkin} & model surface\\\hline Atom sticking & MD &Refs.~\onlinecite{filinov_psst18_1,filinov_psst18_2}, Sec.~\ref{ss:freezout}& classical, no electronic effects\\ & QBE &Ref.~\onlinecite{brenig_zpb79} & model systems\\\hline Cluster/layer growth & MD & Refs.~\onlinecite{abraham_jap16,abraham_cpp18}, Sec.~\ref{ss:spa}& classical, no electronic effects\\ & KMC & Refs.~\onlinecite{bonitz_cpp12,abraham_jap_15} & phenomenological\\\hline Surface reactions & DFT &accounts for electronic structure & el. adiabatic, limited accuracy of approx.XC\\ &TDDFT &electron. non-adiabatic ~[\onlinecite{brenig_2008_reaction}]& small system, few selected trajectories,\\ &&& unknown/limited accuracy of adiabatic approx. to XC\\\hline Sputtering & MD & large system, Ref.~\onlinecite{brault_frontiers_18}& el. adiabatic, unknown accuracy of empirical forces\\ \hline \end{tabular} \caption{Selection of important plasma-surface processes, main surface science methods (as listed in Fig.~\ref{fig:theory}) as well as their quality and limitations. For applications discussed in this paper, the corresponding section number is given. ``XC'' denotes exchange correlation functional, ``el. adiabatic'': electronically adiabatic.} \label{tab:processes-methods} \end{table*} The table shows that there exist a large range of applications of these methods to quite diverse topics. For method development and improvement the best candidates are topics where several methods have been applied in the past and where, therefore, cross checks are possible. \section{Towards integrated plasma-surface modeling}\label{s:integrated} After reviewing the various simulation methods for surface processes we now discuss possible ways how to integrate them into a simulation of the plasma-solid interface. In the first two parts of this section we consider a quantum kinetic theory approach to charge transfer processes across the interface. First, in Sec.~\ref{ss:intgrated_dl} we consider the description of the quasistationary electric double layer that was introduced in Sec.~\ref{ss:psi-processes} and Fig.~\ref{fig:interface-physics} and derive matching conditions between plasma and surface. After this, in Sec.~\ref{ss:intgrated_ion_impact} we outline a time-dependent approach to electron transfer between plasma ions and a solid, including electronic correlation effects. Finally, at the end of this section we briefly return to the coupling of plasma and surface simulations, in Sec.~\ref{ss:plasma-surface}, and to the issue of how to keep track of surface morphology changes, cf. Sec.~\ref{ss:morphology-surface}. \input{EDL_bpf.tex} \subsection{Integrated modeling of ion-surface interaction}\label{ss:intgrated_ion_impact} In the previous subsection the traditional plasma physics approach that assumes that the processes inside the solid occur on too short (small) time (length) scales to affect the plasma~[\onlinecite{Franklin76}]. However, if the plasma impact is spatially localized and due to rare events (such as the impact of a projectiles) substantial deviations from such an approach have to be expected. It is, therefore, of interest to consider a space and time resolved description which, in principle allows to verify the above hypothesis or, on the opposite, to capture processes that are missed by it. While the former approach can be understood as based on space and time averaging over the scales of the solid -- which is essentially a mean field description, the latter concept takes into account correlation effects and fluctuations around mean values, e.g.~[\onlinecite{lacroix_prb14}]. Here, we develop a quantum kinetic description of the coupled electron-ion dynamics across the plasma interface generalizing the methods described in Sec.~\ref{s:negf}. We start from the second-quantized many-body Hamiltonian for the electrons in the interface and separate the system into a plasma ($p$) and solid surface part ($s$) [we denote $\Omega=\{p,s\}$ and do not write the spin index explicitly], \begin{eqnarray} \label{eq.ham} H_{\textup{interface}} &=&\sum_{\alpha\beta\in\Omega}\sum_{ij}H^{\alpha\beta}_{ij}(t)c^{\alpha\dagger}_ic^\beta_j+ \nonumber\\ &&\frac{1}{2}\sum_{\alpha\beta\gamma\delta\in\Omega}\sum_{ijkl}W^{\alpha\beta\gamma\delta}_{ijkl}c^{\alpha\dagger}_ic^{\beta\dagger}_jc^{\gamma}_kc^{\delta}_l\,. \end{eqnarray} Here, the operator $c^{\alpha\dagger}_i$ ($c^{\alpha}_i$) creates (annihilates) an electron in the state $i$ of part $\alpha$. The one-particle Hamiltonian~$H(t)$ contains the electron's kinetic and the time-dependent potential energy, and $W$ accounts for all possible electron-electron Coulomb interactions within and between the two parts. Considering individual energetic plasma ions, which penetrate into the solid and undergo scattering and stopping in the surface layers, we describe the system~(\ref{eq.ham}) by a one-particle nonequilibrium Green function (NEGF) $G^{\alpha\beta}_{ij}(t,t')$, as introduced in Sec.~\ref{sss:negf-defs}, but here with an additional $2\times2$ matrix structure ($\alpha, \beta=\{p,s\}$), \begin{eqnarray} \label{eq.negf} G^{\alpha\beta}_{ij}(t,t') &=&- i \langle T_C c^\alpha_{i}(t)c_{j}^{\beta\dagger}(t')\rangle\,, \\ \rho_{ij}^{\alpha\beta}(t) &=& - i G^{\beta\alpha}_{ji}(t,t^+)\,, \label{eq:g-dm} \end{eqnarray} e.g., Refs.~[\onlinecite{stefanucci_cambridge_2013, balzer-book}], and the time-diagonal elements provide the density matrix (\ref{eq:g-dm}), as discussed in Sec.~\ref{sss:negf-defs}. The diagonal elements, $\rho_{ij}^{pp}$ [$\rho_{ij}^{ss}$], refer to the plasma part, describing the dynamics of free electrons and electrons bound in the ion [to the solid part, describing electrons in bound states of the solid surface]. Moreover, the density matrix component $\rho_{ij}^{ps}$ is related to charge transfer processes between plasma and solid and will be of special interest in the following. The equations of motion for the NEGF are the generalization of Eq.~(\ref{eq.kbe}) to the interface \begin{eqnarray} i\,\partial_tG^{\alpha\beta}_{ij}(t,t') &-& \sum_{\delta\in\Omega,k} H^{\alpha\delta}_{ik}(t)G^{\delta\beta}_{kj}(t,t')=\delta^{\alpha\beta}_{ij}\delta_C(t,t')+ \nonumber\\ &&\sum_{\delta\in\Omega,k} \int_C\!\!\!d\bar{t}\,\Sigma^{\alpha\delta}_{ik}[W,G](t,\bar{t})G^{\delta\beta}_{kj}(\bar{t},t')\,, \label{eq.kbe_interface} \end{eqnarray} where the self-energy $\Sigma^{\alpha\beta}(t,t')$ describes the interaction between the electrons and with phonons. Even though a complete solution of the KBE~(\ref{eq.kbe_interface}) for real materials and with a full quantum treatment of the plasma electrons is out of reach, these equations provide the rigorous starting point for the development of consistent approximations. In the following we show how it is possible to include the electronic states of the ion via an embedding self-energy approach~[\onlinecite{stefanucci_cambridge_2013}], where resonant (neutralization and ionization) processes can be studied. While this embedding approach is based on a formal decoupling of the surface and plasma parts of the KBE, it retains one-electron charge transfer in the Hamiltonian $H^{sp}$, cf. Eq.~(\ref{eq.sigma.ct}), see below. A closed description of the solid can be maintained if correlations in the plasma part and the feedback of the solid on the plasma can be neglected, i.e., for $\Sigma^{sp}=\Sigma^{pp}=0$. This is usually well fulfilled, except for atmospheric pressure plasmas where small correlation corrections should be taken into account. Then, the KBE~(\ref{eq.kbe}) for the plasma part simplify to \begin{eqnarray} \sum_{k} \left\{i\partial_t\delta_{ik}-H^{pp}_{ik}(t)\right\}g^{pp}_{kj}(t,t')=\delta_{ij}\delta_C(t,t')\,, \end{eqnarray} where, with the solution $g^{pp}(t,t')$ we denote the NEGF of the electrons inside the plasma ion, and the time dependence of $H^{pp}(t)$ accounts for possible parametric changes of the energy levels (e.g., as function of the distance of the ion from the surface). The main result is a closed equation for $G^{ss}(t,t')$: \begin{eqnarray} \label{eq.kbe.embedding} \sum_{k} \left\{i\partial_t\delta_{ik}-H^{ss}_{ik}(t)\right\}G^{ss}_{kj}(t,t')=\delta_{ij}\delta_C(t,t') + \\\nonumber \sum_k\int_C\!\!\!d\bar{t}\,\left\{\Sigma^{ct}_{ik}(t,\bar{t})+\Sigma^{ss}_{ik}[G^{ss}](t,\bar{t})\right\}G^{ss}_{kj}(\bar{t},t')\,, \end{eqnarray} with the charge transfer (or embedding) self-energy that involves the charge transfer hamiltonian \begin{eqnarray} \label{eq.sigma.ct} \Sigma^{ct}_{ij}(t,t') &=& \sum_{kl}H^{sp}_{ik}(t)g_{kl}^{pp}(t,t') H^{ps}_{lj}(t')\,, \\ H^{sp}_{ij}(t) &=& \int\!\!\! d^3r\,\phi^s_i(\vec{r})(T+V)\phi_j^p(\vec{r};t)\,. \label{eq:hsp} \end{eqnarray} Equation~(\ref{eq.kbe.embedding}) shows how the many-body description of an isolated (but correlated) solid is altered by the presence of the electronic states of a plasma ion, with the latter giving rise to an additional self-energy $\Sigma^{\textup{ct}}(t,t')$. While, for $\Sigma^{\textup{ct}}=0$, the KBE~(\ref{eq.kbe.embedding}) conserves the particle number [for a conserving approximation of the self-energy $\Sigma^{ss}$, such as Hartree-Fock, second order Born or GW], the inclusion of the embedding self-energy explicitly allows for time-dependent changes of the particle number in the solid and thus accounts for ion charging and neutralization effects. For the practical solution of Eq.~(\ref{eq.kbe.embedding}), the charge transfer Hamiltonian $H^{sp}(t)$ has to be computed by selecting the relevant electronic transitions between solid and plasma and computing the matrix elements of the kinetic and potential energy operators $T$ and $V$, with the electronic single-particle wave functions $\phi^{s}$ ($\phi^p$) in the solid (ion). To test whether this approach can be applied to the plasma-solid interface we use the same concept as for the computation of the stopping power in Sec.~\ref{ss:negf-stopping}. We use, as a model, a nanostructured solid represented by a chain of 20 lattice sites at half filling (i.e. containing 20 electrons). The charge transfer selfenergy (\ref{eq.sigma.ct}) is parametrized by a time-dependent Gaussian coupling term, $\gamma(t)=\gamma_0\textup{e}^{-(t-t_0)^2/(2\tau^2)}$ that models the approach and elastic scattering of a projcetile. In Figure~\ref{fig:doublons1} we analyze how the charge transfere changes upon variation of the amplitude $\gamma_0$ and interaction duration $\tau$. Interestingly, the charge transfer dependes non-monotonically on the amplitude (see top right part of the figure) and, a large couplings, undergoes strong oscillations. On the other hand, increasing the duration of the interaction (bottom right figure) enhances the charge transfer. With this we have formulated the general NEGF framework for integrated plasma-surface modeling. This approach is largely complementary to the one developed in Sec.~\ref{ss:intgrated_dl}: There the electronic double layer and charge transfer processes along the interface were studied. Here we focused on resonant electron transfer processes attempting an accurate description of the solid. However, the plasma particles were treated on a simplified level. Ultimately, a combination of both approaches will be needed. \begin{figure}[h] \begin{center} \hspace{-0.cm}\includegraphics[width=0.25\textwidth]{embedding.pdf}\hspace{0.1cm} \hspace{-0.cm}\includegraphics[width=0.22\textwidth]{embedding1.pdf}\\[0.1cm] \hspace{-0.cm}\includegraphics[width=0.485\textwidth]{embedding2.pdf} \end{center} \vspace{-.40cm} \caption{Example calculation illustrating the embedding scheme An initially half-filled tight-binding chain ($20$ sites, nearest-neighbor hopping $J$, inverse temperature $\beta_\textup{s}=100J^{-1}$) is coupled via a time-dependent parameter $\gamma(t)=\gamma_0\textup{e}^{-(t-t_0)^2/(2\tau^2)}$ to an external energy level $\epsilon_\textup{i}=+J$ giving rise to the transfer of charge. The initial occupation of the energy level is given by $n_\sigma=0.269$ (corresponding to an inverse temperature of $\beta_\textup{i}=1J^{-1}$) and $\bar{n}_\sigma=1-n_\sigma$. Furthermore, $N_\sigma(t)=\sum_{i}n_{i\sigma}(t)$ denotes the total density on the chain.} \label{fig:doublons1} \end{figure} % \subsection{Coupling plasma and surface simulations}\label{ss:plasma-surface} Returning to the overview given in Fig.~\ref{fig:theory}, the coupling between surface and plasma simulations proceeds via the fluxes of particles, momentum and energy. Fluxes of neutrals, electrons and ions from the plasma, $\textbf{J}^p_a$, have to be provided by plasma simulations and serve as an input for surface simulations. Vice versa, surface simulations, ultimately have to provide the the energy or momentum resolved fluxes, $\textbf{J}^s_a$, of atoms, electrons and ions that leave the surface. In the previous sections we have discussed examples of particle fluxes that are produced by surface science simulations. The first example are the fluxes of neutral gas atoms that are scattered from a metal surface and which are obtained from a combined MD--rate equations model as discussed in Sec.~\ref{ss:freezout}. The result of these simulations is the sticking probability in dependence of impact energy and angle of incidence and lattice temperature, $R_{st}(E_{in},\theta_{in};T_s)$, [\onlinecite{filinov_psst18_2}]. Furthermore, these simulations also yield the energy distribution of the reflected atoms. These results can be directly used in PIC-MCC simulations that trace the dynamics of (fast) neutrals, in addition to electrons and ions. Based on the sticking probabilities, a Monte Carlo procedure can be used to determine if the particle is reflected back into the plasma or if it is adsorbed on the surface. In the former case, the energy distribution of the reflected atoms provides valuable information on the collision process. The second example is secondary electron emission (SEE) which is of key relevance, as we discussed in Sec.~\ref{s:intro}, cf. Fig.~\ref{fig:see}. In this paper (Sec.~\ref{s:qkin}) we have presented quantum kinetic results for the energy resolved SEE coefficient. The present approach is expected to be more accurate and potentially more general than previous models and thus may serve as a valuable input for PIC-MCC simulations. \subsection{Taking into account plasma-induced surface morphology changes}\label{ss:morphology-surface} In Sec.~\ref{ss:spa} simulations of metal cluster growth on a polymer substrate were described. Using selective process acceleration the MD simulations could be extended to several minutes. The resulting plasma-induced adsorbate layer strongly differs from the original ``clean'' surface and can be used to re-compute the electronic structure of bulk and surface states via DFT as indicated in Fig.~\ref{fig:theory}. Furthermore, as was discussed in Sec.~\ref{s:influences}, such a modified (``dirty'') surface reacts very differently on the plasma contact leading, in particular, to strongly enhanced secondary electron emission. This could be investigated in detail, e.g. by means of TDDFT (Sec.~\ref{ss:tddft}) using the pre-computed surface morphology as the starting point. Alternatively, this information can be used as input for quantum kinetic simulations as described in Sec.~\ref{s:qkin} which requires to modify the employed model of the solid. Moreover, the MD simulation of neutral atom scattering, cf. Sec.~\ref{ss:freezout} can be repeated using the plasma-modified surface as an input. This will directly impact the plasma properties via the modified fluxes $\textbf{J}^s_a$, as discussed in Sec.~\ref{ss:plasma-surface}. \section{Conclusion and outlook}\label{s:conclusion1} In this paper we considered the interaction of low-temperature plasmas with a solid surface, discussed the broad variety of physical and chemical processes [Fig.~\ref{fig:interface-physics}] and argued that the mutual interaction between both sides requires to develop closely coupled plasma-surface simulations. We presented an overview on a research poject that is under way at Kiel University, in collaboration with scientists from Greifswald, that aims, aomng others, at developing such simulations. Such an approach has the potential for major advances of this field because most of the currently used models are phenomenological using surface coefficients that are poorly known, both experimentally and theoretically. Moreover, these parameters--even if they exist--may carry an (unknown) dependence on the surface conditions or the plasma parameters. We discussed the simulation approaches that are required to treat the plasma and the solid side of the interface and then listed the methods that are suitable to simulate particle, energy and momentum fluxes across the interface, cf. Fig.~\ref{fig:theory}. We highlighted four groups of methods: semiclassical approaches such as MD and KMC, quantum Boltzmann equation-based models and \textit{ab initio} approaches that are either based on density functional theory or on nonequilibrium Green functions. In Sec.~\ref{s:mesoscopic} we discussed two examples of semiclassical molecular dynamics that use accurate force fields to simulate surface processes involving neutral plasma particles. The main challenge here is to extend these simulations that typcially use a time step on the order of one femtosecond to experimentally relevant times of minutes [\onlinecite{bonitz_psst18}]. This can be achieved via selective process acceleration -- which was demonstrated for the metal cluster growth on a polymer surface. Here synchronized acceleration of metal atom diffusion and deposition allows to extend the simulations by more than nine orders of magnitude up to several minutes [\onlinecite{abraham_jap16}]. The second approach that was discussed was dynamical freezout of collective modes (DFCM) [\onlinecite{filinov_psst18_1, bonitz_psst18}]. Here MD simulations were used to reduce the dynamics of atom desorption on a metal surface to a small set of collective modes that obey coupled rate equations. These rate equations completely describe the sticking behavior at time scales larger than a few tens of picoseconds. There are various ways how to extend the present idea. If the surface is inhomogeneous, a straightforward generalization would be to include the space dependence into the densities and the rates. Then, the rate equations turn into hydrodynamic equations. Furthermore, the effect of a plasma environment, such as characteristic particle fluxes or an adsorbate-covered surface, are straightforwardly included into our scheme, as discussed in Ref.~\onlinecite{filinov_psst18_1}. However, semiclassical MD has a limited sphere of applicabiity. In particular for the description of electrons and ions crossing the interface, semiclassical MD fails, and quantum approaches are necessary. This concerns the neutralization of low-energy ions, e.g. [\onlinecite{pamperin_2015_many}] and their stopping in the solid, as well as the electron dynamics across the interface, e.g. [\onlinecite{bronold_2015_absorption}]. Here nonequlibrium quantum methods such as the quantum Boltzmann equation, density functional theory and nonequlibrium Green functions simulations, e.g. [\onlinecite{zhao_2015_comparison, balzer_prb16, balzer-book}] are the methods of choice which we discussed in Secs.~\ref{s:qkin}, \ref{s:abinit} and \ref{s:negf}. In the present article we attempted at giving a comprehensive overview on these methods regarding their existing applications to and future potential for accurately simulating plasma surface interaction. At the same time, these methods have been developed, so far, almost independently of each other, but we hope that the present work will stimulate future comparisons and combinations. Each of these quantum methods, in particular time-dependent DFT and NEGF, is computationally extremely expensive, and still each of them has significant limitations, for an overview and examples, cf. table~\ref{tab:processes-methods}. Interestingly, TDDFT and NEGF are highly complementary, so it will be important to develop suitable combinations that allow one to overcome bottlenecks. Here we mention a recently proposed hybrid scheme [\onlinecite{hopian_prl_16}], as an example. Another important goal will have to be to use the results of TDDFT and NEGF as input to simpler approaches such as the quantum Boltzmann equation, e.g. Ref. \onlinecite{bonitz_qkt}. Moreover, these simulations can also be extended to longer times using the coupling to a reduced system of rate equations (DFDM) as was explained in the context of MD simulations above. In fact, in Sec.~\ref{s:qkin} for the computation of the secondary electron emission coefficient also a system of rate equations was derived that allows to capture the relevant degrees of freedom (electronic states of the helium projectile). Finally, to properly capture the influence of the plasma on the solid, the above surface simulations have to be linked to fluid or kinetic simulations of the plasma, as indicated by the arrows in Fig.~\ref{fig:theory}. Ultimately, an integrated modeling of the plasma and the solid surface will be required [\onlinecite{interface}] to overcome the trial and error character of many experiments and to achieve a predictive modeling of the relevant processes. In Sec.~\ref{s:integrated} we presented two possible approaches that are based on the quantum Boltzmann equation and nonequilibrium Green functions, within an embedding approach, respectively. Even though these have been rather simple examples and model systems, they indicate the way how such an integrated modeling can be constructed in the future. We expect that our results will not only be of relevance for ``traditional'' materials embedded in a plasma such as metals or semiconductors but also for new materials. Of particular interest, in the near future, could be \textit{nanomaterials}, are such as boron nitride structures or carbon nanotubes or graphene sheets and nanoribbons. For these nanomaterials both finitie size effects and electronic correlations will be of particular importance [\onlinecite{son_prl_06, prezzi_prb_08}] and the array of methods oulined in this article should be, in their combination, suitable to describe plasma-surface interaction with such exciting novel materials. \section*{Acknowledgements} The authors acknowledge many fruitful discussions with our colleagues in Kiel -- in particular, M. Bauer, J. Benedikt, J. Golda, B. Hartke, H. Kersten, O. Magnussen, K. Rossnagel, J. Stettner, and T. Trottenberg -- with whom the present concept has been developed. EP is grateful to L. Deuchler for insightful discussions concerning TDDFT-MD simulations. MB and KB are grateful to A. Marini and D. Sangalli for their support in using their \textit{ab initio} NEGF-code \textit{Yambo}. MB is grateful to K. Becker, A. Bogaerts, P. Bruggeman, Z. Donko, J.G. Eden, U. Fantz, I. Kaganovich, M. Kushner, E. Neyts, G. Oehrlein, K. Ostrikov, and Y. Raitses for many stimulating remarks during presentation of early versions of this work. \bibliographystyle{apsrev} \section{DFT-based {\it ab initio} simulations}\label{s:abinit} \subsection{Born-Oppenheimer MD}\label{ss:bomd} Density functional theory (DFT) based methods constitute the most widely applied approach to calculate the electronic structure in physics and chemistry. It can be combined with Born-Oppenheimer Molecular Dynamics (BOMD) to simulate the dynamics of atoms within molecules and solids [\onlinecite{Marx2009}]. {\it Ab initio} BOMD has been applied successfully by many groups, see e.g. Refs.\ \onlinecite{hafner_2008_initio,Gross_O2Pt_2016,Kuehne2014}. The BOMD approach starts from the Born-Oppenheimer (BO) approximation for the coupled system of nuclei and electrons [\onlinecite{Baer2006}], where all non-adiabatic coupling terms between the different BO surfaces of the electronically excited states are neglected. Furthermore, the classical limit for the motion of the nuclei is assumed, ignoring geometrical phases occurring around conical intersections [\onlinecite{Baer2006}]. The equation of motion (EOM) reduces to a set of coupled Newton's equations. Like in semi-classical MD described in Sect.\ IV, in case of an ensemble of $N$ atoms the micro-canonical MD simulation ``boils down'' to the integration of \begin{equation} M_i \ddot {\bf R}_i = {\bf F}_i = -\nabla_{{\bf R}_i} U_0({\bf R}_1, ..., {\bf R}_N)\, \label{eq:BOMD_Newton_EOM} \end{equation} starting from some suitable initial conditions. $U_0({\bf R}_1, ..., {\bf R}_N)$ denotes the ground state potential energy surface (PES) of the system. Canonical ensembles can be simulated by coupling to a Nos\'{e}-Hoover thermostat [\onlinecite{Nose1984}] or Nos\'{e}-Hoover chain thermostat [\onlinecite{Marx2009}]. Alternatively, the Langevin scheme can be used as detailed in Sect.\ IV. As already pointed out, rare events pose a serious problem to any MD scheme, and even more so to the very CPU-time intensive {\it ab initio} based MD programs. Thus, for the determination of reaction and diffusion energy barriers, algorithms like the nudged elastic band method (NEB, CI-NEB) [\onlinecite{JCP:H00_TAN,JCP:H00_CI}] are in use. Together with calculated vibrational frequencies, reaction and diffusion rates can be estimated within transition state theory (TST) [\onlinecite{JPCS:VIN57}]. The rates can be entered into a Master equation approach or kinetic Monte Carlo simulations as described in Sec.\ref{s:mesoscopic}. In addition, Meta-dynamics approaches, as described by Laio and Parrinello [\onlinecite{LaioParrinello2002,Laio2008}] can help to escape local or global potential energy minima and explore further regions of the PES. A general difficulty when applying Meta-dynamics is the judicious choice of few collective variables for the simulation. While not without open basic problems, total energy calculations based on density functional theory (DFT) for the electronic ground state -- together with a suitable approximation to the exchange-correlation energy functional -- constitutes today's standard approach in electronic structure theory -- both in physics and in chemistry [\onlinecite{Martin2004,Marx2009,Burke2012,Becke_Perspective_2014,Yu_Perspective_2016}]. This is in general ascribed to, in case of many applications, a favorable balance between the accuracy of the calculated total energy differences and forces acting on the ions as required for MD on the one hand and the (nonetheless still very large) computational costs on the other hand. DFT describes the exact mapping of the electronic ground state of the interacting many-particle Hamiltonian onto the ground state of a system of fictitious spin-1/2 particles not interacting with each other, which are moving in an effective potential that itself depends on the electron density [\onlinecite{Hohenberg1964, Kohn1965}]. The crucial point for practical applications is the required approximation to the exchange-correlation (XC) energy functional. A careful choice of this approximation for the problem in question is essential for the accuracy of the obtained results. Various approximations are available. A universally applicable approximation to the XC energy functional giving chemical accuracy is lacking -- this constitutes a field of active current research [\onlinecite{Burke2012,Becke_Perspective_2014,Yu_Perspective_2016,Klimes2012}]. In many practical computations one of the generalized gradient approximations GGA-PW91 [\onlinecite{Perdew1992_PW91}] or GGA-PBE [\onlinecite{Perdew1996_PBE}] are applied. Other approximations comprise meta-GGAs [\onlinecite{Tao2003}], hybrid functionals, and many other approaches, see {\it e.g.} Refs. \onlinecite{Burke2012,Becke_Perspective_2014,Yu_Perspective_2016} for reviews. In view of MD simulations relevant for plasma physics of, {\it e.g.}, noble gas atoms interacting with a metal surface, van der Waals forces (which cannot be represented by semi-local XC functionals like the local density approximation, LDA, or the GGAs) have to be considered [\onlinecite{Klimes2012,filinov_psst18_1}]. Furthermore, as soon as open-shell atoms or other spin-polarized systems come into play, Spin-DFT has to be used [\onlinecite{Martin2004}]. Widely used DFT total-energy computer programs ({\it e.g.} VASP [\onlinecite{CMS:VASP96,PRB:VASP93,PRB:VASP94,PRB:VASP96}], the Quantum Espresso package [\onlinecite{Giannozzi_QE_2009, Giannozzi_QE_2017}] and various other programs) employ super-cells, which are repeated periodically in all three spatial directions, to simulate surfaces, clusters, etc., a plane-wave expansion of the Kohn-Sham orbitals and special {\bf k}-points for approximate Brillouin zone integration [\onlinecite{Mattsson2005}]. The electron--ion interaction is represented by pseudopotentials [\onlinecite{Martin2004}] ({\it e.g.} norm-conserving [\onlinecite{Hamann1989,Trouiller1991,Fuchs1999}], ultrasoft [\onlinecite{Vanderbilt1990}], projector augmented wave (PAW) [\onlinecite{PRB:PAW94,PRB:PAW99}] pseudopotentials). In this way only the valence band states have to be computed, while the effect of the (frozen) core electrons is accounted for by the pseudopotentials. Again, special attention is required in case of MD simulations for plasma--surface interactions: If the kinetic energy of a projectile is so large that the inter-atomic separations become small during a collision, the pseudopotential approximation may become invalid and the interaction between the core electrons may have to be accounted for explicitly. \subsection{Time-dependent DFT}\label{ss:tddft} An ion from a plasma interacting with a solid surface [\onlinecite{Modinos1987,Brako1989,Kimmel_1993_alkali_ion_Cu_resonant_charge_transfer,Winter_PhysRep_2002,Race_RepProgPhys_2010,wucher_2011_kinetic}] results in inherently electronically non-adiabatic dynamics -- which is beyond the realm of electronic ground state theory. Already the initial state, corresponding to an ion far away from the surface, is electronically strongly excited, although the individual parts of the system, {\it i.e.} the solid and the projectile, may initially be in their respective electronic ground state. Also the scattering of faster than thermal atoms at a metal surface can result in electronic energy dissipation, the description of which requires a simulation that accounts for electronically nonadiabatic effects (for an example see Ref.\ \onlinecite{Lindenblatt2006b} for 2 -- 10 eV H atoms impinging on an Al(111) surface). However, the deviation from the BO-surface is distinctly smaller in this case. In surface chemistry, electronically non-adiabatic effects have been observed experimentally (see., {\it e.g.} Refs.~\onlinecite{Nienhaus_2002_chemicurrents,Diesing2016,Buenermann1346_2015,Wodtke2016}) and described theoretically by using various approaches including {\it e.g.} electronic friction [\onlinecite{Rittmeyer_2015_LDFA,Rittmeyer2018,Alducin2017,Janke2015,Kroes2014,Monturet2010}], model Hamiltonians [\onlinecite{Mizielinski2005,Mizielinski2007,Mizielinski2008,Mizielinski2010,Bird2008}], or time-dependent density functional theory (TDDFT) [\onlinecite{Lindenblatt2006,Lindenblatt2006a,Lindenblatt2006PRL,Grotemeyer2014,Timmer_2009_H_Al111_perturbation}]. Furthermore, also ion-atom collisions [\onlinecite{Zhao_2015_H+Be}] and ions impinging onto a metal cluster [\onlinecite{Moss_2009_Li+_AlCluster,Castro2012_H+Li4}], colliding with carbon nanostructures [\onlinecite{Krasheninnikov2007}], graphene fragments [\onlinecite{Bubin2012}], graphene or boron nitride (BN) [\onlinecite{zhao_2015_comparison,ojanpera_prb_14}], the collision of Cl, Cl$^-$ with a MoSe$_2$ monolayer [\onlinecite{Wang2015}] as well as the electronic stopping of atoms or ions moving through a solid [\onlinecite{Yost2017,Schleife_2015_H_He_Al,Correa_2012_H_in_Al,Zeb2012,Ullah2015}] have been simulated using TDDFT. TDDFT [\onlinecite{Marques_TDDFT_I,Marques_TDDFT_II,Ullrich_TDDFT,Ullrich2014,Maitra_Perspective_2016}] is based on the Runge--Gross theorem [\onlinecite{Runge1984}]. Under certain restrictions to the single particle potential, and given a fixed initial state $|\Psi(t_0)\rangle$, the electron density $n({\bf r}',t')$, $t' \in [t_0,t]$ of a finite interacting many-particle system determines the time-dependent single particle potential $v({\bf r},t)+c(t)$ apart from an arbitrary function $c(t)$ of the time [\onlinecite{Marques_TDDFT_I,Maitra_Perspective_2016}]. The interacting system can be mapped onto a system of fictitious spin-1/2-fermions with the same density not interacting with each other but moving in a time dependent effective Kohn-Sham potential [\onlinecite{Gross_Kohn_1985,Marques_TDDFT_I}] \begin{eqnarray} v_{\rm KS}({\bf r},t) &=& v({\bf r},t) + \int d^3{\bf r}' \frac{n({\bf r}',t)}{|{\bf r}-{\bf r}'|}\ + \nonumber \\[0.4cm] & & v_{\rm XC}([n],[\Phi_0],[\Psi_0])({\bf r},t) . \end{eqnarray} The XC-potential involves memory, it depends on the charge density history and the initial states $\Psi_0$ and $\Phi_0$ of the interacting and the Kohn-Sham system \cite{Marques_TDDFT_I,Marques_TDDFT_II}. The crucial step, which is limiting the range of applicability of the respective approximate approach [\onlinecite{Provorse2016}], is the approximation applied to the time-dependent XC-potential. Today, in most cases the adiabatic approximation is used (together with an approximation to the ground-state XC potential), {\it i.e.} the instantaneous electron density $n({\bf r},t)$ is inserted into an approximate local ({\it e.g.} LDA or GGA) XC-potential $v_{\rm XC}^{\rm approx}$ from ground state DFT: \begin{equation} v_{\rm XC}[n]({\bf r},t) \approx v_{\rm XC}^{\rm approx}[n(\cdot,t)]({\bf r}) . \end{equation} In the adiabatic approximation memory and initial state dependence are neglected. A critical discussion of these issues can be found in the article by N.T.\ Maitra [\onlinecite{Maitra_Perspective_2016}] and the references cited therein and the article by Provorse and Isborn [\onlinecite{Provorse2016}]. For the effect of different approximations to the ground-state XC-potential see {\it e.g.} Refs.\ \onlinecite{Zhao_2015_H+Be,Nagano2000}. The effect of the approximate XC-potential {\it e.g.} on resonant charge transfer requires careful consideration, and comparison of time-dependent simulations of the many-body system within TDDFT-MD with results obtained from a many-particle NEGF-approach, as described in Sec.~\ref{s:negf}, could provide additional insights to judge the accuracy. The application of time-dependent current density functional theory (TDCDFT) to stopping power and the effect of non-locality is discussed by V.U.\ Nazarov {\it et al.} in Ref. \onlinecite{Nazarov2007}. Another approximation is required for the description of the motion of the ions. The complete equations of motion for the combined system of nuclei and electrons can be written as a coupled system of equations, where the nuclei are moving on the ground and excited state Born Oppenheimer surfaces, depending on the electronic excitation [\onlinecite{Baer2006}]. Tully's surface hopping algorithm [\onlinecite{Tully1990,Shenvi_2009}] allows to simulate such time evolution stochastically. However, as long as the trajectories of the nuclei do not differ qualitatively but, instead, closely follow some average trajectory, the motion of the nuclei can be approximated by means of the much less expensive Ehrenfest dynamics [\onlinecite{Marx2009}], \begin{eqnarray} M_i \frac{d^2 {\bf R}_i} {dt^2} = -\langle \nabla_{{\bf R}_i} \hat V_{\rm ion - el}({\bf R}_1, ..., {\bf R}_n ) \rangle \nonumber\\ - \nabla_{{\bf R}_i} V_{\rm ion - ion}({\bf R}_1, ..., {\bf R}_n ) . \end{eqnarray} The expectation value in above equation has to be calculated with respect to the electronic wave function, which follows, in principle, from the integration of the time-dependent many-particle Schr\"odinger equation of the electrons, or, in practical applications, the time-dependent Kohn-Sham equations. $M_i$ denotes the ion masses, ${\bf R}_i(t)$ the ion positions, and $V_{\rm ion - ion}$ and $V_{\rm ion - el}$ the ion--ion and ion--electron interaction potential energies. Pseudopotentials [\onlinecite{Martin2004}] are in use in TDDFT-Ehrenfest MD simulations [\onlinecite{Marques_octopus_2003}]. This enormously reduces the computational effort as compared to an all-electron approach [\onlinecite{Foglia2017}]. Galilei invariance is preserved if non-local pseudopotentials are multiplied by an ion-velocity dependent gauge factor [\onlinecite{Nagano2000}]. In case the projectile has some finite initial velocity, an initial boost has to be applied to electronic states of the projectile [\onlinecite{AvendanoFranko_Thesis_2013,Nagano2000}] so that the electrons have the same initial velocity as the nuclei. The accuracy of TDDFT-MD simulations has been critically evaluated recently by Yost {\it et al.} [\onlinecite{Yost2017}] for the case of the electronic stopping power of a proton in Si. They report deviations between the all-electron and the pseudopotential approach and suggest a correction scheme. The differences arise at higher proton velocity (beyond 1 a.u.), and the correction is larger at small impact parameter [\onlinecite{Yost2017}]. A pseudoatom method to account for core electron effects has been applied to electronic stopping of {\it e.g.} Li$^+$ ions in graphene for impact energies beyond 10 keV/u in Ref.\ \onlinecite{ojanpera_prb_14}. As the deviation from the all-electron case will depend on the binding energy of the electronic states assumed as frozen core states in the pseudopotential construction, one should care about electron promotion effects missing in the pseudopotential calculation [\onlinecite{German1994}] when carrying through TDDFT-MD simulations including high-energy head-on collisions. TDDFT-Ehrenfest molecular dynamics is implemented, e.g., in the program \textit{octopus} by Rubio {\it et al.} [\onlinecite{Marques_octopus_2003,Castro_octopus_2006,Andrade_octopus_2015}]. In their program, the Kohn-Sham wavefunctions are represented on a real space grid and systems of different dimensionality can be treated. Applications of TDDFT-MD by many authors using different codes can be found in the literature. Examples have already been cited above, including electron stopping of H in Al, see {\it e.g.} Refs.\ \onlinecite{Correa_2012_H_in_Al,Schleife2015,Shukri_2016_electronic-stopping}, the interaction of Li$^+$ with an Al cluster [\onlinecite{Moss_2009_Li+_AlCluster}], the interaction of H/Al in Ref.\ \onlinecite{Lindenblatt2006b}. The H/He anomaly in the electronic stopping power in Au at low kinetic energies [\onlinecite{Markin2009}] and the role of the Au $d$ electrons has been simulated by the authors of Ref.\ \onlinecite{Zeb2012}. These applications refer either to electron stopping in bulk materials or, in case of the ion-discharge processes at surfaces, to resonant charge transfer, which is an inherently single-electron effect -- as opposed to Auger neutralization, which is a two-electron effect [for a discussion, see Sec.~\ref{s:qkin}]). It has been pointed out in the literature before that the energy transfer into electronic excitations can be simulated only for a finite time in case of a finite system due to its discrete energy spectrum [\onlinecite{Mason2007, Race_RepProgPhys_2010}]. Thus, {\it e.g.} in TDDFT-MD simulations of a vibrating molecule interacting with a metal surface [\onlinecite{Grotemeyer2014,Grotemeyer_Dissertation}] the size of the unit cell (the number of atomic layers forming the metal slab or the cluster size) limits the simulation time. Within a simplified tight binding model for the case of a vibrationally excited HCl molecule incident on an Al(111) slab this has been exemplified by M. Grotemeyer [\onlinecite{Grotemeyer_Dissertation}]. A direct simulation of open quantum systems, see {\it e.g.} the article by R. D{'}Agosta and M. Di Ventra [\onlinecite{Agosta2013}] and the article by H. Appel in Ref.~\onlinecite{Marques_TDDFT_II}, would be advantageous. For Ar$^{8+}$--Ar charge transfer collisions Nagano {\it et al.} [\onlinecite{Nagano2000}] observed that no Auger excited electrons occurred in their TDLDA-simulations. Furthermore, as shown by C.A. Ullrich for a model system in Ref. \onlinecite{Ullrich2006}, when applying the ALDA for the XC potential in TDDFT doubly excited configurations are not accounted for. Based on the calculation of an exact XC-potential, V. Kapoor [\onlinecite{Kapoor2016}] has concluded that autoionization requires inclusion of memory effects in the XC-potential, which displays a rather complicated structure. When calculating {\it e.g.} the charge transfer between He ions scattered at an Al-surface, the two-particle character of the Auger transitions leading to the charge transfer between projectile and surface has to be accounted for, and other techniques are therefore used in case of Auger transitions [\onlinecite{Lorente_1994_Auger_neutralization,Monreal_Review_2014}]. To summarize, DFT and TDDFT simulation methods have undergone a very strong development since the formulation of the underlying theorems. We have discussed a number of practical applications and important approximations. However, we are not aware of direct applications to plasma-surface interaction yet. Of course, there have been many simulations of particular processes such as ion stopping, ion neutralization or neutral adsorption, but the influence of a nonequilibrium plasma environment is still an open question. DFT and TDDFT will, without doubt, play a key role in plasma-surface simulations. Their strength will be mostly in testing individual processes, rather than simulating the full dynamics. This knowledge can then be used in other simulations such as those using quantum kinetic equations, cf. Secs.~\ref{s:qkin} and \ref{s:negf}, e.g. as input for models describing the dynamics of electrons in ions, atoms or molecules (such as the Newns-Anderson model used in Sec.~\ref{s:qkin}), as well as for models of solids, such as lattice models used in Secs.~\ref{s:negf} and ~\ref{ss:intgrated_ion_impact}. \section{Quantum Boltzmann equation}\label{s:qkin} In the previous section the focus was on semiclassical or even classical methods for describing the interaction of atoms with surfaces, based on Newton'a equations of motion for the constituents involved and their solution by molecular dynamics techniques. Not all surface scattering processes can be treated in that manner. In particular, charge-transferring processes require a quantum-mechanical approach. The most prominent scattering process (cf. Sec.~\ref{s:influences}), which has to be treated quantum-mechanically and, at the same time, has great relevance for plasma modeling, is secondary electron emission from surfaces/plasma walls due to impacting heavy species. \subsection{Quantum Kinetic Approach to secondary electron emission. Generalized Newns-Anderson model}\label{ss:qke-see} It has been known for a long time that secondary electron emission is an important process in bounded plasmas, affecting the structure of the plasma sheath, the overall charge balance, and the operation modii of basically all types of low-temperature discharges~[\onlinecite{LL05}]. To quantify the process is, however, a rather challenging task. Experimentally it requires sophisticated instrumentation and theoretically it asks for the solution of a scattering problem involving many-body targets and projectiles~[\onlinecite{Rabalais94,Winter02,Winter07}]. It is thus not surprising that little is known quantitatively about secondary electron emission from plasma-exposed surfaces which, moreover, are usually also insufficiently well characterized. The collection and discussion of secondary electron emission data by Petrovi\'c and Phelps~[\onlinecite{PP99}] is still the main reference, cf. Sec.~\ref{s:influences}. Only recently the plasma physics community initiated a number of new investigations devoted to the issue~[\onlinecite{DBS16,MCA15}]. \begin{figure*}[t] \includegraphics[width=\linewidth]{AugerNeutralCartoon.pdf} \caption{(Color online) Illustration of the main ingredients of an Anderson-News model based quantum-kinetic analysis of the neutralization of a helium ion on an aluminum surface characterized by a step potential with depth $E_{\rm F}+\Phi$, where $E_{\rm F}$ is the Fermi energy and $\Phi$ the work function of aluminum. The collision trajectory leading to time-dependent matrix elements enforcing a quantum-kinetic analysis is shown in (a) and the reaction channels included in the modeling are indicated in the (on scale) energy diagram (b) and the reaction diagram (c). For simplicity, the projectile levels shown are the ones far away from the surface, level shifts and broadening due to the interaction with the surface are not visualized. As indicated in (c), there are three routes to the projectile groundstate, each one leading to the emission of an electron. The helium ion may capture an electron from the metal by a single-electron transfer (SET), changing its configuration from $\mathrm{He^+(1s,1^2S_{1/2})}$ to either $\mathrm{He^\ast(1s2s,2^3S_1)}$ or $\mathrm{He^\ast(1s2s,2^1S_0)}$, which may either Auger de-excite (AD) to $\mathrm{He^0(1s^2,1^1S_0)}$ or attract another electron from the metal to form a $\mathrm{He^{\ast-}(1s2s^2,2^2S_{1/2})}$ ion releasing an electron then by auto-detachment (AuD). In addition to these two routes the $\mathrm{He^0(1s^2,1^1S_0)}$ configuration may be also reached by Auger neutralization (AN) of the $\mathrm{He^+(1s,1^2S_{1/2})}$ ion setting also free an electron. Panel (d) finally depicts the three-level system which can be employed to represent the helium configurations taking part in the collision. Depending on the process the levels act as ionization or affinity levels. Auxiliary bosons are used to assign the functionality needed to the levels. Not included in the three level modeling, since it is unaffected by the collision, is the spin-up electron in the 1s-shell common to all configurations listed in (d). For details see Ref.~\onlinecite{PBF18}.} \label{HeNeutralProcesses} \end{figure*} In the following we focus on electron emission at low impact energies where the atomic projectile remains outside the surface and emission is driven by the transfer of internal potential energy from the projectile to the target. In principle there are two theoretical approaches to the problem. The first approach attempts to describe the processes from first principles, using density functional theory or quantum-chemical techniques. There exist various facets of this approach (see, for instance, the work of More and coworkers~[\onlinecite{MMM98}] and the monograph~[\onlinecite{Winter07}]) depending on how many electronic and lattice degrees of freedom are kept at an ab-initio level. Ideally, one keeps all. But this sophistication can be hardly maintained in a realistic description of plasma-surface interaction. The second approach, more modest in theoretical detail and to be followed in this subsection, is to keep only the most important degrees of freedom taking part in the scattering process and to construct effective models for them. Typically these are Anderson-Newns models~[\onlinecite{NMB83,YM86,LG90}] where the collision trajectory of a projectile is prescribed externally, leading to a time-dependent Hamiltonian for the participating electrons. The matrix elements of the Hamiltonian can be, of course, calculated from first principles but it is more in the spirit of an effective modeling to parametrize the matrix elements by physical considerations. The resulting time-dependent Hamiltonian is then fed into a quantum-kinetic treatment, based on nonequilibrium Green functions (NEGF), which, if combined with pseudo-operator techniques, has the advantage that all collision channels, which may open-up when the projectile approaches the target surface, can be treated simultaneously. The NEGF approach is the same as listed in the fourth line of the pink box in Fig.~\ref{fig:theory} and will be discussed in some detail in Sec.~\ref{s:negf}. Here we will concentrate on the main steps that are involved in deriving the quantum Boltzmann equation from the NEGF equations and on the derivation of the rate equations model for the relevant degrees of freedom. The use of Anderson-Newns models for the description of charge-transferring atom-surface collisions is well established and has a long history, recent applications~[\onlinecite{WGM01,VGB05,MBF12a,MBF12b,IGG13,IGG14,PBF15a,PBF15b,PBF18}] differ only in the way the matrix elements are obtained and the number of channels included. Both of course depend on the scattering process to be modelled. It is beyond the scope of this paper to review it here. Instead, we describe the Anderson-News type modeling developed at the University Greifswald~[\onlinecite{MBF12a,MBF12b,PBF15a,PBF15b}] taking a helium ion hitting an aluminum surface as an example. The description is necessarily sparse, details and data for other metals can be found in the work of Pamperin, Bronold, and Fehske~[\onlinecite{PBF18}] to which we refer as PBF. Figure~\ref{HeNeutralProcesses} summarizes the approach as applied to the collision of a positive helium ion with an aluminum surface. In accordance with the general spirit of the Anderson-Newns model~[\onlinecite{NMB83,YM86,LG90}] the center-of-mass motion of the ion, for simplicity assumed to be normal to surface, is put on a prescribed trajectory. Due to the large mass of the projectile, the motion is indeed to a good approximation classical. It leads however to time-dependent single- and two-particle matrix elements and in turn to the necessity of using quantum kinetics to extract from the model the secondary electron emission coefficient $\gamma(\varepsilon_{\vec{q}})$, where $\varepsilon_{\vec{q}}$ is the energy of the emitted electron. The electronic degrees of freedom of the projectile and the target are treated quantum mechanically, using a step potential of depth $E_F+\Phi$ to mimic the aluminum surface, with $E_F$ the Fermi energy and $\Phi$ the work function, and a few-level system to emulate the electronic states of the projectile affected by the collision, that is, the ionization and affinity levels which may take part in an electron transfer. As indicated, there will be typically more than one ionization and affinity levels involved. As a result, electron emission can occur via many channels. In the case depicted there are three main channels: Auger neutralization (AN), direct (DAD) and indirect (IAD) Auger de-excitation, and auto-detachment (AuD). All three can be included in the present NEGF approach and the quantum Boltzmann equation derived from it. It is sensible to parametrize the model Hamiltonian such that, far away from the surface, the energy levels of the projectile coincide with the experimental ionization energies and electron affinities of the isolated helium projectile. As it comes closer to the surface it starts to interact with it leading to level shifts and broadening (not shown in the figure), as well as to electron transfer due to Auger- and single-electron processes. For the results presented below the matrix elements of the model Hamiltonian were obtained from mean-field wave functions and physical considerations, taking image charges~[\onlinecite{Gadzuk67a,Gadzuk67b}] as well as tunneling through potential barriers~[\onlinecite{Propst63,PA90}] into account in case they arise. Details are given in PBF~[\onlinecite{PBF18}]. If instead of the physics-guided manner the matrix elements are obtained from first-principles, the quantum-kinetic treatment of the Hamiltonian described below remains the same. The Hamiltonian describing the physics sketched in Fig.~\ref{HeNeutralProcesses} is best written down in the notation of second quantization, that is, in terms of annihilation and creation operators [cf. Sec.~\ref{s:negf}]. Distinguishing between electron states belonging to the step potential, the few level system, and the unbound continuum, three types of Fermi operators are introduced: $c_{\vec{k}\sigma}^{(\dagger)}$, $c_{\vec{q}\sigma}^{(\dagger)}$, $c_n^{(\dagger)}$ annihilating (creating) an electron with spin $\sigma$ in, respectively, a surface state $|\vec{k}\rangle$, a continuum state $|\vec{q}\rangle$, and a projectile state $|n\rangle$. From a calculational point of view it is advantageous to replace the operators $c_n^{(\dagger)}$ by pseudo-operators, $e^{(\dagger)}$, $d^{(\dagger)}$, and $s_{n\sigma}^{(\dagger)}$ defined by \begin{align} \label{allPseudoStates} \vert000\rangle &= e^\dagger\vert {\rm vac}\rangle \,,\, \vert011\rangle = d^\dagger\vert {\rm vac}\rangle \,, \vert100\rangle = s^\dagger_{1\downarrow}\vert {\rm vac}\rangle \,,\, \nonumber\\ \vert010\rangle &= s^\dagger_{2\uparrow}\vert {\rm vac}\rangle \,,\, \vert001\rangle = s^\dagger_{2\downarrow}\vert {\rm vac}\rangle \,. \end{align} They stand for whole electronic configurations of the projectile and not for single electron states. The number of electrons required for the electronic configurations represented by the pseudo-operators determines their statistics. Is the number odd (even) the operators obey Fermi (Bose) statistics. The labeling of the pseudo-operator is a reminder that the projectile's configurations involve the 1s and 2s shell of helium. In spectroscopic terms, the configurations included are the $\mathrm{He^+(1s,1^2S_{1/2})}$ positive ion, the $\mathrm{He^0(1s^2,1^1S_0)}$ groundstate, the $\mathrm{He^\ast(1s2s,2^3S_1)}$ triplet and $\mathrm{He^\ast(1s2s,2^1S_0)}$ singlet metastable states, and the $\mathrm{He^{\ast-}(1s2s^2,2^2S_{1/2})}$ negative ion. Employing the reasoning developed in~[\onlinecite{MBF12b,PBF15b,PBF15a}], the Hamiltonian describing the neutralization of a $\mathrm{He^+(1s,1^2S_{1/2})}$ ion on an aluminum (or any other metal) surface within the scenario summarized in Fig.~\ref{HeNeutralProcesses} becomes~[\onlinecite{PBF18}] \begin{widetext} \begin{align} \label{pseudoHamiltonian} H(t) &= \varepsilon^0_{1s\downarrow}(t) s^\dagger_{1\downarrow} s^{\phantom{\dag}}_{1\downarrow} + \sum_\sigma \varepsilon^*_{2s\sigma}(t) s^\dagger_{2\sigma} s^{\phantom{\dag}}_{2\sigma} + \big[\varepsilon^-_{2s\uparrow}(t) + \varepsilon^-_{2s\downarrow}(t)\big] d^\dagger d^{\phantom{\dag}} + \sum_\sigma \omega_\sigma(t) b^\dagger_\sigma b^{\phantom{\dag}}_\sigma + \sum_{\vec{k} \sigma} \varepsilon_{\vec{k} \sigma} c^\dagger_{\vec{k} \sigma} c^{\phantom{\dag}}_{\vec{k} \sigma} \nonumber\\ &+ \sum_{\vec{q} \sigma} \varepsilon_{\vec{q}\sigma}(t) c^\dagger_{\vec{q} \sigma} c^{\phantom{\dag}}_{\vec{q} \sigma} + \sum_{\vec{k} \sigma} \big[ V^{\rm SET}_{\vec{k}\sigma}(t) c^\dagger_{\vec{k} \sigma} e^\dagger s^{\phantom{\dag}}_{2\sigma} + \text{H.c.} \big] - \sum_{\vec{k} \sigma} \big[ \sgn(\sigma) V^{\rm SET}_{\vec{k}\sigma}(t) c^\dagger_{\vec{k} \sigma} b^\dagger_\sigma s^\dagger_{2-\sigma} d^{\phantom{\dag}} + \text{H.c.} \big] \nonumber\\ &+ \sum_{\vec{k}_1\vec{k}_2\vec{k}^\prime \sigma} \big[ V^{\rm AN}_{\vec{k}_1\vec{k}_2\vec{k}^\prime \sigma}(t) c^\dagger_{\vec{k}^\prime \sigma} s^\dagger_{1\downarrow} e^{\phantom{\dag}} c^{\phantom{\dag}}_{\vec{k}_1 \downarrow} c^{\phantom{\dag}}_{\vec{k}_2 \sigma} + \text{H.c.} \big] + \sum_{\vec{k}\vec{k}^\prime \sigma} \big[ V^{\rm DAD}_{\vec{k}\vec{k}^\prime \sigma}(t) c^\dagger_{\vec{k}^\prime \sigma} s^\dagger_{1\downarrow} c^{\phantom{\dag}}_{\vec{k} \sigma} s^{\phantom{\dag}}_{2\downarrow} + \text{H.c.} \big] \nonumber\\ &+ \sum_{\vec{k}\vec{q}\sigma} \big[ V^{\rm IAD}_{\vec{k}\vec{q}\sigma}(t) c^\dagger_{\vec{q} \sigma} s^\dagger_{1\downarrow} c^{\phantom{\dag}}_{\vec{k} \downarrow} s^{\phantom{\dag}}_{2\sigma} + \text{H.c.} \big] + \sum_{\vec{q}} \big[ V^{\rm AuD}_{\vec{q}}\, c^\dagger_{\vec{q} \uparrow} s^\dagger_{1\downarrow} d^{\phantom{\dag}} + \text{H.c.} \big] \,. \end{align} \end{widetext} where we included auxiliary Bose operators $b_\sigma^{(\dagger)}$ enabling us to switch between projectile states with defects in their internal energy~[\onlinecite{PBF18}]. The physical meaning of the various terms is easy to grasp. For instance, the second last term stands for indirect Auger de-excitation (IAD) of the projectile (red arrows in Fig.~\ref{HeNeutralProcesses}b), that is, the creation of the groundstate of the projectile ($s^\dagger_{1\downarrow}$) by creating an electron in an unbound continuum state ($c^\dagger_{\vec{q}\sigma}$) and annihilating either a singlet ($s_{2\downarrow}$) or a triplet ($s_{s\uparrow}$) metastable state and an electron bound in the surface ($c^\dagger_{\vec{k}\downarrow}$). The electron in the continuum is the secondary electron released in the course of Auger de-excitation. The time-dependence of the Hamiltonian~(\ref{pseudoHamiltonian}) reflects the dependence of most of its matrix elements on the actual position $z(t)$ of the projectile in front of the surface. The calculation of the matrix elements is analytically and numerically rather demanding. In fact, most of the computation time required for the quantum-kinetic modeling of secondary electron emission is allocated to the numerical evaluation of the matrix elements (and the selfenergies they give rise to). Explicit expressions for the matrix elements worked out along the lines developed in Refs.~\onlinecite{MBF12b,PBF15a} are given in PBF~[\onlinecite{PBF18}]. The time-dependencies of the energies $\varepsilon_{1s\downarrow}^0(t)$, $\varepsilon^*_{2s\sigma}(t)$, $\varepsilon^-_{2s\sigma}(t)$, and $\varepsilon_{\vec{q}\sigma}(t)$ are caused by long-range polarization effects and short-range non-orthogonality corrections. Assuming the projectile staying a few Bohr radii in front of the surface, the latter can be neglected while the former can be approximated by image shifts. The time-dependencies of the Coulomb matrix elements $V^{\rm AN}_{{\vec{k}_1}{\vec{k}_2}{\vec{k}^\prime\downarrow}}(t)$ for Auger neutralization, $V^{\rm DAD}_{\vec{k}\vec{k}\prime\sigma}(t)$ for direct Auger de-excitation, and $V^{\rm IAD}_{\vec{k}\vec{q}\sigma}(t)$ for indirect Auger de-excitation, as well as the time-dependence of the single-electron transfer matrix element $V_{\vec{k}\sigma}^{\rm SET}(t)$ arise from the overlap of projectile and target wave functions which of course also depends on the separation $z(t)$ of the projectile and target. \subsection{Derivation of the quantum Boltzmann equation for SEE}\label{ss:qbe-see} Once the model is constructed it is analyzed within the quantum kinetic approach pioneered by Langreth and co-workers~[\onlinecite{LN91,SLN94a,SLN94b}]. As result one obtains a linear set of ordinary first order differential equations for the probabilities with which the projectile's electronic configurations occur in the course of the collision. In the case under discussion, one obtains equations for the occurrence probabilities of the ion, the groundstate, the two metastable states and the negative ion of the helium projectile. Three steps are required to obtain the rate equations. First, one has to set up two-time Dyson equations for the projectile Green functions $G$, cf. Eq.~(\ref{eq.kbe}). Second, the selfenergies $\Sigma$ on the right hand side contain all interaction effects and have to be calculated. In the absence of strong electron-electron correlations on the projectile, the non-crossing approximation suffices for that purpose. Finally, the time integrals entering the selfenergies and the Dyson equations (\ref{eq.kbe}) are evaluated within a saddle-point approximation utilizing the fact that various functions are peaked around the time-diagonal. For more details, see~Refs.~\onlinecite{LN91,SLN94a,SLN94b} as well as Refs.~\onlinecite{MBF12b,PBF15a,PBF15b,PBF18}. Following this reasoning one obtains the following system of rate equations \begin{widetext} \begin{align*} \label{rateEq} \frac{d}{dt} \begin{pmatrix} n_+ \vspace{1.25mm}\\ n_\uparrow \,\vspace{1.25mm}\\ n_\downarrow \,\vspace{1.25mm}\\ n_- \vspace{1.25mm}\\ n_g\, \end{pmatrix} = \begin{pmatrix} -[\Gamma_\uparrow^<\! +\! \Gamma_\downarrow^<\! +\! \Gamma_{\rm AN}^<] & \hspace*{-5mm}\Gamma_\uparrow^> & \hspace*{-5mm}\Gamma_\downarrow^> & \hspace*{-5mm} 0 & \,\,0 \hspace*{1mm}\vspace{1.25mm}\\ \Gamma_\uparrow^< & \hspace*{-5mm}-[\Gamma_\uparrow^>\! + \!\Gamma^<_{-,\downarrow} \!+\! \Gamma_{\rm IAD\uparrow}^<] & \hspace*{-5mm} 0 & \hspace*{-5mm}\Gamma^>_{-,\downarrow} & \,\,0 \hspace*{1mm}\vspace{1.25mm}\\ \Gamma_\downarrow^< & \hspace*{-5mm}0 & \hspace*{-5mm}-[\Gamma_\downarrow^> \!+\! \Gamma^<_{-,\uparrow} \!+\! \Gamma_{\rm IAD\downarrow}^< \!+\! \Gamma_{\rm DAD\downarrow}^<] & \hspace*{-5mm} \Gamma^>_{-,\uparrow} & \,\,0 \hspace*{1mm}\vspace{1.25mm}\\ 0 & \hspace*{-5mm}\Gamma^<_{-,\downarrow} & \hspace*{-5mm}\Gamma^<_{-,\uparrow} & \hspace*{-5mm} -[\Gamma^>_{-,\uparrow} \!+\! \Gamma^>_{-,\downarrow} \!+\! \Gamma^<_{\rm AuD}] & \,\,0 \hspace*{1mm}\vspace{1.25mm}\\ \Gamma_{\rm AN}^< & \hspace*{-5mm}\Gamma_{\rm IAD\uparrow}^< & \hspace*{-5mm}\Gamma_{\rm IAD\downarrow}^< \!+\! \Gamma_{\rm DAD\downarrow}^< & \hspace*{-5mm} \Gamma^<_{\rm AuD} & \,\,0 \hspace*{1mm} \end{pmatrix} \cdot \begin{pmatrix} n_+ \vspace{1.25mm}\\ n_\uparrow \,\vspace{1.25mm}\\ n_\downarrow \,\vspace{1.25mm}\\ n_- \vspace{1.25mm}\\ n_g \, \end{pmatrix} \, \end{align*} \end{widetext} where $n_+(t)$, $n_\uparrow(t)$, $n_\downarrow(t)$, $n_-(t)$, and $n_g(t)$ denote, respectively, the occurrence probabilities at time $t$ for the positive ion, the triplet and singlet metastable state, the negative ion, and the groundstate of the projectile. Expressions for the time-dependent rates $\Gamma_{...}^\gtrless(t)$, a discussion of the physical content of the rate equations as well as a route for obtaining the secondary electron emission coefficient $\gamma_e$ and the energy spectrum $\gamma_e(\varepsilon_{\vec{q}})$ of the emitted electron from the solution of the rate equations are given in PBF~[\onlinecite{PBF18}]. To indicate the type of data which can be produced by the quantum-kinetic modeling of secondary electron emission, Fig.~\ref{HeNeutralData} shows numerical results for a positive helium ion hitting perpendicularly an aluminum surface with $E_{\rm kin}=60\,{\rm eV}$ using the material parameters listed in PBF~[\onlinecite{PBF18}]. The left panel depicts the instantaneous occurrence probabilities $n_+(t), n_-(t), n_\uparrow(t)$, $n_\downarrow(t)$, and $n_g(t)$ as well as the instantaneous probability $\gamma_e(t)$ for emitting a secondary electron. The energy spectrum of the emitted electron is plotted in the right panel. The projectile starts in the $\mathrm{He^+(1s,1^2S_{1/2})}$ configuration at a distance $z=40\,a_{\rm B}$, where $a_{\rm B}$ is the Bohr radius, moves along the trajectory shown in Fig.~\ref{HeNeutralProcesses}a towards the turning point $z_{\rm tp}=2.27\, a_{\rm B}$, where it is specularly reflected to move back to the distance $z=40\,a_{\rm B}$ to complete the collision. The probabilities for finding the projectile at the end of the collision in any one of the configurations included in the modeling can be read off from the left panel. For instance, the probability for returning in the groundstate configuration $\mathrm{He^0(1s^2,1^1S_0)}$ is $n_g(t_{\rm max})\approx 0.96$ whereas the probability for surviving the collision in the $\mathrm{He^+(1s,1^2S_{1/2})}$ configuration is $n_+(t_{\rm max})\approx 0.04$. There is also a very small probability $n_\uparrow(t_{\rm max})\approx 10^{-5}$ to come back in the $\mathrm{He^\ast(1s2s,2^3S_1)}$ triplet configuration. The probability for electron emission, the secondary electron emission coefficient $\gamma_e=\gamma_e(t_{\rm max})\approx 0.07$. It has the correct order of magnitude suggesting that by careful testing, benchmarking, and comparison with experimental data the semi-empirical approach can be advanced to a level where it can produce very reliable data. The energy spectrum in the right panel indicates that the secondary electron can have rather high energies. It can thus become chemically very active in the gas discharge. A detailed analysis shows that for the aluminum surface secondary electron emission is dominated by Auger neutralization. For other metals, Auger de-excitation may become also important. The relative importance of the two channels depends on the line-up of the projectile's affinity and ionization levels with the Fermi level of the metal. It can thus be controlled by a judicious choice of the metal and the projectile, that is, in the plasma context, of the wall material and the composition of the background gas. \begin{figure*}[t] \begin{minipage}{0.5\linewidth} \includegraphics[width=0.99\linewidth]{occ_all90} \end{minipage}\begin{minipage}{0.5\linewidth} \includegraphics[width=0.99\linewidth]{spectrum_90} \end{minipage} \caption{(Color online) Left panel: Instantaneous probabilities $n_+(t)$, $n_-(t), n_\uparrow(t), n_\downarrow(t)$, and $n_g(t)$ for the projectile to be at time $t$ in the $\mathrm{He^+(1s,1^2S_{1/2})}$, $\mathrm{He^{\ast-}(1s2s^2,2^2S_{1/2})}$, $\mathrm{He^\ast(1s2s,2^3S_1)}$, $\mathrm{He^\ast(1s2s,2^1S_0)}$, and the $\mathrm{He^0(1s^2,1^1S_0)}$ configuration together with the instantaneous probability $\gamma_e(t)$ for emitting an electron. The projectile hits the aluminum surface as a positive ion (solid black line) with a kinetic energy $E_{\rm kin}=60$\, eV and an angle of incidence with respect to the surface of $\varphi=90^\circ$. The thin vertical line denotes the turning point $z_{\rm TP}=2.27$, separating the incoming (left) form the outgoing (right) branch of the collision trajectory. The final probabilities, after the collision is completed, which are also the numbers relevant for plasma modeling, are the values at $z=40~a_{\rm B}$ on the outgoing branch. Right panel: Energy spectrum $\gamma_e(\varepsilon_{\vec{q}})$ of the emitted electron after the collision is completed. } \label{HeNeutralData} \end{figure*} The presentation of the quantum Boltzmann equation approach for charge-transferring atom-surface collisions was couched in a particular application: secondary electron emission. It can be, however, also applied to other atom-surface scattering processes affecting the electronic structure of the projectile-target system. The pseudo-particle representation of the projectile states opens the door for handling, within a single Hamiltonian, complex collisions, involving more than one channel. Numerically the approach is rather involved, preventing thus a quick production of surface parameters needed in plasma modeling. The main obstacle is the calculation of the matrix elements. So far they have been mostly obtained in a semi-empirical manner, guided by physical considerations and experimental data. Using \textit{ab initio} techniques, such as DFT [cf. Sec.~\ref{s:abinit}], instead may eliminate some of the uncertainties of the matrix elements but the numerical effort will remain the same or even increase. Making the quantum Boltzmann equation-based modeling of charge-transferring atom-surface collisions numerically more efficient is one of the challenges for the future. Finally, we notice that the quantum Boltzmann equation cannot resolve ultra-short time and length scales of the solid surface that are related electron correlations (cf. Fig.~\ref{fig:scales}). The extension of quantum kinetic theory into these ranges can be achieved using the method of Nonequilibrium Green functions which is discussed in detail in Sec.~\ref{s:negf}. There we will also present a more accurate treatment of the solid surface that is complementary to the present approach. \section*{References}

960f8e9271043493f59792107fd5d18dc76532e4

\part{Introduction} This paper is mainly occupied with identifying homotopy-algebraic structures on Galois cohomology groups that explain number-theoretic phenomena. More specifically, we describe an $A_\infty$-algebra structure on the Galois cohomology of the adjoint representation of a Galois representation $\rho$ and relate it to the deformation theory of $\rho$. We prove that a certain classical hull of this homotopy algebra represents the classical Galois deformation problem. This gives a presentation, in terms of Galois cohomology, of the Galois deformation rings first studied by Mazur \cite{mazur1989}. It is fair to call this homotopy algebra a ``derived enrichment'' of classical moduli rings of Galois representations, commonly known as universal deformation rings of Galois representations, or \emph{Galois deformation rings}. Therefore, there are some relations between this work and, for example, that of Galatius--Venkatesh \cite{GV2018} on derived Galois deformation rings; for more comments about this, see \S\ref{subsec: GV2018}. However, we do not discuss derived deformation problems here, as our motivation is to identify structures in Galois cohomology that control objects that live squarely in the classical world. Thus, the reader interested in recent developments in derived enrichment to the Langlands correspondence, initiated by Venkatesh and collaborators, may take the perspective that the present paper shows how to explicitly compute classical Galois deformation rings in terms of derived structures. Indeed, the original motivation for this study, introduced in \S\ref{sec: motivation}, is the search for a tangent and obstruction theory for Galois \textit{pseudorepresentations}. This goal is reached in two steps.\label{pg: twosteps} \begin{enumerate}[leftmargin=2em] \item As discussed above, find a presentation for classical moduli rings of Galois representations in terms of cohomological data. The usual tangent and obstruction theory of representations is augmented by computations of $A_\infty$-products, which is needed in order to find this presentation. \item Deduce a tangent and obstruction theory for Galois pseudorepresentations using the author's previous work \cite{WE2018}, which establishes that moduli rings of pseudorepresentations are invariant subrings of moduli rings of representations under a natural adjoint action. This refines work of Bella\"iche \cite{bellaiche2012}. \end{enumerate} The theoretical content behind step (2) comes entirely from \cite{WE2018}. Thus what is new in this paper is to carry out step (1), and then deduce presentations for the adjoint-invariant subrings of the rings presented in (1). After steps (1) and (2) are complete, we \begin{enumerate}[leftmargin=2em, resume] \item demonstrate that the (1) and (2) may be adapted to the deformation theory of Galois (pseudo)representations with a local conditions imposed, and \item as a sample application, prove that the ranks of certain $p$-adic modular Hecke algebras -- both residually Eisenstein and residually cuspidal -- are determined by $A_\infty$-products in Galois cohomology. \end{enumerate} As the presentations of step (1) are novel in number theory, but familiar in non-commutative geometry (see \S\ref{subsec: NCDT}), explaining how these fields interface is a secondary goal of this paper. For this reason, we hope that the reader can gain from the extended introduction -- \S\S1-4 -- what is new to them, and skip over familiar fundamental facts. In particular, the reader interested in the main results can skip ahead to \S\ref{sec: main results} or, alternatively, to the introduction to $A_\infty$-algebras in \S\ref{sec: A-infinity background}. The topics of this paper have been studied from a diverse array of mathematical fields and perspectives. To avoid distracting the reader with references to such connections throughout the text, we concentrate this discussion of ``Complements'' in \S\ref{sec: complements}. \section{Why Galois representations and why $A_\infty$-algebras} \label{sec: motivation} In this section, we introduce number-theoretic motivation to non-commutative geometers and introduce non-commutative geometry techniques to number theorists. Readers familiar with the deformation theory of Galois representations and pseudorepresentations may wish to proceed to \S\ref{subsec: red case} or \S\ref{subsec: illustrate}. \subsection{2-dimensional Galois representations and modular Hecke eigenforms} Consider 2-dimensional representations $\rho$ of the absolute Galois group $G_\mathbb{Q}$ of $\mathbb{Q}$, with coefficients in $\overline{\Q}_p$. We will call these \textit{Galois representations} valued in $\overline{\Q}_p$, and presume that all functions out of $G_\mathbb{Q}$ are continuous without further remark. In fact, up to conjugacy, $\rho$ is valued in the integral closure $\mathcal{O}_E$ of $\mathbb{Z}_p$ in a finite subextension $E/\mathbb{Q}$ of $\overline{\Q}_p/\mathbb{Q}_p$, \[ \rho: G_\mathbb{Q} \longrightarrow \GL_2(\mathcal{O}_E). \] On the other hand, from the theory of Hecke actions on modular forms, we have normalized modular Hecke eigenforms $f$. As these are holomorphic functions of a complex variable $z$ in the upper half of the complex plane with period 1, which we think of as $q$-series in $\mathbb{C}{[\![} q{]\!]}$ for $q = e^{2 \pi i z}$. In fact, these have algebraic integer coefficients $a_n(f)$, that is, \[ f = \sum_{n \geq 0} a_n(f) q^n \in \overline{\mathbb{Z}}{[\![} q{]\!]} \quad \text{ after the normalization } a_1(f) = 1. \] The Fontaine--Mazur conjecture \cite{FM1995} predicts that certain $\rho$ are expected to be \textit{modular}, that is, to be isomorphic to the Galois representation $\rho_f : G_\mathbb{Q} \to \GL_2(\overline{\Q}_p)$ arising from a Hecke eigenform $f$. The key property of $\rho_f$ is that its trace function is characterized by the condition \[ \Tr \rho_f(\mathrm{Frob}_\ell) = a_\ell(f) \] for all but finitely many prime numbers $\ell$, where $\mathrm{Frob}_\ell \in G_\mathbb{Q}$ is a choice of Frobenius element relative to a prime $\ell$. We have implicitly fixed an isomorphism $\overline{\Q}_p \simeq \mathbb{C}$ to make this comparison. Because both the $\rho$ and $f$ have the $p$-integral structure that we have just explained, they are organized into congruence classes modulo $p$. We label the congruence classes of Galois representations and Hecke eigenforms by \[ \bar\rho : G_\mathbb{Q} \rightarrow \GL_2(\mathbb{F}) \text{ and } \bar f \in \mathbb{F}{[\![} q{]\!]}, \] respectively. Here $\mathbb{F}$ is a finite field of characteristic $p$. Work of Mazur \cite{mazur1989, mazur1978} introduced the moduli-theoretic study of these congruence classes, putting them in bijection with \begin{itemize}[leftmargin=2em] \item (Galois representations) homomorphisms $R_{\bar\rho} \rightarrow \overline{\Q}_p$ out of a deformation ring $R_{\bar\rho}$ of Galois representations, designed so that commutative $\mathbb{Z}_p$-algebra homomorphisms $R_{\bar\rho} \rightarrow A$ correspond to strict equivalence classes of Galois representations $\rho_A$ with coefficients in $A$ such that \begin{enumerate}[label=(\roman*), leftmargin=2em] \item $\rho_A : G_\mathbb{Q} \rightarrow \GL_2(A)$ is congruent to $\bar\rho$, and \item $\rho_A$ satisfies properties expected of Galois representations arising from Hecke eigenforms. \end{enumerate} \item (Hecke eigenforms) homomorphisms $\mathbb{T} \rightarrow \overline{\Q}_p$ out of a Hecke algebra $\mathbb{T}$, where $\mathbb{T}$ is the completion of a Hecke algebra (arising from the Hecke action on a finite dimensional $\mathbb{C}$-vector space of modular forms) at a maximal ideal (with finite residue field of characteristic $p$) corresponding to $\bar f$. \end{itemize} When $\bar\rho$ and $\bar f$ may be chosen compatibly, which we now assume, it is natural to ask whether there is a local homomorphism $R_{\bar\rho} \rightarrow \mathbb{T}$ arising from the $p$-adic Galois representations attached to Hecke eigenforms. Then, one is led to ask about ``$R_{\bar\rho} = \mathbb{T}$,'' a statement that Galois representations and Hecke eigenforms interpolate compatibly. \subsection{The irreducible case} \label{sssec: irred case} When there are no Eisenstein series congruent to $\bar f$, or, equivalently, $\bar\rho$ is absolutely irreducible, it is often possible to prove that $R_{\bar\rho} \buildrel\sim\over\ra \mathbb{T}$ -- this was first carried out by Mazur \cite{mazur1989} and Wiles \cite{wiles1995}. Since Wiles's work, one crucial aspect of this argument is control over $R_{\bar\rho}$ in terms of the arithmetic invariants of Galois cohomology. In the most basic setting, these are \[ H^1(G, \mathrm{End}_\mathbb{F}(\rho)) \text{ and } H^2(G, \mathrm{End}_\mathbb{F}(\rho)), \] where $\rho$ may now have arbitrary dimension $d$. This control is called a \textit{tangent and obstruction theory}, which we now explain. For $n \geq 0$, an $n$th-order \textit{lift} of $\rho$ is a representation \[ \rho_n: G \longrightarrow \GL_d(\mathbb{F}[\varepsilon]/\varepsilon^{n+1}) \quad \text{such that } (\rho_n \mod{\varepsilon}) = \rho. \] These lifts are called \emph{strictly equivalent} when they are conjugate by the adjoint action of a matrix of the form $1_{d \times d}+ \varepsilon \cdot M_{d \times d}(\mathbb{F}[\varepsilon]/\varepsilon^{n+1})$. Strict equivalence classes of $n$th-order lifts are called \emph{$n$th-order deformations} of $\rho$. The content of a tangent and obstruction theory is that first-order deformations of $\rho$ (which comprise the \emph{tangent space}) are in bijection with \[ H^1(G, \mathrm{End}_\mathbb{F}(\rho)) \cong \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho). \] and that, for $n \geq 1$, an $n$th-order deformation $\rho_n$ induces an element of the \emph{obstruction space} \[ H^2(G, \mathrm{End}_\mathbb{F}(\rho)) \cong \mathrm{Ext}^2_{\mathbb{F}[G]}(\rho,\rho) \] that is zero if and only if $\rho_n$ can be extended to an $(n+1)$st order deformation, i.e.\ \[ \rho_{n+1} : G \longrightarrow \GL_d(\mathbb{F}[\varepsilon]/\varepsilon^{n+2}) \quad \text{such that } (\rho_{n+1} \mod{\varepsilon^{n+1}}) = \rho_n. \] When one exists, these extensions are a torsor over $H^1(G,\mathrm{End}_\mathbb{F}(\rho))$. In practice, one carries out the Taylor--Wiles method of \cite{wiles1995, TW1995} and subsequent developments. Very roughly speaking, these involve auxiliary cases where the tangent and obstruction theory reduces to the simple case $H^2(\mathrm{aux}, \mathrm{End}_\mathbb{F}(\rho)) = 0$. In contrast, our goal, expressed in the language of deformations and obstructions, is to provide a systematic framework to find the obstruction in $H^2(G,\mathrm{End}_\mathbb{F}(\rho))$ that arises from an $n$th-order deformation $\rho_n$. \subsection{The reducible case} \label{subsec: red case} When $\bar\rho$ is reducible, it is necessary to modify the approach above. Let $d=2$ for simplicity. In order to hope for a correspondence with $\mathbb{T}$ in general, we must replace $R_{\bar\rho}$ by a ring $R^\mathrm{ps}$ that parameterizes 2-dimensional \emph{pseudorepresentations} of Galois groups. So we write $R^\mathrm{ps}$ for clarity. Indeed, isomorphisms $R^\mathrm{ps} \cong \mathbb{T}$ have been proven; see e.g.\ \cite{BK2011, BK2015, deo2017, WWE1}, which also give reasons for studying $R^\mathrm{ps}$ instead of $R_{\bar\rho}$. A \emph{2-dimensional pseudorepresentation} of $G_\mathbb{Q}$ valued in $A$, written $D_A : G \rightarrow A$, amounts to a pair of functions \[ D_A = \{\Tr, \det : G \rightarrow A\} \] obeying properties expected of such functions arising from characteristic polynomials of a 2-dimensional representation with coefficients in $A$. For a precise definition of a $d$-dimensional pseudorepresentation due to Chenevier \cite{chen2014}, see \S\ref{subsec: review RR case}. Given a representation $\rho$, we write $\psi(\rho)$ for the induced pseudorepresentation. For the moment, we take $\bar D = \psi(\bar\rho) : G \rightarrow \mathbb{F}$ to be the pseudorepresentation given by $\{\Tr \bar\rho, \det \bar\rho\}$. Then we let $R^\mathrm{ps}$ be the ``universal pseudodeformation ring'' for ${\bar D}$; it has the universal property that local $\mathbb{Z}_p$-algebra homomorphisms $R^\mathrm{ps} \rightarrow A$ are in bijection with pseudorepresentations $D_A : G \rightarrow A$ such that \begin{enumerate}[label=(\roman*), leftmargin=2em] \item $D_A$ is congruent to ${\bar D}$, i.e.\ the composite of $D_A$ and $A \twoheadrightarrow A/\mathfrak{m}_A \cong \mathbb{F}$ is equal to ${\bar D}$; and \item $D_A$ satisfies properties of Galois representations arising from Hecke eigenforms; such conditions are translated from representations to pseudorepresentations (of any dimension) in the author's work with Preston Wake \cite{WWE4}. \end{enumerate} However, a tangent and obstruction theory for pseudorepresentations has been lacking. For example, Thorne remarks that ``the ring $R^\mathrm{ps}$ is difficult to control using Galois cohomology, a tool which is essential in other arguments'' \cite[pg.\ 786]{thorne2015}. Indeed, to this author's knowledge, no formulation of obstruction theory had been produced. There is a partial characterization of a canonical filtration on the tangent space due to Bella\"iche \cite{bellaiche2012} (following on his work with Chenevier \cite{BC2009}), in the case where the semi-simple $\bar\rho$ inducing ${\bar D}$ has distinct simple factors. This is known as the \emph{residually multiplicity-free} case. However, the tangent space is only characterized when there are two simple factors \cite[Thm.\ A]{bellaiche2012}. The terminal result of this paper, Theorem \ref{thm: main pseudo}, supplies a tangent and obstruction theory for a multiplicity-free residual pseudorepresentation. Theorem \ref{thm: main pseudo} arises from the product of two antecedent theorems and a combinatorial calculation. The two antecedent theorems comprise the two steps mentioned above (pg.\ \pageref{pg: twosteps}). \begin{enumerate}[leftmargin=2em] \item Present moduli rings of Galois representations in terms choices determining an $A_\infty$-algebra structure on the adjoint cohomology. \item Apply the expression for $R^\mathrm{ps}$ as an invariant subring of a moduli ring for Galois representations proved in \cite{WE2018}. \end{enumerate} Step (1), which is carried out in Part 2, is this paper's main theoretical development. Part 3 works out Step (2), which is an application of the theory of \cite{WE2018}, working toward applications to deformations of Galois representations and pseudorepresentations. \begin{rem} In fact we produce a \emph{presentation} for $R^\mathrm{ps}$ to stand in for a conventional tangent and obstruction theory, which only gives the tangent space and the space where obstructions may be computed. Conventional tangent and obstruction theory relies on the residual $\mathbb{F}$-valued object being an object in a $\mathbb{F}$-linear category. But there is no abelian category structure on pseudorepresentations. Thus we rely on deformation theory of representations and the invariant-theoretic step (2) above. \end{rem} \begin{rem} As is well-known, subrings of invariants of a group action on a regular ring may not be regular; that is, step (2) induces obstructions, which are of a combinatorial nature (details in \S\ref{subsec: GIT intro}). Moreover, one sees in the calculation of invariants that the tangent space of the pseudodeformation functor is ``obstructed,'' in the sense that obstructions of deformations of representations cut down the possible first-order deformation of pseudorepresentations. More specifically, if the deformed pseudorepresentation arises from a semi-simple representation $\rho$ with $n$ distinct irreducible summands, then one must know certain deformations of $\rho$ up to $n$th-order in order to merely calculate the tangent space of pseudodeformations. \end{rem} \subsection{Concrete illustration of the role of $A_\infty$-products and adjoint invariants} \label{subsec: illustrate} In this section, we aim to explain to the reader why an $A_\infty$-algebra on adjoint Galois cohomology naturally supplies a presentation of a Galois deformation ring. First, we put forth the perspective that the calculation of Massey products is the natural context to test whether an $n$th-order lift extends to an ($n+1$)-st order lift. Then, we explain that Massey products are efficiently subsumed into products occurring in an $A_\infty$-algebra structure on adjoint Galois cohomology, and outline the method of presenting the Galois deformation ring in terms of the $A_\infty$-structure. \subsubsection{Calculating obstructions inductively} \label{sssec: massey calc intro} Suppose that we want to calculate the $n$th-order lifts of an irreducible representation $\rho: G \to \GL_d(\mathbb{F})$. These are homomorphisms $\rho_n : G \to \GL_d(\mathbb{F}[\varepsilon_n])$ reducing modulo $\varepsilon$ to $\rho$, where $\mathbb{F}[\varepsilon_n] := \mathbb{F}[\varepsilon]/(\varepsilon^{n+1})$. Taking a 1-cocycle $e$ on $G_\mathbb{Q}$ valued in $\mathrm{End}_\mathbb{F}(\mathbb{F}^{\oplus d})$ with action via $\rho$, there is a first-order lift $\rho_1$ determined by $e$ whose expression with respect to the obvious $\mathbb{F}$-basis of $M_d(\mathbb{F}[\varepsilon_1])$ is \[ \rho_1 = \begin{pmatrix} \rho & e \\ & \rho \end{pmatrix}. \] A lift of $\rho_1$ to second order has the form \[ \rho_2 = \begin{pmatrix} \rho & e & f_2 \\ & \rho & e \\ & & \rho \end{pmatrix} \] for some function $f_2 : G_\mathbb{Q} \to M_d(\mathbb{F})$. One should think of $f_2$ as a 1-cochain on $G_\mathbb{Q}$ under the adjoint $G_\mathbb{Q}$ action on $\mathrm{End}_\mathbb{F}(\rho) = M_d(\mathbb{F})$. We write $C^1(G_\mathbb{Q},\mathrm{End}_\mathbb{F}(\rho))$ for this cochain space and we write $d$ for the coboundary map. One may calculate that the function must satisfy \[ df_2 = e \smile e, \] where ``$\smile$'' denotes the case $i = j = 1$ of the cup product map, which is the composite \begin{gather} \label{eq: cup product in cochains} \begin{split} C^i(G_\mathbb{Q},\mathrm{End}_\mathbb{F}(\rho)) \times C^{j}(G_\mathbb{Q}, &\, \mathrm{End}_\mathbb{F}(\rho)) \to C^{i+j}(G_\mathbb{Q},\mathrm{End}_\mathbb{F}(\rho) \otimes_\mathbb{F} \mathrm{End}_\mathbb{F}(\rho))) \\ & \to C^{i+j}(G_\mathbb{Q},\mathrm{End}_\mathbb{F}(\rho)). \end{split} \end{gather} The first arrow is the standard cup product in group cochains, while the second arrow arises from composition in $\mathrm{End}_\mathbb{F}(\rho)$. The coboundary map $d$ along with this multiplication endows $C^\bullet(G_\mathbb{Q},\mathrm{End}_\mathbb{F}(\rho))$ with a differential graded algebra (dg-algebra) structure (see e.g.\ \cite[Prop.\ 1.4.1]{NSW2008}). \begin{rem} \label{rem: intro rem 2nd order def} In particular, we see that $\rho_1$ extends to second order if and only if $e \smile e$ is a 2-coboundary, and that the choices for $f_2$ are a torsor over the 1-cocycles $Z^1(G_\mathbb{Q},\mathrm{End}_\mathbb{F}(\rho))$. In particular, the cohomology class in $H^2(G_\mathbb{Q}, \mathrm{End}_\mathbb{F}(\rho))$ of $e \smile e$ is the obstruction to extending $\rho_1$ to second order. This is a well-known fact, cf.\ \cite[\S1.6, Remark; pg.\ 400]{mazur1989}. \end{rem} A prospective third-order lift extending $\rho_2$ has the form \[ \rho_3 = \begin{pmatrix} \rho & e & f_2 & f_3 \\ & \rho& e & f_2 \\ & & \rho & e \\ & & & \rho \end{pmatrix}. \] One may calculate that $\rho_3$ is a homomorphism if and only if $f_3$ satisfies \[ df_3 = e \smile f_2 + f_2 \smile e. \] For the induction step in general, whenever an $(n-1)$st order lift $\rho_{n-1}$ exists (and has the form \eqref{eq: rho_n def}), a prospective extension to $n$th-order has the form \begin{equation} \label{eq: rho_n def} \rho_n = \begin{pmatrix} \rho & e & f_2 & f_3 & \dotsm & f_n \\ & \rho & e & & & f_{n-1} \\ & & \ddots & \ddots & & \vdots \\ & & & \rho& e & f_2 \\ & && & \rho & e \\ & & && & \rho \end{pmatrix}, \end{equation} and is a bona-fide $n$th-order lift of $\rho$ if and only if \begin{equation} \label{eq: nth massey product} e\smile f_{n-1} + f_2 \smile f_{n-2} + \dotsm + f_{n-2} \smile f_2 + f_{n-1} \smile e \end{equation} is a 2-coboundary. Because this expression is known to be a 2-cocycle, we may think of $H^2(G_\mathbb{Q},\mathrm{End}_\mathbb{F}(\rho))$ as the set of possible obstructions to deforming representations. \subsubsection{Appearance of Massey products} The theory of \textit{Massey products} in dg-algebras, with codomain in tuples of 1-cocycles and valued in $H^2(G_\mathbb{Q},\mathrm{End}_\mathbb{F}(\rho))$, is the proper context for the computation of obstructions as above. Deferring details to \S\ref{sec: massey}, we give an overview of Massey products adapted to apply only to the computations of lifts of $\rho$ above. Massey products of degree $n$ on $C^1(G_\mathbb{Q}, \mathrm{End}_\mathbb{F}(\rho))$ are partially defined and multi-valued $\mathbb{F}$-multi-linear functions \begin{equation} \label{eq: massey adjoint ex} H^1(G_\mathbb{Q}, \mathrm{End}_\mathbb{F}(\rho))^{\times n} \longrightarrow H^2(G_\mathbb{Q},\mathrm{End}_\mathbb{F}(\rho)), \end{equation} where each of the multiple values arises from a choice of \emph{defining system}. The term \emph{Massey product $\langle e_1, \dotsc, e_n\rangle$} for an $n$-tuple in the domain of \eqref{eq: massey adjoint ex} refers to the set of all values over all possible defining systems. In the present context, an $(n-1)$st order lift of the form \eqref{eq: rho_n def} furnishes a particular defining system for the Massey product on the $n$-tuple $(e, e, \dotsc, e)$, and the element of the Massey product $\langle e, \dotsc, e\rangle$ determined by this defining system is the cohomology class of the 2-cocycle written down in \eqref{eq: nth massey product}. Not all values of a Massey product on $n$-tuples of copies of $e$ are realized by the defining systems that test the existence of lifts of $\rho_1$ to $n$th-order, as general defining systems are allowed much more flexibility than the defining systems arising from lifts of $\rho_1$ to $(n-1)$st order. In turn, these other defining systems are related to other lifting questions, and vice versa. Indeed, in principle, Massey products have all of the information needed to determine deformations of $\rho$. This approach is taken in \cite{laudal2002}. However, one disadvantage of the theory of Massey products is its apparent dependence on a multitude of choices that grows inductively with the order of the deformation. This is why we confine ourselves to saying that the theory of Massey products in the dg-algebra $C^\bullet(G_\mathbb{Q},\mathrm{End}_\mathbb{F}(\rho))$ encompasses all of the calculations needed to understand all possible lifts of $\rho$. Instead, we emphasize an alternative to Massey products -- $A_\infty$-products -- coming from the theory of $A_\infty$-algebras. See \S\ref{subsec: A-inf to Massey} for a thoroughgoing explanation of the relationship between Massey products and $A_\infty$-products. \subsubsection{$A_\infty$-algebras are homotopy-associative} \label{sssec: A-inf intro for app} The notion of a $A_\infty$-algebra is a generalization of the notion of dg-algebra, where multiplication, now denoted $m_2$, is no longer required to be associative. Instead, it is required to be associative up to homotopy, and further data keeping track of these homotopies comprises the rest of the $A_\infty$-algebra structure. This notion was first introduced by Stasheff as a ``strong homotopy associative algebra'' \cite{stasheff1963}. We will proceed in this introduction by using the following informal definition of an $A_\infty$-structure on a $\mathbb{Z}$-graded $\mathbb{F}$-vector space $H$, where $\mathbb{F}$ is a field. See \S\ref{subsec: A-inf defn} for details. \begin{itemize}[leftmargin=2em] \item $m_1$ denotes the differential on $H$, of degree 1 \item $m_2$ denotes multiplication (of degree 0), which may be non-associative but satisfies the Leibniz rule with respect to $m_1$ \item The rest of the $A_\infty$-structure consists of a sequence of homotopies $m_n, n \geq 3$, each of degree $2-n$. Together, they track failures of associativity. Namely, $m_3$ sends tuples $(a,b,c) \in H^{\times 3}$ $\mathbb{F}$-multilinearly to elements of $H$ whose boundary (with respect to $m_1$) is \[ (a \cdot b) \cdot c - a \cdot (b \cdot c), \] where the product refers to $m_2$. Subsequent $m_n$ have boundary measuring the failure of a generalized associativity notion for $n$ elements. \item Altogether, an $A_\infty$-algebra structure on $H$, written $m = (m_n)_{n \geq 1}$, consists of $\mathbb{F}$-linear maps \[ m_n : (H^\bullet)^{\otimes n} \longrightarrow H^\bullet \quad \text{of graded degree } 2-n, \ n \in \mathbb{Z}_{\geq 1}, \] satisfying certain compatibility conditions. In particular, it is required that $m_1 \circ m_1 = 0$, and $m_2$ satisfies the Leibniz rule with respect to $m_1$. \item The data of $n$th Massey products above appears in the restriction of $m_n$ to $H^1$, i.e.\ \[ m_n : (H^1)^{\otimes n} \longrightarrow H^2. \] For more details on the relationship between $m_n$ and $n$th Massey products, see \S\ref{subsec: A-inf to Massey}. In summary, we take the perspective that the $A_\infty$-quasi-isomorphism $f$ is a way of ``choosing all defining systems of Massey products in advance.'' \end{itemize} We write $(H,m)$ for such an $A_\infty$-algebra. We defer details on morphisms of $A_\infty$-algebras to \S\ref{subsec: A-inf defn}. In any quasi-isomorphism class of $A_\infty$-algebras, there are two extremal kinds of representatives. \begin{enumerate}[label=(\roman*)] \item One is the \emph{dg-algebra case}, where the ``higher multiplications'' $m_n, n \geq 3$ vanish and, hence, $m_2$ is associative. Then $m_1$ is the differential map and $m_2$ is the multiplication map of a dg-algebra. \item The other is the \emph{minimal} case, where $m_1 = 0$; in other words, this $A_\infty$-algebra is equal to its cohomology. We sometimes write $m = (m_n)_{n \geq 2}$ for a minimal $A_\infty$-structure. \end{enumerate} In fact, Kadeishvili \cite{kadeishvili1982} proved that when one starts with a dg-algebra $C^\bullet$, there is a minimal $A_\infty$-algebra structure $m$ on its cohomology complex $H^\bullet := H^\bullet(C^\bullet)$ that both \begin{itemize}[leftmargin=2em] \item extends to $m = (m_n)_{n \geq 1}$ the graded multiplication $m_2$ on $H^\bullet$ induced by the multiplication on $C^\bullet$, along with the trivial differential $m_1 = 0$ on $H^\bullet$, and \item admits a quasi-isomorphism $f$ to $C^\bullet$. In particular, the quasi-isomorphism includes the data of a map of complexes from $H^\bullet$ to $C^\bullet$ that is a section of projection from cocycles to cohomology. \end{itemize} We use a refinement of Kadeishvili's result due to Merkulov \cite{merkulov1999}: a quasi-isomorphism $f: H^\bullet \to C^\bullet$ as above may be explicitly determined from a choice of section $H^\bullet \to C^\bullet$ of the standard projection from cocycles to cohomology. Such a section may be thought of as the following set of choices, using the differential $m_1$: for each $i \in \mathbb{Z}_{\geq 0}$, choose a decomposition of $C^i$ into $i$-coboundaries, $i$-cocycles that lift cohomology classes, and $i$-cochains that lift $(i+1)$-coboundaries. For details of the formula, see Example \ref{eg:KFT2}. \subsubsection{Application of $A_\infty$-algebras to deformations} We now summarize the application of the framework above to deformations of a $\mathbb{F}$-linear representation $\rho$ of $G_\mathbb{Q}$. We start with adjoint cochains and cohomology \[ C^\bullet = C^\bullet(G_\mathbb{Q},\mathrm{End}_\mathbb{F}(\rho)) \quad\text{and}\quad H^\bullet = C^\bullet(G_\mathbb{Q},\mathrm{End}_\mathbb{F}(\rho)). \] As we noted in \S\ref{sssec: massey calc intro}, $C^\bullet$ is naturally a dg-algebra. The framework gives us an $A_\infty$-algebra structure $m$ on $H^\bullet$ and an $A_\infty$-quasi-isomorphism $f$ from $(H^\bullet,m)$ to $C^\bullet$. The structures $m$ and $f$ may be leveraged to produce a presentation for the deformation ring, using the following argument. \noindent \textbf{1.\ From $C^\bullet$ to deformations of $\rho$ via the Maurer--Cartan equation.} $C^\bullet$ has a direct connection with representations of $G_\mathbb{Q}$: 1-cochains are functions on $G_\mathbb{Q}$. Sacrificing the $\mathbb{F}$-linear formulation of the Massey products, the relationship between $C^\bullet$ and lifts of $\rho$ is concisely expressed by the \emph{Maurer--Cartan equation}. If $A$ is a local Artinian $\mathbb{F}$-algebra with residue field $A/\mathfrak{m}_A \cong \mathbb{F}$, then lifts of $\rho$ to $A$ are in bijection the solutions to the Maurer--Cartan equation, denoted \begin{equation} \label{eq: dg MC intro} \mathrm{MC}(C,A) := \{\xi \in C^1 \otimes_\mathbb{F} \mathfrak{m}_A \mid d\xi + \xi \cdot \xi = 0 \quad \text{in } C^2 \otimes_\mathbb{F} \mathfrak{m}_A\}, \end{equation} where we use shorthand $C^i = C^i(G_\mathbb{Q},\mathrm{End}_\mathbb{F}(\rho))$. Full discussion of the Maurer--Cartan equation begins in \S\ref{subsec: MC functor intro}. \noindent \textbf{2.\ Gauge action.} There is a standard notion of \emph{gauge action} on $\mathrm{MC}(C^\bullet,A)$, which we check amounts to conjugation of lifts to $A$ by elements of $1_{d \times d} + M_{d \times d}(\mathfrak{m}_A)$. Thus, gauge equivalence classes in $\mathrm{MC}(C^\bullet,A)$ are deformations of $\rho$ to $A$. \noindent \textbf{3.\ The homotopy Maurer--Cartan equation} There is an extension of the Maurer--Cartan equation from dg-algebras to $A_\infty$-algebras, sometimes known as the \emph{homotopy Maurer--Cartan equation}. On $\xi \in H^1 \otimes_\mathbb{F} \mathfrak{m}_A$, it has the form \begin{align*} \mathrm{MC}(&(H^\bullet,m), A) := \{x \in H^1 \otimes_\mathbb{F} \mathfrak{m}_A \mid \\ & m_1(\xi) + m_2(\xi^{\otimes 2}) - \dotsm = -\sum_{i = 1}^\infty (-1)^{\frac{i(i+1)}{2}}m_i(\xi^{\otimes i}) = 0 \text{ in } H^2 \otimes_\mathbb{F} \mathfrak{m}_A = 0\}. \end{align*} \noindent \textbf{4.\ Functorality of the Maurer--Cartan functor.} It is well-understood that Maurer--Cartan equation solutions are functorial, giving a map $\mathrm{MC}((H^\bullet,m), -) \to \mathrm{MC}(C^\bullet,-)$ arising from the quasi-isomorphism $f: H^\bullet \to C^\bullet$. Moreover, we show that quasi-isomorphisms of $A_\infty$-algebras induce isomorphisms on the sets of gauge equivalence classes of solutions to the Maurer--Cartan equation. Because the gauge action on $\mathrm{MC}((H^\bullet,m), -)$ turns out to be trivial, we find that $\mathrm{MC}(H^\bullet,-)$ represents the deformation problem for $\rho$. \noindent \textbf{5.\ Presentation for the ring representing $\mathrm{MC}((H^\bullet,m), -)$.} We now use the minimality ``$m_1 = 0$'' of the $A_\infty$-algebra structure $m$ on $H^\bullet$. In the Maurer--Cartan equation, this implies a crucial property that the equation for $C^\bullet$ lacks: the outputs of the equation are contained in $H^2 \otimes_\mathbb{F} \mathfrak{m}_A^2$. From this property, we may read off a presentation for the ring representing the $A \mapsto \mathrm{MC}((H^\bullet,m), A)$ from the maps $m_n$. Also, in order to calculate $n$th-order deformations, only the $m_i$ for $i \leq n$ are required. Altogether, we get a presentation for $R_\rho$, which appears in Theorem \ref{thm: main A-inf}. \section{Background on $A_\infty$-algebras} \label{sec: A-infinity background} In preparation to state the main results, and as this article is intended in part to acquaint number theorists with homotopy-algebraic notions, we now fully introduce the required background on $A_\infty$-algebras. We follow \cite[Ch.\ 9]{LV2012}. See \S\ref{sec:A-inf} for a more conventional introduction. \subsection{Definition via the bar construction} \label{subsec: bar def} Let $\hat T_\mathbb{F}(V)$ denote the complete free (associative) $\mathbb{F}$-algebra on a $\mathbb{F}$-vector space $V$. Let $\hat S_\mathbb{F}(V)$ denote the complete free commutative $\mathbb{F}$-algebra on $V$. A $\mathbb{F}$-basis $\{v_1, \dotsc, v_n\}$ for $V$ determines an isomorphism $\hat S_\mathbb{F}(V) \buildrel\sim\over\ra \mathbb{F}{[\![} v_1, \dotsc, v_n{]\!]}$. That is, $\hat S_\mathbb{F}(V) \cong \prod_{n \geq 0} \Sym_\mathbb{F}^n V$; similarly, $\hat T_\mathbb{F}(V) \cong \prod_{n \geq 0} V^{\otimes n}$. When $H = H^\bullet$ is a $\mathbb{Z}$-graded $\mathbb{F}$-vector space, an $A_\infty$-algebra structure is a sequence $m = (m_n)_{n \geq 1}$ of $\mathbb{F}$-linear maps \begin{equation} \label{eq: m_n} m_n : H^{\otimes n} \longrightarrow H \quad \text{of degree } 2-n, \quad \text{for } n \in \mathbb{Z}_{\geq 1} \end{equation} satisfying many compatibility relations (see \S\ref{subsec: A-inf defn} for this and further details). For the moment, we recall that when $m_n = 0$ for $n \geq 3$, the compatibility relations for $m_1, m_2$ are exactly the axioms of a differential graded algebra with differential $m_1$ and multiplication $m_2$. We introduce morphisms and quasi-isomorphisms of $A_\infty$-algebras. A \emph{morphism of $A_\infty$-algebras} $f : H \rightarrow H'$ is a sequence $f = (f_n)_{n \geq 1}$ of maps \begin{equation} \label{eq: f_n} f_n : H^{\otimes n} \longrightarrow H' \quad \text{of degree } 1-n, \quad \text{for } n \in \mathbb{Z}_{\geq 1}, \end{equation} satisfying certain relations that will be explained later. One of the relations is that $f_1$ is a morphism of complexes $(H, m_1) \rightarrow (H', m'_1)$. We call $f$ a \emph{quasi-isomorphism} when $f_1$ is a quasi-isomorphism of complexes. Both the definition of an $A_\infty$-algebra and the main result will become clearer by reformulating the notion of an $A_\infty$-algebra through a dualized version of the bar construction, which we now explain in a stepwise way. We use the notation $(-)^*$ to denote $\mathbb{F}$-linear duals. That is, $V^*$ is the $\mathbb{F}$-linear dual of a $\mathbb{F}$-vector space $V$. First, we suspend the maps $m_n$, i.e.\ we suspend $H$ using notation $\Sigma$ so that each $m_n$ induces a $\mathbb{F}$-linear map \begin{equation} \label{eq: m_n suspend} \Sigma m_n : \Sigma H^{\otimes n} := (\Sigma H)^{\otimes n} \longrightarrow \Sigma H\quad \text{ of degree } 1. \end{equation} Then, presuming that $H^i$ is finite-dimensional for all $i \in \mathbb{Z}$, we take the linear duals of the composite of \eqref{eq: m_n suspend} with the projection $\Sigma H \rightarrow \Sigma H^i$. Summing these dual maps over their domains yields \[ m_n^* : \Sigma H^* := \bigoplus_{i \in \mathbb{Z}} (\Sigma H^i)^* \longrightarrow (\Sigma H^*)^{\otimes n}, \text{ also of degree } 1. \] Next, we take the product over the codomains as $n$ varies, writing \[ m^* : \Sigma H^* \longrightarrow \prod_{n \geq 1} (\Sigma H^*)^{\otimes n}. \] Finally, we extend the domain of $m^*$ to the complete free associative graded $\mathbb{F}$-algebra $\hat T_\mathbb{F} \Sigma H^*$ via the Leibniz rule, producing \[ m^* : \hat T_\mathbb{F} \Sigma H^* \longrightarrow \hat T_\mathbb{F} \Sigma H^* \quad \text{of degree 1}. \] Altogether, we write \[ \mathrm{Bar}^*(H) = \mathrm{Bar}^*(H, m) := (\hat T_\mathbb{F} \Sigma H^*, m^*, \pi) \] for this complete free graded associative $\mathbb{F}$-algebra with derivation $m^*$, where $\pi$ denotes the standard multiplication operation. Analogously, a sequence of maps $f = (f_n)_{n \geq 1}$ as in \eqref{eq: f_n} induces a homomorphism of complete free graded associative $\mathbb{F}$-algebras \[ \mathrm{Bar}^*(f) : \mathrm{Bar}^*(H') \longrightarrow \mathrm{Bar}^*(H). \] Notice that nothing we have said so far depends on the relations on $m = (m_n)_{n \geq 1}$ defining an $A_\infty$-algebra or the relations on $f$ defining a morphism of $A_\infty$-algebras. The following statement contains a concise and equivalent formulation of these relations. \begin{fact}[Bar construction] \label{fact: bar} Assume that $H^i$ is finite-dimensional for all $i \in \mathbb{Z}$. A sequence of maps $m=(m_n)_{n \geq 1}$ as in \eqref{eq: m_n} make $(H, m)$ an $A_\infty$-algebra if and only if $\mathrm{Bar}^*(H,m)$ is a dg-algebra. That is, $m$ defines an $A_\infty$-structure on $H$ if and only if $m^*$ is a differential, i.e.\ $m^* \circ m^* = 0$. Likewise, a sequence of maps $f = (f_n)_{n \geq 1}$ as in \eqref{eq: f_n} make $f : H \rightarrow H'$ a morphism of $A_\infty$-algebras if and only if the $\mathbb{F}$-algebra homomorphism $\mathrm{Bar}^*(f)$ is a homomorphism of dg-algebras, i.e.\ $\mathrm{Bar}^*(f) \circ m'^* = m^* \circ \mathrm{Bar}^*(f)$. Moreover, \begin{enumerate}[leftmargin=2em] \item One can drop the condition that each $H^i$ is finite-dimensional and produce a co-complete co-free dg-coalgebra (see Definition \ref{defn: bar}). \item This construction induces an \emph{isomorphism} of categories between $A_\infty$-algebras and co-complete co-free dg-coalgebras. \end{enumerate} \end{fact} \begin{proof} By direct computation. See e.g.\ \cite[Lem.\ 9.2.2 and \S9.2.11]{LV2012}. \end{proof} \subsection{A theorem of Kadeishvili} \label{subsec: kad intro} The remainder of the background from homotopical algebra that we require originates in a theorem of Kadeishvili \cite{kadeishvili1982}. \begin{fact}[Kadeishvili] \label{fact: kadeish} Let $(C, d_C, m_{2,C})$ be a dg-$\mathbb{F}$-algebra. Let $H = H^\bullet(C)$ be the graded $\mathbb{F}$-vector space of cohomology of the complex $(C, d_C)$. There is an $A_\infty$-algebra structure $m = (m_n)_{n \geq 1}$ on $H$ and a quasi-isomorphism of $A_\infty$-$\mathbb{F}$-algebras \[ f = (f_n)_{n \geq 1} : (H, m) \rightarrow (C, (m_{n,C})_{n\geq 1}) \] where we let $m_{1,C} = d_C$ and $m_{n,C} = 0$ for $n \geq 3$ (the standard $A_\infty$-algebra structure arising from a dg-algebra structure). Moreover, these satisfy the properties \begin{enumerate}[leftmargin=2em] \item $m_1 = 0$, i.e.\ $(H,m)$ is minimal \item $m_2 = (m_{2,C} \mod{\mathrm{image}(d_C)})$ \item $\mathrm{pr} \circ f_1 = \mathrm{id}_H$, where $\mathrm{pr}$ is the projection from $\ker(d_C)$ to $H$. \end{enumerate} These structures are unique up to non-unique isomorphism. \end{fact} \begin{proof} See e.g.\ \cite[Cor.\ 9.4.8]{LV2012}; see also \S\ref{subsec: dg to A-inf} for more details. \end{proof} The idea is that the standard graded algebra structure $(H, m_2)$ on the cohomology $H$ of a dg-algebra $C$ can be enriched into the structure of a minimal $A_\infty$-algebra $(H,m)$ that does not lose information from $C$. In \S\ref{subsec: dg to A-inf}, we explain that the choice of a homotopy retract between $H$ and $C$ induces a particular choice of $f$ and $m$. \begin{rem} Another way of expressing Fact \ref{fact: kadeish} is that the higher $A_\infty$-products $(m_n, n \geq 3)$ encode the information lost by passing to the cohomology algebra. When no information is lost, i.e.\ there is a choice of $m$ in the statement such that $m_n = 0$ for $n \geq 3$, the dg-algebra is called \emph{formal}. The formal case makes for especially explicitly computable deformation theory (see Remark \ref{rem: cup case}). \end{rem} By way of giving an example of this fact, we introduce the Galois cohomology objects that appear in the main theorem, starting with the standard construction of Galois cohomology. Let $V$ be a $\mathbb{F}[G]$-module. Let \[ C^\bullet(G, V) \cong \bigoplus_{i \geq 0} C^i(G,V) \] denote the complex of inhomogeneous continuous cochains on the profinite group $G$, graded by degree $i$. When $V$ has the structure of a $\mathbb{F}$-algebra, one may check (see e.g.\ \cite[Prop.\ 1.4.1]{NSW2008}) that the composition $m_{2,C}$ of the standard cup product of cochains with the multiplication map $V \otimes_\mathbb{F} V \rightarrow V$, namely, \[ C^i(G, V) \otimes_\mathbb{F} C^j(G, V) \longrightarrow C^{i+j}(G,V \otimes_\mathbb{F} V) \longrightarrow C^{i+j}(G,V), \] makes $C^\bullet(G,V)$ a dg-$\mathbb{F}$-algebra. That is, the Leibniz rule is satisfied. We write $H^\bullet(G,V)$ for the graded $\mathbb{F}$-vector space of cohomology of $C^\bullet(G,V)$. The dg-algebra structure on $C^\bullet(G,V)$ induces a graded algebra multiplication \[ m_2 : H^\bullet(G,V) \otimes_\mathbb{F} H^\bullet(G,V) \longrightarrow H^\bullet(G,V) \] on $H^\bullet(G,V)$. Now we may apply Fact \ref{fact: kadeish}, producing an $A_\infty$-algebra structure $m= (m_n)_{n \geq 1}$ on $H^\bullet(G,V)$ extends the native dg-algebra structure on the graded algebra $(H^\bullet(G,V), 0, m)$. That is, it extends the usual cup product in cohomology \[ m_1 = 0, \quad m_2 = (m_{2,C} \mod{\mathrm{image}(d^C)}). \] For the purposes of this introduction, our case of interest is where $V = \mathrm{End}_\mathbb{F}(\rho)$, where $\rho$ is absolutely irreducible as in \S\ref{sssec: irred case} above. \subsection{The classical hull} \label{subsec: classical hull} Finally, we define the \emph{classical hull} of a dg-algebra. \begin{defn} \label{defn: classical hull} The classical hull $\cA(C) = \cA(C,d,m)$ of a dg-algebra $(C, d, m_{2,C})$ is the ring $\cA(C) := C^0/d(C^{-1})$ concentrated in graded degree zero, taken as a (classical) ring. This functor is left adjoint to the functor sending classical (associative) algebras $(A, m)$ to the dg-algebra $(A[0], 0, m[0])$ concentrated in degree zero as $A$ and with the zero differential. That is, \[ \mathrm{Hom}_{\text{dg-}\mathbb{F}}(C, A[0]) = \mathrm{Hom}_\mathbb{F}(\cA(C), A). \] \end{defn} Since we are most interested in the classical hull of the dg-algebra $(\hat T_F \Sigma H^\bullet, m^*, \pi)$ produced by the bar construction in Fact \ref{fact: bar}, we describe this case in particular. \begin{eg} The classical hull of the dg-algebra $\mathrm{Bar}^*(H) = (\hat T_\mathbb{F} \Sigma H^*, m^*, \pi)$ produced by the dualized bar construction of Fact \ref{fact: bar} is \[ \frac{\hat T_\mathbb{F} (\Sigma H^1)^*} {(m^*((\Sigma H^2)^*)))}. \] Indeed, notice that any map $\hat T_\mathbb{F} \Sigma H^* \rightarrow A$ factors through $\hat T_\mathbb{F} (\Sigma H^1)^* \rightarrow A$, as $(\Sigma H^1)^*$ is the degree zero part of $\Sigma H^*$. Then, calculate using the Leibniz rule that the ideal generated by the projection of $m^*(\Sigma H^*)$ to $\hat T_\mathbb{F} (\Sigma H^1)^*$ is $(m^*((\Sigma H^2)^*)))$. \end{eg} \section{Main results} \label{sec: main results} We present results toward each of the goals (1)-(4) on pg.\ \pageref{pg: twosteps}. In particular, we address deformations of pseudorepresentations in \S\ref{subsec: results2}. \subsection{Results, Part I: Determination of moduli spaces of representations} \label{subsec: results1} As in \S\ref{subsec: red case}, let $\rho : G \rightarrow \GL_d(\mathbb{F})$ be a semi-simple representation with absolutely irreducible summands $\rho_i : G \rightarrow \GL_{d_i}(\mathbb{F})$, $1 \leq i \leq r$. We write \[ \rho \cong \bigoplus_{i=1}^r \rho_i, \] thinking of this as a block diagonal decomposition of the homomorphism $\rho$. We impose the running assumption that $\rho_i \not\simeq \rho_j$ for $i \neq j$ (the \emph{multiplicity-free} condition). We call such $\rho$ a multiplicity-free \emph{residual semi-simplification}, as it turns out that the condition of deforming some $\mathbb{F}$-valued representation whose semi-simplification is $\rho$ is an open and closed condition in the moduli space of $d$-dimensional representations of $G$ valued in appropriately topologized $\mathbb{F}$-algebras. This connected component of moduli space, denoted ${\mathrm{Rep}}_\rho$, is set up in \S\ref{sec: profinite moduli}. Theorem \ref{thm: adapted pres} gives a presentation of ${\mathrm{Rep}}_\rho$ in terms of $A_\infty$-algebra structure on $H^\bullet(G, \mathrm{End}_\mathbb{F}(\rho))$. In this introduction we present the simplest case, where $\rho$ is irreducible, i.e.\ $r=1$. In this case ${\mathrm{Rep}}_\rho \cong \Spec R_\rho$, where $R_\rho$ is the usual deformation ring (see \S\ref{sssec: irred case}). We remark that the case $r=1$ has no less theoretical content than the case of general $r$, but has a simpler expression. \begin{thm}[{Special case ($r=1$) of Theorem \ref{thm: adapted pres}}] \label{thm: main A-inf} Let $\rho$ be an absolutely irreducible $\mathbb{F}$-valued representation of $G$. Assume that $H^i(G, \mathrm{End}_\mathbb{F}(\rho))$ is finite-dimensional for all $i \geq 0$. As described in Fact \ref{fact: kadeish}, there exists a minimal $A_\infty$-algebra structure $m= (m_n)_{n \geq 2}$ on $H = H^\bullet(G, \mathrm{End}_\mathbb{F}(\rho))$ that is compatible with the dg-algebra $C = C^\bullet(G, \mathrm{End}_\mathbb{F}(\rho))$ in that there exists quasi-isomorphism of $A_\infty$-algebras \[ f: (H,m) \longrightarrow (C, d_C, m_{2,C}) \] Let $(\hat T_\mathbb{F} \Sigma H^\bullet(G, \mathrm{End}_\mathbb{F}(\rho))^*, m^*, \pi)$ denote the complete dg-algebra arising from the bar construction on $(H^\bullet(G, \mathrm{End}_\mathbb{F}(\rho)), (m_n)_{n \geq 2})$. These data determine a dg-algebra homomorphism \[ \mathrm{Bar}^*(H,m) = (\hat T_\mathbb{F} \Sigma H^\bullet(G, \mathrm{End}_\mathbb{F}(\rho)), m^*, \pi) \longrightarrow R_\rho[0], \] which factors through an isomorphism of classical complete commutative algebras \[ \frac{\hat S_\mathbb{F} \Sigma H^1(G, \mathrm{End}_\mathbb{F}(\rho))^*} {(m^*(\Sigma H^2(G, \mathrm{End}_\mathbb{F}(\rho))^*))} \buildrel\sim\over\lra R_\rho, \] from the abelianization of the classical hull of $\mathrm{Bar}^*(H,m)$ to $R_\rho$. \end{thm} Conventional tangent and obstruction theory is an immediate corollary of this presentation for $R_\rho$. Let $h^i$ denote the $\mathbb{F}$-dimension of $H^i(G, \mathrm{End}_\mathbb{F}(\rho))$. \begin{cor}[Tangent and obstruction theory] \label{cor: t and o theory intro} There is a tangent and obstruction theory for deformations of $\rho$, as outlined in \S\ref{sssec: irred case}. Moreover, there is a bound on Krull dimension \[ h^1-h^2 \leq \dim(R_\rho) \leq h^1. \] \end{cor} Corollary \ref{cor: t and o theory intro} was originally proved by Mazur \cite[\S1.5, Prop.\ 2, pg.\ 399]{mazur1989}. \begin{proof}[{Proof of Corollary \ref{cor: t and o theory intro}}] The bound on Krull dimension can be read off from the presentation of $R_\rho$ in Theorem \ref{thm: main A-inf}. The tangent and obstruction theory can be derived from the presentation as follows. The formula for the tangent space arises from the fact that $m^*(\Sigma H^2(-)^*)$ is valued in the square of the maximal ideal of $\hat S_\mathbb{F} \Sigma H^1(-)^*$. As far as obstructions, one can produce a class in $H^2(G,\mathrm{End}_\mathbb{F}(\rho))$ from the composition of the presentation of $R_\rho$ with a homomorphism $R_\rho \rightarrow \mathbb{F}[\varepsilon]/\varepsilon^{n+1}$ corresponding to an $n$th-order lift $\rho_n$ of $\rho$. It turns out that this is equal to a class defined directly through the calculations of obstructions using the theory of Massey products, as described in of \S\ref{subsec: illustrate}. \end{proof} Here are some remarks that give further perspective on the presentation for $R_\rho$ in Theorem \ref{thm: main A-inf}. \begin{rem} It is implicit in the theorem statement that there is no canonical choice of $A_\infty$-structure on $H^\bullet(G, \mathrm{End}_\mathbb{F}(\rho))$. Indeed, the tangent theory of \S\ref{sssec: irred case} amounts to a canonical surjection of $\hat S_\mathbb{F} \Sigma H^1(G, \mathrm{End}_\mathbb{F}(\rho))^*$ onto $R_\rho/\mathfrak{m}_\rho^2$, but one may readily calculate by hand to observe that it has no canonical lift to $R_\rho/\mathfrak{m}_\rho^3$. \end{rem} \begin{rem} \label{rem: cup case} Given the surjection $\hat S_\mathbb{F} \Sigma H^1(G, \mathrm{End}_\mathbb{F}(\rho))^* \twoheadrightarrow R_\rho$, we see that the presentation for $R_\rho/\mathfrak{m}_\rho^3$ is determined by the projection of $m^*$ to \[ \Sigma H^2(G, \mathrm{End}_\mathbb{F}(\rho))^* \longrightarrow \Sym^2_\mathbb{F} (\Sigma H^1(G, \mathrm{End}_\mathbb{F}(\rho)^*). \] In fact, this is the $\mathbb{F}$-linear dual of the cup product map $m_2$. This recovers the connection between the cup product and obstructions to second-order deformations that was mentioned in Remark \ref{rem: intro rem 2nd order def}. This observation has been applied to great effect when the minimal $A_\infty$-structure $(m_n)_{n \geq 2}$ may be chosen so that they vanish for $n \geq 3$, which is known as the \emph{formal} case. See, for example, \cite{GM1988}, for an instance of formal deformation theory found in nature, as is discussed a bit more in \S\ref{subsec: kuranishi}. \end{rem} \begin{rem} An abelianization appears in Theorem \ref{thm: main A-inf}. Accordingly, the theorem follows from a more general theorem, Corollary \ref{cor: r-pointed dual}, which is the main theorem of Part 2. Its additional generality consists of the following. \begin{itemize}[leftmargin=2em] \item The source of the representations may be a finitely generated associative $\mathbb{F}$-algebra $E$. In particular, it need not be a Hopf algebra. \item The target coefficients of representations may be non-commutative. That is, we allow for local associative finite-dimensional $\mathbb{F}$-algebras with residue field $\mathbb{F}$ as deformation-theoretic coefficients. \item We may have $r \geq 1$. \end{itemize} Corollary \ref{cor: r-pointed dual} is a new result in non-commutative deformation theory; see \S\ref{subsec: NCDT} for more comments on related results. Corollary \ref{cor: r-pointed dual} specializes to the case $r=1$ in Corollary \ref{cor: irred case A-inf} and yields Theorem \ref{thm: main A-inf} upon applying it to $E = \mathbb{F}[G]$. \end{rem} \subsection{Geometric invariant theory} \label{subsec: GIT intro} This section prepares notation for the statement of Theorem \ref{thm: main pseudo}, which provides presentations for pseudodeformation rings. This suite of notation expresses invariant subrings that arise from geometric invariant theory (GIT) and follows on work of Bella\"iche \cite{bellaiche2012}. We will suppress inexplicit invariant-theoretic formulations in favor of explicit combinatorial expressions. For further details about translation from GIT to combinatorics, see \S\ref{subsec: apply to moduli}. Let the pseudorepresentation $D : G \rightarrow \mathbb{F}$ arise from a semi-simple representation $\rho : G \rightarrow \GL_d(\mathbb{F})$ with distinct absolutely irreducible summands \[ \rho \simeq \bigoplus_{i=1}^r \rho_i. \] \begin{note}[{\cite[\S2.2]{bellaiche2012}}] \label{note: cycles} We set up the following combinatorial objects on the integers from $1$ to $r$. \begin{itemize}[leftmargin=2em] \item Write $\bold r$ for the set $\{1, \dotsc, r\}$. \item A \emph{path} is a function $\gamma: \{0,\dotsc, l\} \rightarrow \bold r$, for some $\ell \geq 0$. We write $l_\gamma = l$ for the \emph{length} of $\gamma$. \item We say that a path $\gamma$ \emph{goes from $i$ to $j$} when $\gamma(0) = i$ and $\gamma(l_\gamma) = j$. \item We call $\gamma$ \emph{closed} if $\gamma(0) = \gamma(l_\gamma)$. In this case, we may consider the domain of $\gamma$ to be $\mathbb{Z}/l_\gamma \mathbb{Z}$. \item We call a path $\gamma$ \emph{simple} if \[ \gamma(i) = \gamma(j) \text{ and } i \neq j \implies \{i,j\} = \{0,l_\gamma\}, \] That is, a path is simple if it is injective, or it is closed and maximally injective. \item Write $SCP(l)$ for the set of simple closed paths in $\bold r$ of length $l$, and write $SCP(\bold r)$ for the set of all simple closed paths in $\bold r$ (of any length). \item A \emph{cycle} is an equivalence class of closed paths under the equivalence relation $\gamma \sim \gamma'$ defined by \[ \gamma \sim \gamma' \buildrel\text{def}\over\iff \left\{ \begin{array}{l} l_{\gamma} = l_{\gamma'} =: l, \text{ and} \\ \exists \ k \in \mathbb{Z}/l\mathbb{Z} \text{ such that } \gamma(i) = \gamma'(i+k)\ \forall\ i \in \mathbb{Z}/l\mathbb{Z}. \end{array}\right. \] \item A cycle is called \emph{simple} if one (equivalently, all) of its constituent closed paths is simple. \item Write $SC(l)$ for the set of simple cycles in $\bold r$ of length $l$, and write $SC(\bold r)$ for the set of all simple cycles in $\bold r$ (of any length). \item For $i,j \in \bold r$, write $SCC(i,j)$ for the set of paths $\gamma$ from $i$ to $j$ such that the concatenation of $\gamma$ with the length 1 path from $j$ to $i$ is a simple closed cycle. In particular, $i = j$ is allowed, but $SCC(i,i) = \varnothing$ in this case. (``SCC'' stands for ``simple closed complements.'') \item Given a cycle represented by a closed path $\gamma$ and $i,j \in \{0, 1, \dotsc, l_\gamma-1\}$, let $SCC_\gamma(i,j)$ be the subpath of $\gamma$ from $\gamma(i)$ to $\gamma(j)$. When $i < j$, this is the path given by the sequence $(\gamma(i), \gamma(i+1), \dotsc, \gamma(j))$. When $i> j$, the situation is similar, but note that this subpath of $\gamma$ (thought of as a loop) includes the sequence $(\gamma(l_\gamma-1), \gamma(0))$ in that order. When $i = j$, $SCC_\gamma(i,j)$ is taken to be the empty path. \end{itemize} \end{note} In the following notation, we express group cohomology as $\mathrm{Ext}$-groups for convenience. These expressions are all contained in the canonical isomorphism \[ \mathrm{Ext}^\bullet_{\mathbb{F}[G]}(\rho,\rho) \cong H^\bullet(G, \mathrm{End}_\mathbb{F}(\rho)). \] For more on this, see \S\ref{subsec: hochs-group}. \begin{note} \label{note: Ext} The following objects enrich the foregoing notation from finite sets to finite-dimensional $\mathbb{F}$-vector spaces coming from $\mathrm{Ext}$. Recall that $(-)^*$ refers to $\mathbb{F}$-linear duality. \begin{itemize}[leftmargin=2em] \item For $i,j \in \bold r$, let \[ \mathrm{Ext}_G^k(j,i) := \mathrm{Ext}^k_{\mathbb{F}[G]}(\rho_j, \rho_i) \cong H^k(G, \rho_i \otimes_\mathbb{F} \rho_j^*). \] \item Given a path $\gamma$ on $\bold r$, we write \[ \Sigma \mathrm{Ext}^1_G(\gamma)^* := \bigotimes_{i=0}^{l_\gamma-1} \left(\Sigma \mathrm{Ext}^1_G(\gamma(i), \gamma(i+1))^*\right), \] where the order of tensor factors may matter. When $\gamma$ is a cycle, we use this same notation only when we are working with symmetric tensors, so that there is no dependence upon the choice of closed path in the cycle. \item Let $\cC(D)$ be the directed graph whose vertices $\{\rho_i\}_{i=1}^r$ and whose arrows from $\rho_i$ to $\rho_j$ are a choice of basis for $\mathrm{Ext}^1_G(\rho_j, \rho_i)$. We will refer to the property of being strongly connected, i.e.\ the existence of a directed path between any two vertices, as well as the decomposition into strongly connected components. \item Let $h_{ij}^k := \dim_\mathbb{F} \mathrm{Ext}^k_G(\rho_j, \rho_i)$. We will mainly use $k = 1, 2$. \item Let $H_1(\cC(D))$ be the simplicial homology of the simplicial 1-complex naturally arising from $\cC(D)$. Let $\mathbb{N}\langle SC(\cC(D))\rangle$ be the free commutative monoid on simple cycles $\cC(D)$. Let $J$ be the kernel of its natural additive injection to $H_1(\cC(D))$. That is, $J$ consists of $\mathbb{N}$-linear combinations of simple cycles that have the same underlying sets-with-multiplicity of arrows. Finally, let $h^2(\cC(D))$ be the set \[ J \smallsetminus \left(J \cdot (\mathbb{N}\langle SC(\cC(D))\rangle \smallsetminus\{0\})\right) \] and let $H^2(\cC(D))$ be the $\mathbb{F}$-vector space with basis $h^2(\cC(D))$. That is, $h^2(\cC(D))$ is a minimal generating set for $J$ as a submonoid of $\mathbb{N}\langle SC(\cC(D))\rangle$. \end{itemize} \end{note} Finally, we define an invariant subring that will appear often in our presentations for pseudodeformation rings. \begin{defn} \label{defn: R1D} Let $R^1_D$ denote the local ring that is the image of the map \begin{equation} \label{eq: R1D} \hat S_\mathbb{F} \displaystyle\bigoplus_{\gamma \in SC(\bold r)} \Sigma\mathrm{Ext}^1_G(\gamma)^* \longrightarrow \hat S_\mathbb{F} \displaystyle\bigoplus_{i,j \in \bold r} \Sigma\mathrm{Ext}^1_G(j,i)^* \cong \hat S_\mathbb{F} \Sigma \mathrm{Ext}_{\mathbb{F}[G]}(\rho,\rho)^*. \end{equation} arising natural from the canonical map from $\Sigma \mathrm{Ext}^1_G(\gamma)^*$ to the codomain. \end{defn} \begin{rem} The image of \eqref{eq: R1D} equals the invariant subring of the codomain induced by the canonical action of the torus ${\mathbb{G}}_m^{\times r}$ on the vector space $\mathrm{Ext}^1_G(j,i)$ by the character labeled by the $r$-tuple $(a_k)_{k=1}^r \in \mathbb{Z}^{\oplus r} \cong X^*({\mathbb{G}}_m^{\times r})$ determined by \[ a_k = \delta_{ik} - \delta_{jk} \] (using the Kronecker delta). That is, thinking of ${\mathbb{G}}_m^{\times r}$ as the standard torus in $\GL_r$, its action on $\mathrm{Ext}^1_G(j,i)$ appears as its action on the $(i,j)$th coordinate of $r$-by-$r$ matrices by the adjoint action. \end{rem} We will supply references for the following results of GIT in \S\ref{subsec: invariants quivers}. \begin{fact} \label{fact: R1D} $R^1_D$ is reduced, normal, and Cohen-Macaulay. If $\cC(D)$ decomposes into strongly connected components $\coprod \cC(D_a)$ where $D = \bigoplus_a D_a$, then there is a canonical isomorphism \[ R^1_D \cong \hat\bigotimes_a R^1_{D_a}, \] to a completed tensor product over $\mathbb{F}$. When $\cC(D)$ is strongly connected, then its Krull dimension is \[ \dim R^1_D = 1 - r + \sum_{1 \leq i, j \leq r} h^1_{ij}. \] \end{fact} We substantiate this next basic fact in \S\ref{subsec: obs theory}. \begin{fact} \label{fact: basic obstruction} Let $K$ denote the kernel of \eqref{eq: R1D} and let $\mathfrak{m}$ the maximal ideal of the codomain. There exists a canonical isomorphism \[ H_2(\cC(D))^* \buildrel\sim\over\lra K/\mathfrak{m} K. \] \end{fact} \begin{rem} The construction $\cC(D)$ is a \emph{quiver}, giving us access to the extensive literature studying representations of quivers and of quivers with relations. For more on this, see \S\ref{subsec: invariants quivers}. \end{rem} \subsection{Results, Part II: determination of moduli spaces of pseudorepresentations} \label{subsec: results2} Let $D = \psi(\rho) : G \rightarrow \mathbb{F}$ be the $d$-dimensional pseudorepresentation induced by $\rho$, as in \S\ref{subsec: red case}. The proofs of the following results are found in \S\S\ref{subsec: PsDef proofs}-\ref{subsec: obs theory}. Here is the main theorem. \begin{thm} \label{thm: main pseudo} Let $D := \psi(\rho)$, where $\rho$ has $r$ distinct absolutely irreducible factors $\rho \simeq \bigoplus_{i=1}^r \rho_i$. Assume that $H^i(G, \mathrm{End}_\mathbb{F}(\rho))$ is finite-dimensional for all $i \geq 0$. Choose a structure of $A_\infty$-algebra $m=(m_n)_{n \geq 1}$ on $H^\bullet(G, \mathrm{End}_\mathbb{F}(\rho))$ and a quasi-isomorphism to the dg-algebra $(C^\bullet(G, \mathrm{End}_\mathbb{F}(\rho)), d_C, m_{2,C})$, as described in Fact \ref{fact: kadeish}, satisfying the condition of compatibility with the decomposition \[ \mathrm{End}_\mathbb{F}(\rho) \cong \bigoplus_{i,j \in \bold r} \mathrm{Hom}_\mathbb{F}(\rho_i, \rho_j) \] explained in Example \ref{eg: r-pointed retract}. The choices above induce an isomorphism \begin{equation} \label{eq: R_D pres} \frac{R^1_D}{ \Bigg(\displaystyle\bigoplus_{i,j \in \bold r} m^* \Sigma \mathrm{Ext}^2_G(\rho_j, \rho_i)^* \otimes \Big(\displaystyle\bigoplus_{\gamma \in SCC(i,j)} \Sigma \mathrm{Ext}^1_G(\gamma)^*\Big)\Bigg) } \buildrel\sim\over\lra R_D. \end{equation} \end{thm} \begin{rem} Because the formula for $R_D$ is rather complex, we supply the following intuitive interpretation. \textit{Generators are cycles.} Only \emph{cycles} $\mathrm{Ext}^1_G(\gamma)^*$ of tensors of $\Sigma \mathrm{Ext}^1_G(\rho_j,\rho_i)^*$ ($i,j \in \bold r$) will be detected by pseudorepresentations. Indeed, a choice of extension class \[ e = (e_{i,j}) \in \mathrm{Ext}^1(\rho,\rho) \cong \bigoplus_{i,j \in \bold r} \mathrm{Ext}^1(\rho_j, \rho_i) \] defines a first order deformation of the form (we take $r = 3$ for concreteness) \[ \rho_e := \rho + \varepsilon e \cong \begin{pmatrix} \rho_1 + \varepsilon e_{11} & \varepsilon e_{12} & \varepsilon e_{13} \\ \varepsilon e_{21} & \rho_2 + \varepsilon e_{22} & \varepsilon e_{23} \\ \varepsilon e_{31} & \varepsilon e_{32} & \rho_3 + \varepsilon e_{33} \end{pmatrix} \] and only products of the elements $e_{ij}$ over a cycle will appear in the diagonal, and thereby be detectable by the trace. The simple cycles generate the monoid of cycles, and $R^1_D$ is generated by these cycles. \textit{Relations are obstructed sub-paths of cycles.} However, not every cycle has factors that can multiply together and still form a homomorphism that is detectable by a central function. The obstructions to the appearance of a cycle represented by $\gamma$ consist precisely of elements of $\mathrm{Ext}^2_G(\rho_j,\rho_i)$ that are the image of $m_n$ on any sub-path $\gamma'$ from $j$ to $i$ of the cycle $\gamma$, that is \[ m_n : \mathrm{Ext}^1_G(\rho_j, \rho_{\gamma(1)}) \otimes \dotsm\buildrel{n = l_{\gamma'} \text{ factors}}\over\dotsm\dotsm \otimes \mathrm{Ext}^1_G(\rho_{\gamma(n-1)}, \rho_i) \longrightarrow \mathrm{Ext}^2_G(\rho_j, \rho_i). \] This is the $m^* \Sigma \mathrm{Ext}^2_G(...)^*$-factor of the denominator of \eqref{eq: R_D pres}. The rest of the denominator accounts for the complement of $\gamma'$ in $\gamma$; that is, we must complete the obstructed path $\gamma'$ to the cycle $\gamma$ to calculate its influence on pseudorepresentations. \end{rem} \begin{rem} This expression for $R_D$ decomposes into the strongly connected components of the directed graph $\cC(D)$ of Notation \ref{note: Ext}. In that notation, we have \[ R_D \cong \bigotimes_a R_{D_a}. \] This is consonant with the fact that each cycle is supported on exactly one strongly connected component. Because this decomposition does not simplify the formulas, we do not use it in the expression of Theorem \ref{thm: main pseudo} and its corollaries. \end{rem} A tangent and obstruction theory can be derived from Theorem \ref{thm: main pseudo}. First we discuss the tangent space. In what follows, we maintain the assumptions of Theorem \ref{thm: main pseudo}. Let \[ \mathfrak{t}_D := (\mathfrak{m}_D/\mathfrak{m}_D^2)^* \] denote the tangent space of $R_D$. \begin{cor}[The tangent space] \label{cor: tangent pseudo} The tangent space $\mathfrak{t}_D$ is isomorphic to \[ \bigoplus_{\gamma \in SC(\bold r)} \ker\left( \Sigma\mathrm{Ext}^1_G(\gamma) \longrightarrow \bigoplus_{0\leq i, j < l_\gamma} \Sigma \mathrm{Ext}^2_G(\rho_{\gamma(j)}, \rho_{\gamma(i)}) \otimes \Sigma \mathrm{Ext}^1_G(SCC_\gamma(i,j))\right) \] where $\gamma$ is taken to be a representative for a simple cycle, and the map parameterized by $(\gamma, i,j)$ sends \[ e(\gamma) := e_0 \otimes \dotsm e_{l_\gamma} \mapsto m_{d_\gamma(i,j)+1}(e_j \otimes e_{j+1} \otimes \dotsm \otimes e_i) \otimes e(SCC_\gamma(i,j)). \] Here $e(SCC_\gamma(i,j))$ denotes the tensor factor of $e(\gamma)$ indexed by $SCC_\gamma(i,j)$, and $d_\gamma(i,j)$ is the number of steps from $\gamma(j)$ to $\gamma(i)$ around $\gamma$. \end{cor} \begin{warn} \label{warn: NC pseudo} It is very important to keep in mind that direct sum expressions in the statement of Theorem \ref{thm: main pseudo} and Corollary \ref{cor: tangent pseudo} are non-canonical expressions of $R_D$ and $\mathfrak{t}_D$. Instead, there is a canonical filtration on $\mathfrak{t}_D$, whose graded pieces can be found among the summands in Corollary \ref{cor: tangent pseudo}. We now explain this canonical filtration. \end{warn} There is a canonical filtration of the tangent space, the ``complexity filtration'' of Bella\"iche \cite[\S3]{bellaiche2012}. This is an increasing filtration, where lower complexity degree corresponds to greater reducibility (or less irreducibility), in the sense of ideals of reducibility defined in Bella\"iche--Chenevier \cite[\S1.5.1]{BC2009} and the derivative notion of complexity of a pseudodeformation of \cite[\S2.4]{bellaiche2012}. Correspondingly, a lower bound on reducibility (equivalently, an upper bound on complexity) produces a closed condition in $\Spec R_D$. In the terms of the presentation for the tangent space given in Corollary \ref{cor: tangent pseudo}, the complexity degree equals the number of tensor factors, i.e.\ the length of $\gamma$. So we may index it from $0$ to $r$ as \[ 0 = \mathrm{Fil}_0 \mathfrak{t}_D \subset \mathrm{Fil}_1 \mathfrak{t}_D \subset \dotsm \subset \mathrm{Fil}_{r-1} \mathfrak{t}_D \subset \mathrm{Fil}_r \mathfrak{t}_D = \mathfrak{t}_D. \] \begin{cor} \label{cor: CFP} The $k$th graded factor of the complexity filtration on $\mathfrak{t}_D$ is canonically isomorphic to the summand of the expression in Corollary \ref{cor: tangent pseudo} labeled by $\gamma \in SC(\bold r)$ such that $l_\gamma = k$. That is, there is a canonical isomorphism \begin{align*} &\frac{\mathrm{Fil}_k \mathfrak{t}_D}{\mathrm{Fil}_{k-1}\mathfrak{t}_D} \cong \\ &\bigoplus_{\substack{\gamma \in SC(\bold r) \\ l_\gamma = k}} \ker\left(\Sigma\mathrm{Ext}^1_G(\gamma) \to \bigoplus_{0\leq i, j < k} \Sigma \mathrm{Ext}^2_G(\rho_{\gamma(j)}, \rho_{\gamma(i)}) \otimes \Sigma \mathrm{Ext}^1_G(SCC_\gamma(i,j))\right) \end{align*} where the map is as in Corollary \ref{cor: tangent pseudo}. \end{cor} \begin{rem} Implicit in the statement is the fact that the kernel does not depend on the choice of $A_\infty$-structure on $\mathrm{Ext}^\bullet_G(\rho,\rho)$. \end{rem} \begin{rem} \label{rem: refine Bel12} This refines the main theorem of \cite{bellaiche2012}: Thm.\ 1 of \textit{loc.\ cit}.\ states that $\mathrm{Fil}_k \mathfrak{t}_D/\mathrm{Fil}_{k-1} \mathfrak{t}_D$ injects into a sum of kernels that is similar to the expression above, but lacks the $A_\infty$-products $m_n$ for $n \geq 3$. That is, only the terms arising from cup products are used in \textit{loc.\ cit}. \end{rem} We derive bounds on the tangent dimension of $R_D$ from its presentation. Let \[ h^k_{ij} := \dim_\mathbb{F} \mathrm{Ext}^k_{\mathbb{F}[G]}(\rho_j, \rho_i) \quad \text{for } 1 \leq i,j \leq r, \] and let \[ h^1(\gamma) := \prod_{0 \leq i < l_\gamma} h^1_{\gamma(i), \gamma(i+1)} = \dim_\mathbb{F} \mathrm{Ext}^1_G(\gamma) \quad \text{for } \gamma \in SC(\bold r). \] \begin{cor}[Tangent dimension] \label{cor: pseudo tangent dim} The dimension of $\mathfrak{t}_D$ satisfies \[ \sum_{\gamma \in SC(\bold r)} \left(h^1(\gamma) - \sum_{0 \leq i, j < l_\gamma} h^2_{ij} \cdot h^1(SCC_\gamma(i,j)) \right) \leq \dim_\mathbb{F} \mathfrak{t}_D \leq \sum_{\gamma \in SC(\bold r)} h^1(\gamma). \] \end{cor} \begin{rem} In some cases, the lower bound can be improved due to symmetry. For example, for a 2-dimensional Galois representation $\rho$ that is a direct sum of two characters $\rho_1$ and $\rho_2$, we have that \[ \mathrm{Ext}^2_{\mathbb{F}[G]}(\rho_1, \rho_1) \cong \mathrm{Ext}^2_{\mathbb{F}[G]}(\rho_2, \rho_2) \cong H^2(G, \mathbb{F}), \] and there is a symmetry in the cup products, as follows. For $b \in \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho_2, \rho_1)$ and $c \in \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho_1, \rho_2)$, \[ b \cup c = 0 \in \mathrm{Ext}^2_{\mathbb{F}[G]}(\rho_2, \rho_2) \iff c \cup b = 0 \in \mathrm{Ext}^2_{\mathbb{F}[G]}(\rho_1, \rho_1). \] \end{rem} Next we present an obstruction theory, expressed for deformations to $\mathbb{F}[\varepsilon]/(\varepsilon^{n+1})$. \begin{cor}[Obstruction theory] \label{cor: obstruction} Via the presentation of $R_D$ in Theorem \ref{thm: main A-inf}, there is associated to an $n$th-order pseudodeformation $D_n : G \rightarrow \mathbb{F}[\varepsilon]/\varepsilon^{n+1}$ of $D$ \begin{enumerate}[leftmargin=2em] \item an element of $\alpha(D_n) \in H_2(\cC(D))$ arising from the map $R^1_D \rightarrow \mathbb{F}[\varepsilon]/\varepsilon^{n+1}$ associated to $D_n$. \item If $\alpha(D_n) = 0$, there is an element $\beta(D_n)$ in \[ \bigoplus_{i,j \in \bold r} \mathrm{Ext}^2_G(\rho_j, \rho_i) \otimes \Big( \bigoplus_{\gamma \in SCC(i,j)} \mathrm{Ext}^1_G(\gamma)\Big) \] associated to $D_n$. \end{enumerate} Moreover, $\alpha(D_n)$ and $\beta(D_n)$ vanish if and only if $D_n$ extends to an $(n+1)$st order pseudodeformation. \end{cor} \begin{rem} \label{rem: bounded deg obs} There exists some $n_0 \in \mathbb{Z}_{\geq 1}$, dependent only on $r$ and $h^1_{ij}$ for $1 \leq i,j \leq r$, such that $\alpha(D_n)$ vanishes for all integers $n \geq n_0$. \end{rem} When $H^2(G, \mathrm{End}_\mathbb{F}(\rho)) \cong \mathrm{Ext}^2_G(\rho,\rho) = 0$, the deformation theory of $\rho$ is smooth, or ``unobstructed.'' This well-known phenomenon is visible in Theorem \ref{thm: main A-inf} in the case $r=1$ (i.e.\ $\rho$ is irreducible), and remains the case for general $r \geq 1$ when we study the stack of representations ${\mathrm{Rep}}_\rho$ mentioned in \S\ref{subsec: results1}. We call the case $\mathrm{Ext}^2_{\mathbb{F}[G]}(\rho,\rho) = 0$ the ``representation-unobstructed case'' for clarity, when our focus is on the deformation theory of pseudorepresentations. When $r > 1$, it is possible for $R_D$ to be non-regular even when $\mathrm{Ext}^2_G(\rho,\rho) = 0$; actually, $R_D$ is not regular for generic choices of the integers $h^1_{ij}$. But we know from invariant theory that some ring-theoretic properties hold. \begin{cor}[The representation-unobstructed case] \label{cor: RU case} Assume $\mathrm{Ext}^2_G(\rho,\rho) = 0$, i.e.\ $H^2(G, \mathrm{End}_\mathbb{F}(\rho)) = 0$. Then the surjection $R^1_D \twoheadrightarrow R_D$ of Theorem \ref{thm: main pseudo} is an isomorphism. In particular, we may read off the following properties of $R^1_D$. \begin{enumerate}[leftmargin=2em] \item There exists an isomorphism \[ \mathfrak{t}_D \buildrel\sim\over\lra \bigoplus_{\gamma \in SC(\bold r)} \Sigma\mathrm{Ext}^1_G(\gamma), \text{ and} \] \item $R_D$ is reduced, normal, and Cohen-Macaulay. \item When $\cC(D)$ is strongly connected, the Krull dimension of $R_D$ is \[ \dim R_D = 1 - r + \sum_{i,j \in \bold r} h^1_{ij}. \] \item When $\cC(D)$ is not strongly connected, then $R_D \cong \otimes_a R_{D_a}$ and $\dim R_D = \sum_a \dim R_{D_a}$, where $D = \bigoplus_a D_a$ is the decomposition of $D$ into strongly connected summands. \item An $n$th-order pseudodeformation $D_n$ of $D$ extends to an $(n+1)$st order pseudodeformation if and only if the obstruction class $\alpha(D_n)$ of Corollary \ref{cor: obstruction} vanishes. That is, the obstruction $\beta(D_n)$ is always zero, when it exists. \end{enumerate} \end{cor} \begin{proof} The first statement is clear in Theorem \ref{thm: main pseudo}. The second statement follows the theorem combined with Fact \ref{fact: R1D}. \end{proof} \begin{rem} While $R^1_D$ is Cohen-Macaulay, for general $r$ and $h^1_{ij}$, it is very rare for $R^1_D$ to be Gorenstein, much less complete intersection or regular. See the discussion of \S\ref{subsec: invariants quivers}. But for small $r$ and $h^1_{ij}$, there are some cases where these ring-theoretic properties hold, and they are well-understood: see Example \ref{eg: R1D LCI}. \end{rem} \begin{rem} The Taylor--Wiles method \cite{TW1995}, and subsequent developments, involve auxiliary deformation problems where one arranges for $H^2(\text{aux},\mathrm{End}_\mathbb{F}(\rho))$ to vanish. This often goes under the moniker ``killing the dual Selmer group.'' This has the effect of making deformation rings $R_\rho^\mathrm{aux}$ isomorphic to a power series ring. We see in Corollary \ref{cor: RU case} what can be deduced about pseudodeformation rings $R_D$ from killing the dual Selmer group. This should be compared with the philosophy that, in situations where ``$\ell_0 = 0$'' so that the Taylor--Wiles method could possibly be applied (see e.g.\ \cite{CG2018} for more on this, including the definition of $\ell_0$), local Galois deformation rings -- when properly set up to correspond to Hecke algebras -- ought to be at least Cohen-Macaulay. See e.g.\ \cite[Thm.\ 4.6.2]{snowden2018}. \end{rem} We conclude with bounds on the Krull dimension of $R_D$. \begin{cor}[Bounds on Krull dimension] \label{cor: pseudo krull bounds} Assume for simplicity that $\cC(D)$ is strongly connected. Then we have the following bounds on the Krull dimension of $R_D$. Letting $h^1_D := 1 - r + \sum_{i,j \in \bold r} h^1_{ij} = \dim R^1_D$, we have \[ h^1_D - \sum_{\substack{\gamma \in SC(\bold r)\\ 0 \leq i,j < l_\gamma}} h^2_{ij} \cdot h^1(\gamma') \leq \dim R_D \leq h^1_D \] \end{cor} \begin{rem} Note that $h^1(\gamma')$ is multiplicative in the $h^1_{ij}$, while $h^1_D$ is additive in the $h^1_{ij}$. Therefore, for large dimensions of $\mathrm{Ext}^1$-groups in the presence of non-zero $\mathrm{Ext}^2$-groups, this lower bound on $\dim R_D$ is trivial. \end{rem} \begin{rem} We remark on the computations involved in the proof of Theorem \ref{thm: main pseudo}. The ring $R^1_D$ is the invariant subring of the codomain of \eqref{eq: R1D} under the adjoint co-action of the $\mathbb{F}$-algebraic torus related to units of \[ \mathrm{End}_{\mathbb{F}[G]}(\rho) \cong \bigoplus_{i=1}^r \mathrm{End}_{\mathbb{F}[G]}(\rho_i) \cong \mathbb{F}^r. \] Similarly, as this action is linearly reductive (in any characteristic), the presentation of $R_D$ in Theorem \ref{thm: main pseudo} follows via a calculation of invariants from Theorem \ref{thm: adapted pres}, which generalizes Theorem \ref{thm: main A-inf} to the case that $\rho$ is semi-simple with distinct simple factors. \end{rem} \subsection{Amplification: Galois representations with conditions} \label{subsec: Gal conditions} In this section, we state a meta-result: all of the theorems and corollaries of \S\ref{subsec: results1} and \S\ref{subsec: results2} may be applied to deformation rings and pseudodeformation rings parameterizing Galois representations \textit{satisfying additional Galois-theoretic conditions} of certain kinds. Let $\cC$ be a condition that applies to finite-length $\mathbb{F}[G]$-modules. We say that $\cC$ is a \emph{stable} condition when the full subcategory of finite-length $\mathbb{F}[G]$-modules satisfying $\cC$ is closed under the formation of subquotients and finite direct sums. It has been understood, since the work of Ramakrishna \cite{ramakrishna1993}, that there exists a quotient $R_\rho \twoheadrightarrow R^\cC_\rho$ parameterizing exactly those deformations with property $\cC$. For stable $\cC$, the author's joint work with Wake \cite{WWE4} (see \S\ref{subsec: stable or CH} for a summary) explains that \begin{enumerate}[leftmargin=2em] \item there exist quotient algebras of $\mathbb{F}[G]$ factoring the action on representations with residual pseudorepresentation $D$ and condition $\cC$, and \item there exists a sensible notion of ``pseudorepresentation of $G$ with property $\cC$'' and a quotient $R_D \twoheadrightarrow R_D^\cC$ parameterizing exactly those pseudodeformations with property $\cC$. \end{enumerate} In the following theorem statement, we use a quotient algebra $E^\cC$ of $\mathbb{F}{[\![} G {]\!]}$ that is constructed in \S\ref{subsec: stable constr}, which has the property that it factors the action map on exactly those finite-length $\mathbb{F}[G]$-modules that satisfy condition $\cC$. \begin{thm} \label{thm: conds} Let $\cC$ be a stable condition on finite length $\mathbb{F}[G]$-modules. Let $\rho$ be a $\mathbb{F}$-linear representation of $G$ that is semi-simple with distinct absolutely irreducible factors satisfying $\cC$. Then the dg-sub-algebra \[ C^\bullet(E^\cC, \mathrm{End}_\mathbb{F}(\rho)) \subset C^\bullet(G, \mathrm{End}_\mathbb{F}(\rho)), \] and a $A_\infty$-algebra structure of Fact \ref{fact: kadeish} on its cohomology $H^\bullet(E^\cC, \mathrm{End}_\mathbb{F}(\rho))$ that induces \begin{enumerate}[leftmargin=2em] \item when $\rho$ is irreducible, a presentation of $R^\cC_\rho$, as in Theorem \ref{thm: main A-inf}, and \item when $D$ is the pseudorepresentation induced by $\psi$, a presentation of $R^\cC_D$, as in Theorem \ref{thm: main pseudo}. \end{enumerate} \end{thm} All of the corollaries to Theorems \ref{thm: main A-inf} and \ref{thm: main pseudo} also apply in the condition $\cC$ context of Theorem \ref{thm: conds}, as they hinge only on the existence of the $A_\infty$-algebra structure on cohomology and its relation to the deformation rings. \begin{eg} Stable conditions $\cC$ of interest in the study of Galois representations include such conditions as \begin{enumerate}[leftmargin=2em] \item when $G = G_F$ is the absolute Galois group of a $p$-adic field $F$, or when $G$ admits a homomorphism $G_F \rightarrow G$, we ask for the property that the $G_F$-action arises from the $\overline{F}$-points of a finite flat group scheme defined over the ring of integers $O_F$ of $F$. \item More generally than (1) when $F/\mathbb{Q}_p$ is unramified, there are Fontaine--Laffaille conditions, which are $p$-integral crystalline conditions. \end{enumerate} \end{eg} \begin{rem} \label{rem: Ext2 for C} For a choice of subcategory $\cC$ of the category of finite-length $\mathbb{F}[G]$-modules, there may exist a notion of $\mathrm{Ext}^\bullet_\cC(\rho, \rho')$ for $\rho, \rho' \in \cC$. We emphasize that the $\mathrm{Ext}^\bullet_{E^\cC}(\rho, \rho)$ above may not be the same as $\mathrm{Ext}^\bullet_\cC(\rho,\rho')$ in all degrees. It is, however, the same in degrees 0 and 1. In degree 2, we have \[ \mathrm{Ext}^2_{E^\cC}(\rho, \rho) \subset \mathrm{Ext}^2_\cC(\rho, \rho). \] Nonetheless, the $A_\infty$-structure on the Hochschild cohomology $H^\bullet(E^\cC,\mathrm{End}_\mathbb{F}(\rho))$ correctly calculates $R_\rho^\cC$ and $R_D^\cC$. In fact, this is a general observation: isomorphism in degrees 0 and 1 and injection in degree 2 result in the same classical deformation problem; see e.g.\ \cite[Thm.\ 2.4]{GM1988}. We explain how this difference arises in Remark \ref{rem: Ext2 discrepancy}. \end{rem} \begin{rem} It is possible to produce an version of $E^\cC$ for some conditions $\cC$ which are of arithmetic importance but are not stable, such as the ``ordinary'' condition on 2-dimensional representations and pseudorepresentations studied in \cite{mazur1989, WWE1, CS2019}. We describe this in \S\ref{subsec: stable or CH}. \end{rem} \section{Complements} \label{sec: complements} In this section we discuss an alternate formulation of the main theorems in terms of Massey products, examples that illustrate the main theorems and computations of the paper, relationships with other works in number theory, related and/or antecedent works outside number theory, and acknowledgements of the influence of colleagues. Here is a list of contents. \begin{itemize}[leftmargin=3em] \item[\S\ref{subsec: massey intro}] Massey products and their relationship to $A_\infty$-products, and previous appearances of Massey products in number theory in work of Sharifi \cite{sharifi2007}. \item[\S\ref{subsec: eg NT}] The computations of ranks of $p$-adic modular Hecke algebras in terms of $A_\infty$-products that appear in \S\ref{sec: ranks}. \item[\S\ref{subsec: NRIT}] The case of the trivial representation. \item[\S\ref{subsec: GV2018}] The derived Galois deformation rings of Galatius--Venkatesh \cite{GV2018}, and the analogue of the cotangent complex in our setting. \item[\S\ref{subsec: NCDT}] Non-commutative geometry and deformation theory. \item[\S\ref{subsec: kuranishi}] The Kuranishi map. \item[\S\ref{subsec: acknow}] Acknowledgements. \end{itemize} In particular, we point out in \S\ref{subsec: NCDT} how the content of Part 2 of this paper relies on and also advances the line of inquiry in non-commutative deformation theory pursued by Laudal \cite{laudal2002} and Segal \cite{segal2008}. \subsection{Massey products} \label{subsec: massey intro} Massey products and their defining systems provide an alternative to products comprising $A_\infty$-structures for the purposes of this paper. (See \S\ref{sec: massey} for an introduction to Massey products.) This is true in a formal sense: the main theorems stated above are the outcome of Part 3 of this paper, which in turn relies on results in non-commutative deformation theory in Part 2. The main result of Part 2, Corollary \ref{cor: r-pointed dual}, gives a presentation of the completion $\mathbb{F}[G]^\wedge_\rho$ of $\mathbb{F}[G]$ at the kernel of $\rho$ in terms of $A_\infty$-structures and certain choices of idempotents. In comparison, the main theorem of \cite{laudal2002} gives an expression of the same algebra in terms of Massey products. We make further comparisons with previous work of Laudal and Segal in \S\ref{subsec: NCDT}. The definition of a Massey product, their defining systems, and the relationship between these and $A_\infty$-algebras is given in \S\ref{sec: massey}. We explain how Massey products determine non-commutative deformation theory in \S\ref{sec: lifts-massey}. The emphasis of \S\ref{sec: lifts-massey} differs from the rest of the paper: we illustrate how Massey products naturally arise when doing explicit computations of deformations in a continuation of \S\ref{subsec: illustrate}. Since any $A_\infty$-product may be realized as a Massey product, it also serves to concretely illustrate the influence of $A_\infty$-products on deformation theory. We discuss other studies of Massey products in Galois cohomology in \S\ref{subsec: NRIT}. Also, Massey products in the Galois cohomology of an endomorphism algebra has been connected to deformations of Galois representations and ranks of Hecke algebras in \cite{WWE3}. We discuss this next. \subsection{Ranks of Hecke algebras} \label{subsec: eg NT} In \S\ref{sec: ranks}, as an example of an application of our main results, we determine the ranks of some $p$-adic modular Hecke algebras in terms of the presentations given by $A_\infty$-products. This relies on a known isomorphism $R^\cC_\star \buildrel\sim\over\ra \mathbb{T}/\mathfrak{m} \mathbb{T}$, where \begin{itemize}[leftmargin=2em] \item $\mathbb{T}$ is the Hecke algebra in question, \item $\mathfrak{m}$ is the maximal ideal, with residue field $\mathbb{F}$, of the regular local base ring over which $\mathbb{T}$ is known to be free and for which we are measuring the rank of $\mathbb{T}$, so we can calculate $\mathrm{rank}\, \mathbb{T} = \dim_\mathbb{F} \mathbb{T}/\mathfrak{m} \mathbb{T} = \dim_\mathbb{F} R^\cC_\rho$ \item ``$\star$'' in $R^\cC_\star$ stands in for either \begin{itemize}[leftmargin=2em] \item $\rho$, a 2-dimensional absolutely irreducible $\mathbb{F}$-valued representation of a global Galois group $G$, which the residual semi-simplification associated to the Hecke eigensystem modulo $p$ cut out by the maximal ideal of $\mathbb{T}$; or \item $D$, a 2-dimensional pseudorepresentation given by $D = \psi(\rho)$ for some 2-dimensional representation $\rho$ of $G$ such that $\rho \cong \chi_1 \oplus \chi_2$ where $\chi_1 \neq \chi_2$, determined similarly by the residual Hecke eigensystem of $\mathbb{T}$ \end{itemize} \item $\cC$ is a condition on finite-length $\mathbb{Z}_p[G]$-modules \item $R^\cC_\star$ is a deformation (resp.\ pseudodeformation) ring of $\rho$ with condition $\cC$ \end{itemize} This rank is an expression of size of a congruence class of modular eigenforms modulo $p$. The result of each example given in \S\ref{sec: ranks} is an expression of $\mathrm{rank}\, \mathbb{T}$ in terms of an arithmetic invariant expressed in terms of the vanishing of $A_\infty$-products (or, equivalently, Massey products). The first two examples that we give in \S\ref{sec: ranks} are drawn from the finite-flat case and the ordinary case of Wiles's $R \cong \mathbb{T}$ theorem \cite{wiles1995}. In contrast, the third example is residually reducible, having to do with the Galois representations and modular forms appearing in Ribet's proof of the converse to Herbrand's theorem \cite{ribet1976}. The fourth case is also residually reducible and set in the case of Mazur's study of the Eisenstein ideal \cite{mazur1978}. In this case, the rank of $\mathbb{T}$ was determined in \cite{WWE3} using Massey products, but it is not possible to use $A_\infty$-products; the contrast is discussed. \subsection{Galois cohomology of the trivial representation} \label{subsec: NRIT} The universal commutative deformation ring of a character (i.e.\ 1-dimensional representation) is relatively straightforward: it does not depend on the character, and is simply a completed and abelianized group algebra, as pointed out by Mazur \cite[\S1.4]{mazur1989}. Thus Theorem \ref{thm: main A-inf} gives a cohomological expression for the completion of $\mathbb{F}[G]^\mathrm{ab}$ at the kernel of the trivial $\mathbb{F}$-valued representation, which is naturally isomorphic to $\mathbb{F}{[\![} G^{\mathrm{ab}, \text{pro-}p}{]\!]}$, the completed group algebra of the maximal abelian pro-$p$ quotient group of $G$. In Part 2 we prove a non-commutative version of Theorem \ref{thm: main A-inf} with exactly the same formula for the deformation ring. This non-commutative deformation ring is simply the completion $\mathbb{F}[G]^\wedge$ of $\mathbb{F}[G]$ at the kernel of the trivial representation, which is the completed group algebra of the image of $G$ in its modulo $p$ Malcev completion. This image retains much more information about $G$ than $G^{\mathrm{ab}, \text{pro-}p}$. There is a particular case where the $A_\infty$-products controlling $\mathbb{F}[G]^\wedge$ are particularly classically well-studied. Namely, let $G = G_F$ be the absolute Galois group of a field $F$ with characteristic different than $p$ and containing the $p$-roots of unity. Upon a choice of $p$th root of unity, the dg-$\mathbb{F}_p$-algebra \[ \bigoplus_{i \geq 0} C^i(G_F, \mu_p^{\otimes i}) \cong \bigoplus_{i \geq 0} C^i(G_F, \mathbb{F}_p) \] has cohomology $\mathbb{F}_p$-algebra whose form is given by Milnor $K$-theory according to the norm residue isomorphism theorem of Rost and Voevodsky \cite{voevodsky2011} (i.e.\ the proved motivic Bloch--Kato conjecture, or Milnor conjecture when $p=2$). As $\mathbb{F}_p \cong \mathrm{End}_{\mathbb{F}_p}(\rho)$ for any 1-dimensional representation $\rho$ over $\mathbb{F}_p$, the $A_\infty$-structure (or higher Massey products) enrich this ring and retain extra information. Since the work of Hopkins--Wickelgren \cite{HW2015}, there has been attention to the vanishing of higher Massey products on $H^1(G_F, \mathbb{F}_p)$ and their links with the arithmetic of $F$. For example, as proved in \cite{HW2015, MT2017} triple Massey products on $H^1(G_F, \mathbb{F}_p)$ vanish. There has also long been interest in determining the structure of pro-$p$ completions of Galois groups, which is clearly very much related. The interested reader can look into the extensive literature on these topics; the introduction of \cite{MT2017} contains a survey. We observe that the natural $A_\infty$-structure on cohomology $H^\bullet(G_F, \mathrm{End}_\mathbb{F}(\rho))$ for arbitrary $\rho$ provides the setting for an ``unstable'' generalization of these questions, at least when $F$ contains the $p$th roots of unity. By ``unstable,'' we mean that $\rho$ becomes trivial after restriction to a finite subgroup of $G_F$. Indeed, just as the existence of $F$-points on an algebraic variety is connected with the vanishing of a triple Massey product in \cite{HW2015}, the study of deformations of Galois representations has been motivated by its applications to arithmetic algebraic geometry. In contrast to the setting of the norm residue isomorphism theorem, the study of Galois groups of global fields with restricted ramification, $G_{F,S}$, is quite different. For one thing, the norm residue isomorphism theorem does not apply. It also is understood, for example, that there are non-vanishing triple Massey products. See in particular \cite[Ex.\ 2.11]{HW2015}, which is due to G\"artner, and the references in \cite[\S1]{HW2015}. Similarly, Sharifi \cite{sharifi2007} works over the cyclotomic $\mathbb{Z}_p$-extension of a number field, so that $p$-power Kummer extensions appear in the first cohomology of the trivial representation. Thus the work of Sharifi can also be interpreted deformation-theoretically as deformations of the trivial character. He relates the vanishing behavior of certain Massey products (in cohomology with restricted ramification) to Iwasawa-theoretic class groups. \subsection{The derived deformation rings of Galatius--Venkatesh} \label{subsec: GV2018} We draw some comparisons between the approaches to deformation theory in present paper and the work of Galatius--Venkatesh \cite{GV2018} on derived Galois deformation rings. On the way to doing this, we explain the limitations and advantages of the setting of $A_\infty$-algebras chosen in this paper. We expect that the dg-algebra $C^\bullet(G, \mathrm{End}_\mathbb{F}(\rho))$ and the $A_\infty$-algebra structure on its cohomology have the information of a derived enrichment of conventional deformation theory that we study in this paper, in the sense of e.g.\ \cite[\S3]{DAG-X}. The coefficient rings of this enrichment are $\mathbb{F}$-augmented dg-Artin algebras that are associative but not commutative in any sense. To this author's knowledge, the most straightforward example of an exposition of such a derived enrichment of a moduli functor of representations has been done by Kapranov \cite{kapranov2001}. He studies to local systems on a finite CW-complex valued in an affine algebraic group over $\mathbb{C}$, with coefficient rings in \emph{commutative dg-$\mathbb{C}$-algebras.} The contrast in coefficient rings between the present work and \cite{kapranov2001} are indicative of the reasons for our choice of setting. \begin{itemize}[leftmargin=2em] \item It is well known that only over $\mathbb{Q}$ do commutative dg-algebras furnish a satisfactory category of coefficient rings for derived algebra geometry. In positive characteristic or mixed characteristic, where the desired applications of this paper are located, formulations of Koszul duality between Lie and commutative operads are topics of contemporary homotopy-theoretic research. Indeed, since the first version of this manuscript appeared, Brantner and Mathew have proposed a notion of \textit{partition Lie algebra} as a Lie notion that furnishes a dual to simplicial commutative rings \cite{BM2019}. \item In contrast, the associative operad is well-understood to be self-dual in any characteristic. In this introduction, Koszul duality of operads is visible in the bar construction of \S\ref{subsec: bar def}: it sends an $A_\infty$-algebra to a dg-coalgebra, but an analogue sends an $L_\infty$-algebra (the Lie version of $A_\infty$) to a commutative dg-coalgebra (see e.g.\ \cite{LV2012}), over $\mathbb{Q}$. Because we work in arbitrary characteristic, we have written $\mathrm{End}_\mathbb{F}(\rho)$ instead of $\ad\rho$ to emphasize that we choose the associative algebra structure on the endomorphism ring, as opposed to its induced Lie algebra structure. Accordingly, our strategy to determine commutative deformation rings and other moduli spaces with commutative coefficients is to stay in the associative (non-commutative) world at least until Koszul duality is applied, and then finally abelianize at the end. \item Consequently, we do not work with representations valued in general algebraic groups, instead focusing on matrix-valued representations. For the cohomology controlling the deformations of algebraic group-valued representations is intrinsically valued in a Lie algebra that is not naturally induced by an associative algebra. \item We also work in constant characteristic. Some additional Bockstein-type map is needed to control deformations to mixed characteristic. \end{itemize} In contrast, Galatius--Venkatesh \cite{GV2018} are able to work with algebraic groups $\cG$ other than $\GL_n$ and mixed characteristic coefficient rings. In order to do this, instead of using commutative dg-algebras, they work with simplicial commutative rings as coefficient rings. They are able to calculate the cotangent complex of their derived deformation problem, which is the analogue of our $H^\bullet(G, \mathrm{End}_\mathbb{F}(\rho))$, and show that it matches the Andr\'e--Quillen cohomology of the derived deformation ring \cite[Lem.\ 5.10]{GV2018}. But as far as we are aware, only the more recent work of Brantner--Mathew, formulating partition Lie algebras, has the potential to supply structure on the cotangent complex that retains the information of the deformation problem. What we gain from a more limited choice of setting than \cite{GV2018} is that there is a well-established and integer-indexed additional structure on $H^\bullet(G, \mathrm{End}_\mathbb{F}(\rho))$ -- the $A_\infty$-structure -- that controls our deformation ring. Finally, we want to highlight the extreme contrast in the terminal results of the present paper compared to those of Galatius--Venkatesh \cite{GV2018}. The main results of this paper express the influence of $H^2(G,\mathrm{End}_\mathbb{F}(\rho))$ on obstructions to (classical) deformations. In particular, the results here are trivial when $H^1(G,\mathrm{End}_\mathbb{F}(\rho))$ is zero, as there are then no non-trivial classical deformations. In contrast, the derived deformation problem to which the terminal results of \cite{GV2018} applies is expected to have no non-trivial classical deformations, yet still have non-trivial derived deformations. For another approach, producing an \emph{analytic} derived moduli space of representations of a profinite group, see \cite{antonio2017}. \subsection{Non-commutative deformation theory} \label{subsec: NCDT} As we have mentioned above, Part 2 has a new result in non-commutative deformation theory that is applied in Part 3, along with some results of \cite{WE2018}, to prove the main theorems stated in this introduction. The main results of Part 2 are Theorem \ref{thm: r-pointed a-inf} and Corollary \ref{cor: r-pointed dual}. We want to make some comments about how these results are related to other work in non-commutative deformation theory. In non-commutative deformation theory, the content of these results addresses what is called the ``deformation theory of $r$-points.'' Here we have an associative $\mathbb{F}$-algebra $E$. A ``point'' of $E$ is a maximal ideal $\mathfrak{m}$ of $E$, and we extend $\mathbb{F}$ if necessary so that $E/\mathfrak{m} \cong M_d(\mathbb{F})$. We now take $r$ points, thought of as surjective representations $\rho_i : E \rightarrow M_{d_i}(\mathbb{F})$ for $i = 1, \dotsc, r$ cutting out distinct maximal ideals. One principal distinction from commutative deformation theory is that distinct points can have extensions between them, i.e.\ $\mathrm{Ext}^1_E(\rho_i, \rho_j)$ can be non-trivial when $\rho_i \not\simeq \rho_i$. This does not happen in the commutative setting. Let $\rho := \bigoplus_{i=1}^r \rho_i$ as usual. In doing non-commutative deformation theory, we are interested in determining the completion $E^\wedge_\rho$ of $E$ at the kernel of $\rho$. Our main result on non-commutative deformation theory (Theorem \ref{thm: r-pointed a-inf}) follows on previous work of Segal \cite[\S2]{segal2008}, which in turn is a development of work of Laudal \cite{laudal2002}. Segal proves a result that is very similar to Corollary \ref{cor: r-pointed dual}(2): in \cite[Thm.\ 2.14]{segal2008}, he proves that what we call $R^\mathrm{nc}_\rho$ is isomorphic to what we call $R$ (in the notation of Corollary \ref{cor: r-pointed dual}). The difference is that we keep track of data that \emph{determines an isomorphism} between $R^\mathrm{nc}_\rho$ and $R$, and determines a presentation for $E^\wedge_\rho$ in terms of cohomology. This is the data of a homotopy retract between the Hochschild cochain complex $C^\bullet(E,\mathrm{End}_\mathbb{F}(\rho))$ and its cohomology, which we explain in \S\ref{subsec: dg to A-inf}. We also carefully keep track of the some choices of idempotents that we use to remove Segal's assumption that $d_i$ (the dimensions of the simple summands of $\rho$) equals 1 for all $i$. Thus we have identified data that determines a presentation of $E^\wedge_\rho$ and the deformation functor, while Segal's approach identifies the isomorphism class of $E^\wedge_\rho$ and its deformation functor. Laudal \cite{laudal2002} also characterizes the isomorphism class of $E^\wedge_\rho$ in terms of Massey products. Thus $E^\wedge_\rho$ is described by Laudal in an inductive way, as we discuss more in \S\ref{subsec: massey express}. As we carry this out, we are careful in \S\ref{subsec: gauge} to make sure that notions of non-commutative gauge equivalence of Maurer--Cartan elements correspond to conjugacy classes of representations. This is well-understood in characteristic zero (see e.g.\ \cite[Defns.\ 2.3 and 2.9]{segal2008}), but appears less often in the literature in arbitrary characteristic because it is often expressed as an exponential. For this purpose, we found Prout\'e's study \cite{proute2011} of twisting morphisms very useful (and also \cite{CL2011}). The statement of the decomposition theorem (Theorem \ref{thm: decomp}) for $A_\infty$-algebras by Chuang--Lazarev \cite{CL2017} was also very helpful. There is a multitude of additional work along these lines in representation theory and non-commutative geometry, of which we can mention only a couple more. \begin{itemize}[leftmargin=2em] \item The monograph of Le Bruyn \cite{lebruyn2008} summarizes results from non-commutative geometry that are related to the content of Part 2, as well as the Cayley--Hamilton algebra theory and pseudodeformation theory dealt with in Part 3. Le Bruyn especially focuses on the representation-unobstructed case (in the terminology of Corollary \ref{cor: RU case}). This is especially relevant (see \S\ref{subsec: GIT intro} and \S\ref{subsec: invariants quivers}) for the study of the ring $R^1_D$ that appears in the main theorem on the pseudodeformation ring (Theorem \ref{thm: main pseudo}). \item Keller \cite{keller2001, keller2002, keller2006} explains the connections between $A_\infty$-algebras and categories of representations. These have much of the same content as our results or Segal's results \cite[\S2]{segal2008}, but are not stated in terms of deformation theory. See especially \cite[\S2]{keller2002}. \item Maurer--Cartan elements and their gauge equivalences were promoted by the work of Kontsevich, e.g.\ \cite{kontsevich2003}. For an introduction to non-commutative deformations and $A_\infty$-algebras similar to our approach, see e.g.\ \cite{KS2009}. \end{itemize} \subsection{The Kuranishi map} \label{subsec: kuranishi} To conclude our discussion of complements, we discuss one more perspective on our presentation of the deformation ring in terms of cohomology in Theorem \ref{thm: main A-inf}. Such results have been sought after in the context of the variation of flat connections on manifolds. In this case, there are functions whose analytic germ is the denominator in the expression of the deformation ring of Theorem \ref{thm: main A-inf}. Indeed, this germ has a natural extension to an analytic function in the neighborhood of the origin in the appropriate cohomology vector space $H^1(\mathrm{End} (\rho))$, known as the Kuranishi obstruction map; see e.g.\ \cite[Ch.\ 12]{MMR1994}. Likewise, in \cite{GM1988}, Goldman--Millson relate moduli spaces of flat connections to moduli spaces of representations. If $G$ is the fundamental group of a compact K\"ahler manifold, the comparison between dg-Lie algebras $C^\bullet(G, \mathrm{End}_\mathbb{C}(\rho))$ and the dg-Lie algebra controlling deformations of the flat connection associated to $\rho$ is exploited by Goldman--Millson to prove that these deformations are \emph{formal} (in the sense of Remark \ref{rem: cup case}). This has a consequence that the passage from $C^\bullet(G, \mathrm{End}_\mathbb{C}(\rho))$ to the graded Lie algebra $H^\bullet(G, \mathrm{End}_\mathbb{C}(\rho))$ loses no information and the presentation for the deformation space as in Theorem \ref{thm: main A-inf} is quadratic. \subsection{Acknowledgements} \label{subsec: acknow} I would like to thank Akshay Venkatesh for his interest and encouragement in this project from its early stages. I would also like to thank the students of Math 202a in Spring 2014 at Brandeis University for their interest as they heard an initial version of the ideas of Part 2. It is also a pleasure to thank Preston Wake for his collaboration, which helped to hone the ideas of this paper through our related joint works \cite{WWE3, WWE4}. It is a pleasure to thank Mark Behrens, John Francis, S{\o}ren Galatius, and Kirsten Wickelgren for helpful conversations about homotopy algebras. Likewise, I thank Bill Goldman and Danny Ruberman for providing perspective on the Kuranishi picture and Ed Segal for comments about non-commutative geometry. I would also like to thank Jo\"el Bella\"iche, Ga\"etan Chenevier, Michael Harris, Romyar Sharifi, and Adam Topaz for helpful conversations and comments. During work on this paper, the author was supported by an AMS-Simons Travel Grant, Engineering and Physical Sciences Research Council grant EP/L025485/1, and research support from Brandeis University and from Imperial College London's Mathematics platform grant. \subsection{Notation and terminology} \label{subsec: notation} $G$ denotes a profinite group. $\mathbb{F}$ denotes a finite field of characteristic $p$, in which $\mathbb{F}[G]$ has the standard profinite topology with completion $\mathbb{F}{[\![} G{]\!]}$. The topology of the codomain of functions with domain $G$ or $\mathbb{F}[G]$ is always from a presentation as a finitely generated (left) module over a topological $\mathbb{F}$-algebra. These topological $\mathbb{F}$-algebras are \begin{itemize}[leftmargin=2em] \item $\cA_\mathbb{F}$, the category of Artinian local associative $\mathbb{F}$-algebras with residue field $\mathbb{F}$, equipped with the discrete topology; \item $\cC_\mathbb{F}$, the full subcategory of $\cA_\mathbb{F}$ consisting of commutative objects; \item the categories of limits $\hat\cA_\mathbb{F}$ and $\hat \cC_\mathbb{F}$, with the resulting profinite topology; \item ${\mathcal{A}f\!f}_\mathbb{F}$, the category of topologically finitely generated $\mathbb{F}$-algebras; this is the opposite category to the category of Noetherian affine $\Spf \mathbb{F}$-formal schemes (see \cite[\S10.1]{ega1}). \end{itemize} We let $E$ denote an associative $\mathbb{F}$-algebra. Often we consider the case $E = \mathbb{F}[G]$ or variants of this, in which case $E$ is topological as discussed above. We let $(V, \rho)$ denote a finite-dimensional representation of $E$, that is, a finite-dimensional $\mathbb{F}$-vector space $V$ with a left $\mathbb{F}$-linear action $\rho$ of $E$. We write $\mathrm{End}_\mathbb{F}(\rho)$ for the adjoint representation of $\rho$, which is an $E$-bimodule. (We write ``$R$-bimodule'' as shorthand for ``$(R,R)$-bimodule.'') In contrast, we use $\mathrm{End}_\mathbb{F}(V)$ for the same $\mathbb{F}$-vector space, but in this case emphasizing its $\mathbb{F}$-algebra structure which receives the homomorphism $\rho$. The difference is only a matter of emphasis. We also write ``$\rho$'' when we identify $V \cong \mathbb{F}^{\oplus d}$, writing $\rho$ as a homomorphism \[ \rho: E \longrightarrow M_d(\mathbb{F}) \quad \text{or} \quad \rho: G \longrightarrow \GL_d(\mathbb{F}) \] in this case. Write $\mathbb{F}[\epsilon_n]$ for $\mathbb{F}[\epsilon]/(\epsilon^{n+1}) \in \cC_\mathbb{F}$. Given a homomorphism $\rho : G \rightarrow \GL_d(\mathbb{F})$, we use the term ``$n$th-order \emph{lift}'' to describe a homomorphism $G \rightarrow \GL_d(\mathbb{F}[\epsilon]/\epsilon^{n+1})$ that reduces to $\rho$ modulo $\epsilon$. In contrast, an ``$n$th-order \emph{deformation}'' refers to an orbit of lifts under the adjoint action of $\ker(\GL_d(\mathbb{F}[\epsilon]/\epsilon^{n+1})) \rightarrow \GL_d(\mathbb{F})$. We use the same terms for lifts of representations of $E$. The term \emph{$\mathbb{F}^r$-algebra} refers to an algebra in the category of $\mathbb{F}^r$-bimodules, where $r \in \mathbb{Z}_{\geq 1}$. In particular, this does \emph{not} refer to algebras receiving a map from $\mathbb{F}^r$ to the center; see \S\ref{subsec: r-pointed data} for more on this. We write $\underline{\otimes}$ for the tensor product in the category of $\mathbb{F}^r$-bimodules, which is also the tensor product for $\mathbb{F}^r$-algebras. All of this reduces to the usual setting of $\mathbb{F}$-algebras when $r=1$. So we will refer to $\mathbb{F}^r$-algebras for the rest of this introduction to notation. Graded objects are graded by $\mathbb{Z}$. Derivations and differentials on a graded object have degree $+1$. A complex is a graded object with a differential: we generally use the notation $(C,d_C)$ for a complex, where $d_C^i : C^i \rightarrow C^{i+1}$. We may denote it by $C$ when the context is clear. Likewise, we may denote the cocycle and coboundary subobjects as $C^i \supset Z^i \supset B^i$ when the context $(C,d_C)$ is understood. Suspension ``$\Sigma$'' on a complex produces the complex where $\Sigma C^i = C^{i+1}$ and $d_{\Sigma C} = -d_C$. We write \emph{dg-algebra} for a differential graded algebra. These are complexes equipped with an associative graded multiplication satisfying the Leibniz rule, and are denoted $(C,d_C, m_{2,C})$, or ``$C$'' for short. These are most often dg-$\mathbb{F}^r$-algebras. We refer to \emph{augmented} dg-$\mathbb{F}^r$-algebras $C$, meaning that there is a augmentation map $C \twoheadrightarrow \mathbb{F}^r$. A \emph{complete} dg-$\mathbb{F}^r$-algebra is an augmented dg-$\mathbb{F}^r$-algebra that is complete with respect to the kernel of the augmentation map. A \emph{free} complete algebra (resp.\ graded algebra, resp.\ dg-algebra) on a (resp.\ graded, resp.\ dg) $\mathbb{F}^r$-bimodule $V$ is the (resp.\ graded, resp.\ dg) tensor algebra \[ \hat T_{\mathbb{F}^r} V := \prod_{n \geq 0} V^{\underline{\otimes} n} \] (resp.\ equipped with the differential produced by extension via the Leibniz rule). We use the following notation for categories of dg-algebras. \begin{itemize}[leftmargin=2em] \item $\cA_{\mathbb{F}^r}^\mathrm{dg}$ finite-dimensional augmented dg-$\mathbb{F}^r$-algebras \item $\hat \cA_{\mathbb{F}^r}^\mathrm{dg}$ limits of finite-dimensional augmented dg-$\mathbb{F}^r$-algebras, such as $\hat T_{\mathbb{F}^r}^\mathrm{dg}$ when $V$ has finite dimension as an $\mathbb{F}$-vector space. These are complete. \end{itemize} Undecorated tensor products ``$\otimes$'' are assumed to be over $\mathbb{F}$. Likewise, we use ``$(-)^*$'' to denote the $\mathbb{F}$-linear dual of a $\mathbb{F}$-linear object or morphism. This applies naturally to objects with $\mathbb{F}^r$-bimodule structure as well. On a graded $\mathbb{F}$-vector space $C$, it is applied graded-piecewise by default: $(C^*)^i := (C^{-i})^*$. We refer to this as a \emph{graded-dual} object when we wish to emphasize this. The duality operation on a complexes produces a complex $(C^*, d_{C^*})$ where $d_{C^*}^i := (d_C^{-i-1})^*$. As is standard when working with tensor products of morphisms of graded vector spaces, the Koszul sign rule \[ (f \otimes f')(x \otimes x') = (-1)^{|f'| |x|} f(x) \otimes f'(x') \] is in force. Coalgebras inherit all of the notions above: codifferential, coaugmentation, cocomplete, cofree, etc., in the standard way. Given a graded $\mathbb{F}$-vector space $C$, we will often refer to the graded vector space denoted ``$\Sigma C^*$.'' This is shorthand for $(\Sigma C)^*$: it is the graded-dual of the suspension. \part{$A_\infty$-algebras and deformation theory} In Part 2, we develop the theory of $A_\infty$-algebras, use them to produce presentations of non-commutative deformation rings, and discuss their relationship with Massey products. \section{$A_\infty$-algebras} \label{sec:A-inf} In this section we recall the definition of an $A_\infty$-algebra and discuss the relationship between $A_\infty$-algebras and dg-algebras. We also discuss the bar equivalence between $A_\infty$-algebras and cocomplete cofree dg-coalgebras, which will provide a useful perspective on $A_\infty$-algebras in the sequel. \subsection{Defining $A_\infty$-algebras} \label{subsec: A-inf defn} Recall that undecorated tensor products ``$\otimes$'' are over the base field $\mathbb{F}$. \begin{defn} \label{defn:A-infinity} An \emph{$A_\infty$-algebra} over $\mathbb{F}$ is a pair $(A, (m_n)_{n \geq 1})$ consisting of a graded $\mathbb{F}$-vector space $A = \bigoplus_{i \in \mathbb{Z}} A^i$ and a sequence of homogenous degree $2-n$ maps \[ m_n: A^{\otimes n} \longrightarrow A, \quad n \geq 1 \] such that \begin{equation} \label{eq:m_conds} \sum (-1)^{r+st} m_u ( 1^{\otimes r} \otimes m_s \otimes 1^{\otimes t}) = 0, \end{equation} where the sum ranges over all decompositions $n = r+s+t$ into non-negative integers, with the conventions that $u = r+1+t$ and $m_0 = 0$. We also use $m$ as shorthand for $(m_n)_{n \geq 1}$. A \emph{morphism} $f: A \rightarrow A'$ of $A_\infty$-algebras $(A, m)$, $(A', m')$ is a sequence of maps \[ f_n : A^{\otimes n} \longrightarrow A', \quad n \geq 1 \] of homogenous degree $1-n$ such that, for all $n \geq 1$, \begin{equation} \label{eq:f_conds} \sum (-1)^{r+st} f_u(1^{\otimes r} \otimes m_s \otimes 1^{\otimes t}) = \sum (-1)^s m'_r(f_{i_1} \otimes f_{i_2} \otimes \dotsm \otimes f_{i_r}) \end{equation} where the first sum runs over all decompositions $n = r+s+t$ into non-negative integers, we maintain $u = r+1+t$, and the second sum runs over all $1 \leq r \leq n$ and all decompositions $n = i_1 + \dotsm + i_r$ into positive integers, where $s = \sum_{j=1}^{r-1} j (i_j-1)$. In particular, the relations above imply that $m_1^2 = 0$, i.e.\ $(A, m_1)$ is a complex. They also imply that $f_1$ is a morphism of complexes $(A,m_1) \rightarrow (A', m'_1)$. A morphism $f$ of $A_\infty$-algebras is called a \emph{quasi-isomorphism} when $f_1$ is a quasi-isomorphism of complexes. It is called an \emph{isomorphism} when $f_1$ is an isomorphism of complexes; indeed, $f$ is an isomorphism if and only if it has an inverse. \end{defn} An $A_\infty$-structure records homotopies that make $m_2$ ``associative up to homotopy,'' in contrast to a dg-algebras, which are (strictly) associative. \begin{eg} \label{eg:A-inf} The following examples illustrate the relationship between this definition and differential graded algebras. Let $(A, m)$ be an $A_\infty$-algebra. \begin{enumerate}[leftmargin=2em] \item We have noted that \eqref{eq:m_conds} states that $(A, m_1)$ is a complex. \item For $n=2$, \eqref{eq:m_conds} states that $m_2(m_1 \otimes 1 + 1 \otimes m) = m_1m_2$, the Leibniz rule. \item For $n = 3$, \eqref{eq:m_conds} states that \[ m_1m_3 + m_3(m_1 \otimes 1 \otimes 1 + 1 \otimes m_1 \otimes 1 + 1 \otimes 1 \otimes m_1) = m_2(1 \otimes m_2 - m_2 \otimes 1). \] We see that the left hand side vanishes when $m_3 = 0$ and the right hand side is an \emph{associator,} as it measures the failure $m_2$ to be associative. The left hand side is a chain homotopy of maps $A^{\otimes 3} \rightarrow A$ that is the boundary of $m_3$, so this relation expresses that $m_2$ is associative up to homotopy. \item The subsequent maps $m_n$ for $m \geq 4$ record associativity up to homotopy among the composition of $n$ elements. \item If $(B,d,m_2)$ is a dg-algebra with differential $d: B \rightarrow B$ of degree 1 and multiplication $m_2: B \otimes B \rightarrow B$ (of degree 0), then assigning $m'_1 := d$, $m'_2 := m_2$, and $m'_n = 0$ for $n \geq 3$ results in an $A_\infty$-algebra $(B,m')$. Likewise, when $(A,m)$ is an $A_\infty$-algebra where $m_n = 0$ for $n \geq 3$, then $(A, m_1, m_2)$ is a dg-algebra. \end{enumerate} Now we discuss a morphism $f = (f_n)_{n \geq 1} : (A, m) \rightarrow (A', m')$. \begin{enumerate}[resume, leftmargin=2em] \item We have noted that \eqref{eq:f_conds} states that $f_1m_1 = m'_1f_1$, i.e.\ $f_1$ is a morphism of degree zero of the complexes. \item For $n = 2$, a rearrangement of \eqref{eq:f_conds} states that \[ m'_1f_2 + f_2(m_1 \otimes 1 + 1 \otimes m_1) = f_1 m_2 - m'_2(f_1 \otimes f_1) \] as maps $A^{\otimes 2} \rightarrow A'$. That is, the right hand side expresses the failure of $f_1$ to be multiplicative with respect to $m_2$, and $f_2: A^{\otimes 2} \rightarrow A'$ is a chain homotopy whose boundary is this failure. \item Consider $A_\infty$-algebra morphisms $f: A \rightarrow A'$ where \begin{itemize}[leftmargin=2em] \item $(A', m')$ is an $A_\infty$-algebra arising from a dg-algebra (i.e.\ $m'_n = 0$ for $n \geq 3$) and \item $m_1 = 0$ (making $(A, m)$ a ``minimal'' $A_\infty$-algebra). \end{itemize} If one also assumes that $m_i = 0$ for $2 \leq i \leq n-1$, then \eqref{eq:f_conds} states \begin{equation} \label{eq:A-inf to Massey} f_1 m_n = m'_1f_n + m'_2(f_1 \otimes f_{n-1} - f_2 \otimes f_{n-2} + \dotsm + (-1)^n f_{n-1} \otimes f_1) \end{equation} Equation \eqref{eq:A-inf to Massey} will be relevant for comparison with Massey products in \S\ref{subsec: A-inf to Massey}. \end{enumerate} \end{eg} \begin{rem} If the reader finds the relations \eqref{eq:m_conds} and \eqref{eq:f_conds} unenlightening, the more compact equivalent formulation of these conditions in terms of the bar complex may be helpful. For this, see \S\ref{subsec: bar}. \end{rem} \subsection{The relationship between dg-algebras and $A_\infty$-algebras} \label{subsec: dg to A-inf} Homotopy retracts relate dg-algebras and $A_\infty$-algebras. \begin{defn} \label{defn: HR} Let $(A,d_A)$, $(C,d_C)$ be complexes. We call $(A,d_A)$ a \emph{homotopy retract} of $(C,d_C)$ when they are equipped with maps \[ \xymatrix{ C \ar@(dl,ul)^h \ar@<1ex>[r]^p & A \ar@<1ex>[l]^i } \] such that $p$ and $i$ are morphisms of complexes, $h : C \rightarrow \Sigma^{-1} C$ is a morphism of graded vector spaces, $\mathrm{id}_C - ip = d_Ch + hd_C$, and $i$ is a quasi-isomorphism. \end{defn} The following example of a homotopy retract will be used extensively. \begin{eg} \label{eg:H-sections} Let $(C,d_C)$ be a cochain complex. Then $(H^\bullet(C), 0)$ is a homotopy retract of $(C,d_C)$, when it is equipped with maps $i, p$ set up in the following way. For each $n$, choose a section $i^n: H^n(C) \rightarrow C^n$ of the standard map from $\ker(d_C\vert_{C^n})$ to $H^n(C)$ and a section $h^n : d_C(C^n) \rightarrow C^n$ of $d_C\vert_{C^n} : C^n \rightarrow C^{n+1}$. These sections produce a decomposition $C^n \buildrel\sim\over\ra \im(d_C\vert_{C^{n-1}}) \oplus i^n(H^n(C)) \oplus h^n(d_C(C^n))$, which we write as \begin{equation} \label{eq:LPWZ decomp} C^n = B^n(C) \oplus \tilde H^n(C) \oplus L^n(C) = B^n \oplus \tilde H^n \oplus L^n. \end{equation} Think of the three summands as \begin{itemize} \item co\textit{B}oundaries, \item cocycles lifting $H^n(C)$, and \item \textit{L}ifts of coboundaries to a cochain inducing it. \end{itemize} Then let $p^n : C^n \rightarrow H^n(C)$ be the projection killing the summands complementing $\tilde H^n$. Combine the above data into maps of complexes $p$ and $i$. Finally, extend $h^n$ to $C^{n+1}$ by killing the summands complementing $B^{n+1}$, so that the $h^n$ together produce a map $h : C \rightarrow \Sigma^{-1} C$ of graded vector spaces. By design, $p$, $i$, and $h$ make $(H^*(C), 0)$ a homotopy retract of $(C,d_C)$. There is also a complimentary complex to $H^\bullet(C)$, which we call $K = K^\bullet(C)$. It consists of \begin{equation} \label{eq: acyclic K} K^n = B^n \oplus L^n \end{equation} with the restricted differential $d_K = d_C\vert_K$. It is acyclic and contractible, decomposing into $L^n \buildrel\sim\over\ra B^{n+1}$ plus the zero map out of $B^n$. \end{eg} When $(A,d_A)$ is a homotopy retract of $(C,d_C)$, there is an isomorphism $H^\bullet(A) \buildrel\sim\over\ra H^\bullet(C)$ induced by $i$. Consequently, when $(C,d_C)$ admits the extra structure of a dg-algebra $(C,d_C, m_C)$, there is an induced graded algebra structure on $H^\bullet(A)$. Moreover, there is a natural lift of this graded algebra structure to a graded linear map $A \otimes A \rightarrow A$, namely, \[ m_2(x \otimes y) := pm_C(i(x) \otimes i(y)), \quad \text{ for } x,y \in A. \] It satisfies the Leibniz rule with respect to $d_A$ but is not necessarily associative. The $m_2$ map is, however, associative up to a homotopy that can be expressed in terms of the homotopy retract structure. This homotopy is a map $m_3 : A^{\otimes 3} \rightarrow A$ of homogenous degree $-1$ that satisfies the relation \eqref{eq:m_conds} for $n=3$, see \cite[Lem.\ 9.4.2]{LV2012}. Iterating this procedure yields the following result. \begin{thm}[{Kontsevich--Soibelman \cite{KS2000}}] \label{thm:KS} Let $(C, (m'_n))$ be an $A_\infty$-algebra such that the complex $(C,m'_1)$ is a homotopy retract of $(A, d_A)$ via \[ \xymatrix{ *[r]{\;(C, m'_1)\;} \ar@<0.5ex>[rr]^{p} \ar@(dl,ul)[]^{h} && *[l]{\;(A,d_A)\,} \ar@<0.5ex>[ll]^{i} } \] The retract determines an $A_\infty$-algebra structure $(m_n)$ on $(A, d_A)$ with $m_1 = d_A$ and a quasi-isomorphism of $A_\infty$-algebras $f = (f_n) : (A, (m_n)) \buildrel\sim\over\ra (C, (m'_n))$ such that $f_1 = i$. \end{thm} The equality $f_1 = i$ will be expressed as ``$f = (f_n)_{n \geq 1}$ is a quasi-isomorphism of $A_\infty$-algebras extending the quasi-isomorphism of complexes $i$.'' \begin{proof} See \cite[Thm.\ 9.4.14]{LV2012} and the references therein. \end{proof} We will especially commonly apply the theorem in the following case. In this statement, recall that any dg-algebra $(C,d_C,m_C)$ is an $A_\infty$-algebra in the sense of Example \ref{eg:A-inf}(4). \begin{cor}[{Kadeishvili \cite{kadeishvili1982}}] \label{cor:A-inf on H} Let $(C,d_C, m_{2,C})$ be a dg-algebra with graded algebra of cohomology $(H^\bullet(C), 0, \bar m_{2,C})$. Then any choice of homotopy retract \[ \xymatrix{ *[r]{\;(C, d_C)\;} \ar@<0.5ex>[rr]^{p} \ar@(dl,ul)[]^{h} && *[l]{\;(H^\bullet(C),0)\,} \ar@<0.5ex>[ll]^{i} }\] as in Example \ref{eg:H-sections} determines an $A_\infty$-structure $(m_n)_{n \geq 1}$ on $H^\bullet(C)$ and a quasi-isomorphism of $A_\infty$-algebras $f = (f_i)_{i \geq 1} : (H^\bullet(C),(m_n)) \rightarrow (C, d_C, m_{2,C})$ such that \begin{itemize} \item $m_1 = 0$, \item $m_2 = \bar m_C$, and \item $f$ extends $i$, that is, $f_1 = i$. \end{itemize} \end{cor} \begin{proof} See \cite[Cor.\ 9.4.8]{LV2012}. We sketch this proof in the following examples, because we require the explicit expression of $f$ in the sequel. \end{proof} \begin{eg} \label{eg:KFT1} For $n \geq 3$, let $PBT_n$ be the set of planar binary rooted trees with $n$ leaves, as in \cite[App.\ C]{LV2012}. For each $T \in PBT_n$, label each leaf of $T$ with $i$, each intermediate branch of $T$ by $h$, each vertex by $m_C$, and the root of $T$ by $p$. With the $j$th leaf corresponding to the $j$th tensor factor of $H^*(C)^{\otimes n}$, each $T \in PBT_n$ determines a homogenous degree $2-n$ map $H^*(C)^{\otimes n} \rightarrow H^*(C)$ (in the sense of e.g.\ \cite[Thm.\ 10.3.8]{LV2012}). We set $m_n$ to be the sum over all $T$, where the summand for $T$ is given sign $(-1)^{s_T+1}$, where $s_T$ is the number of leaves lying above the left of the two branches arriving at the root. The construction of $f = (f_n)$ is similar. We already have $f_1 = i$. For $n \geq 2$, we modify the labeling of $PBT_n$ used to produce $m_n$ by labeling the root by $h$ instead of by $p$. As a result, each $T \in PBT_n$ gives rise to a homogenous degree $n-1$ map $H^*(C)^{\otimes n} \rightarrow C$. We assign $f_n$ to the sum of these maps over all $T \in PBT_n$, with the sign convention as above. \end{eg} \begin{eg} \label{eg:KFT2} For an explicit formula realizing Kadeishvili's theorem but not referring to $PBT_n$, we reproduce \cite[p.\ 2021]{LPWZ2009}, a formula credited to Merkulov \cite{merkulov1999}. Crucially for these formulas, the homotopy retract data induce an identification of $H^n(C)$ with a subspace $\tilde H^n \subset C^n$ as in \eqref{eq:LPWZ decomp}. We will now define versions of $m_n$ on $C$ instead of $H^\bullet(C)$. We denote these by $\tilde m_n$. Initially, we set $\tilde m_1 = d_C$, $\tilde m_2 = m_C$, as usual. For $n \geq 3$ we set define a homogenous degree $2-n$ map $\tilde m_n : C^{\otimes m} \rightarrow C$ by \begin{equation} \label{eq:tilde m} \tilde m_n = \sum_{s+t=n; s,t \geq 1} (-1)^{s+1} m_C ((h\circ \tilde m_s) \otimes (h\circ \tilde m_t)), \end{equation} where we formally define $h\circ \tilde m_1$ to be $-\mathrm{id}_C$. Now, for $n \geq 1$, we may define $m_n : H^\bullet(C)^{\otimes n} \rightarrow H^\bullet(C)$ as $p \circ \tilde m_n \circ i^{\otimes n}$, where $i$ stands in for the inclusion $\tilde H \hookrightarrow C$. Likewise, let $f_n := -(h \circ \tilde m_n) \circ i^{\otimes n}$, resulting in $f_1 = i$. \end{eg} One may check by examining diagrams in $PBT_n$ that the formulations of $m_n$ (resp.\ $f_n$) in Examples \ref{eg:KFT1} and \ref{eg:KFT2} are equal. \begin{defn} Let $(A, m)$ be an $A_\infty$-algebra. It is called \emph{minimal} provided that $m_1 = 0$. A minimal $A_\infty$-algebra equipped with a quasi-isomorphism from $A$ is called a \emph{minimal model} of $A$. \end{defn} The $A_\infty$-algebras $(A, m)$ produced in Examples \ref{eg:KFT1} and \ref{eg:KFT2} are minimal $A_\infty$-algebras. The morphism $f$ produced in each example is the minimal model, as guaranteed by Corollary \ref{cor:A-inf on H}. \subsection{Relationship with the minimal model} In this section, we explain the extent to which the minimal model induced by homotopy retracts is unique. We also decompose an $A_\infty$-algebra into the sum of its minimal model and a linearly contractible factor. To begin, we record this useful fact. \begin{lem} \label{lem: p extends} In the setting of Theorem \ref{thm:KS}, the projection map $p$ extends to a quasi-isomorphism of $A_\infty$-algebras. That is, there exists a quasi-isomorphism $g = (g_n)_{n \geq 1}: (C, m') \rightarrow (A, m)$ such that $g_1 = p$. \end{lem} \begin{proof} See \cite[Thm.\ 3.9(2)]{CL2017}. \end{proof} This is the uniqueness statement we will use. \begin{prop} \label{prop: uniqueness over retract} Let $(C,d_C, m_{2,C})$ be a dg-algebra with graded algebra of cohomology $(H^\bullet(C), 0, \bar m_{2,C})$. Let $(i,p), (i',p')$ be two choices of homotopy retract as in Corollary \ref{cor:A-inf on H}, resulting in two pairs \[ (f, m) \text{ and } (f', m'). \] Then the morphism of $A_\infty$-algebras $h := g \circ f$ (where $g$ is as in Lemma \ref{lem: p extends}) \[ h = (h_n)_{n \geq 1}: (H^\bullet(C), m) \buildrel\sim\over\lra (H^\bullet(C), m'). \] is an isomorphism, and $h_1$ is the identity map on the complex $H^\bullet(C)$. \end{prop} \begin{proof} The claim follows from Lemma \ref{lem: p extends} and Corollary \ref{cor:A-inf on H}. \end{proof} To prepare for the statement of the following theorem, recall that the homotopy retract on $(C,d_C)$ given in Example \ref{eg:H-sections} produced the linearly contractible subcomplex $(K, d_K) \subset (C,d_C)$ defined in \eqref{eq: acyclic K}. The \emph{trivial $A_\infty$-structure} on $K$, which we also denote by $(K,d_K)$, consists of $m_1 = d_K$ along with $m_n = 0$ for $n \geq 2$. \begin{thm}[Decomposition theorem] \label{thm: decomp} Let $(C, m')$ be an $A_\infty$-algebra. Let the cohomology $(H=H^\bullet(C), 0)$ of the complex $(C,m'_1)$ be given the structure of a homotopy retract via \[ \xymatrix{ *[r]{\;(C, m'_1)\;} \ar@<0.5ex>[rr]^{p} \ar@(dl,ul)[]^{h} && *[l]{\;(H,0)\,} \ar@<0.5ex>[ll]^{i} } \] in the form of Example \ref{eg:H-sections}. Let $(H, m)$ be the minimal $A_\infty$-algebra structure on $H$ induced in Theorem \ref{thm:KS} by this data, and let $(K, d_K)$ represent a trivial $A_\infty$-algebra. Then there exists an isomorphism of $A_\infty$-algebras \[ \chi: (C, m') \buildrel\sim\over\lra (H, m) \oplus (K, d_K) \] such that \begin{enumerate}[label=(\roman*), leftmargin=2em] \item $\chi_1 : C \buildrel\sim\over\ra H \oplus K$ is the identity isomorphism of complexes. \item The projection of $\chi$ onto $H$ is equal to the $A_\infty$-quasi-isomorphism $g : C \buildrel\sim\over\ra H$ of Lemma \ref{lem: p extends} \item The restriction to $H$ of the inverse isomorphism $\chi^{-1} : H \oplus K \rightarrow C$ is equal to the $A_\infty$-quasi-isomorphism $f : H \buildrel\sim\over\ra C$ of Theorem \ref{thm:KS}. \end{enumerate} \end{thm} \begin{proof} This is the case of \cite[Thm.\ 3.14]{CL2017} where the differential on $H$ is trivial. Moreover, an explicit formula for $\chi$ is given there. \end{proof} \subsection{The bar equivalence} \label{subsec: bar} In this section, we set up a dual formulation of $A_\infty$-algebras. Recall that $\Sigma$ denotes the suspension operation on complexes $(A, d_A)$, so $(\Sigma A)^i = A^{i+1}$. We write $s$ for the canonical homogeneous degree $-1$ map $s : A \rightarrow \Sigma A$ and $\omega:= s^{-1}$ for its inverse. Recall that $d_{\Sigma A} = -s \circ d_A \circ \omega$. \begin{defn}[Bar construction] \label{defn: bar} Let $A, A'$ be graded vector spaces. For $n \geq 1$, let $m_n : A^{\otimes n} \rightarrow A$ have homogeneous degree $2-n$. Given this data, the \emph{bar construction} consists of \begin{itemize}[leftmargin=2em] \item The homogenous degree $1$ map $b_n : (\Sigma A)^{\otimes n} \rightarrow \Sigma A$ given by $b_n := -s \circ m_n \circ \omega^{\otimes n}$, for $n \geq 1$, \item The coderivation $b = b_A$ on the cofree cocomplete coalgebra \[ \hat T^\mathrm{co}_\mathbb{F} (\Sigma A) := \bigoplus_{i\geq 0} (\Sigma A)^{\otimes i} \] determined by extending $\bigoplus_{n \geq 1} b_n: \hat T^\mathrm{co} (\Sigma A) \rightarrow \Sigma A$ to a coderivation via the co-Leibniz rule. \end{itemize} We write $\mathrm{Bar}(A)$ or $\mathrm{Bar}(A, m)$ for the data $(\hat T^\mathrm{co}_\mathbb{F} (\Sigma A), b_A)$. Likewise, we define the analogue of the bar construction for morphisms. Given the data $f = (f_n)_{n \geq 1}$ where $f_n : A^{\otimes n} \rightarrow A'$ has homogenous degree $1-n$, we define \begin{itemize}[leftmargin=2em] \item the homogeneous degree $0$ map $g_n:= (\Sigma A)^{\otimes n} \rightarrow \Sigma A'$ given by $g_n = s \circ f_i \circ \omega^{\otimes n}$, for $n \geq 1$ \item the morphism $g^*$ of free graded cocomplete coalgebras $\hat T^\mathrm{co}_\mathbb{F} (\Sigma A) \rightarrow \hat T^\mathrm{co} (\Sigma A')$ determined by extending $\bigoplus_{n \geq 1} g_n : \hat T^\mathrm{co}_\mathbb{F} (\Sigma A) \rightarrow \Sigma A'$ co-multiplicatively. \end{itemize} We write $\mathrm{Bar}(f): \mathrm{Bar}(A') \rightarrow \mathrm{Bar}(A)$ for $g^*$. \end{defn} Note that we have not imposed any conditions on the graded linear maps $m_n$ (resp. $f_n$) in the construction above. The following theorem explains exactly when they satisfy the $A_\infty$-compatibility conditions making $(A,m)$ an $A_\infty$-algebra (resp.\ making $f$ a morphism of $A_\infty$-algebras). \begin{thm} \label{thm: bar} Let $A, A'$ be graded vector spaces. For $n \geq 1$, let $m_n : A^{\otimes n} \rightarrow A$ and $m'_n : A'^{\otimes n} \rightarrow A'$ (resp.\ $f_n : A^{\otimes n} \rightarrow A'$) be linear of homogeneous degree $2-n$ (resp.\ $1-n$). \begin{enumerate}[leftmargin=2em] \item $(A, (m_n)_{n\geq 1})$ is an $A_\infty$-algebra if and only if the bar construction $(\hat T^\mathrm{co}_\mathbb{F} (\Sigma A), b)$ is a dg-coalgebra, i.e.\ $b^2 = 0$. \item Assuming that $(A, (m_n)_{n \geq 1})$ and $(A', m'_n)_{n \geq 1})$ are $A_\infty$-algebras, $f := (f_n)_{n \geq 1}$ is a morphism of $A_\infty$-algebras $f : A \rightarrow A'$ if and only if $\mathrm{Bar}(f) : \hat T^\mathrm{co}_\mathbb{F} (\Sigma A) \rightarrow \hat T^\mathrm{co}_\mathbb{F} (\Sigma A')$ is a morphism of dg-coalgebras (i.e.\ $\mathrm{Bar}(f)$ commutes with the codifferentials $b^*_A$ on $\hat T^\mathrm{co}_\mathbb{F} (\Sigma A)$ and $b^*_{A'}$ on $\hat T^\mathrm{co}_\mathbb{F} (\Sigma A')$. \end{enumerate} \end{thm} \begin{proof} See e.g.\ \cite[Lem.\ 9.2.2 and \S9.2.11]{LV2012}. \end{proof} The following amplification of the theorem is also true. \begin{cor}[Bar equivalence] \label{cor: bar equiv} The bar construction defines a functor from $A_\infty$-algebras to the full subcategory dg-coalgebras determined by the cofree cocomplete objects. This functor is an isomorphism of categories. \end{cor} \begin{proof} The first statement follows directly from Theorem \ref{thm: bar}. The second statement follows from the calculations of the bar construction (Definition \ref{defn: bar}): we see that they do not lose any information and that a natural inverse to the bar construction exists. \end{proof} When the graded factors $A^n \subset A$ are finite-dimensional over $\mathbb{F}$, it amounts to the same thing to work with the dg-algebra that is $\mathbb{F}$-dual to the dg-coalgebra $\mathrm{Bar}(A)$. This is the dg-algebra that appears in \S\ref{subsec: bar def}. We will use the notation therein for this complete free dg-algebra: \begin{defn} Let $(A,m)$ be an $A_\infty$-$\mathbb{F}$-algebra. Assume that $A^n$ is finite-dimensional for all $n \in \mathbb{Z}$. Then we write \[ \mathrm{Bar}^*(A, m) := (\hat T_\mathbb{F} \Sigma A^*, m^*, \pi) \] for the free complete dg-algebra that is graded-dual to the cofree cocomplete dg-coalgebra $\mathrm{Bar}(A,m)$. \end{defn} Note that we write $m^*$ for the differential that is dual to the codifferential $b$. \subsection{The Maurer--Cartan functor} \label{subsec: MC functor intro} We change notation: now $(A,d_A, m_{2,A})$ denotes a test object in the category of dg-algebras, while $(C,d_C, m_{2,C})$ denotes a dg-algebra receiving an $A_\infty$ quasi-isomorphism $f : (H, m) \rightarrow (C,d_C, m_{2,C})$ from a minimal $A_\infty$-algebra $(H,)$. We follow \cite{CL2011}. For simplicity, we assume that $A$ is finite-dimensional and complete with maximal ideal $\mathfrak{m}_A$, i.e.\ $A \in \cA_\mathbb{F}^\mathrm{dg}$, which suffices for applications. We denote by $A^* = (A^*, d^*_A, m^*_{2,A})$ the natural dg-coalgebra that is $\mathbb{F}$-dual to $A$, with maximal ideal $\mathfrak{m}^*_A$. Recall that $\mathfrak{m}_A$ is nilpotent. We put an ``$A$-linear'' $A_\infty$-algebra structure $m^A$ on $H \otimes A$, defined by \begin{equation} \label{eq: A-linear m} m_n^A = \mathrm{mult}_n \otimes m_n : (A \otimes H)^{\otimes n} \cong A^{\otimes n} \otimes H^{\otimes n} \to A \otimes H, \end{equation} where $\mathrm{mult}_n$ is the usual $n$-fold multiplication map $A^{\otimes n} \to A$. Likewise, we obtain $b^A_n$. \begin{defn}[{Cf.\ \cite[\S7.6]{keller2001}}] \label{defn:HMC} Let $(H, m)$ be an $A_\infty$-algebra. Given a dg-algebra $A \in \cA_\mathbb{F}^\mathrm{dg}$, a \emph{Maurer--Cartan element} for $H$ valued in $A$ is some $\xi \in (\mathfrak{m}_A \otimes H)^1$ such that \begin{equation} \label{eq: HMC} (d_A \otimes \mathrm{id}_H)(\xi) + \sum_{n =1}^\infty (-1)^{\frac{n(n+1)}{2}}m^A_n(\xi) = 0. \end{equation} Denote the set of such elements by $\mathrm{MC}(H,A)$. \end{defn} \begin{rem} The reader may encounter an alternate definition of the homotopy Maurer--Cartan equation in terms of the $b^A_n$, which has the advantage of lacking signs. Then one must consider Maurer--Cartan elements in the degree zero piece of the suspension. They are $\xi' \in (\mathfrak{m}_A \otimes \Sigma H)^0$ such that \begin{equation} \label{eq: HMC shift} (d_A \otimes \mathrm{id}_{\Sigma H})(\xi') + \sum_{n =1}^\infty b^A_n(\xi') = 0. \end{equation} \end{rem} \begin{rem} For applications, we will restrict to \emph{classical} complete algebras $A$, i.e.\ $A \in \cA_\mathbb{F}$. In this case, Maurer--Cartan elements come from $\mathfrak{m}_A \otimes \Sigma (H^1)$, and the leftmost term of the equations \eqref{eq: HMC} and \eqref{eq: HMC shift} may be dropped. \end{rem} \begin{rem} \label{rem: MC classical} In the case that an $A_\infty$-algebra $C$ arises from a dg-algebra $(C,d_C, m_{2,C})$ and $A \in \cA_\mathbb{F}$ is classical, the Maurer--Cartan equation takes on its classical formulation \[ d_C(\xi) + m_{2,C}(\xi \otimes \xi) = 0. \] Sometimes the equation of Definition \ref{defn:HMC} is called the \emph{homotopy} Maurer--Cartan equation to distinguish it from this case. \end{rem} \begin{prop} \label{prop:MC-corep} Let $(H, m)$ be an $A_\infty$-algebra. The association \[ \cA^\mathrm{dg}_\mathbb{F} \ni A \mapsto \mathrm{MC}(H,A) \in \underline{\text{Sets}} \] is functorial in $A$ and corepresentable by the bar construction of $H$; that is, \[ \mathrm{MC}(H,A) = \mathrm{Hom}_\text{dg-co}(A^*, \mathrm{Bar}(H)). \] Moreover, if $H^n$ is finite-dimensional for all $n \in \mathbb{Z}$, then $\mathrm{MC}(H,A)$ is representable by the dual of the bar construction of $H$, that is, \[ \mathrm{MC}(H,A) = \mathrm{Hom}_\text{dg}(\mathrm{Bar}^*(H), A). \] \end{prop} \begin{proof} By calculation, e.g.\ \cite[Prop.\ 2.2]{CL2011}. \end{proof} \begin{eg} \label{eg: fin dim classical hull} Assume that $H^n$ is finite-dimensional for all $n$. Because of our interest in working with classical coefficient rings, we want to determine the classical complete $\mathbb{F}$-algebra $R$ such that for all classical augmented algebras $A = (A, 0, m_{2,A})$, we have \[ \mathrm{Hom}_\mathrm{dg}(\mathrm{Bar}^*(H), A) = \mathrm{Hom}_\mathrm{alg}(R,A). \] This $R$ is the \emph{classical hull} of $\mathrm{Bar}^*(H) = (\hat T (\Sigma H)^*, m^*, \pi)$, which we set up in \S\ref{subsec: classical hull}. \end{eg} \subsection{The gauge action} \label{subsec: gauge} Here we recall basic facts about the gauge action on Maurer--Cartan sets. \begin{rem} We emphasize that the following version of the gauge action is well-defined over a field $\mathbb{F}$ of any characteristic. One reason we highlight this is that many accounts of the gauge action are expressed in the setting of Lie algebras, where the gauge action originated. There, the action is best expressed as an exponential and therefore only makes sense in characteristic zero. This action readily generalizes to the associative setting, as we will see here; but only some of the literature about this action has denominator-free expressions. \end{rem} \begin{defn}[Unital associative gauge action] Let $(C,d_C, m_{2,C})$ be a unital dg-$\mathbb{F}$-algebra. Let $\beta \in C^1$. Then the gauge action of $\gamma \in (C^0)^\times$ on $\beta$ is \[ \gamma \cdot \beta := \gamma \beta \gamma^{-1} - d_C(\gamma) \gamma^{-1}. \] \end{defn} For non-unital dg-algebras, one can use the same formula to produce the following augmented gauge action. \begin{defn}[Augmented associative gauge action] \label{defn: aug gauge action} Let $(C,d_C, m_{2,C})$ be a non-unital complete dg-$\mathbb{F}$-algebra. Then the gauge action of $\gamma \in C^0$ on $\beta \in C^1$ is the unital gauge action of $(1-\gamma)^{-1}$, which is \[ \gamma \cdot \beta := \beta - (1-\gamma)^{-1}(\beta \gamma - \gamma \beta + d_C(\gamma)), \] where $(1-\gamma)^{-1}$ is interpreted as the standard power series and 1 is interpreted as the identity action on $C$. \end{defn} \begin{prop} The gauge action in the expression above preserves the Maurer--Cartan subset of $C^1$ (consisting of elements such that $d\beta + \beta^2 = 0$). \end{prop} \begin{proof} For lack of a reference, we supply the calculation in the unital case, from which the augmented case can be derived. Given $\gamma \in (C^0)^\times$ and a Maurer--Cartan element $\beta \in C^1$, we calculate that \begin{align*} d(\gamma \cdot \beta) & = d(\gamma \beta \gamma^{-1} - d(\gamma)\gamma^{-1}) \\ & = d(\gamma)\beta\gamma^{-1} - \gamma d(\beta)\gamma^{-1} - \gamma\beta d(\gamma^{-1}) + d(\gamma) d(\gamma^{-1}) \end{align*} while \[ (\gamma \cdot \beta)^2 = - d(\gamma) \beta \gamma^{-1} + \gamma \beta^2 \gamma^{-1} - \gamma \beta\gamma^{-1} d(\gamma) \gamma^{-1} + d(\gamma) \gamma^{-1} d(\gamma)\gamma^{-1}. \] Pairing the terms of these two expressions using the written order of the summands, and in light of the formula $d(\gamma^{-1}) = -\gamma^{-1}d(\gamma)\gamma^{-1}$, we see that each pair has sum zero. \end{proof} Recall that $\mathrm{MC}(C, A)$ denotes the Maurer--Cartan elements in $(C \otimes \mathfrak{m}_A)^1$, when $(C,m)$ is an $A_\infty$-algebra over $\mathbb{F}$ (perhaps arising from a dg-algebra), $A$ is a dg-$\mathbb{F}$-algebra, and $A$ is complete. The $A_\infty$-algebra structure on $C \otimes \mathfrak{m}_A$ is as in \eqref{eq: A-linear m}. We have the Maurer--Cartan equation of \eqref{eq: HMC}. In this setting, we formulate ``strict'' gauge equivalence in analogy with deformation theory: one conjugates by the multiplicative group $1 + M_d(\mathfrak{m}_A)$, as opposed to the entirety of $M_d(A)^\times$. \begin{defn}[Strict gauge equivalence] \label{defn: strict gauge} Let $C$ be a dg-$\mathbb{F}$-algebra and let $A \in \hat\cA_\mathbb{F}^\mathrm{dg}$. We say that $\beta,\beta' \in (C \otimes \mathfrak{m}_A)^1$ are \emph{strictly gauge equivalent} when they are in the same orbit under the augmented gauge action of $(C \otimes A)^0$. Let $(C, m)$ be an $A_\infty$-algebra, and let $A \in \cA_\mathbb{F}^\mathrm{dg}$. We say that $\beta, \beta' \in (C \otimes \mathfrak{m}_A)^1$ are gauge equivalent via $\gamma \in (C \otimes \mathfrak{m}_A)^0$ when \[ \beta - \beta' = (1 \otimes d_A)\gamma + \sum_{n=1}^\infty \sum_{j=1}^n (-1)^j m_n^A(\beta^{\otimes j-1} \otimes \gamma \otimes \beta'^{\otimes i-j}) \] Denote by $\overline{\mathrm{MC}}(C,A)$ the set of strict equivalence classes of the Maurer--Cartan set $\mathrm{MC}(C,A)$. \end{defn} One may check that the $A_\infty$-version of strict gauge equivalence reduces to the dg-algebra case of Definition \ref{defn: aug gauge action}, provided that the $A_\infty$-algebra arises from a dg-algebra. It is shown in \cite[Cor.\ 4.17]{proute2011} that gauge equivalence is an equivalence relation under a ``connectedness'' assumption. The point of this assumption is that certain expressions can be inverted, but since we presently work in the complete case (unlike \emph{loc.\ cit.}), these assumptions can be dropped. \begin{rem} For an expression of an $A_\infty$ strict gauge action with denominators, see \cite[Defn.\ 2.9]{segal2008}. The denominator-free version above comes from \cite[Defn.\ 4.5]{proute2011}, where it is expressed dually in $C$. \end{rem} \section{Non-commutative deformations of a point} \label{sec:assoc pt} The goal of this section is to prove a non-commutative generalization of theorems of \S\ref{subsec: results1} --- the determination of a deformation ring in terms of $A_\infty$-structure --- about deformations of an absolutely irreducible representation. \subsection{Non-commutative deformation theory} \label{subsec:adt setup} We use the conventions of \S\ref{subsec: notation}, which set up the deformation theory of a representation \[ \rho: E \rightarrow \mathrm{End}_\mathbb{F}(V) \] of an associative $\mathbb{F}$-algebra $E$ with finite dimension $d := \dim_\mathbb{F} V$. We especially use the coefficient categories $\cA_\mathbb{F}$ and $\hat \cA_\mathbb{F}$ described there, consisting of objects $(A, \mathfrak{m}_A)$. When a $\mathbb{F}$-basis for $V$ is chosen, we will write $\rho : E \rightarrow M_d(\mathbb{F})$. \begin{defn} \label{defn: lift assoc} Let $A \in \cA_\mathbb{F}$. A \emph{lift of $\rho$ over $A$} is a homomorphism $\rho_A : E \rightarrow \mathrm{End}_\mathbb{F}(V) \otimes A$ such that $\rho_A \otimes_A \mathbb{F} = \rho$. A \emph{deformation of $\rho$ over $A$} is an equivalence class of lifts $\rho_A : E \rightarrow \mathrm{End}_\mathbb{F}(V) \otimes A$ under the equivalence relation of conjugation (i.e.\ inner automorphism) by an element of $\mathrm{End}_\mathbb{F}(V) \otimes A$ whose reduction modulo $\mathrm{End}_\mathbb{F}(V) \otimes \mathfrak{m}_A$ is $\mathrm{id}_V \in \mathrm{End}_\mathbb{F}(V)$ (any such element is a unit). We define the \emph{lifting functor of $\rho$} (resp.\ the \emph{deformation functor of $\rho$}), denoted ${\mathrm{Def}}^{\mathrm{nc},\square}_\rho$ (resp.\ ${\mathrm{Def}}^\mathrm{nc}_\rho$), as the functor from $\cA_\mathbb{F}$ to the category of sets sending $A$ to the set of lifts (resp.\ deformations) of $\rho$ over $A$. \end{defn} To relate ${\mathrm{Def}}^{\mathrm{nc},\square}_\rho$ to homological invariants, we introduce the Hochschild cochain complex. \begin{defn} \label{defn: Hochs} Let $E$ be an associative $\mathbb{F}$-algebra. Let $M$ be an $E$-bimodule. The \emph{Hochschild cochain complex}, denoted $C^\bullet(E,M)$, is determined by the $E$-bimodules \[ C^\bullet(E,M) := \bigoplus_{i \geq 0} C^i(E,M), \qquad C^i(E,M) := \mathrm{Hom}_\mathbb{F}(E^{\otimes i}, M). \] The differential $d = d^i : C^i(E,M) \rightarrow C^{i+1}(E,M)$ is determined by \begin{align*} d^i(f)(x_1, \dotsc, x_{i+1}) = &\ x_1 f(x_2, \dotsc, x_{i+1}) + (-1)^{i+1} f(x_1, \dotsc, x_i) x_{i+1} \\ + &\ \sum_{j = 1}^i (-1)^j f(x_1, \dotsc, x_j x_{j+1}, \dotsc, x_{i+1}). \end{align*} It is standard to check that $d^{i+1} \circ d^i = 0$. We denote by $H^\bullet(E,M)$ the cohomology graded vector space of $C^\bullet(E,M)$, which is called \emph{Hochschild cohomology}. \end{defn} \begin{lem} \label{lem: Hochs of algebra} If $M$ has the structure of an associative $\mathbb{F}$-algebra, then the graded multiplication on $C^\bullet(E,M)$ induced by \begin{align*} C^i(E,M) \otimes &\ C^j(E,M) \longrightarrow C^{i+j}(E,M) \\ f \otimes g &\mapsto \big[ (x_1, \dotsc, x_{i+j}) \mapsto f(x_1, \dotsc x_i) \cdot g(x_{i+1}, \dotsc, x_{i+j}) \big] \end{align*} is a dg-algebra. \end{lem} \begin{proof} This is a standard computation. \end{proof} ${\mathrm{Def}}^{\mathrm{nc},\square}_\rho$ has the following corepresentability. Indeed, the $E$-bimodule structure on $\mathrm{End}_\mathbb{F}(\rho)$ is the natural one, \begin{equation} \label{eq: EE bimodule structure} ((x,y) \cdot f)(v) = \rho(x) \cdot f(\rho(y) \cdot v) \quad \text{for } x,y \in E, v \in V, f \in \mathrm{End}_\mathbb{F}(\rho). \end{equation} For the statement of the proposition, we interpret $\beta \in C^1(E,\mathrm{End}_\mathbb{F}(\rho)) \otimes \mathfrak{m}_A$ as a function $\beta : E \rightarrow \mathrm{End}_\mathbb{F}(\rho) \otimes \mathfrak{m}_A$. \begin{thm} \label{thm: lift corep} We have the dg-$\mathbb{F}$-algebra $C = C^\bullet(E, \mathrm{End}_\mathbb{F}(\rho))$. For $A \in \cA_\mathbb{F}$, there is the following natural bijection between $A$-valued Maurer--Cartan elements for $C$ and lifts of $\rho$ to $A$. That is, $\mathrm{MC}(C,A) \buildrel\sim\over\ra {\mathrm{Def}}^{\mathrm{nc}, \square}_\rho(A)$ via \[ \xi \mapsto \big(\rho \oplus \xi : E \rightarrow \mathrm{End}_\mathbb{F}(V) \otimes A\big). \] In particular, $\mathrm{Bar}(C)$ corepresents ${\mathrm{Def}}^{\mathrm{nc}, \square}_\rho$. \end{thm} \begin{proof} Using Definition \ref{defn:HMC} and Remark \ref{rem: MC classical}, we calculate that an element \[ \xi \in \mathfrak{m}_A \otimes C^1(E, \mathrm{End}_\mathbb{F}(\rho)) \cong \mathrm{Hom}_\mathbb{F}(E, \mathrm{End}_\mathbb{F}(\rho) \otimes \mathfrak{m}_A) \] is Maurer--Cartan if and only if it obeys the relation \[ \xi(x_1 \cdot x_2) = \xi(x_1) \cdot \xi(x_2) + \rho(x_1) \cdot \xi(x_2) + \xi(x_1) \cdot \rho(x_2). \] From this, one readily observes that an element of $\mathrm{Hom}_\mathbb{F}(E, \mathrm{End}_\mathbb{F}(V) \otimes \mathfrak{m}_A)$ is Maurer--Cartan if and only if \[ \rho \oplus \xi : E \rightarrow \mathrm{End}_\mathbb{F}(V) \otimes A \cong (\mathrm{End}_\mathbb{F}(V) \oplus \mathrm{End}_\mathbb{F}(V) \otimes \mathfrak{m}_A) \] is a $\mathbb{F}$-algebra homomorphism. The corepresentability claim follows from Proposition \ref{prop:MC-corep}. \end{proof} We now turn to deformations. We write \begin{equation} \label{eq: def Erho} E^\wedge_\rho := \varprojlim_i E/\ker(\rho)^i. \end{equation} First, we notice that the natural map $E \rightarrow E^\wedge_\rho$ is a deformation of $\rho$. Indeed, because all deformations of matrix algebras are known to be trivial (or apply \cite[Thm.\ 1.2]{laudal2002}), we get that $E/\ker(\rho)^i$ is isomorphic to a matrix algebra $M_d(A_i)$ for some $A_i \in \cA_\mathbb{F}$. In the limit, we choose an isomorphism \begin{equation} \label{eq: isom lift} E^\wedge_\rho \simeq M_d(R^\mathrm{nc}_\rho) \simeq \mathrm{End}_\mathbb{F}(V) \otimes R^\mathrm{nc}_\rho \end{equation} for a chosen $R^\mathrm{nc}_\rho \in \hat \cA_\mathbb{F}$, well-defined up to inner automorphism. This choice is a lift of $\rho$ to $R^\mathrm{nc}_\rho$. So it is fair to call $E \rightarrow E^\wedge_\rho$ is a deformation of $\rho$ valued in $R^\mathrm{nc}_\rho$. The isomorphism \eqref{eq: isom lift} realizes a Morita equivalence between $E^\wedge_\rho$ and $R^\mathrm{nc}_\rho$, as follows. Selecting the idempotent $e^{11} \in M_d(R^\mathrm{nc}_\rho) \simeq E^\wedge_\rho$ (the matrix with $1$ concentrated in the $(1,1)$ coordinate) via isomorphisms above, the Morita equivalence of categories is explicitly given by \begin{align} \label{eq: morita} \begin{split} E^\wedge_\rho\text{-Mod} &\buildrel\sim\over\lra R^\mathrm{nc}_\rho\text{-Mod} \\ W &\mapsto e^{11}W \\ V \otimes Y &\mapsfrom Y \end{split} \end{align} Now the representability of ${\mathrm{Def}}^\mathrm{nc}$ follows from the explicit Morita equivalence. \begin{prop} \label{prop:def rep} Assume that $\rho : E \rightarrow \mathrm{End}_\mathbb{F}(V)$ is absolutely irreducible and let $R^\mathrm{nc}_\rho \in \hat\cA_\mathbb{F}$ as above. Then ${\mathrm{Def}}^{\mathrm{nc}}_\rho$ is isomorphic to the inner automorphism quotient of the $\mathrm{Hom}$-functor on $\cA_\mathbb{F}$ of $R^{\mathrm{nc}}_\rho$. That is, we have a functorial isomorphism \[ {\mathrm{Def}}^{\mathrm{nc}}_\rho(A) \buildrel\sim\over\lra \mathrm{Hom}_\mathbb{F}(R^{\mathrm{nc}}_\rho, A)/\sim_A, \] where $\sim_A$ indicates the equivalence relation of inner $\mathbb{F}$-algebra automorphisms of $A$. This isomorphism is given by applying $(-) \otimes \mathrm{End}_\mathbb{F}(V)$ to (the domain and codomain of) a representative $\eta : R^\mathrm{nc}_\rho \rightarrow A$ of an element of $\mathrm{Hom}_\mathbb{F}(R^\mathrm{nc}_\rho, A)/\sim_A$. \end{prop} \begin{rem} We see that when we restrict the test coefficients of ${\mathrm{Def}}^\mathrm{nc}_\rho$ from $\cA_\mathbb{F}$ to $\cC_\mathbb{F}$, which we call ${\mathrm{Def}}_\rho$, the abelianization of $R^\mathrm{nc}_\rho$ represents ${\mathrm{Def}}_\rho$. \end{rem} Finally, we prove that ${\mathrm{Def}}^\mathrm{nc}_\rho$ is representable by $C$ when taken up to strict gauge equivalence. The salient point is that strict gauge equivalence amounts to strict isomorphism. \begin{prop} \label{prop: MC-gauge of C} We have the dg-$\mathbb{F}$-algebra $C = C^\bullet(E, \mathrm{End}_\mathbb{F}(\rho))$. For $A \in \cA_\mathbb{F}$, there is the following natural bijection between $A$-valued Maurer--Cartan elements for $C$ and lifts of $\rho$ to $A$. That is, $\overline{\mathrm{MC}}(C,A) \buildrel\sim\over\ra {\mathrm{Def}}^{\mathrm{nc}}_\rho(A)$. \end{prop} \begin{proof} In light of Theorem \ref{thm: lift corep}, it remains to prove that the strict gauge action of $\gamma \in C^0 \otimes \mathfrak{m}_A$ on $\xi \in \mathrm{MC}(C,A)$ amounts to conjugation of $\rho \oplus \xi$ by $(1-\gamma)$. Indeed, we calculate the conjugation \begin{align*} (1-\gamma)^{-1}(\rho \oplus \xi)(1-\gamma) &= \sum_{i=0}^\infty \gamma^i(\rho + \xi - \rho\gamma - \xi\gamma) \\ & = \rho + \xi + \sum_{i=0}^\infty \gamma^i(\gamma \xi - \xi \gamma - d\gamma) \\ & = \rho \oplus \xi', \end{align*} where \[ \xi' := \xi - (1-\gamma)^{-1}([\xi,\gamma] + d\gamma). \] This is exactly the strict case of the augmented gauge action of Definition \ref{defn: aug gauge action}. \end{proof} \subsection{Non-commutative deformation theory determined by $A_\infty$-structure} \label{subsec:adt a-inf} We aim for an analogue of Theorem \ref{thm: lift corep}, giving a homological expression for ${\mathrm{Def}}^\mathrm{nc}_\rho$. This analogue will be formulated in terms of an $A_\infty$-structure on Hochschild cohomology. We use the objects defined above, $E$, $\rho$, and $C$, assuming that $\rho$ is absolutely irreducible. Let \[ H := H^\bullet(C) = H^\bullet(E, \mathrm{End}_\mathbb{F}(\rho)) \] denote Hochschild cohomology of $\mathrm{End}_\mathbb{F}(\rho)$. We choose a homotopy retract structure on $(H, 0)$ relative to $(C, d_C)$ as in Example \ref{eg:H-sections} and apply the results of \S\ref{sec:A-inf}, producing \begin{itemize}[leftmargin=2em] \item a minimal $A_\infty$-structure on $H = H^\bullet(G, \mathrm{End}_\mathbb{F}(\rho))$, denoted \[ (H,m) = (H, (m_n)_{n \geq 2}), \] extending its canonical graded algebra structure $m_2$. This comes along with \item a quasi-isomorphism $f : H \rightarrow C$ of $A_\infty$-algebras (Corollary \ref{cor:A-inf on H}) and \item an isomorphism $\chi : C \rightarrow H \oplus K$ of $A_\infty$-algebras, where the projection to $H$ is a left inverse and right quasi-inverse to $f$ (Theorem \ref{thm: decomp}), and $(K,d_K)$ is a trivial $A_\infty$-algebra \end{itemize} We assemble the following two facts. \begin{lem} \label{lem: contractible MC} For a trivial $A_\infty$-algebra $(K,d_K)$, one has $\overline{\mathrm{MC}}(K,-) = \ast$ and \[ \mathrm{MC}(K,A) \cong B^1 \otimes \mathfrak{m}_A \] for $A \in \cA_\mathbb{F}$. For the particular trivial $A_\infty$-algebra $(K,d_K)$ produced from $C = C^\bullet(E, \mathrm{End}_\mathbb{F}(\rho))$, along with a homotopy retract on $C$ given in Example \ref{eg:H-sections}, we have \[ \mathrm{MC}(K, A) \cong \frac{\mathrm{End}_\mathbb{F}(\rho)}{\mathrm{diag}(\mathbb{F})} \otimes \mathfrak{m}_A, \] where $\mathrm{diag}(\mathbb{F})$ denotes the scalar matrix subfield $\mathbb{F} \hookrightarrow \mathrm{End}_\mathbb{F}(\rho)$. \end{lem} \begin{proof} Firstly, recall that $m_{K,1} : B^i \oplus L^i \rightarrow B^{i+1} \oplus L^{i+1}$ arises by restriction of the differential $d_C^i : C^i \rightarrow C^{i+1}$. It is the zero map on $B^i$ along with the isomorphism $L^i \buildrel\sim\over\ra B^{i+1}$. Considering the case $i=1$ and recalling that $m_{K,n} = 0$ for $n \geq 2$, we have the calculation of $\mathrm{MC}(K,A)$. Considering the case $i=0$, we see that the gauge action is a torsor, hence $\overline{\mathrm{MC}}(K,-)$ is a point. The final claim follows from the canonical isomorphism $B^1 \cong C^0/\ker(d_C^0)$, noting that $C^0 \cong \mathrm{End}_\mathbb{F}(\rho)$ and $d_C^0$ kills exactly the scalar matrices. \end{proof} \begin{lem} \label{lem: connected gauge} Assume that $H^i$ is finite-dimensional for all $i \in \mathbb{Z}$, so that $\mathrm{MC}(H,-)$ is pro-represented on $\cA_\mathbb{F}$ by the classical hull $R$ of $\mathrm{Bar}^*(H) \in \hat\cA_\mathbb{F}^\mathrm{dg}$ (see Example \ref{eg: fin dim classical hull}). For $A \in \cA_\mathbb{F}$, this pro-representability maps strict equivalence classes in $\mathrm{MC}(H,A)$ isomorphically onto inner automorphism classes in $\mathrm{Hom}_\mathbb{F}(R, A)$. \end{lem} The key point is that $H$ is \emph{connected}, in the sense that $H^0 \cong \mathbb{F}$, arising from the center of $C$. \begin{proof} Let $\gamma \in H^0 \otimes \mathfrak{m}_A$ and let $\beta, \beta' \in H^1 \otimes \mathfrak{m}_A$. Because $H^0 \cong \mathbb{F}$ and it arises from the center of $C^0$, the higher multiplications are trivial on $H^0$. That is, when $\gamma' \in H^0$ and $\delta_i \in H^i$ for $1 \leq i < n$, then $m_n(\gamma' \otimes \delta_1 \otimes \dotsm \otimes \delta_{n-1}) = 0$ for $n=1$ and $n \geq 3$, and for any other tensor-permutation of the arguments. Therefore the formula expressing $A_\infty$-strict gauge equivalence (Definition \ref{defn: strict gauge}) between $\beta$ and $\beta'$ via $\gamma$ reduces to \[ \beta - \beta' = -\gamma \beta' + \beta \gamma, \quad \text{and thus} \quad \beta(1-\gamma) = (1-\gamma)\beta' \] (where we are implicitly using $m_2^A$ as multiplication). This is the relation of conjugation by $(1-\gamma)$, and only the $\mathfrak{m}_A$ tensor factor of $H^0 \otimes \mathfrak{m}_A$ conjugates non-trivially. \end{proof} With these two lemmas in place, the idea is to use the isomorphism of Maurer--Cartan sets and their compatible gauge relations under the $A_\infty$-isomorphism $\chi : C \buildrel\sim\over\ra H \oplus K$. \begin{thm} \label{thm: a-inf main} Let $E,\rho, C, H$ be as above. Choose a homotopy retract structure on $(H, 0)$ relative to $(C, d_C)$ as in Example \ref{eg:H-sections}, which gives the additional data $(H, m)$, $f$, and $\chi$. For $A \in \cA_\mathbb{F}$, this choice determines isomorphisms \[ \overline{\mathrm{MC}}(H,A) \underset{\bar f_*}{\buildrel\sim\over\lra} \overline{\mathrm{MC}}(C,A) \buildrel\sim\over\lra {\mathrm{Def}}^\mathrm{nc}_\rho(A), \] and they are functorial in $A$. \end{thm} In words, the theorem states that ${\mathrm{Def}}^\mathrm{nc}_\rho$ is corepresented by the $A_\infty$-algebra $(H, m)$ up to gauge equivalence. We also see that for Maurer--Cartan elements of $H$ in $A$, gauge equivalence amounts to inner automorphism in $A$. We write $\bar f_*$ in the statement to indicate the map on gauge equivalence classes induced from the map of Maurer--Cartan sets \[ f_* : \mathrm{MC}(H,-) \rightarrow \mathrm{MC}(C,-) \cong {\mathrm{Def}}^{\mathrm{nc}, \square}_\rho(-) \] induced by $f : H \rightarrow C$. \begin{proof} By Theorem \ref{thm: decomp}, we have available an $A_\infty$-isomorphism \[ \chi: (C, d_C, m_{2,C}) \buildrel\sim\over\lra (H, m) \oplus (K, d_K). \] Clearly this induces isomorphisms of Maurer--Cartan functors \[ \mathrm{MC}(C, -) \buildrel\sim\over\ra \mathrm{MC}(H \oplus K, -), \quad \overline{\mathrm{MC}}(C, -) \buildrel\sim\over\ra \overline{\mathrm{MC}}(H \oplus K, -). \] Combining Proposition \ref{prop: MC-gauge of C} with Lemmas \ref{lem: contractible MC} and \ref{lem: connected gauge}, and noting that gauge equivalence decomposes along the decomposition $H \oplus K$, the claim follows immediately. \end{proof} \begin{rem} We see in the statement of Theorem \ref{thm: a-inf main} an instance of ``the homotopy invariance of the Maurer--Cartan functor'': after gauge equivalence, it is a quasi-isomorphism invariant of $A_\infty$-algebras. This is well-known but rarely stated in arbitrary characteristic. In this generality, it can be derived from Theorem \ref{thm: decomp} from \cite{CL2017}, Lemma \ref{lem: contractible MC}, and a generalization of Lemma \ref{lem: connected gauge} for minimal $A_\infty$-algebras that we do not require here. \end{rem} We are interested in amplifying Theorem \ref{thm: a-inf main} to give a cohomological presentation for $R^\mathrm{nc}_\rho$ and explicit formulas for representations associated to elements of $\mathrm{MC}(H,A)$. The data $f, \chi$ induced by the homotopy retract is suited for this. Under the assumption that $H^i$ is finite-dimensional for all $i \geq 0$, recall that \begin{equation} \label{eq: def R} R := \frac{\hat T_\mathbb{F} (\Sigma H^1)^*}{(m^*((\Sigma H^2)^*))} \in \hat\cA_\mathbb{F} \end{equation} denote the classical hull of the dg-algebra $\mathrm{Bar}^*(H)$ set up in Example \ref{eg: fin dim classical hull}. \begin{cor} \label{cor: irred case A-inf} Let $E,\rho, C, H$ be as above. Choose a homotopy retract structure on $(H, 0)$ relative to $(C, d_C)$ as in Example \ref{eg:H-sections}, which gives the additional data $(H, m)$, $f$, and $\chi$. Under the additional assumption that $H^i$ is finite-dimensional for all $i \geq 0$, these data \begin{enumerate}[leftmargin=2em] \item determine an isomorphism \[ \rho^u: E^\wedge_\rho \buildrel\sim\over\lra \mathrm{End}_\mathbb{F}(V) \otimes R \] given by, for $x \in E$, \[ x \mapsto \rho(x) + \sum_{i =1}^\infty (\underline{e} \mapsto (f_i(\underline{e}))(x)) \in \mathrm{End}_\mathbb{F}(V) \otimes R \] where $\underline{e}$ is a generic element of $(\Sigma H^1)^{\otimes i} = \Sigma H^1(E, \mathrm{End}_\mathbb{F}(V))^{\otimes i}$. \item Upon the additional choice of an idempotent $e^{11} \in E^\wedge_\rho$ used to define $R^\mathrm{nc}_\rho$, $\rho^u$ induces an isomorphism \[ R^\mathrm{nc}_\rho \buildrel\sim\over\lra R. \] \end{enumerate} \end{cor} We give an explanation of the notation $(\underline{e} \mapsto (f_i(\underline{e}))(x))$. By definition of $f = (f_n)_{n \geq 1}$, we find $f_i(\underline{e}) \in C^1(E, \mathrm{End}_\mathbb{F}(V)) = \mathrm{Hom}_\mathbb{F}(E, \mathrm{End}_\mathbb{F}(V))$, which is a function that can be evaluated on $x \in E$. So, altogether, $(\underline{e} \mapsto (f_i(\underline{e}))(x))$ is an element of $(\Sigma H^1(E, \mathrm{End}_\mathbb{F}(V))^*)^{\otimes i} \otimes \mathrm{End}_\mathbb{F}(V)$. This determines an element of $\mathrm{End}_\mathbb{F}(V) \otimes R$ via the surjection $\hat T (\Sigma H^1)^* \twoheadrightarrow R$ of \eqref{eq: def R}. \begin{rem} Theorem \ref{thm: main A-inf} follows more-or-less directly from Corollary \ref{cor: irred case A-inf}. \end{rem} \begin{proof} First we produce $\rho^u$ and justify its formula. We claim that $\rho^u$ is the $R$-valued lift of $\rho$ arising from the map $\mathrm{Bar}(f) : \mathrm{Bar}(H) \rightarrow \mathrm{Bar}(C)$. Indeed, recall from Theorem \ref{thm: lift corep} that $\mathrm{Bar}(C)$ corepresents the lifting functor ${\mathrm{Def}}^{\mathrm{nc}, \square}_\rho$ on Artinian augmented $\mathbb{F}$-algebras. Then recall that that $R$ is a limit of such algebras, and it is also the classical hull of the dual dg-algebra to $\mathrm{Bar}(H)$. To prove the claim, note that the sum of $f_i : (H^1)^{\otimes i} \rightarrow C^1$ over $i \geq 1$ determines an element of \[ \prod_{i \geq 1} (\Sigma H^1)^* \otimes C^1. \] This element reduces to the Maurer--Cartan element $\xi_R$ in $C^1 \otimes \mathfrak{m}_R$ arising from $\mathrm{Bar}(H) \rightarrow \mathrm{Bar}(C)$. Finally, as in the proof of Theorem \ref{thm: lift corep}, $\xi_R$ is a $\mathbb{F}$-linear map from $E$ to $\mathrm{End}_\mathbb{F}(V) \otimes \mathfrak{m}_R$ that determines a homomorphism $\rho \oplus \xi_R : E \rightarrow \mathrm{End}_\mathbb{F}(V) \otimes R$ that appears in the formula in (1). We observe that its codomain is local and complete, so $\rho \oplus \xi_R$ induces the map $\rho^u$. Now that we know that $\rho^u$ is a homomorphism, we produce \begin{equation} \label{eq: Rnc to R} e \rho^u e : R^\mathrm{nc}_\rho = eE^\wedge_\rho e \longrightarrow R \cong \rho^u(e)(\mathrm{End}_\mathbb{F}(V) \otimes R)\rho^u(e). \end{equation} The following collection of facts implies that $e \rho^u e$ is an isomorphism. Firstly, note that $R$ pro-represents $\mathrm{MC}(H,-)$ on $\cA_\mathbb{F}$ by its definition. Next, Theorem \ref{thm: a-inf main} draws an isomorphism $\overline{\mathrm{MC}}(H, -) \buildrel\sim\over\ra {\mathrm{Def}}^\mathrm{nc}_\rho$. Proposition \ref{prop:def rep} shows that $R^\mathrm{nc}_\rho$ represents ${\mathrm{Def}}^\mathrm{nc}_\rho$ after inner automorphism of the coefficients. Lemma \ref{lem: connected gauge} shows that the projection $\mathrm{MC}(H, -) \twoheadrightarrow \overline{\mathrm{MC}}(H, -)$ amounts to inner automorphism classes in the coefficients in $\cA_\mathbb{F}$. Therefore $R^\mathrm{nc}_\rho$ and $R$ pro-represent the functors on $\cA_\mathbb{F}$ that are identified via the map $\bar f_*$ of Theorem \ref{thm: a-inf main} up to inner automorphism. Because the formula for $\rho^u$ realizes the map $f_*$ discussed after Theorem \ref{thm: a-inf main}, we see that $e \rho^u e$ is compatible with this isomorphism of functors, up to inner automorphism. Therefore $e \rho^u e$ is itself an isomorphism. \end{proof} \section{Non-commutative deformations of multiple points} \label{sec:assoc mult pts} The goal of this section is to generalize the results of \S\ref{sec:assoc pt} to the case where we are deforming multiple points. Stated representation-theoretically, we are deforming a semi-simple representation with distinct simple summands. We will follow the approach of \cite{segal2008} in order to study this problem with multiple-pointed coefficients: see \S1.3 and \S2 of \emph{loc.\ cit}. \subsection{Setting up the data for multiple points} \label{subsec: r-pointed data} We adapt the notation established at the outset of \S\ref{subsec:adt setup}. Now $\rho : E \rightarrow \mathrm{End}_\mathbb{F}(V)$ is a semi-simple representation on a $\mathbb{F}$-vector space $V$. We write $\rho \cong \bigoplus_{i=1}^r \rho_i$, where $\rho_i : E \rightarrow \mathrm{End}_\mathbb{F}(V_i)$ and we fix an isomorphism $V \buildrel\sim\over\ra \bigoplus_{i=1}^r V_i$. We assume that the summands $\rho_i$ are absolutely irreducible and pairwise non-isomorphic. As in \cite[\S1.3]{segal2008}, we use coefficient algebras on $r$ points. We write $\mathbb{F}^r$ for the $r$-times product algebra $\mathbb{F} \times \dotsm \times \mathbb{F}$. \begin{defn} \label{defn: r} For $1 \leq i \leq r$, we write $1_i \in \mathbb{F}^r$ for the element with $1$ concentrated in the $i$th coordinate. For a $\mathbb{F}^r$-bimodule $M$ and $1 \leq i,j \leq r$, we write $M_{ij}$ for $1_i \cdot M \cdot 1_j$, so that $M \cong \bigoplus_{i,j} M_{ij}$. Let $\mathrm{Alg}^r_\mathbb{F}$ denote the \emph{category of $\mathbb{F}^r$-algebras}, that is, associative unital algebra objects in the category of $\mathbb{F}^r$-bimodules. We write $\mathbb{F}^r \in \mathrm{Alg}^r_\mathbb{F}$ for the standard $\mathrm{Alg}^r_\mathbb{F}$ structure on the ring $\mathbb{F}^r$. Then we observe that $\mathbb{F}^r$ is a unit for the symmetric monoidal (tensor) product $\underline{\otimes}$ in $\mathrm{Alg}_\mathbb{F}^r$ by assigning to $A, A' \in \mathrm{Alg}_\mathbb{F}^r$ the vector space \[ (A \underline{\otimes} A')_{ij} := A_{ij} \otimes A'_{ij}, \] with coordinate-wise multiplication. Note that $\underline{\otimes}$ is the underlying tensor product in the category of $\mathbb{F}^r$-bimodules. In contrast, given $\mathbb{F}^r$-bimodules $M,N$, we use $M \otimes_{\mathbb{F}^r} N$ to denote the usual tensor product, using the left $\mathbb{F}^r$-module structure on $M$ and the right $\mathbb{F}^r$-module structure on $N$ to produce the $\mathbb{F}^r$-bimodule structure of $M \otimes_{\mathbb{F}^r} N$. Finally, the undecorated symbol ``$\otimes$'' is understood to be over $\mathbb{F}$, as usual. An \emph{augmentation} of $A \in \mathrm{Alg}^r_\mathbb{F}$ is a morphism $A \rightarrow \mathbb{F}^r$ in $\mathrm{Alg}^r_\mathbb{F}$. Let $\cA_\mathbb{F}^r$ denote the category of augmented $\mathbb{F}^r$-algebras that have finite $\mathbb{F}$-dimension. We write $\mathfrak{m}_A \subset A$ for the augmentation ideal of $A \in \cA_\mathbb{F}^r$, so $A/\mathfrak{m}_A \buildrel\sim\over\ra \mathbb{F}^r$. Given a $\mathbb{F}^r$-bimodule $M$, we have the completed tensor product $\mathbb{F}^r$-algebra \[ \hat T_{\mathbb{F}^r} M := \prod_{i \geq 0} M^{\otimes_{\mathbb{F}^r} i}. \] We understand $\mathrm{End}_\mathbb{F}(V)$ to be a $\mathbb{F}^r$-algebra by sending $1_i$ to the projection operator from $V$ to $V_i$. We will also use the $\mathbb{F}^r$-subalgebra of $\mathrm{End}_\mathbb{F}(V)$, \[ \mathrm{End}_\mathbb{F}^r(V) := \bigoplus_{i=1}^r 1_i \mathrm{End}_\mathbb{F}(V) 1_i \cong \bigoplus_{i=1}^r \mathrm{End}_\mathbb{F}(V_i). \] \end{defn} \begin{warn} Note that the present notion of $\mathbb{F}^r$-algebra is not the same as ``an associative ring receiving a homomorphism from $\mathbb{F}^r$ to its center.'' An $\mathbb{F}^r$ algebra does receive a canonical homomorphism from $\mathbb{F}^r$, but it is not central. See \cite[\S1.3]{segal2008} for equivalent formulations of $\mathrm{Alg}^r_\mathbb{F}$. Most useful in the sequel is the following alternate formulation: an associative $\mathbb{F}$-algebra with an ordered complete set of $r$ orthogonal idempotents. \end{warn} \subsection{Dg-algebras, $A_\infty$-algebras, and representability over multiple points} \label{subsec: r-pointed versions} In Theorem \ref{thm: lift corep}, we found a bijection between associative lifts and Maurer--Cartan elements for the Hochschild complex. We have the following $r$-pointed generalizations of the objects. \subsubsection{Dg-algebras} A dg-$\mathbb{F}^r$-algebra amounts to a $\mathbb{F}$-linear dg-category on $r$ objects (labeled by $\{1, \dotsc, r\}$), or, equivalently, additional $r$-pointed structure on a dg-algebra over $\mathbb{F}$. It will suffice to consider the example we are concerned with: morphisms in this category are the Hochschild cochain complexes \[ \mathrm{Hom}(j, i) := C^\bullet(E, \mathrm{Hom}_\mathbb{F}(\rho_j, \rho_i)) \quad \text{for } 1 \leq i,j \leq r \] (the $E$-bimodule structure of $\mathrm{Hom}_\mathbb{F}(\rho_j, \rho_i)$ is just like \eqref{eq: EE bimodule structure}), where the composition of morphisms arises from \[ \mathrm{Hom}_\mathbb{F}(\rho_k, \rho_j) \otimes \mathrm{Hom}_\mathbb{F}(\rho_j, \rho_i) \rightarrow \mathrm{Hom}_\mathbb{F}(\rho_k, \rho_i). \] This composition is compatible with the Hochschild differential for the same reason as Lemma \ref{lem: Hochs of algebra}. Indeed, it follows from applying the statement of Lemma \ref{lem: Hochs of algebra} (verbatim) to the Hochschild complex of $\mathrm{End}_\mathbb{F}(\rho)$, and then using its $\mathbb{F}^r$-algebra structure to deduce the compatibility for the dg-category. We leave the notion of morphisms to the reader. \subsubsection{$A_\infty$-algebras} Similarly to dg-algebras, we may view an $A_\infty$-$\mathbb{F}^r$-algebra as a $\mathbb{F}$-linear $A_\infty$-category on $r$ objects. That is, $\mathrm{Hom}(j,i)$ is a complex with differential $m_1$, and for $n \geq 2$ and any finite sequence $i_0, \dotsc, i_n$ in $\{1, \dotsc, r\}$, there is a $\mathbb{F}$-linear composition law $m_n$ on \begin{equation} \label{eq: r-pointed m_n} m_n : \mathrm{Hom}(i_0, i_1) \otimes \dotsm \otimes \mathrm{Hom}(i_{n-1}, i_n) \longrightarrow \mathrm{Hom}(i_0, i_n) \text{ of degree } 2-n. \end{equation} The $m=(m_n)_{n \geq 1}$ are required to satisfy the compatibility conditions of \eqref{eq:m_conds}. We will mainly discuss $A_\infty$-$\mathbb{F}^r$-algebra structures on $H^\bullet(E, \mathrm{End}_\mathbb{F}(V))$; namely, \[ \mathrm{Hom}(j,i) = H^\bullet(E, \mathrm{Hom}_\mathbb{F}(\rho_j, \rho_i)). \] The composition $m_n$ on $H^\bullet(E, \mathrm{End}_\mathbb{F}(V))$ is a sum of maps of the form \eqref{eq: r-pointed m_n} by applying the direct sum decomposition $\mathrm{End}_\mathbb{F}(\rho) \cong \bigoplus_{1 \leq i,j\leq r} \mathrm{End}_\mathbb{F}(\rho_j, \rho_i)$. We leave the notion of morphisms to the reader. \subsubsection{Bar construction} The bar construction involves taking linear duals and suspensions, all of which naturally respects $\mathbb{F}^r$-structure. The bar equivalence of Corollary \ref{cor: bar equiv} also generalizes, giving an isomorphism of categories between $A_\infty$-$\mathbb{F}^r$-algebras and cofree cocomplete (over $\mathbb{F}^r$, i.e.\ coaugmented over $\mathbb{F}^r$) dg-$\mathbb{F}^r$-coalgebras. The cofree cocomplete dg-$\mathbb{F}^r$-coalgebra corresponding to an $A_\infty$-$\mathbb{F}^r$-algebra $(H, m)$ is the data of a codifferential on $\hat T^\mathrm{co}_{\mathbb{F}^r} \Sigma H \cong \bigoplus_{i \geq 0} \Sigma H^{\otimes_{\mathbb{F}^r} i}$. When $H^i$ has finite $\mathbb{F}$-dimension for all $i \in \mathbb{Z}$, then we form the dual dg-algebra \[ (\hat T_{\mathbb{F}^r} \Sigma H^*, m^*, \pi). \] \subsubsection{Maurer--Cartan functor} For an $A_\infty$-$\mathbb{F}^r$-algebra $(H, (m_n)_{n \geq 1})$ and any $A \in \cA^r_\mathbb{F}$, a Maurer--Cartan element for $H$ valued in $A$ is some $\xi \in (\mathfrak{m}_A \underline{\otimes} \Sigma H)^0$ such that the Maurer--Cartan equation \eqref{eq: HMC} holds. The functor of Maurer--Cartan elements is corepresentable over $\mathbb{F}^r$, in direct analogy to Proposition \ref{prop:MC-corep}. \subsubsection{Kadeishvili's theorem and the decomposition theorem} \label{sssec: r-pointed KT} Next, in order to discuss deformations, we need an $r$-pointed version of Kadeishvili's theorem (Corollary \ref{cor:A-inf on H}). The key point is that the homotopy retract relating the complex $C$ and its cohomology $H$ should respect $\mathbb{F}^r$-bimodule structure. To emphasize this, we state the $r$-pointed generalization of Definition \ref{defn: HR}. \begin{defn} Let $(A,d_A)$, $(C,d_C)$ be complexes of $\mathbb{F}^r$-bimodules. We call $(A,d_A)$ a \emph{$r$-pointed homotopy retract} of $(C,d_C)$ when they are equipped with maps \[ \xymatrix{ C \ar@(dl,ul)^h \ar@<1ex>[r]^p & A \ar@<1ex>[l]^i } \] such that $p$ and $i$ are morphisms of complexes of $\mathbb{F}^r$-bimodules, $h : C \rightarrow \Sigma^{-1} C$ is a morphism of graded $\mathbb{F}^r$-bimodules, $\mathrm{id}_C - ip = d_Ch + hd_C$, and $i$ is a quasi-isomorphism of complexes of $\mathbb{F}^r$-bimodules. \end{defn} Once this is done, Corollary \ref{cor:A-inf on H} and Theorem \ref{thm: decomp} apply to $A_\infty$-$\mathbb{F}^r$-algebras. \begin{eg} \label{eg: r-pointed retract} To illustrate this for the dg-$\mathbb{F}^r$-algebra $C = C^\bullet(E, \mathrm{End}_\mathbb{F}(V))$ and its cohomology $H$, the point is that the retract datum \[ i: (H, 0) \longrightarrow (C,d_C) \] must lift cohomology classes in $H^\bullet(E, \mathrm{Hom}_\mathbb{F}(\rho_j, \rho_i))$ to a cocycle in the $\mathbb{F}$-subspace \[ C^\bullet(E, \mathrm{Hom}_\mathbb{F}(\rho_j, \rho_i)) \subset C^\bullet(E, \mathrm{End}_\mathbb{F}(V)). \] Just as in Example \ref{eg:H-sections}, this induces a direct sum decomposition of $C^\bullet(E, \mathrm{Hom}_\mathbb{F}(\rho_j, \rho_i))$ into coboundaries, cocycles complementing coboundaries, and cochains complementing cocycles. Then one may readily deduce from Example \ref{eg:KFT1} or \ref{eg:KFT2} that the resulting \begin{itemize}[leftmargin=2em] \item $A_\infty$-structure $m$ on $H$, \item $A_\infty$-quasi-isomorphism $f: H \rightarrow C$, and \item $A_\infty$-isomorphism $\chi : C \rightarrow H \oplus K$ \end{itemize} respect $\mathbb{F}^r$-structure, using the formulas for $m$, $f$, and $\chi$. \end{eg} \subsection{Deformation theory of $r$ points} \label{subsec: r pointed defs} We begin by setting up an an $r$-pointed version of the 1-pointed lifting functor ${\mathrm{Def}}^{\mathrm{nc}, \square}_\rho$ and 1-pointed deformation functor ${\mathrm{Def}}^{\mathrm{nc}}_\rho$ that were defined in Definition \ref{defn: lift assoc}. \begin{rem} \label{rem: r conj options} We comment on the appropriate notion of $r$-pointed notion of strict equivalence: there are two possible options, and we will show that they are equivalent (Proposition \ref{prop: segal auts}) when applied to the appropriate notions of lift. At the least, conjugation of a lift $\rho_A$ of $\rho$ should preserve the lifting property. The largest subgroup of $(\mathrm{End}_\mathbb{F}(V) \underline{\otimes} A)^\times$ that does this is $\mathbb{F}^r + \mathrm{End}_\mathbb{F}(V) \underline{\otimes} \mathfrak{m}_A$, where $\mathbb{F}^r \hookrightarrow \mathrm{End}_\mathbb{F}(V)$ arises from its $\mathbb{F}^r$-algebra structure. Within this subgroup, we can also insist on preserving $\mathbb{F}^r$-structure when conjugating, i.e.\ we demand an inner automorphism of $\mathrm{End}_\mathbb{F}(V) \underline{\otimes} A$ as an $\mathbb{F}^r$-algebra. This subgroup is \[ \mathbb{F}^r + \mathrm{End}_\mathbb{F}^r(V) \underline{\otimes} \mathfrak{m}_A \subset \mathbb{F}^r + \mathrm{End}_\mathbb{F}(V) \underline{\otimes} \mathfrak{m}_A. \] The smaller one is more naturally $r$-pointed. However, we are forced to use the larger relation because $E$ has no natural $r$-pointed structure. \end{rem} \begin{defn} Let $A \in \cA_\mathbb{F}^r$. A \emph{lift of $\rho$ over $A$} is a $\mathbb{F}$-algebra homomorphism $\rho_A : E \rightarrow \mathrm{End}_\mathbb{F}(V) \underline{\otimes} A$ such that $\rho_A \underline{\otimes}_A \mathbb{F}^r = \rho$. A \emph{deformation of $\rho$ over $A$} is an equivalence class of lifts $\rho_A : E \rightarrow \mathrm{End}_\mathbb{F}(V) \underline{\otimes} A$ under the equivalence relation of conjugation by $\mathbb{F}^r + \mathrm{End}_\mathbb{F}(V) \underline{\otimes} \mathfrak{m}_A$. We define the \emph{lifting functor of $\rho$} (resp.\ the \emph{deformation functor of $\rho$}) on $\cA_\mathbb{F}^r$, denoted ${\mathrm{Def}}^{\mathrm{nc},\square}_\rho$ (resp.\ ${\mathrm{Def}}^\mathrm{nc}_\rho$), as the functor from $\cA_\mathbb{F}^r$ to the category of sets sending $A$ to the set of lifts (resp.\ deformations) of $\rho$ over $A$. \end{defn} To produce $r$-pointed notions of lift and deformation, we begin with $E \rightarrow E^\wedge_\rho$, defined in \eqref{eq: def Erho}. Let $\bar e_i \in \mathrm{End}_\mathbb{F}(V_i)$ be a projection operator onto a 1-dimensional subspace of $V_i$, and let $\bar e = \sum_1^r \bar e_i \in \mathrm{End}_\mathbb{F}^r(V)$. Choose orthogonal idempotent lifts $e_i \in E^\wedge_\rho$ of $\bar e_i$ via $\rho_i : E^\wedge_\rho \twoheadrightarrow \mathrm{End}_\mathbb{F}(V_i)$. Letting $e = \sum_{i=1}^r e_i$, we get a Morita equivalence between $E^\wedge_\rho$ and \[ R^\mathrm{nc,r}_\rho := e E^\wedge_\rho e \in \hat \cA_\mathbb{F}^r, \] where the $\mathbb{F}^r$-algebra structure on $R^{\mathrm{nc},r}_\rho$ is determined by $(e_i)_{i=1}^r$. The inverse equivalence on algebras is realized by \begin{equation} \label{eq: r isom lift} E^\wedge_\rho \simeq (M_{d_i \times d_j}((R^{\mathrm{nc},r}_\rho)_{i,j}))_{i,j} \simeq \mathrm{End}_\mathbb{F}(V) \underline{\otimes} R^{\mathrm{nc},r}_\rho; \end{equation} for a proof of this, apply \cite[Thm.\ 1.2]{laudal2002} in the limit on $E/\ker(\rho)^i$. Now that we have a choice \eqref{eq: r isom lift} of $\mathbb{F}^r$-algebra structure on $E^\wedge_\rho$, we can set up $r$-pointed lifting and deformation functors. \begin{defn} Let $A \in \cA_\mathbb{F}^r$. A \emph{$r$-lift of $\rho$ over $A$} is a $\mathbb{F}^r$-algebra homomorphism $\rho_A : E^\wedge_\rho \rightarrow \mathrm{End}_\mathbb{F}(V) \underline{\otimes} A$ such that $\rho_A \underline{\otimes}_A \mathbb{F}^r = \rho$. A \emph{$r$-deformation of $\rho$ over $A$} is an equivalence class of $r$-lifts $\rho_A : E^\wedge_\rho \rightarrow \mathrm{End}_\mathbb{F}(V) \underline{\otimes} A$ under the equivalence relation of conjugation by $\mathbb{F}^r + \mathrm{End}_\mathbb{F}^r(V) \underline{\otimes} \mathfrak{m}_A$. We define the \emph{lifting functor of $\rho$} (resp.\ the \emph{deformation functor of $\rho$}) on $\cA_\mathbb{F}^r$, denoted ${\mathrm{Def}}^{\mathrm{nc},\square,r}_\rho$ (resp.\ ${\mathrm{Def}}^{\mathrm{nc},r}_\rho$), as the functor from $\cA_\mathbb{F}^r$ to the category of sets sending $A$ to the set of lifts (resp.\ deformations) of $\rho$ over $A$. \end{defn} While there is clearly a natural proper inclusion of lifting functors on $\cA_\mathbb{F}^r$ \[ {\mathrm{Def}}^{\mathrm{nc}, \square,r}_\rho \hookrightarrow {\mathrm{Def}}^{\mathrm{nc}, \square}_\rho, \] the two notions of deformation are equivalent. \begin{prop}[Segal] \label{prop: segal auts} There is a natural isomorphism of functors on $\cA_\mathbb{F}^r$ \[ {\mathrm{Def}}^{\mathrm{nc},r}_\rho \buildrel\sim\over\lra {\mathrm{Def}}^{\mathrm{nc}}_\rho. \] \end{prop} \begin{proof} This is \cite[Prop.\ 2.11 and Lem.\ 2.12]{segal2008}. \end{proof} Having set up the deformation theory, we examine what arises from the natural $r$-pointed structures on the Hochschild cochain dg-$\mathbb{F}^r$-algebra $C = C^\bullet(E, \mathrm{End}_\mathbb{F}(V))$. We immediately find the following $r$-pointed version of Theorem \ref{thm: lift corep}, where the left isomorphism is simply the representability of the Maurer--Cartan functor. \begin{thm} \label{thm: r lift corep} Let $A \in \cA^r_\mathbb{F}$. Let be the Hochschild cochain dg-$\mathbb{F}^r$-algebra. There are canonical isomorphisms \begin{equation} \label{eq:MC lift rep mult} \mathrm{Hom}_{\mathbb{F}^r\text{-dgca}}(A^\vee, \mathrm{Bar}(C)) \buildrel\sim\over\lra \mathrm{MC}_{\mathbb{F}^r}(C,A) \buildrel\sim\over\lra {\mathrm{Def}}^{\mathrm{nc}, \square}_\rho(A). \end{equation} \end{thm} Similarly, in analogy to Proposition \ref{prop:def rep} and its proof, we can find a tautological construction representing the deformation functor in terms of $E^\wedge_\rho$ the choice of idempotent made to construct $R^{\mathrm{nc},r}_\rho$ and give $E^\wedge_\rho$ the structure of a $\mathbb{F}^r$-algebra. We apply the Morita equivalence of categories explicitly given by \begin{align} \label{eq: r morita} \begin{split} E^\wedge_\rho\text{-Mod} &\buildrel\sim\over\lra R^{\mathrm{nc},r}_\rho\text{-Mod} \\ W &\mapsto eW \\ V \otimes_{\mathbb{F}^r} Y &\mapsfrom Y \end{split} \end{align} generalizing the case $r=1$ of \eqref{eq: morita}. \begin{prop} \label{prop: r def rep} ${\mathrm{Def}}^{\mathrm{nc}, r}_\rho$ is isomorphic to the $\mathbb{F}$-inner automorphism quotient of the $\mathrm{Hom}_{\mathbb{F}^r}$-functor on $\cA_\mathbb{F}^r$ of $R^{\mathrm{nc},r}_\rho$. That is, there is a functorial isomorphism \[ {\mathrm{Def}}^{\mathrm{nc}, r}_\rho(A) \buildrel\sim\over\lra \mathrm{Hom}_{\mathbb{F}^r}(R^{\mathrm{nc},r}_\rho, A)/\sim_{A,r} \] over $A \in \cA_\mathbb{F}^r$, where $\sim_{A,r}$ indicates the equivalence relation of inner $\mathbb{F}^r$-algebra isomorphisms of $A$. This isomorphism is given by applying $(-) \underline{\otimes} \mathrm{End}_\mathbb{F}(V)$ to a representative $\eta : R^{\mathrm{nc},r}_\rho \rightarrow A$ of an element of $\mathrm{Hom}_{\mathbb{F}^r}(R^{\mathrm{nc},r}_\rho, A)/\sim_{A,r}$. \end{prop} Finally, we prove that ${\mathrm{Def}}^\mathrm{nc}_\rho$ is representable by $C$ when taken up to gauge equivalence and $\mathbb{F}^r$-conjugation. The proof is exactly as in the 1-pointed version of Proposition \ref{prop: MC-gauge of C}. The one point of difference is that $\mathbb{F}^r$ is no longer central as it was when $r=1$, so the strict gauge action of conjugation by $1 + \mathrm{End}_\mathbb{F}(\rho) \underline{\otimes} \mathfrak{m}_A$ must be followed by conjugation by $(\mathbb{F}^r)^\times \cong \mathrm{Aut}_E(\rho)$. \begin{prop} \label{prop: r MC-gauge of C} We have the dg-$\mathbb{F}^r$-algebra $C = C^\bullet(E, \mathrm{End}_\mathbb{F}(\rho))$. For $A \in \cA_\mathbb{F}^r$, there is the following natural bijection between $A$-valued Maurer--Cartan elements for $C$ and lifts of $\rho$ to $A$. That is, $\overline{\mathrm{MC}}(C,A)/(\mathbb{F}^r)^\times \buildrel\sim\over\ra {\mathrm{Def}}^{\mathrm{nc}}_\rho(A)$. \end{prop} \subsection{$A_\infty$-algebras and deformations over multiple points} We aim for an analogue of Theorem \ref{thm: r lift corep} expressing ${\mathrm{Def}}^{\mathrm{nc}, r}_\rho$ in terms of cohomological data. We use the setup for the $r$-pointed Kadeishvili's theorem and decomposition theorem from \S\ref{sssec: r-pointed KT}. After assembling a few more lemmas, we apply very similar arguments to the 1-pointed case to prove the main theorems. From above, we have $E$, $\rho$, and $C$. Let \[ H := H^\bullet(C) = H^\bullet(E, \mathrm{End}_\mathbb{F}(\rho)) \] denote Hochschild cohomology of $\mathrm{End}_\mathbb{F}(\rho)$. We choose an $r$-pointed homotopy retract structure on $(H, 0)$ relative to $(C, d_C)$, obtaining all of the decompositions and structures listed in Example \ref{eg: r-pointed retract}. We use all of the notation therein. The following $r$-pointed analogues of Lemmas \ref{lem: contractible MC} and \ref{lem: connected gauge} are needed. \begin{lem} \label{lem: r contractible MC} For a trivial $A_\infty$-algebra $(K,d_K)$, one has $\overline{\mathrm{MC}}(K,-) = \ast$ and \[ \mathrm{MC}(K,A) \cong B^1 \underline{\otimes} \mathfrak{m}_A \] for $A \in \cA_\mathbb{F}^r$. For the particular trivial $A_\infty$-$\mathbb{F}^r$-algebra $(K,d_K)$ produced from $C = C^\bullet(E, \mathrm{End}_\mathbb{F}(\rho))$, along with a homotopy retract on $C$ given in Example \ref{eg: r-pointed retract}, we have \[ \mathrm{MC}(K, A) \cong \frac{\mathrm{End}_\mathbb{F}(\rho)}{\mathrm{diag}(\mathbb{F}^r)} \underline{\otimes} \mathfrak{m}_A, \] where $\mathrm{diag}(\mathbb{F}^r)$ denotes the product of scalar matrices in $\mathrm{End}_\mathbb{F}^r(\rho)$. \end{lem} \begin{proof} Same as that of Lemma \ref{lem: contractible MC}. \end{proof} \begin{lem} \label{lem: r connected gauge} Assume that $H^i$ is finite-dimensional for all $i \in \mathbb{Z}$, so that $\mathrm{MC}(H,-)$ is pro-represented on $\cA_\mathbb{F}$ by the classical hull $R$ of $\mathrm{Bar}^*(H) \in \cA_\mathbb{F}^\mathrm{dg}$ (see Example \ref{eg: fin dim classical hull}). For $A \in \cA_\mathbb{F}$, this pro-representability maps strict equivalence classes in $\mathrm{MC}(H,A)$ isomorphically onto strict $\mathbb{F}^r$-inner automorphism classes in $\mathrm{Hom}_\mathbb{F}(R, A)$. \end{lem} Just as $\mathbb{F}^r$-inner automorphism of $A \in \cA^r_\mathbb{F}$ means conjugation by $\mathbb{F}^r \underline{\otimes} A \cong \bigoplus_{i=1}^r A_{i,i}$, strict $\mathbb{F}^r$-inner automorphism refers to conjugation by $1 + \mathbb{F}^r \underline{\otimes} \mathfrak{m}_A$. \begin{proof} The key point is that $H$ is itself augmented as an $A_\infty$-$\mathbb{F}^r$-algebra, in the sense that $H^0 \cong \mathbb{F}^r$, arising from the center of $C$. The proof then proceeds with the same calculations as in Lemma \ref{lem: connected gauge}. We arrive at conjugation by $1-\gamma$, where \[ \gamma \in H^0 \underline{\otimes} \mathfrak{m}_A \cong \mathbb{F}^r \underline{\otimes} \mathfrak{m}_A \cong \bigoplus_{i=1}^r (\mathfrak{m}_A)_{i,i}. \qedhere \] \end{proof} Now we deduce an $r$-pointed analogue of Theorem \ref{thm: a-inf main}. \begin{thm} \label{thm: r-pointed a-inf} Let the data $E$, $\rho$, $C$, $H$ be as above. Choose an $r$-pointed homotopy retract between $H$ and $C$ as in Example \ref{eg: r-pointed retract}, inducing $m$, $f$, and $\chi$. For any $A \in \cA_\mathbb{F}^r$, these data determine functorial isomorphisms \[ \overline{\mathrm{MC}}_{\mathbb{F}^r}(H,A)/(\mathbb{F}^r)^\times \underset{\bar f_*}{\buildrel\sim\over\lra} \overline{\mathrm{MC}}_{\mathbb{F}^r}(C,A)/(\mathbb{F}^r)^\times \buildrel\sim\over\lra {\mathrm{Def}}^\mathrm{nc}_\rho(A). \] \end{thm} As in the 1-pointed version, we write $\bar f_*$ to indicate the map on gauge equivalence classes induced from the map of Maurer--Cartan sets \[ f_* : \mathrm{MC}(H,-) \rightarrow \mathrm{MC}(C,-) \cong {\mathrm{Def}}^{\mathrm{nc}, \square}_\rho(-) \] induced by $f : H \rightarrow C$. \begin{proof} Analogously to the proof of Theorem \ref{thm: a-inf main}, this follows straightforwardly from the $r$-pointed version of the decomposition Theorem \ref{thm: decomp} discussed in \S\ref{sssec: r-pointed KT}, the conclusions about the gauge action in Lemmas \ref{lem: r contractible MC} and \ref{lem: r connected gauge}, and the rightmost isomorphism of the theorem statement from Proposition \ref{prop: r MC-gauge of C} \end{proof} Under the assumption that $H^i$ is has finite $\mathbb{F}$-dimension for all $i \in \mathbb{Z}$, we consider the augmented $\mathbb{F}^r$-algebra \begin{equation} \label{eq: r pointed hull} R = \frac{\hat T_{\mathbb{F}^r} (\Sigma H^1)^*}{(m^*((\Sigma H^2)^*))}, \end{equation} which is the classical hull of the augmented dg-$\mathbb{F}^r$-algebra $\mathrm{Bar}^*(H)$ that is dual to the dg-$\mathbb{F}^r$-coalgebra produced by the bar construction (Fact \ref{fact: bar}). This directly generalizes the 1-pointed expression of \eqref{eq: def R}. Then, we prove an explicit relationship between $E^\wedge_\rho$ and $R$. \begin{cor} \label{cor: r-pointed dual} We assume the setting of Theorem \ref{thm: r-pointed a-inf}, so that we have the data $E$, $\rho$, $C$, $H$, $m$, $f$, and $\chi$. Under the additional assumption that $H^i$ is finite-dimensional for all $i \geq 0$, these data \begin{enumerate}[leftmargin=2em] \item determine an isomorphism of $\mathbb{F}$-algebras \[ \rho^u: E^\wedge_\rho \buildrel\sim\over\lra \mathrm{End}_\mathbb{F}(V) \underline{\otimes} R \] given by, for $x \in E$, \[ x \mapsto \rho(x) + \sum_{i =1}^\infty (\underline{e} \mapsto (f_i(\underline{e}))(x)) \in \mathrm{End}_\mathbb{F}(V) \otimes R \] where $\underline{e}$ is a generic element of $(\Sigma H^1)^{\otimes i} = \Sigma H^1(E, \mathrm{End}_\mathbb{F}(V))^{\otimes i}$. \item Upon the additional choice of an idempotents used to give $E^\wedge_\rho$ an $\mathbb{F}^r$-algebra structure and define $R^{\mathrm{nc},r}_\rho$ (see \eqref{eq: r isom lift}), $\rho^u$ induces an isomorphism in $\hat\cA_\mathbb{F}^r$ \[ R^\mathrm{nc}_\rho \buildrel\sim\over\lra R. \] \end{enumerate} \end{cor} The notation $(\underline{e} \mapsto (f_i(\underline{e}))(\gamma))$ has mostly the same as in Corollary \ref{cor: irred case A-inf}, with the sole exception that it is correct here to use $\underline{\otimes}$: $f$ will send $\underline{e}_{i,j}$ to $\mathrm{Hom}_\mathbb{F}(V_j, V_i)$, as we insisted that the homotopy retract respect $\mathbb{F}^r$-structure. \begin{rem} As discussed in \S\ref{subsec: NCDT}, Corollary \ref{cor: r-pointed dual} refines results of Segal \cite[Thm.\ 2.14]{segal2008} and Laudal \cite{laudal2002}. \end{rem} \begin{proof} The proof proceeds just as the proof of Corollary \ref{cor: irred case A-inf}. The formula for $\rho^u$ is identical, and respects $\mathbb{F}^r$-structure because the retract structures and (therefore) the $A_\infty$-structures and homomorphisms do so. We deduce the isomorphism (2), from which (1) follows. The choice of idempotents yields $e\rho^u e : R^{\mathrm{nc},r}_\rho \rightarrow R$ exactly as in \eqref{eq: Rnc to R}. Theorem \ref{thm: r-pointed a-inf} draws an isomorphism $\overline{\mathrm{MC}}(H, -)/(\mathbb{F}^r)^\times \buildrel\sim\over\ra {\mathrm{Def}}^\mathrm{nc}_\rho$. Proposition \ref{prop: r def rep} shows that $R^{\mathrm{nc},r}_\rho$ represents ${\mathrm{Def}}^{\mathrm{nc},r}_\rho$ up to $\mathbb{F}^r$-inner automorphism of the coefficients. By Proposition \ref{prop: segal auts}, there is an isomorphism of functors ${\mathrm{Def}}^{\mathrm{nc},r}_\rho \buildrel\sim\over\ra {\mathrm{Def}}^\mathrm{nc}_\rho$. Lemma \ref{lem: r connected gauge} shows that the projection $\mathrm{MC}(H, -) \twoheadrightarrow \overline{\mathrm{MC}}(H, -)$ amounts to $\mathbb{F}^r$-inner automorphism classes in the coefficients in $\cA_\mathbb{F}^r$. Recall that $R$ pro-represents $\mathrm{MC}(H,-)$ on $\cA_\mathbb{F}^r$ by its definition. Putting together these isomorphisms, we deduce that $R^\mathrm{nc}_\rho$ and $R$ pro-represent functors on $\cA_\mathbb{F}^r$ that are isomorphic via the map $\bar f_*$ of Theorem \ref{thm: r-pointed a-inf} (up to $\mathbb{F}^r$-inner automorphism). Because the formula for $\rho^u$ realizes the map $f_*$ discussed after Theorem \ref{thm: r-pointed a-inf}, we see that $e \rho^u e$ is compatible with this isomorphism of functors, up to inner automorphism. Therefore $e \rho^u e$ is itself an isomorphism. \end{proof} \section{Massey products} \label{sec: massey} The point of this section is to introduce Massey products, in preparation for the explanation of \S\ref{sec: lifts-massey} of the relationship between lifts of representations and Massey products. Here, we focus on the relationship between Massey products and $A_\infty$-products; mainly, we follow \cite{LPWZ2009}. \begin{rem} Massey products were first introduced in topology by Massey and Massey--Uehara \cite{massey1958,MU1957}. For introductions relatively similar to our approach, see Kraines \cite{kraines1966}, May \cite{may1969}, and Dwyer \cite[\S2]{dwyer1975}. \end{rem} \subsection{Massey products in dg-algebras} In this section, we depart from the notation of other sections in order to write $C = (C^\bullet ,\partial, \smile)$ for a dg-$\mathbb{F}$-algebra, possibly non-unital as usual. Let $H = H^\bullet(C)$ be its cohomology. A Massey product of degree $n$ is a multi-valued cohomology operation $H^{\otimes n} \rightarrow H$ of cohomological degree $2-n$. They are not always defined: each value arises from a defining system. Presently, we will introduce these notions in detail. The second Massey product $\langle \,, \rangle : H^{\otimes 2} \rightarrow H$ is unambiguously and unconditionally defined: is the reduction of $\smile$ modulo coboundaries, i.e.\ the cup product. Further Massey products are defined as follows. We establish the notation $\bar \sigma = (-1)^{i+1}\sigma$ for $\sigma \in C^i$ for this general definition. Note that in our main case of interest where $d=1$, we have $\bar \sigma = \sigma$. We also remind the reader that we write $C^i \supset Z^i \supset B^i$ for the cocycle and coboundary subobjects. \begin{rem} There are at least two sign conventions used for Massey products. We follow May \cite{may1969} and \cite{LPWZ2009}, in contrast to \cite{kraines1966} and \cite{LV2012}. \end{rem} \begin{defn} \label{defn:massey_product} Let $n \geq 3$. Let $I_n$ be the set of pairs of integers $(i,j)$ such that $1 \leq i \leq j \leq n$ and $(i,j) \neq (1,n)$. Let $\sigma_i \in Z^{d_i} \subset C^{d_i}$ be cocycles for $1 \leq i \leq n$. For $(i,j) \in I \cup \{(1,n)\}$, let \[ d(i,j) = -(i-j) + \sum_{k=i}^j d_i. \] We say that a set $\mathfrak{S}=\{\sigma(i,j) \in C^{d(i,j)} : (i,j) \in I\}$ is a \emph{defining system for the $n$th Massey product} $\dia{\sigma_1,\dots, \sigma_n}$ if \begin{enumerate}[leftmargin=2em] \item $\sigma(i,i)=\sigma_i$ for all $i=1,\dots n$, and \item $\displaystyle \partial\sigma(i,j) = \sum_{k=i}^{j-1} \bar\sigma(i,k)\smile \sigma(k+1,j)$ for all $(i,j) \in I_n$ such that $i < j$. \end{enumerate} When $\mathfrak{S}$ is a defining system for $\dia{\sigma_1,\dots,\sigma_n}$, we note that \[ c(\mathfrak{S}) := \sum_{k=1}^{n-1} \bar\sigma(1,k)\smile \sigma(k+1,n) \] is an element of $Z^{d(1,n)+1}$ and we let $\dia{\sigma_1,\dots, \sigma_n}_\mathfrak{S} \in H^{d(1,n)+1}$ be the class of $c(\mathfrak{S})$. We let \[ \dia{\sigma_1,\dotsc, \sigma_n} = \{\dia{\sigma_1,\dots, \sigma_n}_\mathfrak{S} \} \subset H^{d(1,n)+1} \] where $\mathfrak{S}$ ranges over all defining systems. It may be empty. Call $\dia{\sigma_1,\dots, \sigma_n}$ \emph{defined} if it is non-empty (i.e.\ there exists a defining system), and say that $\dia{\sigma_1,\dots, \sigma_n}$ \emph{contains zero} if $0 \in \dia{\sigma_1,\dots, \sigma_n}$. \end{defn} It is known that the set $\dia{\sigma_1,\dots, \sigma_n}$ only depends on the cohomology classes of $\sigma_1, \dots, \sigma_n$ \cite[Thm.\ 3]{kraines1966}. Also, note that $d(1,n)+1 = 2-n+ \sum_{k=1}^n d_i$, confirming that the $n$th Massey product has cohomological degree $2-n$ (as a multi-valued map). Finally, note that $\dia{\sigma_1, \dotsc, \sigma_n}$ is defined only if \begin{itemize}[leftmargin=2em] \item all lower-degree Massey products on proper sub-words of $\sigma_1\sigma_2\dotsm\sigma_n$ are defined, and \item all of these contain zero. \end{itemize} In the sequel, we have $d_i = 1$ for all $i$; therefore all such Massey products are valued in $H^2$. \subsection{Massey powers in dg-algebras} \label{subsec: Massey powers} Let $C = (C, \partial, \smile)$ continue to represent a dg-algebra. The following (non-standard) notion of \emph{Massey power} will be useful for our applications. Due to our attention to this application, we only discuss Massey powers of elements of $C^1$. \begin{defn} \label{defn: Massey powers} Let $\tau \in Z^1$, and let $\tau_1 := \tau, \tau_2, \dotsc, \tau_{n-1} \in C^1$, where $n \geq 3$. We say that $\mathfrak{T}:=\{\tau_1, \dots, \tau_{n-1}\}$ \emph{is a defining system $\mathfrak{S}$ for the $n$th Massey power} $\dia{\tau}^n$ if the set \[ \mathfrak{S} = \mathfrak{S}(\mathfrak{T}) := \{\sigma(i,j)=\tau_{j-i+1} : 1 \le i \le j \le n, (i,j) \ne (1,n)\} \] is a defining system for the Massey product $\dia{\tau, \dots, \tau}$ (with $\tau$ repeated $n$ times). A defining system (for the product) arising in this manner is called \emph{symmetric}. If $\mathfrak{T}$ is a defining system for the Massey power $\dia{\tau}^n$, then we let $\dia{\tau}^n_\mathfrak{T} :=\dia{\tau,\dots,\tau}_{\mathfrak{S}(\mathfrak{T})}$, and we let $c(\mathfrak{T}):=c(\mathfrak{S})$. We let \[ \dia{a}^n = \{ \dia{a}^n_\mathfrak{T}\} \subset H^2 \] where $\mathfrak{T}$ ranges over defining systems for the Massey powers. \end{defn} We make the following important observations. \begin{itemize}[leftmargin=2em] \item $\dia{\tau}^n \subset \dia{\tau, \dots, \tau}$, properly in general. \item $\mathfrak{T}=\{\tau_1, \dots, \tau_{n-1}\} \subset C^1$ is a defining system for the Massey power $\dia{\tau}^n$ if and only if $\tau_1=\tau$ and, for all $i=1, \dots, n-1$, we have \begin{equation} \label{eq:massey power reln} d\tau_i = \sum_{j=1}^{i-1} \tau_j \smile \tau_{i-j}. \end{equation} \item for such $\mathfrak{T}$, we have \[ c(\mathfrak{T}) = \sum_{j=1}^{n-1} \tau_j \smile \tau_{n-j}. \] \item The cohomology classes of $\dia{a}^n$ do not depend on the choice of $a$ within its cohomology class. \end{itemize} \subsection{The relationship between Massey products and $A_\infty$-products} \label{subsec: A-inf to Massey} Recall from Definition \ref{defn:A-infinity} that an $A_\infty$-algebra structure $m$ on a graded $\mathbb{F}$-vector space $H$ consists of maps $m_n : H^{\otimes n} \rightarrow H$ of homogeneous degree $2-n$. Recall also from Corollary \ref{cor:A-inf on H} that when $C = (C^\bullet, \partial, \smile)$ is a dg-$\mathbb{F}$-algebra, then there are various compatible choices of $A_\infty$-algebra structure $m$ on its cohomology $H = H^\bullet(C)$. For example, homotopy retracts between $H$ and $C$ induce such an $m$, according to Example \ref{eg:KFT2}. This suggests a relationship between Massey products and $A_\infty$-products. Following \cite{LPWZ2009}, we discuss the relationship between these two notions. \begin{prop} \label{prop:A-inf vs Massey} Let $(C, \partial, \smile)$ be a dg-algebra with cohomology $H = H^*(C)$. We fix cohomology classes $a_i\in H^{d_i}$ for $1 \leq i \leq n$, where $n \geq 3$. Choose in addition the data of a homotopy retract \[ \xymatrix{ *[r]{\;(C, d_C)\;} \ar@<0.5ex>[rr]^{p} \ar@(dl,ul)[]^{h} && *[l]{\;(H,0)\,} \ar@<0.5ex>[ll]^{f} } \] as in Example \ref{eg:H-sections}, specifying an $A_\infty$-algebra $(H, (m_n)_{n\geq 2})$ and a quasi-isomorphism $f : (C, \partial, \smile) \rightarrow (H, (m_n)_{n \geq 2})$ as in Example \ref{eg:KFT1}. \begin{enumerate}[label=(\alph*), leftmargin=2em] \item Assume that for all $i$, $2 \leq i \leq n-1$ and all sub-$i$-tuples $(a_j, \dotsc, a_{j+i-1})$ of $(a_1, \dotsc, a_n)$, it is the case that $m_i(a_j \otimes \dotsm \otimes a_{j+i-1}) = 0$. \item Define \begin{equation} \label{eq:massey to A-inf} a(i,j) = f_{j-i+1}(a_i \otimes \dotsm \otimes a_j). \end{equation} \end{enumerate} Then $(-1)^b m_n(a_1 \otimes \dotsm \otimes a_n)$ is the element \begin{equation} \label{eq:mass A-inf equality} \langle a_1, \dotsc, a_n \rangle_D = \sum_{i=1}^{n-1} \bar a(i,n-i) \smile a(n-i, i) \end{equation} of the Massey product $\langle a_1, \dotsc, a_n \rangle$ arising from the defining system $D=\{a(i,j): 1 \le i \le j \le n, (i,j) \ne (1,n)\}$, where \[ b = 1 + d_{n-1} + d_{n-3} + \dotsm \] is a sum with final term $d_1$ or $d_2$. \end{prop} We remark that condition (a) does depend on the choice of retract. \begin{proof} Using induction on $n$, we will show that the proposition follows directly from part (8) of Example \ref{eg:A-inf}. As \emph{loc.\ cit.} notes, assumption (a) implies that the relation \eqref{eq:A-inf to Massey} holds when evaluated on $a_1 \otimes \dotsm \otimes a_n$. Using the induction step, assumption (a) also implies that $\partial f_i(a_j \otimes \dotsm \otimes a_{j+i-1})$ is equal to \[ -m_{2,C} (f_1 \otimes f_{i-1} - f_2 \otimes f_{i-2} + \dotsm + (-1)^i f_{i-1} \otimes f_1)(a_j \otimes \dotsm \otimes a_{j+i-1}) \] (where we use $m_{2,C}$ to represent the $\smile$ map applied to tensors here). Using definition (b) for the $a(i,j)$, we see that \eqref{eq:A-inf to Massey} states that $\langle a_1, \dotsc, a_n \rangle_D$ is a member of the cohomology class $m_n(a_1 \otimes \dotsm \otimes a_n)$, as desired, up to some sign. From \cite[Thm.\ 3.1]{LPWZ2009}, this sign is given by $b$. \end{proof} \begin{eg} \label{eg:massey-Ainf-deg1} In particular, for our case of interest where $d_i = 1$ for all $i$, $b = (-1)^{(n+1)(n+2)/2}$. This coincides with the signs in the homotopy Maurer--Cartan equation \eqref{eq: HMC}. Moreover, when $d_i = 1$ for all $i$, this formula for $b$ extends to the case $n=2$ (from $n\geq 3$ as in the statement of Proposition \ref{prop:A-inf vs Massey}), where we have set the Massey product equal to the cup product. \end{eg} The natural converse to Proposition \ref{prop:A-inf vs Massey} is not true, because there exist defining systems that cannot arise from a fixed set of $(f_i)$ as in \eqref{eq:massey to A-inf}. For example, consider a cup product $\langle a, a\rangle = 0$, and a defining system $D$ for the triple Massey product $\langle a,a,a\rangle_D = a(1,1) \smile a(2,3) + a(1,2) \smile a(3,3)$. If $a(1,2) \neq a(2,3)$, then \eqref{eq:massey to A-inf} is not possible. Of course, sufficient conditions such that a Massey product arises as in \ref{prop:A-inf vs Massey}(b) could be made clear. For our purposes, it suffices to produce conditions guaranteeing that Massey powers arise from an $A_\infty$-structure arising from a homotopy retract, as follows. \begin{prop} Let $(C,\partial, \smile)$ be a dg-algebra and choose $\tau \in H^1(C)$. Let $\mathfrak{T} = \{\tau_1, \tau_2, \dotsc, \tau_{n-1}\}$ be a defining system for the Massey power $\langle \tau\rangle^n$. The following claims are equivalent. \begin{enumerate}[leftmargin=2em] \item Then there exists a choice of retract of $(C, \partial)$ by $(H^*(C), 0)$ as in Examples \ref{eg:H-sections} such that $\tau_i = f_i(a^{\otimes i})$. \item There exists a section $h^2 : B^2(C) \rightarrow C^1$ of $\partial\vert_{C^1}$ such that $h^2(\sum_{j=1}^{i-1} \tau_j \smile \tau_{i-j}) = \tau_i$. In this case, $m_n(a^{\otimes n}) = \langle \tau \rangle^n_\mathfrak{T}$. \end{enumerate} \end{prop} \begin{proof} Given $h^2$ as in (2), one defines $i^1 : H^1(C) \rightarrow C^1$ so that $\tau_1 = i^1(\tau)$ and defines $p^1 : C^1 \rightarrow H^1(C)$ so that it kills $h^2(B^2(C))$ (which is possible because $h^2(B^2(C))$ is linearly disjoint from $\ker(d^1 : C^1 \rightarrow B^2)$). Clearly these $h^2, i^1, p^1$ can be extended to a retract $h, i,p$ between $H$ and $C$. The converse is clear, in view of the Massey power defining system identities \eqref{eq:massey power reln}. \end{proof} \section{Massey products and non-commutative deformations} \label{sec: lifts-massey} The main point of this section is to flesh out the introductory illustration, in \S\ref{subsec: illustrate}, that the Massey products and Massey powers of \S\ref{sec: massey} in the Hochschild cohomology of $\mathrm{End}_\mathbb{F}(\rho)$ are intrinsically related to lifts of $\rho$. More specifically, we will illustrate that the computations involved in Massey products and their defining systems are exactly those calculations that one needs to inductively choose and compute lifts $\mathbb{F}$-linearly. We are motivated in part by known applications in number theory of cup products -- especially Nekov\'a\v{r}'s height pairings (see \cite[11.5.5]{nekovar2006}) -- and an expectation that higher Massey products may be involved in analogous applications. \begin{eg} Massey powers have been used to calculate an invariant that controls congruences of modular forms, answering a question of Mazur. This is a main theorem of the author's joint work with Wake \cite{WWE3}; see \S\ref{subsec: Mazur Eis}. \end{eg} \begin{rem} \label{rem: A-inf vs Massey} On the same token, we expect that the reader will conclude from these computations that the approach using $A_\infty$-algebras in in \S\S\ref{sec:assoc pt}-\ref{sec:assoc mult pts} is more efficient, thanks to its rigidity. By ``rigidity,'' we are referring to Proposition \ref{prop:A-inf vs Massey}'s statement that the choice of a homotopy retract ``chooses all Massey products in advance.'' (Indeed, it fixes all $A_\infty$-products and also induces the data of Massey defining systems of degree $n+1$, when the appropriate degree $\leq n$ products vanish.) \end{rem} \begin{rem} Logically speaking, one could combine \begin{itemize}[leftmargin=2em] \item the results relating $A_\infty$-algebras to non-commutative deformations in \S\S\ref{sec:assoc pt}-\ref{sec:assoc mult pts} with \item the content of \S\ref{subsec: A-inf to Massey}, which shows that any $A_\infty$-product can be computed in terms of a Massey product (but the converse relation is not true) \end{itemize} to prove a theorem computing non-commutative deformations in terms of Massey products, along the lines of \cite{laudal2002}. But we do not leverage this implication; instead, in this section, we directly link Massey products to deformations. \end{rem} At the end of this section, there are remarks on perspectives from which Massey powers and products are superior to $A_\infty$-products. \subsection{Iterated extensions of representations} \label{subsec: iter} Let $\rho_i : E \rightarrow M_{d_i}(\mathbb{F})$, $1 \leq i \leq n$, be a sequence of representations of $E$. Let \[ \sigma_{i+1,i} \in Z^1(E, \mathrm{Hom}_\mathbb{F}(\rho_i, \rho_{i+1})) \quad \text{for } 1 \leq i < n \] be representatives of extension classes. Let $d = \sum_{i=1}^n d_i$. Assume that there exists a representation $\eta_{n} : E \rightarrow M_d(\mathbb{F})$ that realizes the $\sigma_{i+1}$ below the diagonal, in the sense that there exist $\sigma_{i,j}$ such that \begin{equation} \label{eq: iter basis} \eta_{n} = \begin{pmatrix} \rho_1 & & & & \\ \sigma_{2,1} & \rho_2 & & & \\ \sigma_{3,1} & \sigma_{3,2} & \rho_3& & \\ \vdots & & \ddots & \ddots & \\ \sigma_{n,1}& \dotsm & & \sigma_{n,n-1} & \rho_{n} \end{pmatrix}. \end{equation} We form a dg-$\mathbb{F}$-algebra out of the Hochschild complex valued in the $E$-bimodule \[ \bigoplus_{1 \leq i < j \leq n} \mathrm{Hom}_\mathbb{F}(\rho_i, \rho_j) \] with the natural compositions of homomorphisms making this $E$-bimodule a non-unital $\mathbb{F}$-algebra. Now we use Massey products in this dg-algebra. One may readily compute that $\eta_{n}$ is a homomorphism if and only if $\mathfrak{S} = \{\sigma(i,j)\} = \{\sigma_{j,i}\}$ is a defining system for the $n$th Massey product $\langle \sigma_{2,1}, \dotsc, \sigma_{n,n-1}\rangle$ and \[ d\sigma_{n,1} = c(\mathfrak{S}). \] In other words, we have an equivalence, as follows. \begin{prop} \label{prop: matric vanishing} There exists an $\eta_n$ realizing the $\sigma_{i+1,i}$ below the diagonal if and only if the Massey product $\langle \sigma_{2,1}, \dotsc, \sigma_{n,n-1}\rangle$ is defined and contains zero. \end{prop} This idea of such a connection between defining systems and extensions is due to May \cite{may1969}. This may be taken to be a condition on iterated extensions of representations: the condition that $\mathfrak{S}$ is a defining system is equivalent to the existence of ``overlapping'' homomorphisms \[ \eta_{1,n-1} = \begin{pmatrix} \rho_1 & & & \\ \sigma_{2,1} & \rho_2 & & \\ \vdots & \ddots & \ddots & \\ \sigma_{n-1,1}& \dotsm & \sigma_{n-1,n-2} & \rho_{n-1} \end{pmatrix}, \eta_{2,n} = \begin{pmatrix} \rho_2 & & & \\ \sigma_{3,2} & \rho_3 & & \\ \vdots & \ddots & \ddots & \\ \sigma_{n,1}& \dotsm & \sigma_{n,n-1} & \rho_{n} \end{pmatrix} \] Given this, the element $\langle \sigma_{2,1}, \dotsc, \sigma_{n-1,n}\rangle_\mathfrak{S} = c(\mathfrak{S})$ of $\langle \sigma_{2,1}, \dotsc, \sigma_{n-1,n}\rangle$ vanishes (in cohomology) if and only if there exists a common extension $\eta_n$ of $\eta_{1,n-1}$ and $\eta_{2,n}$ exists. For $\eta_n$ exists if and only if there exists $\sigma_{n,1} \in C^1(E, \mathrm{Hom}_\mathbb{F}(\rho_1, \rho_n)$ such that $d\sigma_{n,1} = c(\mathfrak{S})$. \subsection{Lifts of representations} \label{subsec: explicit lifts} We will start with a representation $\rho : E \rightarrow M_d(\mathbb{F})$ as in \S\S\ref{sec:assoc pt}-\ref{sec:assoc mult pts}. We will especially use the following coefficient algebras: for $n \geq 0$, write by $\mathbb{F}[\varepsilon_n]$ for $A[\varepsilon]/(\varepsilon^{n+1}) \in \cC_\mathbb{F}$. An \emph{$n$th-order lift of $\rho$} is a lift $\rho_n$ of $\rho$ (as in Definition \ref{defn: lift assoc}) to $\mathbb{F}[\varepsilon_n]$. We associate to $\rho_n$ an expression as a homomorphism to $M_{nd}(\mathbb{F})$, extending the standard basis $(a_i)_{i=1}^d$ of $\mathbb{F}^{\oplus d}$ to a $\mathbb{F}$-basis of $\mathbb{F}[\varepsilon_n]^{\oplus d}$ consisting of elements $(\varepsilon^j a_i : 1 \leq i \leq d, 0 \leq j \leq n\}$ with the ordering by $j$ and then by $i$. We arrive at the matrix realization \begin{equation} \label{eq: ep basis} \rho_n = \begin{pmatrix} \rho & \sigma_1 & \sigma_2 & \dotsm & \sigma_n \\ & \rho & \sigma_1 & & \vdots \\ & & \ddots & \ddots & \\ & & & \rho & \sigma_1 \\ & & & & \rho \end{pmatrix} : E \rightarrow M_{nd}(\mathbb{F}).. \end{equation} We will render this as \[ \rho_n = \rho + \sum_{i = 1}^n \sigma_i \varepsilon^i : E \rightarrow M_d(\mathbb{F}[\varepsilon_n]), \] where $\sigma_i$ is a function $\sigma_i : E \rightarrow M_d(\mathbb{F}) \cong \mathrm{End}_\mathbb{F}(\rho)$. Let $C = C^\bullet(E, \mathrm{End}_\mathbb{F}(\rho))$, so $\sigma_i \in C^1$. Because $\rho_n$ is a homomorphism, one readily observes that $\sigma_1$ lies in $Z^1$. More generally, the relations \begin{equation} \label{eq:ut_cond} d\sigma_i = \sum_{j=1}^{i-1} \sigma_j \smile \sigma_{i-j} \text{ for } 1 \leq i \leq n \end{equation} are satisfied if and only if the corresponding expression for $\rho_n$ a homomorphism. We may apply the connection between these conditions and Massey products from \S\ref{subsec: iter}. Next, we observe that these are Massey powers, using the symmetry visible by comparing \eqref{eq: ep basis} to \eqref{eq: iter basis}. Namely, the set $\mathfrak{T} = \{\sigma_1, \dotsc, \sigma_n\}$ satisfies \eqref{eq:ut_cond}, and therefore it constitutes a defining system for the $(n+1)$st Massey power $\dia{\sigma_1}_\mathfrak{T}^{n+1}$. We will simply denote this Massey power defining system $\mathfrak{T}$ by $\rho_n$ when it will not cause confusion. Thus the resulting Massey power is written $\langle \sigma_1 \rangle^{n+1}_{\rho_n}$, the cohomology class of the cocycle $c(\rho_n)$. If we increment $n$ to $n+1$, the new relation of \eqref{eq:ut_cond} is \[ d\sigma_{n+1} = \sum_{j=1}^n \sigma_j \smile \sigma_{n-j+1} =: c(\rho_n), \] so there exists some $\sigma_{n+1}$ satisfying \eqref{eq:ut_cond} if and only if $c(\rho_n)$ is a 2-coboundary. We summarize this discussion as follows. \begin{prop} \label{prop: massey lifting} For $n \geq 1$, let $\rho_n$ be an $n$th-order lift of $\rho$, defining cochains $\sigma_i \in C^1(E, \mathrm{End}_\mathbb{F}(\rho))$ for $1 \leq i \leq n$ as above. Then $\sigma_1 \in Z^1(E, \mathrm{End}_\mathbb{F}(\rho))$ and the Massey power $\langle \sigma_1 \rangle^{n+1}_{\rho_n}$ vanishes if and only if there exists an $(n+1)$st order deformation $\rho_{n+1} = \rho + \sum_{i=1}^{n+1} \sigma_i \varepsilon^i$ extending $\rho_n$. In this case, we have an equality of 2-coboundaries $d\sigma_{n+1} = c(\rho_n)$ and the set of possible $\sigma_{n+1}$ is a $Z^1(E, \mathrm{End}_\mathbb{F}(\rho))$-torsor. \end{prop} Varying over possible lifts extending a first order lift $\rho_1$ of $\rho$, we have the following immediate consequence, analogous to Proposition \ref{prop: matric vanishing}. \begin{cor} Let $\rho_1 = \rho + \sigma_1\varepsilon$ be a first order lift of $\rho$. Then $\rho_1$ extends to an $n$th-order lift if and only if the Massey power $\langle \sigma_1 \rangle^n \subset H^2(E, \mathrm{End}_A(\rho))$ is defined and contains zero. \end{cor} \subsection{Expression of moduli spaces using Massey products} \label{subsec: massey express} In light of Proposition \ref{prop: massey lifting}, it is clear, in principle, that Massey products control ${\mathrm{Def}}^\mathrm{nc}_\rho$. In this section, we explain how to compute \emph{universal} lifts in terms of elements of Massey powers. We will only do this up to the point of illustrating the technique --- and illustrating its limitations compared to the $A_\infty$-based expression --- as we have already explained that $A_\infty$-structures give rise to Massey products and $A_\infty$-structures control ${\mathrm{Def}}^\mathrm{nc}_\rho$ (Remark \ref{rem: A-inf vs Massey}). We observe that there is a universal first order lift $\rho_1^u$ of $\rho$. It is induced by the ``universal 1-cocycle'' defined by \begin{equation} \label{eq:universal 1-cocycle} \sigma_1^u : E \longrightarrow M_d(Z^1(E,\mathrm{End}_\mathbb{F}(\rho))^*), \quad \gamma \mapsto (\sigma_1 \mapsto \sigma_1(\gamma)_{i,j})_{i,j} \end{equation} for $\gamma \in E$, $\sigma_1 \in Z^1(E, \mathrm{End}_\mathbb{F}(\rho)$, and $(i,j)$ denoting the matrix coordinate. There is a left and a right $E$-action on $M_d(Z^1(E,\mathrm{End}_\mathbb{F}(\rho))^*) \cong M_d(\mathbb{F}) \otimes Z^1(E,\mathrm{End}_\mathbb{F}(\rho))^*$ given by the usual left and right actions of $E$ on $M_d(\mathbb{F})$ via $\rho$ and the multiplication map of $M_d(\mathbb{F})$. Letting $\mathbb{F}[M]$ denote the square-zero $\mathbb{F}$-algebra extension of $\mathbb{F}$ by an $\mathbb{F}$-vector space $M$, we have \begin{equation} \label{eq:u 1st order lift} \rho_1^u = \rho + \varepsilon \sigma_1^u : E \longrightarrow M_d(\mathbb{F}[Z^1(E,\mathrm{End}_\mathbb{F}(\rho))^*]). \end{equation} This lift is universal in the sense that for any other first-order lift $\rho_A$, there is a unique $\mathbb{F}$-linear map $E^1(E,\mathrm{End}_\mathbb{F}(\rho))^* \rightarrow \mathfrak{m}_A$ such that $\rho^1_u \otimes_{\mathbb{F}[Z^1(E, \mathrm{End}_\mathbb{F}(\rho))^*]} A = \rho_A$. This follows from the fact that $\rho_A - \rho \otimes_\mathbb{F} A : E \rightarrow M_d(\mathfrak{m}_A)$ is a cocycle, i.e.\ valued in $Z^1(E, \mathrm{End}_\mathbb{F}(\rho) \otimes_\mathbb{F} \mathfrak{m}_A)$. For the remainder of this section we produce a universal lift of any order, applying the computations of the previous section. We establish notation for the sake of concision: write $T$ for $Z^1(E, \mathrm{End}_\mathbb{F}(\rho))^*$ and let $\mathbb{F}[T_n] := \bigoplus_{i=0}^n T^{\otimes i}$ be the free associative $\mathbb{F}$-algebra on $T$ truncated at degree $n$, in analogy with $\mathbb{F}[\varepsilon_n]$. Inductively, we construct Massey powers of $\sigma_1^u \in Z^1(E,\mathrm{End}_\mathbb{F}(\rho)) \otimes T$. The base case is the cup product, which is the cohomology class of the unambiguously defined 2-cocycle \[ \sigma_1^u \cup \sigma_1^u \in H^2(E,\mathrm{End}_\mathbb{F}(\rho)) \otimes T^{\otimes 2}. \] Let $I_2 \subset T^{\otimes 2}$ be the minimal subspace such that $\sigma_1^u \cup \sigma_1^u$ vanishes modulo $I_2$. Then we can solve \eqref{eq:ut_cond} modulo $I_2$, i.e.\ there exists $\sigma_2^u \in C^1(E, M_d(T^{\otimes 2}/I_2))$ such that \[ d\sigma_2^u \equiv \sigma_1^u \cup \sigma_1^u \pmod{I_2} \] and the set of possible choices for $\sigma_2^u$ is a torsor under $Z^1(E, \mathrm{End}_\mathbb{F}(\rho)) \otimes_\mathbb{F} T^{\otimes 2}/I_2$. We get a second order lift \[ \rho_2^u = \rho + \sigma_1^u + \sigma_2^u : E \longrightarrow M_d(\mathbb{F}[T_2]/I_2). \] In fact, untangling dualities, we see that $I_2$ is generated by the image in $T^{\otimes 2}$ of the image of the map \[ \cup^* : H^2(E, \mathrm{End}_\mathbb{F}(\rho))^* \longrightarrow Z^1(E, \mathrm{End}_\mathbb{F}(\rho))^* \otimes Z^1(E, \mathrm{End}_\mathbb{F}(\rho))^* \cong T^{\otimes 2} \] that is dual to the cup product map. The inductive step from order $n$ to order $n+1$ is to start with an $n$th-order lift $\rho_n^u = \rho + \sum_{i=1}^n \sigma_n^u$ of $\rho$ with coefficients in \[ \frac{\mathbb{F}[T_n]}{(I_2, I_3, \dotsc, I_n)} \] and calculate Massey power $\langle \sigma_1^u \rangle^{n+1}_{\rho_n^u}$ valued in $H^2(E, \mathrm{End}_\mathbb{F}(\rho)) \otimes_\mathbb{F} T^{\oplus n+1}/I'_{n+1}$, where $I'_{n+1}$ is the ideal of $\mathbb{F}[T^{n+1}]$ generated by the image of $(I_2, I_3, \dotsc, I_n) \subset \mathbb{F}[T_n]$ under the (non-multiplicative) natural map $\mathbb{F}[T_n] \hookrightarrow \mathbb{F}[T_{n+1}]$. Then define $I_{n+1}$ to be the minimal submodule of $T^{\otimes n+1}$ containing the degree $n+1$ projection of $I'_{n+1}$ and such that $\langle \sigma_1 \rangle^{n+1}_{\rho_n}$ vanishes modulo $I_{n+1}$. As in the case $n=1$, we now have $\sigma_{n+1}^u$ valued in $T^{\otimes n}/I_{n+1}$ and $\rho_{n+1}^u$ valued in $\mathbb{F}[T_{n+1}]/(I_2, \dotsc, I_{n+1})$. Because $I_n$ is concentrated in degree $n$, we have limits \[ R^\square_\rho := \varprojlim_n \frac{\mathbb{F}[T_n]}{(I_2, I_3, \dotsc, I_n)}, \quad \rho^u := \varprojlim_n \rho_n^u : E \rightarrow M_d(R^\square_\rho) \] where $(R^\square_\rho, \mathfrak{m}^\square_\rho)$ is a complete local $\mathbb{F}$-algebra quotient of $\hat T_\mathbb{F} T$. We summarize the construction above and state its universal property. \begin{thm} \label{thm:lifting ring} For any local Artinian $\mathbb{F}$-algebra with residue field $\mathbb{F}$ and lift $\rho_A$ of $\rho$ valued in $A$, there exists a unique local $\mathbb{F}$-algebra homomorphism $R^\square_\rho \rightarrow A$ such that $\rho_A = \rho^u \otimes_{R^\square_\rho} A$. That is, ${\mathrm{Def}}^\square_\rho = \Spf R^\square_\rho$ and $R^\square_\rho$ pro-represents ${\mathrm{Def}}^\square_\rho$. Moreover, we have a canonical isomorphism $(\mathfrak{m}^\square_\rho/(\mathfrak{m}^\square_\rho)^2)^* \cong Z^1(\mathrm{End}_\mathbb{F}(\rho))$. \end{thm} We omit the proof, since it amounts to the same argument as the proof of Corollary \ref{cor: irred case A-inf}, but is more complicated due to the inductive construction of $R^\square_\rho$. \begin{rem} \label{rem: Massey over S} One advantage of Massey powers over $A_\infty$-products is that the base coefficient ring $\mathbb{F}$ may be replaced with a general commutative ring $S$. Indeed, all of the calculations of ideals $I_n$ make sense in this case, and the ideals $I_n$ may have non-trivial $S$-part. For example, when $S = \mathbb{Z}/p^2$ for some prime $p$, it is possible for \begin{itemize} \item a non-trivial first-order lift to exist modulo $p^2$, i.e., over $S[\epsilon_1]$, \item an extension to a second-order lift to exist only over $\mathbb{Z}/p$, i.e., over $R := S[\epsilon]/(p\epsilon^3, \epsilon^4)$, \item and no extension to a third-order lift whatsoever. \end{itemize} If this particular first-order lift is unique modulo $p$, then the universal deformation ring is $R$. For an example of an application of Massey products over $S = \mathbb{Z}/p^n\mathbb{Z}$, see \cite{WWE3}. \end{rem} \begin{rem} One obstruction to applying the technology of $A_\infty$-algebras, as in \S\ref{sec:A-inf}, over a general commutative base ring $S$ in place of $\mathbb{F}$ is that homotopy retracts as in Example \ref{eg:H-sections} may not exist. A replacement for Kadeishvili's theorem (Corollary \ref{cor:A-inf on H}) is needed: see \cite{sagave2010}. \end{rem} \part{Moduli of Galois representations and pseudorepresentations} In this part, we apply the results of Part 2 to cases of interest in number theory. First we set up the theory for representations of a profinite group. Continuity of representations is taken to be implicit, and we now discuss only commutative coefficients. Then we adapt this theory for cases of interest in number theory: representations of a profinite group satisfying some additional condition. \section{Moduli of representations of a profinite group} \label{sec: profinite moduli} We recall moduli spaces of representations of a profinite group. These moduli spaces were set up in the author's previous work \cite{WE2018}, which we now recall. In contrast with \emph{loc.\ cit.}, we work in constant positive characteristic. Thus the initial coefficient ring is a finite field. We write $p$ for the characteristic. \subsection{Connected components biject with residual semi-simplification} \label{subsec: residual SS} The main result of \cite[\S3]{WE2018} is that the moduli of (integral) $p$-adic representations of a profinite group is the disjoint union of connected components parameterized precisely by the residual semi-simplification. We set up a precise meaning of the term ``residual semi-simplification.'' \begin{defn} \label{defn: resid SS} Let $G$ be a profinite group and let $p$ be a prime. A \emph{residual semi-simplification} is an equivalence class of semi-simple representations of $G$ valued in a finite field $\mathbb{F}$ of characteristic $p$ such that each simple summand is absolutely irreducible. The equivalence relation is isomorphism of representations, or, equivalently, isomorphism after change of coefficients via $\otimes_\mathbb{F} \overline{\mathbb{F}}$. A residual semi-simplification is called \emph{multiplicity-free} if there are no non-trivial isomorphisms among the simple summands. \end{defn} \begin{rem} Residual semi-simplifications are in bijection with \emph{residual pseudorepresentations} in \cite[Def.\ 3.4]{WE2018}. \end{rem} Fix a representative $\rho : G \rightarrow \GL_d(\mathbb{F})$ of a residual semi-simplification; we take $\mathbb{F}_\rho$ to be the smallest possible base field such that a residual semi-simplification is defined over $\mathbb{F}_\rho$ and each irreducible factor is absolutely irreducible. We set up the equal-characteristic moduli of deformations of all residual representations with residual semi-simplification $\rho$. Without loss of generality, we take $\rho$ to be in block diagonal form in $\GL_d$, with diagonal summands \[ \rho \cong \bigoplus_{i=1}^r \rho_i. \] Here $\rho_i : G \rightarrow \GL_{d_i}(\mathbb{F})$ are the absolutely irreducible factors of $\rho$. Given this data, we write $\GL(\rho)$ for the corresponding Levi sub-$\mathbb{F}$-algebraic group \[ \GL(\rho) := \prod_{i=1}^r \GL_{d_i} \hookrightarrow \GL_d. \] Write $\PGL(\rho)$ for the quotient group of $\GL(\rho)$ by the center of $\GL_d$. The natural equal-characteristic category of coefficient rings are \emph{topologically finitely generated $\mathbb{F}_p$-algebras}, i.e.\ topological quotients of algebras of the form $\mathbb{F}_p{[\![} x_1, \dotsc, x_n{]\!]} \langle y_1, \dotsc, y_m\rangle$. These algebras have the $(x_1, \dotsc, x_n)$-adic topology, and the angle brackets refer to restricted power series in this topology. We denote this category by ${\mathcal{A}f\!f}_{\mathbb{F}_p}$. Equivalently, ${\mathcal{A}f\!f}_{\mathbb{F}_p}^\mathrm{op}$ is the category of finite type affine $\Spf \mathbb{F}_p$-formal schemes. It is natural to replace $\mathbb{F}_p$ by $\mathbb{F}$ when we impose the condition that a representation has residual semi-simplification $\rho$, as follows. \begin{defn} \label{defn: Rep spaces} Let ${\mathrm{Rep}}^\square_d$ denote the functor on $A \in {\mathcal{A}f\!f}_{\mathbb{F}_p}$ valued in sets, given by \[ {\mathrm{Rep}}^\square_d(A) = \{ \rho_A: G \rightarrow \GL_d(A)\}. \] Likewise, we have the quotient groupoid by the adjoint action of $\PGL_d$, \[ {\mathrm{Rep}}_d := [{\mathrm{Rep}}^\square_d / \PGL_d], \] which represents the deformation groupoid defined just as in the definition of ${\mathrm{Rep}}^\square_\rho$, but with $\GL_d(A)$ replaced by the units of a $d$-dimensional Azumaya algebra over $A$ (see \cite[\S2.1]{WE2018}). Let ${\mathrm{Rep}}^\square_{\rho}$ denote the subfunctor of ${\mathrm{Rep}}^\square_d \times_{\mathbb{F}_p} \mathbb{F}_\rho$ given by \[ {\mathrm{Rep}}^\square_{\rho}(A) = \{ \rho_A: G \rightarrow \GL_d(A) \mid \text{for all } f: A \rightarrow \overline{\mathbb{F}}_\rho, (\rho_A \otimes_{A,f} \overline{\mathbb{F}}_\rho)^{ss} \simeq \rho \otimes_{\mathbb{F}_\rho} \overline{\mathbb{F}}_\rho\} \] for $A \in {\mathcal{A}f\!f}_{\mathbb{F}_\rho}$. Here $\simeq$ indicates being in the same orbit under the adjoint action of $\PGL_d(\overline{\mathbb{F}}_\rho)$. \end{defn} A main result of \cite{WE2018} is that the mixed-characteristic versions of such spaces are representable in the category of $\Spf \mathbb{Z}_p$-formal schemes. We state this result specialized to equal-characteristic. \begin{thm}[{\cite[\S3.1]{WE2018}}] \label{thm: algfamS3 main} ${\mathrm{Rep}}^\square_{\rho}$ is representable by a topologically finite type affine $\Spf \mathbb{F}$-formal scheme $\Spf S_{\rho}$. We have \[ {\mathrm{Rep}}^\square_d \times_{\mathbb{F}_p} \overline{\mathbb{F}}_p = \coprod_{\rho} {\mathrm{Rep}}^\square_{\rho} \times_{\mathbb{F}_\rho} \overline{\mathbb{F}}_\rho, \] where $\rho$ varies over all residual semi-simplifications of dimension $d$ and $\mathbb{F}_\rho$ denotes the coefficient field of $\rho$. \end{thm} In additional to representability, the main upshot is that in order to understand the entire moduli space ${\mathrm{Rep}}^\square_d$, we may study it one residual semi-simplification $\rho$ at a time. Similarly, this theorem implies algebraicity and decomposition into connected components parameterized by $\rho$ for the stack quotients ${\mathrm{Rep}}_d$. In the case that the residual semi-simplification $\rho$ is irreducible, both ${\mathrm{Rep}}_\rho^\square$ and ${\mathrm{Rep}}_\rho$ are represented by the formal spectra of complete Noetherian local rings with residue field $\mathbb{F}_\rho$. These are the usual deformation rings for representations of profinite groups whose study was initiated by Mazur \cite{mazur1989}. We proceed with results that will be useful in order to study the case where $\rho$ is not irreducible. \subsection{Theory of pseudorepresentations, Cayley--Hamilton algebras, and generalized matrix algebras} \label{subsec: review RR case} We review the theory of pseudorepresentations due to Chenevier \cite{chen2014}. Because the review of \cite[\S2]{WE2018} is precisely what we need, we refer the reader there. Here, we recall only notation and selected parts of definitions. \begin{itemize}[leftmargin=2em] \item $D : E \rightarrow A$ denotes a \emph{pseudorepresentation}. Using this notation implies that $E$ is an associative unital $A$-algebra, where $A$ is a commutative ring. This $D$ has a dimension $d \in \mathbb{Z}_{\geq 1}$, and is a functor from commutative $A$-algebras $B$ to functions $D_B : E \otimes_A B \rightarrow B$ that are homogeneous of degree $d$ in $B$. \item When $G$ is a group, $D : G \rightarrow A$ is notation for a pseudorepresentation $D : A[G] \rightarrow A$. \item For any commutative $A$-algebra $B$, and $x \in E \otimes_A B$, there is a characteristic polynomial $\chi_D(x,t) \in B[t]$. \item Given $D : E \rightarrow A$, there is a notion of a kernel two-sided ideal and Cayley--Hamilton two-sided ideal of $E$, \[ E \supset \ker(D) \supset \mathrm{CH}(D). \] There exist canonical factorizations of $D$ through $E/\ker(D)$ and $E/\mathrm{CH}(D)$. \item A pseudorepresentation is called \emph{Cayley--Hamilton} when $\mathrm{CH}(D) = 0$. Equivalently, for all commutative $A$-algebras $B$ and all $x \in E \otimes_A B$, $\chi_D(x,x) = 0$. That is, $x$ satisfies its own characteristic polynomial $\chi_D(x,t) \in B[t]$. Collectively, such data $(E, A, D : E \rightarrow A)$ is called a \emph{Cayley--Hamilton $A$-algebra}. \item Given a representation $\eta : G \rightarrow M_d(A)$, there is an induced $d$-dimensional pseudorepresentation, denoted $\psi(\eta) : G \rightarrow A$, given by composing $\eta$ with the determinant pseudorepresentation $\det : M_d(A) \rightarrow A$. \item Similarly, if $E$ is equipped with a pseudorepresentation $D : E \rightarrow A$ and $\eta$ is a homomorphism $G \rightarrow E^\times$, then there is an induced pseudorepresentation $D \circ \eta : G \rightarrow A$. We especially study the case where $(E, A, D)$ is a Cayley--Hamilton representation. Then we call the data \[ (\eta : G \rightarrow E^\times, E, A, D : E \rightarrow A) \] a \emph{Cayley--Hamilton representation of $G$}, and we call $\psi(\eta) := D \circ \eta$ its \emph{induced pseudorepresentation} . \item A \emph{generalized matrix algebra over $A$} or \emph{$A$-GMA} is an associative $A$-algebra $E$ equipped with \begin{itemize}[leftmargin=2em] \item a complete orthogonal set of idempotents $(e_i)_{i=1}^r \subset E$, \item $A$-algebra isomorphisms $e_i E e_i \buildrel\sim\over\ra M_{d_i}(A)$ \end{itemize} that satisfy an extra condition. This notion, due in this form to Bella\"iche--Chenevier \cite[\S1]{BC2009}, was shown to admit a natural Cayley--Hamilton pseudorepresentation $D_\mathrm{GMA} : E \rightarrow A$ in \cite[Prop. 2.23]{WE2018}. \item We call a Cayley--Hamilton representation $(\rho, A, E, D)$ a \emph{GMA representation} over $A$ when $E$ is also equipped with GMA data such that $D = D_\mathrm{GMA}$. \end{itemize} The application of the tools above to the moduli of profinite groups is the main content of \cite[\S3]{WE2018}. We require a few more definitions and results about this situation, which we recall directly from \cite[\S3]{WE2018}. The results about deformation theory of pseudorepresentations are due to Chenevier \cite{chen2014}. However, here we work in constant characteristic $p$. We start with a fixed residual semi-simplification $\rho : G \rightarrow \GL_d(\mathbb{F}_\rho)$, and write $\mathbb{F} = \mathbb{F}_\rho$. It induces a $d$-dimensional pseudorepresentation \[ D = \psi(\rho) : G \rightarrow \mathbb{F}. \] \begin{itemize}[leftmargin=2em] \item There is a deformation functor sending $A \in \cC_\mathbb{F}$ to pseudorepresentations $D_A : G \rightarrow A$ such that the reduction modulo $\mathfrak{m}_A$ \[ G \buildrel{D_A}\over\longrightarrow A \longrightarrow \mathbb{F} \] is equal to $D$. Such a $D_A$ is called a \emph{pseudodeformation of $D$ to $A$}. This gives rise to a deformation functor on $\cC_A$ for $D$. \item There is a universal pseudodeformation ring $R_D$ representing the deformation functor for $D$ on the category $\cC_\mathbb{F}$. It also represents the extension of this deformation problem from $\cC_\mathbb{F}$ to ${\mathcal{A}f\!f}_\mathbb{F}$. Thus $R_D$ supports the universal deformation \[ D^u : G \rightarrow R_D \] of $D$. It is a complete local $\mathbb{F}$-algebra with residue field $\mathbb{F}$. When $G$ satisfies the $\Phi_p$ finiteness condition of \cite[\S1]{mazur1989}, $R_D$ is Noetherian. \item We say that a Cayley--Hamilton representation \[ (\eta : G \rightarrow E^\times, A, E, D_E : E \rightarrow A) \] where $A \in \cC_\mathbb{F}$ \textit{has residual pseudorepresentation $D$} when its induced pseudorepresentation of $G$ \[ \psi(\eta) = D_E \circ \eta : G \longrightarrow A \] is a deformation of $D$. For short, we say that $\rho$ is \emph{over $D$}. These notions have a sensible extension of coefficient algebras from $A \in \cC_\mathbb{F}$ to $A \in {\mathcal{A}f\!f}_\mathbb{F}$. \item There is a universal Cayley--Hamilton representation of $G$ over $D$, produced as follows. We let $E_D$ be the Cayley--Hamilton quotient of the universal pseudodeformation of $D$, that is, \[ E_D := \frac{R_D[G]}{\mathrm{CH}(D^u)}. \] The theory of Cayley--Hamilton algebras recalled above implies that $D^u$ factors through $E_D$ as a Cayley--Hamilton pseudorepresentation; we denote the factorization by $D_{E_D}$. Thus we have a Cayley--Hamilton pseudorepresentation \[ (\rho^u : G \rightarrow E_D^\times, R_D, E_D, D_{E_D} : E \rightarrow R_D) \] over $D$. Its induced pseudorepresentation $\psi(\rho^u) : G \rightarrow R_D$ is equal to the universal pseudodeformation $D^u : G \rightarrow R_D$ of $D$. \item When $G$ satisfies the $\Phi_p$ finiteness property, $E_D$ is finitely generated as a $R_D$-module. Therefore it is a Noetherian ring. \end{itemize} \subsection{Application to moduli spaces of representations} \label{subsec: apply to moduli} In order to apply the theory of Cayley--Hamilton representations to the moduli spaces of representations ${\mathrm{Rep}}^\square_\rho$ and ${\mathrm{Rep}}_\rho$, we make the following observations and additional definitions, which come from \cite[\S3]{WE2018}. \begin{itemize}[leftmargin=2em] \item When $D = \psi(\rho) : G \rightarrow \mathbb{F}$, there is a natural functor $\psi^\square : {\mathrm{Rep}}^\square_\rho \rightarrow \Spf R_D$ sending a representation $\eta : G \rightarrow M_d(A)$ with residual semi-simplification $\rho$ to its induced pseudorepresentation $\psi(\eta) = \det \circ \, \eta : G \rightarrow A$, which is a pseudodeformation of $D$. And $\psi^\square$ factors through \[ \psi : {\mathrm{Rep}}_\rho \rightarrow \Spf R_D, \] which sends an Azumaya algebra valued representation of $G$ to the pseudorepresentation of $G$ that arises from composition with the Azumaya algebra's reduced norm. \item Given a $d$-dimensional Cayley--Hamilton algebra $(E, A, D: E \rightarrow A)$, there exists an affine $A$-scheme ${\mathrm{Rep}}^\square_{E,D}$ of representations of $E$ that are \emph{compatible with $D$}. It is a functor on commutative $A$-algebras $B$ sending \[ B \mapsto \{\eta : E \rightarrow M_d(B) \mid \psi(\eta) := \det \circ \, \eta : E \rightarrow B \text{ equals } D \otimes_A B : E \rightarrow B\}. \] Likewise, there is an moduli groupoid of Azumaya algebra-valued representations of $E$, which is represented by the stack quotient $[{\mathrm{Rep}}^\square_{E,D}/\PGL_d]$. \item Given a $d$-dimensional GMA $(E, A, D_\mathrm{GMA} : E \rightarrow A)$ with idempotents $e_i$ each of dimension $d_i$, there exists a closed sub-$A$-scheme \[ {\mathrm{Rep}}^\mathrm{GMA}_{E,D} \subset {\mathrm{Rep}}^\square_{E,D} \] of \emph{adapted representations}. The notion and moduli of adapted representations were first studied in \cite[\S1.3]{BC2009}. These are matrix algebra-valued representations that fix the data of idempotents, where we choose a diagonal data of idempotents in the matrix algebra. \item Let $Z(e_i)$ be the split torus in $\GL_d$ which centralizes the block diagonal subalgebra \[ \bigoplus_{i=1}^r e_i E e_i \cong \bigoplus_{i=1}^r M_{d_i}(A) \hookrightarrow M_d(A). \] This torus has a natural adjoint action on ${\mathrm{Rep}}^\mathrm{GMA}_{E,D}$, and its stack quotient admits an isomorphism \[ [{\mathrm{Rep}}^\mathrm{GMA}_{E,D}/Z(e_i)] \cong {\mathrm{Rep}}_{E,D}. \] \item Let $e^{11}_i$ denote the idempotent of $e_i E e_i \cong M_{d_i}(A)$ cutting out the $(1,1)$-coordinate of $M_{d_i}(A)$. Let $e^{11} = \sum_{i=1}^r e_i^{11}$. We then get a Morita-equivalent algebra \begin{equation} \label{eq: Morita GMA} e^{11} E e^{11} \end{equation} that naturally admits the structure of a GMA: the idempotents are $e^{11}_i$ and $d_i = 1$ for all $1 \leq i \leq r$. We write $\cA_{i,j}$ for \[ \cA_{i,j} := e^{11}_j E e^{11}_i = e^{11}_j (e^{11} E e^{11}) e^{11}_i. \] \item According to \cite[Prop.\ 1.3.9]{BC2009}, there is an expression for the $A$-algebra $S^\mathrm{GMA}_{E,D}$ representing the affine $A$-scheme ${\mathrm{Rep}}^\mathrm{GMA}_{E,D}$ in terms of the multiplication map on $e^{11}E e^{11}$ decomposed into its idempotent-based coordinates as \[ \varphi_{i,j,k} : \cA_{i,j} \otimes_A \cA_{j,k} \rightarrow \cA_{i,k}. \] The expression for $S^\mathrm{GMA}_{E,D}$ is \begin{equation} \label{eq: adapt ring} S^\mathrm{GMA}_{E,D} \buildrel\sim\over\lra \frac{ \Sym^*_A \left(\bigoplus_{1 \leq i \neq j \leq r} \cA_{i,j}\right) }{ \left(x \otimes y - \varphi(x \otimes y)\right) }, \end{equation} where the denominator stands for the ideal generated over varying $x \in \cA_{i,j}$, $y \in \cA_{j,k}$, and $\varphi = \varphi_{i,j,k}$, over varying $(i,j,k)$. \end{itemize} We summarize the results about these objects given in \cite{WE2018}. \begin{thm}[{\cite[\S3]{WE2018}}] \label{thm: algfam main} Let $\rho$ be a residual semi-simplification with $D = \psi(\rho)$, and assume that $G$ satisfies the $\Phi_p$ finiteness condition. There are natural isomorphisms of $\Spf \mathbb{F}$-spaces \[ {\mathrm{Rep}}^\square_\rho \cong {\mathrm{Rep}}_{E_D, D_{E_D}}^\square, \qquad {\mathrm{Rep}}_\rho \cong {\mathrm{Rep}}_{E_D, D_{E_D}}, \] each of which admits a finite type module over $\Spec R_D$. That is, there is an isomorphism of the moduli of \begin{itemize}[leftmargin=2em] \item representations of $G$ with residual semi-simplification $\rho$ and \item representations of $E_D$ that are compatible with the pseudorepresentation $D^u : E_D \rightarrow R_D$, \end{itemize} which is an isomorphism of $\Spf R_D$-formal spaces. Furthermore, assuming that $\rho$ is multiplicity-free, \begin{enumerate}[leftmargin=2em] \item $E_D$ admits the structure $(e_i)_{i=1}^r$ of an $R_D$-GMA such that $D^u = D_\mathrm{GMA}$ and \[ {\mathrm{Rep}}_\rho \cong [{\mathrm{Rep}}^\mathrm{GMA}_{E,D_{E_D}} / Z(e_i)] \] \item If we let $S^\mathrm{GMA}_\rho$ be the commutative $A$-algebra representing the affine scheme ${\mathrm{Rep}}^\mathrm{GMA}_{E,D_{E_D}}$, then the structure morphism $\psi_\mathrm{GMA} : {\mathrm{Rep}}^\mathrm{GMA}_{E,D_{E_D}} \rightarrow \Spf R_D$ induces an isomorphism from $R_D$ to the invariant subring \[ R_D \buildrel\sim\over\lra (S^\mathrm{GMA}_{E,D})^{Z(e_i)}. \] \end{enumerate} \end{thm} \begin{proof} The initial statements come from \cite[Thm.\ 3.7]{WE2018}. The GMA structure claimed in (1) comes from \cite[Thm.\ 2.22(ii)]{chen2014}. The rest of (1) is proved in \cite[Thm.\ 2.27]{WE2018}, and (2) is \cite[Thm.\ 3.8(4)]{WE2018}. \end{proof} \begin{rem} As discussed in \cite[\S3]{WE2018}, the result (2) means that $\Spec R_D$ is canonically isomorphic to the GIT quotient for the stack ${\mathrm{Rep}}_\rho$. Even when $\rho$ is not multiplicity-free, so that (2) is not known to hold, nonetheless it is proved in \textit{loc.\ cit.}\ that there is a map $R_D \rightarrow (S^\square_\rho)^{\PGL_d}$ (where ${\mathrm{Rep}}^\square_\rho \cong \Spf S^\square_\rho$) that is very close to being an isomorphism. \end{rem} \section{Presentations in terms of $A_\infty$-structure on group cohomology} \label{sec: profinite presentations} We express ${\mathrm{Rep}}^\square_\rho$ as a moduli space of representations of an algebra, so that we may apply the results of Part 2. \subsection{From Hochschild cohomology to group cohomology} \label{subsec: hochs-group} Given a left $G$-module $V$, we write $C^\bullet(G,V)$ for the cochain complex of inhomogeneous group cochains; see \S\ref{subsec: kad intro}, or, for a full introduction, see e.g.\ \cite{brown1982}. Because we are working over a field, group cohomology realizes the $\mathrm{Ext}$-functors in the category of $\mathbb{F}[G]$-modules. There is also a direct compatibility between group cohomology and Hochschild cohomology of left modules for the group algebra. \begin{prop} Let $V, W$ be left $\mathbb{F}[G]$-modules. Give $\mathrm{Hom}_\mathbb{F}(W,V)$ the natural induced $\mathbb{F}[G]$-bimodule structure (left via $V$, right via $W$). \begin{enumerate}[leftmargin=2em] \item There is an isomorphism from the Hochschild complex to the group cochain complex \begin{align*} \theta^n : C^n(\mathbb{F}[G], \mathrm{Hom}_\mathbb{F}(W,V)) & \buildrel\sim\over\lra C^n(G, \mathrm{Hom}_\mathbb{F}(W,V)) \\ (f : \mathbb{F}[G]^{\otimes n} \rightarrow \mathrm{Hom}_\mathbb{F}(W,V)) & \mapsto f\vert_{G^{\times n}}, \end{align*} using the natural embedding $G^{\times n} \hookrightarrow \mathbb{F}[G]^{\otimes n}$. \item There are canonical isomorphisms of graded $\mathbb{F}$-vector spaces \[ H^\bullet(C^\bullet(\mathbb{F}[G], \mathrm{Hom}_\mathbb{F}(W,V))) \buildrel\sim\over\lra H^\bullet(G, \mathrm{Hom}_\mathbb{F}(W, V)) \buildrel\sim\over\lra \mathrm{Ext}^\bullet_{\mathbb{F}[G]}(V,W). \] \end{enumerate} \end{prop} \begin{proof} Because $\mathbb{F}[G]^{\otimes n} \cong \mathbb{F}[G^{\times n}]$, $\theta$ is an isomorphism of graded vector spaces. The differentials are compatible under $\theta$ because the map from $\mathbb{F}[G]$-bimodule actions $\star = (\star_\mathrm{left}, \star_\mathrm{right})$ to left $G$-module actions \[ g \cdot m := g \star_\mathrm{left} m \star_\mathrm{right} g^{-1} \] sends the given $\mathbb{F}[G]$-bimodule action on $\mathrm{Hom}_\mathbb{F}(W,V)$ to its standard left $G$-module action. From there, one observes that the formula for the Hochschild differential (Definition \ref{defn: Hochs}) is the same as the formula for the differential on inhomogeneous group cochains. The leftmost isomorphism of (2) is then clear. The right isomorphism relies on $C^n(G,\mathbb{F})$ being a projective resolution for the $G$-module $\mathbb{F}$, and $\mathbb{F}$ having trivial homological dimension. \end{proof} \begin{rem} The arguments are valid for discrete modules with a continuous action of $G$, when $G$ is a profinite group, using the fact that $\mathbb{F}$ is finite. The key fact is that $C^n(G,\mathbb{F})$ remains a resolution of $\mathbb{F}$; see e.g.\ \cite[Prop.\ 6.2.2]{RZ2010}. Correspondingly, the ambient categories are continuous finite discrete $G$-modules, resp.\ continuous finite discrete $\mathbb{F}[G]$-modules. \end{rem} \subsection{Presentation of the completed group algebra} Note that each absolutely irreducible factor $\rho_i$ of $\rho$ cuts out a maximal ideal of the completed group algebra $\mathbb{F}{[\![} G{]\!]}$ (see e.g.\ \cite[\S5.3]{RZ2010} for the definition of $\mathbb{F}{[\![} G{]\!]}$). We now let \[ E := \mathbb{F}{[\![} G{]\!]} \] and apply the theory of Part 2 as follows. \begin{itemize}[leftmargin=2em] \item We consider only \emph{continuous} representations of $E$, but we leave this implicit without stating it explicitly in the sequel. \item Likewise, we let $C = C^\bullet(E, \mathrm{End}_\mathbb{F}(\rho))$ denote the \emph{continuous} Hochschild cochain complex \[ C^i(E, \mathrm{End}_\mathbb{F}(\rho)) := \mathrm{Hom}_\mathrm{cts}(E^{\otimes i}, \mathrm{End}_\mathbb{F}(\rho)). \] It is a straightforward exercise to check that the differential and multiplication in $C$ preserves continuity. \end{itemize} After setting up these two instances of continuity, there are no additional instances where we must impose it. For consider that the lift $\rho \oplus \xi$ of $\rho$ to $A \in \cA_\mathbb{F}$ associated to a Maurer--Cartan element $\xi \in \mathrm{MC}(C, A) \subset C^1 \otimes \mathfrak{m}_A$ is obviously continuous. As all other representations are ultimately produced out of elements of $C^1$ along with formulas within $C$ that preserve continuity, namely, those of Example \ref{eg:KFT2}. We will implicitly always work in the continuous case in the sequel. \begin{thm} \label{thm: profinite r-pointed dual} Assume that $G$ satisfies the $\Phi_p$ finiteness condition. Choose an $r$-pointed homotopy retract between $C = C^\bullet(E, \mathrm{End}_\mathbb{F}(\rho))$ and its cohomology $H = H^\bullet(C)$, as in Example \ref{eg: r-pointed retract}. Choose also idempotents as in \eqref{eq: r isom lift}, including $e$. These choices induce an $A_\infty$-algebra structure $m$ on $H$ and complete $\mathbb{F}^r$-algebra isomorphisms \begin{align*} \rho^u : \mathbb{F}[G]^\wedge_{\rho} \buildrel\sim\over\lra \mathrm{End}_\mathbb{F}(V)\, \underline{\otimes}\, \frac{\hat T_{\mathbb{F}^r} (\Sigma H^1)^*}{(m^*((\Sigma H^2)^*))},\\ e \rho^u e : R^\mathrm{nc}_\rho := e \mathbb{F}[G]^\wedge_{\rho} e \buildrel\sim\over\lra \frac{\hat T_{\mathbb{F}^r} (\Sigma H^1)^*}{(m^*((\Sigma H^2)^*))} \end{align*} \end{thm} \begin{proof} This is an application of Corollary \ref{cor: r-pointed dual}. \end{proof} In order to concisely list the choices made in main theorems, we set up this \begin{defn} \label{defn: pres datum} Given a profinite group $G$ and a residual semi-simplification $\rho$, a \emph{presentation datum} for the moduli of representations of $G$ with residual semi-simplification $\rho$ is \begin{itemize}[leftmargin=2em] \item A (ordered) basis for $\rho$ making the expression $\rho \cong \bigoplus_{i=1}^r \rho_i$ compatible with the ordering of the factors of the block diagonal subalgebra $\bigoplus_{i=1}^r M_{d_i}(\mathbb{F}) \hookrightarrow M_d(\mathbb{F})$, where $\rho_i : E \rightarrow M_{d_i}(\mathbb{F})$. \item A $r$-pointed homotopy retract between $H$ and $C$, as in Example \ref{eg: r-pointed retract}. \item A choice of $\mathbb{F}^r$-algebra structure on $\mathbb{F}[G]^\wedge_{\ker\rho}$ arising from choices of idempotents as in \eqref{eq: r isom lift}, compatible with the standard matrix idempotents in the codomains of the $\rho_i$. \end{itemize} \end{defn} \subsection{Presentation of the moduli space ${\mathrm{Rep}}_\rho$} Fix a residual semi-simplification $\rho$. Next we want to deduce a presentation for ${\mathrm{Rep}}_\rho$. We will do this in the case that $\rho$ is multiplicity-free. Because representations of $G$ parameterized by ${\mathrm{Rep}}_\rho$ are continuous, it follows that the induced $\mathbb{F}[G]$-action factors through $E := \mathbb{F}{[\![} G{]\!]}$. Moreover, the condition that they have residual semi-simplification $\rho$ implies that they factor through the completion \[ E^\wedge_{\ker \rho}. \] (Note that one can assume that $\rho$ is multiplicity-free without any loss of generality on the algebras $E^\wedge_{\ker\rho}$ we study here.) The only difference from the previous subsection is that we now consider coefficients in $M_d(A)$ for \emph{commutative} algebras $A \in {\mathcal{A}f\!f}_\mathbb{F}$, with its standard $\mathbb{F}^r$-algebra structure. Formerly, in place of $M_d(A)$, we considered the category of coefficient algebras $\cA_\mathbb{F}^r$. We require some definitions in order to state the presentation. Recall from \S\ref{subsec: GIT intro} the notion of a simple closed path $\gamma$, the set of simple closed paths $SCP(\bold r)$, and tensor module of $\mathrm{Ext}^1$-modules $\mathrm{Ext}^1_{\mathbb{F}[G]}(\gamma)^*$. Throughout the following discussion, we use the notation $\mathrm{Ext}^k_{\mathbb{F}[G]}(\rho_j, \rho_i)$ to factor into summands the cohomology elsewhere denoted as $H^k = H^k(G, \mathrm{End}_\mathbb{F}(\rho)) = \mathrm{Ext}^k_{\mathbb{F}[G]}(\rho,\rho)$. \begin{defn} Let \[ I_\mathrm{cyc} \subset T_{\mathbb{F}^r} \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^* \cong T_{\mathbb{F}^r}\bigoplus_{1 \leq i,j \leq r} \Sigma\mathrm{Ext}^1_{\mathbb{F}[G]}(\rho_j, \rho_i)^* \] (note that this free $\mathbb{F}^r$-algebra is not completed) denote the ideal generated by the submodule of \emph{cyclic tensors} \[ \bigoplus_{\gamma \in SCP(\bold r)} \mathrm{Ext}^1_{\mathbb{F}[G]}(\gamma)^* \] (where because we are using non-symmetric tensors, we sum over all of the simple closed paths $SCP(\bold r)$ that constitute the simple cycles $SC(\bold r)$). Note that $SCP(\bold r)$ and $SC(\bold r)$ include the loop $i \rightarrow i$ for each $i \in \bold r$. Let $\hat T_\mathrm{cyc} \Sigma \mathrm{Ext}^1(\rho,\rho)^*$ denote the \textit{cyclic completion} of $T_{\mathbb{F}^r} \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*$, that is, its completion by $I_\mathrm{cyc}$, which admits an inclusion into $\hat T_{\mathbb{F}^r} \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)$. Let $\hat S_\mathrm{cyc} \Sigma \mathrm{Ext}^1(\rho,\rho)^*$ denote the abelianization of $\hat T_\mathrm{cyc} \Sigma \mathrm{Ext}^1(\rho,\rho)^*$. \end{defn} Notice that the submodule \[ m^*\mathrm{Ext}^2_{\mathbb{F}[G]}(\rho,\rho) \subset \hat T_{\mathbb{F}^r} \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho) \] lies within $\hat T_\mathrm{cyc} \Sigma \mathrm{Ext}^1(\rho,\rho)^*$. This inclusion follows from the following fact, which we will use frequently. \begin{fact} \label{fact: cyc bound} Any non-zero simple tensor of degree $s$ on $\mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*$ includes as a tensor-factor at least $\lfloor s/r\rfloor$ simple cyclic tensors, and is expressible as the product of a cyclic tensor with a tensor of degree bounded by $r$. \end{fact} Consequently, we may sensibly define \[ S^\mathrm{GMA}_D := \frac{\hat S_\mathrm{cyc} \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*} {m^* \Sigma \mathrm{Ext}^2_{\mathbb{F}[G]}(\rho,\rho)^*}. \] Here ``GMA'' refers to a moduli problem represented: $S^\mathrm{GMA}_D$ is a commutative $\mathbb{F}^r$-algebra and, in particular, is not a GMA. \begin{thm} \label{thm: adapted pres} Assume that $G$ satisfies the $\Phi_p$ finiteness condition. Assume also that the residual semi-simplification $\rho$ is multiplicity-free. Fix a presentation datum (Definition \ref{defn: pres datum}). These choices induce \begin{enumerate}[leftmargin=2em] \item an $\mathbb{F}^r$-algebra structure on the universal Cayley--Hamilton algebra $E_D$ over $D$, \item a GMA structure on $E_D$ such that $D_\mathrm{GMA} = D^u$, whose idempotents are compatible with the $\mathbb{F}^r$-structure of part (1), and \item a presentation ${\mathrm{Rep}}^\mathrm{GMA}_{E_D} \cong \Spf S^\mathrm{GMA}_D$ (where $S^\mathrm{GMA}_D$ is forgotten down from a $\mathbb{F}^r$-algebra to an object of ${\mathcal{A}f\!f}_\mathbb{F}$). \end{enumerate} \end{thm} \begin{proof} The choice of presentation datum furnishes the choices to which Theorem \ref{thm: profinite r-pointed dual} applies. Thus we have the $\mathbb{F}^r$-algebra structure on $E^\wedge_\rho$ given by applying the composite of \begin{itemize} \item the isomorphism $\rho^u : E^\wedge_\rho \buildrel\sim\over\ra M_d(\mathbb{F}) \underline{\otimes} R^\mathrm{nc}_\rho$ furnished by Theorem \ref{thm: profinite r-pointed dual} \item the block diagonal subalgebra map $\bigoplus_{i=1}^r M_{d_i}(\mathbb{F}) \hookrightarrow M_d(\mathbb{F})$ \end{itemize} to the identity elements of the matrix algebras $M_{d_i}(\mathbb{F})$. The natural map \[ E^\wedge_\rho \longrightarrow E_D \] then gives us result (1). To prove (2), observe that the set of $r$ ordered idempotents of $E^\wedge_\rho$ induced by the $\mathbb{F}^r$-structure induces a GMA structure on $E_D$. Indeed, just like the idempotents of $E_D$ supplied by \cite[Thm.\ 2.22(ii)]{chen2014} in Theorem \ref{thm: algfam main}(1), they lift the standard idempotents of $\mathbb{F}[G]/\ker\rho \cong E_D/\ker\rho \cong \bigoplus_{i=1}^r M_{d_i}(\mathbb{F})$. Therefore, by \cite[Thm.\ 2.9.18(iii), pg.\ 242]{rowen1988} (just as in the argument for \cite[Lem.\ 5.6.8]{WWE1}), there is some conjugation in $E_D$ sending one ordered set of idempotents to the other. We also know that if one set of ordered idempotents supplies a GMA structure, so does its conjugate. We have noted in Theorem \ref{thm: algfam main}(1) that the native pseudorepresentation $D^u : E_D \rightarrow R_D$ is equal to the pseudorepresentation $D_\mathrm{GMA} : E_D \rightarrow R_D$ induced by this GMA structure. This completes part (2). We define an auxiliary moduli functor ${\mathrm{Rep}}^{\mathbb{F}^r}_{\mathbb{F}[G]^\wedge_{\ker\rho}}$ by sending $A \in {\mathcal{A}f\!f}_\mathbb{F}$ to the set of $\mathbb{F}^r$-algebra homomorphisms \[ \mathbb{F}[G]^\wedge_{\rho} \longrightarrow M_d(A) \] that are compatible with the maps from $\bigoplus_{i=1}^r M_{d_i}(\mathbb{F})$ into each of them. (The map to $E^\wedge_\rho$ comes from the inverse of $\rho^u$.) We claim that the map $\mathbb{F}[G]^\wedge_\rho \rightarrow E_D$ induces an isomorphism of functors on ${\mathcal{A}f\!f}_\mathbb{F}$ \[ {\mathrm{Rep}}^{\mathbb{F}^r}_{\mathbb{F}[G]^\wedge_{\rho}} \cong {\mathrm{Rep}}^\mathrm{GMA}_{E_D}. \] Indeed, because the residual semi-simplification of all representations parameterized by ${\mathrm{Rep}}^{\mathbb{F}^r}_{\mathbb{F}[G]^\wedge_{\ker\rho}}$ is $\rho$, Theorem \ref{thm: algfam main} provides that these representations factor through $E_D$. Because of the compatibility of idempotents arranged above, these factorizations $E_D \rightarrow M_d(A)$ preserve the GMA structure. Therefore, each $A$-point of ${\mathrm{Rep}}^{\mathbb{F}^r}_{\mathbb{F}[G]^\wedge_{\rho}}$ induces an $A$-point of ${\mathrm{Rep}}^\mathrm{GMA}_{E_D}$ via this factorization. There is a left inverse map ${\mathrm{Rep}}^\mathrm{GMA}_{E_D} \hookrightarrow {\mathrm{Rep}}^{\mathbb{F}^r}_{\mathbb{F}[G]^\wedge_{\rho}}$ given by $\mathbb{F}[G]^\wedge_\rho \rightarrow E_D$, which is also a right inverse because ${\mathrm{Rep}}^\mathrm{GMA}_{E_D}$ and ${\mathrm{Rep}}^{\mathbb{F}^r}_{\mathbb{F}[G]^\wedge_{\rho}}$ admit compatible monomorphisms into ${\mathrm{Rep}}^\square_\rho$. Now we apply the Morita equivalence of $\mathbb{F}[G]^\wedge_\rho$ and $e\mathbb{F}[G]^\wedge_\rho e$ to draw an isomorphism of functors on ${\mathcal{A}f\!f}_\mathbb{F}$ \[ {\mathrm{Rep}}^{\mathbb{F}^r}_{e\mathbb{F}[G]^\wedge_{\rho}e} \cong {\mathrm{Rep}}^{\mathbb{F}^r}_{\mathbb{F}[G]^\wedge_{\rho}}, \] where ${\mathrm{Rep}}^{\mathbb{F}^r}_{e\mathbb{F}[G]^\wedge_{\rho}e}$ sends $A \in {\mathcal{A}f\!f}_\mathbb{F}$ to $\mathbb{F}^r$-algebra homomorphisms \[ e \mathbb{F}[G]^\wedge_\rho e \longrightarrow M_r(A), \] where the $\mathbb{F}^r$-structure on $M_r(A)$ comes from the standard diagonal idempotents with exactly one non-zero entry. (The same procedure is applied to GMAs in \cite[\S1.3.2]{BC2009}.) The augmented $\mathbb{F}^r$-algebra isomorphism of Theorem \ref{thm: profinite r-pointed dual} \[ (e\rho^u e)^{-1} : R^\mathrm{nc}_\rho = \frac{\hat T_{\mathbb{F}^r} \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*}{m^* (\Sigma \mathrm{Ext}^2_{\mathbb{F}[G]}(\rho,\rho)^*)} \buildrel\sim\over\lra e\mathbb{F}[G]^\wedge_{\rho}e \] allows us to calculate ${\mathrm{Rep}}^{\mathbb{F}^r}_{e\mathbb{F}[G]^\wedge_{\rho}e}$. Firstly we work in the case $\mathrm{Ext}^2_{\mathbb{F}[G]}(\rho,\rho) = 0$ and consider $\mathbb{F}^r$-homomorphisms \begin{equation} \label{eq: Fr Homs from R} R^\mathrm{nc}_\rho \cong \hat T_{\mathbb{F}^r} \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^* \longrightarrow M_r(A). \end{equation} Write $I_A \subset A$ for an ideal that is maximal among ideals of definition for $A$. By considering the standard coordinate-wise formulas for matrix multiplication, we observe that homomorphisms \eqref{eq: Fr Homs from R} biject with the $A$-submodule of \begin{equation} \label{eq: Ext1 coordinate decomp} \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho) \otimes A \cong \bigoplus_{1 \leq i, j \leq r} \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho_j, \rho_i)) \otimes A \end{equation} that is the intersection of the kernels of the maps \[ f(\gamma, x) : \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho) \otimes A \rightarrow A/I_A \] parameterized by $\gamma \in SCP(\bold r)$ and $x \in \mathrm{Ext}^1_{\mathbb{F}[G]}(\gamma)^*$. This submodule corresponds precisely to maps $T_{\mathbb{F}^r} \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^* \rightarrow M_r(A)$ that factor through its completion at $I_\mathrm{cyc}$, that is, $\hat T_\mathrm{cyc}\Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*$. \begin{rem} For the sake of clarity, we write out the definition of $f(\gamma,x)$. Given an element $y = (y_{ij})$ in $\Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho) \otimes A$, where the coordinates $y_{ij}$ arise from the decomposition \eqref{eq: Ext1 coordinate decomp}, we obtain the element \[ \bigotimes_{i=0}^{l_\gamma-1} x_{\gamma(i)\gamma(i+1)}(y_{\gamma(i)\gamma(i+1)}) \in A^{\otimes l_\gamma}. \] Multiplication in $A$ yields the desired element of $A$. \end{rem} Next, admit the case that $\mathrm{Ext}^2_{\mathbb{F}[G]}(\rho,\rho) \neq 0$. From the properties of free algebras, we deduce that ${\mathrm{Rep}}^{\mathbb{F}^r}_{e\mathbb{F}[G]^\wedge_\rho e}(A)$ is naturally isomorphic to \[ \mathrm{Hom}_{\mathbb{F}^r}(\frac{\hat T_\mathrm{cyc}\Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*}{m^* \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*}, M_r(A)). \] This in turn is naturally isomorphic to \[ \mathrm{Hom}_{\mathbb{F}}(\frac{\hat T_\mathrm{cyc}\Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*}{m^* \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*}, A), \] where we simply forget the $\mathbb{F}^r$-algebra structure on the domain. Finally, because $A$ is assumed to be commutative, we deduce the desired result (3). \end{proof} \begin{rem} One can derive from Theorem \ref{thm: adapted pres} the relationship established in \cite[\S1.5.3-1.5.4]{BC2009} between the structure of the GMA $E_D$ and various $\mathrm{Ext}$-groups, mainly $\mathrm{Ext}^1$. See especially \cite[Rem.\ 1.5.7]{BC2009}, which discusses the relationship with $\mathrm{Ext}^2$: the ambiguities there are controlled by the $m^*$ map on $\mathrm{Ext}^2_{\mathbb{F}[G]}(\rho,\rho)^*$. \end{rem} \subsection{Presentation of the pseudodeformation ring} \label{subsec: PsDef proofs} Next, we apply Theorem \ref{thm: algfam main}(2) in order to present the pseudodeformation ring $R_D$. We recall from Definition \ref{defn: R1D} the complete Noetherian local ring $R^1_D$, which will be shown to present $R_D$ when $\mathrm{Ext}^2_{\mathbb{F}[G]}(\rho,\rho) = 0$. \begin{thm} \label{thm: R_D pres body} Assume that $G$ satisfies the $\Phi_p$ finiteness condition and assume that the residual semi-simplification $\rho$ is multiplicity-free. Fix a presentation datum (Definition \ref{defn: pres datum}). These choices produce a presentation of $R_D$ as a complete Noetherian local $\mathbb{F}$-algebra with residue field $\mathbb{F}$, \begin{equation} \label{eq: R_D pres body} \frac{R^1_D}{ \Bigg(\displaystyle\bigoplus_{i,j \in \bold r} m^* \Sigma \mathrm{Ext}^2_G(\rho_j, \rho_i)^* \otimes \Big(\displaystyle\bigoplus_{\gamma \in SCC(i,j)} \Sigma \mathrm{Ext}^1_G(\gamma)^*\Big)\Bigg) } \buildrel\sim\over\lra R_D. \end{equation} \end{thm} \begin{proof} The presentation for $R_D$ follows from combining \begin{itemize}[leftmargin=2em] \item the presentation for $S^\mathrm{GMA}_{E_D}$ of Theorem \ref{thm: adapted pres} with \item the result of \cite[\S2-3]{WE2018} stated in Theorem \ref{thm: algfam main}(2), that $R_D$ is the invariant subring of $S^\mathrm{GMA}_{E_D}$ under the adjoint action of the torus $Z(\rho)$. \end{itemize} Because $Z = Z(\rho)$ is linearly reductive over $\mathbb{F}$ (even in positive characteristic, as it is a torus), its invariant functor is exact. In particular, for any affine $\mathbb{F}$-scheme $\Spec S$ that is a closed subscheme $\Spec S \cong \Spec S'/I \subset \Spec S'$ admitting a $Z$-action on $S'$ that preserves $S$, one has (see e.g.\ \cite[Rem.\ 4.11]{alper2013}) \[ S^Z = (S'/I)^Z \cong S'^Z/I^Z. \] We apply this to the presentation of $S^\mathrm{GMA}_{E_D} = S'/I$ as a quotient, where \[ S' := \hat S_\mathrm{cyc} \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*, \qquad I := (m^* \Sigma \mathrm{Ext}^2_{\mathbb{F}[G]}(\rho,\rho)^*); \] indeed, $I$ is $Z$-stable because it has generators that are isotypic for certain characters of $Z$ (these are in bijection with weights of the adjoint action of $Z$ as the torus in $\PGL_r$). We claim that $S'^Z \cong R^1_D$. Indeed, we see that a simple tensor in $S'$ is fixed by the adjoint action if and only if it is a cyclic tensor, and cyclic tensors generate precisely the image of \eqref{eq: R1D} within the codomain $\hat S \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*$. Clearly the image is contained in the subring $\hat S_\mathrm{cyc} \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*$. Similarly, we see that a generating set for $I^Z \subset I$ is formed as follows. Choose $\mathbb{F}$-basis $\{b_{i,j,k}\}$ for the generating vector space $m^*\Sigma \mathrm{Ext}^2_{\mathbb{F}[G]}(\rho,\rho)^*$ of $I$, such that its subset with fixed $(i,j)$ is a basis for $m^*\Sigma \mathrm{Ext}^2_{\mathbb{F}[G]}(\rho_j,\rho_i)^*$. Thus each $b_{i,j,k}$ is isotypic for the $Z$-action with the action depending only on $(i,j)$; call this character $\chi_{i,j}$. Then we observe that $I_Z$ is generated by $b_{i,j,k} \otimes c_{i,j,l}$, where for fixed $(i,j)$ the $c_{i,j,l}$ are a basis for a generating vector space of the $\chi_{i,j}^{-1} = \chi_{j,i}$-isotypic part of $\hat S_\mathrm{cyc} \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)$ as an $R^1_D$-module. The minimal such vector space is \[ \bigoplus_{\gamma \in SCC(i,j)} \Sigma \mathrm{Ext}^1_G(\gamma)^*. \] Then observe that this generating set $\{b_{i,j,k} \otimes c_{i,j,l}\}$ is a basis for the vector space in the denominator of \eqref{eq: R_D pres body}. \end{proof} Having presented $R_D$, we can prove the rest of the results of \S\ref{subsec: results2}, which are corollaries of the presentation. From now on, we refer to \S\ref{subsec: results2} for the statement of these corollaries. \begin{proof}[{Proof of Corollary \ref{cor: tangent pseudo} (Presentation of the tangent space $\mathfrak{t}_D$ of $R_D$)}] Firstly we observe directly from the definition of $R^1_D$ (Definition \ref{defn: R1D}) that when $\mathrm{Ext}^2_{\mathbb{F}[G]}(\rho,\rho) = 0$, then the presentation datum induces a presentation of its tangent space $\mathfrak{t}^1_D$ as \[ \mathfrak{t}^1_D \cong \bigoplus_{\gamma \in SC(\bold r)} \Sigma \mathrm{Ext}^1_{\mathbb{F}[G]}(\gamma). \] Using the presentation of $R_D$ in Theorem \ref{thm: R_D pres body} and unraveling the definition of $m^*$ in terms of the $A_\infty$-operations $m_n$ from the bar equivalence (\S\ref{subsec: bar}) for $n \leq r$, we produce the maps out of $\mathfrak{t}^1_D$ that appear in the statement of the corollary. Their common kernel is $\mathfrak{t}_D \subset \mathfrak{t}^1_D$. \end{proof} \subsection{Canonical structure of the tangent space} \label{subsec: tangent} As emphasized in Warning \ref{warn: NC pseudo}, the direct sum expression of $\mathfrak{t}_D$ in Corollary \ref{cor: tangent pseudo} is highly non-canonical, being dependent on the presentation datum. In contrast, the \emph{complexity filtration} of \cite[\S3]{bellaiche2012} is a canonical filtration on $\mathfrak{t}_D$ whose graded pieces are summands appearing in Corollary \ref{cor: tangent pseudo}. \begin{defn}[Bella\"iche] \label{defn: complexity} The \emph{complexity} of a pseudodeformation $D_A$ of $D$ is the minimal integer $c(D_A)$ such that for all $\gamma \in SC(\bold r)$ with length strictly greater than $c(D_A)$, the image of $\Sigma\mathrm{Ext}^1_{\mathbb{F}[G]}(\gamma)^*$ in $R_D$ under the presentation \eqref{eq: R_D pres body} are sent to zero under the induced map \[ R_D \rightarrow A. \] Equivalently, the map of tangent spaces $(\mathfrak{m}_A/\mathfrak{m}_A^2)^* \rightarrow \mathfrak{t}_D$ has image contained in \[ \mathrm{Fil}_{c(D_A)} \mathfrak{t}_D := \bigoplus_{\substack{\gamma \in SC(\bold r) \\ l(\gamma) \leq c(D_A)}} \Sigma\mathrm{Ext}^1_{\mathbb{F}[G]}(\gamma). \] The \emph{complexity filtration} on $\mathfrak{t}_D$ is the increasing exhaustive filtration consisting of the sums above. We denote it by $\mathrm{Fil}_k \mathfrak{t}_D \subset \mathfrak{t}_D$; observe that $\mathrm{Fil}_0 \mathfrak{t}_D = 0$ and $\mathrm{Fil}_r \mathfrak{t}_D = \mathfrak{t}_D$. \end{defn} \begin{lem} \label{lem: CF is can} The complexity filtration of $\mathfrak{t}_D$ is canonical. \end{lem} \begin{proof} To see that the complexity filtration is canonical, first observe that the ideal of cycles is a canonical ideal $I_\mathrm{cyc} \subset S^\mathrm{GMA}_{E_D}$. Indeed, it is $I_\mathrm{cyc} = \mathfrak{m}_D \cdot S^\mathrm{GMA}_{E_D}$, where $\mathfrak{m}_D \subset R_D \hookrightarrow S^\mathrm{GMA}_{E_D}$ as the invariant subring of the $Z$-action (see the proof of Theorem \ref{thm: R_D pres body}). Then, we have a decreasing filtration of $S^\mathrm{GMA}_{E_D}$ given by $I_\mathrm{cyc}^n$, $n \geq 0$. The decreasing filtration on the cotangent space $\mathfrak{m}_D/\mathfrak{m}_D^2$ of $R_D$ given by the intersection of $\mathfrak{m}_D$ with $I_\mathrm{cyc}^n$ is perfectly dual to the complexity filtration of $\mathfrak{t}_D$. \end{proof} As we pointed out in Remark \ref{rem: refine Bel12}, the use of $A_\infty$-products or Massey higher products refines the result \cite[Thm.\ 1]{bellaiche2012}, which only used cup products. It results in a canonical determination of $\mathrm{gr}_k \mathfrak{t}_D := \mathrm{Fil}_k \mathfrak{t}_D / \mathrm{Fil}_{k-1} \mathfrak{t}_D$. \begin{proof}[{Proof of Corollaries \ref{cor: CFP} (determination of $\mathrm{gr}_k \mathfrak{t}_D$) and \ref{cor: pseudo tangent dim} (tangent dimension)}] Notice that the sought-after presentation of $\mathrm{gr}_k \mathfrak{t}_D$ in the statement of Corollary \ref{cor: CFP} appears as the sub-summand of the direct sum expression for $\mathfrak{t}_D$ that appears in Corollary \ref{cor: tangent pseudo}. This sub-summand is cut out by restricting the indices $\mathfrak{t}_D = \bigoplus_{\gamma \in SC(\bold r)} (-)$ to those $\gamma$ with length $k$. Therefore, after applying Corollary \ref{cor: tangent pseudo} and Lemma \ref{lem: CF is can}, we get a map \begin{equation} \label{eq: gr_k embedding} \mathrm{gr}_k \mathfrak{t}_D \hookrightarrow \bigoplus_{\substack{\gamma \in SC(\bold r) \\ l_\gamma = k}} \Sigma\mathrm{Ext}^1_G(\gamma). \end{equation} It remains to show that the image, as claimed in Corollary \ref{cor: CFP}, is canonical (i.e.\ not dependent upon the choice of presentation datum). The content of \cite[\S3.3]{bellaiche2012} is the proof that there is a canonical injection exactly as in \eqref{eq: gr_k embedding}. It remains to check that this injection is compatible with \eqref{eq: gr_k embedding} (which arises from Corollary \ref{cor: tangent pseudo}). Both of these injections ultimately arise from the canonical $Z$-equivariant isomorphism $\ker(\rho)/\ker(\rho)^2 \cong \Sigma\mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*$ (where $\ker\rho \subset \mathbb{F}[G]$) by taking symmetric tensor powers and $Z$-invariants, we have the desired compatibility. Counting the dimensions of the vector spaces appearing in the expression for $\mathrm{gr}_k \mathfrak{t}_D$, we deduce Corollary \ref{cor: pseudo tangent dim}. \end{proof} \subsection{Input from invariant theory and quiver representation theory} \label{subsec: invariants quivers} Like the tangent dimension, the bounds on the Krull dimension of $R_D$ claimed in Corollary \ref{cor: pseudo krull bounds} follow mostly from counting dimensions in the presentation of $R_D$ of Theorem \ref{thm: R_D pres body}. The additional ingredient is our extra knowledge about the ring $R^1_D$ from invariant theory. These have been stated in Fact \ref{fact: R1D}, which is a brief and partial summary of extensive literature about invariant subrings of regular rings under (linearly) reductive group actions. The point is that there are finite combinatorial objects controlling $R^1_D$. In particular, the relations cutting out these rings are polynomial, in contrast to the power series arising from $A_\infty$-products. \begin{proof}[{References for Fact \ref{fact: R1D}}] It is clear that $R^1_D$ is reduced, as it is a subring of a domain. The fact that a subring of invariants of a regular commutative algebra under the action of a linearly reductive algebraic group is normal and Cohen-Macaulay is due to Hochster \cite{hochster1972}. The decomposition into tensor factors, each arising from a strongly connected component, follows from the generation of $R^1_D$ by cyclic tensors. The claim about the Krull dimension is \cite[Thm.\ 6]{LBP1990}. There, the authors use quivers and their representations; their paper begins with an introduction to these notions. The representation-unobstructed setting we are in can be translated to theirs by observing that the representation theory of $\hat T_{\mathbb{F}^r} \Sigma\mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*$ is the same as the representation theory of the quiver with $r$ vertices labeled by $\{1, \dotsc, r\}$, and with $h^1_{i,j}$ directed arrows from $i$ to $j$. The dimension vector $\alpha$ (in the notation of \emph{loc.\ cit.}) is $\alpha = (1, \dotsc, 1) \in \mathbb{N}^{\oplus r}$, because each $\rho_i$ appears with multiplicity 1 in $\rho$. The ``Ringel bilinear form'' $R$ on $\mathbb{Z}^{\oplus r} \times \mathbb{Z}^{\oplus r}$ is represented with the matrix \[ R = (R_{i,j}) = \dim_\mathbb{F} \mathrm{Hom}_{\mathbb{F}[G]}(\rho_i, \rho_j) - \dim_\mathbb{F} \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho_i, \rho_j) = \delta_{i,j} - h^1_{i,j}. \] The we see in \emph{loc.\ cit.}\ that the Krull dimension of $R^1_D$ is $R(\alpha, \alpha)$, which is equal to $1-r + \sum_{1 \leq i,j \leq r} h^1_{i,j}$, as claimed. The final claim follows from Krull dimension being additive under tensor products of commutative $\mathbb{F}$-algebras. \end{proof} \begin{rem} It has been determined when $R_D$ is regular \cite[\S1.1, Thm.~4]{KKMS1973}, complete intersection \cite{nakajima1986} and Gorenstein \cite{stanley1978}. \end{rem} \begin{rem} For a broader perspective on quiver representations in relation to this discussion, see \cite[\S\S5.7-5.8]{lebruyn2008}, which relies on work of Bocklandt. The statement on Krull dimension is reproduced in [\textit{loc.\ cit.}, Lem.\ 5.13]. \end{rem} \begin{eg} \label{eg: R1D LCI} There is a single simple cycle $\gamma = (12)$, and the only relation that must be imposed is the extra commutativity relation \begin{equation} \label{eq:r=2_obs} (x_{12} \otimes x_{21})\cdot (y_{12} \otimes y_{21}) = (x_{12} \otimes y_{21}) \cdot (y_{12} \otimes x_{21}), \end{equation} which results in an quadratic obstruction of dimension $\binom{d_{12}}{2} \binom{d_{21}}{2}$. In particular, $R_D$ is regular when either of $d_{12}$ or $d_{21}$ is $\leq 1$. Consider the case $d_{12} = d_{21} = 2$. We can then take $\{x_{ij}, y_{ij}\} \subset \mathrm{Ext}^1_G(\rho_j, \rho_i)^*$ to be a $k$-basis, so that the space of relations is $1$-dimensional, generated by \eqref{eq:r=2_obs} as written. If we write \[ W = x_{12} \otimes x_{21}, \quad X = x_{12} \otimes y_{21}, \quad Y = y_{12} \otimes x_{21}, \quad Z = y_{12} \otimes y_{21}, \] then we readily see that \[ R_D \cong \frac{k{[\![} W,X,Y,Z{]\!]}}{(WZ-XY)}. \] \end{eg} \begin{eg} The singularity of Example \ref{eg: R1D LCI} is the only possible singularity of $R^1_D$ when its Krull dimension is 3. For a classification of singularities in $R^1_D$ when its Krull dimension is $\leq 6$, see \cite{BLV2003}. \end{eg} \subsection{Obstruction theory} \label{subsec: obs theory} Here we prove the remaining statements of \S\ref{subsec: results2}. All that is left to add is the following basic formulation of obstruction theory. While the obstructions $\beta$ are representation-theoretic, the obstructions $\alpha$ are combinatorial and can be calculated using the content of \S\ref{subsec: invariants quivers}. \begin{proof}[{Proof of Fact \ref{fact: basic obstruction}}] It is a standard fact that whenever $(S, \mathfrak{m}_S)$ is a regular local $\mathbb{F}$-algebra surjecting onto $S'$ with kernel $I$, then the obstruction to lifting a homomorphism $\ell_n : S' \rightarrow \mathbb{F}[\varepsilon]/(\varepsilon)^n$ to a homomorphism $S' \rightarrow \mathbb{F}[\varepsilon]/(\varepsilon)^{n+1}$ is an element of $I/\mathfrak{m} I$ that can be produced from $\ell_n$. See, for example, \cite[\S1.5, Prop.\ 2, pg.\ 399]{mazur1989}. Then Fact \ref{fact: basic obstruction} follows from observing that $h^2(\cC(D))$, as defined in Notation \ref{note: Ext}, is a basis for $I/\mathfrak{m} I$ in this case. Remark \ref{rem: bounded deg obs} follows from observing that $J$ of Notation \ref{note: Ext} is finitely generated as a monoid. \end{proof} \begin{proof}[{Proof of Corollary \ref{cor: obstruction}}] In light of the presentation for $R_D$ of Theorem \ref{thm: R_D pres body}, part (1) follows from Fact \ref{fact: basic obstruction}. Given that the obstruction $\alpha(D_n)$ of part (1) vanishes, Part (2) follows from the principle of Fact \ref{fact: basic obstruction}; the following computation realizes this principle. Choose $i,j \in \bold r$ and element $\omega \in \Sigma\mathrm{Ext}^2_{\mathbb{F}[G]}(\rho_j, \rho_i)^*$. We have $m^*(\omega)$ in the $(i,j)$-part of $\hat T_{\mathbb{F}^r} \mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*$. For each $\gamma \in SCC(i,j)$ and $\kappa \in \Sigma\mathrm{Ext}^1_{\mathbb{F}[G]}(\gamma)^*$, $m^*(\omega) \otimes \kappa$ is an element of the $(k,k)$-part of $\hat T_{\mathbb{F}^r} \Sigma\mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*$. Therefore, taken as an element of $\hat S_\mathrm{cyc} \Sigma\mathrm{Ext}^1_{\mathbb{F}[G]}(\rho,\rho)^*$, $m^*(\omega) \otimes \kappa$ is in the subring $R^1_D$. It was killed by $\psi_n^1$ (because $\psi_n$ exists), therefore $\psi_{n+1}^1(m^*(\omega) \otimes \kappa) \in \epsilon^{n+1}\mathbb{F} \cong \mathbb{F}$. By duality, we see that $\psi_{n+1}^1$ gives rise to an element of the $\mathbb{F}$-linear dual of the denominator of the presentation \eqref{eq: R_D pres body}, as desired. \end{proof} \section{Galois representations satisfying arithmetic conditions} \label{sec: conds} The goal of this section is to prove Theorem \ref{thm: conds}. This theorem claims that there exists a dg-algebra to which one can apply the theory of \S\ref{sec: profinite presentations} to compute these deformation spaces with an extra condition $\cC$. Indeed, so far we have considered only the ``unrestricted case,'' presenting deformation spaces for representations of $G$. To do this, we produce an algebra quotient $E^\cC$ of $\mathbb{F}{[\![} G{]\!]}$ that factors exactly those actions with condition $\cC$. Then we use the Hochschild cochain complex of the endomorphism module of a representation of this algebra to compute the deformations, applying the theory of \S\ref{sec: profinite presentations}. \subsection{Stable conditions and Cayley--Hamilton conditions} \label{subsec: stable or CH} We begin by recalling previous results that construct the moduli spaces of representations of $G$ with condition $\cC$. We begin with a description of the conditions $\cC$ that we will consider. Instead of working with particular conditions, we set up two different sorts of conditions to which we can apply our theorems. In this particular section, we work with mixed characteristic coefficients, using $\mathbb{Z}_p$ for simplicity. \begin{defn}[Stable condition] \label{defn: stable} A \emph{stable condition} $\cC$ is a subcategory of finite length $\mathbb{Z}_p[G]$-modules that is closed under subquotients and finite direct sums. \end{defn} Stable conditions first appeared in the study of Galois representations by Ramakrishna \cite{ramakrishna1993}, in the context of deformations rings of representations. When $\rho$ has scalar endomorphisms, he produced a quotient $R_\rho \twoheadrightarrow R^\cC_\rho$ parameterizing exactly those deformations satisfying condition $\cC$. In the author's joint work with Preston Wake \cite{WWE4}, stable conditions were shown to naturally extend to Cayley--Hamilton representations, thereby allowing for \begin{itemize}[leftmargin=2em] \item a sensible definition of pseudorepresentations of a profinite group $G$ with condition $\cC$, \item a universal pseudodeformation ring $R^\cC_D$ parameterizing deformations of $D$ satisfying $\cC$, which cuts out a closed condition $R_D \twoheadrightarrow R^\cC_D$ in the whole deformation space, \item the construction of a Zariski-closed subspaces ${\mathrm{Rep}}^{\square, \cC}_\rho \subset {\mathrm{Rep}}^\square_\rho$, ${\mathrm{Rep}}^\cC_\rho \subset {\mathrm{Rep}}_\rho$ of representations with residual semi-simplification $\rho$ and condition $\cC$. \end{itemize} Here is the second kind of condition that we will study. In order to formulate this notion, we use the category of Cayley--Hamilton representations over a residual semi-simplification $\rho$; see \cite[Defn.\ 2.1.5]{WWE4}. \begin{defn} \label{defn: CH cond} Let $\rho$ be a residual semi-simplification of $G$. We say that $\cC$ is a \emph{Cayley--Hamilton condition (over $\rho$)} when it applies to any Cayley--Hamilton representation over $\rho$ and is ``representable,'' in the sense that there exists a universal object in the category of Cayley--Hamilton representations over $\rho$. We will denote such a universal object by $E^\cC_D$. \end{defn} In particular, a Cayley--Hamilton condition $\cC$ over $\rho$ is not assumed to apply to arbitrary finite length $\mathbb{F}[G]$-modules, nor even to those with composition factors among the factors of $\rho$. A structure of a Cayley--Hamilton representation is required. \begin{eg}[Ordinary representations] \label{eg: ordinary} A 2-dimensional representation of $G_\mathbb{Q}$ is called \emph{ordinary} when its restriction to a decomposition group at $p$ admits a 1-dimensional unramified quotient. In \cite[\S5]{WWE1}, as well as \cite[\S7]{WE2018}, this condition was extended to a Cayley--Hamilton condition. In each of these works, a ``residually $p$-distinguished'' condition was required in order to use the structure of a GMA on all Cayley--Hamilton representations (in contrast, see \cite{CS2019}). The condition was then that the GMA representation must be upper-triangular after restriction to a decomposition group at $p$, with the quotient character being unramified. Following \cite{CS2019}, a GMA-structure-free notion of the ordinary Cayley--Hamilton condition was produced in \cite[\S3.7]{WWE5}. \end{eg} \subsection{Construction of a universal associative algebra with condition $\cC$} \label{subsec: stable constr} In this section, provided that $\cC$ is a stable condition, we generalize the construction of \cite[\S2.4]{WWE4} in order to produce an associative algebra $E^\cC$. This algebra $E^\cC$ play the role of $\mathbb{F}{[\![} G{]\!]}$ after imposing condition $\cC$. Once we have $E^\cC$, then for any representation $\rho$ of $G$ satisfying $\cC$, the completion $E^\cC_\rho := (E^\cC)^\wedge_{\ker \rho}$ is a desirable analogue of $\mathbb{F}{[\![} G{]\!]}_{\ker \rho}$. \begin{rem} In the case that $\cC$ is a Cayley--Hamilton condition over $\rho$, we only have the analogue $E^\cC_D$ of $\mathbb{F}[G]^\wedge_{\ker\rho}$. This is a limitation of the Cayley--Hamilton case relative to the stable case. \end{rem} Let $\cC$ be a stable condition. Notice that $\mathbb{Z}_p{[\![} G{]\!]}$ is a profinite algebra, which is a topological limit of its finite quotient algebras \[ E(a,b) := \mathbb{Z}_p/p^a\mathbb{Z}_p[G_b], \] where $G \cong \varprojlim_b G_b$ is a profinite presentation for $G$. Therefore, condition $\cC$ may be sensibly applied to $E(a,b)$, taking it as a left $\mathbb{Z}_p{[\![} G{]\!]}$-module. By \cite[Lem.\ 2.3.5]{WWE4}, there is a maximal quotient module $E(a,b) \twoheadrightarrow E(a,b)^\cC$ satisfying $\cC$. Therefore, for any $a' \geq a, b' \geq b$, the $\mathbb{Z}_p{[\![} G{]\!]}$-module quotient $E(a', b') \twoheadrightarrow E(a,b)^\cC$ factors through $E(a',b')^\cC$. Therefore we can produce a new limit \[ E^\cC := \varprojlim_{a,b} E(a,b)^\cC. \] Now, as in \cite[Lem.\ 2.4.3(2)]{WWE4}, considering the the right action of $\mathbb{Z}_p{[\![} G{]\!]}$ on $E(a,b) \twoheadrightarrow E(a,b)^\cC$ allows one to find that $E(a,b)^\cC$ has the structure of an algebra quotient of $E(a,b)$. Naturally, in the limit this makes $E^\cC$ an algebra quotient of $\mathbb{Z}_p{[\![} G{]\!]}$. When $\rho$ has property $\cC$, then $\rho : \mathbb{F}{[\![} G{]\!]} \rightarrow \mathrm{End}_\mathbb{F}(V)$ factors through $E^\cC$, and we let \[ E^\cC_\rho := \varprojlim_i E^\cC/\ker(\rho)^i. \] \begin{rem} This generalizes \cite[\S2.4]{WWE4} only in that the construction discussed there is applied to a general profinite algebra instead of only Cayley--Hamilton algebras. \end{rem} \subsection{Proof of Theorem \ref{thm: conds}} Let $\cC$ be a stable condition and let $E^\cC$ be as constructed above. We choose a representation $\rho$ of $G$ with condition $\cC$. Notice that we now have a natural inclusion of Hochschild cochain complexes \[ C^\bullet(E^\cC, \mathrm{End}_\mathbb{F}(\rho)) \subset C^\bullet(\mathbb{F}[G], \mathrm{End}_\mathbb{F}(\rho)) \] that is an inclusion of dg-$\mathbb{F}$-algebras. Choose compatible presentation data (as in Definition \ref{defn: pres datum}) for $E$ and $\mathbb{F}[G]$, meaning that the homotopy retract data $(i,p,h)$ on $C^\bullet(\mathbb{F}[G], \mathrm{End}_\mathbb{F}(\rho))$ restricts to that on $C^\bullet(E^\cC, \mathrm{End}_\mathbb{F}(\rho))$; and that the choices of idempotents of $E^\wedge_\rho$ and $\mathbb{F}[G]^\wedge_\rho$ are compatible under $\mathbb{F}[G] \rightarrow E^\cC$. Write $C^\bullet_\cC$ for $C^\bullet(E^\cC, \mathrm{End}_\mathbb{F}(\rho))$, and write $H^\bullet_\cC$ for its cohomology. We now have a more specific statement of Theorem \ref{thm: conds}. \begin{thm} \label{thm: conds body} Let $G$ be a profinite group satisfying finiteness condition $\Phi_p$, let $\rho$ be a multiplicity-free residual semi-simplification, and let $\cC$ be a stable condition. The compatible presentation data above induce \begin{itemize}[leftmargin=2em] \item a presentation for $E^\cC_\rho$ as a $\mathbb{F}^r$-algebra, exactly as in Theorem \ref{thm: profinite r-pointed dual}, \item presentations for ${\mathrm{Rep}}^{\cC,\mathrm{GMA}}_\rho$ and ${\mathrm{Rep}}^\cC_\rho$ as formal moduli spaces over $\Spf \mathbb{F}$, exactly as in Theorem \ref{thm: adapted pres}, and \item a presentation for $R^\cC_D$ as an object of $\hat\cA_\mathbb{F}$, exactly as in Theorem \ref{thm: R_D pres body} \end{itemize} equipped with morphisms to (resp.\ from) the analogous unrestricted objects $\mathbb{F}[G]^\wedge_{\ker\rho}$, ${\mathrm{Rep}}^{\square}_\rho$, ${\mathrm{Rep}}_\rho$, and $R_D$. All are closed immersions (resp.\ quotients of rings). \end{thm} \begin{proof} Using the argument for \cite[Lem.\ 2.4.3(3)]{WWE4}, we know that a $\mathbb{F}{[\![} G{]\!]}$-module has property $\cC$ if and only if the action factors through $E^\cC$. With this understood, we may repeat the proof of Theorem \ref{thm: adapted pres}. \end{proof} Naturally, the formulas for the tangent space, representation-unobstructed case, obstruction theory, and Krull dimension also follow from the presentations of Theorem \ref{thm: conds body}, just as in \S\ref{subsec: tangent} and \S\ref{subsec: obs theory}. \begin{rem} \label{rem: Ext2 discrepancy} We discuss the contrast between the two cases \begin{itemize} \item $\cC$ is a stable condition \item $\cC$ is a Cayley--Hamilton condition over $\rho$ and $D = \psi(\rho)$. \end{itemize} In order to compare them, it is instructive to consider the Cayley--Hamilton condition $\cC^\mathrm{CH}$ over $\rho$ arising from a stable condition, and the difference in the Hochschild complex for $E^\cC$ vs.\ $E^{\cC^\mathrm{CH}}_D$. (Because the maximal Cayley--Hamilton quotient of $E^\cC$ over $\rho$ is $E^\cC_D$, we write $E^\cC_D$ for $E^{\cC^\mathrm{CH}}_D$.) It follows from the theorem that they will each compute exactly the same deformation space. It is also the case that $\mathrm{Ext}^k_{\cC}(\rho_j, \rho_i) \cong \mathrm{Ext}^k_{E^\cC_D}(\rho_j, \rho_i)$ for $k = 0,1$. But there is a difference when $k=2$. The classes of $\mathrm{Ext}^2_{E^\cC_D}(\rho_j, \rho_i)$ are only those which appear as Massey products of representations with filtrations with graded pieces among the $(\rho_i)_{i=1}^r$. But $E^\cC$, having been constructed without any completion at $\ker\rho$, has the full $\mathrm{Ext}^2$-groups of the subcategory $\cC$ of $\mathbb{F}[G]$-modules. We regard the $\mathrm{Ext}$-groups arising from $E^\cC$ as being meaningful even in a derived sense, while those that arise from $E^{\cC^\mathrm{CH}}$ as being useful only for computations of classical deformations. \end{rem} \section{Ranks of $p$-adic Hecke algebras} \label{sec: ranks} In this section, we give examples of $p$-adic completions (or interpolations) $\mathbb{T}$ of classical Hecke algebras acting on modular forms. These are known to be free of finite rank over a regular local complete Noetherian $\mathbb{Z}_p$-algebra. Here, we give an expression for this rank in terms of $A_\infty$-products. This is a measure of the size of the module of congruent modular forms. It would be interesting to relate this quantity to some sort of analytic invariant, i.e.\ an appropriate modulo $p$ version of an adjoint $L$-function. The approach is to use a known isomorphism $R \buildrel\sim\over\ra \mathbb{T}$ where $R$ is some deformation ring of Galois representations, and then use $A_\infty$-products to calculate the rank of $R$. We discuss four cases. \begin{itemize}[leftmargin=2em] \item The finite-flat residually non-Eisenstein setting of Wiles \cite{wiles1995}, \item the ordinary residually non-Eisenstein setting of Wiles \textit{op.\ cit}., \item the ordinary Eisenstein setting of Ribet's converse to Herbrand's theorem \cite{ribet1976}, and \item the finite-flat residually Eisenstein setting of Mazur's Eisenstein ideal \cite{mazur1978}, following \cite{WWE3}. \end{itemize} ``Finite-flat'' and ``ordinary'' are conditions on Galois representations. The residually Eisenstein/non-Eisenstein distinction is more serious, because on the Galois side this is the residually irreducible/reducible distinction. Let $G = G_{\mathbb{Q},S}$ be the Galois group of $\mathbb{Q}$ with ramification only at a finite set of places $S$, where $S$ is the support of $Np\infty$ and where $N \in \mathbb{Z}_{\geq 1}$ satisfies $p \nmid N$. When $\ell$ is a prime number, let $G_\ell = \mathrm{Gal}(\overline{\Q}_\ell/\mathbb{Q}_\ell) \to G$ be the homomorphism arising from a choice of decomposition subgroup for $\ell$ in $G_{\mathbb{Q},S}$. \subsection{Weight 2 non-Eisenstein non-ordinary Hecke algebras} \label{subsec: flat Wiles} In this section and the next, we discuss the Hecke algebras proved to be isomorphic to a Galois deformation ring by Wiles \cite{wiles1995}. Here, we focus on the finite-flat case, which is case (ii) on [pg.\ 456, \textit{loc.\ cit.}]. Write $\rho$ here for the residual representation \[ \rho : G_{\mathbb{Q},S} \longrightarrow \GL_2(\mathbb{F}). \] written $\rho_0$ in \textit{loc.\ cit.}; in particular, it is odd and absolutely irreducible, and $\rho\vert_{G_p}$ is finite-flat. We assume that $\rho$ is ramified at all primes $\ell \mid N$, and that $\rho\vert_{I_\ell}$ satisfies cases (B) or (C) of [pg.\ 458, \textit{loc.\ cit.}]. (We do not permit case (A).) This makes for a deformation condition denoted $\cD$ there, which includes the finite-flat condition on restriction to $G_p$. We let $G$ be the maximal quotient of $G_{\mathbb{Q},S}$ in which $\ker(\rho\vert_{I_\ell}) \subset I_\ell$ vanishes for all primes $\ell \mid N$. Then, let $E_\cD$ be the maximal quotient of $\mathbb{F}{[\![} G{]\!]}$ that is finite-flat upon restriction to $G_p$, in the sense of \S\ref{subsec: stable constr}. \begin{prop} The Hochschild cochain complex $C_\cD := \cC^\bullet(E_\cD, \mathrm{End}_\mathbb{F}(\rho))$ calculates the deformation ring $R_\cD/pR_\cD$ of \cite[pg.\ 458]{wiles1995} via Theorems \ref{thm: main A-inf} and \ref{thm: conds body}. In particular, the $A_\infty$-products in $H^\bullet(C_\cD)$ determine the rank of $R_\cD \cong \mathbb{T}$. \end{prop} \begin{proof} First we establish that a deformation $\tilde \rho$ of $\rho$ as a $G_{\mathbb{Q},S}$-representation satisfies Wiles's deformation problem $\cD$ if and only if $\tilde \rho$ factors through $\mathbb{F}{[\![} G_{\mathbb{Q},S}{]\!]} \twoheadrightarrow \mathbb{F}{[\![} G{]\!]}$. Firstly, we note that the deformation condition at $\ell \mid N$ is exactly that $\tilde \rho(I_\ell) \cong \rho(I_\ell)$ in cases $B$ and $C$. Thus these deformation conditions amount to factoring through $G_{\mathbb{Q},S} \twoheadrightarrow G$. The remaining condition is the finite-flat condition on $\tilde \rho\vert_{G_p}$, which is satisfied exactly when $\tilde \rho$ factors through $\mathbb{F}{[\![} G{]\!]} \twoheadrightarrow E_\cD$. Now we may apply Theorems \ref{thm: main A-inf} and \ref{thm: conds body}. Because $\rho$ is irreducible, one has $r=1$ and $S^{\cD,\mathrm{GMA}}_\rho \cong R_\cD$ represents ${\mathrm{Rep}}_\rho$. Because $R_\cD$ is free of finite rank over $W(\mathbb{F})$ by the $R_\cD \buildrel\sim\over\ra \mathbb{T}$ theorem of \cite[Thm.\ 3.3]{wiles1995}, the $\mathbb{F}$-dimension of $R_\cD/p R_\cD$ is equal to this rank. This $\mathbb{F}$-dimension can be read off from the presentation for $R_\cD/pR_\cD$ furnished by Theorems \ref{thm: main A-inf} and \ref{thm: conds body}. \end{proof} Using \cite[(1.5), pg.\ 460]{wiles1995}, we find that $H^1(C_\cD)$ is canonically isomorphic to the $p$-torsion of the ``adjoint Selmer group'' $H^1_\cD(\mathbb{Q}_S, \mathrm{End}_\mathbb{F}(\rho))$ given there. The $A_\infty$-products or Massey products $m_n : H^1(C_\cD)^{\otimes n} \rightarrow H^2(C_\cD)$ are obstruction classes in the category of finite-flat left $\mathbb{F}[G]$-modules. \begin{eg} \label{eg: 1-dim H1} For instance, if $\dim_\mathbb{F} H^1(C_\cD) = 1$, we know that the rank is at least 2, i.e.\ there is some non-trivial congruence between modular forms. Then $R_\cD/pR_\cD \simeq \mathbb{F}[\epsilon_n]$ for some $n \geq 1$. The rank of $\mathbb{T}$ is then $n+1$. Let $a \in H^1(C_\cD)$ be a basis. Then $n$ is the greatest $i \geq 1$ such that $m_i(b^{\otimes i}) = 0$ in $H^2(C_\cD)$. This cohomology class also can be calculated as a Massey power (Definition \ref{defn: Massey powers}). \end{eg} \subsection{Ordinary non-Eisenstein Hecke algebras} \label{subsec: ord Wiles} We now assume that $\rho$ is as in \S\ref{subsec: flat Wiles}, with the exception that we now move to case (i) on [pg.\ 456, \textit{loc.\ cit.}], the ordinary case. We form the maximal quotient $G'$ of $G_{\mathbb{Q},S}$ such that $\ker(\rho\vert_{I_\ell})$ for all primes $\ell \mid N$, just as in \S\ref{subsec: flat Wiles}. Next, we take the universal Cayley--Hamilton quotient $E$ of $\mathbb{F}[G']$ over $\rho$, and then take its ordinary Cayley--Hamilton quotient $E_{\cD}$ of \cite[\S5]{WWE1} (see Example \ref{eg: ordinary}). This matches the ordinary condition (i-b) on [pg.\ 457, \textit{loc.\ cit.}]. Let $E_{\cD'}$ be the further Cayley--Hamilton quotient with some fixed determinant valued in $W(\mathbb{F})$; denote the corresponding deformation ring by $R_{\cD'}$. This is essentially the ``Selmer'' deformation condition of (i-a) on [pg.\ 456, \textit{loc.\ cit.}]. Note also that our use of $\cD$ and $\cD'$ matches that of Wiles [bottom of pg.\ 458, \textit{loc.\ cit.}]. Under a mild condition, we claim that we can express condition $\cD'$ (modulo $p$) as a stable condition in a somewhat limited sense (which we discuss at the end of this section). We thank Shaunak Deo for conversations giving rise to this idea. The mild condition is the requirement that $\chi_1 \chi_2^{-1}$ has order at least 3, where $\chi_1, \chi_2$ are the Jordan-H\"older factors of $\rho\vert_{G_p}$. The stable condition is given as follows. Let $\chi_{i,A} : G_p \rightarrow \mathbb{F}{[\![} t_i{]\!]}$ be any unramified deformations of $\chi_i$ to $A \in \cC_\mathbb{F}$, and let $e_A$ be any extension of $\chi_{2,A}$ by $\chi_{1,A}$. Let $H \subset G_p$ be the intersection of the kernels of the $G_p$-action on all such extensions. Then we observe that a deformation of $\rho\vert_{G_p}$ is ordinary if and only if $H$ is in its kernel. Thus we may express the ``ordinary with fixed determinant'' condition by considering deformations of $\rho$ that kill $H$: let $G$ be the maximal quotient of $G'$ in which $H$ vanishes. Let $E_{\cD'} := \mathbb{Z}_p[G]^\wedge_\rho$. In this setting, $R_{\cD} \cong \mathbb{T}$ is a ``big'' ordinary Hecke algebra, originally constructed by Hida. This is a finite rank free $\Lambda$-module, where $\Lambda \simeq W(\mathbb{F}){[\![} t{]\!]}$ and $\Lambda \rightarrow \mathbb{T}$ is the weight map induced by the inclusion of the Hecke algebra of diamond operators. Equivalently, $\Lambda \rightarrow \mathbb{T}$ parameterizes the determinant of the Galois representation valued in $\mathbb{T}$, so $R_{\cD'} \cong R_{\cD} \otimes_\Lambda W(\mathbb{F})$ for a map $\Lambda \rightarrow W(\mathbb{F})$ explained in [pg.\ 459, \textit{loc.\ cit.}]. Thus the $\Lambda$-rank of $R_{\cD} \cong \mathbb{T}$ is equal to the $\mathbb{F}$-dimension of $R_{\cD'}/pR_{\cD'}$. \begin{prop} The Hochschild cochain complex $C_{\cD} := \cC^\bullet(E_{\cD}, \mathrm{End}_\mathbb{F}(\rho))$ calculates the deformation ring $R_{\cD}/pR_\cD$ via Theorem \ref{thm: conds body}. Similarly, the Hochschild cochain complex $C_{\cD'} := \cC^\bullet(E_{\cD'}, \mathrm{End}_\mathbb{F}(\rho))$ calculates the deformation ring $R_{\cD'}/pR_{\cD'}$. In particular, the $A_\infty$-products in $H^\bullet(C_{\cD'})$ determine the $\Lambda$-rank of $R_\cD \cong \mathbb{T}$. \end{prop} As in \S\ref{subsec: flat Wiles}, one can check that $H^1(C_\cD)$ is canonically isomorphic to the reduction modulo $p$ of the ordinary $H^1_\cD(\mathbb{Q}_S, \mathrm{End}_\mathbb{F}(\rho))$ of \cite[(1.5), pg.\ 460]{wiles1995}. While $H^2(C_\cD)$ functions correctly in deformation-theoretic computations, the category $\mathcal{E}$ for which it computes an $\mathrm{Ext}^2_\mathcal{E}(\rho,\rho)$ is limited, in the sense of Remark \ref{rem: Ext2 discrepancy}. Indeed, $\mathcal{E}$ is the category of $E_\cD$-actions on finite dimensional $\mathbb{F}$-vector spaces. \subsection{Ordinary residually Eisenstein Hecke algebras} \label{subsec: ord Eis 1} We follow \cite[\S2]{WWE1} (and the references therein) for the setting of ordinary modular forms we work with, and the assumptions of the main theorems. We work in the case $S = \{p,\infty\}$, i.e.\ we work with ordinary modular forms of level 1, for simplicity. It is possible to work with more general level, but we leave this out so that the assumptions of the main theorems reduce to the Kummer--Vandiver conjecture. Let $\omega$ denote the modulo $p$ cyclotomic character of $G_{\mathbb{Q},S}$, and let $\kappa$ denote its $p$-adic cyclotomic character. The relevant residual semi-simplification is $\rho \cong \omega^{k-1} \oplus 1$ over $\mathbb{F}_p$, where $2 \leq a \leq p-3$ is even. Ribet \cite{ribet1976} proved that there is a non-trivial $\omega^{1-k}$-part $A[\omega^{1-k}]$ of the $p$-cotorsion $A$ of the class group of $\mathbb{Q}(\zeta_p)$ if and only if $p$ divides the numerator of the Bernoulli number $B_k$. Ribet's idea was to use a modular eigenform produce a $\mathbb{F}_p$-linear Galois representation of the form \[ \ttmat{\omega^{k-1}}{0}{*}{1}: G_{\mathbb{Q},S} \rightarrow \GL_2(\mathbb{F}_p), \] which is not semi-simple, but is semi-simple after restriction to $G_p$. Let $\mathbb{T}$ be the Hida Hecke algebra with residual eigensystem corresponding to $\rho$. It is a finite free $\Lambda$-algebra (in the same fashion as in \S\ref{subsec: ord Wiles}) admitting a homomorphism $\mathbb{T} \rightarrow \Lambda$ corresponding to the interpolation of ordinary $p$-stabilizations of Eisenstein series of level 1 and weight congruent to $k$ modulo $p-1$. Assume also that $p \mid B_k$, so that $\mathbb{T}$ has rank at least 2, reflecting that it parameterizes some cusp forms with a congruence with these Eisenstein series. Upon the assumption that $A[\omega^{-k}]$ is zero --- which follows from Vandiver's conjecture, as $-k$ is even --- there is proved in \cite[Thm.\ 4.2.8]{WWE2} an isomorphism \[ R_D^\ord \buildrel\sim\over\ra \mathbb{T}, \] where $R_D^\ord$ is an ordinary pseudodeformation ring for $D := \psi(\rho)$, as in \S\ref{subsec: stable or CH}. Because $\omega^{k-1}$ is not of order 2, the ``ordinary with fixed determinant $\kappa^{k-1}$'' condition can be expressed as a stable condition, just as in \S\ref{subsec: ord Wiles}. We let $\cD$ be this deformation condition, in equal characteristic $p$. Under this assumption, the $\mathrm{Ext}^1$-group for the Cayley--Hamilton condition $\cD$ (writing $\chi = \omega^{k-1}$ for convenience) is \[ \mathrm{Ext}^1_{E_\cD}(\rho,\rho) \cong \ttmat{\mathrm{Ext}^1_{E_\cD}(\chi,\chi)} {\mathrm{Ext}^1_{E_\cD}(1,\chi)} {\mathrm{Ext}^1_{E_\cD}(\chi,1)} {\mathrm{Ext}^1_{E_\cD}(1,1)}, \] and one can compute that \begin{align*} \mathrm{Ext}^1_{E_\cD}(\chi,\chi) \cong \mathrm{Ext}^1_{\mathbb{F}[G_{\mathbb{Q},S}]}(\chi,\chi)^{I_p\text{-triv}} \cong H^1(G_{\mathbb{Q},S}, \mathbb{F}_p)^{I_p\text{-triv}} \cong 0,&\\ \mathrm{Ext}^1_{E_\cD}(1,\chi) \cong \mathrm{Ext}^1_{\mathbb{F}[G_{\mathbb{Q},S}]}(1,\chi) \cong H^1(G_{\mathbb{Q},S}, \chi),& \quad \dim_{\mathbb{F}_p} = t\\ \mathrm{Ext}^1_{E_\cD}(\chi,1) \cong \mathrm{Ext}^1_{\mathbb{F}[G_{\mathbb{Q},S}]}(\chi,1)^{G_p\text{-triv}} \cong H^1_{(p)}(G_{\mathbb{Q},S}, \chi^{-1}) \cong A[\chi^{-1}]^*,& \quad \dim_{\mathbb{F}_p} = s\\ \mathrm{Ext}^1_{E_\cD}(1,1) \cong \mathrm{Ext}^1_{\mathbb{F}[G_{\mathbb{Q},S}]}(1,1)^{I_p\text{-triv}} \cong 0.& \end{align*} Here $H^i_{(p)}$ denotes cohomology supported at $p$, which is the same as cohomology with compact support since $S = \{p,\infty\}$; see e.g.\ \cite[\S6.2]{WWE1} and the references there. We have assumed that $p \mid B_k$, so $A[\chi^{-1}] \neq 0$, i.e.\ $s \geq 1$. The assumption $A[\omega^{-k}] = 0$ implies that $t=1$. Writing $\{b\}$, $\{c_1, \dotsc, c_s\}$ for a $\mathbb{F}_p$-basis for the $\mathbb{F}_p$-duals of the suspensions of the two non-zero $\mathbb{F}_p$-vector spaces, we find that the auxiliary ring $R^{1,\cD}_D$ has the presentation \[ R^{1,\cD}_D \cong \mathbb{F}_p{[\![} bc_1, \dotsc, bc_s{]\!]}. \] In particular, this ring is formally smooth and all obstructions to the pseudodeformations parameterized by $R_D^\cD$ are representation-theoretic; i.e.\ $\alpha = 0$ in Corollary \ref{cor: obstruction}. Using an injection from $\mathrm{Ext}^2_{E_\cD}(\rho,\rho)$ into appropriate cohomology groups over $\mathbb{Q}$, we also find that $\mathrm{Ext}^2_{E_\cD}(\rho,\rho)$ is non-trivial only in its $\mathrm{Ext}^2_{E_\cD}(\chi,1)$-term, namely, \begin{equation} \label{eg: ord Ext2} \mathrm{Ext}^2_{E_\cD}(\chi,1) \hookrightarrow H^2_{(p)}(G_{\mathbb{Q},S}, \chi^{-1}) \cong H^1_{(p)}(G_{\mathbb{Q},S}, \chi^{-1}) \cong A[\chi^{-1}]^*, \end{equation} where the injection may not (a priori) be an isomorphism for the reasons given in Remark \ref{rem: Ext2 discrepancy}. Letting $\gamma_1, \dotsc, \gamma_s$ be a basis for $\Sigma H^2_{(p)}(G_{\mathbb{Q},S}, \chi^{-1})^*$, Theorem \ref{thm: main pseudo} yields a presentation \[ R_D^\cD \cong \frac{\mathbb{F}_p{[\![} a, bc_1, \dotsc, bc_s{]\!]}} {(m^*(\gamma_1)b, \dotsc, m^*(\gamma_s)b)}. \] Due to the isomorphism $R_D^\ord \buildrel\sim\over\ra \mathbb{T}$ and the fact that $R_D^\cD \cong R_D^\ord/\mathfrak{m}_\Lambda R_D^\ord$, we know that $R_D^\cD$ has Krull dimension 0, and that its $\mathbb{F}_p$-dimension is the $\Lambda$-rank of $\mathbb{T}$. In particular, we know that the injection of \eqref{eg: ord Ext2} is an isomorphism. It is expected that $s=1$; this is a consequence of the assumption $A[\omega^k] = 0$, which follows from the Kummer--Vandiver conjecture but is different than our running assumption that $A[\omega^{-k}] = 0$. In this case, the presentation for $R_D^\cD$ is of the form $\mathbb{F}_p[\epsilon_n]$. Here $n$ may be computed as follows. Let $c = c_1$. \begin{prop} The Hochschild cochain complex $C_\cD := \cC^\bullet(E_\cD, \mathrm{End}_\mathbb{F}(\rho))$ calculates the deformation ring $R_\cD$ via Theorem \ref{thm: conds body}. Assume \begin{itemize}[leftmargin=2em] \item Vandiver's conjecture, and \item assume that $p \mid B_k$, so that the $\Lambda$-rank of $\mathbb{T}$ is at least 2 and $s \geq 1$. \end{itemize} Then we may read off the presentation for $R_\cD$ that the $\Lambda$-rank of $\mathbb{T}$ is equal to $n+1$ and $\mathbb{T}/\mathfrak{m}_\Lambda \mathbb{T} \cong \mathbb{F}[\epsilon_n]$, where $n \in \mathbb{Z}_{\geq 2}$ is the greatest such that \[ m_{2n-3}(c \otimes b \otimes c \otimes \dotsm \otimes b \otimes c) = 0. \] \end{prop} \begin{rem} As a complement to the discussion above, we see that a failure of the Kummer--Vandiver conjecture in the form of the existence of $p$ and $k$ such that \begin{itemize}[leftmargin=2em] \item $t > 1$, when $s \geq 1$ as well; or \item $s > 1$ \end{itemize} is detectable by the tangent dimension of $R_D^\cD$ being greater than 1. See \cite[\S1.2]{wake2} for a discussion of known relationships between the Kummer--Vandiver conjecture (and its consequences) and the Gorenstein property of Hecke algebras. It is interesting to compare these, because we might guess that $R^\ord_D \buildrel\sim\over\ra \mathbb{T}$ even when the hypothesis $A[\omega^{-k}] = 0$ fails. \end{rem} \subsection{The finite-flat residually Eisenstein Hecke algebra of Mazur} \label{subsec: Mazur Eis} In this example of a Hecke algebra, we give an example of a Cayley--Hamilton condition that cannot be expressed as a stable condition, but nonetheless has deformation theory controllable by Massey products with defining systems chosen to match the Cayley--Hamilton condition, according to \cite{WWE3}. Let $\mathbb{T}^0$ be the Hecke algebra arising from the Hecke action on weight 2 level $\Gamma_0(N)$ cusp forms with a congruence with the Eisenstein series modulo $p$, where $N$ is a prime. Mazur proved that $\mathbb{T}^0 \neq 0$ if and only if $p$ divides the numerator of $(N-1)/12$, and asked for an expression for the rank of $\mathbb{T}^0$ \cite[\S19]{mazur1978}. In \cite[Thm.\ 1.3.1]{WWE3}, an expression is given in terms of Massey products in Galois cohomology for the $\mathbb{Z}_p$-rank of $\mathbb{T}^0$. This theorem relies on an isomorphism $R^\cC_D \buildrel\sim\over\ra \mathbb{T}$, where $\mathbb{T} \twoheadrightarrow \mathbb{T}^0$ is the cuspidal quotient of the full Eisenstein-congruent Hecke algebra $\mathbb{T}$, and $\mathrm{rank}_{\mathbb{Z}_p} \mathbb{T} - \mathrm{rank}_{\mathbb{Z}_p} \mathbb{T}^0 = 1$. We will now explain the Cayley--Hamilton condition condition $\cC$ that determines $R^\cC_D$. In this setting, the residual semi-simplification is the representation $\rho = \omega \oplus 1 : G_{\mathbb{Q},S} \rightarrow \GL_2(\mathbb{F}_p)$, where $S = \{p,N,\infty\}$. Let $D = \psi(\rho)$ be the induced pseudorepresentation. We want to study the deformations of $\rho$ and $D$ satisfying $\cC$, where $\cC$ is the combination of the following two conditions: \begin{itemize}[leftmargin=2em] \item the finite-flat condition upon restriction to $G_p$, which is a stable condition (in the sense of Definition \ref{defn: stable}), and \item the Cayley--Hamilton condition that the restriction to $G_N$ induces a trivial pseudodeformation on $I_N$; see \cite[Defn.\ 10.1.2]{WWE3}. We will call this condition ``pseudo-unramified at $N$.'' \end{itemize} The Massey products of \cite[Thm.\ 1.3.1]{WWE3} are valued in $H^2(G_{\mathbb{Q},S}, \mathrm{End}_{\mathbb{F}_p}(\rho))$ and are given in terms of defining systems of lifts of $\rho$ (to coefficients in $\mathbb{F}_p[\epsilon_n]$) that are chosen inductively: see \cite[\S10]{WWE3}. This is basically the same process as the inductively constructed lifts and Massey powers in \S\ref{subsec: massey express}. These defining systems are designed so that the lifts satisfy $\cC$. In contrast with the previous three examples, it does not seem that the pseudo-unramified at $N$ condition can be imposed as a stable condition on deformations of $\rho$. However, we can apply our theory to the part of condition $\cC$ that is stable, that is, the finite-flat part. This gives a formulation of the finite-flat Galois cohomology referred to in \cite[Rem.\ 10.6.3]{WWE3}. In particular, it values the Massey products in a cohomology group $H^2_\mathrm{flat}(G_{\mathbb{Q},S}, \mathrm{End}_{\mathbb{F}_p}(\rho))$ realized as the Hochschild cohomology we now define. Namely, just as in \S\ref{subsec: flat Wiles}, we let $E_\mathrm{flat}$ be the maximal quotient of $\mathbb{F}_p{[\![} G_{\mathbb{Q},S}{]\!]}$ that is finite-flat upon restriction to $G_p$. \begin{prop} \label{prop: finite-flat cohom} The cohomology groups \[ H^i_\mathrm{flat}(G_{\mathbb{Q},S}, \mathrm{End}_{\mathbb{F}_p}(\rho)) := H^i(E_\mathrm{flat}, \mathrm{End}_{\mathbb{F}_p}(\rho)). \] satisfy the desiderata of finite-flat cohomology of \cite[Rem.\ 10.6.3]{WWE3} and support an $A_\infty$-algebra structure quasi-isomorphic to the dg-algebra $C^\bullet(E_\mathrm{flat}, \mathrm{End}_{\mathbb{F}_p}(\rho))$. \end{prop} Then the Massey product computations of \cite[\S10]{WWE3} can be repeated in the cochain complex $C^\bullet(E_\mathrm{flat}, \mathrm{End}_{\mathbb{F}_p}(\rho))$. When this is done, the finite-flat condition on a deformation will be automatic, when the Massey product vanishes. This contrasts with how the arguments of \emph{loc.\ cit.}\ must arrange for an unrestricted deformation to be ``adjusted'' so that it becomes finite-flat. It remains that the pseudo-unramified-at-$N$ condition must be arranged for by hand. \begin{rem} It was possible, for the purposes of \cite{WWE3}, to make the aforementioned adjustments and forgo constructing this finite-flat cohomology because it was understood that once it was produced, there would be an injection \[ H^2_\mathrm{flat}(G_{\mathbb{Q},S}, \mathrm{End}_{\mathbb{F}_p}(\rho)) \hookrightarrow H^2(G_{\mathbb{Q},S}, \mathrm{End}_{\mathbb{F}_p}(\rho)), \] so that the vanishing of Massey products could be calculated in unrestricted Galois cohomology without additional complications. However, this does not always hold for similar deformation problems of interest, so statements like Proposition \ref{prop: finite-flat cohom} are useful for formulation computations in such settings. \end{rem} \bibliographystyle{alpha}

6494a42474b5ca6835c072e48780d84796576839

\section{Introduction} \subsection{Stuxnet and Cyber-Physical Security} The Stuxnet worm was identified in 2010 as a piece of malware that targeted a very specific Industrial Control System (ICS) -- namely, uranium enrichment infrastructure \cite{nourian18jsr,symantec11tr}. This may not have been the first cyber attack to cause physical damage to an ICS, but it was highly publicized. As such, Stuxnet brought the potential physical consequences of cyber attacks into the public eye. Stuxnet was highly sophisticated. Part of its sophistication lay in its strategy for obtaining access to its targets: it exploited four 0-day vulnerabilities, compromised two digital certificates, and propagated itself through networks and removable devices \cite{symantec11tr}. Once it reached a control system, it continued to act stealthily. Stuxnet fed fake data to the ICS to disguise malicious actions \cite{symantec11tr,howser14cp} and limited its attacks to avoid detection \cite{karnouskos11cp}. The goal of Stuxnet was not to cause catastrophic failure, which would have been easier. Rather, it exploited the physical vulnerabilities as well as the cyber vulnerabilities inherent in the ICS. Stuxnet forced analysts to consider the risk associated with these kinds of cyber attacks. If we understand risk as the product of consequence, vulnerability, and threat, we can address each of those components separately. The potential for significant consequence is clear: many industrial processes, including critical infrastructure systems (e.g., the power grid), rely on Supervisory Control and Data Acquisition (SCADA) software and ICSs. These systems are also vulnerable. Updates can be risky because they may cause previously functional systems to produce new errors \cite{karnouskos11cp}, and even if this is not the case, taking the system in question offline to perform the updates may be difficult or infeasible \cite{nourian18jsr}. There is a tradeoff between security and ease of use, and a knowledge gap between cyber security specialists and control engineers can compound this. There are two more factors that increase the vulnerability of ICSs to cyber attack. Firstly, industrial systems are often serviced by outside contractors, and the devices (computers, USB drives, etc.) used by those contractors can provide a malware vector that bypasses traditional cyber security measures such as air gaps \cite{symantec11tr}. Secondly, industry standardization also reduces uncertainty for potential attackers; complexity, heterogeneity, and uncertainty make it more difficult for attackers to design successful attacks. Most of the uncertainty regarding the risk of cyber attacks on ICSs has to do with threat. The old consensus was that these systems were too specialized to attack \cite{karnouskos11cp}. Stuxnet, for example, required a great deal of specialized knowledge about the control systems in question \cite{symantec11tr}. In the case of terrorism, for example, it is easier to build a bomb than to write code that will cause comparable physical destruction. However, Stuxnet showed that these kinds of attacks are possible for those determined to carry them out. \subsection{Hypergames} Game theory is a branch of mathematics that looks at strategic interactions between rational entities. It has seen considerable use in economic \cite{roth02jsr} and security \cite{sandler2003terrorism} applications. A fundamental premise of strategic games in game theory is that all of the players are seeing and playing the same game. This is not always true, though. Belief manipulation plays a key role in some strategic interactions. In other cases, not all player objectives may be common knowledge. This necessitates understanding more completely players' perceptions of the game(s) they are playing; one way to model this is through hypergames \cite{bennett80jsr}. Hypergames allow players to play different games and can account for differences in their perceptions of the same game without considering uncertainty probabilistically. For example, one group of players may distinguish between certain actions while another group considers those actions all to be identical. On the other hand, some players may not be aware of the existence of other players in the game (or may not be aware of all of those other players' actions). Hypergames essentially enable us to extend the concept of rationality to a bounded information situation. This, in turn, makes it possible for a given player to exploit another player's misperceptions. In analyzing the (potentially) different games that each player is playing, though, we are still able to apply game theoretic concepts and thus build on existing game theory research. We can describe a two-player game as \begin{gather} G_{A,B} = \left(\mathcal{P},\mathcal{S},\mathcal{U}\right) \end{gather} \begin{gather} \mathcal{P} = \left\{A,B\right\} \\ \mathcal{S} = \left\{S_A,S_B\right\} \\ \mathcal{U} = \left\{u_A,u_B\right\} \end{gather} \noindent where $A$ and $B$ are the players, $S_A$ and $S_B$ are those players' respective action spaces, and $u_A,u_B: S_A \times S_B \rightarrow \Re$ are their respective payoff functions, which provide a partial ordering over $S_A \times S_B$ for each player. We can describe a first level hypergame as \begin{gather} H_{A,B} \left(A,B,G_{A,B}\right) = \left\{ p\left(A,G_{A,B}\right),p\left(B,G_{A,B}\right)\right\} \end{gather} \noindent where $p\left(A,G_{A,B}\right)$ is $A$'s perception of $G_{A,B}$. The condition $p\left(A,G_{A,B}\right) \neq p\left(B,G_{A,B}\right)$ could be caused by discrepancies such as $p \left(A,\left\{A,B\right\}\right) = \left\{A\right\}$, which would indicate that $A$ is not aware of $B$'s presence. We can also describe perceptions about perceptions. For example, $p\left(AB,u_A\right)$ is $A$'s perception of $B$'s perception of $A$'s utility function. For a first level hypergame, there are misperceptions, but the players are not aware of those misperceptions: \begin{gather} p\left(A,G_{A,B}\right) \neq p\left(B,G_{A,B}\right) \\ p\left(AB,G_{A,B}\right) = p\left(A,G_{A,B}\right) \\ p\left(BA,G_{A,B}\right) = p\left(B,G_{A,B}\right) \end{gather} For a second level hypergame, at least one player is aware of the misperceptions. For example, if $A$ is aware of the misperceptions but $B$ is not, we have \begin{gather} p\left(AB,G_{A,B}\right) \neq p\left(A,G_{A,B}\right) \\ p\left(BA,G_{A,B}\right) = p\left(B,G_{A,B}\right) \end{gather} Player $B$ then plays $p\left(B,G_{A,B}\right)$ while $A$ plays the hypergame \begin{gather} H_{A,AB} \left(A,AB,G_{A,B}\right) = \left\{p\left(A,G_{A,B}\right),p\left(AB,G_{A,B}\right)\right\} \end{gather} The overall solution to a hypergame can then be calculated by correctly aggregating the equilibrium solutions to the players' perceived (hyper)games. In a first level hypergame, for example, the equilibrium solution is $\left(x_A,x_B\right)$, where $x_A$ is $A$'s equilibrium strategy for the game $p\left(A,G_{A,B}\right)$ and $x_B$ is $B$'s equilibrium strategy for the game $p\left(B,G_{A,B}\right)$. For the second level hypergame described above, $x_A$ would be $A$'s optimal strategy for $H_{A,AB}\left(A,AB,G_{A,B}\right)$, while $x_B$ would still be $B$'s equilibrium strategy for $p\left(B,G_{A,B}\right)$. These concepts extend naturally to higher level hypergames and additional players. See Kovach et al. \cite{kovach15jsr} and Gutierrez et al. \cite{gutierrez15jsr} for more details. Approaches such as reflexive control \cite{novikov14bk}, Mirage Equilibria \cite{sakovics01jsr}, and $k$-level reasoning \cite{camerer04jsr,stahl95jsr} have been applied to systems that may not have common knowledge (and thereby incorporate a kind of bounded rationality). Despite some differences in notation and nomenclature, these approaches all incorporate hierarchies of beliefs (e.g., Player 1's beliefs about Player 2's beliefs). However, the first two, along with hypergames, differ somewhat from $k$-level reasoning with respect to the accuracy of the player perceptions. In $k$-level reasoning, the focus is on the degree to which one player anticipates another. In principle, this approach does not rule out the possibility that a given player might misperceive the nature of the game (payoff structure, available actions, etc.), but in practice, this is not a key consideration. For hypergames, this \emph{is} a key consideration. The concept of a subjective game (i.e., $p\left(A,G_{A,B}\right)$) is central to hypergame analysis, and belief hierarchies exist to support that; the same is true for reflexive control and Mirage Equilibria. For example, a key hypergame result is that hypergame equilibrium solutions can be stable under misperceptions \cite{sasaki08cp}. In these cases, each player does what the other players expect -- which can happen even when the players' perceptions differ or are erroneous -- and thus there is no motivation for players to update their perceptions. This is similar to a conjectural equilibrium \cite{sakovics01jsr} in that players do not know what they do not know. In a repeated hypergame context, then, these equilibria are stable, and extending belief hierarchies to higher and higher levels would not necessarily change that. Using the formalism we employed previously, a hypergame equilibrium is stable if $p\left(A,x_B\right) = x_B$ and $p\left(B,x_A\right) = x_A$, which need not imply that $p\left(A,G_{A,B}\right) = p\left(B,G_{A,B}\right)$. Hypergames have been used to study water resource management \cite{okada85jsr}, supply chain relationships \cite{graham92jsr}, and cyber attacks \cite{house10jsr}. Some research has also looked at connecting hypergames with other branches of game theory. Kanazawa et al. \cite{kanazawa07jsr} studied an evolutionary version of hypergames. This included calculating evolutionarily stable strategies and defining hypergame replicator dynamics. Sasaki and Kijima \cite{sasaki12jsr,sasaki16jsr} showed how hypergames can be reformulated as Bayesian games (at least in some cases). In doing so, though, they identified reasons why it may be advantageous to avoid that reformulation. Firstly, hypergames can provide a simpler and more natural epistemic representation of the game's players; the treatment of unawareness, for example, can be more convincing than in the Bayesian case. Secondly, there are some hypergame solution concepts, such as stability under misperception, that do not map to the Bayesian reformulation. The topic of misperception has also led to research into how repeated hypergames can be used to improve or update perceptions \cite{sasaki08cp}. House and Cybenko used both hidden Markov models and a maximum entropy approach \cite{house10jsr}. Takahashi et al., on the other hand, used a genetic algorithm \cite{takahashi99cp}. Generally speaking, though, the hypergame literature is relatively small; Kovach et al. provided a review of the field \cite{kovach15jsr}. Moreover, all of the examples that we have seen have involved hypergames with a relatively small number of discrete choices. Solving for the equilibrium solutions, then, has involved hand calculations and/or exhaustive enumeration. \subsection{Aim and Motivation} The goal of this paper is to show how hypergames can be used in optimal control where the control system in question is subject to adversarial perturbations and to demonstrate how this analysis can apply to Stuxnet-like attacks. This research contributes to ongoing work in optimal control by showing how manipulating controller perceptions can function as an attacker strategy; the attacker actually uses the control system against itself. These analyses then highlight weaknesses in the control system -- weaknesses that are vulnerable to attack even if they might not be vulnerable to random events. This research also advances hypergame research in two ways. Firstly, it brings hypergames to bear on a new application area (i.e., optimal control) -- one rather different than the examples in previous papers. Secondly, it applies hypergame concepts to systems of significantly greater complexity than previous hypergame research has used. The examples in this paper have continuous variables, and the second example is a discrete-time optimal control problem with time-varying variables. Both problems, moreover, require using numerical optimization methods to find hypergame equilibria. Taking hypergames to this level of complexity makes the hypergame concept more viable as a tool for analyzing real systems and not just toy problems. This kind of investigation is highly relevant to addressing Stuxnet-like attacks from a control perspective. Leaving aside IT-based cyber security concerns, let us assume that an attacker has access to at least part of an ICS. Can we then characterize the kind of damage that that attacker could produce, and can we design control procedures that minimize that damage? In this paper, we focus primarily on the former but touch upon the latter; we intend to address the latter more fully in later work. ICSs provide examples of (potentially high-impact) cyber-physical systems where control provides the connection between the `cyber' and 'physical' components. The idea behind this research, then, is not to replace traditional cyber security methods but rather to recognize that control systems can be used to provide another layer of robustness to attack if those control systems are designed to do so and that the physical weaknesses accessible through cyber means can be analyzed by looking at the control model. \section{Static Problem Formulation} To demonstrate some of the concepts of this paper, we consider a static optimization problem constrained within an operating envelope, which is represented as an inequality constraint: \begin{gather} \min \limits_{u} J\left(u,\theta\right) \\ g(u,c)\leq 0 \label{operating constraint} \end{gather} \noindent where $u$ is the vector of decision variables, $\theta$ is the vector of objective function parameters, and $c$ is the vector of operating envelope parameters. Note that $g$ may be a vector of constraint equations $g_l$, $l = 1,2,\ldots$, in which case (\ref{operating constraint}) is equivalent to $g_l\left(u,c\right) \leq 0 \ \forall \ l$. \subsection{Objective Function Manipulation} Here, we will consider a situation where the attacker can manipulate the defender's observation of objective function parameters; $\hat{\theta} = \theta + \Delta \theta$, where the vector $\hat{\theta}$ denotes the quantities that the defender observes. The attacker optimization is then \begin{gather} \max \limits_{\Delta \theta} J\left(\hat{u}^*,\theta\right) \label{att start} \\ \frac{1}{2} \left\|\Delta \theta\right\|^2 \leq \delta_{\theta,max} \label{att theta cons} \\ \hat{u}^* = \argmin_{\hat{u}} \left( J\left(\hat{u},\hat{\theta}\right) : g(\hat{u},c) \leq 0 \right) \label{def opt} \end{gather} \noindent where (\ref{def opt}) describe what the attacker expects the defender's optimization to be and (\ref{att theta cons}) is a constraint on the attacker's manipulations, which is a reasonable assumption in a context of limited attack budgets or when attack detection mechanisms are present in the system. This constitutes a second level hypergame. If $A$ represents the attacker and $D$ represents the defender, we have \begin{gather} p\left(D,\theta\right) = \hat{\theta} \neq \theta = p\left(A,\theta\right) \\ p\left(D,\left\{A,D\right\}\right) = \left\{D\right\} = p\left(AD,\left\{A,D\right\}\right) \end{gather} If the defender knows of the attacker, this leads to a higher level hypergame, where \begin{equation} p\left(DAD,\left\{A,D\right\}\right) = p\left(AD,\left\{A,D\right\}\right) = \left\{D\right\} \end{equation} The defender's optimization is \begin{gather} \min \limits_{u} J\left(u,\hat{\theta} - \Delta \theta\right) \\ g(u,c) \leq 0 \end{gather} \noindent The true $\theta$ values are unknown to the defender, but the defender calculates the $\Delta \theta$ values by solving what is believed to be the attacker's problem: (\ref{att start})-(\ref{def opt}). \begin{gather} \max \limits_{\Delta \theta} J\left(\hat{u}^*,\theta\right) \\ \frac{1}{2}\left\|\Delta \theta\right\|^2 \leq \delta_{\theta,max}\\ \hat{u}^* = \argmin_{\hat{u}} \left( J\left(\hat{u},\hat{\theta}\right) : g(\hat{u},c) \leq 0 \right) \end{gather} Given that the defender only knows $\hat{\theta}$, not $\theta$, solving the attacker's problem to determine $\Delta \theta$ will require using $\theta = \hat{\theta} - \Delta \theta$. As a further extension, we consider the scenario where the attacker manipulates the defender's perceptions of $\theta$, the defender knows that the attacker is doing this, and the attacker knows that the defender is anticipating the attacker's perturbations. We refer to this as a `double-bluff' manipulation here and in the rest of the paper. This problem leads us to a multi-level optimization problem: \begin{gather} \max_{\Delta \theta} J\left(u^*,\theta\right) \\ \frac{1}{2} \left\| \Delta \theta\right\|^2 \leq \delta_{\theta,max} \\ u^* = \argmin_u \left(J\left(u,\tilde{\theta}\right) : g(u,c) \leq 0 \right) \end{gather} \noindent subject to \begin{gather} \max_{\Delta \hat{\theta}} J\left(\hat{u}^*,\tilde{\theta}\right) \\ \frac{1}{2}\left\|\Delta \hat{\theta}\right\|^2 \leq \delta_{\theta,max} \\ \hat{u}^* = \argmin_{\hat{u}} \left(J\left(\hat{u},\tilde{\theta} + \Delta \hat{\theta}\right): g(\hat{u},c) \leq 0\right) \end{gather} \noindent where $p\left(D,\theta\right) = \tilde{\theta} = \hat{\theta} - \Delta \hat{\theta}$ is the defender's estimate of the true value of $\theta$. There are many other potential combinations of misperceptions that we could also consider. Note that the defender's perceived cost (i.e., objective function value) may differ from the true cost in some cases. For the purpose of comparison, we can model the attacker manipulating the true value of $\theta$: \begin{gather} \max_{\Delta \theta} \min_u J\left(u,\theta + \Delta \theta\right) \\ \left\| \Delta \theta\right\|^2 \leq \delta_{\theta,max} \\ g\left(u,c\right)\leq 0 \end{gather} In this case, there are no misperceptions, and the situation is simply a zero-sum game, not a hypergame. \subsection{Constraint Manipulation} \label{Constraint Manipulation} The previous section involved the attacker manipulating parameters in the objective function. In this case, there is a significant difference between manipulating the true values and the defender's perceptions. If the attacker is manipulating the constraints, however, then the distinction changes. If the attacker alters the constraint to be more restrictive, then it does not matter whether the manipulation is of the real constraint or of the defender's perceptions -- both actions lead to the same result (assuming that the defender abides by the constraint), and the perceived cost is the true cost in both cases. If the attacker alters the constraint to be less restrictive, the results are less clear. If the attacker manipulates the defender perception, the control process may hit a physical limit and/or damage the system trying to reach an infeasible state. This could be modelled by having some kind of large penalty function for violations of the true constraint. Manipulating the true constraint in such a way as to relax it may be impossible if the constraint is a physical limitation of the system. For this section, we specify that the attacker can manipulate the defender's perception of parameters in the constraint ($\hat{c} = c + \Delta c$ are the quantities that the defender perceives). These perturbations are then subject to a cost function as with the objective function parameter perturbations. \begin{gather} \frac{1}{2} \left\| \Delta c\right\|^2 \leq \delta_{c,max} \end{gather} \subsubsection{Maximizing Cost} If the attacker is manipulating the defender's perceptions to maximize defender cost, this results in a series of multi-level optimization problems, corresponding to second or higher level hypergames, analogous to those described in the previous section. If the attacker is deceiving an unsuspecting defender, we have \begin{gather} \max \limits_{\Delta c} J\left(u^*,\theta\right) \\ \frac{1}{2} \left\|\Delta c\right\|^2 \leq \delta_{c,max}\\ u^* = \argmin_u \left( J\left(u,\theta\right) : g\left(u,\hat{c}\right) \leq 0 \right) \end{gather} If the defender is aware of the attack, we have \begin{gather} \min \limits_{u} J\left(u,\theta\right) \\ g\left(u,\hat{c} - \Delta c\right) \leq 0 \end{gather} \noindent subject to \begin{gather} \max \limits_{\Delta c} J\left(\hat{u}^*,\theta\right) \\ \frac{1}{2}\left\|\Delta c\right\|^2 \leq \delta_{c,max}\\ \hat{u}^* = \argmin_{\hat{u}} \left(J\left(\hat{u},\theta\right): g\left(\hat{u},\hat{c}\right) \leq 0 \right) \end{gather} In a situation analogous to that described in the previous section, the defender only knows $\hat{c}$, not $c$, so solving the attacker's problem to determine $\Delta c$ will require using $c = \hat{c} - \Delta c$. If the attacker is aware that the defender is anticipating an attack, the resulting problem is \begin{gather} \max_{\Delta c} J\left(u,\theta\right) \\ \frac{1}{2} \left\| \Delta c\right\|^2 \leq \delta_{c,max} \\ u^* = \argmin_u \left(J\left(u,\theta\right) : g(u,\tilde{c}) \leq 0 \right) \end{gather} \noindent subject to \begin{gather} \max_{\Delta \hat{c}} J\left(\hat{u}^*,\theta\right) \\ \frac{1}{2}\left\|\Delta \hat{c}\right\|^2 \leq \delta_{c,max} \\ \hat{u}^* = \argmin_{\hat{u}} \left(J\left(\hat{u},\theta \right): g(\hat{u},\tilde{c} + \Delta \hat{c}) \leq 0\right) \end{gather} \noindent where $p\left(D,c\right) = \tilde{c} = \hat{c} - \Delta c$ is the defender's estimate of the true value of $c$. \subsubsection{Breaking the System} The attacker could also try to cause the defender to deviate maximally from the operating envelope constraint in the interest of causing a catastrophic failure. We refer to this as attempting to break the system. If the attacker is deceiving an unsuspecting defender, we have \begin{gather} \max \limits_{\Delta c} \gamma^T g\left(u^*,c\right) \\ \frac{1}{2} \left\|\Delta c\right\|^2 \leq \delta_{c,max}\\ u^* = \argmin_u \left(J\left(u,\theta\right) : g\left(u,\hat{c}\right) \leq 0 \right) \end{gather} \noindent where $\gamma^T g\left(u,c\right)$ indicates a weighted sum for a vector-valued $g$. If the defender is aware of the attack, we have \begin{gather} \min \limits_{u} J\left(u,\theta\right) \\ g\left(u,\hat{c} - \Delta c\right) \leq 0 \end{gather} \noindent subject to \begin{gather} \max \limits_{\Delta c} \gamma^T g\left(\hat{u}^*,c\right) \\ \frac{1}{2}\left\|\Delta c\right\|^2 \leq \delta_{c,max}\\ \hat{u}^* = \argmin_{\hat{u}} \left( J\left(\hat{u},\theta\right) : g\left(\hat{u},\hat{c}\right) \leq 0 \right) \end{gather} If the attacker is aware that the defender is anticipating an attack, the resulting problem is \begin{gather} \max_{\Delta c} \gamma^T g\left(u^*,c\right) \\ \frac{1}{2} \left\| \Delta c\right\|^2 \leq \delta_{c,max} \\ u^* = \argmin_u \left(J\left(u,\theta\right) : g(u,\tilde{c}) \leq 0 \right) \end{gather} \noindent subject to \begin{gather} \max_{\Delta \hat{c}} \gamma^T g\left(\hat{u}^*,c\right) \\ \frac{1}{2}\left\|\Delta \hat{c}\right\|^2 \leq \delta_{c,max} \\ \hat{u}^* = \argmin_{\hat{u}} \left(J\left(\hat{u},\theta \right): g(\hat{u},\tilde{c} + \Delta \hat{c}) \leq 0\right) \end{gather} \noindent where $\tilde{\theta}$ is defined as before. There are various other possibilities in the same vein involving asymmetric information or false beliefs. \subsection{Analytical Results} \subsubsection{Objective Function Perturbations} In this section, we will show that the defender can be robust with respect to manipulated perceptions of $\theta$. Let us assume that $g\left(u,c\right)$ is convex for $c\geq0$ and \begin{gather} J\left(u,\theta\right) = \sum_k \theta_k f_k\left(u\right) \end{gather} \noindent where each $f_k\left(u\right)$ is convex. The optimization is therefore convex for $\theta \geq 0$, and the optimality conditions \begin{gather} \sum_k \frac{\partial f_k}{\partial u} \theta_k + \sum_l \lambda_l \frac{\partial g_l}{\partial u} = \theta^T \frac{\partial f}{\partial u} + \lambda^T \frac{\partial g}{\partial u} = 0 \\ 0 \leq \lambda_l \perp g_l\left(u,c\right) \leq 0 \ \forall \ l \end{gather} \noindent are both necessary and sufficient; $\lambda$ is the vector of Kuhn-Tucker multipliers. Let us also define \begin{gather} R\left(u\right) = \left\{ \left. \frac{\partial f_k}{\partial u}\right|_u : k = 1,2,\ldots, n_{\theta}\right\} \\ S\left(u\right) = \left\{ l : \lambda_l > 0\right\} \\ S'\left(u\right) = \left\{ l : g_l\left(u,c\right) = 0\right\} \\ T\left(u\right) = \left\{\left.\frac{\partial g_l}{\partial u}\right|_u : l \in S\left(u\right) \right\} \\ T'\left(u\right) = \left\{\left.\frac{\partial g_l}{\partial u}\right|_u : l \in S'\left(u\right) \right\} \end{gather} \noindent where $R$ and $T$ are sets of vectors, $S$ is a set of indices denoting the positive $\lambda_l$ values at $u$, and $S'$ is a set of indices denoting the active set at $u$. Note that $S\left(u\right) \subseteq S'\left(u\right)$, and $S\left(u\right) \neq S'\left(u\right)$ only if there are active constraints with corresponding multipliers that are zero. \begin{lemma} \label{lem 1} Assume that $u^* \in \argmin \limits_u \left(J\left(u,\theta\right): g\left(u,c\right) \leq 0 \right)$ and that $\hat{\theta} = \theta + \Delta \theta \geq 0$. If there exists $\Delta \lambda \geq -\lambda$ such that \begin{gather} \Delta \theta^T \left.\frac{\partial f}{\partial u}\right|_{u^*} + \Delta \lambda^T \left.\frac{\partial g}{\partial u}\right|_{u^*} = 0 \\ \Delta \lambda_l g_l \left(u^*,c\right) = 0 \ \forall \ l \end{gather} \noindent then $u^* \in \argmin \limits_u \left(J\left(u,\hat{\theta}\right): g\left(u,c\right) \leq 0 \right)$ and $\hat{\lambda} = \lambda + \Delta \lambda$ are the new Kuhn-Tucker multipliers. \end{lemma} \begin{proof} If \begin{gather} \theta^T \left.\frac{\partial f}{\partial u}\right|_{u^*} + \lambda^T \left.\frac{\partial g}{\partial u}\right|_{u^*} = 0 \\ \Delta \theta^T \left.\frac{\partial f}{\partial u}\right|_{u^*} + \Delta \lambda^T \left.\frac{\partial g}{\partial u}\right|_{u^*} = 0 \end{gather} \noindent then for $\hat{\theta} = \theta + \Delta \theta$ \begin{gather} \left(\hat{\theta}^T - \Delta \theta^T \right) \left.\frac{\partial f}{\partial u}\right|_{u^*} + \left(\lambda^T - \Delta \lambda^T + \Delta \lambda^T \right) \left.\frac{\partial g}{\partial u}\right|_{u^*} = 0 \\ \hat{\theta}^T \left.\frac{\partial f}{\partial u}\right|_{u^*} + \left(\lambda^T + \Delta \lambda^T \right) \left.\frac{\partial g}{\partial u}\right|_{u^*} = 0 \end{gather} Furthermore, since $\Delta \lambda \geq - \lambda$ and $\Delta \lambda_l g_l \left(u^*,c\right) = 0 \ \forall \ l$, \begin{gather} \hat{\theta}^T \left.\frac{\partial f}{\partial u}\right|_{u^*} + \hat{\lambda}^T \left.\frac{\partial g}{\partial u}\right|_{u^*} = 0 \label{opt lem 1} \\ 0 \leq \hat{\lambda}_l \perp g_l\left(u^*,c\right) \leq 0 \ \forall \ l \label{opt lem 2} \end{gather} \noindent where $\hat{\lambda} = \lambda + \Delta \lambda$. Since $\hat{\theta}\geq 0$ and $J\left(u,\hat{\theta}\right)$ and $g\left(u,c\right)$ are convex, the optimization \begin{gather} \min_u J\left(u,\hat{\theta}\right) \\ g\left(u,c\right) \leq 0 \end{gather} \noindent is convex, and the optimality conditions (\ref{opt lem 1})-(\ref{opt lem 2}) are necessary and sufficient. $u^*$ satisfies these conditions, so $u^* \in \argmin \limits_u \left(J\left(u,\hat{\theta}\right): g\left(u,c\right) \leq 0 \right)$. \end{proof} \begin{lemma} \label{cor 2} If $\text{span} \left(R\left(u^*\right)\right) \subseteq \text{span}\left(T\left(u^*\right)\right)$, then there exists $r>0$ such that for $\left\| \Delta \theta \right\|_p \leq r$, $p>0$, $u^* \in \argmin \limits_u \left(J\left(u,\theta\right): g\left(u,c\right) \leq 0 \right)$ implies that $u^* \in \argmin \limits_u \left(J\left(u,\hat{\theta}\right): g\left(u,c\right) \leq 0 \right)$, where $\hat{\theta} = \theta + \Delta \theta$. \end{lemma} \begin{proof} Let us define the matrix $A$ such that the rows of $A$ are the vectors $\frac{\partial g_l}{\partial u} \in T\left(u^*\right)$. If $\text{span} \left(R\left(u^*\right)\right) \subseteq \text{span}\left(T\left(u^*\right)\right)$, then any linear combination of $\frac{\partial f_k}{\partial u} \in R\left(u^*\right)$ exists within $\text{span}\left(T\left(u^*\right)\right)$, which is the rowspace of $A$. This implies that for any $\Delta \theta$, there exists $\Delta \lambda$ such that \begin{gather} \sum_k \Delta \theta_k \left.\frac{\partial f_k}{\partial u}\right|_{u^*} + \sum \limits_{l \in S\left(u^*\right)} \Delta \lambda_l \left. \frac{\partial g_l}{\partial u} \right|_{u^*} = \Delta \theta^T \left. \frac{\partial f}{\partial u} \right|_{u^*} + b^T A = 0 \\ \Delta \lambda_l = 0, \ l \notin S \left(u^*\right) \end{gather} \noindent and if $A^+$ is the Moore-Penrose pseudo-inverse of $A$, then \begin{gather} b^T = - \Delta \theta^T \left. \frac{\partial f}{\partial u} \right|_{u^*} A^+ \end{gather} \noindent satisfies this exactly because $\Delta \theta^T \left. \frac{\partial f}{\partial u} \right|_{u^*}$ is in the rowspace of $A$. Define \begin{gather} \lambda_{min} = \min \limits_{l \in S\left(u^*\right)} \lambda_l \end{gather} By definition, $\lambda_{min} > 0$. If $\left\| \Delta \lambda \right\|_p \leq \lambda_{min}$, then \begin{gather} \max_l \left|\Delta \lambda_l\right| = \left\| \Delta \lambda\right\|_{\infty} \leq \left\| \Delta \lambda \right\|_p \leq \lambda_{min} , \ p >0 \end{gather} Therefore, $\left\| \Delta \lambda \right\|_p \leq \lambda_{min}$ implies that $\max \limits_l \left|\Delta \lambda_l\right| \leq \lambda_{min} $ and thus $\Delta \lambda_l \geq - \lambda_{min} \geq - \lambda_l \ \forall \ l$. If \begin{gather} \left\| \Delta \theta\right\|_p \leq \frac{\lambda_{min}}{\left\| \frac{\partial f}{\partial u} A^+\right\|_p} = r \end{gather} \noindent then \begin{gather} \left\| \Delta \lambda \right\|_p = \left\| \Delta \theta^T \frac{\partial f}{\partial u} A^+ \right\|_p \leq \left\| \Delta \theta \right\|_p \left\| \frac{\partial f}{\partial u} A^+ \right\| \leq \lambda_{min} \end{gather} By Lemma \ref{lem 1}, $u^* \in \argmin \limits_u \left(J\left(u,\hat{\theta}\right): g\left(u,c\right) \leq 0 \right)$. \end{proof} \begin{corollary} \label{cor 1} Define the matrix $A$ such that the rows of $A$ are the vectors $\frac{\partial g_l}{\partial u} \in T\left(u^*\right)$. If $A$ is invertible, then there exists $r>0$ such that for $\left\| \Delta \theta \right\|_p \leq r$, $p>0$, $u^* \in \argmin \limits_u \left(J\left(u,\theta\right): g\left(u,c\right) \leq 0 \right)$ implies that $u^* \in \argmin \limits_u \left(J\left(u,\hat{\theta}\right): g\left(u,c\right) \leq 0 \right)$, where $\hat{\theta} = \theta + \Delta \theta$. \end{corollary} \begin{proof} If $A$ is invertible, then the rows of $A$ are linearly independent and $\text{span} \left(T \left(u^*\right)\right) = R^{n_u}$, where $u \in R^{n_u}$, and thus $\text{span} \left(R\left(u^*\right)\right) \subseteq \text{span}\left(T\left(u^*\right)\right)$. This satisfies the conditions of Lemma \ref{cor 2}, and thus the same conclusions follow. \end{proof} \begin{lemma} \label{lem 2} The set $\Theta \left(u^*\right) = \left\{ \hat{\theta} : u^* \in \argmin \limits_u \left(J\left(u,\hat{\theta}\right): g\left(u,c\right) \leq 0 \right), \hat{\theta} \geq 0\right\}$ is unbounded and convex if it is non-empty. \end{lemma} \begin{proof} $J\left(u,\theta\right)$ is linear in $\theta$, so $J\left(u,c\hat{\theta}\right) = c J\left(u,\hat{\theta}\right)$ for any positive scalar $c$. Optimal solutions are invariant with respect to scalar multiples of the objective function: \begin{gather} \argmin_u \left(J\left(u,\hat{\theta}\right): g\left(u,c\right) \leq 0 \right) = \argmin_u \left(c J\left(u,\hat{\theta}\right): g\left(u,c\right) \leq 0 \right) \nonumber \\ = \argmin_u \left(J\left(u,c\hat{\theta}\right): g\left(u,c\right) \leq 0 \right) \end{gather} Therefore, for any $\hat{\theta} \in \Theta \left(u^*\right)$ and any positive scalar $c$, $c \hat{\theta} \in \Theta \left(u^*\right)$. Thus, $\Theta \left(u^*\right)$ is unbounded if it is non-empty. Furthermore, for fixed $u^*$, the optimality conditions \begin{gather} \hat{\theta}^T \left.\frac{\partial f}{\partial u}\right|_{u^*} + \hat{\lambda}^T \left.\frac{\partial g}{\partial u}\right|_{u^*} = 0 \\ \hat{\lambda}_l = 0 \ l \notin S'\left(u^*\right) \\ \hat{\lambda}_l \geq 0 \ l \in S'\left(u^*\right) \end{gather} \noindent form a set of linear inequalities in $\hat{\lambda}$ and $\hat{\theta}$; because $u^*$ is fixed, we can disregard $g\left(u^*,c\right) \geq 0$. The space of $\hat{\lambda}$ and $\hat{\theta}$ that satisfy these constraints is therefore convex. Since this space is convex, for any $\left(\hat{\theta}_1,\hat{\lambda}_1\right)$ and $\left(\hat{\theta}_2,\hat{\lambda}_2\right)$ in this space \begin{gather} \left(\alpha\hat{\theta}_1 + \left(1-\alpha\right) \hat{\theta}_2,\alpha\hat{\lambda}_1 + \left(1-\alpha\right) \hat{\lambda}_2\right) , \ \alpha \in \left[0,1\right] \end{gather} \noindent remains in $\Theta \left(u^*\right)$. Thus for any $\hat{\theta}_1, \hat{\theta_2} \in \Theta \left(u^*\right)$, $\left(\alpha\hat{\theta}_1 + \left(1-\alpha\right) \hat{\theta}_2\right) \in \Theta \left(u^*\right)$, so $\Theta \left(u^*\right)$ is convex. \end{proof} \begin{theorem} If $\text{span} \left(R\left(u^*\right)\right) \subseteq \text{span}\left(T\left(u^*\right)\right)$ and $u^* \in \argmin_u \left(J\left(u,\theta\right): g\left(u,c\right) \leq 0 \right)$, there exists a convex, unbounded set of $\Delta \theta$ such that $u^* \in \argmin \limits_u \left(J\left(u,\theta + \Delta \theta \right): g\left(u,c\right) \leq 0 \right)$. \end{theorem} \begin{proof} By Lemma \ref{cor 2}, if $\text{span} \left(R\left(u^*\right)\right) \subseteq \text{span}\left(T\left(u^*\right)\right)$ and $u^* \in \argmin_u \left(J\left(u,\theta\right): g\left(u,c\right) \leq 0 \right)$, then there exists $r>0$ such that for $\left\| \Delta \theta \right\|_p \leq r$, $p>0$, $u^* \in \argmin \limits_u \left(J\left(u,\theta + \Delta \theta\right): g\left(u,c\right) \leq 0 \right)$. Therefore, the set \begin{gather} \Theta \left(u^*\right) = \left\{ \hat{\theta} : u^* \in \argmin \limits_u \left(J\left(u,\hat{\theta}\right): g\left(u,c\right) \leq 0 \right), \hat{\theta} \geq 0\right\} \end{gather} \noindent is non-empty. By Lemma \ref{lem 2} if $\Theta \left(u^*\right)$ is non-empty, it is unbounded and convex. \end{proof} \begin{lemma} If $\text{span} \left(R\left(u^*\right)\right) \nsubseteq \text{span} \left(T'\left(u^*\right)\right)$ for $u^* \in \argmin \limits_u \left(J\left(u,\theta\right): g\left(u,c\right) \leq 0 \right)$, then for any $\epsilon > 0$, there exists $\Delta \theta$ such that $\left\| \Delta \theta\right\| < \epsilon$ and $u^* \notin \argmin \limits_u \left(J\left(u,\hat{\theta}\right): g\left(u,c\right) \leq 0 \right)$. \end{lemma} \begin{proof} Assume that for sufficiently small $\epsilon>0$, there is no $\Delta \theta$ such that $0 < \left\| \Delta \theta\right\| < \epsilon$ and $u^* \notin \argmin \limits_u \left(J\left(u,\hat{\theta}\right): g\left(u,c\right) \leq 0 \right)$. Then for sufficiently small $\Delta \theta$, there exists $\Delta \lambda$ such that \begin{gather} \Delta \theta^T \left.\frac{\partial f}{\partial u}\right|_{u^*} + \Delta \lambda^T \left.\frac{\partial g}{\partial u}\right|_{u^*} = 0 \\ \Delta \lambda_l \geq - \lambda_l \ l \in S'\left(u^*\right) \\ \Delta \lambda_l = 0 \ l \notin S'\left(u^*\right) \end{gather} Since $u^*$ is fixed, the active set cannot change. Let us define the matrix $A$ such that the rows of $A$ are the vectors $\frac{\partial g_l}{\partial u}\in T'\left(u^*\right)$ and define the vector $b$ such that the elements of $b$ are $\Delta \lambda_l, \ l \in S'\left(u^*\right)$. Then \begin{gather} \Delta \lambda^T \left.\frac{\partial g}{\partial u}\right|_{u^*} = b^T A \\ \Delta \theta^T \left.\frac{\partial f}{\partial u}\right|_{u^*} + b^T A = 0 \end{gather} If $\text{span} \left(R\left(u^*\right)\right) \nsubseteq \text{span} \left(T\left(u^*\right)\right), \ l \in S'\left(u^*\right)$, then there exists $\Delta \theta_0$ such that $\Delta \theta_0^T \frac{\partial f}{\partial u} \notin \text{span} \left(T\left(u^*\right)\right)$ and therefore \begin{gather} \Delta \theta_0^T \left.\frac{\partial f}{\partial u}\right|_{u^*} + b^T A \neq 0 \ \forall \ b \end{gather} Moreover, for any such $\Delta \theta_0$, there exists $\Delta \theta = c \Delta \theta_0$ such that for any $c > 0$ \begin{gather} c \Delta \theta_0^T \left.\frac{\partial f}{\partial u}\right|_{u^*} + b^T A \neq 0 \ \forall \ b \end{gather} Since $\left\| c \Delta \theta_0\right\| = c \left\| \Delta \theta_0\right\|$, for any $\epsilon > 0$, there exists $\Delta \theta = \frac{\epsilon}{\left\| \Delta \theta_0\right\|} \Delta \theta_0$ such that \begin{gather} \Delta \theta^T \left.\frac{\partial f}{\partial u}\right|_{u^*} + b^T A \neq 0 \ \forall \ b \end{gather} Because the optimality conditions are necessary and sufficient, and because these conditions cannot be satisfied, $u^* \notin \argmin \limits_u \left(J\left(u,\theta + \Delta \theta \right): g\left(u,c\right) \leq 0 \right)$, and thus the lemma is proved by contradiction. \end{proof} If small $\Delta \theta$ values change the value of $u^*$ but not the active set, it is possible to calculate the $\frac{\partial u}{\partial \Delta \theta}$ for the optimal solution by differentiating the optimality conditions. This provides us with a linear system that we can solve to calculate $\frac{\partial u}{\partial \Delta \theta}$, and $u^*\left(\Delta \theta\right)$ will be smooth and well-defined as long as the active set does not change. We can therefore compare this kind of system with one that is impervious to these small changes. For such a system, the measure of the `safe' range is conservative, but outside of it, continuous changes in $\hat{\theta}$ could result in discrete jumps in $u^*$ as the active set changes. Furthermore, if $J\left(u,\theta\right)$ is nonlinear in $\theta$ but still convex for all $\theta \geq 0$, then it may possible to produce similar proofs for this case, but this would require further assumptions regarding the dependence of $J$ on $\theta$. \subsubsection{Constraint Function Manipulations} Unfortunately, manipulations of $c$ are not subject to the same kinds of robustness that manipulations of $\theta$ are. This is essentially a consequence of the discussion at the beginning of Section \ref{Constraint Manipulation}: manipulating the defender's perception of the constraints produces the same change in the decision variables as changing the true constraints would as long as the defender abides by the perceived constraints. For example, \begin{gather} g_l \left(u,c\right) = 0, \ l \in S\left(u\right) \\ \sum_i \frac{\partial g_l}{\partial u_i} \frac{\partial u_i}{\partial c_j} + \frac{\partial g_l}{\partial c_j} = 0 \end{gather} Therefore, if $\frac{\partial g_l}{\partial c_j} \neq 0$, then $\frac{\partial u_i}{\partial c_j} \neq 0$. \subsection{Test Problem} As a demonstration, we consider minimizing power consumption for a fan in an HVAC system. A problem like this could form a component in a larger HVAC system, possibly as a subsystem subject to repeated optimization under changing parameter values. The baseline defender optimization problem is \begin{gather} \min \limits_{m,p} \theta_1 m + \theta_2 m^2 + \theta_3 p \\ \frac{1}{2} \left[ \left(m-c_m\right)^2 + \left(p-c_p\right)^2 - c_r^2\right] \leq 0 \label{static cons} \end{gather} \noindent where $m$ is the mass flow rate, $p$ is the static pressure, the $\theta$ values are power consumption parameters for the fan, and $c_m$, $c_p$, and $c_r$ are parameters defining the operating envelope. The attacker can introduce perturbations $\Delta \theta_i$ such that $\hat{\theta}_i = \theta_i + \Delta \theta_i$ and $\frac{1}{2} \left\| \Delta \theta\right\|^2_2 \leq \delta_{\theta,max}$ or perturbations $\Delta c_m,\Delta c_p,\Delta c_r$ such that $\hat{c}_m = c_m + \Delta c_m$, $\hat{c}_p = c_p + \Delta c_p$, $\hat{c}_r = c_r - \Delta c_r$, and $\frac{1}{2} \left\| \Delta c \right\|^2_2 \leq \delta_{c,max}$. Note the negative sign in $\hat{c}_r$. This deviates slightly from our convention above, but it also helps to simplify later calculations in some ways, and it does not ultimately change the results. In our computations in the rest of the paper, we use $\theta_1 = \theta_2 = 1$, $\theta_3 = 2$, $c_m = c_p = 5$, and $c_r^2 = 10$. The $\frac{1}{2}$ constant in (\ref{static cons}) does not change the mathematical properties of the optimization, but it, too, simplifies some of the calculations used later in this paper; see Appendix \ref{Fan Description} for these calculations. \section{Dynamic Optimization} \subsection{Model Formulation} We now bring hypergames to bear on a Model Predictive Control (MPC) problem, where the control objective is to minimize a cost function subject to state dynamics constraints and operational constraints over a time horizon of length $\tau$: \begin{gather} \min_{u^t} \sum \limits_{t=1}^{\tau} J(u^t,x^t,\theta) \label{dyn generic start}\\ x^t = f(x^{t-1},u^t,\alpha^t,\beta) \label{dynamics} \\ x^{\tau} - x^0 = 0 \\ g\left(x^t,u^t,\alpha^t,\beta\right)\leq 0 \label{dyn generic end} \end{gather} \noindent where $u^t$ are the control decision variables, $x^t$ are the states of the system, $\alpha^t$ are the system disturbances, and $\beta$ are the model parameters. We assume that $\beta$ and $\alpha^t$ can be affected by adversarial perturbations. The attacker can either perturb the defender's perception of parameters $\beta$ to maximize cost (`Static Attack') or perturb the defender's perception of $\alpha^t$ to maximize cost ('Dynamic Attack'). The perturbations denoted $\Delta \beta$, and $\Delta \alpha^t$ are bounded by constraints, normalized as appropriate if they have different orders of magnitude; such constraints are then with respect to \textit{relative} perturbations on those parameters. \begin{gather} \frac{\Delta \beta}{\beta} \equiv \left[ \frac{\Delta \beta_1}{\beta_1} \ \frac{\Delta \beta_1}{\beta_1} \ \ldots \right]^T \\ \frac{1}{2} \left\|\frac{\Delta \beta}{\beta}\right\|^2 \leq \delta_{\beta,max} \\ \frac{1}{2} \sum \limits_{t=1}^{\tau} \|\Delta \alpha^t\|_2^2 \leq \delta_{\alpha,max} \end{gather} The static attack problem is \begin{gather} \max_{\Delta \beta} \sum \limits_{t=1}^{\tau} J(u^t,x^t,\theta) \\ \frac{1}{2} \left\|\frac{\Delta \beta}{\beta}\right\|^2 \leq \delta_{\beta,max} \\ x^t = f(x^{t-1},u^t,\alpha^t,\beta) \\ \hat{x}^0 = x^0 \end{gather} \noindent subject to \begin{gather} \min_{u^t} \sum \limits_{t=1}^{\tau} J(u^t,\hat{x}^t,\theta) \\ \hat{x}^t = f(\hat{x}^{t-1},u^t,\alpha^t,\hat{\beta}) \\ \hat{x}^{\tau} - \hat{x}^0 = 0 \\ g\left(\hat{x}^t,u^t,\alpha^t,\hat{\beta}\right)\leq 0 \end{gather} This is a second level hypergame where $p\left(D,\beta\right) = \hat{\beta} \neq \beta$. The defender optimization is with respect to perceived values, not real values; the attacker perturbations mean that $p\left(D,x^t\right) = \hat{x}^t \neq x^t$ even though the attacker does not directly manipulate the state variables. The dynamic attack problem is \begin{gather} \max_{\Delta \alpha^t} \sum \limits_{t=1}^{\tau} J(u^t,x^t,\theta) \\ \frac{1}{2} \sum_t \left\|\Delta \alpha^t\right\|_2^2 \leq \delta_{\alpha,max} \\ x^t = f(x^{t-1},u^t,\alpha^t,\beta) \\ \hat{x}^0 = x^0 \end{gather} \noindent subject to \begin{gather} \min_{u^t} \sum \limits_{t=1}^{\tau} J(u^t,\hat{x}^t,\theta) \\ \hat{x}^t = f(\hat{x}^{t-1},u^t,\hat{\alpha}^t,\beta) \\ \hat{x}^{\tau} - \hat{x}^0 = 0 \\ g\left(\hat{x}^t,u^t,\hat{\alpha}^t,\beta\right)\leq 0 \end{gather} This, similarly, is a second level hypergame where $p\left(D,\alpha^t\right) = \hat{\alpha}^t \neq \alpha^t$. As before, we could consider many variations on the dynamic and static attacks, but we will only look at these two scenarios here. \subsection{Analytical Results} The analytical results derived for the static optimization problem are applicable here as well. If the dynamic optimization is convex, there are analogous results for perturbations to $\theta$, and it can similarly be shown that constraint perturbations (to $\beta$ and $\alpha^t$, in this case) cannot exhibit the same kind of local robustness as objective function perturbations. \subsection{Test Problem} Our MPC test problem is a single-zone HVAC system with a fan, heater, and chiller. The objective is to minimize power consumption subject to physical constraints (e.g., the zonal temperature evolution) and operational constraints (e.g., remaining within comfort-defined temperature limits). The baseline optimal control problem for the system is \begin{gather} \min \sum \limits_{t=1}^{\tau} \left[\theta_1 m^t + \theta_2 \left(m^t\right)^2 + \nu_h c_p m^t \left(T^t_i - d^t T^t_0 - \left(1-d^t\right) T^t_n\right) \right. \nonumber \\ \left. + \nu_n c_p m^t \left(T^t_{s,n} - T^t_s\right) + \nu_c c_p m^t \left(T^t_i - T^t_s\right)\right] \\ T^t_n = \left(1-\gamma\right) T^{t-1}_n + \beta m^t \left(T^t_{s,n} - T^t_n \right) + \gamma T^t_0 + Q^t_n \\ T^{\tau}_n - T^0_n = 0 \\ m_l \leq m^t \leq m_u \\ T^t_{s,n} - T^t_s \geq 0 \\ T^l_n \leq T^t_n \leq T^u_n \\ d_l \leq d^t \leq d_u \\ T^l_{s,n} \leq T^t_{s,n} \leq T^u_{s,n} \\ T^t_i - d^t T^t_0 - \left(1-d^t\right) T^t_n \geq 0 \\ T^t_i - T^t_s \geq 0 \end{gather} \noindent where $m^t$ is the mass flow rate, $T^t_i$ is the internal duct temperature, $T^t_s$ is the temperature of the air put out by the chiller, $T^t_{s,n}$ is the temperature of the air supplied to the zone, $T^t_n$ is the temperature of the zone, and $d^t$ is the damper position. All of these are control variables. $T^t_0$ is the external temperature (set to $25^{\circ}$C in this instantiation of the model); $\beta$ and $\gamma$ are scalar parameters that capture the room thermal properties. Other quantities listed in the problem description are parameters that are not affected by any adversarial perturbations. See Appendix \ref{HVAC Description} for more details. The fan, heater, and chiller power consumption levels at each time step are \begin{gather} \theta_1 m^t + \theta_2 \left(m^t\right)^2 \\ \nu_h c_p m^t \left(T^t_i - d^t T^t_0 - \left(1-d^t\right) T^t_n\right) \nu_n c_p m^t \left(T^t_{s,n} - T^t_s\right) \\ \nu_c c_p m^t \left(T^t_i - T^t_s\right) \end{gather} \noindent respectively. In this model, the static pressure is almost constant, and thus we omit it from the fan component of the model. The static attack manipulates the defender perception of $\beta$ and $\gamma$. The attacker goal is to maximize power consumption over the entire time horizon given that the defender observes $\hat{\beta} = \beta + \Delta \beta$ and $\hat{\gamma} = \gamma + \Delta \gamma$ and the attacker is constrained by \begin{gather} \frac{1}{2} \left[\left(\frac{\Delta \beta}{\beta} \right)^2 + \left(\frac{\Delta \gamma}{\gamma}\right)^2 \right] \leq \delta_{max} \end{gather} \noindent subject to the defender optimization of the original baseline problem. Because $\beta$ and $\gamma$ are of different magnitudes, using relative perturbations, not absolute ones, avoids some potential problems. We also highlight the previously mentioned differences between the perceived and actual state variables values. For example, the true zone temperature, $T^t_n$, and the defender perception of the zone temperature, $\hat{T}^t_n$, will evolve according to the equations, respectively, \begin{gather} T^t_n = \left(1-\gamma\right) T^{t-1}_n + \beta m^t \left(T^t_{s,n} - T^t_n \right) + \gamma T^t_0 + Q^t_n \\ \hat{T}^t_n = \left(1-\hat{\gamma}\right) \hat{T}^{t-1}_n + \hat{\beta} m^t \left(T^t_{s,n} - \hat{T}^t_n \right) + \hat{\gamma} T^t_0 + Q^t_n \label{perceived thermal eqn} \end{gather} There will be a similar discrepancy between $T^t_i$ and $\hat{T}^t_i$. The dynamic attack manipulates the defender's perception of $T^t_0$ so that $\hat{T}^t_0 = T^t_0 + \Delta T^t_0$ and $\frac{1}{2} \sum \limits_t \left(\Delta T^t_0\right)^2 \leq \Delta T_{max}$. As in the static parameter manipulation case, the defender will misperceive both $T^t_n$ and $T^t_i$. The full formulations for the static and dynamic manipulation problems are provided in Appendix \ref{HVAC Description}. \section{Computational Implementation} The specific calculations to turn each hypergame problem into a tractable nonlinear program (NLP) are provided in Appendices \ref{Fan Description} and \ref{HVAC Description}. We summarize our general approach here. Each hypergame produces a multi-level optimization problem. To solve this, we write the optimality conditions of the lower level problems as complementarity conditions. In the case of the fan optimization, we can transform these complementarity conditions into equality constraints and then solve the resulting problem as an NLP. For the HVAC problem, we cannot do this, and this leaves us with a Mathematical Program with Equilibrium Constraints (MPEC) \cite{ruiz14jsr}. We can solve the MPEC as a series of NLPs by relaxing the complementarity constraints and penalizing the relaxation with a progressively increasing weight. For the work described in this paper, this was both reliable and efficient. To implement our approach, we derived the necessary optimality conditions by hand, coded up the NLPs in MATLAB \cite{MATLAB}, and solved the NLPs using \textit{fmincon}. \section{Results} \subsection{Fan Optimization} \begin{table}[htp] \caption{Objective Function Manipulation Results ($\delta_{\theta,max} = 0.1$)} \label{tab:objfun results} \centering \setlength{\tabcolsep}{0.5em} \begin{tabular}{c|cccccc} Case &$m$ &$p$ &$\Delta \theta_1$ &$\Delta \theta_2$ &$\Delta \theta_3$ &Power \\ \hline Baseline &2.06 &3.85 &- &- &- &13.97 \\ True Manipulation &2.02 &3.94 &0.150 &0.303 &0.292 &16.68 \\ Perception Manipulation &2.29 &3.38 &-0.090 &-0.411 &0.151 &14.26 (12.42) \\ Faulty Defender Anticipation&1.95 &4.16 &- &- &- &14.08 (14.71) \\ Double-Bluff Manipulation &1.89 &4.42 &0.00684 &0.259 &-0.358 &14.30 (13.76) \end{tabular} \end{table} Table \ref{tab:objfun results} shows the results for the attacker manipulation of the objective function parameters; power consumption values in parentheses indicate the power usage perceived by the defender where it differs from the actual usage. Manipulating the true $\theta_i$ values produced a notable increase in power consumption compared with the baseline. Manipulating defender perceptions, though, proved less effective. For example, when the attacker manipulated the perceptions of an unsuspecting defender (Perception Manipulation), the gap between the perceived and actual power usage was noticeable, but the actual increase in power relative to the baseline case was small. Similarly, if the defender erroneously thought that the attacker was manipulating the perceived values of $\theta_i$ (Faulty Defender Anticipation), the true power usage was almost identical to the baseline case, though the perceived power consumption was somewhat higher. When manipulating the defender's perceptions, the attacker got the defender to increase $m$ and decrease $p$ (relative to the baseline case) by decreasing the perceived value of $\theta_1$ and $\theta_2$ ($\Delta \theta_1 < 0$, $\Delta \theta_2 < 0$) and increasing the perceived value of $\theta_3$ ($\Delta \theta_3 > 0$). This approach is more beneficial for the attacker than decreasing $m$ and increasing $p$ because the objective is quadratic in $m$ but only linear in $p$. In the double-bluff situation, however, the defender expects the attacker to employ this optimal strategy, and so the attacker does the exact opposite (i.e., encourages the defender to increase $p$ and decrease $m$), which provides a slight additional benefit over the simple manipulation case. \begin{figure}[htp] \centering \includegraphics[width=0.7\textwidth]{manip_plot.png} \caption{Visualization of `Perception Manipulation' attack.} \label{fig:manip plot} \end{figure} Fig. \ref{fig:manip plot} shows the `Perception Manipulation' case and why it produces so little payoff for the attacker. There, we see how the perceived objective function contours are essentially a rotated version of the original objective function contours. That rotation, produced by changes in the relative magnitudes of the $\theta_i$ parameters, produces a perceived (i.e., false) optimum point that is noticeably different from the true optimum point. However, even a significant difference in the solution location does not necessarily translate to a large difference in the true objective function value because neither the constraint nor the objective function contours have large curvatures near the true optimum -- most of the translation between the two points is parallel to the contours of the true objective function. Manipulating constraints gave the attacker more options than manipulating the objective function parameters. As Table \ref{tab:cons results I} shows, constraint manipulation was also much more effective as an attacker strategy. For example, when the attacker attempted to maximize power consumption against a defender who did not believe an attack was underway (Power Max, Normal), the attacker was able to increase power consumption by almost 30\% compared with the baseline. Attempting to maximize the constraint violation (Break System, Normal) resulted in a significant level of violation, too. \begin{table}[htp] \caption{Constraint Manipulation Results ($\delta_{c,max} = 0.1$)} \label{tab:cons results I} \centering \setlength{\tabcolsep}{0.5em} \begin{tabular}{cc|cccc} Attacker Action &Defender Belief &$m$ &$p$ &Power &Violation\\ \hline No Attack &Normal &2.06 &3.85 &13.97 &-\\ Power Max &Normal &2.59 &4.22 &17.76 &-\\ No Attack &Power Max &1.57 &3.37 &10.79 &4.92\\ No Attack &Break System &2.59 &2.24 &17.76 &-\\ Break System&Power Max &1.17 &2.78 &8.11 &4.85\\ Power Max &Break System &3.16 &4.53 &22.21 &-\\ Break System&Normal &1.58 &3.36 &10.79 &2.20\\ Power Max (Double-Bluff)&Power Max &2.16 &3.94 &14.71 &0.406\\ Break System (Double-Bluff)&Break System &2.05 &3.87 &13.97 &0.003 \end{tabular} \end{table} \begin{table}[htp] \caption{Constraint Manipulation Results ($\delta_{c,max} = 0.1$)} \label{tab:cons results II} \centering \setlength{\tabcolsep}{0.5em} \begin{tabular}{cc|ccc} Attacker Action &Defender Belief &$\Delta c_m$ &$\Delta c_p$ &$\Delta c_r$ \\ \hline Power Max &Normal &0.301 &0.097 &0.316 \\ Break System&Power Max &-0.285 &-0.137 &-0.316 \\ Power Max &Break System &0.301 &0.097 &0.316 \\ Break System&Normal &-0.285 &-0.137 &-0.316 \\ Power Max (Double-Bluff)&Power Max &0.419 &0.157 &0.000 \\ Break System (Double-Bluff)&Break System &-0.295 &-0.113 &-0.316 \end{tabular} \end{table} In this case, there were also major consequences for wrongly anticipating an attack. Anticipating a power maximization attack when there was no attack resulted in a worse constraint violation than when the attacker was deliberately trying to break the system. Conversely, anticipating a `break system' attack when the actual attack was a `power max' attack led to an increase in power consumption of almost 60\% compared with the baseline. Note that in these false anticipations, the attacker is assuming that the defender is just playing normally (i.e., the attacker is not taking advantage of the defender's mistake). The double-bluff strategies did not provide much benefit to the attacker, though. Table \ref{tab:cons results II} also shows the perturbations used by the attacker. We can see that the attacker strategies for maximizing power consumption and breaking the system are almost exactly mirror opposites, which makes sense. The double-bluff strategies are not that much different than the regular strategies that they correspond to, though, so it is not surprising that the double-bluff approach is not very effective. Switching attack modes would be a better option if the defender is anticipating an attack, and though we did not calculate this here, it would be possible to calculate an optimal attack for one mode given that the defender is expecting the other mode. Given how the two modes produce almost exactly opposite attacker strategies, the attacker strategy would likely be quite similar to the same attack mode employed against an unsuspecting defender. In general, changes in constraint parameters may result in larger objective function changes than changes in objective function parameters for two reasons. Firstly, the changes in constraints will be multiplied by the dual variables (Lagrange or Kuhn-Tucker) associated with those constraints to produce a final change in the objective function. Secondly, changing constraint values may result in the active set at the optimum also changing, and that could produce large, nonlinear changes in the objective function. All in all, this likely makes constraint manipulation a much more attractive target for a would-be attacker than objective function manipulation. \subsection{Single-Zone HVAC Control} \begin{figure}[ht] \centering \includegraphics[width=0.6\textwidth]{baseline_results.png} \caption{Baseline temperature results.} \label{fig:baseline} \end{figure} In the baseline case, and for all of the adversarial perturbations, $m^t$ and $d^t$ were both at their lower bounds for the entire optimization. Fig. \ref{fig:baseline} shows the defender strategy in more detail for different optimization horizon lengths. There, we see that the defender essentially allows the zone to evolve without manipulation until the last time step. Because $T^t_0 > T^t_n$, this means that the zone warms over time, but because $\gamma$ is very small, this happens slowly. At the last time step, the defender then chills the zone back to the initial temperature. We can see this in the sudden drop in $T^t_s$ at the end of each time horizon; note that the optimization produces $T^t_s = T^t_{s,n}$ for each optimization. This general behaviour is seen when the attacker manipulates defender perceptions, too. The longer the optimization time horizon, the larger the drop in $T^t_s$ at the last time step. If the length of the time horizon were increased sufficiently, eventually the system would require multiple steps of cooling, because $T^t_s$ would hit its lower bound. $T^t_n$ never hit its upper bound, but if it did, this would also require additional cooling prior to the end of the optimization horizon. \begin{table}[ht] \caption{Static Parameter Manipulation Results ($\delta_{max} = 0.1$)} \label{tab:static parameter} \centering \setlength{\tabcolsep}{0.5em} \begin{tabular}{c|ccc} &5-step &10-step &20-step \\ \hline Baseline Power &14.76 &29.48 &58.77 \\ Actual Power &15.08 &30.27 &60.95 \\ Defender Perceived Power &15.00 &29.97 &59.80 \\ $\Delta \beta$ &-1.81e-3 &-1.84e-3 &-1.94e-3 \\ $\Delta \gamma$ &1.64e-5 &1.52e-5 &9.74e-6 \\ $\lambda_{mean}$ &367 &370 &383 \end{tabular} \end{table} Table \ref{tab:static parameter} shows that manipulating the defender's perception of $\beta$ and $\gamma$ resulted in small power increases, relative to the baseline, and small discrepancies between the actual and perceived power use. The perturbations themselves also change slightly as the length of the time horizon changes; there is a greater emphasis on $\Delta \beta$ as the time horizon gets longer. In this model, $\beta$ essentially measures how hard it is to change the zone temperature with the HVAC system. Setting $\Delta \beta < 0$ makes the defender think that the zone is harder to adjust than it actually is. The $\gamma$ parameter then captures the heat transfer between the zone and the outside environment. Setting $\Delta \gamma > 0$ makes the defender think that there is more heat transfer than there actually is. All of this combines to increase the amount of cooling that the defender thinks is necessary at the end. The $\Delta T$ plots in Figs. \ref{fig:percep static} and \ref{fig:percep dynamic} show this kind of behaviour: the defender thinks that the temperatures are higher than they actually are and therefore overcompensates at the end. This overcompensation leads to an increase in power use and a final $T^t_n$ value that is actually slightly lower than it should be. Next, we can look at the $\lambda_{mean}$ values given in Table \ref{tab:static parameter}. $\lambda_{mean}$ is the average of the Lagrange multipliers associated with (\ref{perceived thermal eqn}) and therefore provides a measure of how the $\Delta \beta$ and $\Delta \gamma$ perturbations get multiplied. This value increases as the time horizon lengthens, which makes sense: as the time horizon lengthens, the importance of the thermal evolution process increases. An attacker perturbing $\beta$ and $\gamma$ would want this value to be as large (positive or negative) as possible. \begin{figure}[htp] \centering \begin{subfigure}[t]{0.6\textwidth} \centering \includegraphics[width=\textwidth]{percep1_deviation.png} \caption{Static attack.} \label{fig:percep static} \end{subfigure} \\ ~\begin{subfigure}[t]{0.6\textwidth} \centering \includegraphics[width=\textwidth]{percep2_deviation.png} \caption{Dynamic attack.} \label{fig:percep dynamic} \end{subfigure} \caption{Temperature deviations, $\Delta T = \left(T_{true} - T_{perceived}\right)$.} \end{figure} \begin{table}[ht] \caption{Dynamic Attack Results ($\Delta T_{max} = 0.1n$ for $n$-step problem)} \label{tab:dynamic parameter} \centering \setlength{\tabcolsep}{0.5em} \begin{tabular}{c|ccc} &5-step &10-step &20-step \\ \hline Baseline Power &14.76 &29.48 &58.77 \\ Actual Power &16.35 &32.85 &65.68 \\ Defender Perceived Power &15.58 &31.20 &62.27 \\ $\lambda_{mean}$ &219 &218 &216 \end{tabular} \end{table} \begin{figure}[htp] \centering \includegraphics[width=0.6\textwidth]{percep2_perturb.png} \caption{Dynamic parameter manipulation temperature perturbations.} \label{fig:percep2 perturb} \end{figure} \begin{table}[ht] \caption{Power Consumption Comparisons relative to Baseline (\%)} \label{tab:comparisons} \centering \setlength{\tabcolsep}{0.5em} \begin{tabular}{c|ccc} &5-step &10-step &20-step \\ \hline Static Attack (Perceived) &1.6 &1.7 &1.8 \\ Static Attack (Actual) &2.2 &2.7 &3.7 \\ Dynamic Attack (Perceived) &5.6 &5.8 &6.0 \\ Dynamic Attack (Actual) &10.1 &11.4 &11.8 \end{tabular} \end{table} Table \ref{tab:dynamic parameter} shows that manipulating $T^t_0$ provided a much larger increase in power consumption as well as a larger difference between the perceived and actual power consumption. $\lambda_{mean}$ is also much smaller, and these phenomena are related. The static parameters could only affect the power consumption indirectly through the temperature evolution equation. $T^t_0$, however, shows up in the objective function and another constraint in addition to the temperature evolution equation, so increasing $\lambda_{mean}$ becomes less important. In this case, misperceptions of $T^t_i$ and $T^t_n$ become smaller (see Figs. \ref{fig:percep static} and \ref{fig:percep dynamic}) and less important to the attacker. Instead, the attacker uses $\Delta T^t_0 > 0$ to get the defender to increase $T^t_i$, and thus the defender ends up engaging the heater (because $T^t_i - d^t T^t_0 - \left(1-d^t\right) T^t_n > 0$ even though $\hat{T}^t_i - d^t \hat{T}^t_0 - \left(1-d^t\right) \hat{T}^t_n = 0$) as well as the chiller. The perturbations themselves follow a clear pattern, as shown in Fig. \ref{fig:percep2 perturb}. They increase very slightly over time until the last time step, at which point they drop to nearly zero. The last step is less valuable to the attacker because there are no more thermal evolution steps left in the optimization at that point. Table \ref{tab:comparisons} provides an overall summary of the power consumption results. Generally speaking, the relative payoff for the attacker increases with the length of the time horizon. The actual power consumed in the static attack scenario, relative to the baseline, is roughly proportional to the length of the time horizon, but the other three cases in Table \ref{tab:comparisons} all seem to plateau. \section{Discussion} \subsection{Stuxnet-like Attacks and Hypergames} In this paper, we showed examples of how an attacker with knowledge of the system in question could manipulate the optimization processes of that system. These problems were relatively small, but they were sufficient to show how the analysis works. Hypergames are about strategic interactions when there are misperceptions and/or information asymmetries. In this case, we were able to show how those asymmetries or misperceptions could affect system performance. For example, getting the defender to respond to a non-existent threat could actually prove to be a very effective attacker strategy. Conversely, it is possible for the defender system to have a natural robustness to perturbations (though that was not the case in these test problems). We could consider more complex interactions, and we intend to do so in future work, but that future work will need to build upon the basics outlined here. When we look at Stuxnet as a motivating example for this work, we can see that there are many similarities as well as some key differences between Stuxnet and the cases considered here. In both Stuxnet and our case studies, the attacker employed limited deviations to avoid detection; we modelled this using the concept of an attacker budget. Both also involved fake sensor signals ($\Delta T^t_0$) and manipulated calibration values ($\Delta \theta$, $\Delta c$, $\Delta \beta$, $\Delta \gamma$). Our examples each had two different kinds of attack modes, and for the fan optimization, there were two different attack objectives for one of the modes, but these all involved negatively impacting the defender's control system in some way. Finally, Stuxnet and the attacks considered in this paper all utilized deep knowledge of an automated decision-making system to determine how to perform the attack. There are two primary sets of differences between this paper's case studies and Stuxnet. Firstly, to the best of our knowledge, Stuxnet was not optimization-based, and the centrifuge control systems did not employ optimal control. As such, the decision-making processes for both the attacker and the defender were different than in our paper. Secondly, Stuxnet actually overrode the control signals and software to manipulate the centrifuges \cite{symantec11tr}, whereas our attacks only altered sensor and calibration data. If we were trying to model the Stuxnet attack itself, these discrepancies would be problematic. Given the more general nature of our investigation here, though, this is less of an issue. Moreover, the key similarities identified above are ones we believe to be relevant to a wide range of control systems that might be threatened by cyber attacks in general and APTs in particular. \subsection{Scalability Considerations} A big question in applying these techniques to real-world problems is scalability. These problems were relatively small; even the 20-step HVAC problem had only 120 variables (six per time step) in the baseline problem. How easy would it be to propagate the optimality conditions and solve the resulting MPECs for larger systems? The answer has two parts. Firstly, there is the question of the optimality conditions. If those optimality conditions are necessary but not sufficient, as in general continuous NLP problems, propagating the optimality conditions to turn the multi-level optimization into an MPEC may run into difficulties; multiple optima would be one example of this. That being said, the single-zone HVAC system presented here was a nonconvex problem, and it had no such problems. If there are more than two levels to the optimization, that can also cause difficulties, as the optimization conditions from lower levels compound. This then leads into the question of tractability. Adding the dual variables of lower level optimizations to the problem description in order to solve the system as an MPEC can greatly increase the number of variables involved; having multiple levels may exacerbate the issue. However, it is sometimes possible to simplify the optimality conditions and thereby remove some of the dual variables (as was done for the fan optimization problem). The NLP sequential relaxation of the MPEC also scales well and handles the complementarity constraints efficiently. On the whole, the scalability of this approach will depend on the problem in question and how many levels of (mis)perception are of interest. Hypergames where the individual players' games are differentiable, convex optimization problems are likely to have the greatest amount of success with this approach. Problems with known or constant active constraint sets will also generally be more amenable to the multi-level optimizations than problems with active sets that change. \subsection{Future Work} Some authors writing on Stuxnet suggest the use of heuristics to identify attacks \cite{karnouskos11cp,bencsath12jsr}. One area of future work would be to take existing research on learning in repeated hypergames \cite{takahashi99cp,gharesifard10cp} and apply it to this context. For this, we would consider the defender's ability to detect attacks as well as the attacker's behaviour when the non-detection constraint is endogenous rather than exogenous; the attacker budget imposed here would be an example of an exogenous detection constraint. Another area of interest would be the defender's decision-making more generally. Given the possibility of attack and the potential consequences (as calculated in this paper), how should a defender respond if an attack is undetectable beforehand? Hypergame results here should enable us to to evaluate and prescribe control policies more broadly. Finally, we intend to extend this work to larger, real-world systems. Working on such systems may then also involve more complicated attacker manipulations, but we anticipate being able to use the same techniques demonstrated here. \section{Conclusions} In this paper, we showed how hypergames can be extended to situations with continuous and time-varying variables. That extension allowed us to consider the effects of adversarial perturbations in an optimal control context, which can give us insights into the control aspects of a Stuxnet-like attack. Manipulating constraints can be a more effective attacker strategy than directly manipulating objective function parameters; our analytical results showed why we would expect this to be true more generally. Moreover, the attacker need not change the underlying system in any way to attack successfully -- it may be sufficient to deceive the defender controlling the system. It is possible to scale our approach up to larger systems, but the ability to do so will depend on the characteristics of the system in question, and we identified several characteristics that will make larger systems amenable to hypergame analysis. \bibliographystyle{ieeetr}

3c832733c295ef816f7ce5a26ed62d2b4cc871b0

\section{Introduction} In a very interesting recent paper \cite{Ribault}, a proposal was put forward for some of the four-point correlation functions of the percolation problem in two dimensions. This proposal was part of a more general conjecture addressing various geometrical objects involving four points in the diagrammatic formulation \cite{FK72} of the $Q$-state Potts model \cite{Potts52}. The case $Q=1$ corresponds to percolation, and the proposal in \cite{Ribault} covers such objects as the probability that two of these points belong to one cluster, and the two others to another cluster. Obtaining closed-form expressions for such objects is one of the holy grails in the field. It is a far from obvious endeavour because the conformal field theory (CFT) describing percolation (and more generally geometrical features of the $Q$-state Potts model) is not well understood: it is non-unitary, probably involves logarithms (even for $Q$ generic), and involves operators which are not degenerate, precluding the use of the differential equations approach \`a la BPZ \cite{yellowpages}. The construction in \cite{Ribault} is elegant and powerful. It starts with a seemingly reasonable hypothesis for the spectrum of operators appearing in the fusion channels for the fusion of two order operators, and determines, using a clever code, the whole set of structure constants based on our knowledge of conformal blocks \cite{Zamoblocks1,Zamoblocks2} together with the imposition of crossing symmetry. The results are then checked against Monte Carlo simulations, with, it is claimed, reasonably good agreement. Although the results in \cite{Ribault} are appealing, they are not really consistent with what is known about the Potts model CFT and, in particular, percolation. Early work \cite{DFSZ} has revealed indeed a much richer spectrum than the one postulated in \cite{Ribault}, which covers only a very tiny set of the known full operator content of the theory. Of course, it could be that by some accident, the order operator in the $Q$-state Potts model does not couple to as many fields as one would expect, at least in the scaling limit. But it could also be that something is simply missing in the work of \cite{Ribault}, despite the apparent numerical effectiveness of their proposal. To investigate this question requires a long and detailed analysis, of which we present the results here. In a nutshell, we have gathered direct, in our opinion unquestionable evidence that the spectrum of the $Q$-state Potts model is as complex as could have been feared, that many more fields appear in the OPE of order operators in the Potts model than was conjectured in \cite{Ribault}, and that the proposal in that paper, appealing as it may be, simply cannot be correct. It is, at best, a good numerical approximation to the true expressions for the four-point functions. Our paper is organised as follows. In section \ref{sec:Potts-corr} we remind the reader of basic facts and results about the $Q$-state Potts model and its geometrical formulation. Algebraic aspects---which constitute a crucial part of our approach---are discussed in section \ref{sec:latt-alg}. Section \ref{sec:four-point} summarises our method of analysis, and how we extract exponents as well as amplitudes, from lattice data. Section \ref{sec:results} discusses our results for the spectra in the intermediate channels of four-point functions. A comparison with results in \cite{Ribault} is provided in section \ref{sec:comp-ribault}. In section \ref{sec:conclusion}, we return to the issue of divergences in the amplitudes, and re-analyse briefly our results as well as those of \cite{Ribault} from the point of view of degeneracies, and, potentially, logarithmic CFTs. Since a good part of our analysis is based on extracting amplitudes from lattice data, a lot of technical aspects have to be considered both to make the program possible, and to check its validity and its limits. We have thus gathered quite a bit of material in a series of appendices. Appendix \ref{sec:appA} discusses in detail many aspects of the numerical algorithms and other techniques used to obtain our results, while appendix \ref{sec:appB} goes over a series of checks, including detailed comparisons, in particular, with known results for $Q=0,2,4$. \section{Potts model and its correlation functions} \label{sec:Potts-corr} We consider the $Q$-state Potts model \cite{Potts52} defined on a graph $G=(V,E)$ with vertices $V$ and edges $E$. There is a spin $\sigma_i = 1,2,\ldots,Q$ attached to each vertex $i \in V$ and an interaction energy $-K \delta_{\sigma_i,\sigma_j}$ attached to each edge $(ij) \in E$. The partition function (in units where the inverse temperature is absorbed into the coupling constant $K$) reads \begin{equation} Z = \sum_{\{ \sigma \}} \prod_{(ij) \in E} {\rm e}^{K \delta_{\sigma_i,\sigma_j}} \,. \end{equation} Note that this initial formulation supposes $Q$ to be a positive integer, $Q \in \mathbb{N}$. This constraint can be lifted in a rewriting of $Z$ due to Fortuin and Kasteleyn (FK) \cite{FK72}. Indeed, write ${\rm e}^{K \delta_{\sigma_i,\sigma_j}} = 1 + v \delta_{\sigma_i,\sigma_j}$ with temperature parameter $v = {\rm e}^K - 1$, expand the product $\prod_{(ij) \in E}$, and perform the sum over all spins $\{ \sigma \}$ to obtain \begin{equation} Z = \sum_{A \subseteq E} v^{|A|} Q^{k(A)} \,, \label{Z_FK} \end{equation} where the sum is over all $2^{|E|}$ subsets of $E$, and $|A|$ denotes the number of edges in the subset. Moreover, $k(A)$ denotes the number of connected components (also called FK clusters) in the subgraph $G_A = (V,A)$. In the remainder of the paper we take $G$ to be the two-dimensional square lattice. The temperature parameter will be taken at its critical value, $v_{\rm c} = \sqrt{Q}$ \cite{Potts52,Baxter73}, so that the model is conformally invariant in the continuum limit. In this latter limit, we are interested in the geometry of the infinite plane, so that boundary effects are immaterial. We shall often use the trick of transforming this into the geometry of a cylinder, via an appropriate conformal mapping (details will be given below). This cylinder geometry is convenient for imposing the lattice discretisation, which is our main tool of algebraic and numerical investigations. In that case we always take the square lattice $G$ to be axially oriented with respect to the cylinder axis, so that the row-to-row transfer matrix describes the (imaginary) time evolution of $L$ Potts spins. \subsection{Loop model} An equivalent formulation of $Z$ is given \cite{BKW76} by the loop model on the medial lattice ${\cal M}(G) = (V_{\cal M},E_{\cal M})$. The vertices $V_{\cal M}$ of ${\cal M}(G)$ are situated at the mid-points of the original edges $E$, and two vertices in $V_{\cal M}$ are connected by an edge in $E_{\cal M}$ whenever the former stand on edges in $E$ that are incident on a common vertex from $V$. In particular, when $G$ is a square lattice, ${\cal M}(G)$ is just another square lattice, tilted through an angle $\frac{\pi}{4}$ and scaled down by a factor of $\sqrt{2}$. There is a bijection between edge subsets $A \subseteq E$ and completely-packed loops on ${\cal M}(G)$. The loops are defined so that they turn around the FK clusters and their internal cycles (alternatively they separate the FK clusters from their duals). One has then \cite{BKW76} \begin{equation} Z = Q^{|V|/2} \sum_{A \subseteq E} \left( \frac{v}{n} \right)^{|A|} n^{\ell(A)} \,, \label{Z_loop} \end{equation} where $\ell(A)$ denotes the number of loops. The loop fugacity is \begin{equation} n = Q^{1/2} = \mathfrak{q} + \mathfrak{q}^{-1} \,, \label{loop-fugacity} \end{equation} where the (quantum group related) parameter $\mathfrak{q}$ will be used intensively below. The loop model will be convenient to make contact with the Temperley-Lieb (TL) algebra \cite{TL71}, which will be discussed below. Note also that on a square lattice, we have simply $\frac{v_{\rm c}}{n} = 1$, so at the critical point $Z$ depends only on $\ell(A)$. \subsection{Correlation functions} The Potts model allows for the definition of various correlation functions, depending on whether one uses its formulation in terms of Potts spins, FK clusters or TL loops.% \footnote{One can even use the spin and FK cluster formulations simultaneously to define new correlation functions \cite{VJ_spin_FK}.} The spin correlators are naturally defined in terms of the order parameter (or spin) operator \begin{equation} {\cal O}_a(\sigma_i) \equiv Q \delta_{\sigma_i,a} - 1 \,. \label{spin-op} \end{equation} More interesting and general results can however be obtained by moving to the cluster or loop formulations, in which the correlation functions acquire a geometrical content. In the same vein, the spin correlators can be analytically continued from $Q \in \mathbb{N}$ to arbitrary real values, in which case they also acquire a geometrical interpretation \cite{DJS_spin1,DJS_spin2,DelfinoViti} to which we shall return in a short while. Such generalisations to $Q \in \mathbb{R}$ are not only useful, but actually indispensable in our case, since our main objective is to study correlation functions in the generic case where $\mathfrak{q}$ is not a root of unity. Correlation functions in the loop formulation are of either electric or magnetic type, where the terminology refers to the Coulomb gas approach to CFT \cite{Nienhuis_CG,DFSZ}. Let $i_1,i_2,\ldots,i_N$ be a number of distinct marked vertices. Electric correlators are defined for $i_k \in V$ by appropriately modifying the weight of loops that contain a subset of marked vertices on their inside, and the remainder on their outside. Magnetic correlators are defined for $i_k \in V_{\cal M}$ by specifying whether given vertices belong to the same or different loops; one can also increase the set of possibilities by allowing for topological defects that insert a number of open loop strands at each marked vertex \cite{DS_watermelon}. While these electromagnetic correlation functions have been intensively studied for $N=1,2$ in a variety of contexts, we wish here to recall only one recent result. Namely, the electric $N=3$ correlation functions have been shown to be related, for generic values of $n \in [0,2]$, to the so-called DOZZ formula for the structure constants within Liouville field theory \cite{DelfinoViti,PiccoSantachiaraVitiDelfino,IkhlefJacobsenSaleur}. Our main interest here is however correlation functions defined in terms of the FK clusters. Let again $i_1, i_2,\ldots, i_N \in V$ be a number of distinct marked vertices, and let ${\cal P}$ be a partition of a set of $N$ elements. We then define \begin{equation} P_{\cal P} = \frac{1}{Z} \sum_{A \subseteq E} v^{|A|} Q^{k(A)} {\cal I}_{\cal P}(i_1,i_2,\ldots,i_N | A) \,, \label{P_corr_def} \end{equation} where $Z$ is given by (\ref{Z_FK}), and ${\cal I}_{\cal P}(i_1,i_2,\ldots,i_N | A)$ is the indicator function that, $\forall k,l \in \{1,\ldots,N\}$, vertices $i_k$ and $i_l$ belong to the same connected component in $A$ if and only if $k$ and $l$ belong to the same block of the partition ${\cal P}$. It is convenient to denote ${\cal P}$ by an ordered list of $N$ symbols ($a,b,c,\ldots$) so that identical symbols refer to the same block. For instance, with $N=2$, $P_{aa}$ is the probability that vertices $i_1,i_2$ belong to the same FK cluster, whereas $P_{ab} = 1 - P_{aa}$ is the probability that $i_1,i_2$ belong to two distinct FK clusters. In the context of four-point functions, we are therefore interested in the 15 probabilities $P_{aaaa}, P_{aabb},\ldots,P_{abcd}$. The combinatorial properties of FK correlation functions were further discussed in \cite{DelfinoViti}. It is natural to relate $P_{\cal P}$ to correlation functions of the spin operator. Define \begin{equation} G_{a_1,a_2,\ldots,a_N} = \left \langle {\cal O}_{a_1}(\sigma_{i_1}) {\cal O}_{a_2}(\sigma_{i_2}) \cdots {\cal O}_{a_N}(\sigma_{i_N}) \right \rangle \,, \label{order-param-corr} \end{equation} where $a_1,a_2,\ldots,a_N$ is a list of (identical or different) symbols defining a set partition ${\cal P}$, and the expectation value $\langle \cdots \rangle$ is defined with respect to the normalisation $Z$. It is straightforward to formally relate the $G_{\cal P}$ to $P_{\cal P}$. Indeed, to evaluate the expectation value of a product of Kronecker deltas, we initially suppose that $Q$ is integer, and use that spins on the same FK cluster are equal, while spins on different clusters are statistically dependent. This leads to $Q$-dependent relations, which can finally be extended to real values of $Q$ by analytical continuation. For instance, with $N=2$, one readily finds that \begin{equation} G_{a_1,a_2} = \left(Q \delta_{a_1,a_2} -1\right) P_{aa} \,. \label{G_2-point} \end{equation} In other words, the two-point function of the spin operator is proportional to the probability that the two points belong to the same FK cluster. Therefore ${\cal O}_a(\sigma_i)$ effectively ``inserts'' an FK cluster at position $i \in V$ and ensures its propagation until it is ``taken out'' by another spin operator. \begin{remark} In a recent series of works \cite{tensor1,tensor2,tensor3} we have introduced a more general class of operators ${\cal O}_{a_1,a_2,\ldots,a_N}(\sigma_{i_1},\sigma_{i_2},\ldots,\sigma_{i_N})$ that act on $N$ spins according to given irreducible representations of the symmetric groups $S_Q$ and $S_N$. These operators can enforce the propagation of more than one FK cluster, with the set of propagating clusters having specific symmetry properties. Some of the four-point functions to be considered below (namely $P_{abab} \pm P_{abba}$), with the points being considered as regrouped in two pairs, actually coincide with two-point functions of such operators, each acting on a pair of spins ($N=2$). \end{remark} In the remainder of this paper we shall focus on the same subset of four-point correlation functions as was studied in \cite{Ribault}. They are the functions $P_{\cal P}$ where the partition ${\cal P}$ contains only one or two blocks, namely: $P_{aaaa}$, $P_{aabb}$, $P_{abba}$ and $P_{abab}$. The relation with the corresponding $G_{\cal P}$ read \cite[eqs.~(19)--(22)]{DelfinoViti} \begin{subequations} \label{sumrules} \begin{eqnarray} G_{aaaa} &=&(Q-1)(Q^2-3Q+3) P_{aaaa}+(Q-1)^2(P_{aabb}+P_{abba}+P_{abab}) \,, \\ G_{aabb} &=&(2Q-3) P_{aaaa}+(Q-1)^2P_{aabb}+P_{abba}+P_{abab} \,, \\ G_{abba} &=&(2Q-3)P_{aaaa}+P_{aabb}+(Q-1)^2P_{abba}+P_{abab} \,, \\ G_{abab} &=&(2Q-3)P_{aaaa}+P_{aabb}+P_{abba}+(Q-1)^2P_{abab} \,. \end{eqnarray} \end{subequations} As already stated above, for $Q$ arbitrary, the left-hand sides of these equations are only formally defined: it is in fact the right-hand sides that give them a meaning. Note that this linear system has determinant $Q^4 (Q-1)(Q-2)^3 (Q-3)$, so it cannot be fully inverted for $Q=0,1,2,3$. By analogy with (\ref{G_2-point}) one would expect that, in the scaling limit, the four $P_{\cal P}$ of interest would be described by combinations of conformal blocks for the spin operator. In particular, the function $P_{aaaa}$ corresponding to the four points being in the same cluster should become, in the scaling limit, a crossing-invariant such combination. The other three would maybe not be crossing-invariant individually, but might be related with each other by crossing (or give rise, after proper combinations, to other crossing-invariant objects). Clearly, to implement the bootstrap programme, one needs an idea of the set of conformal blocks that may appear in these geometrical correlation functions. The key question in this problem---the one that we shall pursue in the remainder of this paper---is therefore what happens in the $s$-channel of each of these four correlation functions, when two order operators are brought close to each other. Note that, since the conformal field theories we are dealing with are not unitary, the behavior of the $G$ or $P$ functions might be more complicated than in the unitary cases, and involve, in particular, logarithmic terms. Examples of such behaviours are already known for two- and three-point functions \cite{tensor1,tensor2,tensor3}. \begin{remark} An important note: unless otherwise specified we will use the same notation (such as $P_{\cal P}$ and $G_{\cal P}$) for correlation functions defined on the lattice and for their scaling limits. \end{remark} \section{Lattice algebras} \label{sec:latt-alg} As mentioned above, our main exploratory tool for unravelling the structure of four-point functions is to impose a lattice discretisation and study the Potts model in the cylinder geometry. The algebraic object that propagates a row of $L$ Potts spins axially along the cylinder axis is a linear operator called the row-to-row transfer matrix $T$. In this section we discuss how $T$ can be used to build the partition function $Z$, and defer the more technical question about how to build the correlation functions $P_{\cal P}$ to Appendix~\ref{sec:appA1}. Both the algebraic definition of $T$ and the space of states on which it acts depend subtly on the degrees of freedom defining the model. We are here interested in two different representations, viz.\ in terms of FK clusters and TL spins, which we now describe in turn. The key technical point is to impose a weight $Q$ per cluster in the former case, or a weight $n$ per loop in the latter. \subsection{FK clusters and the join-detach algebra} \label{sec:join-detach} To build a row of an axially oriented square lattice, the transfer matrix $T$ must first add $L$ ``horizontal'' edges in some row of constant imaginary time $t$, and then propagate to the next row at time $t+1$ by adding $L$ ``vertical'' edges. It is convenient to introduce more elementary operators that add just a single edge to the lattice. Concretely, ${\sf H}_i$ adds a horizontal edge between sites $i$ and $i+1$ (mod $L$), while ${\sf V}_i$ adds a vertical edge on top of site $i$. We can thus write \begin{subequations} \label{sparse_matrix_factorisation} \begin{eqnarray} T &=& {\sf V} {\sf H} \,, \\ {\sf H} &=& {\sf H}_L \cdots {\sf H}_2 {\sf H}_1 \,, \\ {\sf V} &=& {\sf V}_L \cdots {\sf V}_2 {\sf V}_2 \,. \end{eqnarray} \end{subequations} The operators ${\sf H}_i$ and ${\sf V}_i$ must ensure the correct building of the sum $\sum_{A \subseteq E}$ in (\ref{Z_FK}). They can be written \begin{subequations} \label{HV} \begin{eqnarray} {\sf H}_i = {\sf I} + v {\sf J}_i \,, \\ {\sf V}_i = v {\sf I} + {\sf D}_i \,, \end{eqnarray} \end{subequations} where ${\sf I}$ denotes the identity operator, while ${\sf J}_i$ and ${\sf D}_i$ will be defined shortly. Each expression has two terms depending on whether the given edge $e$ belongs to the subset $A$ or not. In the former case, a weight $v$ is applied. The subtle point is to obtain also the non-local weight of $Q$ per completed cluster. To that end, $T$ acts on states $\{s_1,s_2,\ldots,s_L\}$ which are set partitions of $L$ points describing how the sites of a row are interconnected via the parts of the FK clusters living at times prior to $t$. The join operator ${\sf J}_i$ amalgamates the blocks of the partition corresponding to sites $i$ and $i+1$ (mod $L$). The detach operator ${\sf D}_i$ transforms site $i$ into a singleton, applying a weight $Q$ if it was already a singleton beforehand. The join-detach algebra is defined by the algebraic rules emanating from these requirements: \begin{subequations} \label{join-detach} \begin{eqnarray} {\sf J}_i^2 &=& {\sf J}_i \,, \\ {\sf D}_i^2 &=& Q {\sf D}_i \,, \\ {\sf J}_i {\sf D}_j {\sf J}_i &=& {\sf J}_i \mbox{ for } j=i,i+1 \,, \\ {\sf D}_i {\sf J}_j {\sf D}_i &=& {\sf D}_i \mbox{ for } j=i-1,i \,, \end{eqnarray} \end{subequations} where all indices are considered modulo $L$. Operators associated with sites that are farther apart than in the relations given simply commute. In two dimensions, the join-detach algebra is closely related to the Temperley-Lieb (TL) algebra \cite{TL71} that we describe next. For a more general discussion, see \cite{HalversonRam}. The question of how the join-detach algebra must be adapted to accommodate the computation of correlation functions is deferred to Appendices~\ref{sec:appA1}--\ref{sec:appA2}. For some applications (see Appendix~\ref{sec:appA2} in particular) we shall also need to consider the transpose of the join-detach algebra, which furnishes another geometrical representation of the TL algebra that we shall call the {\em split-attach algebra} and describe in some detail in Appendix~\ref{sec:split-attach}. \subsection{Loops and the Temperley-Lieb algebra} Another option is to define the transfer matrix in terms of the loops that separate the FK clusters from their duals. We recall that these loops now live on a tilted square lattice ${\cal M}(G)$. At each vertex of ${\cal M}(G)$ two pieces of loop, labelled $i$ and $i+1$ according to their horizontal position, can either bounce off a ``vertical'' or a ``horizontal'' edge of $G$ (or its dual $G^*$), operations that are described in imaginary time by respectively the identity operator ${\sf I}$ and the so-called braid monoid $e_i$: \begin{equation} {\sf I} = \ \begin{tikzpicture}[scale=1/4, baseline = {(current bounding box.center)},yscale=-1] \draw[black, line width = 1pt] (0,0) .. controls (0.4,0.5) and (0.4,1.5) .. (0,2); \draw[black, line width = 1pt] (1,0) .. controls (0.6,0.5) and (0.6,1.5) .. (1,2); \end{tikzpicture} \qquad \qquad e_i = \ \begin{tikzpicture}[scale=1/4, baseline = {(current bounding box.center)},yscale=-1] \draw[black, line width = 1pt] (0,0) .. controls (0,1) and (1,1) .. (1,0); \draw[black, line width = 1pt] (0,2) .. controls (0,1) and (1,1) .. (1,2); \end{tikzpicture} \qquad \end{equation} The $e_i$ generate the Temperley-Lieb (TL) algebra which has a long history \cite{TL71} and is deeply associated with work on the Potts model \cite{Baxter,PPMartin}. We note that a horizontal cut through ${\cal M}(G)$, in between two rows of vertices, will intersect the loop pieces in $N = 2L$ points. If we set ${\sf J}_i = Q^{-1/2} e_{2i}$ and ${\sf D}_i = Q^{1/2} e_{2i-1}$ for any $i=1,2,\ldots,L$, the algebraic relations (\ref{join-detach}) become simply \begin{subequations} \label{TLpdef} \begin{eqnarray} e_i^2 &=& n e_i \,, \label{TLpdef-a} \\ e_i e_{i \pm 1} e_i &=& e_i \,, \\ \left[ e_i , e_j \right] &=&0 \mbox{ for } \left|i-j \right| \geq 2 \,, \label{TLpdef-c} \end{eqnarray} where $n$ is given by (\ref{loop-fugacity}). These are precisely the defining relation of the TL algebra. Up to this point we have deliberately been rather loose about specifying the boundary conditions. Indeed, the TL algebra per se is associated with the Potts model on a strip---i.e., with open boundary conditions (that is, free boundary conditions for the Potts spins and reflecting boundary conditions for the loops)---and the generators $e_i$ are defined for $i=1,2,\ldots,N-1$. In the cylinder geometry---i.e., with periodic boundary conditions---a tempting possibility is to merely add a last generator ``closing'' the system, $e_{N}$, and define the labels modulo $N$ in the defining relations (\ref{TLpdef}), so that in particular $e_{N} e_1 e_{N} = e_1$ and $e_1 e_N e_1 = e_1$. This natural generalisation however takes one into a sticky mathematical problem: the corresponding algebra is then seen to be infinite-dimensional, even for finite $N$. In a nutshell, this occurs because of through-lines or loops that can wind around the system. While what must be done with these objects is clear in the Potts model itself, this requires providing extra information that is not present in the definition of the ``periodicised'' Temperley-Lieb algebra. This extra information takes the mathematical form of {\em quotients}. To define these quotients more precisely, it is useful to also introduce a translation generator $u$ that shifts the label of the $e_i$ generators, giving rise to the following extra relations---in addition to (\ref{TLpdef-a})--(\ref{TLpdef-c})---with integer indices considered modulo $N$ (that is, $i\in\mathbb{Z}_N$): \begin{eqnarray} u e_i u^{-1} &=& e_{i+1}\label{TLpdef-d}\\ u^2 e_{N-1} &=& e_{1} e_2 \dots e_{N-1}\label{TLpdef-e}\ . \end{eqnarray} \end{subequations} The translation operator has the diagrammatic representation \begin{equation*} u = \begin{tikzpicture}[scale=1/3, baseline = {(current bounding box.center)},yscale=-1] \foreach \r in {1,2,5,6}{ \draw[black, line width = 1pt] (\r,0) .. controls (\r,1) and (\r-1,2) .. (\r-1,3); }; \draw[black, line width = 1pt] (0,0) .. controls (0,1) and (-1,2) .. (-1,3); \draw[black, line width = 1pt] (7,0) .. controls (7,1) and (6,2) .. (6,3); \filldraw[white] (-.5,0) rectangle (-1.5,3); \filldraw[white] (6.5,0) rectangle (7.5,3); \node[anchor = north] at (3,1) {$\hdots$}; \end{tikzpicture}. \end{equation*} The last relation (\ref{TLpdef-e}) is easily understood in terms of diagrams, for example for $N=4$, \begin{equation*} e_{1}e_{2}e_{3} = \quad \begin{tikzpicture}[scale=1/4, baseline = {(current bounding box.center)},yscale=-1] \draw[black, line width = 1pt] (0,0) .. controls (0,1) and (1,1) .. (1,0); \draw[black, line width = 1pt] (0,3) .. controls (0,2) and (1,2) .. (1,3); \draw[black, line width = 1pt] (2,0) -- (2,3); \draw[black, line width = 1pt] (3,0) -- (3,3); \draw[black, line width = 1pt] (1,3) .. controls (1,4) and (2,4) .. (2,3); \draw[black, line width = 1pt] (1,6) .. controls (1,5) and (2,5) .. (2,6); \draw[black, line width = 1pt] (0,3) -- (0,6); \draw[black, line width = 1pt] (3,3) -- (3,6); \draw[black, line width = 1pt] (2,6) .. controls (2,7) and (3,7) .. (3,6); \draw[black, line width = 1pt] (2,9) .. controls (2,8) and (3,8) .. (3,9); \draw[black, line width = 1pt] (0,6) -- (0,9); \draw[black, line width = 1pt] (1,6) -- (1,9); \end{tikzpicture} \quad = \quad \begin{tikzpicture}[scale=1/4, baseline = {(current bounding box.center)},yscale=-1] \draw[black, line width = 1pt] (0,0) .. controls (0,1) and (1,1) .. (1,0); \draw[black, line width = 1pt] (2,3) .. controls (2,2) and (3,2) .. (3,3); \draw[black, line width = 1pt] (2,0) .. controls (2,1) and (0,2) .. (0,3); \draw[black, line width = 1pt] (3,0) .. controls (3,1) and (1,2) .. (1,3); \end{tikzpicture} \quad = \quad \begin{tikzpicture}[scale=1/4, baseline = {(current bounding box.center)},yscale=-1] \foreach \r in {0,...,4}{ \draw[black, line width = 1pt] (\r,0) .. controls (\r,1) and (\r-1,2) .. (\r-1,3); \draw[black, line width = 1pt] (\r,3) .. controls (\r,4) and (\r-1,5) .. (\r-1,6); }; \filldraw[white] (-.5,0) rectangle (-1.5,6); \filldraw[white] (3.5,0) rectangle (4.5,6); \draw[black, line width = 1pt] (2,6) .. controls (2,7) and (3,7) .. (3,6); \draw[black, line width = 1pt] (2,9) .. controls (2,8) and (3,8) .. (3,9); \draw[black, line width = 1pt] (0,6) -- (0,9); \draw[black, line width = 1pt] (1,6) -- (1,9); \end{tikzpicture} \quad = u^{2}e_{3}. \end{equation*} Note also that $u^{N}$ is central: it commutes with all the generators $e_i$. The resulting algebra is called the \textit{affine Temperley-Lieb} algebra $\ATL{N}(n)$. In the following we shall draw extensively on known results about its representation theory \cite{MartinSaleur,GrahamLehrer} and its relation with conformal field theory \cite{GRSV1}. \subsection{The transfer matrix sectors} While $\ATL{N}(n)$ is infinite-dimensional, it is easy to define the finite-dimensional modules which are relevant to us. First, we fix the number of {\em through-lines}, which are the pieces of loops connecting the bottom and the top of the diagrams. The number of through-lines is denoted $2j$, with $j \in \mathbb{N}/2$---the factor of $2$ comes about because one can relate each loop strand to a $\mathfrak{q}$-deformed representation of spin-$1/2$. Second, we stipulate that whenever $2j$ through-lines wind counterclockwise around the axis of the cylinder $l$ times, we can unwind them at the price of a complex phase factor ${\rm e}^{2ijlK}$; similarly, for clockwise winding, the phase will be ${\rm e}^{-i 2jlK}$ \cite{MartinSaleur}. This unwinding means more precisely that we equate the words in the algebra corresponding to the winding configurations with a numerical factor (the phase) times the related words without winding. This operation is known to give rise to a generically irreducible module over $\ATL{N}(n)$, which we denote by $\AStTL{j}{z^2=\mathrm{e}^{2 iK}}$ and call the \textit{standard module} \cite{GrahamLehrer}. A key point is that inside the modules $\AStTL{j}{z^2}$ one has the identity \begin{equation} u^{N}=z^{2j} \,, \end{equation} meaning that the central element $u^N$ of $\ATL{N}(n)$ is replaced by the complex number $z^{2j}$. The dimensions of the standard modules $\AStTL{j}{\mathrm{e}^{2 iK}}$ are given by \begin{equation}\label{eq:dj} \hat d_{j}= \binom{N}{N/2+j},\qquad j>0\ . \end{equation} Note that the dimensions do not depend on $K$, although the representations with different ${\rm e}^{iK}$ are not isomorphic. The case $j=0$ is a bit special, due to the absence of through-lines. There is no pseudomomentum, but representations are characterised by another parameter, related with the weight given to {\em non-contractible} loops (i.e., loops that close around the periodic direction). Parameterising this weight as $z+z^{-1}$, the corresponding standard module of $\ATL{N}(n)$ is denoted $\AStTL{0}{z^2}$ and has dimension $\binom{N}{N/2}$. These modules are irreducible for generic $z$. As in the case $j>0$, we indicate only the $z^2$ value, though it does not mean that the two standard modules with $\pm z$ are isomorphic. We will indicate the sign of $z$ when it is necessary. \subsection{Potts model} \label{sec:Potts-model} The fact that we wish to apply $\ATL{N}(n)$ to study the Potts models entails a few minor modifications of the general setup. First, since the number of sites $N=2L$ is even, the number of through-lines is also even, so that $j \in \mathbb{N}$. Second, the translation operator in the Potts model shifts the $L$ spins cyclically, meaning that the TL sites must be shifted by two units. Therefore, we are actually going to use the subalgebra in $\ATL{N}$ generated by the $e_i$'s and by $u^2$ (instead of $u$ itself), which is why the above notation refers to $z^2$ rather than to $z$ itself. The basic ingredient is the finite-dimensional modules of $\ATL{N}$ described above, with the loop weight $n = \sqrt{Q}$ parameterised as in (\ref{loop-fugacity}) in order to match (\ref{Z_loop}). However, to account for the particularities of the Potts model, the algebra that we are mostly interested in is a quotient of $\ATL{N}$, known as the Jones-Temperley-Lieb algebra $\rJTL{N}(n)$ \cite{Jones,ReadSaleur01}. It is obtained by: \begin{itemize} \item[(i)] replacing non-contractible loops by the same weight $n=\sqrt{Q}$ as for the contractible ones; \item [(ii)] identifying diagrams connecting the same sites, even if they are non-isotopic on the cylinder; and \item[(iii)] setting $u^N=1$, which allows one to unwind through-lines. \end{itemize} Rules (i) and (ii) are only relevant in the case $j=0$, where through-lines are not present. The first rule leads to $z^2=\mathfrak{q}^{\pm 2}$. In this case, in fact, the affine TL module $\AStTL{0}{\mathfrak{q}^2}$ is reducible, and contains a unique simple submodule isomorphic to $\AStTL{1}{1}$. The reason for this is that the general $\ATL{N}(n)$ allows (in diagrammatic terms) to distinguish loop arcs that connect two given sites on the front or on the back of the cylinder, meaning that a closed loop can be given different weights depending on whether it is contractible or not. When this distinction is not needed we must identify arcs only according to which sites are being connected, as in rule (ii). Identifying non-isotopic diagrams in this way corresponds to replacing $\AStTL{0}{\mathfrak{q}^2}$ by the (unique) simple quotient $\AStTL{0}{\mathfrak{q}^2}/\AStTL{1}{1}\equiv \bAStTL{0}{\mathfrak{q}^2}$ of dimension \begin{equation} \overline{d}_0 = \dim \bAStTL{0}{\mathfrak{q}^2}= \binom{N}{N/2} - \binom{N}{N/2+1}\ , \end{equation} In technical terms, this quotient is precisely the standard module of $\rJTL{N}(n)$ for $j=0$. \begin{remark} The quotient $\bAStTL{0}{\mathfrak{q}^2}$ is but one example of representations that appear more generally in $\ATL{N}$ when $\mathfrak{q}$ is still generic, but $z$ takes particular values \cite{GrahamLehrer,PerioFusion}. Indeed, the standard module $\AStTL{j'}{z'}$ has a non-zero homomorphism to $\AStTL{j}{z}$, \begin{equation}\label{cell-emb} \StJTL{j'}{z'}\hookrightarrow\StJTL{j}{z} \,, \end{equation} if and only if $j'-j \in \mathbb{N}_0$ and the pairs $(j',z')$ and $(j,z)$ satisfy \begin{equation}\label{eq:emb-cond} (z')^2 = (-\mathfrak{q})^{2\epsilon j} \qquad \text{and}\qquad z^2 = (-\mathfrak{q})^{\epsilon 2j'},\qquad \text{for} \quad \epsilon=\pm1. \end{equation} When $\mathfrak{q}$ is not a root of unity, there is at most one solution to (\ref{cell-emb}). When there is one, the module $\StJTL{j}{z}$ is not irreducible, but has a unique proper irreducible submodule isomorphic to $\StJTL{j'}{z'}$. One can then obtain a simple module by taking the quotient \begin{equation} \bAStTL{j}{z}\equiv\StJTL{j}{z}/\StJTL{j'}{z'}\label{Wbar-def} \end{equation} of dimension \begin{equation} \overline{d}_j = \dim \bAStTL{j}{\mathfrak{q}^2} = \binom{N}{N/2+j} - \binom{N}{N/2+j'} \,. \end{equation} The quotient $\bAStTL{0}{\mathfrak{q}^2}$ appearing above is but the simplest example (with $j=0$, $z=\mathfrak{q}^2$ and $j'=1$, $z'=1$) of this situation, and it is the {\em only} such quotient that is relevant for the Potts model at generic $\mathfrak{q}$. \end{remark} Whenver $j\neq 0$, rule (iii) leads to a quantisation of the momentum: $K=\pi p/M$, with $M$ a divisor of $j$ (i.e., $M | j$) and with a greatest common divisor $p \wedge M =1$. The modules encountered so far are thus $ \bAStTL{0}{\mathfrak{q}^2}=\bAStTL{0}{\mathfrak{q}^{-2}}$, and $\AStTL{j}{\mathrm{e}^{2 i\pi p/M}}$ for $j\neq 0$ and $M|j$. On top of this, there is a small subtlety having to do with the relation between through-lines and through-clusters, by which we mean clusters that propagate along the imaginary time direction. For $j \ge 2$, each of the $2j$ through-lines alternatingly separates a propagating FK cluster and an propagating dual cluster, implying the existence of precisely $j$ through-clusters. However, when we wish to impose just one through-cluster, the situation is different. Since nothing prevents this cluster from wrapping the periodic direction of the cylinder, it will in fact do so with probability one, implying that through-lines will be absent ($j=0$). On the other hand, there cannot be any non-contractible loops either, since this would prevent the propagation of the through-cluster. The correct module is thus obtained by giving a vanishing weight to non-contractible loops. This is easily accomplished by setting $z=\pm i$, leading to $\AStTL{0}{-1}$. The three types of modules that we have just introduced: \begin{equation} \bAStTL{0}{\mathfrak{q}^{\pm 2}}\oplus~\AStTL{j}{\mathrm{e}^{2 i\pi p/M}}{\scriptstyle (M|j,j\ge 2)}\oplus~\AStTL{0}{-1}\label{modules} \end{equation} are known to encode the full Potts model partition function on the torus \cite{DFSZ,ReadSaleur01,RichardJacobsen}. Formal multiplicities for these modules are also known. For $Q$ non-integer, they are real numbers with group-theoretical significance \cite{tensor2,tensor3}. The crucial observation that will be made below is that the modules (\ref{modules}) are also the sufficient objects to describe the four-point correlation functions in the geometrical Potts model. An important additional fact is that actually only the modules with {\em even values} of $j$ are necessary for the description of four-point functions. By contrast, any $j \ge 2$ contributes\footnote{Note that the value $j=1$ is perfectly allowed in algebraic terms, and is crucial for the description of the statistics of cluster hulls. It does not, however, appear in the torus partition function of the Potts model, nor in the connectivity correlations functions in the bulk.} to the partition function on the torus, as has been verified in details for finite systems \cite{JacobsenSalas-torus}. This last result was not totally obvious a priori. Indeed, one must in general be careful with geometrical questions in models such as the Potts model, where the set of observables is seemingly not limited, if one sways far enough from locality. It is well-known, for instance, that correlations involving several independent paths along clusters---the case of two such paths defines the celebrated backbone exponents---cannot be described using $\ATL{N}$, and the corresponding exponents have never been identified using Coulomb gas techniques. Indeed, the numerical measurements of \cite{JacobsenZinn1,JacobsenZinn2,DengBloteNienhuis} appear to convincingly rule out any such identification for this whole class of so-called monochromatic path-crossing exponents. Similar remarks can be made about other seemingly reasonable observables, such as the shortest-path exponent \cite{ZhouYangDengZiff}, to mention but one example. Fortunately, then, the matters seem to be (relatively) simpler for the four-point correlation functions. \subsection{Summary of notations} For the reader's convenience we summarise here some of the notations used in this paper. They are, as far as possible, the same as those used in \cite{GRS1,GRS2,GRS3,GJSV,GJRSV,BGJSV,GRSV1,PerioFusion}. \begin{itemize} \item$\AStTL{j}{z^2=e^{2iK}}$ --- the standard modules over $\ATL{N}$, \item$\StJTL{j}{z^2}$ --- the same, with $P=e^{2iK}$, \item$\bAStTL{0}{\mathfrak{q}^2}$ --- the standard module over $\rJTL{N}$ for $j=0$, \item $\left[j, e^{2iK}\right]$ or $\IrrJTL{j}{z^2}$ --- simple modules over $\ATL{N}(n)$. \end{itemize} Moreover, when discussing the transfer matrices for the Potts model correlation functions (see Appendix~\ref{sec:appA}, and section~\ref{sec:momentum_sectors} in particular) we shall sometimes need a lighter notation $V_{\ell,k,m}$ for the sector with $\ell$ propagating clusters, an integer momentum variable $k=0,1,\ldots,\ell-1$ for the through-lines (if any), and a lattice momentum $m=0,1,\ldots,L-1$ which is the precursor of the conformal spin for a system of finite size $\ell$. The notations $V_{\ell,k}$ or $V_{\ell}$ with some of the variables omitted mean that these take indiscriminate values. The correspondence with the standard modules is then: \begin{itemize} \item $V_0$ is the sector with no through-lines, and non-contractible loops have weight $\sqrt{Q}$: $V_0=\bAStTL{0}{z^2=\mathfrak{q}^2}$, \item $V_1$ is the sector with no through-lines, and non-contractible loops have weight zero: $V_1=\AStTL{0}{z^2=-1}$, \item $V_{\ell,k}$ is the sector with $j=\ell \ge 2$ {\em pairs} of through-lines and phases $z^2=e^{2i\pi k/j}$: $V_{\ell,k}=\AStTL{j}{z^2=e^{2i\pi k/j}}$. \end{itemize} \section{Four-point functions in the $s$-channel} \label{sec:four-point} \subsection{Generalities} \label{sec:generalities} We consider a general four-point function of primary operators in a CFT, which we write in the following convenient form in the plane: \begin{equation} \langle\Phi_1(z_1,\bar{z}_1)\Phi_2(z_2,\bar{z}_2)\Phi_3(z_3,\bar{z}_3)\Phi_4(z_4,\bar{z}_4)\rangle=\prod_{i<j} z_{ij}^{\delta_{ij}} \bar{z}_{ij}^{\bar{\delta}_{ij}} \mathrm{G}(z,\bar{z}) \,, \label{GenFptf} \end{equation} where we have denoted $z_{ij} \equiv z_i - z_j$, with the exponents \begin{subequations} \begin{eqnarray} \delta_{12} &=& 0 \,, \\ \delta_{13} &=& -2h_1 \,, \\ \delta_{14} &=& 0 \,, \\ \delta_{23} &=& h_1-h_2-h_3+h_4 \,, \\ \delta_{24} &=& -h_1-h_2+h_3-h_4 \,, \\ \delta_{34} &=& h_1+h_2-h_3-h_4 \,. \end{eqnarray} \end{subequations} The antiholomorphic exponents $\bar{\delta}_{ij}$ are obtained from the holomorphic ones $\delta_{ij}$ by the replacement $h\to\bar{h}$, and the same convention henceforth applies to any other quantity. The parameter $z$ denotes the anharmonic ratio \begin{equation} z\equiv {z_{12}z_{34}\over z_{13}z_{24}} \,. \end{equation} One finds easily that \begin{equation} \hbox{Lim}_{\Lambda\to\infty} \Lambda^{2h_3}\bar{\Lambda}^{2\bar{h}_3} \langle \Phi_1(z,\bar{z})\Phi_2(0,0)\Phi_3(\Lambda,\bar{\Lambda})\Phi_4(1,1)\rangle=\mathrm{G}(z,\bar{z}) \,. \label{z01inf} \end{equation} This function $G(z,\bar{z})$ is what one usually refers to as $ \langle \Phi_1(z,\bar{z})\Phi_2(0,0)\Phi_3(\infty,\infty)\Phi_4(1,1)\rangle$. It is known \cite{BPZ} to expand as a sum over conformal blocks \begin{equation} \mathrm{G}(z,\bar{z})=\sum_{\Delta,\bar{\Delta}\in {\cal S}} C_{\Phi_1\Phi_2\Phi_{\Delta\bar{\Delta}}}C_{\Phi_{\Delta\bar{\Delta}}\Phi_3\Phi_4}{\cal F}_\Delta^{(s)}(z)\overline{{\cal F}}_{\bar{\Delta}}^{(s)}(\bar{z}) \,, \end{equation} where $(\Delta,\bar{\Delta})$ are the conformal weights of the primary fields appearing in the operator product expansion relevant at small $z$. They define the scaling dimension (eigenvalue of the dilatation operator) $\Delta + \bar{\Delta}$ and the conformal spin (eigenvalue of the rotation operator) $\Delta - \bar{\Delta}$. We shall be particularly interested in the limit $z_1\to z_2$ and $z_3\to z_4$: this is called the $s$-channel (borrowing a standard terminology of particle physics due to Mandelstam). This limit corresponds to taking $z \to 0$ in (\ref{z01inf}), so we can write the expansion \begin{equation} \mathrm{G}(z,\bar{z})=\sum_{\Delta,\bar{\Delta}\in {\cal S}} C_{\Phi_1\Phi_2\Phi_{\Delta\bar{\Delta}}}C_{\Phi_{\Delta\bar{\Delta}}\Phi_3\Phi_4}z^{(\Delta-h_1-h_2)}\bar{z}^{\bar{\Delta}-\bar{h}_1-\bar{h}_2}\left[1+O(z,\bar{z})\right] \,. \label{leadingexpan} \end{equation} One could of course similarly consider the $t$-channel ($z \to \infty$) and in the $u$-channel ($z \to 1$), by expanding in powers of $\frac{1}{z}$ and $(z-1)$ respectively. The idea of the conformal bootstrap programme is that all these expansions determine the same function $\mathrm{G}(z,\bar{z})$ and hence will impose constraints. The first step in any further discussion is therefore to establish the fields intervening in one of these channels, which we take here to be the $s$-channel. The key question we want to address in this paper is thus to determine the set ${\cal S}$ of values of $\Delta,\bar{\Delta}$, which we will tackle by a combination of algebraic and numerical methods. In its crudest form, the latter involves the brute force numerical determination of a (very) large number of terms appearing in the right-hand side of (\ref{leadingexpan}). Note that the determination of the set ${\cal S}$ from the knowledge of these terms will only be fully possible in ``generic'' cases, where none of the $\Delta,\bar{\Delta}$ differ by integers. Otherwise, there will be ambiguities, as a term such as $\Delta+n,\bar{\Delta}+\bar{n}$ (with $n,\bar{n}$ integer) may arise from a genuine primary field, or from a Virasoro descendent of some primary field with weights $\Delta+p,\bar{\Delta}+\bar{p}$, with $p \le n$ and $\bar{p} \le \bar{n}$ (with at least one of the inequalities being strict). The non-generic case requires $Q$ to take particular values (with $\mathfrak{q}$ being a root of unity); it is clearly more complicated than the generic case and will typically lead to at least some correlation functions having logarithmic behaviour. A few non-generic cases (not all of them logarithmic) will be discussed in Appendix~\ref{sec:appB}. But the main text is henceforth dedicated to the generic case, for which we shall determine ${\cal S}$ fully. Our strategy is to study the expansion (\ref{leadingexpan}) on the cylinder, where we will be able to use, on the numerical side, transfer matrix techniques, and, on the analytic side, algebraic results. The four-point function on the cylinder follows from (\ref{GenFptf}) via the conformal map \begin{equation} w={L\over 2\pi}\ln z \,. \end{equation} Using the fact that the fields are primary, and restricting here to $i=j=k=l$ for notational simplicity, we find \begin{equation} \langle \Phi(w_1,\bar{w}_1)\Phi(w_2,\bar{w}_2)\Phi(w_3,\bar{w}_3)\Phi(w_4,\bar{w}_4)\rangle_{\rm cyl}=\left({2\pi\over L}\right)^{4(h+\bar{h})} {1\over \left|4\sinh{\pi w_{13}\over L}\sinh{\pi w_{24}\over L}\right|^{2(h+\bar{h})} }~\mathrm{G}(w,\bar{w}) \,, \end{equation} where the subscript ``cyl'' on the left-hand side refers to the cylinder geometry, and we have set $w_{ij} \equiv w_i - w_j$ as before. The expansion variable is now \begin{equation} w={\sinh{\pi w_{12}\over L}\sinh{\pi w_{34}\over L}\over \sinh{\pi w_{13}\over L}\sinh{\pi w_{24}\over L}} \,. \end{equation} Using (\ref{leadingexpan}) we can write this as \begin{eqnarray} \langle \Phi(w_1,\bar{w}_1)\Phi(w_2,\bar{w}_2)\Phi(w_3,\bar{w}_3)\Phi(w_4,\bar{w}_4)\rangle_{\rm cyl}=\left({2\pi\over L}\right)^{4(h+\bar{h})}{1\over |4\sinh{\pi w_{12}\over L}\sinh{\pi w_{34}\over L}|^{2(h+\bar{h})}}\nonumber\\ \sum_{\Delta,\bar{\Delta}\in {\cal S}} C_{\Phi\Phi\Phi_{\Delta\bar{\Delta}}}C_{\Phi_{\Delta\bar{\Delta}}\Phi\Phi}\left[\left({\sinh{\pi w_{12}\over L}\sinh{\pi w_{34}\over L}\over \sinh{\pi w_{13}\over L}\sinh{\pi w_{24}\over L}}\right)^{\Delta}\left({\sinh{\pi \bar{w}_{12}\over L}\sinh{\pi \bar{w}_{34}\over L}\over \sinh{\pi \bar{w}_{13}\over L}\sinh{\pi \bar{w}_{24}\over L}}\right)^{\bar{\Delta}}+O(w,\bar{w})\right]\label{CylExpan} \end{eqnarray} In practice, to access the $s$-channel properties, we will take the points $w_1,w_2$ on a given slice of imaginary time, and $w_3,w_4$ on another, distant, slice along the cylinder of finite circumference $L$. This geometry is shown in Figure~\ref{fig:cylinder}. In other words, $w_{12}$ and $w_{34}$ will be fixed, while $w_{13}$ and $w_{24}$ will be large and vary. In this limit, it will then be possible to compare the expansion (\ref{CylExpan}) with the results of transfer matrix calculations, and identify, in particular, the set ${\cal S}$. \begin{figure} \begin{center} \begin{tikzpicture}[scale=0.7] \draw[black,thick] (0,0) arc (0:360:0.5 and 1.5); \draw[black,thick] (-0.5,1.5)--(8,1.5); \draw[black,thick] (-0.5,-1.5)--(8,-1.5); \draw[black,thick,dotted] (1.5,0)--(4.2,0); \draw[black,thick,dotted] (1.5,0) arc (0:360:0.5 and 1.5); \draw[black,fill] (1.42,0.8) circle (2pt); \draw[black,fill] (1.42,-0.8) circle (2pt); \draw (1.42,0.8) node[left] {$w_1$}; \draw (1.42,-0.8) node[left] {$w_2$}; \draw (1.5,0.4) node[right] {$a$}; \draw (1.5,-0.4) node[right] {$a$}; \draw[black,thick,dotted] (6.5,0) arc (0:360:0.5 and 1.5); \draw[black,thick] (8.0,-1.5) arc (-90:90:0.5 and 1.5); \draw[black,thick,dashed] (8.0,1.5) arc (90:270:0.5 and 1.5); \draw[black,thick,dotted] (3.8,-0.5)--(6.5,-0.5); \draw[black,fill] (6.50,0.3) circle (2pt); \draw[black,fill] (6.27,-1.3) circle (2pt); \draw (6.50,0.3) node[right] {$w_3$}; \draw (6.27,-1.3) node[right] {$w_4$}; \draw (6.50,-0.1) node[left] {$a$}; \draw (6.50,-0.9) node[left] {$a$}; \draw[black,thin,<->] (4,0)--(4,-0.5); \draw (4,-0.25) node[right] {$x$}; \draw[black,thin,<->] (1.0,-1.8)--(6.0,-1.8); \draw (4,-2.1) node {$l$}; \end{tikzpicture} \end{center} \caption{Four-point functions in the cylinder geometry.} \label{fig:cylinder} \end{figure} Let us now be more precise. We set \begin{subequations} \label{eq:w1w2w3w4} \begin{eqnarray} w_1 &=& ia \,, \\ w_2 &=& -ia \,, \\ w_3 &=& i(a+x)+l \,, \\ w_4 &=& i(-a+x)+l \,, \end{eqnarray} \end{subequations} which means that the points $w_{1,2}$ and $w_{3,4}$ are a certain distance% \footnote{Note that $a$ is not the lattice spacing, but some arbitrary parameter. It will always occur in the combination $\frac{2a}{L}$.} $2a$ apart on the vertical axis, $l$ is the horizontal distance (imaginary time) between the two groups, and on top of this we have the centre of mass of $w_{3,4}$ shifted by $x$. A short calculation then gives \begin{eqnarray} \langle \Phi(w_1,\bar{w}_1)\Phi(w_2,\bar{w}_2)\Phi(w_3,\bar{w}_3)\Phi(w_4,\bar{w}_4)\rangle_{\rm cyl} &=& \left({2\pi\over L}\right)^{4(h+\bar{h})} e^{-8\pi h l/L} \label{CylExpan1} \\ &\times& \left(1-e^{-2\pi (l+ix)/L}\right)^{-4h}\left(1-e^{-2\pi (l-ix)/L}\right)^{-4h} \nonumber \\ & \times& \mathrm{G}\left({4e^{i\pi}\sin^2 {2\pi a\over L} e^{-2\pi(l+ix)/L}\over (1-e^{-2\pi(l+ix)/L})^2},{4e^{-i\pi} \sin^2 {2\pi a\over L} e^{-2\pi(l-ix)/L}\over (1-e^{-2\pi(l-ix)/L})^2}\right) \,. \nonumber \end{eqnarray} We can then expand this from (\ref{leadingexpan}): \begin{eqnarray} \langle \Phi(w_1,\bar{w}_1)\Phi(w_2,\bar{w}_2)\Phi(w_3,\bar{w}_3)\Phi(w_4,\bar{w}_4)\rangle_{\rm cyl} &=& \left({2\pi\over L}\right)^{4(h+\bar{h})} {1\over (4\sin^2 {2\pi a\over L})^{4h}} \sum_{\Delta,\bar{\Delta}\in {\cal S}} C_{\Phi\Phi\Phi_{\Delta\bar{\Delta}}}C_{\Phi_{\Delta\bar{\Delta}}\Phi\Phi} \nonumber\\ &\times& \left(4\sin^2{2\pi a\over L}\right)^{\Delta+\bar{\Delta}} (-1)^{\Delta-\bar{\Delta}}\xi^\Delta\bar{\xi}^{\bar{\Delta}}[1+O(\xi,\bar{\xi})] \,, \label{mainexp} \end{eqnarray} where we have set \begin{equation} \xi\equiv e^{-2\pi (l+ix)/L},~\bar{\xi}\equiv e^{-2\pi (l-ix)/L} \,. \end{equation} The bracket $[1+O(\xi,\bar{\xi})]$ contains now contributions from the conformal blocks and contributions from the hyperbolic functions in the conformal map. The expansion (\ref{mainexp}) is the crucial tool that we will use systematically in our analysis below. In the following, we will sometimes use the short-hand notation \begin{equation} A_{\Phi_{\Delta\bar{\Delta}}}\equiv C_{\Phi\Phi\Phi_{\Delta\bar{\Delta}}}C_{\Phi_{\Delta\bar{\Delta}}\Phi\Phi} \,. \label{ampldef} \end{equation} We now discuss this in more detail. \begin{remark} We also see that if we exchange $w_1$ and $w_2$ in (\ref{CylExpan}), the leading contributions for a given $\Delta,\bar{\Delta}$ is multiplied by $(-1)^{\Delta-\bar{\Delta}}$. Hence primary fields with odd integer spin should contribute an opposite weight upon making this exchange. \end{remark} For future reference, the definition of the channels is \begin{eqnarray} \mbox{$s$-channel} &:& z_1\to z_2\nonumber\\ \mbox{$t$-channel} &:& z_1\to z_4\nonumber\\ \mbox{$u$-channel} &:& z_1\to z_3 \end{eqnarray} Henceforth, when denoting the probabilities $P_{a_1,a_2,a_3,a_4}$, the four labels specifying the partition ${\cal P}=\{a_1,a_2,a_3,a_4\}$ refer to the points $z_1,z_2,z_3,z_4$ in that order. Clearly, then, $P_{aaaa}$ should have the same spectrum (and structure constants) in all channels, while, for instance, $P_{aabb}$ should have the same spectrum (and structure constants) in the $t$- and $u$-channels, while the spectrum should be different in the $s$-channel. \subsection{Exponents} Contrarily to what is implied in \cite{Ribault}, the exponents of percolation---and more generally the $Q$-state Potts model in the FK cluster representation---are essentially known (with the exception of certain ``exotic'' exponents, see \cite{JacobsenZinn1,JacobsenZinn2,DengBloteNienhuis,ZhouYangDengZiff}). This knowledge relies on two stages. First, the transfer matrix sectors of the Potts model can be described in terms of standard modules of the affine TL algebra, as described in section~\ref{sec:Potts-model}. Second, the continuum limits of these objects are known in the form of spectrum generating functions within the corresponding CFT, as we now review. This is of course not yet the solution of the $s$-channel conundrum, but since we are able to formulate the four-point functions in terms of the FK transfer matrix (see section~\ref{sec:algorithm} and Appendix~\ref{sec:appA1}), the results on the generating functions will narrow down the set of states than can possibly be part of the spectrum ${\cal S}$. Extensive numerical analysis---corroborated by the solvability of a few special cases (see Appendix~\ref{sec:appB})---will then lead to the results that we give in section~\ref{sec:results}. The local FK connectivities in the geometrical Potts model and their evolution along the cylinder are described by a transfer matrix or, in the familiar extreme anisotropic limit, a Hamiltonian. Both transfer matrix and Hamiltonian exhibit the same conformal content---that is, eigenstates associated with local CFT operators, and the corresponding conformal weights $h,\bar{h}$, together with their multiplicities. It is convenient to encode the latter into spectrum generating functions. Using for instance the Hamiltonian language and setting \begin{equation} H=-\lambda\sum_{i=1}^{2L} e_i \,, \end{equation} with $\lambda$ adjusted so that the sound velocity is unity as usual, we define the generating function of levels (eigenenergies of $H$) and lattice momentum $P$ as as traces of lattice operators, with the scaling limit \cite{PS} \begin{equation} \hbox{Tr}\, \left[e^{-\beta_R(H-N\varepsilon_0)}e^{-i\beta_I P}\right]\;\xrightarrow{\, \rm scaling}\; \hbox{Tr}\, q^{L_0-c/24}\bar{q}^{\bar{L}_0-c/24} \,. \label{charform} \end{equation} % Here $\varepsilon_0$ is the (non-universal) ground state energy per site in the limit $N\to\infty$. The scaling limit is defined by taking $N,\beta_R,\beta_I\to\infty$ while keeping the modular parameters% \footnote{Here and elsewhere a notation of the type $q(\bar{q})$ means that $q$ is given by the first expression on the right-hand side (the one with a $+$ sign), and $\bar{q}$ by the second expression (with a $-$ sign).} $q(\bar{q})=\exp\left[-{2\pi\over N}(\beta_R\pm i\beta_I)\right]$ (with $\beta_{R,I}$ real and $\beta_R>0$) finite. The parameters $\beta_R$ and $\beta_I$ define the size of the system in the two principal directions of the torus, while the trace ensures the periodic boundary conditions in the imaginary time direction. We recall that $N=2L$ is the number of sites in the system, often referred to as the length of the spin chain in the Hamiltonian limit. In other words, only even chains are relevant in our problem. On the right-hand side of (\ref{charform}), $L_0$ and $\bar{L}_0$ are of course Virasoro generators, while $c$ denotes the central charge. The generating function (\ref{charform}) calculated in the modules $\mathcal{W}_{j,\mathrm{e}^{2 i K}}$ is \cite{AGR,PS} \begin{equation}\label{eq_F} \displaystyle \mathrm{Tr}_{\mathcal{W}_{j,\mathrm{e}^{2iK}}}\left[e^{-\beta_R(H-N\varepsilon_0)}e^{-i\beta_I P}\right] \;\xrightarrow{\, \rm scaling\,}\; F_{j,\mathrm{e}^{2iK}} \equiv\frac{q^{-c/24}\bar{q}^{-c/24}}{P(q) P(\bar{q})} \sum_{e \in \mathbb{Z}} q^{h_{e+{K\over\pi},-j}} \bar{q}^{h_{e+{K\over\pi},j}}, \end{equation} where \begin{equation} \displaystyle P(q) = \prod_{n=1}^{\infty} (1 - q^n) = q^{-1/24} \eta (q) \end{equation} is the (inverse of) the generating function for integer partitions, and $\eta(q)$ is Dedekind's eta function. Instead of (\ref{loop-fugacity}) we shall find it convenient to parameterise the number of states in the Potts model by \begin{equation} \sqrt{Q}=2\cos \left( {\pi\over m+1} \right) \,, \mbox{ with } m \in [1,\infty] \,, \end{equation} so that $\mathfrak{q} = {\rm e}^{\frac{i \pi}{m+1}}$. Note that to access the generic case ($\mathfrak{q}$ not a root of unity) we do {\em not} restrict $m$ to be integer, as would be the case for the minimal models. The corresponding central charge is then \begin{equation} c=1-{6\over m(m+1)} \,, \end{equation} and we also use the Kac table parameterisation of conformal weights \begin{equation} h_{rs}={[(m+1)r-ms]^2-1\over 4m(m+1)} \,. \end{equation} In ``usual'' CFT the labels $(r,s)$ are positive integers, but as for the parameter $m$ we shall here allow them to take more general values, as is already evident from (\ref{eq_F}). To make contact with standard references, it is also convenient in the following to introduce the {\em Coulomb gas coupling constant} $g={m\over m+1}$ and the {\em background electric charge} $e_0={1\over m+1}$. The operator associated with the order parameter has conformal weight $h_{1/2,0}$ \cite{denNIJS,Nienhuis_CG}, the primordial example of an ``unusual'' $h_{rs}$, with $r$ here being non-integer and $s$ non-positive. This operator belongs to the generating function $F_{0,-1}$. \begin{remark} \label{rem:switch} To compare with \cite{Ribault} one must set in their equation (1.1) $\beta^2={m\over m+1}$ (so that ${1\over 2}\leq \beta^2\leq 1$), and $q=Q$. Moreover, the conventions used in their paper for the exponents are {\em switched} with respect to ours. In other words, they call $\Delta_{sr}$ what we call $h_{rs}$ (or $\Delta_{rs}$). \end{remark} Restricting now to the cases of interest with momentum $K=\pi p/M$ and $M|j$ gives \begin{equation} \displaystyle F_{j,\mathrm{e}^{2i\pi p/M}} =\frac{q^{-c/24}\bar{q}^{-c/24}}{P(q) P(\bar{q})} \sum_{e \in \mathbb{Z}} q^{h_{e+{p\over M},-j}} \bar{q}^{h_{e+{p\over M},j}} \,, \mbox{ with } M|j \mbox{ and } j \in \mathbb{Z} \,. \label{bspec1} \end{equation} On top of this we also have to consider the generating function of levels in $\mathcal{W}_{0,-1}$, which reads \begin{equation} \displaystyle F_{0,-1} =\frac{q^{-c/24}\bar{q}^{-c/24}}{P(q) P(\bar{q})} \sum_{e \in \mathbb{Z}} q^{h_{e+1/2,0}} \bar{q}^{h_{e+1/2,0}} \,. \label{bspec2} \end{equation} Finally we need the generating function for the quotient module $\bAStTL{0}{\mathfrak{q}^2}$. The twist $e^{iK}=\mathfrak{q}$ corresponds in our notation to ${K\over\pi}=e_0$, so we have first \begin{equation} \displaystyle F_{0,\mathfrak{q}^2} =\frac{q^{-c/24}\bar{q}^{-c/24}}{P(q) P(\bar{q})} \sum_{e \in \mathbb{Z}} q^{h_{e+e_0,0}} \bar{q}^{h_{e+e_0,0}} \,. \end{equation} The subtraction necessary to obtain the module $\bAStTL{0}{\mathfrak{q}^2}=\AStTL{0}{\mathfrak{q}^2}/\AStTL{1}{1}$ leads to the expression for the generating function of the corresponding levels: \begin{equation} \displaystyle \bar{F}_{0,\mathfrak{q}^2} =\frac{q^{-c/24}\bar{q}^{-c/24}}{P(q) P(\bar{q})} \left[\sum_{e \in \mathbb{Z}} q^{h_{e+e_0,0}} \bar{q}^{h_{e+e_0,0}}-\sum_{e \in \mathbb{Z}} q^{h_{e,1}} \bar{q}^{h_{e,-1}}\right] \,. \end{equation} The subtraction can actually be implemented term by term. Introducing the characters of the so-called Kac modules---which are Verma modules where a single singular vector at level $r s$ has been removed---, \begin{equation} K_{rs}=q^{-c/24} {q^{h_{rs}}-q^{h_{r,-s}}\over P(q)}=q^{-c/24} q^{h_{rs}}{1-q^{rs}\over P(q)} \,, \end{equation} we have \begin{equation} \bar{F}_{0,\mathfrak{q}^2}=\sum_{r=1}^\infty K_{r1}(q)K_{r1}(\bar{q}) \,. \label{bspec3} \end{equation} To summarise, corresponding to the modules (\ref{modules}) we expect (and confirm below) that the set of exponents contributing to the four-point connectivities is encoded into \begin{equation} F_{j,\mathrm{e}^{2i\pi p/M}}{\scriptstyle( M|j,~j\geq 2)}\oplus~F_{0,-1}\oplus~\bar{F}_{0,\mathfrak{q}^2} \,. \end{equation} Moreover, we shall see below that only $j$ even contributes, which cannot be foreseen at this stage. We note that for generic values of $Q$ or $\mathfrak{q}$ (i.e, with $m$ irrational), there are no coincidences of exponents in the different sectors (generating functions)---this is also true on the lattice, where there are no coincidences of eigenvalues. Moreover, in a given sector, no two exponents differ by integers. \subsubsection{Numerical validation of the generating functions} It appears useful at this stage to test the internal coherence of the ingredients brought together this far. On one hand, in section \ref{sec:Potts-model} we have related the sectors of the Potts model transfer matrix $T$ to certain standard modules, $\mathcal{W}_{j,z^2}$ and $\bAStTL{0}{\mathfrak{q}^2}$, of the affine TL algebra. On the other hand, we have just given their corresponding spectrum generating functions, $F_{j,z}$ and $\bar{F}_{0,\mathfrak{q}^2}$. This means that a numerical diagonalisation of $T$ in the various sectors should produce---after a proper extrapolation to the continuum limit $L \to \infty$---the primaries and descendents (with multiplicities) of these generating functions. We are not aware of a previous careful study that this is indeed so. To this end, we first outline in Appendix~\ref{sec:appA3} the extraction of the eigenvalue spectrum of $T$ in the various sectors relevant for the Potts model. Fixing the values of the momentum and the conformal spin is a non-trivial operation that is expounded in Appendix~\ref{sec:momentum_sectors}. Once this has been done, we can examine the spectrum of $T$; this is done first for a generic value of $Q$ in Appendix~\ref{sec:specQ12}, and then for a few non-generic values: $Q=4$ in Appendix~\ref{sec:specQ4}, and $Q=2$ in Appendix~\ref{sec:appA_Ising}. When combined, these three cases permit us to test examples of Verma modules with zero, one, or infinitely many singular vectors. In all cases we find that the agreement with the expected spectrum generating functions is perfect. In the generic case, we are able to see descendents up to level 6 for the identity operator, and up to level 3 for other operators. Moreover, the set of primaries fully agree with the expectations from the affine TL algebra. In the Ising case we are able to follow the first 29 scaling levels and observe descendents up to level 9 for both operators ($I$ and $\epsilon$) in the even sector, with an agreement better than $10^{-4}$ for almost all scaling dimensions. Moreover, the degeneracy observed for each ``completed'' level is in perfect accord with the spectrum generating functions. The techniques used in the numerical analysis may be of independent interest and can be consulted in the appendices (see also Appendix~\ref{sec:practical-remarks} for a few practical remarks). \subsection{The numerical algorithm} \label{sec:algorithm} The geometrical setup for four-point functions is shown in Figure~\ref{fig:cylinder}. As stated earlier, our lattice discretisation consists in embedding a periodic square lattice $G=(V,E)$ of width $L$ Potts spins into the cylinder, with the edges $E$ being either horizontal or vertical with respect to the figure (axial geometry). We possess two different strategies for obtaining numerical results for the correlation functions. The first strategy gives access to the most general FK correlation functions, namely the 15 different $P_{a_1,a_2,a_3,a_4}$. It is practically feasible up to size $L=7$, after a considerable numerical effort. It applies for both generic and non-generic values of $Q$, meaning that in the latter case it can determine the indecomposable structure of correlation functions. The second strategy applies to a smaller set of correlation functions, namely the four order-parameter correlators $G_{aaaa}$, $G_{aabb}$, $G_{abba}$ and $G_{abab}$. Its advantage is that it gives access to larger sizes, in practice up to $L=11$, at a much smaller computational expenditure than the first method. However, it applies only to generic values of $Q$ and, at least in its present form, cannot determine the Jordan block structure at non-generic $Q$-values. We now briefly outline the two methods, while relegating all technical details to Appendices~\ref{sec:appA1}--\ref{sec:appA2}. \subsubsection{First method} \label{sec:1st_method} It is shown in the appendix that all $P_{a_1,a_2,a_3,a_4}$ can be computed, for fixed values of the distances $a$, $x$ and $l$, via a suitable modification of the FK representation of the transfer matrix in which certain clusters (viz., the ones touching one or more of the points $w_1,w_2,w_3,w_4$) carry specific marks. The spectrum of this modified transfer matrix is contained within that of the original one, namely the one described in section~\ref{sec:Potts-model} in terms of affine TL representations. The spectrum can be proven to be real, so we can order the {\em distinct} eigenvalues as $\Lambda_0 > \Lambda_1 > \cdots > \Lambda_i > \cdots$. The correlation function then takes the following form, for generic values of $Q$, \begin{equation} P_{a_1,a_2,a_3,a_4} = \sum_i A_i \left( \frac{\Lambda_i}{\Lambda_0} \right)^l \,, \label{PAeval} \end{equation} where the amplitudes $A_i = A_i(a,x,L)$ are to be determined. The corresponding expression for non-generic values has the same form, but with the replacement \begin{equation} A_i \longrightarrow \sum_{j=0}^{r_i-1} A_i^{(j)} l^j \,, \label{generalised_amplitudes} \end{equation} whenever the eigenvalue $\Lambda_i$ is associated with a Jordan block of rank $r_i$. In the latter case the generalised amplitudes $A_i^{(j)} = A_i^{(j)}(a,x,L)$ are again independent of $l$. \begin{remark} \label{rem:degeneracy} There is of course an exact degenerescence of the scaling dimension of a CFT operator with non-zero spin $\Delta-\bar{\Delta}$ and that of its conjugate (i.e., obtained by the exchange $\xi \to \bar{\xi}$). This is prefigured in the lattice discretisation by the exact degenerescence of the eigenvalues corresponding to eigenstates of $T$ with opposite non-zero lattice momenta $\pm m$ (cf.\ Appendix~\ref{sec:momentum_sectors}). Because of the regrouping of degenerate eigenvalues in (\ref{PAeval}) it should thus be remembered to divide the amplitude $A_i$ of such states by a factor of two when comparing to the CFT predictions (see Appendix~\ref{sec:appB} for many examples of this phenomenon). \end{remark} For finite $L$ the set of eigenvalues of $T$---and hence the number ${\cal N} \equiv \sum_i r_i$ of (generalised) amplitudes to be determined---is finite. It follows that the form (\ref{PAeval}) is an exact expression, not merely an asymptotic expansion. Therefore, to determine the amplitudes $A_i$---or the generalised amplitudes $A_i^{(j)}$ for the cases with Jordan blocks---for given separations $a$, $x$ and size $L$, it suffices in principle to numerically determine the spectrum $\{ \Lambda_i \}$, compute the correlation functions $P_{a_1,a_2,a_3,a_4}$ for ${\cal N}$ different values $l$, and to invert the linear system (\ref{PAeval}). In practice, of course, things are more complicated, and several remarks must be made (see Appendix~\ref{sec:practical-remarks}). The most important of those is that the magnitude of the terms in (\ref{PAeval}) decreases exponentially fast, in particular when ${\cal N}$, and hence $l$, is large. Therefore both $\{ \Lambda_i \}$ and $P_{a_1,a_2,a_3,a_4}$ must be computed to an exceedingly high numerical precision. For instance, our most demanding computation (see section~\ref{sec:big-computation} for details) required a 4000-digit numerical precision. Another remark is that when ${\cal N} \gg 1$ we often wish to determine only the first ``few'' amplitudes (corresponding to $i \le$ some $i_{\rm max}$). This can be done by using the expression (\ref{PAeval}), truncated to the first $i_{\rm max}$ terms, as an asymptotic expression, i.e., by solving for $A_i$ the linear system provided by the numerically computed $P_{a_1,a_2,a_3,a_4}$ with $l = l_{\rm min}, l_{\rm min}+1,\ldots,l_{\rm min}+i_{\rm max}$, where $l_{\rm min}$ is taken sufficiently large. One then has to carefully check that the desired $A_i$ are stable, within the desired numerical precision, to small changes in $l_{\rm min}$. \subsubsection{Second method} \label{sec:2nd_method} Our other method applies to the computation of the order-parameter correlators $G_{a_1,a_2,a_3,a_4}$, defined in (\ref{order-param-corr}). This requires another variant of the FK transfer matrix with marked clusters, as described in details in Appendix~\ref{sec:appA2}. The number of different marks allowed must be chosen as the number of different symbols among $a_1,a_2,a_3,a_4$, and the dimension of the transfer matrix grows with this number. In practice we have employed two different marks, in order to gain access to the four correlators $G_{aaaa}$, $G_{aabb}$, $G_{abba}$ and $G_{abab}$. The spin operator ${\cal O}_a(\sigma_k)$, defined in (\ref{spin-op}), can be expressed within this basis and has essentially the effect of attributing the label $a$ to the spin situated at vertex $k \in V$. Our geometrical setup is such that vertices $w_1$ and $w_2$ belong to the same time slice, while $w_3$ and $w_4$ belong to another time slice, with the relative positions within these two time slices being specified by Figure~\ref{fig:cylinder}. We henceforth denote the spin operators simply by ${\cal O}_{a_k}$, for $k=1,2,3,4$, and keep implicit their point of insertion in the relevant time slices. Let $\langle v_i |$ and $| v_i \rangle$ denote the left and right eigenvectors of $T$. In the case of simple eigenvalues we then have \begin{subequations} \begin{eqnarray} \langle v_i | T &=& \Lambda_i \langle v_i | \,, \\ T | v_i \rangle &=& \Lambda_i | v_i \rangle \,. \end{eqnarray} \end{subequations} But even for the generic values of $Q$ that we consider here, some of the eigenvalues are degenerate, due to symmetries of the lattice and of the order parameter symbols $a_k$. In that case we denote the multiplicity of $\Lambda_i$ by $d_i$, and we endow the corresponding eigenvectors with an extra label, $| v_{i,j} \rangle$, where $j=1,2,\ldots,d_i$. The left and right eigenvectors can be obtained efficiently within an iterative diagonalisation scheme, such as the Arnoldi method (see again Appendix~\ref{sec:practical-remarks} for details). Left and right vectors will obviously be orthogonal if they correspond to different eigenvalues, but the Arnoldi method does not guarantee orthogonality within the degenerate subspaces. It is however possible to perform an additional diagonalisation step that will ensure that the orthogonality holds with respect to both labels: \begin{equation} \langle v_{i,j} | v_{i',j'} \rangle \propto \delta_{i,i'} \delta_{j,j'} \,. \label{orthogonality} \end{equation} \begin{remark} Some readers are likely to be well acquainted with the representation theory of the TL algebra, in which ``geometrical'' scalar products (see e.g.\ \cite{JS_combi,dGJP}) are introduced between basis states (often called link patterns in the context of the loop representation). These scalar products ``count'' loops and clusters formed by the gluing of the states, leading to $Q$-dependent results. We must therefore stress that the scalar products appearing in this section are simply the standard Euclidean scalar products between ordinary vectors. Moreover, all eigenvectors turn out to be real, so there is no issue of complex conjugation. \end{remark} We now claim that the amplitudes $A_i$ for the generic case without Jordan blocks can be obtained as \begin{equation} A_i = \sum_{j=1}^{d_i} \frac{ \langle v_0 | {\cal O}_{a_3} {\cal O}_{a_4} | v_{i,j} \rangle \, \langle v_{i,j} | {\cal O}_{a_1} {\cal O}_{a_2} | v_0 \rangle} {\langle v_0 | v_0 \rangle \, \langle v_{i,j} | v_{i,j} \rangle} \,. \label{scalar_prod_method} \end{equation} It is crucial for the validity of this result that the orthogonalisation in degenerate subspaces has been performed (see Appendix~\ref{sec:appA2}). We have performed extensive checks that the first and second methods give the same results, in situations where they are both applicable. The advantage of the second method is that it is numerically much more efficient, and hence gives access to larger sizes $L$. The sources for this gain in efficiency are explained in Appendix~\ref{sec:appA2}. Formula (\ref{scalar_prod_method}) has a nice geometrical interpretation that makes direct contact with Figure~\ref{fig:cylinder}. Indeed the numerator of the formula (read from right to left) and the figure (read from left to right) are completely analogous: \begin{enumerate} \item The propagation from the free boundary condition at imaginary time $t \to -\infty$ to the time slice containing $w_1$ and $w_2$ corresponds to the production of the ground state $| v_0 \rangle$. \item The first two operators are then inserted by ${\cal O}_{a_1} {\cal O}_{a_2}$. \item The piece $| v_{i,j} \rangle \, \langle v_{i,j} |$ corresponds to the projection on a definite state appearing in the $s$-channel of the four-point function. \item This is followed by the insertion of the two remaining operators, at $w_3$ and $w_4$. \item The projection on $\langle v_0 |$ matches the propagation to the free boundary condition at the other extreme of the cylinder ($t \to +\infty$). \end{enumerate} We stress that the validity of (\ref{scalar_prod_method}) depends crucially on the orthogonality (\ref{orthogonality}). \subsubsection{Continuum limit} \label{sec:num_cont_limit} To extract the continuum limit ($L \to \infty$) of the amplitudes it is important to be able to associate each $A_i$ with a definite field in the continuum limit.% \footnote{But see remark~\ref{rem:degeneracy} above.} For instance, one question that one might want to answer is what would be the amplitude contributing to a given primary $\Phi_{\Delta \bar{\Delta}}$, i.e., to identify the $A_i$ that will converge to the amplitude of $\Phi_{\Delta \bar{\Delta}}$ in the expansion (\ref{mainexp}). Several remarks are in order in this respect. Obviously, it is the ratio between two amplitudes---rather than each amplitude taken individually---that is universal and hence related to CFT. Therefore we assume tacitly in what follows that each amplitude of interest is measured via its ratio to the one that gives the leading contribution to the considered correlation function. Moreover, the conformal mapping to the cylinder implies that we should correct the raw lattice result by a conformal factor, namely the powers of $\sin \frac{2 \pi a}{L}$ appearing in (\ref{mainexp}). Once this has been done, our main claim is that we are capable of analysing the numerical results so as to establish the convergence \begin{equation} A_i \to A_{\Phi_{\Delta \bar{\Delta}}} \mbox{ as } L \to \infty \,, \end{equation} where the conformal amplitude has been defined in (\ref{ampldef}). More generally, we can obtain the corresponding results for subdominant contributions from the conformal blocks, corresponding to the amplitude multiplying a term of the type $\xi^{\cal N} \bar{\xi}^{\bar{\cal N}}$ in the square bracket of (\ref{mainexp}). This is interpreted as the (total) amplitude of the descendents at level ${\cal N},\bar{\cal N}$ of the primary $\Phi_{\Delta \bar{\Delta}}$. The challenge involved in making this identification is to make sure that we possess enough information about the lattice model to unambiguously associate a given field in the continuum limit with its ``corresponding'' eigenvalue of the transfer matrix in finite size $L$. In the lattice model, each $A_i$ is unambiguously associated with the eigenvalue $\Lambda_i$. A careful study of the transfer matrix (see Appendix~\ref{sec:momentum_sectors}) enables us to attribute to each eigenvalue three labels $\ell,k,m$, formally restricting to a representation denoted $V_{\ell,k,m}$. The first label $\ell$ gives the number of propagating FK clusters, so in the notation of the standard modules $\AStTL{j}{z^2}$ we have $j = \ell$ for all $\ell \neq 1$, while $\ell = 1$ corresponds to $j = 0$. The second label $k$ is directly related with the momentum of through-lines, via $z^2=e^{2i\pi k/j}$. And finally the third label $m$ is the lattice momentum that gives directly the conformal spin, $h - \bar{h} = m$, at least if $L$ is large enough to accommodate the desired spin. While certainly very helpful, the three labels $\ell,k,m$ are not quite enough to determine which conformal block to associate with $A_i$, nor at which level ${\cal N},\bar{\cal N}$. Roughly speaking, the trouble is that the $k$'th smallest scaling dimension in the continuum limit does not necessarily correspond to the $k$'th largest eigenvalue of $T$ in finite size $L$. While this is certainly true for $L \gg 1$, there are numerous crossovers in finite size, and these have to be monitored carefully in order to make to correct identifications. How we overcame this delicate problem is described in Appendix~\ref{sec:specT_CFTlimit}. Finally, once the finite-size approximation $A_i$ to a given CFT amplitude has been determined, for several different sizes $L$, the numerical value of the latter is determined by finite-size extrapolation techniques. This is again discussed in Appendix~\ref{sec:specT_CFTlimit}. A large number of concrete applications of the entire method can be found in section~\ref{sec:results} and (with more details provided) in Appendix~\ref{sec:appB}. \section{Results} \label{sec:results} \subsection{Checks} Our approach, being based on properties of the lattice model, requires a careful control of the continuum limit. There are several aspects to this. The most obvious one is that, since we are studying four-point functions of a CFT, we should, ideally, have all ratios $|z_{ij}| \gg 1$ (where all distances are measured in units of the lattice spacing). On the cylinder, we have chosen to take points far apart along the cylinder axis, but placed, pairwise, on identical imaginary time slices. Since the cylinder widths $L$ are limited for technical reasons, this means that $|w_{12}|,|w_{34}|$ will be limited, in fact, to a few lattice spacings.% \footnote{We could take the two points within each pair to reside on different time slices, of course, but this would only allow us to get real parts of $w_{ij}$ bigger, with the imaginary parts similarly limited.} \medskip We first observe that the dependence $A(x)$ of the amplitudes on the shift $x$ in the space-like direction (see Figure~\ref{fig:cylinder}) between the two groups of points is in fact trivial. Taking into account that all amplitudes have been normalised as ratios with respect to the leading one, as well as remark~\ref{rem:degeneracy}, it is seen from (\ref{mainexp}) that \begin{equation} A(x) = \cos \left( \frac{2 \pi s x}{L} \right) A(0) \,, \label{amp_dep_x} \end{equation} where $s = \Delta - \bar{\Delta}$ denotes the conformal spin (which coincides with the lattice momentum, $m = s$). An alternative means of deriving (\ref{amp_dep_x}) goes through the inspection of (\ref{scalar_prod_method}). Imagine evaluating the first scalar product in the numerator in a geometry where the cylinder has been rotated by the amount $-x$. This rotation will re-align the second pair of operators ${\cal O}_{a_3} {\cal O}_{a_4}$ with the first pair ${\cal O}_{a_1} {\cal O}_{a_2}$, like in the computation of $A(0)$. The ground state $\langle v_0 |$ is obviously rotationally invariant, but an intermediate state $| v_{i,j} \rangle$ of lattice momentum $m \neq 0$ is not, and will therefore pick up a corresponding phase factor under the rotation. Summing this over the degenerate contributions $\pm m$ reproduces (\ref{amp_dep_x}). We have checked numerically on explicit examples that (\ref{amp_dep_x}) holds true exactly in finite size. As a matter of fact, determining numerically the dependence $A(x)$ is a convenient means of establishing the lattice momentum $m$ of a given state, complementary to the techniques explained in Section~\ref{sec:momentum_sectors}. \medskip A maybe more subtle aspect is that the lattice observables are not in general pure scaling fields. This means that the conformal field whose four-point functions we want to study, is identified on the lattice as the Potts spin operator only up to additional corrections (``higher (or excited) spin operators''), whose contributions become negligible only when all distances are once again much larger than the lattice spacing: put otherwise, measured four-point functions on the lattice are a mixture of four-point functions of pure scaling fields. For our approach to be useful, it is necessary to perform many tests in order to control these potential drawbacks. As discussed extensively in Appendix~\ref{sec:appB}, we have checked that, for the sizes we were able to access: \begin{itemize} \item{} The mixture of excited spin operators can be neglected (see Appendix~\ref{sec:appB1}); and \item{} The values of the extrapolated amplitudes $A_{\Phi_{\Delta\bar{\Delta}}}$---see (\ref{ampldef})---as well as those of the first few (in practice, a handful) subdominant contributions to each conformal tower, extracted via the method outlined in section~\ref{sec:algorithm} are in fine agreement with their exact CFT values in three exactly solvable cases ($Q=0,2,4$), which are treated respectively in Appendices~\ref{sec:appB-Q0}, \ref{sec:appB-Q2} and \ref{sec:appB-Q4}. In the most favourable cases the relative deviations are as small as $10^{-4}$. \end{itemize} Moreover, even for operators higher in the spectrum ${\cal S}$ where the lattice determination of amplitudes may not have fully converged, our approach, combined with the algebraic understanding of transfer matrix sectors, indicates unambiguously which coupling constants will remain non-zero in the scaling limit, even if error bars on their extrapolated values are not negligible. \subsection{The $s$-channel spectrum of $P_{abab}$ or $P_{abba}$: $F_{j,\mathrm{e}^{2i\pi p/M}}{\scriptstyle( M|j,~j\geq 2)}$, $j$ even} We first study the cases where (at least) two distinct clusters are forced to propagate between the two distinguished time slices in Figure~\ref{fig:cylinder}. Specifically, $P_{abab}$ is the probability that points $w_1,w_3$ and $w_2,w_4$ respectively belong to the same clusters. This quantity is also called $P_2$ in \cite{Ribault}. Similarly, $P_{abba}$ is the probability that $w_1,w_4$ and $w_2,w_3$ belong to the same clusters. These two correlation functions are depicted in Figure~\ref{fig:2clusters}. \begin{figure} \definecolor{hazygray}{gray}{0.9} \definecolor{darkergray}{gray}{0.83} \begin{center} \begin{tikzpicture}[scale=0.55] \draw[hazygray,fill,rotate=175,yshift=-63.0] (0,1.4) arc(0:360:4.0 and 0.3); \draw[hazygray,fill,rotate=175,yshift=-23.0] (-0.2,1.4) arc(0:360:3.6 and 0.3); \draw[black,thick] (0,0) arc (0:360:0.5 and 1.5); \draw[black,thick] (-0.5,1.5)--(8,1.5); \draw[black,thick] (-0.5,-1.5)--(8,-1.5); \draw[black,thick,dotted] (1.5,0) arc (0:360:0.5 and 1.5); \draw[black,fill] (1.42,0.8) circle (2pt); \draw[black,fill] (1.42,-0.8) circle (2pt); \draw (1.42,0.8) node[left] {$w_1$}; \draw (1.42,-0.8) node[left] {$w_2$}; \draw[black,thick,dotted] (6.5,0) arc (0:360:0.5 and 1.5); \draw[black,thick] (8.0,-1.5) arc (-90:90:0.5 and 1.5); \draw[black,thick,dashed] (8.0,1.5) arc (90:270:0.5 and 1.5); \draw[black,fill] (6.50,0.3) circle (2pt); \draw[black,fill] (6.27,-1.3) circle (2pt); \draw (6.50,0.3) node[right] {$w_3$}; \draw (6.27,-1.3) node[right] {$w_4$}; \draw (4,-1.5) node[below] {$P_{abab}$}; \end{tikzpicture} \qquad \qquad \begin{tikzpicture}[scale=0.55] \draw[darkergray,fill] (3,-1.5)--(5,1.5)--(4,1.5)--(2,-1.5)--cycle; \draw[hazygray,fill,rotate=159,yshift=-73.0] (0,1.4) arc(0:360:3.6 and 0.3); \draw[hazygray,fill] (1.3,-0.4) -- (3,-1.5) -- (2,-1.5) -- (1.3,-1.0) .. controls (1.1,-0.8) and (0.8,-0.6) .. (1.3,-0.4); \draw[hazygray,fill] (4,1.5)--(5,1.5)--(7,0.5) .. controls (7.2,0.4) and (6.7,0.0) .. (6.0,0.1) -- cycle; \draw[black,thick] (0,0) arc (0:360:0.5 and 1.5); \draw[black,thick] (-0.5,1.5)--(8,1.5); \draw[black,thick] (-0.5,-1.5)--(8,-1.5); \draw[black,thick,dotted] (1.5,0) arc (0:360:0.5 and 1.5); \draw[black,fill] (1.42,0.8) circle (2pt); \draw[black,fill] (1.42,-0.8) circle (2pt); \draw (1.42,0.8) node[left] {$w_1$}; \draw (1.42,-0.8) node[left] {$w_2$}; \draw[black,thick,dotted] (6.5,0) arc (0:360:0.5 and 1.5); \draw[black,thick] (8.0,-1.5) arc (-90:90:0.5 and 1.5); \draw[black,thick,dashed] (8.0,1.5) arc (90:270:0.5 and 1.5); \draw[black,fill] (6.50,0.3) circle (2pt); \draw[black,fill] (6.27,-1.3) circle (2pt); \draw (6.50,0.3) node[right] {$w_3$}; \draw (6.27,-1.3) node[right] {$w_4$}; \draw (4,-1.5) node[below] {$P_{abba}$}; \end{tikzpicture} \end{center} \caption{FK cluster configurations that contribute to the correlation functions $P_{abab}$ and $P_{abba}$.} \label{fig:2clusters} \end{figure} \subsubsection{Results in finite size} It is evident from Figure~\ref{fig:2clusters} that the leading term in this sector should be the term corresponding to the propagation of two different clusters, that is, four cluster boundaries. The affine TL modules $\AStTL{j}{z^2}$ (or their continuum counterparts $F_{j,z^2}$) correspond to $2j$ through-lines, so the propagation of two clusters must have a contribution with $j=2$. The corresponding generating function of levels has two sectors, depending on whether a pair of boundaries going around the system picks up a phase $z^2=1$ or $z^2=-1$. No other choice is possible since for two pairs of boundaries (picking up a phase $z^4$), we do not want a phase.% \footnote{The reader might wonder why we have used the parameter $z$ since only $z^2$ seems to appear in the discussions. This is in part to conform with the literature, although $z$ itself may also have meaning for other questions in the Potts model. For instance, the single shift operator $u$ appplied to the system amounts to performing a duality transformation.} Hence we expect the contribution of modules \begin{equation} \AStTL{2}{1}\oplus\AStTL{2}{-1} \,. \label{2modcontr} \end{equation} We have first checked that {\em all} the eigenvalues associated with these two modules contribute to the probabilities $P_{abab}$ and $P_{abba}$ for all finite sizes. This is however not all. Sectors with a higher number of clusters than the two imposed by the choice of indices might also be thought to contribute to these correlation functions. One could think of several mechanisms for such contributions. First, there might be more clusters, distinct from the two imposed by the boundary conditions, that ``by chance'' connect the two time slices. Second, the two clusters might have more complicated topologies, with for instance one of them (say, the one containing points $w_1$ and $w_3$ in the left part of Figure~\ref{fig:2clusters}) starting out at one insertion point (here $w_1$), and wrapping all around the other cluster (containing points $w_4$ and $w_2$), before arriving at its terminal point (here $w_3$). We have found it difficult to provide a convincing argument that certain subclasses of configurations will necessarily lead to further contributions to the correlation functions, in terms of the modules $\AStTL{j}{z^2}$; we think there are underlying symmetry and branching rules considerations that may answer this riddle on general grounds, and that we do not yet control. Fortunately the numerical results are completely clear. We find that, for all finite sizes, all eigenvalues associated with the modules $\AStTL{j}{\mathrm{e}^{2 i\pi p/M}},M|j$, with $j$ even, and only those, contribute to the probabilities $P_{abab}$ and $P_{abba}$. For $j=4$ for instance, this allows contributions from the momentum sectors $z^2=1,\exp(i\pi/4), \exp(i\pi/2),\exp(3i\pi/4)$, and thus the following modules, in addition to those of (\ref{2modcontr}), \begin{equation} \AStTL{4}{1}\oplus \AStTL{4}{i}\oplus \AStTL{4}{-1}\oplus \AStTL{4}{-i} \,. \end{equation} Note that for a given width $L$, the maximum value of $j$ is bounded from above, $j \le L$. We have checked that, as $L$ increases, higher values of $j$ start contributing to the probabilities, provided that the separation $2a$ between the insertion points is sufficiently large. More precisely, we have observed that: \begin{itemize} \item For $L=5$ and separation $2a=2$, all the eigenvalues with $j=2,4$ and none of the eigenvalues with $j=0,1,3$ contribute to the probabilities. \item Still for $L=5$, but diminishing to separation $2a=1$, the contributions from $j=4$ disappear. \item For $L=7$ and separation $2a=3$, the two probabilities get contributions from all eigenvalues with $j=2,4,6$ and none of the eigenvalues with $j=0,1,3,5$. \item Still for $L=7$, but diminishing to separation $2a=2$, the contributions from $j=6$ disappear. Diminishing further to $2a=1$, the contributions from $j=4$ disappear as well. \end{itemize} The above result is corroborated by a closer study of the spectrum of the transfer matrix described in section~\ref{sec:2nd_method}, namely the one that produces the correlation functions of order parameter operator. Its eigenvalues are precisely those corresponding to the modules (\ref{modules}) with $j$ even, while those corresponding to $j$ odd are not observed at all. Motivated by the above list of observations, we conjecture that in fact a given sector $j \in 2 \mathbb{N}$ contributes only when $2a \ge j/2$. \subsubsection{Non-zero coupling to the sector $j=6$} \label{sec:big-computation} The computation with $L=7$ and $2a=3$ establishing that the sector $j=6$ does contribute to the correlation functions $P_{\cal P}$ is the most demanding among all of those made for this paper. For the benefit of readers interested in computational aspects (and those wanting to scrupulously assess the validity of our conclusions) we wish to describe it in some more detail---other readers may wish to skip this section and resume the reading below. This computation was performed for the generic value $Q=\frac32$. There is a total of $3\,932$ distinct eigenvalues in the sectors with $j=0,2,4,6$, corresponding to all possible momenta. Using the methods of Appendix \ref{sec:momentum_sectors} we can classify them in sectors $V_{\ell,k,m}$ corresponding to $\ell$ propagating clusters with cluster momentum $k$ and lattice momentum $m$. Ordering all the eigenvalues as $\Lambda_1 > \Lambda_2 > \cdots > \Lambda_{3932}$, with $\Lambda_1$ being the dominant eigenvalue in $V_0$ (i.e., the ground state), the dominant eigenvalues in the sectors $V_1$, $V_2$, $V_4$ and $V_6$---which obviously have vanishing momenta, $k=m=0$---are respectively $\Lambda_2$, $\Lambda_5$, $\Lambda_{205}$ and $\Lambda_{2390}$. Suppose first that we considered some correlation function coupling to all of these eigenvalues, and we wished to isolate the amplitude corresponding to $\Lambda_{2390}$ by using the first method of Appendix~\ref{sec:appA1}. The ratio $r = \Lambda_{2390}/\Lambda_1 \simeq 3.451 \cdot 10^{-4}$, and since we need to determine 2\,390 coefficients $A_i$ in (\ref{PAeval}) we will need the same number of equations, obtained by choosing the distance $l = l_{\rm min}+1,\ldots,l_{\rm min}+2390$. We would need (at the very least) to take $l_{\rm min} = 100$ in order to be in the asymptotic regime. Then, since $r^{2490} \simeq 2.6 \cdot 10^{-8821}$, we see that the terms entering (\ref{PAeval}) would differ by almost nine thousand orders of magnitude, so allowing some margin for numerical instabilities we would have to compute the correlation function for (at least) 2\,500 different values of $l$ to a numerical precision of (at least) 10\,000 digits. This task is hopelessly impossible, given that the transfer matrix of Appendix~\ref{sec:appA1} is of dimension $\sim 10^6$ in this case. To do better, we need to consider a particular well-chosen combination of correlation functions, designed so that it decouples from a sufficient number of low-lying states in the spectrum. The quantity \begin{equation} P^* = P_{aaaa} + P_{aabb} + \frac{1}{Q-1} \left( P_{abab} + P_{abba} \right) \end{equation} is a such a good combination. On symmetry grounds, it decouples from the sectors with odd momenta $k$. The term $P_{aaaa}$ is rather easily checked to pick up contributions from the sectors $V_1$, $V_2$ and $V_4$ (and maybe higher values of $\ell$), whereas $P_{aabb}$ couples in addition to $V_0$. Meanwhile, $P_{abab}$ and $P_{abba}$ get contributions only from $V_2$ and $V_4$ (and maybe higher values of $\ell$); it is indeed clear that since these terms impose two long clusters (see Figure~\ref{fig:2clusters}) they cannot couple to $V_0$ and $V_1$. The surprising property of $P^*$ is now that, with the above choice of the two coefficients in its definition, all contributions from $V_1$ and $V_2$ disappear from the combination. In other words, $P^*$ couples to $V_0$, $V_{40}$, $V_{42}$, and maybe $V_{\ell,k}$ with higher values of $\ell$ and even $k$. We have ${\rm dim}\, V_0 = 232$, ${\rm dim} \, V_{40} = 190$ and ${\rm dim} \, V_{42} = 182$. Therefore we shall be able to settle whether there is a non-zero amplitude for the 6-cluster sector provided we can look beyond the first $604$ eigenvalues. To that end, we have computed $P^*$ for $l=100,101,\ldots,900$ to be on the safe side. Noting that $r^{900} \simeq 1.3 \cdot 10^{-3116}$ we have performed the computations to a numerical precision of $4\,000$ digits. This required about $3 \times 10^5$ hours of single-processor CPU time. The conclusion is that we have unambiguously established that $P^*$ picks up non-zero contributions from the first few eigenvalues in each of the sectors $V_{60}$, $V_{62}$ and $V_{64}$, with amplitudes in the range $\sim 10^{-10}$. While these numbers may seem small, they follow the clear trend (observed in all cases) that the non-zero amplitudes decay exponentially with the index of the corresponding eigenvalue. Moreover, these amplitudes are numerically stable towards changing $l_{\rm min}$ throughout the range $l_{\rm min} \in [100,200]$. We have also checked that the absolute values of contributions which are genuinely supposed to be zero (such as those from sectors $V_{61}$, $V_{63}$ and $V_{65}$) come out numerically as $\ll 10^{-500}$, which is fully compatible with the above estimates of the required numerical precision. \subsubsection{Exponents} Introducing our usual notation $F_{j,z^2}$, the spectrum in the sector with $j=2$ propagating clusters is encoded in the generating functions \begin{equation} F_{2,1}\oplus F_{2,-1} \end{equation} This corresponds to the conformal weights given by (\ref{bspec1}): \begin{subequations} \begin{eqnarray} (h_{e,-2};h_{e,2}) \,, & & e\in \mathbb{Z} \,, \\ (h_{e+1/2,-2};h_{e+1/2,2}) \,, & & e\in \mathbb{Z} \,. \end{eqnarray} \end{subequations} Note that for the first part of the spectrum, $h-\bar{h}$ is an even integer, while it is an odd integer for the first part. We will denote these two contributions by $2S$ and $2A$ respectively, where $S$ stands for symmetric and $A$ for antisymmetric. Going back to an earlier remark, this means that we should have, for primary fields \begin{subequations} \begin{eqnarray} \Delta-\bar{\Delta} &=& \hbox{even in S part} \,, \\ \Delta-\bar{\Delta} &=& \hbox{odd in A part} \,. \end{eqnarray} \end{subequations} We thus have, for the $2S$ part \begin{subequations} \begin{eqnarray} h_{e,2} &=& {[(m+1)e-2m]^2-1\over 4m(m+1)} \,, \\ h_{e,-2} &=& h_{e,2}+2e \,, \end{eqnarray} \end{subequations} while for the $2A$ part \begin{subequations} \begin{eqnarray} h_{e+1/2,2} &=& {[(m+1)e+{1-3m\over 2}]^2-1\over 4m(m+1)} \,, \\ h_{e+1/2,-2} &=& h_{e+1/2,2}+2e+1 \,. \end{eqnarray} \end{subequations} Similarly, in the sector with $j=4$ propagating clusters we have \begin{equation} F_{4,1}\oplus F_{4,i}\oplus F_{4,-1}\oplus F_{4,-i} \end{equation} The corresponding conformal weights from (\ref{bspec1}) are \begin{subequations} \begin{eqnarray} (h_{e,-4};h_{e,4}) \,, & & e\in\mathbb{Z} \,, \\ (h_{e+1/4,-4};h_{e+1/4,4}) \,, & & e\in\mathbb{Z} \,, \\ (h_{e+1/2,-4};h_{e+1/2,4}) \,, & & e\in\mathbb{Z} \,, \\ (h_{e+3/4,-4};h_{e+3/4,4}) \,, & & e\in\mathbb{Z} \,, \end{eqnarray} \end{subequations} but the fourth set is identical to the second one. As the number of clusters increases, so does the number of allowed sectors. Our finite-sizes observations are clearly in favour of the extension of this pattern, invariably with all even values of $j$. Clearly, we find exponents $(h_{r,s},h_{r,-s})$ in the $s$-channel with $r\in \mathbb{Z}$ and $s\in2\mathbb{Z}$, and $r\in \mathbb{Z}+1/2,s\in 2\mathbb{Z}$. We will refer to these exponents as sets $S_{\mathbb{Z},2\mathbb{Z}}$ and $S_{\mathbb{Z}+1/2,2\mathbb{Z}}$, in analogy with \cite{Ribault}. These sub-spectra arise from the modules $\AStTL{j}{1}$ and $\AStTL{j}{-1}$ with $j$ even. But we find that the $s$-channel, in finite size at least, contains many more exponents, arising from phases $z^2\neq 1,-1$ and, in terms of exponents, corresponding to rational values of the first label with higher denominators, such as those with first label $e+1/4$. The next key question is whether some sort of simplification might occur in the scaling limit---for instance, whether some sectors that contribute to the probabilities in finite size might do so with amplitudes that go to zero as $L\to\infty$. We have seen absolutely no evidence of this. To make the point as clear as possible, we illustrate it on the case of the antisymmetric combination $P_{abab}-P_{abba}$. \subsubsection{Amplitudes and the antisymmetric combination $P_{abab}-P_{abba}$} The antisymmetry of the combination implies that only modules with $z^{j}=-1$ contribute, which translates into primaries with $h-\bar{h}$ an odd number---what we have called earlier the $j$ even, A sectors. Let us now focus on how this contributes to $P_{abab}-P_{abba}$. We have the first fields at spin $|h-\bar{h}|=1,3$ with weights $(h_{1/2,\mp2},h_{1/2,\pm2})$ and $(h_{3/2,-\mp 2},h_{3/2,\pm2})$ in the sector ${\cal S}_{\mathbb{Z}+{1\over 2},2\mathbb{Z}}$. But according to our earlier analysis we also expect contributions from, in particular, $(h_{1/4,\mp4},h_{1/4,\pm4})$. To make things concrete, we can take for instance $Q=1/2$ (so $m$ is irrational), in which case we find \begin{subequations} \begin{eqnarray} (h_{1/2,-2},h_{1/2,2}) &=& (1.156405\cdots,0.156405\cdots) \,, \\ (h_{3/2,-2},h_{3/2,2}) &=& (2.969378\cdots,-0.030621\cdots) \,, \\ (h_{1/4,-4},h_{1/4,4}) &=& (2.925269\cdots,1.925269\cdots) \,. \end{eqnarray} \end{subequations} We observe that the field with $(h_{1/2,-2}+2,h_{1/2,2})$ has total dimension larger than $(h_{3/2,-2},h_{3/2,2})$. Therefore, at momentum 3, the field $(h_{3/2,-2},h_{3/2,2})$ will be the first contribution, and so will be $(h_{1/2,-2},h_{1/2,2})$ at momentum one. It is therefore very easy to identify the corresponding contributions to the four-point function: % \begin{eqnarray} P_{aabb}-P_{abba}\propto (z\bar{z})^{-2h_{1/2,0}}\left( A_{\Phi_{h_{1/2,-2},h_{1/2,2}}}z^{h_{1/2,-2}}\bar{z}^{h_{1/2,2}}+ A_{\Phi_{h_{3/2,-2},h_{3/2,2}}}z^{h_{3/2,-2}}\bar{z}^{h_{3/2,2}}+ \ldots\right.\nonumber\\ \left.A_{\Phi_{h_{1/4,-4},h_{1/4,4}}}z^{h_{1/4,-4}}\bar{z}^{h_{1/4,4}}+\ldots\right) \,. \end{eqnarray} Since $m$ is irrational, there is no mixing in the conformal mapping, and we have on the cylinder \begin{eqnarray} P_{aabb}-P_{abba}\propto A_{\Phi_{h_{1/2,-2},h_{1/2,2}}}\left(4\sin^2{2\pi a\over L}\right)^{h_{1/2,-2}+h_{1/2,2}}\xi^{h_{1/2,-2}}\bar{\xi}^{h_{1/2,2}}+\ldots\nonumber\\ +A_{\Phi_{h_{3/2,-2},h_{3/2,2}}}\left(4\sin^2{2\pi a\over L}\right)^{h_{3/2,-2}+h_{3/2,2}}\xi^{h_{3/2,-2}}\bar{\xi}^{h_{3/2,2}}+\ldots\nonumber\\ +A_{\Phi_{h_{1/4,-4},h_{1/4,4}}}\left(4\sin^2{2\pi a\over L}\right)^{h_{1/4,-4}+h_{1/4,4}}\xi^{h_{1/4,-4}}\bar{\xi}^{h_{1/4,4}}+\ldots \end{eqnarray} To restate the obvious, what we do then is measure the combination of probabilities on the left, identify the various terms on the right (via the exponential $l$-dependence of $\xi, \bar{\xi}$), and account for the geometrical factors (the powers of $4\sin^2 {2\pi a\over L}$) to extract, for a given sizes $L$, an estimate of the amplitudes. \subsubsection{The $(1/4,\mp 4)$ amplitude} We give in figure (\ref{quarteramp}) the results for the ratio $A_{\Phi_{h_{1/4,-4},h_{1/4,4}}}/A_{\Phi_{h_{1/2,-2},h_{1/2,2}}}$ as a function of $Q$ for various sizes $L$. While this amplitude is small (amplitudes typically decay very fast with the dimension of the associated primaries), it is clearly not zero in general, nor does it show any indication of going to zero as $L$ increases. We note however that, for all finite sizes, $A_{\Phi_{h_{1/4,-4},h_{1/4,4}}}=0$ for $Q=0,3,4$. This is well expected, as discussed in the Appendices \ref{sec:appB-Q0} and \ref{sec:appB-Q4} in particular. We find on the other hand that $A_{\Phi_{h_{1/4,-4},h_{1/4,4}}}\neq 0$ for $Q=2$ (cf.\ Appendix~\ref{sec:appB-Q2}). \begin{figure}[ht] \begin{center} \includegraphics[width=12cm]{Plot-ratio1} \caption{The ratio $A_{\Phi_{h_{1/4,-4},h_{1/4,4}}}/A_{\Phi_{h_{1/2,-2},h_{1/2,2}}}$ as a function of $Q$ for $L=5,6,7,8$ (blue, orange, green and red dots respectively). This ratio is generically non zero. It exhibits (in finite size) simple poles at $Q=4 \cos^2{\pi\over 8}$, $Q=4\cos^2{4\pi\over 8}$, and vanishes exactly (in finite size) for $Q=0,3,4$. }\label{quarteramp} \end{center} \end{figure} While the amplitude is small in general, it is found to become large---nay divergent---for two special values: \begin{subequations} \begin{eqnarray} Q &=& 4\cos^2{\pi\over 8} = 3.414213\cdots \,, \\ Q &=& 4\cos^2{3\pi\over 8} = 0.585786\cdots \,. \end{eqnarray} \end{subequations} There are several ways to understand this. We will discuss a CFT analysis in the conclusion. From the lattice point of view, the divergence arises because the transfer matrix exhibits a Jordan cell in the lowest level of $\AStTL{4}{\pm i}$. This Jordan cell arises from representation theory of the Jones algebra for $\mathfrak{q}=e^{i\pi/8},\mathfrak{q}=e^{3i\pi/8}$. To illustrate this, take for instance the case $\mathfrak{q}=e^{3i\pi/8}$. The module $\AStTL{2}{-1}$ becomes reducible for this value of $\mathfrak{q}$, and admits a sequence of submodules as represented in Figure~\ref{figmodjor}. The presence of submodules $\AStTL{4}{i},\AStTL{4}{-i}$ (in particular) suggests\footnote{While the structure of modules $\AStTL{j}{z^2}$ in degenerate cases is well under control, what happens here is the glueing of two standard modules for $\mathfrak{q}$ a root of unity. The understanding of which modules glue with which ones for a given transfer matrix is a bit more complicated, and involves more representation theory; see \cite{GRSV1} for a discussion of this point.} that excited states in $\AStTL{2}{-1}$ (a module with two propagating clusters) mix with states in $\AStTL{4}{i},\AStTL{4}{-i}$ within the module involving four propagating clusters. This mixture leads to Jordan cells in the transfer matrix. As shown in (\ref{generalised_amplitudes})---and further discussed in Appendix~\ref{sec:appB-Q0} in the case $Q=0$---a Jordan cell in turns translates into a contribution to the correlation function that is linear in (imaginary) time on the cylinder. This, finally, corresponds formally to an infinite amplitude. \begin{figure}[ht] \begin{center} \includegraphics[width=4cm]{crudefig4.pdf} \caption{Sub-module structure of $\AStTL{2}{-1}$ for $\mathfrak{q}=e^{3i\pi/8}$. Note the appearance of sub-modules isomorphic to $\AStTL{4}{\pm i}$, which lead to glueing of standard modules into bigger, indecomposable modules, and Jordan cells for the transfer matrix.}\label{figmodjor} \end{center} \end{figure} \subsection{The $s$-channel spectrum of $P_{aaaa}$: $F_{j,\mathrm{e}^{2i\pi p/M}}{\scriptstyle( M|j,~j\geq 2)}$, $j,jp/M$ even and $F_{0,-1}$ } We now turn to $P_{aaaa}$: this is the probability that all four points belong to the same cluster. It is called $P_0$ in \cite{Ribault}. For all finite sizes, we find that the modules $\AStTL{j}{e^{2i\pi p/M}}$ with $M|j,~j\geq 2$ and $j$ even contribute when $jp/M$ is even: this corresponds to the sectors with an even number of clusters propagating, and values of $z$ obeying $z^j=1$, what we have called earlier the $j$ even, S sectors. Geometrically, these contributions arise from configurations where for instance the points $1,3$ and $2,4$ are joined by two clusters which are only connected outside of the interval between their two (imaginary) time slices. On top of this, we also have the contribution where the four points belong to a single cluster arising between their two (imaginary) time slices. As discussed earlier, having a cluster propagating along the cylinder does not imply that there are boundaries around the cluster. The corresponding module of the Jones algebra is thus not a module with $j=1$: rather, it occurs as $\AStTL{0}{-1}$, i.e., as a module with no through-lines, but for which non-contractible loops (which would cut the connection between $1,2$ and $3,4$) are forbidden. Like for $P_{abab}$ and $P_{abba}$ we find that all eigenvalues in these modules do contribute in finite size, and that none of the amplitudes seem to vanish as $L\to \infty$. This suggests that the spectrum of critical exponents is given by $F_{j,\mathrm{e}^{2i\pi p/M}}{\scriptstyle( M|j,~j\geq 2)}$, $j,jp/M$ even and $F_{0,-1}$. In the two clusters ($j=2$) sector, this leads to \begin{equation} (h_{e,-2};h_{e,2}),~~ e\in\mathbb{Z} \end{equation} while in the four clusters ($j=4$) sector we find \begin{eqnarray} (h_{e,-4};h_{e,4}),~~ e\in\mathbb{Z}\nonumber\\ (h_{e+1/2,-4};h_{e+1/2,4}),~~ e\in\mathbb{Z} \end{eqnarray} These two contributions occur as well in $S_{\mathbb{Z}+1/2,2\mathbb{Z}}$. New contributions appear for higher even values of $j$. For instance we find also \begin{eqnarray} (h_{e,6};h_{e,-6}),~~ e\in\mathbb{Z}\nonumber\\ (h_{e\pm1/3,6};h_{e\pm1/3,-6}),~~ e\in\mathbb{Z} \end{eqnarray} On top of this we have the `one-cluster sector', which is described by $F_{0,-1}$ (i.e., non-contractible loops are killed). This corresponds to the set of conformal weights \begin{equation} (h_{e+1/2,0};h_{e+1/2,0}),~~e\in\mathbb{Z} \end{equation} which is also in $S_{\mathbb{Z}+1/2,2\mathbb{Z}}$. \subsection{The $s$-channel spectrum of $P_{aabb}$: $F_{j,\mathrm{e}^{2i\pi p/M}}{\scriptstyle( M|j,~j\geq 2)}$, $j,jp/M$ even, $F_{0,-1}$ and $\bar{F}_{0,\mathfrak{q}^2}$} The quantity $P_{aabb}$ is the probability for two ``short clusters'' (as opposed to the ``long clusters'' shown in Figure~\ref{fig:2clusters}): points $1,2$ belonging to one cluster, points $3,4$ to the other. It is called $P_1$ in \cite{Ribault}. We find that all the eigenvalues occurring in $P_{aaaa}$ also contribute to $P_{aabb}$. On top of these, we also find the eigenvalues from the module $\bAStTL{0}{\mathfrak{q}^2}=\bAStTL{0}{\mathfrak{q}^{-2}}$. This module corresponds to a sector with no (forced) propagating cluster, which is obtained simply by giving non-contractible loops their bulk weight. As usual now, none of the corresponding amplitudes seem to vanish in the limit $L\to\infty$. The operator content from $\bAStTL{0}{\mathfrak{q}^{\pm2}}$ involves diagonal primaries, with weights \begin{equation} (h_{e+e_0,0};h_{e+e_0,0}),~~e\in\mathbb{Z} \,, \end{equation} where \begin{equation} h_{e+e_0,0}={[(m+1)(e+1)-1]^2-1\over 4m(m+1)} \,. \end{equation} Of course this is the same set as the set \begin{equation} (h_{e,1};h_{e,1}),~~e\in\mathbb{Z} \,, \end{equation} after a shift of the electric charge. We will denote this set as $\mathcal{S}^d_{\mathbb{Z},1}$. \subsection{Summary} We can now summarise our spectra in the $s$-channel \begin{align} \setlength{\arraycolsep}{6mm} \renewcommand{\arraystretch}{1.2} \begin{array}{c|ll} & \mbox{$s$-channel} & \mbox{Parities} \\ \hline P_{aaaa} & \mathcal{S}_1\equiv\mathcal{S}_{\mathbb{Z}+1/2,0}\cup\{\mathcal{S}_{\mathbb{Z}+{p\over M},j}\} & j \in 2 \mathbb{Z}, \ jp/M\hbox{ even} \\ P_{aabb} &\mathcal{S}_2\equiv \mathcal{S}^d_{\mathbb{Z},1}\cup \mathcal{S}_{\mathbb{Z}+1/2,0}\cup\{\mathcal{S}_{\mathbb{Z}+{p\over M},j}\} & j \in 2 \mathbb{Z}, \ jp/M\hbox{ even} \\ P_{abab/abba} & \mathcal{S}_3\equiv \{\mathcal{S}_{\mathbb{Z}+{p\over M},j} \}& j \in 2 \mathbb{Z}, \ jp/M\hbox{ integer} \\ \hline \end{array} \end{align} where we have allowed $j$ to take positive or negative values, since the sets of exponents are invariant under $j\to -j$. Recall that e.g. the set $\mathcal{S}_{\mathbb{Z},2\mathbb{Z}}$ refers to {\sl pairs} of exponents $(h_{r,s},h_{r,-s})$ with $r\in \mathbb{Z},s\in2\mathbb{Z}$, while $\mathcal{S}^d_{\mathbb{Z},1}$ denotes pairs $(h_{r,1},h_{r,1})$, with $r\in\mathbb{Z}$. Recall also that $p,M$ are coprime integers, and that the value $p=0$ in particular is allowed. The case ${p\over M}={1\over 2}$ appears already in \cite{Ribault}. Note that these are the generic results, i.e., those valid for $m$ irrational. Some contributions vanish for special values of $Q$, such as $Q=0$, $Q=2$ and $Q=4$ (see Appendix~\ref{sec:appB}) and in some cases Jordan blocks appear. The spectra in the other channels follow from simple geometrical considerations: \begin{align} \setlength{\arraycolsep}{6mm} \renewcommand{\arraystretch}{1.2} \begin{array}{c|llc} & \mbox{$t$-channel} & \mbox{$u$-channel} \\ \hline P_{aaaa}& \mathcal{S}_1 & \mathcal{S}_1 \\ P_{aabb} &\mathcal{S}_3 & \mathcal{S}_3 \\ P_{abab} & \mathcal{S}_3 & \mathcal{S}_2 & \\ P_{abba} & \mathcal{S}_2 & \mathcal{S}_3 & \\ \hline \end{array} \end{align} An important property of our spectra in the case of $P_{aabb}$ is that only states with positive conformal weights propagate along the cylinder: no ``effective central charge'' appears, despite the non-unitarity of the CFT. This is contrast with what would be observed, for instance, in the case of minimal models corresponding to $m+1\equiv {p\over p'}$, $p,p'$ integer, where the effective ground state with $c_{\rm eff}=1-{6\over pp'}$ would appear. It is our understanding that a similar phenomenon takes place in the conjectured expressions of \cite{Ribault}. \section{Comparison with results in \cite{Ribault}} \label{sec:comp-ribault} The comparison with the proposal in \cite{Ribault} requires some discussion, since the authors in this reference did not, in particular, provide conjectured results for $P_{aaaa}$. The simplest quantity to consider is $P_{abab}-P_{abba}\equiv P_2-P_3$ in the notations of that reference. Indeed, from eq. (3.2) in \cite{Ribault} \begin{equation} R_\sigma=\lambda(P_0+\mu P_\sigma) \end{equation} we see that, in their notations, $R_2-R_3=\lambda\mu(P_2-P_3)$. The spectrum in the $s$-channel of $P_2-P_3=P_{abab}-P_{abba}$, according to our analysis, is made of the fields $(h_{e+p/M,-j};h_{e+p/M,j})$ for $e\in \mathbb{Z}$, with $j$ even and $pj/M$ an odd integer. Note that all these fields have $h-\bar{h}$ odd. In \cite{Ribault}, meanwhile, the spectrum is $S_{\mathbb{Z}+1/2,2\mathbb{Z}}$ (after switching indices in\cite{Ribault} to make their conventions the same as ours), restricted like for us to odd spin $h-\bar{h}$. So for instance the field with weights $(h_{1/4,\mp 4};h_{1/4,\pm 4})$ for which we have seen that the amplitude is generically non-zero, is absent in the solution proposed in \cite{Ribault}. This suggests that their solution is, generically, not the correct one, and that {\bf an infinity of fields is missing in their proposal}. We illustrate this qualitatively in Figures \ref{figexpmiss},\ref{figexpmiss1}. \begin{figure}[ht] \begin{center} \includegraphics[width=10cm]{NewPlotExponents} \caption{A sample of the full spectrum in the s-channel for $P_{abab}-P_{abba}$ represented by the pairs $(r,s)$ of the $h_{r,s}$ exponents ($r$ is on the y-axis, and $s$ on the x-axis. The spectrum considered in \cite{Ribault}, depicted as crosses, is seen to be a tiny subset of the full spectrum: the projections of the dots on the y-axis in fact should cover it densely (represented here are exponents for $M=2,4,6,8,10,12$ only).}\label{figexpmiss} \end{center} \end{figure} \begin{figure}[ht] \begin{center} \includegraphics[width=3cm]{Dense} \caption{The projection on the vertical axis of the exponents represented in Figure \ref{figexpmiss}}\label{figexpmiss1} \end{center} \end{figure} Meanwhile, it is fascinating to compare results for amplitudes that are predicted in \cite{Ribault} and which are also found to occur in our analysis. A good example of this is the first amplitudes for the sector with $j=2$, namely $A_{\Phi_{h_{3/2,-2},h_{3/2,2}}}$ and $A_{\Phi_{h_{1/2,-2},h_{1/2,2}}}$. The bootstrap in \cite{Ribault} produces amplitudes which are in fact simply related with those of Liouville field theory at $c<1$, and thus admit analytical expressions \cite{EstienneIkhlef,RibaultMigliaccio}. In particular, their conjecture is% \begin{equation} {A_{\Phi_{h_{3/2,-2},h_{3/2,2}}}\over A_{\Phi_{h_{1/2,-2},h_{1/2,2}}}}= 2^{-\frac{4}{\beta^2}} \frac{\Gamma(\frac32+\frac{1}{4\beta^2})} {\Gamma(\frac{1}{4\beta^2})} \frac{\Gamma(\frac32+\frac{3}{4\beta^2})}{\Gamma(1+\frac{3}{4\beta^2})} \frac{\Gamma(-1-\frac{1}{4\beta^2})}{\Gamma(\frac32-\frac{1}{4\beta^2})} \frac{\Gamma(2-\frac{3}{4\beta^2})}{\Gamma(\frac32-\frac{3}{4\beta^2})}\label{exactconj} \end{equation} where $\beta^2={m\over m+1}$, $\sqrt{Q}=2\cos{\pi\over m+1}$, $m\in[1,\infty]$. This conjecture reproduces results which are believed to be exact at $Q=0,3,4$---the result for $Q=0$ is discussed in our Appendix \ref{sec:appB-Q0}; the result for $Q=4$ follows from a work by A.\ Zamolodchikov (as discussed in \cite{Ribault}), and the result for $Q=3$ is unpublished work of R.\ Santachiara. Numerical results for this ratio are given in figure \ref{figconjsyl}. They are intriguingly close ---after reasonable extrapolation---to the formula (\ref{exactconj}). The agreement is worse near $Q=4$, but as commented elsewhere in this paper, this discrepancy can possibly be attributed to the presence of a marginal operator affecting corrections to scaling. We do not know whether (\ref{exactconj}) might actually be exact, or whether it is just very close to the exact result. Numerics, at this stage, do not really allow us to settle this issue. \begin{figure}[ht] \begin{center} \includegraphics[width=10cm]{Plot-ratio4} \caption{Results for the ratio $A_{\Phi_{h_{3/2,-2},h_{3/2,2}}}/ A_{\Phi_{h_{1/2,-2},h_{1/2,2}}}$, as a function of $Q$, for sizes $L=5,6,7,8,9,10,11$ corresponding to colours blue, orange, green, red, purple, mauve, clear blue. The points in purple are obtained by extrapolating data for $L=5,7,9,11$ and those in yellow by extrapolating data for $L=6,8,10$. The green curve is the conjecture (\ref{exactconj}).}\label{figconjsyl} \end{center} \end{figure} Meanwhile, the uncertainty of the numerical determination shown in Figure~\ref{figconjsyl} can be estimated from the difference between the extrapolations through even and odd system sizes $L$. Given that this uncertainty is (for most values of $Q$) comparable to the distance to the conjectured result (\ref{exactconj}) is certainly a strong motivation for further improving the numerical algorithm and gain access to a few more sizes. This could maybe be achieved if one could impose the sector and momentum constraints within our scalar product method (see Appendix~\ref{sec:appA2}). \section{Conclusion} \label{sec:conclusion} We believe that the numerical and algebraic evidence presented in this paper invalidates the results in \cite{Ribault}. This is a very intriguing conclusion, since, in particular, the authors of \cite{Ribault} presented Monte Carlo simulations of four-point functions in the plane that were in good agreement with their bootstrap prediction. It is possible that the conjecture in \cite{Ribault}, while not the correct answer to the problem of describing geometrical correlations in the Potts model, is indeed a solution to the bootstrap, and moreover captures numerically the essential features of the four-point functions, failing only at an accuracy, or for values of the cross-ratio $z$, not accessible using the Monte-Carlo approach. If this is the case, this raises several questions, in particular about the number of possible solutions to the bootstrap,% \footnote{Recall that there are cases where several solutions to the bootstrap are known to exist, for instance the Liouville theory at $c=1$ and the Runkel-Watts limit of minimal models \cite{RunkelWatts,RibaultSantachiara}.} and what, if anything, is truly described by the proposal in \cite{Ribault}. To shed more light on this issue, an obvious route is to build four-point functions following the methodology in \cite{Ribault} but based on our spectra. This is quite challenging technically, because of the large number of primary fields with dimensions of the same order of magnitude we would have to involve. Another, more fundamental aspect worth mentioning is that, in our spectrum, many of the conformal weights have degenerate values, with singular conformal blocks. It is not clear whether the regularisation procedure used in the bootstrap approach \cite{Ribault} is actually the relevant one for the $Q$-state Potts model. This, we believe, could be answered by numerical studies in the spirit of the present paper and \cite{KooSaleur}. A particularly intriguing fact is that we found numerically a ratio $A_{\Phi_{h_{3/2,-2},h_{3/2,2}}}/ A_{\Phi_{h_{1/2,-2},h_{1/2,2}}}$ which is not incompatible with the proposal in \cite{Ribault}; see Figure~\ref{figconjsyl}. It could be that the solution to the bootstrap relevant for the Potts model involves for this ratio a value close to the one in \cite{Ribault} and yet different, over the whole range $Q\in [0,4]$. It could also be that the amplitudes in \cite{Ribault}---which, to the best of our understanding, are actually given by standard formulae for Liouville at $c<1$, naively extended to the case of fields with $h\neq \bar{h}$---are exact, but that something has to be added. Adding ``something'' to the spectrum in \cite{Ribault} is definitely necessary if one wishes to avoid correlation functions with many singularities as $Q$ is varied. To see why this is the case, we consider the contributions to the antisymmetric combination of probabilities: including now higher order terms in the conformal blocks \begin{eqnarray} P_{aabb}-P_{abba} &\propto& (z\bar{z})^{-2h_{1/2,0}}\left( A_{\Phi_{h_{1/2,-2},h_{1/2,2}}}{\cal F}_{h_{1/2,-2}}^{(s)}\overline{{\cal F}}_{h_{1/2,2}}^{(s)}\right.+\nonumber\\ & & \left.A_{\Phi_{h_{3/2,-2},h_{3/2,2}}}{\cal F}_{h_{3/2,-2}}^{(s)}\overline{{\cal F}}_{h_{3/2,2}}^{(s)}+ A_{\Phi_{h_{1/4,-4},h_{1/4.4}}}z^{h_{1/4,-4}}\bar{z}^{h_{1/4,4}}+\ldots\right) \,. \label{basicform} \end{eqnarray} Note that we used here conformal blocks where the dependency $ z^{-2h_{1/2,0}}$ (resp. $\bar{z}^{-2h_{1/2,0}}$) has been factored out. When $Q\to 4\cos^2{3\pi\over 8}\equiv Q^*$, we find that $h_{3/2,2}\to h_{1,2}$, a degenerate value. The conformal block in the four-point function coming from \begin{equation} (h_{3/2,-2},h_{3/2,2})=(h_{3/2,-2},h_{1,2})=\left(-{1\over32}+3,-{1\over 32}\right) \end{equation} has a null-state at level 2 for the $\bar{z}$ components, with weights \begin{equation} (h_{3/2,-2},h_{1,-2})=\left(-{1\over32}+3,-{1\over 32}+2\right) \,. \end{equation} The appearance of the null-state means that the conformal block $\overline{{\cal F}}_{h_{3/2,2}}$ has a pole of the form ${1\over Q-Q^*}$ multiplying the term $z^{h_{3/2,-2}}\bar{z}^{h_{1,-2}}$. Setting \begin{equation} {\cal F}_{h_{3/2,-2}}^{(s)}\overline{{\cal F}}_{h_{3/2,2}}^{(s)}\approx \ldots+{r^*\over Q-Q^*}z^{h_{3/2,-2}}\bar{z}^{h_{3/2,2}+2}+\ldots \,, \end{equation} we see that the amplitude of the singular term in the bracket in (\ref{basicform}) is \begin{equation} A_{\Phi_{h_{3/2,-2},h_{3/2,2}}}{r^*\over Q-Q^*} z^{h_{3/2,-2}}\bar{z}^{h_{3/2,2}+2} \,. \end{equation} Meanwhile we observe that the weights for the singular term coincide with the weights from the $1/4$ sector: \begin{equation} (h_{1/4,-4},h_{1/4,4})=\left({-1\over 32}+3,{-1\over 32}+2\right) \,. \end{equation} Recall that we have found numerically that the amplitude of this field also has a simple pole when $Q\to Q^*$: \begin{equation} A_{\Phi_{h_{1/4,-4},h_{1/4,4}}}\approx {R\over Q-Q^*} \,, \end{equation} so the amplitude of the second singular term in (\ref{basicform}) is \begin{equation} {R\over Q-Q^*}z^{h_{1/4,-4}}\bar{z}^{h_{4,1/4}} \,. \end{equation} For technical reasons, we normalise all quantities by $A_{\Phi_{h_{1/2,-2},h_{1/2,2}}}$ (this amplitude is not expected to be singular), so we set \begin{equation} {A_{\Phi_{h_{1/4,-4},h_{1/4,4}}}\over A_{\Phi_{h_{1/2,-2},h_{1/2,2}}}} \approx {r\over Q-Q^*} \,. \end{equation} We find numerically that on ``resonance'', there is a Jordan cell of rank two mixing the two terms, but no singularity. This means that we should have the condition \begin{equation} r+{A_{\Phi_{h_{3/2,-2},h_{3/2,2}}}\over A_{\Phi_{h_{1/2,-2},h_{1/2,2}}}} r^*=0 \,. \end{equation} To put things differently, the appearance of a null-state in the conformal block $\overline{{\cal F}}_{h_{3/2,2}}$ leads to a divergence in the four-point function (assuming the formula for $A_{\Phi_{h_{3/2,-2},h_{3/2,2}}}$ given earlier is correct). To cancel this divergence, a contribution $A_{\Phi_{h_{1/4,-4},h_{1/4,4}}}$ is {\bf necessary}. Moreover, this contribution must exhibit a simple pole, as we have observed numerically. It is possible that this picture generalises, with singularities in the proposal of \cite{Ribault} exactly cancelled out by the additional terms we find in our lattice analysis. This will be discussed elsewhere \cite{JRSS}. To conclude this paper, we re-iterate the remark that the eigenvalues contributing to the probabilities are (a subset of) those appearing in the Potts model partition function \cite{DFSZ}. While this would be a well expected fact for a model defined locally such as the Ising model or any kind of height model, this is not so obvious in our case. Indeed, in a model where correlations are defined non-locally there is no clear connection between the partition function and at least some of the observables. To give a simple example, we know well that the probability that two points are connected with a cluster allowing two independent paths involves exponents not present in the partition function \cite{JacobsenZinn1,DengBloteNienhuis}. The fact that no such exponents are needed for the $P_{a_1 a_2 a_3 a_4}$ suggests that these are rather close to ``ordinary'' observables, and that we may be able to understand them in terms of fully local operators acting on the space of states. One of the main ``elementary'' mysteries in this description is why only sectors with an even number of clusters contribute. We believe that thinking more deeply about algebraic aspects of the problem on the lattice will shed some light on this question. \paragraph{Acknowledgements.} We thank the authors of \cite{Ribault} and especially S.\ Ribault for inspiring discussions. We also thank C.R.\ Scullard for discussions about numerical aspects and especially for the use of his unpublished {\sc C++} implementation of the {\sc Arpack} version of the Arnoldi algorithm. This work was supported by the ERC Advanced Grant NuQFT.

32f7c06791cad9fce5291ca0d29a2c9b55c2838a

\section{Introduction} \label{S:Intro} Many academic studies on fault detection need to assess the quality of their technique using seeded faults. One of the widely-used systematic ways to seed simulated faults into the programs is mutation testing~\cite{DeMillo1978}. Mutation testing is the process of injecting faults into software (i.e. creating a mutant), and counting the number of these faults that make at least one test fail (i.e. kill the mutant). The process of creating a mutant consists of applying a predefined transformation on the code (i.e. mutation operator) that converts a version of the code under test into a faulty version. It has been shown that mutation testing is an appropriate method to simulate real faults and perform comparative analysis on testing techniques~\cite{Andrews2005,Andrews2006,Just2014}. There has been many studies to optimize the process of mutation testing by following the maxim \textit{\{do faster, do smarter, do fewer\}}~\cite{Offutt2001}. In particular, \textit{do fewer} aims to reduce the number of produced mutants. There are several techniques that implement this logic (e.g. selective mutation~\cite{Mathur1991,Offutt1993,Offutt1996}, and mutant sampling~\cite{Wong1995,Zhang2010,Zhang2013,Parsai2016}). However, only recently the academics began to investigate the threats to validity the redundant mutants introduce in software testing experiments~\cite{Papadakis2016}. Papadakis et al. demonstrate that the existence of redundant mutants introduces a significant threat by ``artificially inflating the apparent ability of a test technique to detect faults''~\cite{Papadakis2016}. One of the recent solutions to alleviate this problem is to use mutant subsumption~\cite{Ammann2014}. Mutant \emph{A} \textit{truly} subsumes mutant \emph{B} if and only if all inputs that kill \emph{A} also kill \emph{B}. This means that mutant \emph{B} is redundant, since killing \emph{A} is sufficient to know that \emph{B} is also killed. It is possible to provide a more accurate analysis of a testing experiment by determining and discarding the redundant mutants. However, it is often impossible to check mutants for every possible input to the program in practice. Therefore, as a compromise, dynamic mutant subsumption is used instead~\cite{Ammann2014}. Mutant \emph{A} \textit{dynamically} subsumes mutant \emph{B} with regards to test set \emph{T} if and only if there exists at least one test that kills \emph{A}, and every test that kills \emph{A} also kills \emph{B}. Given the fact that mutant subsumption only recently has been at the center of attention, there are no mature tools that can perform dynamic mutant subsumption analysis on real-life Java programs. This, however, is necessary to facilitate further research on the topic. Therefore we aim to fill this void by developing such tool. We used LittleDarwin\footnote{\url{https://littledarwin.parsai.net/}} mutation testing framework to implement the features needed to perform dynamic mutant subsumption analysis. LittleDarwin is an extensible and easy to deploy mutation testing tool for Java programs~\cite{Parsai2017}. LittleDarwin has been used previously in several other studies~\cite{Parsai2016,Parsai2016M}, and it is shown to be capable of analyzing large and complicated Java software systems~\cite{Parsai2015T}. The rest of the paper is organized as follows: In Section~\ref{S:Background}, background information about mutation testing is provided. In Section~\ref{S:SOTA}, the current state of the art is discussed. In Section~\ref{S:DMSA}, we provide details on how LittleDarwin can help performing dynamic mutant subsumption analysis. Finally, we present our conclusions in Section~\ref{S:Conclusion}. \section{Background} \label{S:Background} The idea of mutation testing was first mentioned by Lipton, and later developed by DeMillo, Lipton and Sayward~\cite{DeMillo1978}. The first implementation of a mutation testing tool was done by Timothy Budd in 1980~\cite{Budd1980}. Mutation testing is performed as follows: First, a faulty version of the software is created by introducing faults into the system \textit{(Mutation)}. This is done by applying a known transformation \textit{(Mutation Operator)} on a certain part of the code. After generating the faulty version of the software \textit{(Mutant)}, it is passed onto the test suite. If there is an error or failure during the execution of the test suite, the mutant is marked as killed \textit{(Killed Mutant)}. If all tests pass, it means that the test suite could not catch the fault, and the mutant has survived \textit{(Survived Mutant)}~\cite{Jia2011}. If the output of a mutant for all possible input values is the same as the original program, it is called an \emph{equivalent mutant}. It is not possible to create a test case that passes for the original program and fails for an equivalent mutant, because the equivalent mutant is indistinguishable from the original program. This makes the creation of equivalent mutants undesirable, and leads to false positives during mutation testing. In general, detection of equivalent mutants is undecidable due to the halting problem~\cite{Offutt1997}. Manual inspection of all mutants is the only way of filtering all equivalent mutants, which is impractical in real projects due to the amount of work it requires. Therefore, the common practice within today's state-of-the-art is to take precautions to generate as few equivalent mutants as possible, and accept equivalent mutants as a threat to validity (accepting a false positive is less costly than removing a true positive by mistake~\cite{Fawcett2006}). \begin{equation} Mutation\ Coverage = \frac{Number\ of\ killed\ mutants}{Number\ of\ all\ non\mbox{-}equivalent\ mutants} \label{coverageequation} \end{equation} Mutation testing allows software engineers to monitor the fault detection capability of a test suite by means of mutation coverage (see Equation~\ref{coverageequation})~\cite{Jia2011}. A test suite is said to achieve \textit{full mutation test adequacy} whenever it can kill all the non-equivalent mutants, thus reaching a mutation coverage of 100\%. Such test suite is called a \textit{mutation-adequate test suite}. \section{State of the Art} \label{S:SOTA} \begin{figure} \centering \fbox{ \includegraphics[width=0.9\linewidth]{images/dms-example.png}} \caption{An Example Mutated Method} \label{fig:dmsexample} \end{figure} Mutant subsumption is defined as the relationship between two non-equivalent mutants \texttt{A} and \texttt{B} in which \texttt{A} subsumes \texttt{B} if and only if all inputs that kill \texttt{A} is guaranteed to kill \texttt{B}~\cite{Kurtz2015}. The subsumption relationship for faults has been defined by Kuhn in 1999~\cite{Kuhn1999}, but its use for mutation testing has been popularized by Jia et al. for creating hard to kill higher-order mutants~\cite{Jia2008}. Later on, Ammann et al. tackled the theoretical side of mutant subsumption~\cite{Ammann2014}. In their paper, Ammann et al. define \textit{dynamic} mutant subsumption, which redefines the relationship using test cases. Mutant \texttt{A} dynamically subsumes Mutant \texttt{B} if and only if (i) \texttt{A} is killed, and (ii) every test that kills \texttt{A} also kills \texttt{B}. Kurtz et al.~\cite{Kurtz2015} use the notion of dynamic mutant subsumption graph (DMSG) to visualize the concept of dynamic mutant subsumption. Each node in a DMSG represents a set of all mutants that are mutually subsuming. Edges in a DMSG represent the dynamic subsumption relationship between the nodes. They introduce the concept of static mutant subsumption graph, which is a result of determining the subsumption relationship between mutants using static analysis techniques. \begin{table} \centering \caption{Range of Input Values that Kill Mutants of the Example Mutated Method (left), DMSG for the Example Mutated Method (right)} \label{table:dmsexamplerange} \begin{minipage}{0.84\linewidth} \begin{tabular}{|c|c|c|} \hline \textbf{Mutants} & \textbf{Range of $a$} & \textbf{Range of $b$} \\ \hline \textbf{M0} & $\varnothing$ & $\varnothing$ \\ \hline \textbf{M1} & $(-\infty, \infty) - \{0\}$ & $(-\infty, 0)$ \\ \hline \textbf{M2} & $(-\infty, \infty) - \{0\}$ & $(-\infty, \infty) - \{0\}$ \\ \hline \textbf{M3} & $(-\infty, \infty) - \{0\}$ & $(-\infty, \infty) - \{0\}$ \\ \hline \textbf{M4} & $(-\infty, \infty) - \{0\}$ & $(-\infty, 0)$ \\ \hline \textbf{M5} & $(-\infty, \infty) - \{0\}$ & $(-\infty, \infty)$ \\ \hline \textbf{M6} & $(-\infty, \infty) - \{0\}$ & $(-\infty, \infty) - \{0\}$ \\ \hline \textbf{M7} & $\varnothing$ & $\varnothing$ \\ \hline \end{tabular} \end{minipage} \begin{minipage}{0.15\linewidth} \includegraphics[width=\textwidth]{images/DMSExample-graph-crop.pdf} \end{minipage} \end{table} Figure~\ref{fig:dmsexample} shows a Java method and its set of mutants. This method takes $a$ and $b$ as input, and returns $a \times b$ as output. To do this, $a$ is added $b$ times. If $b$ is negative, both $a$ and $b$ are negated so that $b$ becomes positive. Table~\ref{table:dmsexamplerange} shows the range of input values that kills each mutant. As the table shows, M0 and M7 are equivalent mutants, since the change they introduce does not impact the program semantically. M1 and M4 are killed by the same range of inputs. The same holds true for M2, M3, and M6. It can be seen that \{M1,M4\} truly subsume \{M2,M3,M6\}, since any input that kills M1 or M4, also kills M2, M3, and M6; however, the opposite does not hold. Also, \{M2,M3,M6\} truly subsume \{M5\} for the same reason. Using a test suite that includes a test case from each of the input ranges in Table~\ref{table:dmsexamplerange}, it is possible to draw the DMSG for this method. The main purpose behind the use of mutant subsumption is to reliably detect redundant mutants, which create multiple threats to the validity of mutation testing~\cite{Papadakis2016}. This is often done by determining the dynamic subsumption relationship among a set of mutants, and keeping only those that are not subsumed by any other mutant. In our example, keeping only M1 (or M4) suffices, since it subsumes all the other mutants. \begin{figure*}[!h] \centering \includegraphics[width=\linewidth]{images/JTerminal} \caption{Dynamic Mutant Subsumption Graph for JTerminal} \label{fig:jterminal} \end{figure*} \section{Dynamic Mutant Subsumption Analysis with LittleDarwin} \label{S:DMSA} \begin{figure} \centering \includegraphics[width=0.7\linewidth]{images/LittleDarwinDMS-crop} \caption{Dynamic Mutant Subsumption Component I/O} \label{fig:dms} \end{figure} Figure~\ref{fig:dms} shows the input and output of LittleDarwin's dynamic mutant subsumption (DMS) component. To facilitate dynamic mutant subsumption analysis in LittleDarwin, we retain all the output provided by the build system for each mutant. As a result, we can parse this output and extract useful information, e.g. which test cases kill a particular mutant. LittleDarwin's DMS component can then use this information to determine dynamic subsumption relation between each mutant pair. This component then outputs the results in two different ways: (i) the dynamic mutant subsumption graph, to visualize the subsumption relation, and (ii) % a detailed report is generated in CSV\footnote{Comma-separated Values} format that contains all the information processed by the DMS component. For each mutant, mutant ID, mutant path, source path, mutated line number, whether it is a subsuming mutant, number of failed tests, the mutants it subsumes, the mutants that it is subsumed by, and the mutants that are mutually subsuming with it are provided in this report. Since LittleDarwin is a Java mutation testing framework, the application of the DMS component is also restricted to Java programs. \begin{figure} \centering \fbox{ \includegraphics[width=0.9\linewidth]{images/mutatedcode2}} \caption{Mutants 45 and 56 of JTerminal} \label{fig:code} \end{figure} \begin{figure} \centering \fbox{ \includegraphics[width=0.9\linewidth]{images/testone2}} \caption{The Test that Kills Mutant 45, but Not Mutant 56} \label{fig:testcode} \end{figure} To showcase the ability of LittleDarwin in performing dynamic mutant subsumption analysis, we use JTerminal\footnote{\url{https://www.grahamedgecombe.com/projects/jterminal}} as a subject. We improved the test suite of JTerminal by automatically generating test cases using EvoSuite~\cite{SBST17_competition}. % The information about characteristics of JTerminal is shown in Table~\ref{table:cases}. The DMSG for JTerminal is depicted in Figure~\ref{fig:jterminal}. In this figure, each number represents a single killed mutant, each node represents a group of mutants that are killed by exactly the same set of test cases, and each edge shows the dynamic subsumption relationship between each node where the node at the end is subsumed by the node at the start. The survived mutants are not shown in this figure. The double-circled nodes contain the subsuming mutant groups. In order to remove the redundant mutants, one only needs to keep one mutant from each subsuming mutant group and discard the rest. Take M45 and M56 as an example. According to the DMSG, M56 subsumes Mutant M45. Using the CSV report, we can locate the actual mutation of the source code (Figure~\ref{fig:code}). Both M45 and M56 belong to method \texttt{parse} of class \texttt{AnsiControlSequenceParser}, and mutate the same statement on line 99. % M45 acts as a negation of the conditional statement. This means that any input character (except -1) that used to trigger "else if" and "else", now trigger this branch. Since this branch contains a "break" statement, it avoids the rest of the iteration of the loop to be executed. If the input is -1, the "else" branch would be executed, which wrongfully appends -1 to "text". M56, however, changes only two special cases. If the input is +1, the "if" branch would be executed, and the current iteration breaks. If the input is -1, the same thing as M45 happens. For any other input, the program executes as it should. This means that M56 truly subsumes M45. Figure~\ref{fig:testcode} shows the test case that kills M45, but not M56. The input value here is a single control sequence, which is neither -1 or +1, and therefore cannot kill M56. However, since it should have been handled by "else if" branch and M45 does not allow that, it kills M45. Hence, in Figure~\ref{fig:jterminal} (on the left side) , we can see that M56 dynamically subsumes M45. Analysis such as this allows researchers to understand the relations between the mutants and reduce the effects of redundant mutants on their results. \begin{table} \centering \caption{JTerminal Software Information} \label{table:cases} \adjustbox{max width=\linewidth}{ \begin{tabular}{|c|c|c|c|c|c|c|c|c|c|} \hline \multirow{2}{*}{\textbf{Project}} & \multirow{2}{*}{\textbf{Ver.}} & \multicolumn{2}{c|}{\textbf{Size (LoC)}} & \multirow{2}{*}{\textbf{\#C}} & \multirow{2}{*}{\textbf{TS}} & \multirow{2}{*}{\textbf{SC}} & \multirow{2}{*}{\textbf{BC}} & \multirow{2}{*}{\textbf{MC}} & \multirow{2}{*}{\textbf{\#M}}\\ \hhline{~~--~~~~} & & \textbf{Prod.} & \textbf{Test} & & & & & & \\ \hline \hline JTerminal & 1.0.1 & 687 & 428 & 8 & 2 & 66\% & 56\% & 60.0\% & 160\\ \hline % \multicolumn{10}{c}{} \\ \multicolumn{1}{l}{Acronyms:} & \multicolumn{9}{r}{Version (Ver.), Line of code (LoC), Production code (Prod.),} \\ \multicolumn{10}{r}{Number of commits (\#C), Team size (TS), Statement coverage (SC),}\\ \multicolumn{10}{r}{Branch coverage (BC), Mutation coverage (MC), Number of Mutants (\#M)} \end{tabular} } \end{table} \section{Conclusion} \label{S:Conclusion} Many academic studies in the field of software testing rely on mutation testing to use as their comparison criteria, and the existence of redundant mutants is a significant threat to their validity. We developed a component for our mutation testing tool, LittleDarwin, to facilitate the detection of redundant mutants using dynamic mutant subsumption analysis. We performed dynamic mutant subsumption analysis on a small, real-world project to demonstrate the capabilities of our tool. Using our tool, it is possible to detect and filter out redundant mutants, and help in increasing the confidence in results of experiments using mutation testing as a comparison criteria. \balance \bibliographystyle{ACM-Reference-Format}

4b6269507da4aa555f018290cdb00b26d9dab462

\section{Introduction} The non-zero electric dipole moment (EDM) of elementary particles is a direct evidence for violations of both parity ($P$) and time-reversal ($T$) symmetries~\cite{purcell:1950,landau:1957}; $T$ violation also implies charge-parity ($CP$) violation assuming the $CPT$ invariance~\cite{Lehnert:2016}. The demand of new $CP$-violation sources~\cite{Sakharov:1991} to explain the matter-antimatter asymmetry in the Universe~\cite{Canetti:2012} sparked the great interest in the EDM measurements. The EDMs of elementary particles have been searched for decades~\cite{Chupp:2017} since the first neutron EDM experiment in 1950~\cite{purcell:1950}. The extremely small upper limit of the neutron EDM~\cite{Baker:2006,pendlebury:2015} implied an unsolved strong $CP$ problem in the quantum chromodynamics (QCD)~\cite{peccei:1977}. The axion, a hypothetical elementary particle, was introduced to potentially resolve the strong $CP$ problem by Peccei and Quinn~\cite{peccei:19771,weinberg:1978}. The axion is a prominent candidate for dark matter in the Universe~\cite{Duffy:2009}. The evidence for dark matter came from cosmological and astrophysical observations, including the cosmic microwave background (CMB) power spectrum~\cite{Efstathiou:1992}, cluster and galactic rotation curves~\cite{zwicky:1933, rubin:1983}, gravitational lensing~\cite{walsh:1979,clowe:2006}, and large-scale structure formation~\cite{springel:2006}. The invisible dark matter has been known to compose more than 25\% of total mass-energy in the Universe~\cite{Aghanim:2018eyx}. Other possible candidates of dark matter include weakly interacting massive particles (WIMPs)~\cite{steigman:1985}, sterile neutrinos~\cite{kusenko:2009} and others~\cite{patrignani:2017}. Despite many experiments, the nature of the dark matter still remains unknown. The axion resolving the strong $CP$ problem is the so-called QCD axion. The mass of the QCD axion is given by \begin{align} m_a \sim 6\times10^{-10}~\text{eV}\Big(\frac{10^{16}~\text{GeV}}{f_a}\Big) \end{align} where $f_a$ is the symmetry-breaking energy scale~\cite{peccei:1977}. The traditional range of axion mass is considered to be from $10^{-6}$ to $10^{-2}$~eV, the so-called ``axion window," according to constraints from observations of current experiments, astrophysics, and cosmology~\cite{Raffelt:2006}. The lower bound of the axion window is derived from cosmology dark matter abundance~\cite{Turner:1990}. However, theories beyond the Standard Model of particle physics, e.g., supersymmetry and super string~\cite{svrcek:2006,jaeckel:2010}, suggest that ultralight axions with mass much smaller than $10^{-10}$~eV can be generated by a broken global symmetry at high-energy scales such as the grand-unified ($f_a\sim10^{16}$~GeV) or the Planck scales ($f_a\sim10^{19}$~GeV). Besides, if the axion mechanism happened before the cosmic inflation, the inflation could have highly suppressed the axion field and generate ultralight axions~\cite{Pi:1984,Linde:1991}. These axions, the so-called axion-like-particles (ALPs), do not have a specific relation between their field and mass, unlike the QCD axion. Axions, in principle, can couple with ordinary particles such as photons, electrons, and nucleons. Axions have been searched for decades in various systems; however, a large area of parameter space of the axion field amplitude and the axion mass has not been constrained, especially at the sub-eV range of axion mass. To this end, further sensitive experimental searches for axions are necessary. The axion field was first suggested as a classical field oscillating at a frequency equal to the axion mass $m_a$~\cite{preskill:1983,dine:1983} (the axion Compton frequency in natural units where $c=\hbar=1$), expressed as $a_0\cos(m_a t)$. Here $a_0$ is the local amplitude of the axion field. Recently, several papers have suggested that the interactions of the coherently oscillating axion field with particles of gluons and fermions will induce an EDM of the particles oscillating at the same frequency as the axion field~\cite{graham:2011,graham:2013,stadnik:2014,hill:2016,Alexander:2017}. The laboratory experiments detecting these oscillating effects will be sensitive to the axion mass~\cite{dine:1983,Duffy:2009}. Besides, while the effects of several other axion searches, such as axion helioscopes and light-shining-through-walls experiments, scale quadratically with the axion coupling constant~\cite{Graham:2015}, the effects of the oscillating axion field scale linearly with the axion coupling constant, as indicated in Eqs.~\ref{eq:oedm_nuclear} and~\ref{eq:oedm_electron} described below~\cite{budker:2014,abel:2017}. The cosmic axion spin precession experiment (CASPEr) has been proposed to search for a nuclear oscillating EDM (OEDM) induced by axions~\cite{budker:2014}. This experiment is based on the nuclear magnetic resonance technique. The nuclear OEDM, $d_n$, is given by~\cite{graham:2013} \begin{align} d_n =& g_d a(t) = g_d a_0 \cos(m_a t) \label{eq:oedm_nuclear} \end{align} where $g_d$ is the coupling strength of the axion-gluon interaction and $a(t)$ is the local axion field oscillating at the axion mass. On the assumption that this axion field composes all dark matter, its density will be equal to~\cite{graham:2013} \begin{align} \rho_{DM} = \frac{1}{2}m_a^2 a_0^2 \end{align} where $\rho_{DM}\sim 0.3$ GeV/cm$^3$~\cite{patrignani:2017}. Then, the nuclear OEDM can be rewritten as \begin{align} d_n =& g_d\frac{\sqrt{2\rho_{DM}}}{m_a} \cos(m_a t). \label{eq:dn} \end{align} The CASPEr experiment utilizes nuclear spins in a solid sample, pre-polarized by an external strong magnetic field $B_0$. The experiment is conducted in a low temperature in order to reduce thermal noise. In the presence of an electric field, perpendicular to $B_0$, the polarized nuclear spins can be rotated away from the direction of $B_0$ due to the interaction of the nuclear OEDM with the electric field. It causes the spins to precess at the spin's Larmor frequency in $B_0$, which can be detected by a sensitive magnetometer such as a superconducting quantum interference device (SQUID)~\cite{budker:2014} or an atomic magnetometer (AM)~\cite{wang:2018}, both of which can reach a 1 fT/$\sqrt{\text{Hz}}$ sensitivity. Although the main focus of CASPEr is axion coupling to nucleon~\cite{budker:2014}, the axion could also couple to electrons~\cite{hill:2016} or other particles~\cite{stadnik:2014}. According to Refs.~\cite{Hill:2015,hill:2016,Hill:2017}, the interaction between an electron spin and the axion field can induce a $CP$-violating, non-zero electron OEDM. A classical derivation in the Maxwell equations is described in Ref.~\cite{Hill:2016zos}. The axion-induced electron OEDM is written as \begin{align} \vec{d_e}(t) =& g_{a\gamma\gamma} 2\mu_{\text{B}}a_0\cos(m_a t)\hat{\sigma} \label{eq:oedm_electron} \end{align} where $g_{a\gamma\gamma}$ is the strength of the axion-photon coupling~\footnote{The electron OEDM is defined as $d_e(t) = 2g_a \Big(\frac{a_0}{f_a}\Big)\mu_{B} \cos(m_at)$ in~\cite{hill:2016}. The author also suggested that the electron OEDM could be three orders of magnitude larger than that of the nucleon owing to larger magnetic moment of electrons than nucleons. Here we define $g_{a\gamma\gamma} \equiv g_{a}/f_a$ in order to compare with other experiments. For the QCD axion in the KSVZ model ~\cite{patrignani:2017}, $g_{a\gamma\gamma}\sim 4\times 10^{-10} (\frac{m_a}{\text{eV}}) [\text{GeV}^{-1}]$.} and $\mu_\text{B}$ is the Bohr magneton. The OEDM is collinear with electron spin $\vec{\sigma}$. The amplitude of the OEDM can be estimated from Eq.~(3): \begin{align} d_e=8.3\times10^{-23}~e\cdot\text{cm}\Big(\frac{g_{a\gamma\gamma}}{\text{GeV}^{-1}}\Big)\Big(\frac{\text{eV}}{m_a}\Big). \label{eq:de-Amp} \end{align} The coupling between axions and electrons, the $g_{aee}$ term, can also induce the spin precession without the need of an electric field as described in Eq.~21 of Ref.~\cite{graham:2013}. Based on the current constraint $g_{aee} < 10^{-10}~\text{GeV}^{-1}$, the effective energy shift due to the axion-electron coupling is about 5 orders of magnitude smaller than the sensitivity of the energy shift of the proposed experiment (see Eq.~9). Therefore, we neglect the contribution from the axion-electron coupling. The static electron EDM has been searched in various systems such as atoms~\cite{Chin:2001,Regan:2002}, molecules~\cite{Hudson:2011,Baron:2014}, ions~\cite{Cairncross:2017}, and solid-state materials~\cite{Eckel:2012, Kim:2015}. So far the best constraint for electron EDM is $8.7\times 10^{-29}~ \text{e}\cdot\text{cm}$ using polar molecules~\cite{Baron:2014}. The rubidium (Rb) and potassium (K) AMs have been also suggested to employ for hunting the static electron EDM because of their high sensitivity~\cite{Weiss:2003}. While SQUIDs have been proposed as a magnetic-field sensor~\cite{Alexander:2018}, in this paper, we propose to search for the axion-induced electron OEDM using a new technique based on AMs, a non-cryogenic magnetic-field sensor. \section{Experimental Approach} \begin{figure}[b] \includegraphics[width=0.45\textwidth]{Scheme.pdf} \caption{Scheme of detecting an axion-induced electron OEDM using an AM with pumping in the $z$ direction and probing in the $x$ direction. A magnetic field and an electric field are externally applied to the AM vapor along the $z$ and $y$ direction, respectively. The interaction of the electron OEDM with the electric field generates an effective oscillating magnetic field, $\vec{B}_{\text{eff}}$, in the $y$ direction, which causes the electron spins to precess at the spin's Larmor frequency. The projection of the spin precession onto the $x$ direction is the response of the AM to the electron OEDM. } \label{fig:bloch} \end{figure} The basic principle of the axion-induced electron OEDM search using an AM is illustrated in Fig.~\ref{fig:bloch}. One important advantage of this experiment is that an AM will serve as both a source of polarized electrons and a very sensitive detector, so it does not require an additional solid sample and a pickup coil system, for example, the one in the CASPEr experiment. Electron spins of an AM alkali vapor are polarized by a pump laser beam in the $z$ direction. To determine the Larmor frequency for the polarized electron spins, a holding magnetic field $\vec{B}_0$ is externally applied to the vapor along the pump beam $z$ direction. Unlike the CASPEr experiment, this experiment does not require a several Tesla superconducting magnet for polarization of electron spins, greatly simplifying the experimental design. In order to observe the effect of the electron OEDM, a static electric field $\vec{E}$ is externally applied to the vapor in the direction perpendicular to $\vec{B}_0$, in the $y$ direction. The electron OEDM interacts with the electric field, $-\vec{d_e}(t)\cdot\vec{E}$, which induces an effective oscillating magnetic field \begin{align} \vec{B}_{\text{eff}} = [{E} g_{a\gamma\gamma} 2\mu_{\text{B} }a_0 \cos(m_a t)/\gamma]\hat{y}\equiv B_1 \cos(m_a t)\hat{y} \label{eq:beff} \end{align} where $\gamma$ is the gyromagnetic ratio of the AM alkali atoms. The total magnetic field at the AM vapor is written as $\vec{B}=B_0\hat{x}+B_1\cos(m_a t)\hat{y}$. The effective field $\vec{B}_{\text{eff}}$ causes the electron spins to tilt away from the direction of $\vec{B}_0$ and to precess. Such spin precession can be sensitively detected with a probe laser beam in the $x$ direction. When the spin's Larmor frequency is equal to the axion mass, the precession signal will be maximized because a resonance occurs. The goal of this experiment aims to precisely detect the axion-induced spin precession with an AM, which will determine or constrain the axion-photon interaction. For application of an electric field to an AM vapor, two conductive planar electrodes will be placed on the vapor cell. To reduce the magnetic Johnson noise, electrodes will be made from machinable ceramics coated with graphite that has large electrical resistivity~\cite{Kim:2015}. The sensitivity of the OEDM experiment to the axion-photon interaction can be improved by increasing the strength of the electric field (see Eq.~\ref{eq:beff}). One electrode is connected to a high-voltage source and the other electrode to ground. To prevent electrical breakdown of the vapor cell, the cell can be filled with some pressures of nitrogen gas~\cite{Meek1953}. According to Ref.~\cite{Seltzer2008}, it was measured that the amplitude of an electric field inside a vapor cell is smaller than the applied field amplitude. This requires to precisely monitor the field amplitude during the OEDM experiment by, for example, measuring the Stark shift of the optical resonance frequency of a probe laser beam ~\cite{Seltzer2008}. In addition, it was found that a high electric field can adversely affect the alkali vapor density~\cite{Seltzer2008}; the vapor density in a cell made of uncoated aluminosilicate glass slowly decreased in the presence of the electric field in the cell, depending on the strength of the field, and the vapor density did not immediately recover when the field was removed. Possible solutions for the density decrease are to coat the interior cell wall with multilayer of octadecyltrichlorosilane or make the cell with Pyrex glass~\cite{Seltzer2008}. The electric field homogeneity in such glass vapor cells can be improved by a silane coating on the interior cell walls~\cite{Hunter1988}. The OEDM experiment eliminates common systematics in static EDM experiments, which are associated with spurious magnetic fields that mimics an EDM. The systematic fields are produced mainly due to electric-field reversals or movement of target particles~\cite{Kim:2015}. A possible systematic in the OEDM experiments will be transverse fluctuations of the holding field $\vec{B}_0$, which can be suppressed by using a low-noise voltage source, such as dc batteries, and by designing a better coil system such as Maxwell coils or multi-coil arrays~\cite{Wang2002,Merritt1983,Garrett1967}. The transverse component of the holding field produced by such coil systems can be reduced, suppressing its fluctuations. Another possible systematic will be mechanical vibrations at the target frequency, which can be eliminated by making a rigid experimental setup. The dc magnetic field generated from a leakage current across the vapor cell is the most important systematic in static EDM experiments, while in OEDM experiments, it can only change the resonance frequency of an AM, given by $f_0=\gamma \sqrt{B_x^2+B_y^2+B_z^2}$ where $B_{x,y,z}$ are the dc magnetic field components in the $x$, $y$, and $z$ direction ~\cite{Karaulanov:2016}. For example, in an experiment using an AM with a 1~cm-diameter Rb cell, a 100~Hz bandwidth, and $\gamma$ of $7\times10^{9}$~Hz/T: if a leakage current flows in a loop around the cell, the field from the leakage current should be lower than $1.4\times 10^{-8}$~T to have the shift on the resonance frequency within 100~Hz. This implies that the leakage current below 100~$\mu$A is acceptable for operation of the OEDM experiment. High-voltage supplies for the electric field $\vec{E}$ can introduce an additional systematic bias to the observable signal. The nonzero voltage drift of the supplies, $dV/dt$, produces a magnetic field induced by the displacement current, flowing in and out of the electrodes, through the relation of $I=CdV/dt$ where $C$ is the capacitance between the vapor cell and electrodes~\cite{Kim:2015}. In case that the voltage drift is very slow, i.e., the first-order drift, it generates a dc magnetic field. This will only affect the shift of the AM resonance frequency, like the leakage current. The oscillating voltage drift at the target frequency will originate a significant systematic in OEDM experiments. Such ripple voltages of high-voltage supplies can be suppressed by, for example, adding a voltage regulator or selecting low-distortion supplies. Possible sources of noise are expected to be very small and can be controlled; the ambient high or low frequency noise can be suppressed by using radio-frequency (RF), ferrite, or $\mu$-metal shields. \section{Sensitivity Estimate} For the AM-based electron OEDM search, we propose to use a spin-exchange relaxation-free (SERF) AM containing Rb atoms for low frequencies below $\sim$1~kHz and a RF AM containing K atoms for high frequencies above $\sim$1~kHz~\cite{Savukov:2005} for a wide frequency range between 1~Hz to 1~MHz. The AMs are available at Los Alamos National Laboratory. The SERF AM can reach a 1~fT/$\sqrt{\text{Hz}}$ field sensitivity by operating in a near-zero magnetic field and by sufficiently heating the Rb vapor cell to achieve the SERF regime~\cite{Happer:1973,Allred:2002}. Similar high sensitivity of the RF AM is achieved by pumping spins to the stretched state reducing the spin-exchange relaxation (light-narrowing effect)~\cite{Savukov:2005}. Cs magnetometers can also be employed for this electron OEDM search due to the $\sim$ 4 times larger EDM enhancement factor for Cs atoms than Rb atoms~~\cite{nataraj:2008} (discussed below). However, because Cs magnetometers achieved a sensitivity of 40~ $\text{fT}/\sqrt{\text{Hz}}$~[64], roughly 40 times worse than the Rb SERF AM, we do not consider Cs magnetometers here. \begin{figure}[t] \centering \includegraphics[width=0.5\textwidth]{Setup.png} \caption{Simplified schematic drawing of the axion-induced electron OEDM search using an AM-based experimental method (not scaled). An electric and magnetic fields will be applied to an AM vapor by two electrodes and a Helmholtz coil, respectively. The axion-induced spin precession will be measured with a probe-laser-beam polarimeter, containing a photo-detector.} \label{fig:setup} \end{figure} The schematic representation of a proposed experimental setup is shown in Fig.~\ref{fig:setup}. Two laser beams overlap in an AM vapor cell: a pump beam orients electron spins along its direction and a probe beam reads out the state of the spins. The angle between the pump and probe beam can be arbitrary, e.g., parallel~\cite{Karaulanov:2016} or orthogonal. The conventional configuration of orthogonal beams shown in Fig.~\ref{fig:setup} provides the best field sensitivity. The externally applied holding magnetic field $\vec{B}_0$ is along the pump beam direction, and the external electric field $\vec{E}$ is along the direction perpendicular to both $\vec{B}_0$ and the probe beam. The axion-induced electron spin precession is sensitively detected with a probe-laser-beam polarimeter, containing a polarizing beam splitter and two photo-detectors, as illustraged in Fig.~\ref{fig:setup}, or two polarizers at small angle and a photo-detector. The AM system is placed into magnetic shields made of mu-metal or ferrite to reduce ambient magnetic noise. For the Rb and K atoms in the AMs, the effective oscillating magnetic field $\vec{B}_{\text{eff}}$ becomes \begin{align} \gamma\vec{B}_{\text{eff}}(t)=\vec{E}Rd_e (t) \end{align} where $R$ is the EDM enhancement factor for the atoms. Similar to the static EDM measurements of atoms and molecules, the OEDM can be also enhanced by the corresponding Schiff moment~\cite{Sandars:1966} for several orders of magnitude~\cite{stadnik:2014}. The gyromagnetic ratio of both Rb and K atoms is $7\times10^{9}$~Hz/T which corresponds to the energy sensitivity of $2.9\times 10^{-20}$~eV based on the AM field sensitivity of 1~fT in an one second measurement. We estimate the sensitivity of our proposed experiment to the electron OEDM. Based on the energy sensitivity of the AMs and Eq.~8, the sensitivity of the energy shift due to the electron OEDM is limited by \begin{align} E R_{\text{K/Rb}} d_e=2.9\times 10^{-20}~\text{eV} \end{align} where $R_{\text{K/Rb}}$ is the EDM enhancement factor for K and Rb atoms, respectively. Using Eq.~6 and with $E=5$~kV/cm, the sensitivity to the axion-photon coupling, $g_{a\gamma\gamma}$, becomes \begin{align} g_{a\gamma\gamma}=0.07 \Big(\frac{1}{R_{\text{K/Rb}}}\Big) \Big(\frac{m_a}{\text{eV}}\Big) \text{GeV}^{-1}. \label{eq:EneSen} \end{align} For the K AM with $R_{\text{K}}=2.5$~\cite{Sandars:1966}, the sensitivity to $g_{a\gamma\gamma}$ is given by \begin{align} g_{a\gamma\gamma} =2.8\times10^{-2}\Big(\frac{m_a}{\text{eV}}\Big)\text{GeV}^{-1},\label{eq:grrSen1} \end{align} while for the Rb AM with $R_{\text{Rb}}=25$~\cite{nataraj:2008}, \begin{align} g_{a\gamma\gamma} =2.8\times10^{-3}\Big(\frac{m_a}{\text{eV}}\Big)\text{GeV}^{-1}.\label{eq:grrSen2} \end{align} \begin{figure}[t] \includegraphics[width=0.5\textwidth]{./g_agg_all.pdf} \caption{Estimated sensitivity of our electron OEDM experiment based on Rb and K AMs to $g_{a\gamma\gamma}$ on the axion mass range with 1 s integration time (blue dashed line). The experiment will improve the current limit set by the CAST experiment~\cite{Anastassopoulos:2017} (red region) for the axion masses between $10^{-14}$ and $10^{-10}$ ~eV, specifically by 5 orders of magnitude at $m_a\sim10^{-14}$~eV. The gray region is constrained by astrophysical sources such as High Energy Stereoscopic System (H.E.S.S.)~\cite{Abramowski:2013} and supernova SN 1987A~\cite{Payez:2015}. The sensitivities of other proposed experiments such as ABRACADABRA~\cite{Kahn:2016} with 1~yr integration time, DM Radio~\cite{Chaudhuri:2015,Chaudhuri:2018} with 1.5~yr integration time and LC-circuit~\cite{Chu:2018} with 7~h integration time are also included. The black curve above the CAST limit shows the first result from the experiment of ABRACADABRA-10 cm~\cite{Ouellet:2018}.} \label{fig:gagg} \end{figure} Figure~\ref{fig:gagg} shows the projected sensitivity of the proposed experiment to $g_{a\gamma\gamma}$, based on Eqs.~\ref{eq:grrSen1} and ~\ref{eq:grrSen2}. The proposed experiment can set a new experimental limit on a significant axion mass range between $10^{-15}$ to $10^{-10}$~eV. In particular, the experiments will improve the current experimental bound set by the CERN Axion Solar Telescope (CAST) experiment~\cite{Anastassopoulos:2017} by 5 orders of magnitude at $m_a\sim10^{-14}$~eV. The SERF Rb AM is sensitive at low frequencies from 1 to 1500~Hz~\cite{Savukov:2017}, corresponding to the axion masses from $4.13\times 10^{-15}$ to $6.20\times 10^{-12}~\text{eV}$. This requires the $B_0$ of $2\times 10^{-10}$ to $2\times 10^{-7}~\text{T}$ to match the spin's Larmor frequency to the axion mass. The upper bound of axion mass range $10^{-10}$~eV, corresponding to the frequency of $\sim$ 10~kHz, requires $B_0$ $\sim3\times10^{-6}$~T. The sensitivity in the experiment can be significantly improved by orders of magnitude with a long data integration time. \section{Conclusion} In conclusion, we proposed a novel, original experimental method to search for the axion-induced oscillating EDM of electrons using very sensitive atomic magnetometers. We estimated that the experiment is sensitive to the axion-photon coupling in the axion masses from $10^{-15}$ to $10^{-10}$~eV. The electron OEDM experiment can improve the current experimental limit by a few orders of magnitude and push the sensitivity closer to the QCD axion. This approach can be applied to search for OEDMs of other particles such as xenon~\cite{graham:2013}, neutron~\cite{abel:2017}, and molecules~\cite{stadnik:2014} with a non-zero magnetic dipole moment. The method of co-magnetometer readout~\cite{Chupp:1998,Tullney:2013} may be useful to improve the sensitivity in some OEDM experiments as well. In the future, improving technologies for static EDM experiments would also benefit to OEDM experiments. \section*{ACKNOWLEDGMENTS} The authors thank Dr. Mike Snow and Dr. Steven Clayton for useful discussions. This work was supported by the U.S. DOE through the LANL/LDRD program.

96545113662842e185117ade4b89015ddc31e6e0

\section{Introduction and results} This paper is devoted to the scattering theory of mathematical models arising from Quantum Field Theory (QFT). One of our main concerns is to establish asymptotic completeness of the wave operators for models involving massless fields. In the recent literature, this problem has been notably studied for Pauli-Fierz Hamiltonians describing confined non-relativistic particles interacting with a quantized radiation field \cite{Ge02_01,DeKu13_01,FaSi14_01,DeGrKu15_01}. Asymptotic completeness has been proven for the massless spin-boson model, but proving this property for more general Pauli-Fierz Hamiltonians remains an important open problem. In this paper, among other results, asymptotic completeness for a simplified model of QFT involving a massless field is proven, thanks to the particular structure of the model. We consider the weak interaction between the vector bosons $W^{\pm}$ and the full family of leptons. The latter involves the electron $e^-$, the positron $e^+$, the muon $\mu^-$, the antimuon $\mu^+$, the tau $\tau^-$, the antitau $\tau^+$, the associated neutrinos $\nu_{e}$, $\nu_{\mu}$, $\nu_{\tau}$ and the antineutrinos $\bar{\nu}_{e}$, $\bar{\nu}_{\mu}$, $\bar{\nu}_{\tau}$. Typical examples of processes we are interested in are the weak decay of the $W^{\pm}$ bosons into a lepton $l^\pm$ and its associated neutrino $\nu_l$ or antineutrino $\bar{\nu}_l$, \begin{equation} \label{process} W^- \, \to \, l^{-} + \bar{\nu}_{l}, \quad W^+ \, \to \, l^+ + \nu_l. \end{equation} In what follows, the mass of a particle $p$ will be denoted by $m_p$. It is equal to the mass of the corresponding antiparticle. Physically, the following inequalities hold: \begin{equation*} m_{e} < m_{\mu} < m_{\tau} < m_W. \end{equation*} Neutrinos were usually assumed to be massless in the classical form of the standard model of particle physics, but recent experiments have provided evidences for nonzero neutrino masses (see, e.g., \cite{Ol14_01} and references therein). Since the latter are extremely small, however, it is legitimate -- and conceptually interesting -- to consider models where neutrinos are supposed to be massless. In this paper, $m_e$, $m_\mu$, $m_\tau$ and $m_W$ will be treated as strictly positive parameters (we will not use the inequalities above), and we will consider separately two cases: i) $m_{\nu_e}>0$, $m_{\nu_\mu}>0$, $m_{\nu_\tau}>0$ and ii) $m_{\nu_e}=m_{\nu_\mu}=m_{\nu_\tau}=0$. The interaction term for the specific process \eqref{process} is given, in the Lagrangian formalism and for each lepton channel $l$, by (see, e.g., \cite{Gr00_01,GrMu00_01} and references therein) \begin{equation} \label{Interac} I = \int\left[ \Psi_l(x)^{\dagger}\gamma^0 \gamma^{\alpha} (1-\gamma_{5}) \Psi_{\nu_l}(x)W_{\alpha}(x) + \Psi_{\nu_l}(x)^{\dagger}\gamma^0 \gamma^{\alpha} (1-\gamma_{5})\Psi_{l}(x)W_{\alpha}(x)^* \right] d^3x, \end{equation} with \begin{align} &\Psi_l(x)=(2\pi)^{-\frac{3}{2}} \sum_{s_1=\pm\frac{1}{2}}\int\left[ \frac{u(p_1,s_1)e^{\mathrm{i}p_1 \cdot x}}{(2(|p_1|^2+m_l^2)^{\frac{1}{2}})^{\frac{1}{2}}} b_{l,+}(p_1,s_1) + \frac{v(p_1,s_1)e^{-\mathrm{i}p_1 \cdot x}}{(2(|p_1|^2+m_l^2)^{\frac{1}{2}})^{\frac{1}{2}}} b^*_{l,-}(p_1,s_1) \right]d^3p_1, \label{lept} \\ & \Psi_{\nu_l}(x)=(2\pi)^{-\frac{3}{2}} \sum_{s_2=\pm\frac{1}{2}}\int\left[ \frac{u(p_2,s_2)e^{\mathrm{i}p_2\cdot x}}{(2(|p_2|^2+m_{\nu_l}^2)^{\frac{1}{2}})^{\frac{1}{2}}} c_{l,+}(p_2,s_2) + \frac{v(p_2,s_2)e^{-\mathrm{i}p_2 \cdot x}}{(2(|p_2|^2+m_{\nu_l}^2)^{\frac{1}{2}})^{\frac{1}{2}}} c^*_{l,-}(p_2,s_2) \right]d^3p_2, \label{neu} \\ & W_{\alpha}(x)=(2\pi)^{-\frac{3}{2}} \sum_{\lambda=-1, 0, 1} \int\left[ \frac{\epsilon_{\alpha}(p_3,\lambda)e^{\mathrm{i}p_3 \cdot x} }{(2(|p_3|^2+m_{W}^2)^{\frac{1}{2}})^{\frac{1}{2}}} a_+(p_3,\lambda) + \frac{ \epsilon^*_{\alpha}(p_3,\lambda) e^{-\mathrm{i}p_3 \cdot x} }{(2(|p_3|^2+m_{W}^2)^{\frac{1}{2}})^{\frac{1}{2}}} a^*_-(p_3,\lambda) \right] d^3p_3. \label{Boson} \end{align} Here, $u$ and $v$ are the solutions to the Dirac equation (normalized as in \cite[(2.13)]{GrMu00_01}), $\epsilon_{\alpha}$ is a polarisation vector, $\gamma^{\alpha}$, $\alpha = 0 , \dots , 3$ and $\gamma_{5}$ are the usual gamma matrices. Moreover, the index $l \in \{1,2,3\}$ labels the lepton families, $p_1,p_2,p_3 \in \mathbb{R}^3$ stand for the momentum variables of fermions and bosons, $s_i \in \{-\frac{1}{2},\frac{1}{2}\}$ denotes the spin of fermions and $\lambda \in \{-1,0,1\}$ the spin of bosons. The operators $b_{l,+}(p_1,s_1)$ and $b^*_{l,+}(p_1,s_1)$ are annihilation and creation operators for the electron if $l=1$, muon if $l=2$ and tau if $l=3$. The operators $b_{l,-}(p_1,s_1)$ and $b^*_{l,-}(p_1,s_1)$ are annihilation and creation operators for the associated antiparticles. Likewise, $c_{l,+}(p_2,s_2)$ and $c^*_{l,+}(p_2,s_2)$ (respectively $c_{l,-}(p_2,s_2)$ and $c^*_{l,-}(p_2,s_2)$) stand for annihilation and creation operators for the neutrinos of the $l$-family (respectively antineutrinos) and the operators $a_+(p_3,\lambda)$ and $a^*_+(p_3,\lambda)$ (respectively $a_-(p_3,\lambda)$ and $a^*_-(p_3,\lambda)$) are annihilation and creation operators for the boson $W^-$ (respectively $W^+$). It should be mentioned that, when neutrinos are supposed to be massive, a slightly different interaction term can be found in the literature (see, e.g., \cite{Xi14_01}). More precisely, massive neutrinos fields $(\tilde{\Psi}_{\nu_1},\tilde{\Psi}_{\nu_2},\tilde{\Psi}_{\nu_3})$ may be defined by applying a $3\times3$ unitary matrix transformation to the fields $(\Psi_{\nu_1},\Psi_{\nu_2},\Psi_{\nu_3})$ in \eqref{neu}. Our results can be proven without any noticeable change if one considers such interaction terms. We will not do so in the present paper. For shortness, we denote by $\xi_i = (p_i,s_i)$, $i=1,2$, the quantum variables for fermions, and $\xi_3 = (p_3,\lambda)$ for bosons. The following canonical commutation and anticommutation relations hold: \begin{align*} \{b_{l,\epsilon}(\xi_1),b^*_{l',\epsilon'}(\xi_2)\} & = \{c_{l,\epsilon}(\xi_1),c^*_{l',\epsilon'}(\xi_2)\} = \delta_{ll'}\delta_{\epsilon \epsilon'} \delta(\xi_1-\xi_2) ,\\ ~[a_{\epsilon}(\xi_1),a^*_{\epsilon'}(\xi_2)] & = \delta_{\epsilon \epsilon'} \delta(\xi_1-\xi_2) , \\ \{b_{l,\epsilon}(\xi_1),b_{l',\epsilon'}(\xi_2)\} & = \{c_{l,\epsilon}(\xi_1),c_{l',\epsilon'}(\xi_2)\} = 0 , \\ ~[a_{\epsilon}(\xi_1),a_{\epsilon'}(\xi_2)] & = 0 , \\ \{b_{l,\epsilon}(\xi_1),c_{l',\epsilon'}(\xi_2)\} & = \{b_{l,\epsilon}(\xi_1),c^*_{l',\epsilon'}(\xi_2)\} = 0 , \\ \{b^*_{l,\epsilon}(\xi_1),c_{l',\epsilon'}(\xi_2)\} & = \{b^*_{l,\epsilon}(\xi_1),c^*_{l',\epsilon'}(\xi_2)\} = 0 , \\ ~[a_{\epsilon}(\xi_1),c_{\epsilon'}(\xi_2)] & = [a_{\epsilon}(\xi_1),b_{\epsilon'}(\xi_2)] = 0 , \\ ~[a_{\epsilon}(\xi_1),c^*_{\epsilon'}(\xi_2)] & = [a_{\epsilon}(\xi_1),b^*_{\epsilon'}(\xi_2)] = 0 ,\\ ~[a^*_{\epsilon}(\xi_1),c_{\epsilon'}(\xi_2)] & = [a^*_{\epsilon}(\xi_1),b_{\epsilon'}(\xi_2)] = 0 , \\ ~[a^*_{\epsilon}(\xi_1),c^*_{\epsilon'}(\xi_2)] & = [a^*_{\epsilon}(\xi_1),b^*_{\epsilon'}(\xi_2)] = 0 , \end{align*} with $l,l'\in\{1,2,3\}$, $\epsilon,\epsilon'=\pm$. Inserting \eqref{lept}--\eqref{Boson} into \eqref{Interac}, integrating with respect to $x$, and using the convention \begin{equation*} \int d\xi_1d\xi_2d\xi_3 = \sum_{s_1=\pm \frac{1}{2}} \sum_{s_2=\pm \frac{1}{2}} \sum_{\lambda=-1, 0, 1}\int d^3p_1d^3p_2d^3p_3, \end{equation*} we arrive at the formal expression \begin{align} H_I := \sum_{j=1}^4 H^{(j)}_I := \sum_{l=1}^3 \sum_{\epsilon=\pm} \int & \left\{ \left[ G^{(1)}_{l,\epsilon} (\xi_1,\xi_2,\xi_3)b^*_{l,\epsilon}(\xi_1)c^*_{l,-\epsilon}(\xi_2)a_{\epsilon}(\xi_3) + \mathrm{h.c.} \right]\right. \notag \\ & + \left[ G^{(2)}_{l,\epsilon} (\xi_1,\xi_2,\xi_3)b^*_{l,-\epsilon}(\xi_1)c^*_{l,\epsilon}(\xi_2)a^*_{\epsilon}(\xi_3) + \mathrm{h.c.} \right] \notag \\ &+\left[ G^{(3)}_{l,\epsilon} (\xi_1,\xi_2,\xi_3)b^*_{l,-\epsilon}(\xi_1)c_{l,-\epsilon}(\xi_2)a^*_{\epsilon}(\xi_3) + \mathrm{h.c.} \right] \notag \\ &+ \left. \left[ G^{(4)}_{l,\epsilon} (\xi_1,\xi_2,\xi_3)b^*_{l,\epsilon}(\xi_1)c_{l,\epsilon}(\xi_2) a_{\epsilon}(\xi_3) + \mathrm{h.c.} \right] \right\}d\xi_1d\xi_2d\xi_3 , \label{interaction_term} \end{align} where we set $-\epsilon=\mp$ if $\epsilon=\pm$. The kernels $G_{l,\epsilon}^{(j)}$, $j = 1 , \dots , 4$, are of the form \begin{align} &G_{l,\epsilon}^{(1)}( \xi_1 , \xi_2 , \xi_3 ) = f^{(1)}_{l,\epsilon,1}( \xi_1 ) f^{(1)}_{l,\epsilon,2}( \xi_2 ) f^{(1)}_{l,\epsilon,3}( \xi_3 ) \delta ( - p_1 - p_2 + p_3 ) , \label{eq:kernel1} \\ &G_{l,\epsilon}^{(2)}( \xi_1 , \xi_2 , \xi_3 ) = f^{(2)}_{l,\epsilon,1}( \xi_1 ) f^{(2)}_{l,\epsilon,2}( \xi_2 ) f^{(2)}_{l,\epsilon,3}( \xi_3 ) \delta ( p_1 + p_2 + p_3 ) , \\ &G_{l,\epsilon}^{(3)}( \xi_1 , \xi_2 , \xi_3 ) = f^{(3)}_{l,\epsilon,1}( \xi_1 ) f^{(3)}_{l,\epsilon,2}( \xi_2 ) f^{(3)}_{l,\epsilon,3}( \xi_3 ) \delta ( - p_1 + p_2 - p_3 ) , \\ &G_{l,\epsilon}^{(4)}( \xi_1 , \xi_2 , \xi_3 ) = f^{(4)}_{l,\epsilon,1}( \xi_1 ) f^{(4)}_{l,\epsilon,2}( \xi_2 ) f^{(4)}_{l,\epsilon,3}( \xi_3 ) \delta ( - p_1 + p_2 + p_3 ) , \label{eq:kernel4} \end{align} where the maps $p_i \mapsto f^{(j)}_{l,\epsilon,i}( \xi_i )$ are bounded in any compact set of $\mathbb{R}^3$. Their explicit expressions are given in Appendix \ref{app:interaction}. An important property of the interaction Hamiltonian \eqref{interaction_term} is that it preserves the lepton number in the sense that, formally, $N_{l^-} + N_{\nu_l} - N_{l^+} - N_{\bar{\nu}_l}$ commutes with $H_I$. Here, $N_p$ stands for the number operator corresponding to a particle $p$. We observe that the first term in \eqref{interaction_term}, $H_I^{(1)}$, describes explicitly processes like \eqref{process}, while $H_I^{(2)}$ prevents the bare vacuum from being a bound state, as expected from physics. To study a process like \eqref{process}, it is reasonable, in a first approximation, to keep only these first two terms, thus considering the simpler interaction Hamiltonian $H_I = H_I^{(1)} + H_I^{(2)}$. Under the assumption that the masses of the neutrinos vanish, we will make this approximation. The advantage is that the differences of number operators $N_{l^-} - N_{\bar{\nu}_l}$ and $N_{l^+} - N_{\nu_l}$ are preserved by the Hamiltonian. This property will be essential in some of our arguments. On the other hand, if we assume that the masses of the neutrinos do not vanish, our argument applies without requiring that such a quantity be conserved, and therefore the full interaction Hamiltonian \eqref{interaction_term} can be studied. Now, the free Hamiltonian is given by \begin{align} H_0 = & \sum_{l=1}^{3} \sum_{\epsilon=\pm} \int \omega_l^{(1)}(\xi_1) {b^*_{l,\epsilon}}(\xi_1){b_{l,\epsilon}}(\xi_1) d\xi_1 + \sum_{l=1}^{3} \sum_{\epsilon=\pm} \int \omega_l^{(2)}(\xi_2) {c^*_{l,\epsilon}}(\xi_2){c_{l,\epsilon}}(\xi_2) d\xi_2 \notag \\ & + \sum_{\epsilon=\pm} \int \omega^{(3)}(\xi_3) {a^*_{\epsilon}}(\xi_3){a_{\epsilon}}(\xi_3) d\xi_3, \label{Hlibre} \end{align} with the dispersion relations \begin{equation} \label{disprela} \omega^{(1)}_l(\xi_1) = \sqrt{p^2_1+m_l^2}, \quad \omega^{(2)}_l(\xi_2) = \sqrt{p^2_2+m_{\nu_l}^2}, \quad \omega^{(3)}(\xi_3) = \sqrt{p^2_3+m_{W^{\pm}}^2} . \end{equation} The total Hamiltonian is defined by \begin{equation} \label{total_hamiltonian} H = H_0 + H_I . \end{equation} Since the kernels $G^{(j)}_{l,\epsilon}$ are singular, the formal expressions \eqref{interaction_term}--\eqref{total_hamiltonian} do not define a self-adjoint operator in Fock space (see the next section for the precise definition of the Hilbert space that we consider). In order to obtain such a self-adjoint operator, following a standard procedure in constructive QFT (see e.g. \cite{GlJa77_01} and references therein), we introduce ultraviolet and spatial cut-offs in the interaction Hamiltonian. Of course, eventually, it would be desirable to find a renormalization procedure allowing one to remove those cut-offs. This constitutes an important open problem which is beyond the scope of this paper. Let $\Lambda > 0$ be a fixed ultraviolet parameter and let $B( 0 , \Lambda )$ denotes the ball centered at $0$ and of radius $\Lambda$ in $\mathbb{R}^3$. In the formal expression \eqref{interaction_term}, we introduce ultraviolet cut-offs, i.e., we replace $f^{(j)}_{l,\epsilon,i}( \xi_i )$ by $\chi_{ B( 0 , \Lambda ) }( p_i )f^{(j)}_{l,\epsilon,i}( \xi_i )$, for some smooth function $\chi_{ B( 0 , \Lambda ) }$ supported in $B( 0 , \Lambda )$, and we replace the Dirac delta function $\delta(p)$ by an approximation, $\delta_n(p) = n^3 \delta_1( np )$, for some smooth and compactly supported function $\delta_1$. The resulting kernels are still denoted by the same symbols $G^{(j)}_{l,\epsilon}$. In particular, $G^{(j)}_{l,\epsilon}$ are now square integrable. As will be shown in the next section, square integrability of the kernels is actually sufficient to prove that $H=H_0+H_I$ defines a self-adjoint operator in Fock space. Some of our results will be proven in the weak coupling regime. We will therefore study in this paper an abstract class of Hamiltonians given by \begin{equation} H = H_0 + g H_I , \label{total_hamiltonian_g} \end{equation} where $g$ is a real coupling parameter, $H_0$ is defined in \eqref{Hlibre}, and $H_I$ is given by \eqref{interaction_term} with abstract kernels $G^{(j)}_{l,\epsilon}$. The latter will always be supposed to be square integrable and, in some cases, stronger regularity assumptions on $G^{(j)}_{l,\epsilon}$ will be required. This will be made more precise below. The spectral theory of such models of the weak interaction has been studied in particular in \cite{AmGrGu07_01,AsBaFaGu11_01,BaFaGu16_01,BaFaGu16_02,BaFaGuappear,BaGu09_01} (see also \cite{BaDiGu04_01,Ta14_01} for related models of QED). Without entering into details, the results established in these references show that, for weak coupling, and under suitable assumptions on the kernels, $H$ is self-adjoint and has a ground state (i.e. $E:=\inf \sigma( H )$ is an eigenvalue of $H$), and the essential spectrum of $H$ coincides with the semi axis $\sigma_{\mathrm{ess}}(H) = [ E + m_{\nu} , \infty )$, with $m_\nu=\min(m_{\nu_e},m_{\nu_\mu},m_{\nu_\tau})$. In particular, if the masses of the neutrinos vanish, the ground state energy is an eigenvalue of $H$ embedded into its essential spectrum. Moreover, except for the ground state energy, the spectrum of $H$ below the electron mass is purely absolutely continuous. In this paper, we complement the previous spectral results by studying the structure of the essential spectrum in the whole semi-axis $[ E + m_{\nu} , \infty )$ (not only below the electron mass) and by relaxing the weak coupling assumption in the case where the masses of the neutrinos do not vanish. Our main purpose is then to study scattering theory for models of the form \eqref{interaction_term}--\eqref{total_hamiltonian_g} and, in particular, to prove asymptotic completeness. Scattering theory for models of non-relativistic matter coupled to a massive, bosonic quantum field -- massive Pauli-Fierz Hamiltonians -- has been considered by many authors. See, among others, \cite{Ar83_01,De03_01,DeGe99_01,DeGe04_01,Fr74_01,FrGrSc01_01,FrGrSc02_01,FrGrSc04_01,FrGrSc07_01,HuSp95_01,Ho68_01,Ho69_01,Sp97_01}; see also \cite{Am04_01} for fermionic Pauli-Fierz systems, \cite{DeGe00_01} for spatially cut-off $P(\varphi)_2$ Hamiltonians, \cite{GePa09_01} for abstract QFT Hamiltonians, and \cite{BoFaSi12_01,DeKu13_01,DeKu15_01,DeGrKu15_01,FaSi14_01,FaSi14_02,Ge02_01} for massless Pauli-Fierz Hamiltonians. A large part of the techniques used in the present paper are adapted from the ones developed in these references. The first step of the approach to scattering theory that we follow consists in establishing the existence and basic properties of the asymptotic creation and annihilation operators \begin{equation} a^{\pm,\sharp}_{\epsilon} (h) := \underset{t \to +\infty}{\slim} e^{ \pm \mathrm{i} t H} e^{\mp \mathrm{i} tH_0} a^\sharp_{\epsilon}(h) e^{\pm \mathrm{i} tH_0} e^{\mp \mathrm{i} t H } , \label{eq:def_asympt_a} \end{equation} for any $h$ in $L^2( \mathbb{R}^3 \times \{ -1,0,1 \} )$, where $a^{\sharp}$ stands for $a$ or $a^*$ and \begin{equation} a_\epsilon^*(h) = \int h(\xi_3) a_\epsilon^*(\xi_3) d\xi_3, \quad a_\epsilon(h) = \int \bar{h}(\xi_3) a_\epsilon(\xi_3) d\xi_3. \label{eq:def_a} \end{equation} The fermionic asymptotic creation and annihilation operators $b^{\pm,\sharp}_{l,\epsilon}(h)$ and $c^{\pm,\sharp}_{l,\epsilon}(h)$ are defined similarly, for $h \in L^2( \mathbb{R}^3 \times \{-\frac12,\frac12\})$. Let $\mathscr{H}$ be the Hilbert space of the model, defined as a tensor product of Fock spaces, see the next section for precise definitions. The space of asymptotic vacua is defined by \begin{equation*} \mathscr{K}^\pm := \{ u , \, d^\pm( h ) u = 0 \text{ for all asymptotic annihilation operator } d^\pm( h ) \}, \end{equation*} where $d^\pm(h)$ stands for either $a_\epsilon^\pm(h)$, with $h \in L^2( \mathbb{R}^3 \times \{ -1,0,1 \} )$, or $b^{\pm}_{l,\epsilon}(h)$ or $c^{\pm}_{l,\epsilon}(h)$, with $h \in L^2( \mathbb{R}^3 \times \{-\frac12,\frac12\})$. There is a natural definition of isometric wave operators \begin{equation} \Omega^\pm : \mathscr{K}^\pm \otimes \mathscr{H} \to \mathscr{H}, \label{eq:def_Omega_1} \end{equation} with the property that \begin{equation} \Omega^\pm ( \mathds{1} \otimes d^{\sharp}(h) ) = d^{\pm,\sharp}(h) \Omega^\pm, \label{eq:def_Omega_2} \end{equation} where, again, $d^{\sharp}(h)$ stands for any kind of creation or annihilation operator. Asymptotic completeness of $\Omega^\pm$ is the statement that $\Omega^\pm$ are unitary and that $\mathscr{K}^\pm = \mathscr{H}_{\mathrm{pp}}( H )$, where $\mathscr{H}_{\mathrm{pp}}( H )$ denotes the pure point spectral subspace of $H$. An interpretation of asymptotic completeness is that any evolving state $e^{-\mathrm{i}tH}u$, with $u \in \mathscr{H}$, can be decomposed, asymptotically as time $t$ goes to infinity, into a bound state together with asymptotically free particles. Recall that the parameter $j\in\{1,\dots,4\}$ labels the different interaction terms in \eqref{interaction_term} and that the index $l \in \{1,2,3\}$ labels the lepton families. Moreover $\epsilon=\pm$. In what follows, for shortness, we say that \begin{equation*} \text{``$G \in L^2$'' if, for all $j$, $l$ and $\epsilon$, $G_{l,\epsilon}^{(j)}$ is square integrable}. \end{equation*} Recall that $s_1,s_2$ denote the spin variables for fermions and that $\lambda$ denotes the spin variable for bosons. We say that \begin{equation*} \text{``$G \in \mathbb{H}^\mu$'' if, for all $j$, $l$, $\epsilon$, $s_1$, $s_2$ and $\lambda$, $G_{l,\epsilon}^{(j)}( s_1 , \cdot , s_2 , \cdot , \lambda , \cdot )$ belongs to the Sobolev space $\mathbb{H}^{\mu}( \mathbb{R}^9 )$}. \end{equation*} Remembering that the dispersion relation $\omega_l^{(i)}$, $i=1,2$, $l=1,2,3$ and $\omega^{(3)}$ are defined in \eqref{disprela}, we set \begin{align} & a_{(i),l} := \frac{\mathrm{i}}{2} \big ( \nabla_{p_i} \cdot \nabla \omega^{(i)}_l(p_i) + \nabla \omega^{(i)}_l( p_i ) \cdot \nabla_{p_i} \big ), \quad i=1,2 , \quad l = 1 ,2 , 3 , \label{eq:defa12} \\ & a_{(3)} := \frac{\mathrm{i}}{2} \big ( \nabla_{p_3} \cdot \nabla \omega^{(3)}(p_3) + \nabla \omega^{(3)}( p_3 ) \cdot \nabla_{p_3} \big ), \label{eq:defa3} \\ & b_{(i),l} := \frac{\mathrm{i}}{2} \big ( ( p_i \cdot \nabla \omega^{(i)}_l (p_i ) )^{-1} p_i \cdot \nabla_{p_i} + \nabla_{p_i} \cdot p_i ( p_i \cdot \nabla \omega^{(i)}_l (p_i ) )^{-1} \big ) , \quad i=1,2 , \quad l = 1 ,2 , 3 , \label{eq:defb12} \\ & b_{(3)} := \frac{\mathrm{i}}{2} \big ( ( p_3 \cdot \nabla \omega^{(3)} (p_3 ) )^{-1} p_3 \cdot \nabla_{p_3} + \nabla_{p_3} \cdot p_3 ( p_3 \cdot \nabla \omega^{(3)} (p_3 ) )^{-1} \big ), \label{eq:defb3} \end{align} as partial differential operators acting on $L^2( d\xi_1 d\xi_2 d\xi_3 )$. See Section \ref{sec:spectral} for details concerning the domains and properties of these operators. It should be noted that, in the case where the masses of the neutrinos vanish, we have that $a_{(2),l} = b_{(2),l}$. To shorten the statement of some of our results below, we introduce the notation ``$a_{(i),\cdot} G \in L^2$'' with the following meaning: for $i=1,2$, we say that \begin{equation*} \text{``$a_{(i),\cdot} G \in L^2$'' if, for all $j$, $l$ and $\epsilon$, $a_{(i),l} G_{l,\epsilon}^{(j)}$ is square integrable}, \end{equation*} and, for $i=3$, we say that \begin{equation*} \text{``$a_{(3),\cdot} G \in L^2$'' if, for all $j$, $l$ and $\epsilon$, $a_{(3)} G_{l,\epsilon}^{(j)}$ is square integrable}. \end{equation*} The notation $a_{(i),\cdot} a_{(i'),\cdot} G \in L^2$ is defined analogously, and likewise for $b_{(i),\cdot} G \in L^2$ and $b_{(i),\cdot} b_{(i'),\cdot} G \in L^2$. Given these conventions, the notation $|p_3|^{-1} a_{(i),\cdot} G \in L^2$ have an obvious meaning. Our main results can be stated as follows. \begin{Th}\label{thm:main} $\quad$ \begin{itemize} \item[(i)] Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ are positive and consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. Assume that \begin{equation*} G \in L^2 , \quad a_{(i),\cdot} G \in L^2, \quad i = 1,2,3, \end{equation*} and that $G \in \mathbb{H}^{1+\mu}$ for some $\mu > 0$. Then the wave operators $\Omega^\pm$ exist and are asymptotically complete. Suppose in addition that \begin{equation*} b_{(i),\cdot} G \in L^2, \quad i = 1,2,3, \quad b_{(i),\cdot} b_{(i'),\cdot} G \in L^2 , \quad i,i'=1,2,3. \end{equation*} Then there exists $g_0 > 0$, which does not depend on $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$, such that, for all $|g| \le g_0$, $H-E$ is unitarily equivalent to $H_0$. \item[(ii)]Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ vanish and consider the Hamiltonian $H = H_0 + g ( H_I^{(1)} + H_I^{(2)} )$ with $H_I^{(1)}$ and $H_I^{(2)}$ given by \eqref{interaction_term}. Assume that \begin{equation*} G \in L^2, \quad a_{(i),\cdot} G \in L^2, \quad |p_3|^{-1} a_{(i),\cdot} G \in L^2, \quad i = 1,2,3 , \end{equation*} and that $G \in \mathbb{H}^{1+\mu}$ for some $\mu > 0$. Then there exists $g_0>0$ such that, for all $|g| \le g_0$, the wave operators $\Omega^\pm$ exist and are asymptotically complete. Suppose in addition that \begin{equation*} b_{(i),\cdot} G \in L^2, \quad i = 1,2,3, \quad b_{(i),\cdot} b_{(i'),\cdot} G \in L^2 , \quad i,i'=1,2,3. \end{equation*} Then there exists $g'_0>0$ such that, for all $|g|\le g'_0$, $H-E$ is unitarily equivalent to $H_0$. \end{itemize} \end{Th} \begin{Rk} \begin{itemize} \item[(i)] As mentioned above, physically, the masses of the neutrinos are extremely small. We emphasize that the second part of Theorem \ref{thm:main} (i) holds for $|g|$ small enough, uniformly with respect to the masses of the neutrinos. As a consequence, from the observation that the Hamiltonian with massive neutrinos converges to that with massless neutrinos, in the norm resolvent sense, as the masses of the neutrinos go to $0$, one can deduce that, in the massless case, $H-E$ is approximately unitarily equivalent to $H_0$. This means that there exists a sequence of unitary operators $(U_n)$ such that $U_n H_0 U_n^* \to H-E$, as $n\to\infty$, in the norm resolvent sense. The conclusion of Theorem \ref{thm:main} (ii), which concerns also the massless case, is, of course, significantly stronger since it shows that $H-E$ and $H_0$ are unitarily equivalent (not only approximately unitarily equivalent). However, Theorem \ref{thm:main} (ii) only holds for the simplified Hamiltonian $H = H_0 + g ( H_I^{(1)} + H_I^{(2)} )$. \item[(ii)] Suppose that the kernels $G^{(j)}_{l,\epsilon}$ are of the forms \eqref{eq:kernel1}--\eqref{eq:kernel4}, where, as explained above, the $\delta$-functions are regularized and ultraviolet cut-offs are introduced. Suppose in addition that the functions $f^{(j)}_{l,\epsilon,i}$ belong to $\mathrm{C}^\infty( \mathbb{R}^3 \setminus \{ 0 \} )$ and satisfy \begin{equation*} \big | \partial^\alpha_{p_{i,\ell} } f^{(j)}_{l,\epsilon,i} ( \xi_1 , \xi_2 , \xi_3 ) \big| \le \mathrm{C}_\alpha |p_i|^{\nu-\alpha} , \quad \alpha \in \mathbb{N} , \quad p_i = ( p_{i,1} , p_{i,2} , p_{i,3} ), \end{equation*} for some real parameter $\nu$. Then, in the massive case (Theorem \ref{thm:main} (i)), the conditions $G \in L^2$ and $a_{(i),\cdot} G \in L^2$ are satisfied for any $\nu > - 3/2$, and $G \in \mathbb{H}^{1+\mu}$ is satisfied provided that $\nu > -1/2 + \mu$. The conditions $ b_{(i),\cdot} G \in L^2$ are satisfied for $\nu > 1/2$, and $b_{(i),\cdot} b_{(i'),\cdot} G \in L^2$ for $\nu > 5/2$. The same holds in the massless case (Theorem \ref{thm:main} (ii)), except that $a_{(2),\cdot} G \in L^2$ is satisfied for $\nu > -1/2$ (recall that $a_{(2),l} G = b_{(2),l} G$ in the massless case), and, if $i=2$ or $i'=2$, $b_{(i),\cdot} b_{(i'),\cdot} G \in L^2$ is satisfied for $\nu > 3/2$. Besides, $|p_3|^{-1} a_{(i),\cdot} G \in L^2$ holds as soon as $\nu > -1/2$. \item[(iii)] The assumption that $|p_3|^{-1} a_{(i),\cdot} G \in L^2$ in Theorem \ref{thm:main} (ii) can be replaced by $|p_1|^{-1} a_{(i),\cdot} G \in L^2$. \end{itemize} \end{Rk} To prove Theorem \ref{thm:main}, we follow the general approach of \cite{DeGe99_01,DeGe00_01,Am04_01} that we adapt to the present context. In addition to the fact that the model we study involves both bosons and fermions, one of our main achievements, compared to \cite{DeGe99_01,DeGe00_01,Am04_01}, is that the second part of our results in (i) hold with a restriction on the coupling constant which is uniform w.r.t. the masses of the neutrinos, and our results in (ii) are proven for an Hamiltonian involving massless particles. As mentioned above, for Pauli-Fierz Hamiltonians, contributions to scattering theory involving massless particles include \cite{Ge02_01,DeKu13_01,FaSi14_01,FaSi14_02,DeKu15_01,DeGrKu15_01}. In particular, for the massless spin-boson model, asymptotic completeness has been established in \cite{FaSi14_01,DeGrKu15_01}, using the uniform bound on the number of emitted particles proven in \cite{DeKu13_01}. This property, however, has not been proven for more general massless Pauli-Fierz Hamiltonians, yet. In our setting, controlling the number of emitted particles is made possible thanks to the fact that $N_{\mathrm{lept}} - N_{\mathrm{neut}}$ commutes with $H$, with $N_{\mathrm{lept}}$ the number of leptons and antileptons and $N_{\mathrm{neut}}$ the number of neutrinos and antineutrinos. In our proof of Theorem \ref{thm:main}, as in previous works, one of the main issues consists in finding a good choice of a ``conjugate operator'' $A$, such that the commutator $[H,\mathrm{i}A]$ is positive in the sense of Mourre \cite{Mo80_01}, or in a related weaker sense \cite{GeGeMo04_01,GeGeMo04_02}. From such a positive commutator estimate, one deduces spectral properties such that the absence of singular continuous spectrum for $H$, and suitable propagation estimates allowing one to establish asymptotic completeness. In the case where all particles are massive ((i) in Theorem \ref{thm:main}), our conjugate operator is the sum of the second quantizations of the operators \eqref{eq:defa12}--\eqref{eq:defa3}. This is a natural generalization of the conjugate operator chosen for instance in \cite{DeGe99_01,DeGe00_01,Am04_01}. Our main achievement here is that we show that the operators $H-E$ and $H_0$ are unitary equivalent for small $|g|$, uniformly in the masses of the neutrinos. To prove this, we use, in an essential way, the fact that neutrinos are fermions, together with a suitable application of the $N_\tau$ estimates of Glimm and Jaffe \cite{GlJa77_01} and an extension of Mourre's theory allowing for non-self-adjoint conjugate operators \cite{GeGeMo04_01}. The proof of Theorem \ref{thm:main} (ii) constitutes the main novelty compared to previous results in the literature. Here we cannot follow directly the approach of \cite{DeGe99_01,DeGe00_01,Am04_01} because of the presence of massless particles. To obtain a useful Mourre estimate, we combine singular Mourre's Theory \cite{GeGeMo04_01} together with an induction argument of \cite{DeGe99_01} and smallness of the coupling constant $g$. Using this Mourre estimate and the fact that $N_{\mathrm{lept}} - N_{\mathrm{neut}}$ commutes with $H$, we then establish propagation estimates. Our propagation estimates resemble those proven in \cite{DeGe99_01}, but with a different time-dependent propagation observable. Our choice of the (one-particle) propagation observable is inspired in part by that used in \cite{FaSi14_01,FaSi14_02}: it is a time-dependent modification of the usual ``position operator'', especially fitted to handle singularities due to the presence of massless particles. We do not take the same propagation observable as in \cite{FaSi14_01,FaSi14_02} because we follow a different approach, closer to that of \cite{DeGe99_01}, to prove asymptotic completeness. As in \cite{BoFaSi12_01,FaSi14_01,FaSi14_02}, an important ingredient to prove the propagation estimates is the control of the observable $\mathrm{d}\Gamma( |p_2|^{-1} )$ along the time evolution (where $p_2$ is the momentum of a neutrino). More precisely, we prove that, for suitable initial states, the expectation of this observable along the evolution grows slower than linearly in time $t$, which is crucial to estimate some remainder terms in the propagation estimates. Our paper is organized as follows. In Section \ref{sec:selfadjoint}, we show that the Hamiltonian \eqref{total_hamiltonian_g} defines a self-adjoint operator on a Hilbert space given as a tensor product of antisymmetric and symmetric Fock spaces. The result holds without any restriction on the coupling constant $g$. Section \ref{sec:selfadjoint} also contains the proof of some technical estimates that are used in the subsequent sections. In Section \ref{sec:spectral}, we recall results giving the existence of a ground state and the location of the essential spectrum of $H$, and we study the essential spectrum by means of suitable versions of Mourre's conjugate operator method. Section \ref{sec:propag} is devoted to the proof of several propagation estimates. Finally, in Section \ref{sec:AC}, we prove some properties of the asymptotic fields and wave operators, and we use the results of Sections \ref{sec:spectral} and \ref{sec:propag} to prove Theorem \ref{thm:main}. Technically, our main contributions, compared to previous works, are the proof of the Mourre estimate in Theorem \ref{MourreI_3}, and the proof of propagation estimates in Theorems \ref{th:propag_massless} and \ref{th:propag_massless_2}. For the convenience of the reader, the complete expression of the formal interaction Hamiltonian \eqref{interaction_term} is given in Appendix \ref{app:interaction}, and the definitions and properties of some operators in Fock space are recalled in Appendices \ref{sec:standard_def} and \ref{sec:asympt_a}. Technical computations are gathered in Appendix \ref{app:technical}. Throughout the paper, the notation $a \lesssim b$, for positive numbers $a$ and $b$, stands for $a \le \mathrm{C} b$ where $\mathrm{C}$ is a positive constant independent of the parameters involved. \\ \noindent \textbf{Acknowledgements}. We thank J.-C. Guillot for interesting remarks. J.F. is grateful to J.-M. Barbaroux and J.-C. Guillot for many discussions and fruitful collaborations. \section{Self-adjointness of the Hamiltonian and technical estimates}\label{sec:selfadjoint} In this section we show that the Hamiltonian of the model, formally defined in \eqref{total_hamiltonian_g}, identifies with a self-adjoint operator in an appropriate Hilbert space. The definition of the Hilbert space is given in Section \ref{2intro} and self-adjointness of the Hamiltonian is proven in Section \ref{sec:self-adj}. In Section \ref{subsec:estimates}, we prove some technical estimates that will be used in the next sections. \subsection{Hilbert space}\label{2intro} The Hilbert space of the model is a tensor product of Fock spaces for fermions and bosons. For fermions, we define $\Sigma_1:=\mathbb{R}^3 \times \{-\frac{1}{2},\frac{1}{2}\}$ and, for bosons, $\Sigma_2:=\mathbb{R}^3 \times \{-1,0,1\}$. The one-particle Hilbert space for fermions is $\mathfrak{h}_1:=L^2(\Sigma_1)$ and for bosons $\mathfrak{h}_2:=L^2(\Sigma_2)$. The anti-symmetric Fock space for fermions is denoted by $\mathfrak{F}_a:=\oplus^{\infty}_{n=0}\otimes^n_{a}\mathfrak{h}_1$, where $\otimes_a$ stands for the antisymmetric tensor product and where we use the usual convention $\otimes^0_{a}\mathfrak{h}_1=\mathbb{C}$. The symmetric Fock space for bosons is $\mathfrak{F}_s:=\oplus^{\infty}_{n=0}\otimes^n_{s}\mathfrak{h}_2$, where $\otimes_s$ stands for the symmetric tensor product and $\otimes^0_{s}\mathfrak{h}_2=\mathbb{C}$. Every family $l$ of leptons contains either an electron, a muon or a tau, the associated antiparticle, and a neutrino and its antineutrino. Consequently, the Hilbert space for each lepton family is \begin{equation*} \mathfrak{F}_l:=\bigotimes^4 \mathfrak{F}_a , \end{equation*} and we denote the full leptonic Hilbert space by \begin{equation*} \mathfrak{F}_L:=\bigotimes^3 \mathfrak{F}_l. \end{equation*} Analogously, the bosonic Hilbert space is given by \begin{equation*} \mathfrak{F}_W:=\bigotimes^2 \mathfrak{F}_s. \end{equation*} The total Hilbert space is \begin{equation*} \mathscr{H}:=\mathfrak{F}_W \otimes \mathfrak{F}_L. \end{equation*} In other words, $\mathscr{H}$ is the tensor product of $14$ Fock spaces, $2$ symmetric Fock spaces for the bosons $W^\pm$ and $12$ anti-symmetric Fock spaces for the fermions. The number operators for neutrinos and antineutrinos are defined by \begin{equation*} N_{\nu_l} := \sum_{\epsilon = \pm} \int c^*_{l,\epsilon}(\xi_2) c_{l,\epsilon}(\xi_2) d\xi_2 , \quad N_{\mathrm{neut}} := \sum_{l=1}^3 N_{\nu_l}, \end{equation*} and likewise \begin{equation*} N_{l}:=\sum_{\epsilon = \pm} \int b^*_{l,\epsilon}(\xi_1) b_{l,\epsilon}(\xi_1) d\xi_1, \quad N_{\mathrm{lept}} := \sum_{l=1}^3 N_l, \quad N_{W}:=\sum_{\epsilon = \pm} \int a^*_{\epsilon}(\xi_3) a_{\epsilon}(\xi_3) d\xi_3. \end{equation*} The total number operator is \begin{align*} N = N_{\mathrm{lept}} + N_{\mathrm{neut}} + N_W. \end{align*} \subsection{Self-adjointness }\label{sec:self-adj} Using the Kato-Rellich theorem together with $N_\tau$ estimates \cite{GlJa77_01}, it is proven in \cite{BaGu09_01} that the total Hamiltonian $H$ defined in \eqref{total_hamiltonian_g} is a self-adjoint operator in $\mathscr{H}$, with domain $\mathscr{D}\left( H\right)=\mathscr{D}\left( H_0\right) $, provided that the kernels $G_{l,\epsilon}^{(j)}$ are square integrable and $g \ll 1$. In this section, we extend this result to any value of $g$. Recall that the notation ``$G \in L^2$'' means that, for all $j$, $l$ and $\epsilon$, $G_{l,\epsilon}^{(j)}$ is square integrable. We denote by $\|G\|_2$ the sum over $j$, $l$ and $\epsilon$ of the $L^2$-norms of $G_{l,\epsilon}^{(j)}$, \begin{equation*} \| G \|_2 := \sum_{j,l,\epsilon} \big \| G_{l,\epsilon}^{(j)} \big \|_2. \end{equation*} \begin{Th} \label{Th2M} Suppose that $G \in L^2$. Then, for all $g \in \mathbb{R}$, the Hamiltonian $H$ in \eqref{total_hamiltonian_g} is self-adjoint with domain $\mathscr{D}\left( H\right)=\mathscr{D}\left( H_0\right) $. \end{Th} The proof of Theorem \ref{Th2M} will be a consequence of the following two lemmas. The first one is a direct application of the $N_\tau$ estimates of Glimm and Jaffe (see \cite[Proposition 1.2.3(c)]{GlJa77_01}). \begin{lem} \label{lemSCW_1} Suppose that $G \in L^2$. Then \begin{equation*} \|H_I (N_{\mathrm{lept}} + N_W + 1)^{-1} \| \lesssim \| G \|_2. \end{equation*} \end{lem} \begin{proof} Consider for instance the term \begin{equation} H_{I,1,+}^{(1)} := \int G^{(1)}_{1,+}(\xi_1, \xi_2, \xi_3) b^*_{1,+}(\xi_1)c^*_{1,-}(\xi_2)a_{+}(\xi_3) d \xi_1 d \xi_2 d \xi_3 , \label{H_Il1+-} \end{equation} occurring in $H_I$. Proceeding as in Proposition 1.2.3(b) of \cite{GlJa77_01}, one easily verifies that \begin{equation*} \Big\|N_{\mathrm{lept}}^{-\frac12} H_{I,1,+}^{(1)} N_W^{-\frac12} \Big \| \lesssim \| G^{(1)}_{1,+} \|_2. \end{equation*} Using that $N_{\mathrm{lept}}^{-\frac12} H_{I,1,+}^{(1)} = H_{I,1,+}^{(1)} (N_{\mathrm{lept}}+1)^{-\frac12}$, we obtain that \begin{equation*} \| H_{I,1,+}^{(1)} (N_{\mathrm{lept}} + N_W + 1)^{-1} \| \lesssim \| G \|_2. \end{equation*} The other terms occurring in $H_I$ can be treated in an analogous way. \end{proof} The second lemma is a slight generalization of \cite[Proposition 1.2.3(c)]{GlJa77_01}. \begin{lem} \label{lemSCW} Suppose that for all $j$, $l$, $\epsilon$, $s_1$, $s_2$ and $\lambda$, $G_{l,\epsilon}^{(j)}( s_1 , \cdot , s_2 , \cdot , \lambda , \cdot )$ belongs to the Schwartz space $\mathscr{S}( \mathbb{R}^9 )$. Then \begin{equation*} \|H_I (N_W+1)^{-\frac{1}{2}}\| \le \mathrm{C}(G ) , \end{equation*} where $\mathrm{C}(G)$ is a positive constant depending on $G_{l,\epsilon}^{(j)}$. \end{lem} \begin{proof} Let $\{e_i\}$ be an orthonormal basis of $L^2(\mathbb{R})$ composed of eigenvectors, corresponding to the eigenvalues $\lambda_{i}=(2i+1)$, of the one-dimensional harmonic oscillator $h_{\mathrm{ho}} := - \frac{d^2}{dx^2} + x^2$. We consider the orthonormal basis $\{ e_{i_1} \otimes \cdots \otimes e_{i_9} \}$ in $L^2( \mathbb{R}^9 )$. Below we use the notation $\mathrm{i}_1 = ( i_1 , i_2 , i_3)$, $\mathrm{i}_2 = ( i_3 , i_4 , i_5)$, $\mathrm{i}_3 = ( i_6 , i_7 , i_9)$ and a sum over $(\mathrm{i}_1,\mathrm{i}_2,\mathrm{i}_3)$ corresponds to a sum over $( i_1 , \dots , i_9 ) \in \mathbb{N}^9$. Moreover $e_{\mathrm{i}_1} := e_{i_1}\otimes e_{i_2} \otimes e_{i_3}$ and likewise for $e_{\mathrm{i}_2}$ and $e_{\mathrm{i}_3}$. As in the proof of the previous lemma, we consider for instance the term $H_{I,1,+}^{(1)}$ occurring in $H_I$, see \eqref{H_Il1+-}. Let \begin{equation*} H_{I,1,+}^{(1)}(s_1,s_2,\lambda) := \int G^{(1)}_{1,+}(s_1, p_1 , s_2 , p_2, \lambda , p_3) b^*_{1,+}( p_1 , s_1 )c^*_{1,-}( p_2 , s_2 )a_{+}( p_3 , \lambda ) d p_1 d p_2 d p_3 , \end{equation*} so that \begin{equation*} H_{I,1,+}^{(1)} = \sum_{s_1,s_2,\lambda} H_{I,1,+}^{(1)}(s_1,s_2,\lambda). \end{equation*} To prove the lemma, it suffices to verify that, for any fixed $s_1$, $s_2$ and $\lambda$, \begin{equation*} \| H_{I,1,+}^{(1)}(s_1,s_2,\lambda) (N_W+1)^{-\frac{1}{2}}\| \le \mathrm{C}(G ) . \end{equation*} Decomposing $G^{(1)}_{1,+}(\cdot,s_1,\cdot,s_2,\cdot,s_3)$ into the orthonormal basis $\{ e_{i_1} \otimes \cdots \otimes e_{i_9} \}$, we see that \begin{equation*} H_{I,1,+}^{(1)}(s_1,s_2,\lambda) = \sum_{\mathrm{i}_1,\mathrm{i}_2,\mathrm{i}_3} \alpha_{\mathrm{i}_1,\mathrm{i}_2,\mathrm{i}_3} b^*_{1,+,s_1}(e_{\mathrm{i}_1}) \otimes c^*_{1,-,s_2}(e_{\mathrm{i}_2}) \otimes a_{+,\lambda}(e_{\mathrm{i}_3}) , \end{equation*} where we have set $\alpha_{\mathrm{i}_1,\mathrm{i}_2,\mathrm{i}_3} := \langle e_{\mathrm{i}_1} \otimes e_{\mathrm{i}_2} \otimes e_{\mathrm{i}_3} , G^{(1)}_{1,+}(\cdot,s_1,\cdot,s_2,\cdot,s_3) \rangle$, $b^*_{1,+,s_1}(h)= \int_{\mathbb{R}^3} h(p_1) b^*_{1,+}(s_1,p_1) dp_1$, for $h \in L^2(\mathbb{R}^3)$, and likewise for $c^*_{1,-,s_2}(h)$ and $a_{+,\lambda}(h)$. This yields \begin{align*} & \left\| H_{I,1,+}^{(1)}(s_1,s_2,\lambda) (N_W+1)^{-\frac{1}{2}} \right\| \\ & = \sum_{\mathrm{i}_1,\mathrm{i}_2,\mathrm{i}_3} \left\| \alpha_{\mathrm{i}_1,\mathrm{i}_2,\mathrm{i}_3} b^*_{1,+,s_1}(e_{\mathrm{i}_1}) \otimes c^*_{1,-,s_2}(e_{\mathrm{i}_2}) \otimes a_{+,\lambda}(e_{\mathrm{i}_3}) (N_W+1)^{-\frac{1}{2}} \right\| \leq \sum_{\mathrm{i}_1,\mathrm{i}_2,\mathrm{i}_3} | \alpha_{\mathrm{i}_1,\mathrm{i}_2,\mathrm{i}_3} | . \end{align*} Observe that \begin{align*} \alpha_{\mathrm{i}_1,\mathrm{i}_2,\mathrm{i}_3} = \Big ( \prod_{\ell=1}^9 \frac{1}{(2i_\ell+1)} \Big ) \big \langle e_{\mathrm{i}_1} \otimes e_{\mathrm{i}_2} \otimes e_{\mathrm{i}_3} , ( \otimes_{\ell=1}^9 h_{\mathrm{ho}} ) G^{(1)}_{1,+}(\cdot,s_1,\cdot,s_2,\cdot,s_3) \big \rangle. \end{align*} The Cauchy-Schwarz inequality then gives \begin{align*} \sum_{\mathrm{i}_1,\mathrm{i}_2,\mathrm{i}_3} | \alpha_{\mathrm{i}_1,\mathrm{i}_2,\mathrm{i}_3} | \le \Big ( \sum_{\mathrm{i}_1,\mathrm{i}_2,\mathrm{i}_3} \Big ( \prod_{\ell=1}^9 \frac{1}{(2i_\ell+1)^2} \Big ) \Big )^{\frac12} \big\|( \otimes_{\ell=1}^9 h_{\mathrm{ho}} ) G^{(1)}_{1,+}(\cdot,s_1,\cdot,s_2,\cdot,s_3) \big \|_{L^2(\mathbb{R}^9)} , \end{align*} which concludes the proof. \end{proof} Now we are ready to prove Theorem \ref{Th2M}. \begin{proof}[Proof of Theorem \ref{Th2M}.] In this proof, we underline the dependence of the interaction Hamiltonian on the kernels $G_{l,\epsilon}^{(j)}$ by writing $H_I = H_I( G )$. According to Lemma \ref{lemSCW_1}, for $\Psi \in \mathscr{D} (N_{\mathrm{lept}} + N_W)$, there exists $a>0$ such that, for all $G \in L^2$, \begin{equation*} \|H_I(G) \Psi\| \leq a \| G \|_2 \left( \| (N_{\mathrm{lept}} + N_W ) \Psi \| + \|\Psi\| \right). \end{equation*} Let $\delta > 0$. There exist $G_\delta \in L^2$ such that for all $j$, $l$, $\epsilon$, $s_1$, $s_2$ and $\lambda$, $G_{\delta,l,\epsilon}^{(j)}( s_1 , \cdot , s_2 , \cdot , \lambda , \cdot )$ belongs to the Schwartz space $\mathscr{S}( \mathbb{R}^9 )$ and \begin{equation*} \| G - G_\delta \|_2 \leq \frac{\delta}{a}. \end{equation*} Hence \begin{equation} \| ( H_I(G) - H_I(G_\delta) ) \Psi\| \leq \varepsilon \left( \| (N_{\mathrm{lept}} + N_W ) \Psi \| + \|\Psi\| \right). \label{a1} \end{equation} Moreover, by Lemma \ref{lemSCW}, \begin{equation} \| H_I( G_\delta ) \Psi\| \leq \mathrm{C}( G_\delta ) \| (N_W+1)^{\frac{1}{2}} \Psi\| , \label{a2} \end{equation} with $\mathrm{C}( G_\delta )>0$. For $\mu > 0$, we have that \begin{align*} \| (N_W+1)^{\frac{1}{2}} \Psi\|^2 \le \mu \| N_W \Psi \|^2 + ( ( 4 \mu )^{-1} + 1 ) \| \Psi \|^2 \le \big ( \mu^{\frac12} \| N_W \Psi \| + ( ( 4 \mu )^{-1} + 1 )^{\frac12} \| \Psi \| \big )^2. \end{align*} Inserting this into \eqref{a2} and choosing $\mu^{\frac12} = \varepsilon \mathrm{C}( G_\delta )^{-1}$, this implies that \begin{equation} \| H_I( G_\delta ) \Psi\| \leq \delta \|N_W \Psi\| + \mathrm{c}_\delta \|\Psi\| , \label{a3} \end{equation} for some positive constant $\mathrm{c}_\delta$. Equations \eqref{a1} and \eqref{a3} show that $H_I(G)$ is relatively $(N_{\mathrm{lept}} + N_W)$-bounded with relative bound $0$. Since the masses $m_l$ and $m_W$ are positive, it is not difficult to deduce that $N_{\mathrm{lept}} + N_W$ is relatively $H_0$-bounded. Therefore $H_I(G)$ is relatively $H_0$-bounded with relative bound $0$. Applying the Kato-Rellich theorem concludes the proof. \end{proof} \subsection{Technical estimates}\label{subsec:estimates} This section is devoted to some technical lemmas and properties which will be used later. Proofs and notations in this section are close to those of \cite{DeGe99_01}. \subsubsection{Number-energy estimates} We begin with the following result in the case where all particles are supposed to be massive. \begin{lem}\label{lemV1Numbestimates} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ are positive and consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. Assume that $G \in L^2$. \begin{itemize} \item[(i)] For all $m \in \mathbb{Z}$, uniformly for z in a compact set of $\left\{ z \in \mathbb{C}, \pm |\Im{z}| >0 \right\}$, the operator $(N+1)^{-m}(H-z)^{-1}(N+1)^{m+1}$ extends to a bounded operator satisfying \begin{equation*} \big\|(N+1)^{-m}(H-z)^{-1}(N+1)^{m+1}\big\| = \mathcal{O}(|\Im(z)|^{ - \alpha_{m}}) , \end{equation*} where $\alpha_{m}$ denotes an integer depending on $m$. \item[(ii)] Let $\chi \in \mathrm{C}^{\infty}_0(\mathbb{R})$. Then, for all $m,p \in \mathbb{N}$, $N^m \chi(H) N^p$ extends to a bounded operator. \end{itemize} \end{lem} \begin{proof} As in the proof of Theorem \ref{Th2M}, we underline the dependence of the interaction Hamiltonian on the kernels $G_{l,\epsilon,\epsilon'}^{(j)}$ by writing $H_I = H_I( G )$. First, observe that \begin{equation*} [H,N] = [H_I(G),N]. \end{equation*} A direct computation gives \begin{align} & [H_I^{(1)}(G) , N_{\mathrm{lept}} ] = [H_I^{(1)}(G) , N_{\mathrm{neut}} ] = - [H_I^{(1)}(G) , N_{W} ] = \mathrm{i} H_I^{(1)}( \mathrm{i} G ), \label{commut_1} \\ & [H_I^{(2)}(G) , N_{\mathrm{lept}} ] = [H_I^{(2)}(G) , N_{\mathrm{neut}} ] = [H_I^{(2)}(G) , N_{W} ] = \mathrm{i} H_I^{(2)}( \mathrm{i} G ), \label{commut_2} \\ & [H_I^{(3)}(G) , N_{\mathrm{lept}} ] = - [H_I^{(3)}(G) , N_{\mathrm{neut}} ] = [H_I^{(3)}(G) , N_{W} ] = \mathrm{i} H_I^{(3)}( \mathrm{i} G ), \label{commut_3} \\ & [H_I^{(4)}(G) , N_{\mathrm{lept}} ] = - [H_I^{(4)}(G) , N_{\mathrm{neut}} ] = - [H_I^{(4)}(G) , N_{W} ] = \mathrm{i} H_I^{(4)}( \mathrm{i} G ). \label{commut_4} \end{align} In particular, since $G \in L^2$, Lemma \ref{lemSCW_1} together with the fact that $N_{\mathrm{lept}} + N_W$ is relatively $H$-bounded show that $\| [H,N](H+z)^{-1} \| = \mathcal{O}( |\Im{z}|^{-1} )$. Likewise, \begin{align} \| \mathrm{ad}_N^j(H) (H-z)^{-1} \| = \mathcal{O}( |\Im{z}|^{-1} ), \quad j \in \mathbb{N}. \label{est_commutator} \end{align} Next, for $m \in \mathbb{N}$, commuting $N^m$ through $(H-z)^{-1}$, we obtain that \begin{equation*} (H-z)^{-1}N^m = N^m (H-z)^{-1}+ \sum_{k=1}^m N^{m-l}(H-z)^{-1}B_{m,k}(z) , \end{equation*} where, by \eqref{est_commutator}, the operator $B_{m,k}(z)$ satisfies $\|B_{m,k}(z)\| = \mathcal{O}( |\Im{z}|^{-c_{m,k}} )$, with $c_{m,k}$ a positive integer. Therefore \begin{align*} (N+1)^{-m} (H-z)^{-1} (N+1)^{m+1} &= (N+1) (H-z)^{-1} + (N+1)^{-m} \sum_{i=1}^m N^{m-l} (H-z)^{-1} B_{m,k}(z) \\ &= \mathcal{O}( |\Im z|^{-\alpha_m}), \end{align*} where in the second equality we used that $N$ is $H$-bounded. This proves (i). To prove (ii), let $\chi \in \mathrm{C}^{\infty}_0(\mathbb{R})$ and $m,p \in \mathbb{N}$. Then \begin{align*} N^m \chi(H) N^p = \Big ( \prod_{k=0}^{m-1} N^{m-k} (H+\mathrm{i})^{-1} N^{k-m+1} \Big ) (H+\mathrm{i})^m \chi(H)(H+\mathrm{i})^p \Big ( \prod_{k=0}^{p-1} N^{k-p+1} (H+\mathrm{i})^{-1} N^{p-k} \Big ) , \end{align*} where we used the convention $\prod_{k=0}^{m-1} A_k := A_0 A_1 \dots A_{m-1}$ for any operators $A_0,\dots,A_{m-1}$. Hence (i) implies (ii). \end{proof} In the case where the masses of the neutrinos vanish, the statement of the previous lemma can be modified as follows. \begin{lem}\label{lemV1NumbestimatesV2} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ vanish and consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. Assume that $G \in L^2$. \begin{itemize} \item[(i)] For all $m \in \mathbb{Z}$, uniformly for z in a compact set of $\left\{ z \in \mathbb{C}, \pm |\Im{z}| >0 \right\}$, the operator $(N+1)^{-m}(H-z)^{-1}(N+1)^{m}$ extends to a bounded operator satisfying \begin{equation*} \big\|(N+1)^{-m}(H-z)^{-1}(N+1)^{m}\big\| = \mathcal{O}(|\Im(z)|^{ - \alpha_{m}}) , \end{equation*} where $\alpha_{m}$ denotes an integer depending on $m$. Moreover, the operator $(N_{\mathrm{lept}} + N_W+1)^{-m}(H-z)^{-1}(N_{\mathrm{lept}} + N_W+1)^{m+1}$ extends to a bounded operator satisfying \begin{equation*} \big\|(N_{\mathrm{lept}} + N_W +1)^{-m}(H-z)^{-1}( N_{\mathrm{lept}} + N_W + 1)^{m+1}\big\| = \mathcal{O}(|\Im(z)|^{ - \beta_{m}}) , \end{equation*} where $\beta_{m}$ denotes an integer depending on $m$. \item[(ii)] Let $\chi \in \mathrm{C}^{\infty}_0(\mathbb{R})$. Then, for all $m,p \in \mathbb{N}$, $(N_{\mathrm{lept}} + N_W)^m \chi(H) (N_{\mathrm{lept}} + N_W)^p$ extends to a bounded operator. \end{itemize} \end{lem} \begin{proof} The proof is analogous to that of Lemma \ref{lemV1Numbestimates}. The only difference is that $N_{\mathrm{lept}}$ and $N_W$ are still relatively $H$-bounded, but $N_{\mathrm{neut}}$ is not anymore. \end{proof} \subsubsection{Number energy estimates in the ``extended'' setting} In the remainder of this section, we give results that concern an auxiliary ``extended Hamiltonian''. The latter is introduced similarly as in \cite{DeGe99_01}. It will be used in Section \ref{sec:AC}. The ``extended Hilbert space'' is \begin{equation*} \mathscr{H}^{\mathrm{ext}} := \mathscr{H} \otimes \mathscr{H} \end{equation*} and the extended Hamiltonian, acting on $\mathscr{H}^{\mathrm{ext}}$, is defined by \begin{equation*} H^{\mathrm{ext}} := H\otimes \mathds{1}+\mathds{1}\otimes H_0. \end{equation*} More details on the extended objects are given in Appendix \ref{sec:standard_def}. Roughly speaking, the idea is that the first component of $\mathscr{H}^{\mathrm{ext}}$ corresponds to bound states, while the second component corresponds to states localized near infinity. The number operators in the extended setting are defined by \begin{equation*} N_{\mathrm{lept},0} := N_{\mathrm{lept}}\otimes\mathds{1}, \quad N_{\mathrm{lept},\infty}:=\mathds{1}\otimes N_{\mathrm{lept}}, \end{equation*} and likewise for $N_{\mathrm{neut}}$, $N_W$ and the total number operator $N$. The proof of Lemma \ref{lemV1Numbestimates} can be adapted in a straightforward way to obtain the following results. \pagebreak[2] \begin{lem} \label{lemV2Numbestimates} $\quad$ \begin{itemize} \item[(i)] Under the conditions of Lemma \ref{lemV1Numbestimates}, for all $m \in \mathbb{Z}$, we have that \begin{equation*} \big\|(N_0+N_\infty+1)^{-m} (H^{\mathrm{ext}} - z )^{-1} ( N_0 + N_\infty + 1 )^{m+1} \big \| = \mathcal{O}(|\Im(z)|^{-\alpha_{m}}) , \end{equation*} uniformly for z in a compact set of $\left\{ z \in \mathbb{C}, \pm |\Im{z}| >0 \right\}$, where $\alpha_{m}$ denotes an integer depending on $m$. Moreover, for all $\chi \in \mathrm{C}^{\infty}_0(\mathbb{R})$ and $m,p \in \mathbb{N}$, $(N_0 + N_\infty )^m \chi(H^{\mathrm{ext}}) (N_0 + N_\infty )^p$ extends to a bounded operator. \item[(ii)] Under the conditions of Lemma \ref{lemV1NumbestimatesV2}, for all $m \in \mathbb{Z}$, we have that \begin{equation*} \big\|(N_0+N_\infty+1)^{-m} (H^{\mathrm{ext}} - z )^{-1} ( N_0 + N_\infty + 1 )^{m} \big \| = \mathcal{O}(|\Im(z)|^{-\alpha_{m}}) , \end{equation*} uniformly for z in a compact set of $\left\{ z \in \mathbb{C}, \pm |\Im{z}| >0 \right\}$, where $\alpha_{m}$ denotes an integer depending on $m$. Moreover, \begin{align*} & \big\| ( N_{\mathrm{lept},0}+N_{W,0} + N_{\mathrm{lept},\infty}+N_{W,\infty} +1 )^{-m} (H^{\mathrm{ext}}-z)^{-1} \\ & \qquad \qquad \qquad (N_{\mathrm{lept},0}+N_{W,0} + N_{\mathrm{lept},\infty}+N_{W,\infty} +1)^{m+1} \big \| = \mathcal{O}(|\Im(z)|^{-\alpha_{m}}), \end{align*} and for all $\chi \in \mathrm{C}^{\infty}_0(\mathbb{R})$ and $m,p \in \mathbb{N}$, $( N_{\mathrm{lept},0}+N_{W,0} + N_{\mathrm{lept},\infty}+N_{W,\infty} )^m \chi(H^{\mathrm{ext}}) ( N_{\mathrm{lept},0}+N_{W,0} + N_{\mathrm{lept},\infty}+N_{W,\infty} )^p$ extends to a bounded operator. \end{itemize} \end{lem} Let $(j_{1,0}, \dots, j_{14,0}) \in \mathrm{C}^{\infty}_0(\mathbb{R}^3)$ and $(j_{1,\infty}, \dots, j_{14,\infty}) \in \mathrm{C}^{\infty}(\mathbb{R}^3)$ be families of function satisfying $j_{\ell,0} \ge 0$, $j_{\ell,\infty} \ge 0$, $j^2_{\ell,0}+j^2_{\ell,\infty} = 1$ and $ j_{\ell,0}=1 $ near $0$. Recall that the total Hilbert space $\mathscr{H}$ is a tensor product of $14$ Fock spaces. We set $j=\left((j_{1,0},j_{1,\infty}), \dots, (j_{14,0},j_{14,\infty}) \right)$ and, for any $R\geq 1$, $j^R = ((j^R_{1,0},j^R_{1,\infty}), \dots, (j^R_{14,0},j^R_{14,\infty}) )$, with $j^R_{\ell,\sharp}=j_{\ell,\sharp}(\frac{x}{R})$, $x = \mathrm{i} \nabla_{p}$, and $p$ is the momentum of the particle labeled by $\ell \in \{ 1 , \dots , 14 \}$. The definition of the partition of unity $\check{\Gamma}(j^R) : \mathscr{H} \to \mathscr{H}^{\mathrm{ext}} $, adapted from \cite{DeGe99_01} and \cite{Am04_01}, is recalled in Appendix \ref{sec:standard_def}. Recall that we write ``$G \in \mathbb{H}^\mu$'' if, for all $j$, $l$, $\epsilon$, $s_1$, $s_2$ and $\lambda$, $G_{l,\epsilon}^{(j)}( s_1 , \cdot , s_2 , \cdot , \lambda , \cdot )$ belongs to the Sobolev space $\mathbb{H}^{\mu}( \mathbb{R}^9 )$. Note that it is equivalent to assume that, for all $j$, $l$, $\epsilon$, $s_1$, $s_2$ and $\lambda$, \begin{equation*} \forall R >1, \, \big \| \mathds{1}_{[R,\infty)}(|x_i|)G_{l,\epsilon}^{(j)}( s_1 , \cdot , s_2 , \cdot , \lambda , \cdot )\big\|_2 \lesssim R^{-\mu}, \end{equation*} where $x_i = \mathrm{i} \nabla_{p_i}$, $i=1,2,3$. \begin{lem} \label{numberestimates3} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ are positive and consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. Assume that $G \in L^2$. Let $j^R$ be defined as above. Then \begin{equation}\label{1st_estim} (H^{\mathrm{ext}}+\mathrm{i})^{-1} \check{\Gamma}(j^R)-\check{\Gamma}(j^R)(H+\mathrm{i})^{-1} = o(R^{0}) , \quad R \to \infty. \end{equation} In particular, for any $\chi, \chi' \in \mathrm{C}_0^\infty( \mathbb{R} )$, we have that \begin{equation}\label{3rd_estim} \big \{ \chi(H^{\mathrm{ext}}) \check{\Gamma}(j^R)-\check{\Gamma}(j^R)\chi(H) \big \} \chi'(H) = o(R^0), \quad R \to \infty. \end{equation} If $G \in \mathbb{H}^\mu$ with $\mu > 0$, then \begin{equation}\label{2nd_estim} (H^{\mathrm{ext}}+\mathrm{i})^{-1} \check{\Gamma}(j^R)-\check{\Gamma}(j^R)(H+\mathrm{i})^{-1} = \mathcal{O}(R^{-\min(1,\mu)}) , \quad R \to \infty , \end{equation} and in particular, for any $\chi, \chi' \in \mathrm{C}_0^\infty( \mathbb{R} )$ and $\mu \ge 1$, we have that \begin{equation}\label{3rd_estim_2} \big \{ \chi(H^{\mathrm{ext}}) \check{\Gamma}(j^R)-\check{\Gamma}(j^R)\chi(H) \big \} \chi'(H) \in \mathcal{O}(R^{-1}) , \quad R \to \infty. \end{equation} \end{lem} \begin{proof} We first note that \begin{align*} & (H^{\mathrm{ext}}+\mathrm{i})^{-1} \check{\Gamma}(j^R)-\check{\Gamma}(j^R)(H+\mathrm{i})^{-1} = (H^{\mathrm{ext}}+\mathrm{i})^{-1} \big ( \check{\Gamma}(j^R) H - H^{\mathrm{ext}} \check{\Gamma}(j^R) \big ) (H+\mathrm{i})^{-1} , \end{align*} and a direct computation gives \begin{equation*} H^{\mathrm{ext}}_0 \check{\Gamma}(j^R) - \check{\Gamma}(j^R)H_0 = \mathrm{d} \check{\Gamma} (j^R,[\omega,j^R]) , \end{equation*} where the operator $\mathrm{d} \check{\Gamma} (j^R,[\omega,j^R])$ is defined in Appendix \ref{sec:standard_def}. It follows from Lemma \ref{estimationN2} that \begin{align*} \left\| \mathrm{d} \check{\Gamma} (j^R,[\omega,j^R]) ( N_{0} + N_{\infty} +1 )^{-1} \right\| & = \left\| ( N_{0} + N_{\infty} +1 )^{-\frac12} \mathrm{d} \check{\Gamma} (j^R,[\omega,j^R]) ( N + 1 )^{-\frac12} \right\| \\ & \le \big \| [\omega,j^R]^*[\omega,j^R] \big \|^{\frac12} = \mathcal{O}( R^{-1} ). \end{align*} Moreover, we have that \begin{equation*} ( H_I(G)\otimes\mathds{1}) \check{\Gamma}(j^R)-\check{\Gamma}(j^R) H_I(G) = \sum_{j=1}^4 ( H_I^{(j)}(G) \otimes\mathds{1}) \check{\Gamma}(j^R)-\check{\Gamma}(j^R) H_I^{(j)}(G). \end{equation*} Considering for instance the term \begin{equation*} H_{I,1,+}^{(1)}(G) = \int G^{(1)}_{1,+}(\xi_1, \xi_2, \xi_3) b^*_{1,+}(\xi_1)c^*_{1,-}(\xi_2)a_{+}(\xi_3) d \xi_1 d \xi_2 d \xi_3 , \end{equation*} occurring in $H_I^{(1)}(G)$. Using the intertwining properties of Lemma \ref{lem:B6}, one verifies that \begin{equation*} ( H_{I,1,+}^{(1)}(G) \otimes\mathds{1}) \check{\Gamma}(j^R)-\check{\Gamma}(j^R) H_{I,1,+}^{(1)}(G) \end{equation*} can be expressed as a sum of operators of the form \begin{equation*} \int j^R( x_1 , x_2 , x_3 ) G^{(1)}_{1,+}(\xi_1, \xi_2, \xi_3) b^{*,\sharp}_{1,+}(\xi_1) c^{*,\sharp}_{1,-}(\xi_2) a^\sharp_{+}(\xi_3) d \xi_1 d \xi_2 d \xi_3 , \end{equation*} where $j^R(x_1,x_2,x_3)$ stands for either $1-j^R_{\ell_1,0}(x_{1})j^R_{\ell_2,0}(x_2)j^R_{\ell_3,0}(x_3)$ or $j^R_{\ell_1,\sharp}(x_1)j^R_{\ell_2,\sharp}(x_2)j^R_{\ell_3,\sharp}(x_3)$, with at least one of the $j^R_{\ell,\sharp}$ equal to $j^R_{\ell,\infty}$. Moreover $a^\sharp_{+}(\xi_3)$ stands for either $a_{+}(\xi_3) \otimes \mathds{1}$ or $\mathds{1} \otimes a_{+}(\xi_3)$ and likewise for $b^{*,\sharp}_{1,+}(\xi_1)$ and $c^{*,\sharp}_{1,-}(\xi_2)$. Therefore, proceeding as in Lemma \ref{lemSCW_1}, one deduces that if $G \in L^2$, then \begin{equation*} \big\|\big\{(H_I(G)\otimes\mathds{1}) \check{\Gamma}(j^R)-\check{\Gamma}(j^R) H_I(G) \big\} (N_0+N_\infty+1)^{-1}\big\| = o(R^{0}) , \quad R \to \infty. \end{equation*} Similarly, if $G \in \mathbb{H}^\mu$, we obtain that \begin{equation*} \big\|\big\{(H_I(G)\otimes\mathds{1}) \check{\Gamma}(j^R)-\check{\Gamma}(j^R) H_I(G) \big\} (N_0+N_\infty+1)^{-1}\big\| = \mathcal{O}(R^{-\mu}), \quad R \to \infty. \end{equation*} Putting together the previous estimates proves \eqref{1st_estim} and \eqref{2nd_estim}. To prove \eqref{3rd_estim} and \eqref{3rd_estim_2}, let $\tilde{\chi}\in \mathrm{C}^{\infty}_{0}(\mathbb{\mathbb{C}})$ be an almost analytic extension of $\chi$ satisfying: \begin{align*} & \tilde{\chi}|_{\mathbb{R}} = \chi , \quad |\partial_{\bar{z}}\tilde{\chi}(z)| \leq \mathrm{C}_n |\Im(z)|^n,~~n\in \mathbb{N}. \end{align*} Using the Helffer-Sj{\"o}strand functional calculus, we have that \begin{align*} & \left(\chi(H^{\mathrm{ext}}) \check{\Gamma}(j^R)-\check{\Gamma}(j^R)\chi(H) \right)\chi'(H) \\ & \int \partial_{\bar{z}}\tilde{\chi}(z) \left((z-H^{\mathrm{ext}})^{-1} \check{\Gamma}(j^R)-\check{\Gamma}(j^R)(z-H)^{-1} \right)\chi'(H)dz \wedge d\bar{z}. \end{align*} Combining \eqref{1st_estim}, \eqref{2nd_estim} and Lemma \ref{lemV1Numbestimates}, we obtain \eqref{3rd_estim} and \eqref{3rd_estim_2}. \end{proof} In the case where the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ vanish, the statement of Lemma \ref{numberestimates3} is modified as follows. \begin{lem} \label{numberestimates13} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ vanish and consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. Assume that $G \in L^2$. Let $j^R$ be defined as above. Then \begin{equation}\label{1st_estim_massless} (H^{\mathrm{ext}}+\mathrm{i})^{-1} \check{\Gamma}(j^R)-\check{\Gamma}(j^R)(H+\mathrm{i})^{-1} (N_{\mathrm{neut}}+1)^{-1} \in o(R^{0}). \end{equation} If $G \in \mathbb{H}^\mu$ with $\mu > 0$, then \begin{equation}\label{2nd_estim_massless} (H^{\mathrm{ext}}+\mathrm{i})^{-1} \check{\Gamma}(j^R)-\check{\Gamma}(j^R)(H+\mathrm{i})^{-1} (N_{\mathrm{neut}}+1)^{-1} \in \mathcal{O}(R^{-\min(1,\mu)}). \end{equation} In particular, for any $\chi, \chi' \in \mathrm{C}_0^\infty( \mathbb{R} )$ and $\mu \ge 1$, we have that \begin{equation}\label{3rd_estim_massless} \left(\chi(H^{\mathrm{ext}}) \check{\Gamma}(j^R)-\check{\Gamma}(j^R)\chi(H) \right) \chi'(H) (N_{\mathrm{neut}}+1)^{-1} \in \mathcal{O}(R^{-1}). \end{equation} \end{lem} % \begin{proof} It suffices to adapt the proof of Lemma \ref{numberestimates3}, using Lemma \ref{lemV1NumbestimatesV2}. \end{proof} \section{Spectral theory}\label{sec:spectral} In this section, we do the spectral analysis of the Hamiltonian $H$. Section \ref{subsec:spectral_massive} is devoted to the case where the masses of the neutrinos are supposed to be positive while Section \ref{subsec:spectral_massless} concerns the case where the neutrinos are supposed to be massless. In both cases, we begin with recalling results giving the existence of a ground state, next we study the structure of the essential spectrum by means of suitable versions of Mourre's conjugate operator method. \subsection{Massive neutrinos}\label{subsec:spectral_massive} In this section, we do the spectral analysis of $H$ under the assumption that \begin{equation*} m_\nu = \min ( m_{\nu_e} , m_{\nu_\mu} , m_{\nu_\tau} ) > 0. \end{equation*} \subsubsection{Existence of a ground state and location of the essential spectrum} Recall the notation $E= \inf \sigma( H )$. The next theorem shows, in particular, that $H$ has a ground state, i.e., that $E$ is an eigenvalue of $H$. \begin{Th} \label{Thess2} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ are positive and consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. Assume that $G \in L^2$. Then \begin{equation*} \sigma_{\mathrm{ess}}(H)=[E + m_{\nu}, \infty) . \end{equation*} In particular, $E$ is a discrete eigenvalue of $H$. \end{Th} \begin{proof} The fact that \begin{equation}\label{first_incl} \sigma_{\mathrm{ess}} ( H ) \subset [ E + m_\nu , \infty ) \end{equation} is a consequence of \eqref{3rd_estim} in Lemma \ref{numberestimates3}. Indeed, if $\chi$ belongs to $\mathrm{C}_0^\infty( ( - \infty , E + m_\nu ))$, using that the operator $\check{\Gamma}(j^R)$ defined in the previous section is isometric (see Appendix \ref{sec:standard_def}), we have that \begin{align*} \chi ( H ) = \check{\Gamma}(j^R)^* \check{\Gamma}(j^R) \chi ( H ) = \check{\Gamma}(j^R)^* \chi( H^{ \mathrm{ext} } ) \check{\Gamma}(j^R) + o(R^{0}) , \quad R \to \infty. \end{align*} Since $N_\infty = \mathds{1} \otimes N$ commutes with $H^{ \mathrm{ext} }$ in $\mathscr{H} = \mathscr{H}^{\mathrm{ext}} \otimes \mathscr{H}^{\mathrm{ext}}$, and since \begin{equation*} H^{ \mathrm{ext} } \mathds{1}_{ [ 1 , \infty ) }( N_\infty ) \ge ( E + m_\nu ) \mathds{1}_{ [ 1 , \infty ) }( N_\infty ), \end{equation*} this yields \begin{align*} \chi ( H ) & = \check{\Gamma}(j^R)^* ( \mathds{1} \otimes \Pi_\Omega ) \chi ( H^{ \mathrm{ext} } ) \check{\Gamma}(j^R) + o(R^0) = \check{\Gamma}(j^R)^* ( \mathds{1} \otimes \Pi_\Omega ) \check{\Gamma}(j^R) \chi( H ) + o(R^0) , \quad R \to \infty , \end{align*} where $\Pi_\Omega$ denotes the projection onto the vacuum in $\mathscr{H}$. The second equality in the previous equation is another consequence of \eqref{3rd_estim}. The inclusion \eqref{first_incl} then follows from the observation that the operator $\check{\Gamma}(j^R)^* ( \mathds{1} \otimes \Pi_\Omega ) \check{\Gamma}(j^R) \chi( H )$ is compact (see e.g. \cite{DeGe99_01} for a similar argument). The converse inclusion can be proven by constructing a Weyl sequence associated to $\lambda$ for any $\lambda \in [ E + m_\nu , \infty )$ in the same way as in \cite[Theorem 4.1]{DeGe99_01} or \cite[Theorem 4.3]{Am04_01}. \end{proof} \subsubsection{Spectral analysis for any value of the coupling constant}\label{subsubsec:Mourre1} In this section, we study the structure of the essential spectrum of $H$ using Mourre's conjugate operator theory \cite{Mo80_01,AmBoGe96_01}. We begin by recalling the main facts on which we will rely. We refer the reader to \cite{AmBoGe96_01} for more details. Given two self-adjoint operators $A$, $B$ on a Hilbert space $\mathscr{H}$, $B$ is said to be of class $\mathrm{C}^n(A)$ if and only if the map \begin{equation*} s \mapsto e^{- \mathrm{i} s A} ( B - z )^{-1} e^{\mathrm{i} s A} \Phi, \end{equation*} is of class $\mathrm{C}^n(\mathbb{R})$ for all $\Phi \in \mathscr{H}$. One says that $B$ satisfies a Mourre estimate with respect to $A$ (with compact remainder) on an interval $[a,b]\subset\mathbb{R}$ if there exist a positive constant $\mathrm{c}_0$ and a compact operator $K$ such that \begin{equation} \mathds{1}_{[a,b]}(B)[B,\mathrm{i}A]\mathds{1}_{[a,b]}(B) \ge \mathrm{c}_0\mathds{1}_{[a,b]}(B) + K. \label{recall_Mourre} \end{equation} If $B \in \mathrm{C}^1( A )$ and $B$ satisfies a Mourre estimate w.r.t. $A$ on $[a,b]$, then $B$ has only finitely many eigenvalues, with finite multiplicities, in $[a,b]$. If the Mourre estimate is strict (i.e. if $K=0$ in \eqref{recall_Mourre}), then $B$ does not have eigenvalues in $[a,b]$. Moreover, one can prove that if $B$ satisfies a Mourre estimate with compact remainder on $[a,b]$, then, on any interval $[c,d]\subset[a,b]$ such that $[c,d]\cap\sigma_{\mathrm{pp}}(B)=\emptyset$, $B$ satisfies a strict Mourre estimate. If $B \in \mathrm{C}^2(A)$ and satisfies the Mourre estimate \eqref{recall_Mourre}, then $B$ has no singular continuous spectrum in $[a,b]$, i.e. $\sigma_{\mathrm{sc}}(B)\cap[a,b]=\emptyset$. More precisely, one can prove that $B$ satisfies a limiting absorption principle in any interval $[c,d]\subset[a,b]$ such that $[c,d]\cap\sigma_{\mathrm{pp}}(B)=\emptyset$. As for this last result, we mention that the condition $B\in\mathrm{C}^2(A)$ can be weakened to ``$B\in\mathrm{C}^{1,1}(A)$'', for instance. To simplify the presentation, we will not use such a weaker notion of regularity in this paper. Recall that the operators $a_{(i),l}$, $i=1,2$, $l=1,2,3$, and $a_{(3)}$ have been defined in \eqref{eq:defa12}--\eqref{eq:defa3}. In particular, the operators $a_{(i),l}$ are self-adjoint and their domains are given by $\mathscr{D}(a_{(i),l})=\{ h \in \mathfrak{h}_{1} , a_{(i),l}h \in \mathfrak{h}_{1} \}$, where $a_{(i),l}h$ should be understood in the sense of distributions. Likewise, $a_{(3)}$ is self-adjoint with domain $\mathscr{D}(a_{(3)})=\{ h \in \mathfrak{h}_{2} , a_{(3)} h \in \mathfrak{h}_{2} \}$. Using the notation \begin{equation*} \mathrm{d}\Gamma( q ) = \sum_{i=1}^{14} \mathrm{d}\Gamma( q_i ) \end{equation*} as operators on $\mathscr{H}$, with $q = ( q_1 , \dots , q_{14} )$, $q_1, q_2$ operators on $\mathfrak{h}_2$ and $(q_3, \dots , q_{14})$ operators on $\mathfrak{h}_1$ (see Appendix \ref{sec:standard_def} for details regarding this notation and recall that the total Hilbert space $\mathscr{H}$ is the tensor product of $2$ symmetric Fock spaces for bosons and $12$ anti-symmetric Fock spaces for fermions), we set \begin{align} & A:=\mathrm{d}\Gamma(a) , \notag \\ & a:= (a_{(3)}, a_{(3)}, a_{(1),1}, a_{(1),1} , a_{(2),1} , a_{(2),1} , a_{(1),2} , a_{(1),2} , a_{(2),2} , a_{(2),2} , a_{(1),3} , a_{(1),3} , a_{(2),3} , a_{(2),3} ). \label{eq:conjugate_1} \end{align} Hence the following notations will be used: $a_1 = a_{(3)}$, $a_2=a_{(3)}$, $a_3=a_{(1),1}$, \dots, $a_{14} = a_{(2),3}$. Now, a direct computation gives \begin{equation} [H_0,\mathrm{i} A] = \mathrm{d}\Gamma(|\nabla\omega|^2), \label{commut_H0} \end{equation} in the sense of quadratic forms, with (see \eqref{disprela}) $$\nabla\omega := (\nabla\omega^{(3)} , \nabla\omega^{(3)} , \nabla\omega^{(1)}_1 , \nabla\omega^{(1)}_1 , \nabla\omega^{(2)}_1 , \nabla\omega^{(2)}_1 , \nabla\omega^{(1)}_2 , \nabla\omega^{(1)}_2 , \nabla\omega^{(2)}_2 , \nabla\omega^{(2)}_2 , \nabla\omega^{(1)}_3 , \nabla\omega^{(1)}_3 , \nabla\omega^{(2)}_3 , \nabla\omega^{(2)}_3 ).$$ Moreover, writing $H_I=H_I(G)$, we have that \begin{align} [H_I(G),\mathrm{i}A] = [H^{(1)}_I(G)+H^{(2)}_I(G)+H^{(3)}_I(G)+H^{(4)}_I(G),\mathrm{i}A], \label{commut_HI} \end{align} where \begin{align} & [H_I^{(1)}(G) , \mathrm{i}A ] = \sum_{i=1}^{2} \mathrm{i} H_I^{(1)}( \mathrm{i} a_i G ) - \sum_{i=3}^{14} \mathrm{i} H_I^{(1)}( \mathrm{i} a_i G ) , \label{commut_11} \\ & [H_I^{(2)}(G) , \mathrm{i}A ] = -\sum_{i=1}^{14} \mathrm{i} H_I^{(2)}( \mathrm{i} a_i G ), \label{commut_21} \\ & [H_I^{(3)}(G) , \mathrm{i}A ] = -\sum_{i\in\{1,2,5,6,9,10,13,14\}}^{14} \mathrm{i} H_I^{(3)}( \mathrm{i} a_i G ) + \sum_{i\in\{3,4,7,8,11,12\}}^{14} \mathrm{i} H_I^{(3)}( \mathrm{i} a_i G ), \label{commut_31} \\ & [H_I^{(4)}(G) , \mathrm{i}A ] =- \sum_{i\in\{5,6,9,10,13,14\}}^{14} \mathrm{i} H_I^{(4)}( \mathrm{i} a_i G ) + \sum_{i\in\{1,2,3,4,7,8,11,12\}}^{14} \mathrm{i} H_I^{(4)}( \mathrm{i} a_i G ). \label{commut_41} \end{align} The main result of this section is Theorem \ref{MourreI} below, which proves that $H$ satisfies a Mourre estimate with respect to $A$ in any interval that does not intersect the set of thresholds (see \eqref{def_threshold} below for the definition of thresholds in our context). As recalled at the beginning of this section, in order to be able to deduce useful spectral properties of $H$, in addition to the Mourre estimate, one needs to establish that $H$ is regular enough w.r.t. $A$. This is the purpose of the following lemma. Recall that the notation ``$a_{(i),\cdot} G \in L^2$'' has been introduce above the statement of Theorem \ref{thm:main}. \begin{lem} \label{C1} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ are positive and consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. Assume that \begin{equation} G \in L^2, \quad a_{(i),\cdot} G \in L^2, \quad i = 1,2,3. \label{eq:hypclassC1} \end{equation} Then $H$ is of class $\mathrm{C}^{1}(A)$. If in addition \begin{equation} a_{(i),\cdot} a_{(i'),\cdot} G \in L^2, \quad i,i' = 1,2,3, \label{eq:hypclassC2} \end{equation} then $H$ is of class $\mathrm{C}^2( A )$. \end{lem} \begin{proof} Proceeding as in \cite[Section 4]{FrGrSi08_01}, using that $\mathscr{D}( H ) = \mathscr{D}( H_0 )$, it is not difficult to verify that, for all $s \in \mathbb{R}$, $e^{-\mathrm{i}sA}\mathscr{D}(H)\subset\mathscr{D}(H)$. In order to prove that $H$ is of class $\mathrm{C}^1( A )$ if \eqref{eq:hypclassC1} holds, it then suffices to verify that $[ H , \mathrm{i} A ]$ extends to an $H$-bounded operator. From \eqref{commut_H0} and the fact that $|\nabla \omega|^2$ is bounded, it follows that $[H_0,\mathrm{i}A]$ is relatively $N$-bounded, and therefore relatively $H_0$-bounded since the masses of all the particles are positive. Moreover, it follows from Lemma \ref{lemSCW_1} and \eqref{commut_HI}--\eqref{commut_41} that $[H_I(G),\mathrm{i}A]$ is also $H_0$-bounded under assumption \eqref{eq:hypclassC1}. Likewise, to prove that $H$ is of class $\mathrm{C}^2( A )$ if assumption \eqref{eq:hypclassC2} holds, it suffices to verify that the second commutator $[ [ H , \mathrm{i} A ] , \mathrm{i} A ]$ extends to a relatively $H$-bounded operator. The result then follows from computing $[ [ H , \mathrm{i} A ] , \mathrm{i} A ]$ in the same way as in \eqref{commut_H0}--\eqref{commut_41} and using the same arguments as before. Details are left to the reader. \end{proof} The set of thresholds is defined by \begin{equation}\label{def_threshold} \tau = \sigma_{\mathrm{pp}}(H) + \Big \{ \sum^{14}_{i=1} m_i n_i, \, n_i \in \mathbb{N}, \, (n_1,\dots,n_{14}) \neq (0,\dots,0) \Big \}, \end{equation} where $m_i$ denotes the mass of the particle $i$ (i.e. $m_1=m_2=m_W$, $m_3=m_4=m_e$, $m_5=m_6=m_{\nu_e}$, \emph{etc}). \begin{Th}\label{MourreI} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ are positive and consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. Assume that \eqref{eq:hypclassC1} holds. Let $\lambda \in \mathbb{R}\setminus \tau$. There exists $\varepsilon>0$, $\mathrm{c}_0>0$ and a compact operator $K$ such that \begin{equation} \mathds{1}_{[\lambda-\varepsilon,\lambda+\varepsilon]}(H) [H,\mathrm{i}A] \mathds{1}_{[\lambda-\varepsilon,\lambda+\varepsilon]}(H) \geq \mathrm{c}_0 \mathds{1}_{[\lambda-\varepsilon,\lambda+\varepsilon]}(H)+K. \label{eq:Mourre_estimate_1} \end{equation} In particular, for all interval $[\lambda_1,\lambda_2]$ such that $[\lambda_1,\lambda_2]\cap \tau = \emptyset$, $H$ has at most finitely many eigenvalues with finite multiplicities in $[\lambda_1,\lambda_2]$ and, as a consequence, $\sigma_{\mathrm{pp}}(H)$ can accumulate only at $\tau$, which is a countable set. If in addition \eqref{eq:hypclassC2} holds, then $\sigma_{\mathrm{sc}}(H)=\emptyset$. \end{Th} The proof of \eqref{eq:Mourre_estimate_1} uses arguments developed in \cite{DeGe99_01} and \cite{Am04_01} (see also \cite{DeGe97_01}). In Theorem \ref{MourreII_1} below, we will give a complete proof of an analogous Mourre estimate in the more difficult case where neutrinos are supposed to be massless. It is not difficult to adapt the proof of Theorem \ref{MourreII_1} to obtain \eqref{eq:Mourre_estimate_1}, under the hypotheses of Theorem \ref{MourreI}. We do not give the details. The fact that \eqref{eq:Mourre_estimate_1} implies that $H$ has at most finitely many eigenvalues with finite multiplicities in any compact interval disjoint from $\tau$ is a consequence of Lemma \ref{C1} together with the abstract results of Mourre's theory recalled at the beginning of this section. The same holds for the absence of singular continuous spectrum assuming \eqref{eq:hypclassC2}. \subsubsection{Spectral analysis for small coupling constant and regularized kernels}\label{subsubsec:massive_smallg} In this section, we improve the results of Theorem \ref{MourreI} by imposing stronger conditions on the kernels and treating the coupling constant $g$ as a small parameter. The main idea consists in considering a different, non-self-adjoint conjugate operator, in order to obtain a global Mourre estimate without compact remainder. Extensions of the original Mourre's theory \cite{Mo80_01} to settings with non-self-adjoint conjugate operators have been considered by different authors (see, in particular, \cite{HuSp95_01,AmBoGe96_01,GeGeMo04_01,FaMoSk11_01,BaFaGu16_01}). Yet a further extension, sometimes called singular Mourre's theory, concerns the case where the commutator of the Hamiltonian with the conjugate operator is not relatively bounded w.r.t. the Hamiltonian itself (see \cite{GeGeMo04_01,GeGeMo04_02,FaMoSk11_01}). Singular Mourre's theory will not be used in this section, but it will turn out to be an important tool in order to treat the situation where neutrinos are supposed to be massless, see Section \ref{subsec:spectral_massless}. Before stating the results, we briefly recall the definitions and main elements that will be used in the context of Mourre's theory with a non-self-adjoint conjugate operator. For more details, we refer the reader to \cite[Appendix B]{BaFaGu16_01} for a short presentation, and to \cite{GeGeMo04_01} for a detailed and complete theory. Let $H$ be a self-adjoint operator on a Hilbert space $\mathscr{H}$ and let $A$ be a closed and maximal symmetric operator on $\mathscr{H}$. Assuming that $\mathrm{dim}( \mathrm{Ker}( A^* - \mathrm{i} ) )= 0$, it is known that $A$ generates a semigroup of isometries $\{ W_t \}_{ t \ge 0 }$. Let $\mathscr{G} := \mathscr{D}( | H |^{\frac{1}{2}} )$ be equipped with the norm defined by $\| \phi \|^2_{ \mathscr{G} } := \| |H|^{\frac{1}{2}} \phi \|^2 + \| \phi \|^2$, for all $\phi \in \mathscr{G}$, and let $\| \phi \|^2_{ \mathscr{G}^* } := \| ( |H| + 1 )^{-\frac{1}{2}} \phi \big \|^2$. Observe that the dual space $\mathscr{G}^*$ of $\mathscr{G}$ identifies with the completion of $\mathscr{H}$ with respect to the norm $\| \cdot \|_{\mathscr{G}^*}$. In particular, $H$ identifies with an element of $\mathscr{L} ( \mathscr{G} ; \mathscr{G}^* )$. Suppose that, for all $t > 0$, $W_t$ and $W_t^*$ preserve $\mathscr{G}$ and that, for all $\phi \in \mathscr{G}$, \begin{equation*} \sup_{0<t<1} \| W_t \phi \|_\mathscr{G} < \infty, \quad \sup_{0<t<1} \| W_t^* \phi \|_\mathscr{G} < \infty. \end{equation*} This condition implies, in particular, that the restriction of $W_t$ to $\mathscr{G}$ is a $\mathrm{C}_0$-semigroup and that $W_t$ extends to a $\mathrm{C}_0$-semigroup on $\mathscr{G}^*$ which is denoted by the same symbol. One says that $H$ is of class $\mathrm{C}^1( A ; \mathcal{G} ; \mathcal{G}^* )$ if there exists a positive constant $\mathrm{c}$ such that, for all $0 \le t \le 1$, \begin{equation} \| W_{t} H - H W_{t} \|_{ \mathscr{L} ( \mathscr{G} ; \mathscr{G}^* ) } \le \mathrm{c} t . \label{def_C1GG*} \end{equation} In this case there is an operator $H' \in \mathscr{L} ( \mathscr{G} ; \mathscr{G}^* )$ such that, for all $\phi \in \mathscr{D}( H )$, \begin{equation}\label{defH'} \lim_{ t \to 0^+ } \frac{1}{t} \big ( \langle \phi , W_t H \phi \rangle - \langle H \varphi , W_t \phi \rangle \big ) = \langle \phi , H' \phi \rangle , \end{equation} and one can verify that, in a suitable sense, $H'$ identifies with the quadratic form $[ H , \mathrm{i} A ]$. One says that $H \in \mathrm{C}^2( A ; \mathscr{G} ; \mathscr{G}^* )$ if $H$ belongs to $\mathrm{C}^1( A ; \mathscr{G} ; \mathscr{G}^* )$ and $H'$ belongs to $\mathrm{C}^1( A ; \mathscr{G} ; \mathscr{G}^* )$. If $H$ is of class $\mathrm{C}^1( A ; \mathscr{G} ; \mathscr{G}^* )$, one says that $H$ satisfies a (strict) Mourre estimate on an open interval $I$ if there exist constants $\mathrm{c}_0>0$ and $\mathrm{c} \in \mathbb{R}$, such that, in the sense of quadratic forms on $\mathscr{D}( H )$, \begin{equation*} H' \ge \mathrm{c}_0 \mathds{1} - \mathrm{c} ( \mathds{1} - \mathds{1}_I(H) ) ( \mathds{1} + H^2 )^{\frac12}. \end{equation*} As in the setting of Section \ref{subsubsec:Mourre1}, if $H$ is of class $\mathrm{C}^1( A ; \mathscr{G} ; \mathscr{G}^* )$ and satisfies a strict Mourre estimate on an interval $I$, then $H$ does not have eigenvalues in $I$. If $H$ is of class $\mathrm{C}^2( A ; \mathscr{G} ; \mathscr{G}^* )$, then a limiting absorption principle holds in any interval where $H$ satisfies a strict Mourre estimate. In particular $H$ does not have singular continuous spectrum in such an interval. Recall that the operators $b_{(i),l}$, $i=1,2$, $l=1,2,3$, acting on $\mathfrak{h}_1$, and $b_{(3)}$ acting on $\mathfrak{h}_2$ have been defined in \eqref{eq:defb12}--\eqref{eq:defb3}. In particular, the operators $b_{(i),l}$ with domains $\mathrm{C}_0^{\infty}(\mathbb{R}^3 \setminus \{ 0 \}) \times \{-\frac12,\frac12\}$ are symmetric, and likewise $b_{(3)}$ with domain $\mathrm{C}_0^{\infty}(\mathbb{R}^3 \setminus \{ 0 \}) \times \{-1,0,1 \}$ is symmetric. Their closures are denoted by the same symbols. Using the notations of Appendix \ref{sec:standard_def} as before, we set \begin{align}\label{conjugate_2} &B:=\mathrm{d}\Gamma(b) , \notag \\ & b:= (b_{(3)}, b_{(3)}, b_{(1),1}, b_{(1),1} , b_{(2),1} , b_{(2),1} , b_{(1),2} , b_{(1),2} , b_{(2),2} , b_{(2),2} , b_{(1),3} , b_{(1),3} , b_{(2),3} , b_{(2),3} ) , \end{align} (hence $b_1 = b_{(3)}$, $b_2=b_{(3)}$, $b_3=b_{(1),1}$, \dots, $b_{14} = b_{(2),3}$.). Then $\mathrm{dim}( \mathrm{Ker}( B^* - \mathrm{i} ) )= 0$, $B$ generates a $\mathrm{C}_0$-semigroup of isometries $\{W_t\}_{t\ge0}$ and a direct computation gives \begin{equation} [H_0,\mathrm{i} B] = N, \label{commut_H0_2} \end{equation} in the sense of quadratic forms. Moreover, the commutators $[H_I^{(j)}(G) , \mathrm{i}B]$, $j=1,\dots,4$ are given by \eqref{commut_11}--\eqref{commut_41} with $b_i$ instead of $a_i$. A straightforward modification of \cite[Section 5.1]{BaFaGu16_01} shows that $W_t$ and $W^*_t$ preserve $\mathscr{D}(|H|^{\frac{1}{2}})$, and that for all $\phi \in \mathscr{D}(|H|^{\frac{1}{2}})$, \begin{align} \sup_{0<t<1} \|W_t \phi\|_{ \mathscr{D}(|H|^{\frac{1}{2}})} < \infty, \quad \sup_{0<t<1} \|W^*_t \phi\|_{\mathscr{D}(|H|^{\frac{1}{2}})} < \infty. \label{eq:bpreserve} \end{align} Before proving a Mourre estimate and deducing from it spectral properties of $H$, we must show, as in the previous section, that $H$ is regular enough with respect to the conjugate operator. \begin{lem} \label{lem:C2} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ are positive and consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. Assume that \begin{equation} G \in L^2, \quad b_{(i),\cdot} G \in L^2, \quad i = 1,2,3. \label{eq:hypclassC1_3} \end{equation} Then $H$ is of class $\mathrm{C}^{1}(B ; \mathscr{D}(|H|^{\frac{1}{2}}) ; \mathscr{D}(|H|^{\frac{1}{2}})^*)$. If in addition \begin{equation} b_{(i),\cdot} b_{(i'),\cdot} G \in L^2, \quad i,i' = 1,2,3, \label{eq:hypclassC2_4} \end{equation} then $H$ is of class $\mathrm{C}^2( B ; \mathscr{D}(|H|^{\frac{1}{2}}) ; \mathscr{D}(|H|^{\frac{1}{2}})^* )$. \end{lem} \begin{proof} Suppose that \eqref{eq:hypclassC1_3} holds. In order to verify that $H$ is of class $\mathrm{C}^{1}(B ; \mathscr{D}(|H|^{\frac{1}{2}}) ; \mathscr{D}(|H|^{\frac{1}{2}})^*)$, since \eqref{eq:bpreserve} holds, it suffices to prove (see \cite[Proposition 5.2]{GeGeMo04_01}) that the quadratic form $[H,B]$ defined on $\mathscr{D}( |H|^{\frac12} )\cap\mathscr{D}(B)$ by $\langle \phi, [H,B] \phi \rangle = \langle \phi , H B \phi \rangle - \langle B^* \phi , H \phi \rangle$ extends to an element of $\mathscr{L} ( \mathscr{D}(|H|^{\frac{1}{2}}) ; \mathscr{D}(|H|^{\frac{1}{2}})^* )$. By \eqref{commut_H0_2} and the fact that $N$ is relatively $H$-bounded, it is clear that $[H_0,B]$ extends to an element of $\mathscr{L} ( \mathscr{D}(|H|^{\frac{1}{2}}) ; \mathscr{D}(|H|^{\frac{1}{2}})^* )$. That $[H_I,B]$ also extends to an element of $\mathscr{L} ( \mathscr{D}(|H|^{\frac{1}{2}}) ; \mathscr{D}(|H|^{\frac{1}{2}})^* )$ follows from the expression of the commutator (given by \eqref{commut_11}--\eqref{commut_41} with $b_{i}$ instead of $a_{i}$, as mentioned above) together with Lemma \ref{lemSCW_1}, which can be applied under the hypothesis \eqref{eq:hypclassC1_3}. To prove that $H$ is of class $\mathrm{C}^2( B ; \mathscr{D}(|H|^{\frac{1}{2}}) ; \mathscr{D}(|H|^{\frac{1}{2}})^* )$ under the further assumption \eqref{eq:hypclassC2_4}, it suffices to verify similarly that $[[H,B],B]$ extends to an element of $\mathscr{L} ( \mathscr{D}(|H|^{\frac{1}{2}}) ; \mathscr{D}(|H|^{\frac{1}{2}})^* )$. But we have that $[[H_0,B],B]=0$ and a similar computation as before shows that $[[H_I,B],B]$ extends to an element of $\mathscr{L} ( \mathscr{D}(|H|^{\frac{1}{2}}) ; \mathscr{D}(|H|^{\frac{1}{2}})^* )$ by Lemma \ref{lemSCW_1}. \end{proof} The next theorem establishes a global Mourre estimate for $H$, from which, using the regularity properties proven in the previous lemma, we can deduce the desired spectral properties of $H$. \begin{Th} \label{MourreII} Consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term} and assume that \eqref{eq:hypclassC1_3} holds. There exist $g_0>0$, $\mathrm{c}>0$ and $\mathrm{d}>0$ such that, for all values of the masses of the neutrinos $m_{\nu_e}>0$, $m_{\nu_{\mu}}>0$, $m_{\nu_{\tau}}>0$, \begin{equation} \label{eq:Mourreunif_B2} [H,\mathrm{i}B]\ge \mathrm{c}\mathds{1}-\mathrm{d} \Pi_\Omega , \end{equation} where $\Pi_\Omega$ denotes the projection onto the vacuum in $\mathscr{H}$. In particular, $E = \inf \sigma ( H )$ is the only eigenvalue of $H$, and $E$ is non-degenerate. If, in addition, \eqref{eq:hypclassC2_4} holds, then the spectrum of $H$ in $[E+m_\nu,\infty)$ is purely absolutely continuous. \end{Th} \begin{proof} Recall that, in the sense of quadratic forms on $\mathscr{D}(B) \cap \mathscr{D}(H_0)$ we have that $[H_0,\mathrm{i}B]=N$ and that $[H_I,\mathrm{i}B]$ is relatively $N$-bounded by Lemma \ref{lemSCW_1}. Therefore there exists $\mathrm{c}_1>0$ and $\mathrm{c}_2>0$, which do not depend on the masses of the neutrinos, such that \begin{equation*} \langle \psi , [H_I, \mathrm{i} B] \psi \rangle \leq \mathrm{c}_1 \langle \psi , N \psi \rangle + \mathrm{c}_2 \|\psi\|^2. \end{equation*} This yields \begin{align*} [H, \mathrm{i}B] = [H_0,\mathrm{i}B] + g [H_I, \mathrm{i}B] & \geq N - \mathrm{c}_1 g N - \mathrm{c}_2 g \\ & \geq (1 - \mathrm{c}_1 g)( N + \Pi_{\Omega} - \Pi_{\Omega} ) - \mathrm{c}_2 g \\ & \geq (1 - \mathrm{c}_1 g) \mathds{1} - (1 - \mathrm{c}_1 g)\Pi_{\Omega} - \mathrm{c}_2 g \\ & \geq (1 - (\mathrm{c}_1+\mathrm{c}_2) g ) \mathds{1} - (1 - \mathrm{c}_1 g)\Pi_{\Omega} , \end{align*} which proves \eqref{eq:Mourreunif_B2}. The fact that \eqref{eq:Mourreunif_B2} implies that $H$ has at most one eigenvalue is a consequence of the virial theorem (see Lemma 10 of \cite{HuSp95_02}), which holds since $H \in \mathrm{C}^{1}(B ; \mathscr{D}(|H|^{\frac{1}{2}}) ; \mathscr{D}(|H|^{\frac{1}{2}})^*)$ by Lemma \ref{lem:C2}, together with the fact that $\mathrm{dim}( \mathrm{Ran} ( \Pi_\Omega ) ) = 1$. See, e.g., \cite[Lemma 10]{HuSp95_01}. That the spectrum of $H$ in $[E+m_\nu,\infty)$ is purely absolutely continuous if \eqref{eq:hypclassC2_4} is satisfied is a consequence of the abstract results recalled at the beginning of this section together with the fact that $H \in \mathrm{C}^{2}(B ; \mathscr{D}(|H|^{\frac{1}{2}}) ; \mathscr{D}(|H|^{\frac{1}{2}})^*)$, by Lemma \ref{lem:C2}. \end{proof} \subsection{Massless neutrinos}\label{subsec:spectral_massless} In this section, we suppose that \begin{equation*} m_{\nu_e} = m_{\nu_\mu} = m_{\nu_\tau} = 0. \end{equation*} As in the previous section, we first recall suitable assumptions implying existence of a ground state for $H$, next we study the structure of the essential spectrum using a suitable version of Mourre's theory. \subsubsection{Existence of a ground state and location of the essential spectrum} We recall here, without proof, a result due to \cite{BaGu09_01,AsBaFaGu11_01}. \begin{Th} \label{Thess} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ vanish and consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. Assume that $G \in L^2$ and that $|p_2|^{-1} G \in L^2$. Then, there exists $g_0 > 0$ such that, for all $|g| \le g_0$, $H$ has a unique ground state, i.e., $E=\inf \sigma( H )$ is a non-degenerate eigenvalue of $H$. Moreover, \begin{equation*} \sigma(H)=\sigma_{\mathrm{ac}}(H)=[E, \infty). \end{equation*} \end{Th} Let us mention that the conclusion of Theorem \ref{Thess} does not exclude the presence of eigenvalues or singular continuous spectrum in the interval $[E,\infty)$. Proving absence of embedded eigenvalues and of singular continuous spectrum will be one of the main purposes of the next section. Theorem \ref{Thess} is proven in \cite{BaGu09_01,AsBaFaGu11_01} for $H = H_0 + g ( H_I^{(1)} + H_I^{(2)} )$, but the proof goes through without any substantial modification for the full Hamiltonian $H = H_0 + g ( H_I^{(1)} + H_I^{(2)} + H_I^{(3)} + H_I^{(4)} )$. \subsubsection{Spectral analysis} Now, we turn to the study of the essential spectrum $[E,\infty)$ of $H$. Let us mention that, in \cite{AsBaFaGu11_01}, it is proven that the spectrum of $H$ in $[E,m_e)$ is purely absolutely continuous (except for the ground state energy, which is an eigenvalue as recalled in Theorem \ref{Thess}). This result is proven using Mourre's theory, with a conjugate operator given as the generator of dilatations ``restricted to low-energies''. The idea of employing such a conjugate operator originated in \cite{FrGrSi08_01}, see also \cite{ChFaFrSi12_01}. The method of \cite{FrGrSi08_01} is particularly efficient in that it only requires that the kernels of the interaction Hamiltonian belong to the domain of the generator of dilatations, which does not require much regularity of the kernels in the low-energy regime. However, it is presently not known how to extend this approach to prove the absence of singular continuous spectrum in the whole interval $[E,\infty)$ and not only in $[E,m_e)$. On the other hand, if one assumes that the kernels belong to the domain of the operator $b_{i}$ in \eqref{eq:defb12}--\eqref{eq:defb3}, one can proceed as in Section \ref{subsubsec:massive_smallg}. The conjugate operator is still given by \eqref{conjugate_2}. Note that $b_{(2),l} = a_{(2),l}$ when the masses of the neutrinos vanish. As in Section \ref{subsubsec:massive_smallg}, one verifies that $B$ is the generator of a $\mathrm{C}_0$-semigroup of isometries and that $[H_0,\mathrm{i}B]=N$. In particular, $[H_0,\mathrm{i}B]$ is not relatively $H$-bounded anymore and therefore one cannot apply the abstract setting recalled at the beginning of Section \ref{subsubsec:massive_smallg}. Nevertheless, one can use singular Mourre's theory developed in \cite{GeGeMo04_01} (see also \cite{FaMoSk11_01}). Before stating our results, we briefly explain the setting of this theory, focusing on the differences with that of Section \ref{subsubsec:massive_smallg}. Let $\mathscr{H}$ be a complex Hilbert space and consider two self-adjoint operators $H$ and $M$ on $\mathscr{H}$, with $M \ge 0$, and a symmetric operator $R$ relatively $H$-bounded. We suppose that $H \in \mathrm{C}^1( M )$ and that $[ H , \mathrm{i} M ]$ is relatively $H$-bounded. Let $H' := M+R$ on $\mathscr{D}(M) \cap \mathscr{D}(H)$. We set $\mathscr{G} = \mathscr{D}( M^{\frac{1}{2}} )\cap \mathscr{D}( |H|^{\frac{1}{2}} )$, equipped with the norm defined by $\| \phi \|^2_{ \mathscr{G} } := \| M^{\frac{1}{2}} \phi \|^2_{ \mathscr{H} } + \| |H|^{\frac{1}{2}} \phi \|^2_{ \mathscr{H} } + \| \phi \|^2_{ \mathscr{H} }$, for all $\phi \in \mathscr{G}$. As in Section \ref{subsubsec:massive_smallg}, letting $\| \phi \|_{ \mathscr{G}^* } := \| ( M + | H | + 1 )^{-\frac{1}{2}} \phi \|_{ \mathscr{H} }$, one verifies that $\mathscr{G}^*$ identifies with the completion of $\mathscr{H}$ with respect to $\| \cdot \|_{ \mathscr{G}^* }$. Hence $H$ and $M$ identify with elements of $\mathscr{L}( \mathscr{G} ; \mathscr{G}^* )$. As in Section \ref{subsubsec:massive_smallg}, the conjugate operator $A$ is supposed to be closed, maximal symmetric and such that $\dim \mathrm{Ker} (A^* - \mathrm{i}) = 0$. The $\mathrm{C}_0$-semigroup of isometries generated by $A$ is denoted by $\{ W_t \}_{t \ge 0}$ and we suppose that, for all $\phi \in \mathscr{G}$, \begin{equation*} \sup_{0<t<1} \| W_t \phi \|_\mathscr{G} < \infty, \quad \sup_{0<t<1} \| W_t^* \phi \|_\mathscr{G} < \infty. \end{equation*} As before, $H$ is said to belong to $\mathrm{C}^1( A ; \mathscr{G} ; \mathscr{G}^* )$ if \eqref{def_C1GG*} holds. In this case there is an operator $H' \in \mathscr{L} ( \mathscr{G} ; \mathscr{G}^* )$ satisfying \eqref{defH'} and one requires that $H' = M + R$. Given these conditions, assuming that $H$ belongs to $\mathrm{C}^1( A ; \mathscr{G} ; \mathscr{G}^* )$ (or to $\mathrm{C}^2( A ; \mathscr{G} ; \mathscr{G}^* )$) and that the Mourre estimate holds, the conclusions concerning the spectrum of $H$ are the same as in Section \ref{subsubsec:massive_smallg}. Therefore, choosing $M = [ H_0 , \mathrm{i} B ] = N$ and $R = [ H_I , \mathrm{i} B ]$ and proceeding exactly as in the proofs of Lemma \ref{lem:C2} and Theorem \ref{MourreII}, we obtain the following result. \begin{Th} \label{MourreII_1} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ vanish and consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. Assume that \eqref{eq:hypclassC1_3} holds. There exist $g_0>0$, $\mathrm{c}>0$ and $\mathrm{d}>0$ such that, \begin{equation*} [H,\mathrm{i}B]\ge \mathrm{c}\mathds{1}-\mathrm{d} \Pi_\Omega , \end{equation*} where $\Pi_\Omega$ denotes the projection onto the vacuum in $\mathscr{H}$. In particular, $H$ has at most one eigenvalue, which is non-degenerate. If, in addition, \eqref{eq:hypclassC2_4} holds, then, except for the ground state energy $E$, the spectrum of $H$ in $[E,\infty)$ is purely absolutely continuous. \end{Th} Theorem \ref{MourreII_1} provides a complete description of the spectrum of $H$ for small enough values of $g$ and under strong assumptions on the kernels $G$. However, in view of applications to scattering theory in Section \ref{sec:AC} -- more precisely, in order to prove the propagation estimates in Section \ref{sec:propag} that will be subsequently used in Section \ref{sec:AC} -- the conjugate operator $B$ chosen in Theorem \ref{MourreII_1} is too singular. The singularity here comes from the presence of massive particles. Indeed, for massive particles, the operators $b_i$ are strongly singular near the origin $p_i=0$ because of the factor $( p_i \cdot \nabla \omega^{(i)}_l (p_i ) )^{-1}$ in the definitions \eqref{eq:defb12}--\eqref{eq:defb3}. For this reason, we need to prove a Mourre estimate with another conjugate operator, namely the operator $A$ defined in \eqref{eq:conjugate_1}. Notice that this operator is not self-adjoint when neutrinos are supposed to be massless, because in this case the operators $a_{(2),l}$ are not self-adjoint, only maximal symmetric. The domains of the operators $a_{(2),l}$ are explicitly given as follows: let $T:L^2(\mathbb{R}^3 \times \{-\frac12 , \frac12\}) \to L^2( \mathbb{R}_+ ) \otimes L^2( \mathrm{S}^2 ) \otimes \mathbb{C}^2$ be the unitary operator, going from cartesian coordinate to polar coordinate, defined by $(Tu)(r,\theta,s) := ru(r\theta,s)$. Then \begin{equation}\label{eq:domain_conjugate} \mathscr{D}( a_{(2),l} ) = T^{-1} \big ( \mathbb{H}_0^1( \mathbb{R}_+ ) \otimes L^2( \mathrm{S}^2 ) \otimes \mathbb{C}^2 \big ) , \quad a_{(2),l} = T^{-1} ( \mathrm{i} \partial_r ) T , \end{equation} where $\mathbb{H}_0^1( \mathbb{R}_+ )$ is the usual Sobolev space with Dirichlet boundary condition at $0$. As in Section \ref{subsubsec:massive_smallg}, one verifies that $\mathrm{dim}( \mathrm{Ker}( A^* - \mathrm{i} ) )= 0$, and that $A$ generates a $\mathrm{C}_0$-semigroup of isometries $\{ \tilde{W}_t \}_{ t \ge 0 }$ such that $\tilde{W}_t$ and $\tilde{W}^*_t$ preserve $\mathscr{G} := \mathscr{D}(|H|^{\frac{1}{2}}) \cap \mathscr{D}( N_{\mathrm{neut}}^{\frac12} )$. Moreover, for all $\phi \in \mathscr{G}$, \begin{align*} \sup_{0<t<1} \| \tilde{W}_t \phi\|_{ \mathscr{G} } < \infty, \quad \sup_{0<t<1} \| \tilde{W}^*_t \phi\|_{\mathscr{G} } < \infty . \end{align*} A direct computation gives \begin{equation} [H_0,\mathrm{i} A] = N_{\mathrm{neut}} + \sum_{l=1}^3 \mathrm{d} \Gamma( | \nabla \omega^{(1)}_l( p_1 ) |^2 ) + \mathrm{d} \Gamma( | \nabla \omega^{(3)}( p_3 ) |^2 ) , \label{commut_H0_8} \end{equation} in the sense of quadratic forms, where \begin{align*} & \mathrm{d} \Gamma( | \nabla \omega^{(1)}_l( p_1 ) |^2 ) = \sum_{\epsilon=\pm} \int | \nabla \omega^{(1)}_l( p_1 ) |^2 {b^*_{l,\epsilon}}(\xi_1){b_{l,\epsilon}}(\xi_1) d\xi_1 , \\ & \mathrm{d} \Gamma( | \nabla \omega^{(3)}( p_3 ) |^2 ) = \sum_{\epsilon=\pm} \int | \nabla \omega^{(3)}( p_3 ) |^2 {a^*_{\epsilon}}(\xi_3){a_{\epsilon}}(\xi_3) d\xi_3 . \end{align*} Moreover, the commutators $[H_I^{(j)}(G) , \mathrm{i} A]$, $j=1,\dots,4$, are given by \eqref{commut_11}--\eqref{commut_41}. In particular, the commutator $[H,\mathrm{i}A]$ is not relatively $H$-bounded. For this reason, we work in the setting of singular Mourre's theory. The following lemma can be proven in the same way as Lemma \ref{lem:C2}. \begin{lem}\label{lem:C8} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ vanish and consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. Assume that \eqref{eq:hypclassC1} holds. Then $H$ is of class $\mathrm{C}^{1}( A ; \mathscr{G} ; \mathscr{G}^* )$. If in addition \eqref{eq:hypclassC2} holds, then $H$ is of class $\mathrm{C}^2( A ; \mathscr{G} ; \mathscr{G}^* )$. \end{lem} Recall that the set of thresholds, $\tau$, is defined in \eqref{def_threshold}. We are now ready to prove the main result of this section. \begin{Th}\label{MourreI_3} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ vanish and consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. Assume that \begin{equation} G \in L^2, \quad a_{(i),\cdot} G \in L^2, \quad |p_3|^{-1} a_{(i),\cdot} G \in L^2, \quad i = 1,2,3. \label{eq:hypclassC1_3_1_1_1} \end{equation} There exists $g_0>0$ such that, for all $|g| \le g_0$ and $\lambda \in \mathbb{R}\setminus \tau$, there exist $\varepsilon>0$, $\mathrm{c}_0>0$, $\mathrm{d}>0$ and a compact operator $K$ such that \begin{equation} [H,\mathrm{i} A ] \geq \mathrm{c}_0 ( N_{\mathrm{neut}} + \mathds{1} ) - \mathrm{d} ( \mathds{1} - \mathds{1}_{[\lambda-\varepsilon,\lambda+\varepsilon]}(H) ) ( 1 + H^2 )^{\frac12} + K. \label{eq:Mourre_estimate_1_4} \end{equation} In particular, for all interval $[\lambda_1,\lambda_2]$ such that $[\lambda_1,\lambda_2]\cap \tau = \emptyset$, $H$ has at most finitely many eigenvalues with finite multiplicities in $[\lambda_1,\lambda_2]$ and, as a consequence, $\sigma_{\mathrm{pp}}(H)$ can accumulate only at $\tau$, which is a countable set. If in addition \begin{equation} a_{(i),\cdot} a_{(i'),\cdot} G \in L^2, \quad i,i' = 1,2,3, \label{eq:hypclassC2_2_4} \end{equation} then $\sigma_{\mathrm{sc}}(H)=\emptyset$. \end{Th} \begin{Rk} The following weaker version of the Mourre estimate, \begin{equation*} [H,\mathrm{i} A ] \geq \mathrm{c}_0 \mathds{1} - \mathrm{d} ( \mathds{1} - \mathds{1}_{[\lambda-\varepsilon,\lambda+\varepsilon]}(H) ) ( 1 + H^2 )^{\frac12} + K, \end{equation*} would be sufficient for the conclusions of Theorem \ref{MourreI_3} to hold. But the Mourre estimate with the operator $N_{\mathrm{neut}}$ in right-hand-side of \eqref{eq:Mourre_estimate_1_4} will be important in our proof of some propagation estimates in Section \ref{sec:propag}. \end{Rk} \begin{proof}[Proof of Theorem \ref{MourreI_3}] In this proof, for any interval $\Delta$, $\chi_{\Delta}$ will refer to a function in $\mathrm{C}_{0}^{\infty}(\mathbb{R})$ such that $\Delta \subset \mathrm{supp}\left(\chi_{\Delta} \right)$. For all $\lambda \in \mathbb{R}$, we set \begin{align*} d(\lambda) = \inf \Big \{ & \mu \in \mathbb{R} \, \text{ such that } \, \mu = \sum_{ \substack{ l=1,2,3 \\ \epsilon =\pm }} \sum_{i=1}^{n_{l,\epsilon}} | \nabla \omega^{(1)}_l( p_{1,i,l,\epsilon} ) |^2 + \sum_{\epsilon=\pm} \sum_{i=1}^{n_{\epsilon}} | \nabla \omega^{(3)}( p_{3,i,\epsilon} ) |^2 , \\ & \lambda_1 + \sum_{ \substack{ l=1,2,3 \\ \epsilon =\pm }} \sum_{i=1}^{n_{l,\epsilon}} \omega^{(1)}_l( p_{1,i,l,\epsilon} ) + \sum_{\epsilon=\pm} \sum_{i=1}^{n_{\epsilon}} \omega^{(3)}_l( p_{3,i,\epsilon} ) =\lambda, \, \lambda_1 \in \sigma_{\mathrm{pp}}(H) , \\ & n_{l,\epsilon} , n_\epsilon \in \mathbb{N} \text{ and at least one of the } n_{l,\epsilon} \text{ or } n_\epsilon \text{ is} \neq 0, \, p_{1,i,l,\epsilon} , p_{3,i,\epsilon} \in \mathbb{R}^3 \Big \} , \end{align*} with the convention that $\inf \emptyset = 0$. The definition of $\tilde{d}(\lambda)$ is the same, except that we do not impose the restriction that at least one of the $n_{l,\epsilon}$ or $n_\epsilon$ is $\neq0$. One can then verify that $\tilde{d}(\lambda) = d(\lambda)$, if $\lambda \notin \sigma_{\mathrm{pp}}(H)$, and $\tilde d(\lambda)=0$ if $\lambda \in \sigma_{\mathrm{pp}}(H)$. We also introduce, for $\kappa > 0$, \begin{equation*} \Delta^{\kappa}_{\lambda} = [\lambda-\kappa,\lambda+\kappa], \quad d^{\kappa}(\lambda) = \inf_{\mu \in \Delta^{\kappa}_{\lambda}} d(\mu), \quad \tilde{d}^{\kappa}(\lambda) = \inf_{\mu \in \Delta^{\kappa}_{\lambda}} \tilde{d}(\mu). \end{equation*} Recalling \eqref{commut_H0_8}, we set \begin{equation*} H'_0 := \frac12 N_{\mathrm{neut}} + \sum_{l=1}^3 \mathrm{d} \Gamma( | \nabla \omega^{(1)}_l( p_1 ) |^2 ) + \mathrm{d} \Gamma( | \nabla \omega^{(3)}( p_3 ) |^2 ) , \end{equation*} so that \begin{equation*} [H_0,\mathrm{i}A] = \frac12 N_{\mathrm{neut}} + H'_0. \end{equation*} We also set $H'_I := [ H_I , \mathrm{i} A ]$. We follow the general strategy of the proof of \cite[Theorem 4.3]{DeGe99_01}. Let $m_1 := \inf ( m_e , m_\mu , m_\tau , m_W )>0$. We will prove by induction that the following properties hold for any $n \in \mathbb{N}^*$. \begin{itemize} \item[$H_1(n):$] Let $\varepsilon>0$ and $\lambda \in [E,E+n m_1 )$. There exist a constant $\mathrm{c}$, a compact operator $K_0$ and an interval $\Delta$ containing $\lambda$ such that \begin{equation*} H'_0 + g H'_I \ge \Big (\min \big ( \frac14 , d(\lambda) \big ) - \varepsilon \Big ) \mathds{1} - \mathrm{c} \big ( \mathds{1} - \chi_{\Delta} (H) \big ) ( 1 + H^2 )^{\frac12} + K_0. \end{equation*} \item[$H_2(n):$] Let $\varepsilon > 0 $ and $\lambda \in [E,E+n m_1 )$. There exist a constant $\mathrm{c}$ and an interval $\Delta$ containing $\lambda$ such that \begin{equation*} H'_0 + g H'_I \ge \Big (\min \big ( \frac14 , \tilde{d}(\lambda) \big ) -\varepsilon \Big )\mathds{1} - \mathrm{c} \big ( \mathds{1} - \chi_{\Delta} (H) \big ) ( 1 + H^2 )^{\frac12}. \end{equation*} \item[$H_3(n):$] Let $\kappa > 0$, $\varepsilon_0>0$ and $\varepsilon > 0$. There exist a constant $\mathrm{c}$ and $\delta>0$ such that, for all $\lambda \in [E,E+n m_1 - \varepsilon_0] $, one has \begin{equation*} H'_0 + g H'_I \ge \Big (\min \big ( \frac14 , \tilde{d}^{\kappa}(\lambda) \big ) -\varepsilon \Big ) \mathds{1} - \mathrm{c} \big ( \mathds{1} - \chi_{\Delta^{\delta}_{\lambda}} (H) \big )( 1 + H^2 )^{\frac12}. \end{equation*} \item[$S_1(n):$] $\tau$ is a closed countable set in $[E,E+n m_1 ]$. \item[$S_2(n):$] for all $\lambda_1$, $\lambda_2$ such that $\lambda_1 < \lambda_2 \leq E + n m_1 $ and $[\lambda_1,\lambda_2]\cap \tau = \emptyset$, $H$ has finitely many eigenvalues, with finite multiplicities, in $[\lambda_1,\lambda_2]$. \end{itemize} We claim that, for all $n \in \mathbb{N}^*$, \begin{align} & S_2(n-1) \Rightarrow S_1(n) \label{Statement4} \\ & S_1(n) ~~\text{and} ~~ H_3(n-1) \Rightarrow H_1(n). \label{Statement5} \\ & H_1(n) \Rightarrow H_2(n) \label{Statement1} \\ & H_2(n) \Rightarrow H_3(n) \label{Statement2} \\ & H_1(n) \Rightarrow S_2(n) \label{Statement3} \end{align} By definition, $\tau \cap [E , E+ m_1 ) = \emptyset$ and hence $S_1(1)$ is obviously satisfied. We refer to \cite{Mo80_01,AmBoGe96_01,DeGe97_01} for the proofs of \eqref{Statement4}, \eqref{Statement1}, \eqref{Statement2} and \eqref{Statement3}. Note that \eqref{Statement1} uses the compactness of $K_0$ and finite rank operator estimates, and that \eqref{Statement3} is a consequence of the virial theorem. It remains to prove $H_1(1)$ and \eqref{Statement5}. Using that the commutators $[H_I^{(j)}(G) , \mathrm{i} A]$, $j=1,\dots,4$, are given by \eqref{commut_11}--\eqref{commut_41}, a direct application of the $N_\tau$ estimates of Glimm and Jaffe (see \cite[Proposition 1.2.3(b)]{GlJa77_01}) shows that there exist $\mathrm{c}_1>0$ and $\mathrm{c}_2>0$ such that \begin{equation}\label{estimH'I} \big|\langle \Psi , H'_I \Psi \rangle \big | \leq \mathrm{c}_1 \langle \Psi , H'_0 \Psi \rangle + \mathrm{c}_2 \|\Psi\|^2, \end{equation} for all $\Psi \in \mathscr{D}( H'_0 )$. Here it should be noticed that $| \nabla \omega^{(3)}( p_3 ) |^2 = p_3^2 ( p_3^2 + m_W^2 )^{-1}$. Hence, according to \cite[Proposition 1.2.3(b)]{GlJa77_01}, the constant $\mathrm{c}_1$ can be chosen to be proportional to $\|| p_3 |^{-1} G \|_2 + \| G \|_2$, which is finite by \eqref{eq:hypclassC1_3_1_1_1}. Recall that the operator $\check{\Gamma}(j^R) : \mathscr{H} \to \mathscr{H}^{\mathrm{ext}}$ has been defined in Section \ref{subsec:estimates} (see the paragraph after the statement of Lemma \ref{lemV2Numbestimates} and see also Appendix \ref{sec:standard_def}). Using that $\check{\Gamma}(j^R)$ is an isometry, we can write \begin{align*} H'_0 + g H'_I &= \check{\Gamma}(j^R)^*\check{\Gamma}(j^R) (H'_0 + g H'_I ) \\ &= \check{\Gamma}(j^R)^* (H'_0\otimes\mathds{1}+\mathds{1}\otimes H'_0 + g H'_I \otimes \mathds{1} + \mathrm{Rem}_R ) \check{\Gamma}(j^R) , \end{align*} with $\mathrm{Rem}_R ( N_0 + N_\infty + 1 )^{-1} = o( R^0 )$. The last equality can be proven in the same way as in the proofs of Lemmas \ref{numberestimates3}--\ref{numberestimates13}. We decompose \begin{align} [H, \mathrm{i} A] &= \check{\Gamma}(j^R)^* (H'_0\otimes\mathds{1}+\mathds{1}\otimes H'_0 + g H'_I \otimes \mathds{1} + \mathrm{Rem}_R ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \check{\Gamma}(j^R) \notag \\ &\quad + \check{\Gamma}(j^R)^* (H'_0\otimes\mathds{1}+\mathds{1}\otimes H'_0 + g H'_I \otimes \mathds{1} + \mathrm{Rem}_R ) \mathds{1}_{ [1,\infty) }( N_{\mathrm{neut},\infty} ) \check{\Gamma}(j^R) , \label{eq:decomp_1} \end{align} and estimate the two terms separately. For the second one, we notice that \begin{equation} ( \mathds{1}\otimes H'_0 ) \mathds{1}_{ [1,\infty) }( N_{\mathrm{neut},\infty} ) \ge \frac12 ( \mathds{1}\otimes N_{\mathrm{neut},\infty} ) \mathds{1}_{ [1,\infty) }( N_{\mathrm{neut},\infty} ) \ge \frac12 \mathds{1}_{ [1,\infty) }( N_{\mathrm{neut},\infty} ) . \label{eq:lsif1} \end{equation} By \eqref{estimH'I}, we have that \begin{align*} & \check{\Gamma}(j^R)^* (H'_0\otimes\mathds{1}+\mathds{1}\otimes H'_0 + g H'_I \otimes \mathds{1} + \mathrm{Rem}_R ) \mathds{1}_{ [1,\infty) }( N_{\mathrm{neut},\infty} ) \check{\Gamma}(j^R) \\ &\ge \check{\Gamma}(j^R)^* ((1 - \mathrm{c}_1 g) H'_0 \otimes\mathds{1} - \mathrm{c}_2 g + \mathds{1}\otimes H'_0 + \mathrm{Rem}_R ) \mathds{1}_{ [1,\infty) }( N_{\mathrm{neut},\infty} ) \check{\Gamma}(j^R) \\ &\ge \check{\Gamma}(j^R)^* ((1 - \mathrm{c}_1 g) H'_0 \otimes\mathds{1} - \mathrm{c}_2 g + \mathds{1}\otimes H'_0 \\ &\quad + \mathrm{Rem}_R (N_0+N_{\infty}+1)^{-1}(N_0+N_{\infty}+1) ) \mathds{1}_{ [1,\infty) }( N_{\mathrm{neut},\infty} ) \check{\Gamma}(j^R) \\ &\ge \check{\Gamma}(j^R)^* ((1 - \mathrm{c}_1 g) H'_0 \otimes\mathds{1} - \mathrm{c}_2 g + \mathds{1}\otimes H'_0 + o(R^{0})(N_0+N_{\infty}+1) ) \mathds{1}_{ [1,\infty) }( N_{\mathrm{neut},\infty} ) \check{\Gamma}(j^R). \end{align*} For $g$ small enough and $R$ large enough, since $N_{\mathrm{lept}}$ and $N_W$ are relatively $H$-bounded, this yields \begin{align} & \check{\Gamma}(j^R)^* (H'_0\otimes\mathds{1}+\mathds{1}\otimes H'_0 + g H'_I \otimes \mathds{1} + \mathrm{Rem}_R ) \mathds{1}_{ [1,\infty) }( N_{\mathrm{neut},\infty} ) \check{\Gamma}(j^R) \notag \\ & \ge \check{\Gamma}(j^R)^*((1-\mathrm{c}_1 g)H'_0\otimes\mathds{1}-\mathrm{c}_2g+\mathds{1}\otimes H'_0+o(R^0)(N_{\mathrm{neut},0}+N_{\mathrm{neut},\infty})) \mathds{1}_{ [1,\infty) }( N_{\mathrm{neut},\infty} ) \check{\Gamma}(j^R) \notag \\ & + o( R^0 ) \check{\Gamma}(j^R)^* \mathds{1}_{ [1,\infty) }( N_{\mathrm{neut},\infty} ) \check{\Gamma}(j^R) ( 1 + H^2 )^{\frac12} \notag \\ & \ge \frac14 \check{\Gamma}(j^R)^* \mathds{1}_{ [1,\infty) }( N_{\mathrm{neut},\infty} ) \check{\Gamma}(j^R) + o( R^0 ) \check{\Gamma}(j^R)^* \mathds{1}_{ [1,\infty) }( N_{\mathrm{neut},\infty} ) \check{\Gamma}(j^R) ( 1 + H^2 )^{\frac12}. \label{eq:comput_Mourre_1} \end{align} In the last inequality, we used \eqref{eq:lsif1} and the fact that $H'_0\ge0$. Now, we consider the first term in \eqref{eq:decomp_1}. As in the proof of Theorem \ref{Thess2}, we use the fact that, for all bounded interval $I$ and $R>0$, $\check{\Gamma}(j^R)^* ( \mathds{1} \otimes \Pi_\Omega ) \check{\Gamma}(j^R) \chi_I(H)$ is compact. Thus, using that $N_{\mathrm{neut}}\le H'_0$, that $H'_I$ is relatively $(N_{\mathrm{lept}}+N_W)$-bounded and that $N_{\mathrm{lept}}$ and $N_W$ are relatively $H$-bounded, we can write \begin{align} & \check{\Gamma}(j^R)^* (H'_0\otimes\mathds{1}+\mathds{1}\otimes H'_0 + g H'_I \otimes \mathds{1} + \mathrm{Rem}_R ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{\{0\}}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \notag \\ &= \check{\Gamma}(j^R)^* (H'_0\otimes\mathds{1} + g H'_I \otimes \mathds{1} + \mathrm{Rem}_R ) (\mathds{1} \otimes \Pi_\Omega ) )\check{\Gamma}(j^R) \notag \\ &\ge \check{\Gamma}(j^R)^* ( H'_0 \otimes \mathds{1} + g H'_I \otimes \mathds{1} + o( R^0 )( N_{\mathrm{neut}} \otimes \mathds{1} + N_{\mathrm{lept}} \otimes \mathds{1} + N_W \otimes \mathds{1} ) ) (\mathds{1} \otimes \Pi_\Omega ) )\check{\Gamma}(j^R) \notag \\ &\ge \check{\Gamma}(j^R)^* ( (1+o( R^0 )) H'_0 \otimes \mathds{1} + g H'_I \otimes \mathds{1} + o( R^0 )( N_{\mathrm{lept}} \otimes \mathds{1} + N_W \otimes \mathds{1} ) ) (\mathds{1} \otimes \Pi_\Omega ) )\check{\Gamma}(j^R) \notag \\ &\ge \check{\Gamma}(j^R)^* ( g H'_I \otimes \mathds{1} + o( R^0 )( N_{\mathrm{lept}} \otimes \mathds{1} + N_W \otimes \mathds{1} ) ) (\mathds{1} \otimes \Pi_\Omega ) )\check{\Gamma}(j^R) \notag \\ & \ge K_{R,I} - \mathrm{c}_3 \check{\Gamma}(j^R)^* (\mathds{1} \otimes \Pi_\Omega ) \check{\Gamma}(j^R) ( \mathds{1} - \chi_I( H ) ) ( 1 + H^2 )^{\frac12} , \label{eq:comput_Mourre_2} \end{align} for any bounded interval $I$ and $R>0$ large enough, where $K_{R,I}$ is compact and $\mathrm{c}_3$ is a positive constant. It remains to consider \begin{align*} & \check{\Gamma}(j^R)^* (H'_0\otimes\mathds{1}+\mathds{1}\otimes H'_0 + g H'_I \otimes \mathds{1} + \mathrm{Rem}_R ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \\ &\ge \check{\Gamma}(j^R)^* ( H'_0\otimes\mathds{1} + g H'_I \otimes \mathds{1} + o(R^0)(N_0+N_{\infty}+1) ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R). \end{align*} Here we used that $\mathds{1} \otimes H'_0 \ge 0$. We introduce $\mathds{1} = \chi_I(H) + ( \mathds{1} - \chi_I(H) )$ on the right. Using that $H'_0 \otimes \mathds{1} \ge 0$, we can write \begin{align} & \check{\Gamma}(j^R)^* (H'_0\otimes\mathds{1}+\mathds{1}\otimes H'_0 + g H'_I \otimes \mathds{1} + \mathrm{Rem}_R ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \notag \\ &\ge \check{\Gamma}(j^R)^* (( 1 - o( R^0 ) ) ( H'_0 + g H'_I ) \otimes\mathds{1} ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \chi_I(H) \notag \\ &\quad + \check{\Gamma}(j^R)^* (o(R^0)g H'_I \otimes \mathds{1} + o(R^0)(N_0+N_{\infty}+1) ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R)\chi_I( H ) \notag \\ &\quad + \check{\Gamma}(j^R)^* ( H'_0\otimes\mathds{1} + g H'_I \otimes \mathds{1} + o(R^0)(N_0+N_{\infty}+1) ) \notag \\ & \qquad \qquad \qquad \qquad \qquad \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) ( \mathds{1} - \chi_I( H ) ) . \notag \end{align} Hence, using again that $H'_I$ is relatively $(N_{\mathrm{lept}}+N_W)$-bounded and that $N_{\mathrm{lept}}$ and $N_W$ are relatively $H$-bounded, we obtain that \begin{align} & \check{\Gamma}(j^R)^* (H'_0\otimes\mathds{1}+\mathds{1}\otimes H'_0 + g H'_I \otimes \mathds{1} + \mathrm{Rem}_R ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \notag \\ &\ge \check{\Gamma}(j^R)^* (( 1 - o( R^0 ) ) ( H'_0 + g H'_I ) \otimes\mathds{1} ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \chi_I(H) \notag \\ &\quad + o(R^0) \check{\Gamma}(j^R)^* \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \chi_I(H) \notag \\ &\quad - \mathrm{c}_4 \check{\Gamma}(j^R)^* \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} ) \check{\Gamma}(j^R) ( \mathds{1} - \chi_I( H ) ) ( 1 + H^2 )^{\frac12} , \label{eq:comput_Mourre_9} \end{align} for any bounded interval $I$ and $R>0$, where $\mathrm{c}_4$ is a positive constant. Now, we can prove $H_1(1)$ and \eqref{Statement5}. To prove $H_1(1)$, let $\varepsilon>0$, $\lambda \in [E,E+ m_1)$ and $\Delta$ be an interval containing $\lambda$ and supported in $[E,E+m_1)$. The function $\chi_\Delta \in \mathrm{C}_0^\infty( \mathbb{R} )$ is chosen such that $\mathrm{supp} (\chi_\Delta) \subset ( - \infty , E+m_1)$ and we consider in addition $\tilde\chi_\Delta \in \mathrm{C}_0^\infty( \mathbb{R} )$ such that $\mathrm{supp} (\tilde\chi_\Delta) \subset ( - \infty , E+m_1)$ and $\tilde\chi_\Delta \chi_\Delta = \chi_\Delta$. Then, using that $H'_0 \ge 0$, the fact that $H'_I$ is $(N_{\mathrm{lept}}+N_W)$-bounded as before, and Lemma \ref{numberestimates13}, the first term in the right-hand-side of \eqref{eq:comput_Mourre_9} (with $I=\Delta$) can be estimated, for $R$ large enough, in the following way: \begin{align} &\check{\Gamma}(j^R)^* (( 1 - o( R^0 ) ) ( H'_0 + g H'_I ) \otimes\mathds{1} ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \chi_\Delta(H) \notag \\ &\ge\check{\Gamma}(j^R)^* (( 1 - o( R^0 ) ) H'_0 \otimes\mathds{1} ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \chi_\Delta(H) \notag \\ &\quad - \mathrm{c}_5 \check{\Gamma}(j^R)^* \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \chi_\Delta(H) \notag \\ &\ge\check{\Gamma}(j^R)^* (( 1 - o( R^0 ) ) H'_0 \otimes\mathds{1} ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \chi_\Delta(H) \notag \\ & - \mathrm{c}_5 \check{\Gamma}(j^R)^* \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} ) \tilde \chi_\Delta(H^{\mathrm{ext}}) \check{\Gamma}(j^R) \chi_\Delta(H) \notag \\ &\ge - \mathrm{c}_5 \check{\Gamma}(j^R)^* \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} ) \tilde \chi_\Delta(H^{\mathrm{ext}}) \check{\Gamma}(j^R) \chi_\Delta(H) \notag \\ &=0. \label{eq:comput_Mourre_10} \end{align} The last equality comes from the fact that $H^{\mathrm{ext}} = H \otimes \mathds{1}+\mathds{1} \otimes H_0 \ge E + m_1$ on the range of the operator $\mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )$, and hence that $\tilde \chi_\Delta(H^{\mathrm{ext}})\mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )=0$, since $\tilde\chi_\Delta$ is supported in $(-\infty,E+m_1)$. Since $d(\lambda)=0$ for $\lambda \in (-\infty,E+m_1)$, fixing $R$ large enough, $I=\Delta$ in \eqref{eq:comput_Mourre_2} and \eqref{eq:comput_Mourre_9}, and combining Equations \eqref{eq:decomp_1}--\eqref{eq:comput_Mourre_10}, we obtain $H_1(1)$. To prove \eqref{Statement5}, let $\varepsilon>0$ and $\lambda \in [E,E+ n m_1 )$. We go back to \eqref{eq:comput_Mourre_9}, with $I=\Delta^{\delta}_\mu$, and consider again the first term in the right-hand-side. By $S_1(n)$, $\tau$, where $d$ vanishes, is a closed countable set in $[E,E+n m_1 ]$, which implies that there exists $\kappa$ such that $\tau \cap \Delta^{\kappa}_{\lambda} \neq \emptyset$ and then $d(\lambda) = \sup_{\kappa > 0} d^{\kappa}(\lambda)$. Hence there exists $\kappa > 0$ such that $d^{\kappa}(\lambda)> d(\lambda)-\varepsilon/3$. Let $\varepsilon_0>0$ be such that $\lambda \in [E,E+ n m_1 -\varepsilon_0 ]$. By $H_3(n-1)$, we know that there exist $\mathrm{c}_6\in\mathbb{R}$ and $\delta >0$ such that \begin{equation} H'_0 + gH'_I \geq ( \min ( \frac14 , \tilde{d}^{\kappa}(\mu) )- \frac{\varepsilon}{3}) \mathds{1} - \mathrm{c}_6 ( \mathds{1} - \tilde\chi_{\Delta^{\delta}_{\mu}} (H) )( 1 + H^2 )^{\frac12} , \label{eq:H3(n)} \end{equation} for all $\mu \in [E , E+(n-1)m_1 - \varepsilon_0 ]$. Here $\tilde\chi_{\Delta^{\delta}_{\lambda}}$ is chosen such that $\tilde\chi_{\Delta^{\delta}_{\lambda}}\chi_{\Delta^{\delta}_{\lambda}}=\tilde\chi_{\Delta^{\delta}_{\lambda}}$, where $\chi_{\Delta^{\delta}_{\lambda}}$ is the function appearing in \eqref{eq:comput_Mourre_9}. We begin by estimating the first term in the right-hand-side of \eqref{eq:comput_Mourre_9} as \begin{align} & \check{\Gamma}(j^R)^* (( 1 - o( R^0 ) ) ( H'_0 + g H'_I ) \otimes\mathds{1} ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \chi_{\Delta_\lambda^\delta}(H) \notag \\ &\geq \check{\Gamma}(j^R)^* (( 1 - o( R^0 ) ) ( \min ( \frac14 , \tilde{d}^{\kappa}(\mu) )- \frac{\varepsilon}{3}) \mathds{1} \otimes\mathds{1} ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \chi_{\Delta_\lambda^\delta}(H) \notag \\ & - \mathrm{c}_6 \check{\Gamma}(j^R)^* (( 1 - o( R^0 ) ) (( \mathds{1} - \tilde\chi_{\Delta^{\delta}_{\mu}} (H) )( 1 + H^2 )^{\frac12} \otimes\mathds{1} ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \chi_{\Delta_\lambda^\delta}(H) .\notag \end{align} On the range of $\mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )$, we have that $\mathds{1} \otimes H_0 \ge m_1$. Therefore, by \eqref{eq:H3(n)} and the functional calculus, with $\mu = \lambda - 1 \otimes H_0$, we obtain that \begin{align} & ( 1 - o(R^0)) \check{\Gamma}(j^R)^* ( \min ( \frac14 , \tilde{d}^\kappa( \lambda - \mathds{1} \otimes H_0 ) ) - \frac{\varepsilon}{3} ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \chi_{\Delta_\lambda^\delta}(H) \notag \\ & = ( 1 - o(R^0)) \check{\Gamma}(j^R)^* ( \min ( \frac14 , d^\kappa( \lambda ) ) - \frac{\varepsilon}{3} ) \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \chi_{\Delta_\lambda^\delta}(H) \notag \\ & \ge ( 1 - o(R^0)) ( \min ( \frac14 , d(\lambda) ) - \frac{2\varepsilon}{3} ) \check{\Gamma}(j^R)^* \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \chi_{\Delta_\lambda^\delta}(H) \notag \\ & \ge ( \min ( \frac14 , d(\lambda) ) - \varepsilon ) \check{\Gamma}(j^R)^* \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \chi_{\Delta_\lambda^\delta}(H) , \label{eq:comput_Mourre_12} \end{align} where the equality comes from the definitions of $\tilde{d}^\kappa$ and $d^\kappa$ and the fact that $\tilde{d}^\kappa( \lambda - \mathds{1} \otimes H_0 )$ acts on the range of $\mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )$. Moreover, let $\tilde{\tilde{\chi}}_{\Delta_\lambda^\delta} \in \mathrm{C}_0^{\infty}(\mathbb{R})$ be such that $\tilde{\tilde{\chi}}_{\Delta_\lambda^\delta}\chi_{\Delta_\lambda^\delta}= \chi_{\Delta_\lambda^\delta}$. Using Lemma \ref{numberestimates3}, we write \begin{align} & \mathrm{c}_6 \check{\Gamma}(j^R)^* (( 1 - o( R^0 ) ) ( ( \mathds{1} - \tilde\chi_{\Delta^{\delta}_{\mu}} (H) )( 1 + H^2 )^{\frac12} \otimes\mathds{1} ) \notag \\ &\qquad \qquad \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} )\check{\Gamma}(j^R) \chi_{\Delta_\lambda^\delta}(H) \tilde{\tilde{\chi}}_{\Delta_\lambda^\delta}(H) \notag \\ & = \mathrm{c}_6(( 1 - o( R^0 ) ) \check{\Gamma}(j^R)^* \mathds{1}_{\{0\}}( N_{\mathrm{neut},\infty} ) \mathds{1}_{[1,\infty)}( N_{\mathrm{lept},\infty} + N_{W,\infty} ) \notag \\ &\qquad \qquad ((( \mathds{1} - \tilde{\chi}_{\Delta^{\delta}_{\mu}} (H) )( 1 + H^2 )^{\frac12} \otimes\mathds{1} )\chi_{\Delta_\lambda^\delta}(H^{\mathrm{ext}})\check{\Gamma}(j^R) \tilde{\tilde{\chi}}_{\Delta_\lambda^\delta}(H) + o(R^0) \notag \\ & = o(R^0). \label{eq:comput_Mourre_13} \end{align} The last equality comes from $\mu = \lambda - \mathds{1}\otimes H_0$ and $\tilde\chi_{\Delta^{\delta}_{\lambda}}\chi_{\Delta^{\delta}_{\lambda}}=\tilde\chi_{\Delta^{\delta}_{\lambda}}$. Fixing $R$ large enough and $I = \Delta_\lambda^\delta$, Equations \eqref{eq:decomp_1}, \eqref{eq:comput_Mourre_1}, \eqref{eq:comput_Mourre_2}, \eqref{eq:comput_Mourre_9}, \eqref{eq:comput_Mourre_12} and \eqref{eq:comput_Mourre_13} prove \eqref{Statement5}. This concludes the proof of \eqref{eq:Mourre_estimate_1_4}. The fact that $\sigma_{\mathrm{sc}}(H) = \emptyset$ if \eqref{eq:hypclassC2_2_4} holds follows from Lemma \ref{lem:C8} and the abstract results recalled above. \end{proof} \section{Propagation estimates}\label{sec:propag} In this section, we use the method of propagation observables that was developed in $N$-body scattering theory (see e.g. \cite{SiSo87_01,HuSiSo99_01,Gr90_01,De93_01,DeGe97_01} and references therein). This method was adapted to the context of Pauli-Fierz or $P(\varphi)_2$ Hamiltonians in several papers (see, in particular, \cite{DeGe99_01,DeGe00_01,FrGrSc02_01,FrGrSc04_01,Am04_01,FrGrSc07_01,BoFaSi12_01,FaSi14_01}). In the case where all particles are massive, the propagation estimates that we prove in this section are straightforwardly adapted from \cite{Gr90_01,DeGe99_01}. In the case where neutrinos are supposed to be massless, however, the propagation estimates have to be substantially modified. The main idea consists in replacing the usual one-particle position operator by a suitably modified, time-dependent one-particle ``position'' operator. A related trick was used in \cite{FaSi14_01} for Pauli-Fierz Hamiltonians, but the details of the analysis here are different, in particular because we rely heavily on the Mourre estimate proven in Theorem \ref{MourreI_3}. The basic approach that we follow to prove our propagation estimates is the following: let $H$ be a self-adjoint operator on a Hilbert space $\mathscr{H}$ and let $\Phi(t)$ be a time-dependent family of self-adjoint operators. Suppose that for some $u \in \mathcal{H}$, \begin{align*} & \big \langle e^{ - \mathrm{i} t H } u , \Phi( t ) e^{ - \mathrm{i} t H } u \big \rangle \le \mathrm{C}_u , \quad \text{ uniformly in } t \ge 1 , \end{align*} and that one of the following to conditions holds, \begin{align} & \partial_t \big \langle e^{ - \mathrm{i} t H } u , \Phi( t ) e^{ - \mathrm{i} t H } u \big \rangle \ge \big \langle e^{ - \mathrm{i} t H } u , \Psi(t) e^{ - \mathrm{i} t H } u \big \rangle - \sum_{j=1}^n \big \langle e^{ - \mathrm{i} t H } u , B^*_j(t) B_j(t) e^{ - \mathrm{i} t H } u \big \rangle, \label{eq:condPO} \\ & \partial_t \big \langle e^{ - \mathrm{i} t H } u , \Phi( t ) e^{ - \mathrm{i} t H } u \big \rangle \le - \big \langle e^{ - \mathrm{i} t H } u , \Psi(t) e^{ - \mathrm{i} t H } u \big \rangle + \sum_{j=1}^n \big \langle e^{ - \mathrm{i} t H } u , B^*_j(t) B_j(t) e^{ - \mathrm{i} t H } u \big \rangle, \label{eq:condPO_1} \end{align} where $\Psi(t)$ are positive operators and $B_i(t)$ are families of time-dependent operators such that \begin{equation} \int_1^\infty \big \| B_i(t) e^{-\mathrm{i}tH}u\big\|^2dt \le\mathrm{C}_u. \label{eq:cond2PO} \end{equation} Then, integrating with respect to $t$, one obtains that \begin{align} \int_1^\infty \big \langle e^{ - \mathrm{i} t H } u , \Psi( t ) e^{ - \mathrm{i} t H } u \big \rangle dt \lesssim \mathrm{C}_u , \label{eq:concPO} \end{align} which is sometimes called a weak propagation estimate for the family of observables $\Psi(t)$. Observe that the left-hand-sides of \eqref{eq:condPO}--\eqref{eq:condPO_1} can be rewritten as \begin{equation*} \big \langle e^{ - \mathrm{i} t H } u , \mathbf{D} \Phi( t ) e^{ - \mathrm{i} t H } u \big \rangle , \end{equation*} where $\mathbf{D}$ stands for the Heisenberg derivative $\mathbf{D} \Phi ( t ) = \partial_t \Phi( t ) + [ H , \mathrm{i} \Phi( t ) ]$. Therefore, to prove the propagation observable \eqref{eq:concPO}, it suffices to find a family of operators $\Phi(t)$ whose Heisenberg derivative ``dominates'' $\Psi(t)$, in the sense that $\mathbf{D} \Phi ( t ) \ge \Psi(t)$ or $\mathbf{D} \Phi ( t ) \le - \Psi(t)$, up to remainder terms that are integrable in the sense of \eqref{eq:cond2PO}. The strategy usually consists in comparing the time derivative $\partial_t \Phi( t )$ and the commutator $[ H , \mathrm{i} \Phi( t ) ]$, possibly by means of a ``commutator expansion" of $[ H , \mathrm{i} \Phi( t ) ]$. We refer the reader to e.g. \cite{DeGe97_01} for details on the method of propagation observables, see also \cite[Section 2]{FaSi14_01} for a description of the method of propagation observables for Pauli-Fierz Hamiltonians, closely related to the approach that we will follow here. It is useful to introduce the following notations \begin{align*} \mathbf{d}^{(i)}_{0l} b(t) = \frac{\partial b}{\partial t}(t) + [\omega^{(i)}_l(p_i), \mathrm{i} b(t) ] , \quad i=1,2 , \quad \mathbf{d}^{(3)}_{0} b(t) = \frac{\partial b}{\partial t}(t) + [\omega^{(3)}(p_i), \mathrm{i} b(t) ] , \end{align*} if $b(t)$ is a family of operators, acting on $\mathfrak{h}_1$ if $i=1,2$, or acting on $\mathfrak{h}_2$ if $i=3$. Likewise, we set \begin{align*} \mathbf{D}_0 B(t) = \frac{\partial B}{\partial t}(t) + [H_0, \mathrm{i} B(t) ] , \quad \mathbf{D} B(t) = \frac{\partial B}{\partial t}(t) + [H, \mathrm{i} B(t) ] , \end{align*} if $B(t)$ is a family of operators acting on $\mathscr{H}$. Note that if $B(t) = (b_1(t) , \dots , b_{14}(t) )$, with the notations of Appendix \ref{sec:standard_def} (remembering that the total Hilbert space $\mathscr{H}$ is the tensor product of $14$ Fock spaces), then, as functions of $t$, \begin{align*} \mathbf{D}_0 \mathrm{d}\Gamma(B) =: \mathrm{d}\Gamma( \mathbf{d}_0 b ) := \mathrm{d}\Gamma( & \mathbf{d}_{0,1} b_1 , \dots , \mathbf{d}_{0,14} b_{14} ) \\ := \mathrm{d}\Gamma\big( & \mathbf{d}^{(3)}_{0} b_1 , \mathbf{d}^{(3)}_{0} b_2 , \mathbf{d}^{(1)}_{01} b_3 , \mathbf{d}^{(1)}_{01} b_4 , \mathbf{d}^{(2)}_{01} b_5 , \mathbf{d}^{(2)}_{01} b_6 , \mathbf{d}^{(1)}_{02} b_7, \\ &\mathbf{d}^{(1)}_{02} b_8 , \mathbf{d}^{(2)}_{02} b_9 , \mathbf{d}^{(2)}_{02} b_{10} , \mathbf{d}^{(1)}_{03} b_{11} ,\mathbf{d}^{(1)}_{03} b_{12} , \mathbf{d}^{(2)}_{01} b_{13} , \mathbf{d}^{(2)}_{01} b_{14} \big) \end{align*} In the remainder of this section, we prove the propagation estimates that will be used in Section \ref{sec:AC}. We begin with the case where the masses of the neutrinos are supposed to be positive in Section \ref{subsec:massive_propag}, next we turn to the much more difficult case where neutrinos are supposed to be massless (Section \ref{subsec:massless_propag}). \subsection{Massive neutrinos}\label{subsec:massive_propag} As mentioned before, the proofs of the propagation estimates of this section are almost straightforward adaptations of the ones in \cite{DeGe99_01,DeGe00_01}. We state the results and emphasize the differences with \cite{DeGe99_01,DeGe00_01} but the details of the proofs are left to the reader. To shorten expressions below, we set $x_i = (x_i^{(1)},x_i^{(2)},x_i^{(3)}) = \mathrm{i} \nabla$, $i=1,\dots,14$, and, for $R=(R_1,\dots,R_{14})$ and $R'=(R'_1,\dots,R'_{14})$, \begin{equation}\label{eq:1Rx} \mathds{1}_{[R,R']}(|x|)=\big(\mathds{1}_{[R_1,R'_1]}(|x_1|),\dots,\mathds{1}_{[R_{14},R'_{14}]}(|x_{14}|)\big)=\big(\mathds{1}_{[R_1,R'_1]}(\sqrt{-\Delta}),\dots,\mathds{1}_{[R_{14},R'_{14}]}(\sqrt{-\Delta})\big), \end{equation} and likewise for other functions of $x$. The operators $\mathrm{d}\Gamma(\mathds{1}_{[R,R']}(|x|))$, $\Gamma(\mathds{1}_{[R,R']}(|x|))$ are then defined, as in the previous sections, following the conventions of Appendix \ref{sec:standard_def}. We also set \begin{equation}\label{eq:not_omega} \omega := (\omega^{(3)} , \omega^{(3)} , \omega^{(1)}_1 , \omega^{(1)}_1 , \omega^{(2)}_1 , \omega^{(2)}_1 , \omega^{(1)}_2 , \omega^{(1)}_2 , \omega^{(2)}_2 , \omega^{(2)}_2 , \omega^{(1)}_3 , \omega^{(1)}_3 , \omega^{(2)}_3 , \omega^{(2)}_3 ). \end{equation} Recall that, for $i=1,2$, the notation ``$a_{(i),\cdot} G \in L^2$'' means that, for all $j$, $l$ and $\epsilon$, $a_{(i),l} G_{l,\epsilon}^{(j)}$ is square integrable, and that ``$a_{(3),\cdot} G \in L^2$'' means that for all $j$, $l$ and $\epsilon$, $a_{(3)} G_{l,\epsilon}^{(j)}$ is square integrable. \begin{Th}\label{th:propag_massive} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ are positive and consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. Assume that \begin{equation*} G \in L^2, \quad a_{(i),\cdot} G \in L^2, \quad i = 1,2,3. \end{equation*} and that \begin{equation*} G \in \mathbb{H}^{1+\mu} \text{ for some } \mu > 0. \end{equation*} \begin{itemize} \item[(i)] Let $\chi \in \mathrm{C}_0^{\infty}(\mathbb{R})$, $R=(R_1,\dots,R_{14})$ and $R'=(R'_1,\dots,R'_{14})$ be such that $R'_{i}>R_i>1$. There exists $\mathrm{C} > 0$ such that, for all $u \in \mathscr{H}$, \begin{equation*} \int^{\infty}_{1} \Big \| \mathrm{d} \Gamma \Big ( \mathds{1}_{[R,R']} \Big (\frac{|x|}{t} \Big ) \Big)^{\frac{1}{2}}\chi(H) e^{-\mathrm{i}tH} u \Big \|^2 \frac{dt}{t} \leq \mathrm{C} \|u\|^{2}. \end{equation*} \item[(ii)] Let $0<\mathrm{v}_0<\mathrm{v}_1$ and $\chi \in \mathrm{C}_0^{\infty}(\mathbb{R})$. There exists $\mathrm{C}>0$ such that, for all $\ell \in \{ 1,2,3\} $ and $u \in \mathscr{H}$, \begin{align*} & \int^{\infty}_{1} \Big \| \mathrm{d}\Gamma \Big( \Big\langle \big( \frac{x}{t} - \nabla \omega \big ) , \mathds{1}_{ [ \mathrm{v}_0 , \mathrm{v}_1 ] }( \frac{x}{t} ) \big( \frac{x}{t} - \nabla \omega \big ) \Big\rangle \Big)^{\frac12} \chi(H)e^{-\mathrm{i} t H} u \Big\|^{2} \frac{dt}{t} \le \mathrm{C} \| u \|^2. \end{align*} \item[(iii)] Let $0<\mathrm{v}_0<\mathrm{v}_1$, $J \in \mathrm{C}_0^{\infty}( \{ x \in \mathbb{R}^3 , \mathrm{v}_0 < |x| < \mathrm{v}_1\})$ and $\chi \in \mathrm{C}_0^{\infty}(\mathbb{R})$. There exists $\mathrm{C}>0$ such that, for all $\ell \in \{ 1,2,3\} $ and $u \in \mathscr{H}$, \begin{equation*} \int^{\infty}_{1} \Big \| \mathrm{d}\Gamma \Big( \Big| J\Big(\frac{x}{t}\Big) \Big( \frac{x^{(\ell)}}{t} - \partial_{(\ell)} \omega \Big) + \mathrm{h.c }\Big|\Big)^{\frac{1}{2}}\chi(H)e^{-\mathrm{i} t H} u \Big\|^{2} \frac{dt}{t} \le \mathrm{C} \| u \|^2. \end{equation*} \item[(iv)] Let $\chi \in \mathrm{C}_{0}^{\infty}(\mathbb{R})$ be supported in $\mathbb{R}\backslash(\tau \cup \sigma_{\mathrm{pp}}(H))$. There exists $\epsilon>0$ and $\mathrm{C} > 0$ such that, for all $u \in \mathscr{H}$, \begin{equation*} \int^{\infty}_{1} \Big\| \Gamma \Big ( \mathds{1}_{[0,\epsilon]} \Big(\frac{|x|}{t}\Big)\Big) \chi(H) e^{-\mathrm{i} t H} u \Big\|^2 \frac{dt}{t} \leq \mathrm{C}\| u \|^2. \end{equation*} \end{itemize} \end{Th} \begin{proof} We give a sketch of the proof of (i), underlying the differences with \cite{DeGe99_01}. The proofs of (ii)--(iv) can be achieved by similar arguments, adapting \cite[Section 6]{DeGe99_01} in a straightforward way. By definition of the operator $\mathrm{d}\Gamma$, we have that \begin{align} \int^{\infty}_{1} \Big \| \mathrm{d}\Gamma \Big( \mathds{1}_{[R,R']} \Big(\frac{|x|}{t}\Big)\Big)^{\frac{1}{2}}\chi(H)e^{-\mathrm{i}tH} u \Big\|^2 \frac{dt}{t} & = \int^{\infty}_{1} \Big\| \sum^{14}_{i=1} \mathrm{d}\Gamma_i \Big(\mathds{1}_{[R_i,{R}'_i]}\Big(\frac{\sqrt{-\Delta}}{t}\Big)\Big)^{\frac{1}{2}}\chi(H) e^{-\mathrm{i}tH} u \Big\|^2 \frac{dt}{t} \notag \\ & \lesssim \sum^{14}_{i=1} \int^{\infty}_{1} \Big\|\mathrm{d}\Gamma_i\Big(\mathds{1}_{[R_i,{R}'_i]}\Big(\frac{\sqrt{-\Delta}}{t}\Big)\Big)^{\frac{1}{2}}\chi(H) e^{-\mathrm{i}tH} u \Big\|^2 \frac{dt}{t}, \label{eq:hdt1} \end{align} where we have set \begin{equation}\label{eq:defdgammaj} \mathrm{d}\Gamma_{i}(B) = \underset{i-1}{\underbrace{\mathds{1}\otimes \dots \otimes \mathds{1}}}\otimes \mathrm{d}\Gamma(B) \otimes \underset{14-i}{\underbrace{\mathds{1}\otimes \dots \otimes\mathds{1}}}. \end{equation} Next we proceed as in \cite{DeGe99_01,Am04_01}. Let $F \in \mathrm{C}^{\infty}(\mathbb{R})$ be equal to 1 near $\infty$, to 0 on a compact set near the origin, and such that $F'(s) \geq \mathds{1}_{[R_i,R'_i]}(s) $. We define in addition \begin{align*} \Phi(t) & = \chi(H) \mathrm{d}\Gamma_i\Big(F\Big(\frac{|x_i|}{t}\Big)\Big)\chi(H) , \qquad b(t) = \mathbf{d}_{0, i} F\left(\frac{|x_i|}{t}\right). \end{align*} A direct computation then shows that \begin{equation*} \mathbf{D} \Phi(t) = \chi(H) b(t) \chi(H) + \chi(H) \, \mathrm{i} \Big[ H_I,\mathrm{d}\Gamma_i\Big(F\Big(\frac{|x_i|}{t}\Big)\Big) \Big ]\chi(H) . \end{equation*} Using pseudo-differential calculus (or a commutator expansion at second order, see, e.g., \cite{DeGe97_01}) gives \begin{align*} b(t) \leq \frac{ - \mathrm{C}_0}{t} \mathds{1}_{[R_i,R'_i]}(t) + \mathcal{O}(t^{-2}) . \end{align*} Besides, since $F_i$ vanishes near $0$ and since $G \in \mathbb{H}^{1+\mu}$, we have that \begin{align*} \Big\|F_i\Big(\frac{|x_i|}{t}\Big)G\Big\|_2=\mathcal{O}(t^{-1-\mu}). \end{align*} Therefore one can deduce from Lemma \ref{lemSCW_1} together with the fact that $N$ is relatively $H$-bounded that \begin{align*} \chi( H ) \Big [ H_I , \mathrm{d}\Gamma_i\Big(F\Big(\frac{|x_i|}{t}\Big)\Big) \Big] \chi( H ) = \mathcal{O}(t^{-1-\mu}). \end{align*} Hence the condition \eqref{eq:condPO} is satisfied and it suffices to apply the abstract method recalled at the beginning of Section \ref{sec:propag}. \end{proof} \subsection{Massless neutrinos}\label{subsec:massless_propag} In this section we prove the counterpart of Theorem \ref{th:propag_massive} in the case where the masses of the neutrinos vanish. As mentioned at the beginning of Section \ref{sec:propag}, for massless neutrinos we cannot directly adapt \cite{DeGe99_01,DeGe00_01}, hence some parts of the proof require substantial modifications. A first issue comes from the fact that, in order to control remainder terms in some commutator expansions, one needs to control the second quantization of expressions of the form $[ \mathrm{i} \nabla_{p_2} , [ \mathrm{i} \nabla_{p_2} , \omega^{(2)}_l(p_2) ] ]$, where $\omega^{(2)}_l$ are the dispersion relations of neutrinos. In the massive case, such commutators are bounded. Their second quantizations are therefore relatively $N$-bounded and, hence, relatively $H$-bounded. In the massless case, however, $[ \mathrm{i} \nabla_{p_2} , [ \mathrm{i} \nabla_{p_2} , \omega^{(2)}_l(p_2) ] ]$ is of order $\mathcal{O}(|p_2|^{-1})$ near $0$. To overcome this difficulty, the idea is to show that $\langle e^{ - \mathrm{i} t H } u , \mathrm{d} \Gamma( |p_2|^{-1} ) e^{ - \mathrm{i} t H } u \rangle \lesssim t^\gamma$ for some $\gamma<1$ and for suitable states $u \in \mathscr{H}$. This idea was used in \cite{BoFaSi12_01,FaSi14_01,FaSi14_02} for Pauli-Fierz Hamiltonians, modifying an argument of \cite{Ge02_01}. In Lemma \ref{lem:evolp2-1}, we adapt \cite{BoFaSi12_01,FaSi14_01,FaSi14_02} to our context. A second issue that has to be dealt with, in the massless case, to prove a suitable minimal velocity estimate, is that we have to use the Mourre estimate stated in Theorem \ref{MourreI_3}, with a positive term proportional to the number operator $N_{\mathrm{neut}}$. This is again required in order to be able to control some remainder terms. Some care must also be taken because of the fact that the considered conjugate operator (see \eqref{eq:conjugate_1}) is not self-adjoint. Propagation estimates analogous to those of Theorem \ref{th:propag_massive} are established in Theorem \ref{th:propag_massless}. An important difference with the massive case is that the propagation estimates hold only for states in a dense subset of the total Hilbert space. In Theorem \ref{th:propag_massless_2}, we prove a second version of propagation estimates involving a time-dependent modified one-particle ``position'' operator. Theorem \ref{th:propag_massless_2} will be a crucial input in our proof of asymptotic completeness of the wave operators in Section \ref{sec:AC}. For shortness, we set $\mathrm{d} \Gamma ( | p_2 |^{-\alpha} ) = \mathrm{d} \Gamma ( h^\alpha )$ where $h^\alpha=(h^\alpha_1,\dots,h^\alpha_{14})$, according to the notations of Appendix \ref{sec:standard_def}, with $h^\alpha_i=0$ if $i\in\{1,2,3,4,7,8,11,12\}$ (corresponding to the label of a massive particle), and $h^\alpha_i$ is the operator of multiplication by $|p_2|^{-\alpha}$ if $i \in \{5,6,9,10,13,14\}$ (corresponding to the label of a neutrino). The first part of the following lemma (see \eqref{eq:estimp2-1}) is adapted from \cite[Lemma 4.1]{BoFaSi12_01}. For technical reasons, that will appear in the proofs of Theorems \ref{th:propag_massless} and \ref{th:propag_massless_2} below, we also need a new, related estimate, see \eqref{eq:estimp2-1_a}. \begin{lem}\label{lem:evolp2-1} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ vanish and consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. Let $0 \le \alpha \le 1$ and assume that \begin{equation*} G \in L^2 , \quad |p_2|^{-1-\mu} G \in L^2 \text{ for some } \mu > 0. \end{equation*} Let $\chi \in \mathrm{C}_0^\infty ( \mathbb{R} )$. There exists $\mathrm{C}>0$ such that, for all $u \in \mathscr{D}( \mathrm{d}\Gamma( |p_2|^{-\alpha} )^{\frac12} )$, \begin{equation} \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{d} \Gamma ( | p_2 |^{-\alpha} ) e^{ - \mathrm{i} t H } \chi( H ) u \Big \rangle \le \mathrm{C} t^{\frac{\alpha}{1+\mu}} \big ( \| \mathrm{d} \Gamma( |p_2|^{-\alpha} )^{\frac12} u \|^2 + \| N_{\mathrm{neut}}^{\frac12} u \|^2 + \| u \|^2 \big ) , \label{eq:estimp2-1} \end{equation} and for all $u \in \mathscr{D}( \mathrm{d}\Gamma( |p_2|^{-\alpha} ) )$, \begin{equation} \Big \| \mathrm{d} \Gamma ( | p_2 |^{-\alpha} ) e^{ - \mathrm{i} t H } \chi( H ) u \Big \| \le \mathrm{C} t^{\frac{\alpha}{1+\mu}} \big ( \| \mathrm{d} \Gamma( |p_2|^{-\alpha} ) u \| + \| N_{\mathrm{neut}} u \| + \| u \| \big ) . \label{eq:estimp2-1_a} \end{equation} \end{lem} \begin{proof} To prove \eqref{eq:estimp2-1}, we consider a function $f_0 \in \mathrm{C}^\infty( [0,\infty) ; \mathbb{R} )$ such that $f_0$ is decreasing, $f_0(r) = 1$ on $[ 0 , 1 ]$ and $f_0(r) = 0$ on $[ 2 , \infty )$. Let $f_\infty = 1 - f$. For $\nu>0$, we decompose \begin{equation} \mathrm{d} \Gamma ( | p_2 |^{- \alpha} ) = \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_0( t^{\nu} | p_2 | ) \big) + \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_\infty (t^{\nu} | p_2 | ) \big) , \label{eq:decompp-1} \end{equation} and insert this into the left-hand-side of \eqref{eq:estimp2-1}. Since $f_\infty$ is supported in $[1,\infty)$, the second term can be estimated as \begin{align} \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_\infty (t^{\nu} | p_2 | ) \big) e^{ - \mathrm{i} t H } \chi ( H ) u \Big \rangle &\leq t^{\alpha\nu} \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{d} \Gamma \big( f_\infty ( t^{\nu} | p_2 | ) \big) e^{ - \mathrm{i} t H } \chi ( H ) u \Big \rangle \notag \\ &\leq t^{\alpha\nu} \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , N_{\mathrm{neut}} e^{ - \mathrm{i} t H } \chi ( H ) u \Big \rangle \notag \\ &\lesssim t^{\alpha\nu} \big( \| N_{\mathrm{neut}}^{\frac12} u \|^2 + \| u \|^2 \big) . \label{eq:1st_p2-1} \end{align} The second equality comes from the facts that $N_{\mathrm{neut}}-N_{\mathrm{lept}}$ commutes with $H$ and that $N_{\mathrm{lept}}$ is relatively $H$-bounded. To estimate the evolution of the first term in \eqref{eq:decompp-1}, we differentiate \begin{align*} & \partial_t \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) \big) e^{ - \mathrm{i} t H } \chi ( H ) u \Big \rangle \\ &= \nu t^{\nu-1} \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{d} \Gamma \big( | p_2 |^{1-\alpha} f'_0 (t^{\nu} | p_2 | ) \big) e^{ - \mathrm{i} t H } \chi ( H ) u \Big \rangle \\ &\quad + \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \big [ H , \mathrm{i} \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) \big) \big ] e^{ - \mathrm{i} t H } \chi ( H ) u \Big \rangle . \end{align*} Since $f_0' \le 0$ and since $H_0$ commutes with $\mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) \big)$, this implies that \begin{align} & \partial_t \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) \big) e^{ - \mathrm{i} t H } \chi ( H ) u \Big \rangle \notag \\ &\le \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \big [ H_I(G) , \mathrm{i} \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) \big) \big ] e^{ - \mathrm{i} t H } \chi ( H ) u \Big \rangle . \label{eq:estimp2-1_1} \end{align} Similarly as in \eqref{commut_1}--\eqref{commut_4}, a direct computation gives \begin{align*} & \big [H_I^{(1)}(G) , \mathrm{i} \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) \big ) \big ] = - H_I^{(1)}( \mathrm{i} | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) G ), \\ & \big [H_I^{(2)}(G) , \mathrm{i} \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) \big ) \big ] = - H_I^{(2)}( \mathrm{i} | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) G ), \\ & \big [H_I^{(3)}(G) , \mathrm{i} \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) \big ) \big ] = H_I^{(3)}( \mathrm{i} | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) G ), \\ & \big [H_I^{(4)}(G) , \mathrm{i} \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) \big ) \big ] = H_I^{(4)}( \mathrm{i} | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) G ). \end{align*} By Lemma \ref{lemSCW_1}, the support of $f_0$ and the assumption that $|p_2|^{-1-\mu} G \in L^2$, we obtain that \begin{align} \big \| \big [ H_I(G) , \mathrm{i} \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) \big) \big ] (N_{\mathrm{lept}} + N_W + 1)^{-1} \big \| & \lesssim \| (1 + | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) ) G \|_2 \notag \\ & \lesssim \| (1 + | p_2 |^{1+\mu-\alpha} f_0 (t^{\nu} | p_2 | ) |p_2|^{-1-\mu}) G \|_2 \notag \\ & \lesssim t^{-(1+\mu-\alpha) \nu } \| (1 + |p_2|^{-(1+\mu)} ) G \|_2. \label{eq:ruqb1} \end{align} Since $N_{\mathrm{lept}} + N_W$ is relatively $H$-bounded, integrating \eqref{eq:estimp2-1_1} shows that \begin{align} & \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) \big) e^{ - \mathrm{i} t H } \chi ( H ) u \Big \rangle \lesssim t^{1-(1+\mu-\alpha)\nu} \big ( \| \mathrm{d} \Gamma( |p_2|^{-\alpha} )^{\frac12} u \|^2 + \| u \|^2 \big ). \label{eq:1st_p2-2} \end{align} Equations \eqref{eq:1st_p2-1} and \eqref{eq:1st_p2-2} yield \begin{align*} & \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} \big) e^{ - \mathrm{i} t H } \chi ( H ) u \Big \rangle \lesssim \max( t^{1-(1+\mu-\alpha)\nu} , t^{\alpha \nu} ) \big ( \| \mathrm{d} \Gamma( |p_2|^{-1} )^{\frac12} u \|^2 + \| N_{\mathrm{neut}}^{\frac12} u \|^2 + \| u \|^2 \big ). \end{align*} Choosing $\nu =( 1+\mu)^{-1}$ concludes the proof of \eqref{eq:estimp2-1}. To establish \eqref{eq:estimp2-1_a}, we modify the proof as follows. On one hand, we have that \begin{equation} \Big\| \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_\infty (t^{\nu} | p_2 | ) \big) e^{ - \mathrm{i} t H } \chi ( H ) u \Big\| \leq t^{-\nu\alpha} \Big\| N_{\mathrm{neut}} e^{ - \mathrm{i} t H } \chi ( H ) u \Big\| \lesssim t^{-\nu\alpha} \Big( \big\| N_{\mathrm{neut}} u \big\| +\| u \|\Big) , \label{eq:jske1} \end{equation} and on the other hand \begin{align} & \Big\| \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) \big) e^{ - \mathrm{i} t H } \chi ( H ) u \Big\| \notag \\ &=\Big\| e^{\mathrm{i}t H } \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_0 (t^{\nu} | p_2 | ) \big) e^{ - \mathrm{i} t H } \chi ( H ) u \Big\| \notag \\ &\le \Big\| \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_0 ( | p_2 | ) \big) e^{ - \mathrm{i} H } \chi ( H ) u \Big\| + \int_1^t \Big\| e^{\mathrm{i}s H } \Big [ \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} f_0 (s^{\nu} | p_2 | ) \big) , \mathrm{i} H \Big ] e^{ - \mathrm{i} s H } \chi ( H ) u \Big\| ds \notag \\ & \lesssim \big\| \mathrm{d} \Gamma \big( | p_2 |^{-\alpha} \big) u \big\|+\|u\| + \int_1^t s^{-(1+\mu-\alpha) \nu } \big \| (1 + |p_2|^{-(1+\mu)} ) G \big \|_2 \| u \| ds , \label{eq:jske2} \end{align} where we used \eqref{eq:ruqb1} and the fact that $N_{\mathrm{lept}} + N_W$ is relatively $H$-bounded in the last inequality. Combining \eqref{eq:jske1} and \eqref{eq:jske2}, we deduce \eqref{eq:estimp2-1_a} similarly as above. \end{proof} Now we are ready to prove a first version of the propagation estimates in the massless case. The general strategy is the same as in Theorem \ref{th:propag_massive}, with some important technical modifications. In particular, we have to use Lemma \ref{lem:evolp2-1}, the particular form of the Mourre estimate stated in Theorem \ref{MourreI_3} and the fact that $N_{\mathrm{lept}} - N_{\mathrm{neut}}$ commutes with $H$. Recall that the notation $\mathds{1}_{[R,R']}(|x|)$ has been introduced in \eqref{eq:1Rx}. \begin{Th}\label{th:propag_massless} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ vanish and consider the Hamiltonian $H = H_0 + g ( H_I^{(1)} + H_I^{(2)} )$ with $H_I^{(1)}$ and $H_I^{(2)}$ given by \eqref{interaction_term}. Assume that \begin{equation*} G \in L^2, \quad a_{(i),\cdot} G \in L^2, \quad |p_3|^{-1} a_{(i),\cdot} G \in L^2, \quad i = 1,2,3 , \end{equation*} and that \begin{equation*} G \in \mathbb{H}^{1+\mu} \text{ for some } \mu > 0. \end{equation*} \begin{itemize} \item[(i)] Let $\chi \in \mathrm{C}_0^{\infty}(\mathbb{R})$, $R=(R_1,\dots,R_{14})$ and $R'=(R'_1,\dots,R'_{14})$ be such that $R'_{i}>R_i>1$. There exists $\mathrm{C} > 0$ such that, for all $u \in \mathscr{D}(N_{\mathrm{neut}}^{\frac12})\cap\mathscr{D}( \mathrm{d}\Gamma( |p_2|^{-1} )^{\frac12} )$, \begin{align*} & \int^{\infty}_{1} \Big \| \mathrm{d} \Gamma \Big ( \mathds{1}_{[R,R']} \Big (\frac{|x|}{t} \Big ) \Big)^{\frac{1}{2}}\chi(H) e^{-\mathrm{i}tH} u \Big \|^2 \frac{dt}{t} \leq \mathrm{C} \big ( \| N_{\mathrm{neut}}^{\frac12} u \|^2 + \| \mathrm{d}\Gamma( |p_2|^{-1} )^{\frac12} u \|^2 + \|u\|^{2} \big ). \end{align*} \item[(ii)] Let $0<\mathrm{v}_0<\mathrm{v}_1$ and $\chi \in \mathrm{C}_0^{\infty}(\mathbb{R})$. There exists $\mathrm{C}>0$ such that, for all $u \in \mathscr{D}(N_{\mathrm{neut}}^{\frac12})\cap\mathscr{D}( \mathrm{d}\Gamma( |p_2|^{-1} )^{\frac12} )$, \begin{align*} & \int^{\infty}_{1} \Big \| \mathrm{d}\Gamma \Big( \Big\langle \big( \frac{x}{t} - \nabla \omega \big ) , \mathds{1}_{ [ \mathrm{v}_0 , \mathrm{v}_1 ] }( \frac{x}{t} ) \big( \frac{x}{t} - \nabla \omega \big ) \Big\rangle \Big)^{\frac12} \chi(H)e^{-\mathrm{i} t H} u \Big\|^{2} \frac{dt}{t} \\ & \le \mathrm{C} \big ( \| N_{\mathrm{neut}}^{\frac12} u \|^2 + \| \mathrm{d}\Gamma( |p_2|^{-1} )^{\frac12} u \|^2 + \|u\|^{2} \big ). \end{align*} \item[(iii)] Let $0< \mathrm{v}_0<\mathrm{v}_1$, $J \in \mathrm{C}_0^{\infty}( \{ x \in \mathbb{R}^3 , \mathrm{v}_0 < |x| < \mathrm{v}_1\})$ and $\chi \in \mathrm{C}_0^{\infty}(\mathbb{R})$. There exists $\mathrm{C}>0$ such that, for all $\ell \in \{ 1,2,3\} $ and $u \in \mathscr{D}(N_{\mathrm{neut}}^{\frac12})\cap\mathscr{D}( \mathrm{d}\Gamma( |p_2|^{-1} )^{\frac12} )$, \begin{align*} & \int^{\infty}_{1} \Big \| \mathrm{d}\Gamma \Big( \Big| J\Big(\frac{x}{t}\Big) \Big( \frac{x^{(\ell)}}{t} - \partial_{(\ell)} \omega \Big) + \mathrm{h.c }\Big|\Big)^{\frac{1}{2}}\chi(H)e^{-\mathrm{i} t H} u \Big\|^{2} \frac{dt}{t} \\ & \le \mathrm{C} \big ( \| N_{\mathrm{neut}}^{\frac12} u \|^2 + \| \mathrm{d}\Gamma( |p_2|^{-1} )^{\frac12} u \|^2 + \|u\|^{2} \big ). \end{align*} \item[(iv)] There exists $g_0>0$ such that, for all $|g|\le g_0$, the following holds: let $\chi \in \mathrm{C}_{0}^{\infty}(\mathbb{R})$ be supported in $\mathbb{R}\backslash(\tau \cup \sigma_{\mathrm{pp}}(H))$. There exist $\delta>0$ and $\mathrm{C} > 0$ such that \begin{equation*} \int^{\infty}_{1} \Big\| \Gamma \Big ( \mathds{1}_{[0,\delta]} \Big(\frac{|x|}{t}\Big)\Big) \chi(H) e^{-\mathrm{i} t H} u \Big\|^2 \frac{dt}{t} \le \mathrm{C} \big\| ( \mathrm{d}\Gamma(|p_2|^{-1}) + N_{\mathrm{neut}} + 1 )^{\frac12} (N_{\mathrm{neut}}+1)^{\frac32} u \big\|^2 , \end{equation*} for all $u \in \mathscr{D}( ( \mathrm{d}\Gamma(|p_2|^{-1}) + N_{\mathrm{neut}} + 1 )^{\frac12} (N_{\mathrm{neut}}+1)^{\frac32} )$. \end{itemize} \end{Th} \begin{proof} We prove (i) and (iv), underlying the differences with \cite{DeGe99_01}. The proofs of (ii) and (iii) can be deduced by adapting \cite[Section 6]{DeGe99_01}, using furthermore arguments similar to those used to prove (i). (i) Let $F_i \in \mathrm{C}^\infty ( \mathbb{R} ; \mathbb{R} )$ be a non-decreasing function such that $F_i=0$ near $0$, $F_i=\mathrm{const}$ near $+\infty$ and $F'_i \ge \mathds{1}_{[R_i,R'_i]}$. As in the proof of Theorem \ref{th:propag_massive}, \eqref{eq:hdt1} holds and therefore it suffices to prove that \begin{equation*} \int^{\infty}_{1} \Big\|\mathrm{d}\Gamma_i\Big(\mathds{1}_{[R_i,R'_i]}\Big(\frac{ |x_i| }{t}\Big)\Big)^{\frac{1}{2}}\chi(H) e^{-\mathrm{i}tH} u \Big\|^2 \frac{dt}{t} \lesssim \big ( \| N_{\mathrm{neut}}^{\frac12} u \|^2 + \| \mathrm{d}\Gamma( |p_2|^{-1} )^{\frac12} u \|^2 + \|u\|^{2} \big ) \end{equation*} for all $u \in \mathscr{D}(N_{\mathrm{neut}}^{\frac12})\cap\mathscr{D}( \mathrm{d}\Gamma( |p_2|^{-1} )^{\frac12} )$ and $i \in \{1,\dots,14\}$, where $\mathrm{d}\Gamma_i$ is defined in \eqref{eq:defdgammaj}. Let \begin{align} f_i(t) := \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{d}\Gamma_{i} \Big( F_i \Big(\frac{ |x_i| }{t}\Big)\Big) e^{ - \mathrm{i} t H } \chi(H) u \Big \rangle . \label{eq:def_fj} \end{align} Note that, since $\mathrm{d}\Gamma_{i} ( F_i (\frac{ |x_i| }{t})) \le \mathrm{C} N$, for some positive constant $\mathrm{C}$, since $N_{\mathrm{lept}}$ and $N_W$ are relatively $H$-bounded and since $N_{\mathrm{lept}} - N_{\mathrm{neut}}$ commutes with $H$, we can write \begin{align*} f_i(t) & \le \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{C} N e^{ - \mathrm{i} t H } \chi(H) u \Big \rangle \\ & = \mathrm{C} \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \big ( 2 N_{\mathrm{lept}} + N_W + N_{\mathrm{neut}} - N_{\mathrm{lept}} \big ) e^{ - \mathrm{i} t H } \chi(H) u \Big \rangle \\ & \lesssim \| u \|^2 + \Big \langle \big ( N_{\mathrm{neut}} - N_{\mathrm{lept}} \big )^{\frac12} u , \chi(H)^2 \big ( N_{\mathrm{neut}} - N_{\mathrm{lept}} \big )^{\frac12} u \Big \rangle \\ & \lesssim \| N_{\mathrm{neut}}^{\frac12} u \|^2 + \| u \|^2. \end{align*} Hence $f_i(t)$ is uniformly bounded. Differentiating $f_i$ gives \begin{align} \partial_t f_i(t) &= \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathbf{D} \mathrm{d}\Gamma_{i} \Big( F_i \Big( \frac{ |x_i| }{t}\Big)\Big) e^{ - \mathrm{i} t H } \chi(H) u \Big \rangle \notag \\ &= \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{d}\Gamma_{i} \Big( \mathbf{d}_0 F_i \Big(\frac{ |x_i| }{t}\Big)\Big) e^{ - \mathrm{i} t H } \chi(H) u \Big \rangle \notag \\ &\quad + \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \Big [ H_I^{(1)}(G) + H_I^{(2)}(G) , \mathrm{i} \mathrm{d}\Gamma_{i} \Big( F_i \Big(\frac{|x_i| }{t}\Big) \Big) \Big ] e^{ - \mathrm{i} t H } \chi(H) u \Big \rangle , \label{eq:propagestim1} \end{align} where $\mathbf{D}$ and $\mathbf{d}_0$ denotes the Heisenberg derivatives defined at the beginning of Section \ref{sec:propag}. Using that $F'_i \ge \mathds{1}_{[R_i,R'_i]}$ and $R_i>1$, a commutator expansion at second order (see, e.g., \cite{DeGe97_01} and \cite[Lemma 5.2]{BoFaSi12_01}) gives \begin{align*} \mathbf{d}_0 F_i \Big(\frac{|x_i|}{t}\Big) \le \frac{-\mathrm{C}_0}{t}\mathds{1}_{[R_i,R'_i]} \Big(\frac{|x_i|}{t}\Big) + \mathcal{O}(t^{-2}) , \end{align*} if $i\in\{1,2,3,4,7,8,11,12\}$ (corresponding to the label of a massive particle), and \begin{align*} \mathbf{d}_0 F_i \Big(\frac{|x_i|}{t}\Big) \le \frac{-\mathrm{C}_0}{t} \mathds{1}_{[R_i,R'_i]} \Big(\frac{|x_i|}{t}\Big) + \mathcal{O}(t^{-2}) |p_2|^{-1} , \end{align*} if $i \in \{5,6,9,10,13,14\}$ (corresponding to the label of a neutrino). Here $\mathrm{C}_0>0$. In the first case, using as above that $N_{\mathrm{lept}}-N_{\mathrm{neut}}$ commutes with $H$, this gives \begin{align} & \Big\langle e^{-\mathrm{i}tH}\chi(H)u,\mathrm{d}\Gamma_{i}\Big(\mathbf{d}_0F_i\Big(\frac{|x_i|}{t}\Big)\Big)e^{-\mathrm{i}tH}\chi(H)u\Big\rangle \notag \\ &\le - \frac{ \mathrm{C}_0 }{ t } \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{d}\Gamma_{i} \Big ( \mathds{1}_{[R_i,R'_i]} \Big(\frac{|x_i|}{t}\Big) \Big ) e^{ - \mathrm{i} t H } \chi(H) u \Big \rangle + \mathcal{O}(t^{-2}) \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , N e^{ - \mathrm{i} t H } \chi(H) u \Big \rangle \notag \\ &\le - \frac{ \mathrm{C}_0 }{ t } \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{d}\Gamma_{i} \Big ( \mathds{1}_{[R_i,R'_i]} \Big(\frac{|x_i|}{t}\Big) \Big ) e^{ - \mathrm{i} t H } \chi(H) u \Big \rangle + \mathcal{O}(t^{-2}) \big ( \| N_{\mathrm{neut}}^{\frac12} u \|^2 + \| u \|^2 \big ) , \label{eq:propagestim2} \end{align} if $i\in\{1,2,3,4,7,8,11,12\}$. In the second case (if $i$ corresponds to the label of a neutrino), we use in addition Lemma \ref{lem:evolp2-1}, yielding \begin{align} & \Big\langle e^{-\mathrm{i}tH}\chi(H)u,\mathrm{d}\Gamma_{i}\Big(\mathbf{d}_0F_i\Big(\frac{|x_i|}{t}\Big)\Big)e^{-\mathrm{i}tH}\chi(H)u\Big\rangle \notag \\ &\le - \frac{ \mathrm{C}_0 }{ t } \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{d}\Gamma_{i} \Big ( \mathds{1}_{[R_i,R'_i]} \Big(\frac{|x_i|}{t}\Big) \Big ) e^{ - \mathrm{i} t H } \chi(H) u \Big \rangle + \mathcal{O}(t^{-2}) \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{d}\Gamma(|p_2|^{-1}) e^{ - \mathrm{i} t H } \chi(H) u \Big \rangle \notag \\ &\le - \frac{ \mathrm{C}_0 }{ t } \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{d}\Gamma_{i} \Big ( \mathds{1}_{[R_i,R'_i]} \Big(\frac{|x_i|}{t}\Big) \Big ) e^{ - \mathrm{i} t H } \chi(H) u \Big \rangle + \mathcal{O}(t^{-2 + \frac{2}{2+\mu} }) \big ( \| \mathrm{d} \Gamma( |p_2|^{-1} )^{\frac12} u \|^2 + \| u \|^2 \big ) . \label{eq:propagestim3} \end{align} Note that Hardy's inequality implies that, if $G \in \mathbb{H}^{1+\mu}$ for some $\mu>0$, then $|p_2|^{-1-\mu}G \in L^2$ (if $\mu<1/2$), and hence Lemma \ref{lem:evolp2-1} can indeed by applied. The commutators $ [ H_I^{(j)}(G) , \mathrm{d}\Gamma_{i} ( F_i (\frac{|x_i|}{t}) ) ]$, $j=1,2$, are given by expressions similar to \eqref{commut_11}--\eqref{commut_21}, with the operator $F_i (\frac{|x_i|}{t})$ instead of $a_i$. Since $F_i$ vanishes near $0$ and $G \in \mathbb{H}^{1+\mu}$, we have that \begin{align*} \Big\|F_i\Big(\frac{|x_i|}{t}\Big)G\Big\|_2=\mathcal{O}(t^{-1-\mu}). \end{align*} Therefore we deduce from Lemma \ref{lemSCW_1} that \begin{align} & \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \Big [ H_I^{(1)}(G) + H_I^{(2)}(G) , \mathrm{i} \mathrm{d}\Gamma_{i} \Big( F_i \Big(\frac{|x_i|}{t}\Big) \Big) \Big ] e^{ - \mathrm{i} t H } \chi(H) u \Big \rangle \notag \\ & \le \mathrm{C} t^{-1-\mu} \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , ( N_W + N_{\mathrm{lept}} ) e^{ - \mathrm{i} t H } \chi(H) u \Big \rangle = \mathcal{O}( t^{-1-\mu} ) \| u \|^2. \label{eq:propagestim4} \end{align} Integrating \eqref{eq:propagestim1} over $[1,\infty)$, using that $t^{-2+\frac{2}{2+\mu}}$ and $t^{-1-\mu}$ are integrable, we obtain the statement of (i) by combining \eqref{eq:propagestim2}--\eqref{eq:propagestim3} and \eqref{eq:propagestim4}. \vspace{0,2cm} (iv) Let $\lambda \in \mathbb{R}\setminus\{ \tau \cup \sigma_{\mathrm{pp}}(H) \}$. Clearly, since $N_{\mathrm{lept}}-N_{\mathrm{neut}}$ commutes with $H$, it suffices to prove (iv) for all $u \in \mathrm{Ran}( \mathds{1}_{\{n\}}(N_{\mathrm{lept}}-N_{\mathrm{neut}}))$, $n\in\mathbb{Z}$. Hence we fix $n\in\mathbb{Z}$ and $u \in \mathrm{Ran}( \mathds{1}_{\{n\}}(N_{\mathrm{lept}}-N_{\mathrm{neut}}))$. Recall that, by Theorem \ref{MourreI_3}, there exist $\varepsilon>0$, $\mathrm{c}_0>0$ and $\mathrm{C}_\lambda > 0$ such that \begin{equation*} [H,\mathrm{i} A ] \geq \mathrm{c}_0 ( N_{\mathrm{neut}} + \mathds{1} ) - \mathrm{C}_\lambda ( \mathds{1} - \mathds{1}_{[\lambda-\varepsilon,\lambda+\varepsilon]}(H) ) ( 1 + H^2 )^{\frac12} . \end{equation*} For $\tilde\chi \in \mathrm{C}_0^\infty( ( \lambda-\varepsilon,\lambda+\varepsilon) )$, this implies that \begin{equation} [H,\mathrm{i} A ] \geq \mathrm{c}_0 ( N_{\mathrm{neut}} + \mathds{1} ) - \mathrm{C}_\lambda ( \mathds{1} - \tilde{\chi}^2(H) ) ( 1 + H^2 )^{\frac12} . \label{eq:Mourre_for_propag} \end{equation} Let $\chi \in \mathrm{C}_0^\infty( ( \lambda-\varepsilon,\lambda+\varepsilon) )$ be such that $\chi \tilde\chi = \chi$. Let $\delta > 0 $ and $q_i$, $i \in \{1, \dots, 14\}$ be functions in $\mathrm{C}_0^{\infty} (\{ x \in \mathbb{R}^3 , |x| \le 2 \delta \})$ such that $0 \le q_i \le 1$, $q_i=1$ on $\{ x \in \mathbb{R}^3 , |x| \leq \delta \}$ and let $q^t=(q_1(\frac{\mathrm{i}\nabla}{t}),\dots,q_{14}(\frac{\mathrm{i}\nabla}{t}))$. Let \begin{equation*} h(t):=\Big\langle e^{-\mathrm{i}tH}\chi(H)u,\Gamma(q^t)\frac{A}{t}\Gamma(q^t)e^{-\mathrm{i}tH}\chi(H)u\Big\rangle. \end{equation*} Since $\nabla \omega^{(i)}_l$, $i=1,2$, $l=1,2,3$ and $\nabla \omega^{(3)}$ are bounded, it is not difficult to observe that \begin{equation} \label{Step1} \pm \Gamma(q^t) \frac{A}{t} \Gamma(q^t) \le \mathrm{C} \delta (N+1) \quad \text{ and } \quad \big \| \frac{A}{t} \Gamma(q^t) ( N+1 )^{-1}\big\|\le \mathrm{C} \delta. \end{equation} Here it should be noticed that $\Gamma(q^t)$ maps $\mathscr{H}$ to $\mathscr{D}(A)$. Indeed a vector in $\mathrm{Ran}(\Gamma(q^t))$ belongs to the Sobolev space $\mathbb{H}^s(\mathbb{R}^3)$, for any $s>0$, as a function of any of the momentum variable $p_i$. It is then regular in a neighborhood of zero and using, in particular, \eqref{eq:domain_conjugate}, it is not difficult to verify that $\mathrm{Ran}( \Gamma(q^t) ) \subset \mathscr{D}(A)$. In the second estimate of \eqref{Step1}, Hardy's inequality in $\mathbb{R}^3$, $\| |p_2|^{-1} u \|_2 \lesssim \| \nabla_{p_2} u \|_2$, is used. From \eqref{Step1} and using as before that $N_{\mathrm{lept}} - N_{\mathrm{neut}}$ commutes with $H$, we obtain that \begin{align} | h(t) | \le \Big\langle e^{-\mathrm{i}tH}\chi(H)u,\Gamma(q^t)\frac{A}{t}\Gamma(q^t)e^{-\mathrm{i}tH}\chi(H)u\Big\rangle \lesssim \|N_{\mathrm{neut}}^{\frac12}u\|^2 + \|u\|^2. \label{eq:estimh(t)} \end{align} In particular, $h(t)$ is uniformly bounded. Furthermore, one can compute \begin{align} \partial_t h(t) &= \Big\langle e^{-\mathrm{i}tH} u, \Big ( \chi(H) \mathrm{d}\Gamma(q^t, \mathbf{d}_0 q^t) \frac{A}{t}\Gamma(q^t) \chi(H) + \mathrm{h.c.} \Big ) e^{-\mathrm{i}tH}u\Big\rangle \notag \\ & \quad + \Big\langle e^{-\mathrm{i}tH}u, \Big ( \chi(H) [ H_I^{(1)} (G) + H_I^{(2)} (G) , \mathrm{i} \Gamma(q^t) ] \frac{A}{t}\Gamma(q^t) \chi(H) + \mathrm{h.c.} \Big ) e^{-\mathrm{i}tH}u \Big\rangle \notag \\ & \quad + t^{-1} \Big\langle e^{-\mathrm{i}tH}u, \chi(H) \Gamma(q^t) [ H , \mathrm{i}A ] \Gamma(q^t) \chi(H) e^{-\mathrm{i}tH}u\Big\rangle \notag \\ & \quad - t^{-1} \Big\langle e^{-\mathrm{i}tH}u, \chi(H) \Gamma(q^t)\frac{A}{t}\Gamma(q^t) \chi(H) e^{-\mathrm{i}tH}u\Big\rangle \notag \\ & =: R_1(t) + R_2(t) + R_3(t) + R_4(t). \label{eq:R1-R4} \end{align} Observe that $A$ being symmetric, $A^*$ is an extension of $A$ so that the second terms in $R_2(t)$ and $R_3(t)$ are indeed the hermitian conjugates of the first ones. In what follows we consider each term $R_\sharp(t)$ separately. We begin with $R_2(t)$. From the assumption that $G\in\mathbb{H}^{1+\mu}$, the commutation relations of Appendix \ref{sec:standard_def} and Lemma \ref{lemSCW_1}, one can show that \begin{align*} \big \| [H_I^{(1)} (G) + H_I^{(2)} (G) , \Gamma(q^t)] (N_{\mathrm{lept}}+N_W)^{-1} \big \| = \mathcal{O}(t^{-1-\mu}) . \end{align*} Together with \eqref{Step1} this implies that \begin{equation*} |R_2(t)|\lesssim \mathcal{O}(t^{-1-\mu})\big\|(N+1)\chi(H)e^{-\mathrm{i}tH}u\big\|^2, \end{equation*} since $N_{\mathrm{lept}}+N_W$ is relatively $H$-bounded. Since $N_{\mathrm{lept}}-N_{\mathrm{neut}}$ commutes with $H$, we deduce as before that \begin{equation} |R_2(t)|\lesssim \mathcal{O}(t^{-1-\mu})\|(N_{\mathrm{neut}}+1)u\|^2. \label{eq:estim_R2} \end{equation} To evaluate $R_1(t)$, we compute, by means of a commutator expansion (see \cite{DeGe97_01} and \cite[Section 5]{BoFaSi12_01}), \begin{align} \mathbf{d}_0 q_i^t(x_i) = -\frac{1}{2 t} \Big \langle \frac{x_i}{t} - \nabla\omega_i , \nabla q_i \big( \frac{x_i}{t} \big) \Big \rangle + \mathrm{h.c.} + \mathcal{O}(t^{-2}) , \label{eq:comput_rem_0} \end{align} if $i\in\{1,2,3,4,7,8,11,12\}$ (corresponding to the label of a massive particle), and \begin{align} \mathbf{d}_0 q_i^t(x_i) = -\frac{1}{2 t} \Big \langle \frac{x_i}{t} - \nabla\omega_i , \nabla q_i \big( \frac{x_i}{t} \big) \Big \rangle + \mathrm{h.c.} + \mathcal{O}(t^{-\frac32}) |p_2|^{-\frac12} , \label{eq:comput_rem} \end{align} if $i \in \{5,6,9,10,13,14\}$ (corresponding to the label of a neutrino). Here it should be noted that that the remainder term in \eqref{eq:comput_rem} can be computed to be of order $|p_2|^{-1+\gamma} \mathcal{O}(t^{-2+\gamma})$ for any $0\le\gamma\le1$ (see \cite[Lemma 5.2]{BoFaSi12_01}. Here we choose $\gamma=\frac12$ for convenience, see below. Let $g^t_i := -\frac{1}{2} \langle \frac{x_i}{t} - \nabla\omega_i , \nabla q_i \big( \frac{x_i}{t} \big) \rangle + \mathrm{h.c.}$ and let $r^t_i$ be the remainder in \eqref{eq:comput_rem_0} or \eqref{eq:comput_rem}. For $i\in\{1,2,3,4,7,8,11,12\}$, by \eqref{Step1}, we deduce that \begin{align} \Big|\Big\langle e^{-\mathrm{i}tH} u, \Big ( \chi(H) \mathrm{d}\Gamma(q^t_i, r^t_i) \frac{A}{t}\Gamma(q^t) \chi(H) + \mathrm{h.c.} \Big ) e^{-\mathrm{i}tH}u\Big\rangle\Big| & \lesssim \mathcal{O}(t^{-2})\big\|N\chi(H)e^{-\mathrm{i}tH}u\big\|^2 \notag \\ & \lesssim \mathcal{O}(t^{-2}) \|(N_{\mathrm{neut}}+1)u\|^2 , \label{eq:dd0} \end{align} where we used again that $N_{\mathrm{lept}}-N_{\mathrm{neut}}$ commutes with $H$ in the second inequality. The case $i \in \{5,6,9,10,13,14\}$ corresponding to massless particles is more difficult. Using \eqref{eq:comput_rem} and Lemma \ref{estimationN2}, we write \begin{align*} & \Big|\Big\langle e^{-\mathrm{i}tH} u, \Big ( \chi(H) \mathrm{d}\Gamma(q^t_i, r^t_i) \frac{A}{t}\Gamma(q^t) \chi(H) + \mathrm{h.c.} \Big ) e^{-\mathrm{i}tH}u\Big\rangle\Big| \\ & \Big|\Big\langle e^{-\mathrm{i}tH} u, \Big ( \chi(H) \mathrm{d}\Gamma(q^t_i , r^t_i) (N+1)^{-\frac12}(N+1)^{\frac12}\frac{A}{t}\Gamma(q^t) \chi(H) + \mathrm{h.c.} \Big ) e^{-\mathrm{i}tH}u\Big\rangle\Big| \\ & \lesssim \mathcal{O}(t^{-\frac32}) \Big \| \mathrm{d}\Gamma(|p_2|^{-1})^{\frac12} \chi(H) e^{-\mathrm{i}tH} u \Big \| \big \| (N+1)^{\frac32} \chi(H) e^{-\mathrm{i}tH} u \big \| . \end{align*} Applying Lemma \ref{lem:evolp2-1}, and using that $N_{\mathrm{lept}}-N_{\mathrm{neut}}$ commutes with $H$, it thus follows that \begin{align} & \Big|\Big\langle e^{-\mathrm{i}tH} u, \Big ( \chi(H) \mathrm{d}\Gamma(q^t_i , r^t_i) \frac{A}{t}\Gamma(q^t) \chi(H) + \mathrm{h.c.} \Big ) e^{-\mathrm{i}tH}u\Big\rangle\Big| \notag \\ & \lesssim \mathcal{O}(t^{-\frac32+\frac{1}{2+\mu}}) \big ( \| \mathrm{d}\Gamma(|p_2|^{-1})^{\frac12} u \| + \| N_{\mathrm{neut}}^{\frac12}u \| + \|u\| \big ) \big ( \big \| (N+1)^{\frac32} u \big \| + \| u \| \big ). \label{eq:dd1} \end{align} Summing over $i$, we obtain from \eqref{eq:dd0} and \eqref{eq:dd1} that \begin{align} & \Big|\Big\langle e^{-\mathrm{i}tH} u, \Big ( \chi(H) \mathrm{d}\Gamma(q^t , r^t) \frac{A}{t}\Gamma(q^t) \chi(H) + \mathrm{h.c.} \Big ) e^{-\mathrm{i}tH}u\Big\rangle\Big| \notag \notag \\ & \lesssim \mathcal{O}(t^{-2})\|(N_{\mathrm{neut}}+1)u\|^2 +\mathcal{O}(t^{-\frac32+\frac{1}{2+\mu}}) \big ( \| \mathrm{d}\Gamma(|p_2|^{-1})^{\frac12} u \| + \|u\| \big ) \big ( \big \| (N+1)^{\frac32} u \big \| + \| u \| \big ). \label{eq:estim_rt} \end{align} To estimate the term corresponding to $\frac1t\mathrm{d}\Gamma(q^t, g^t)$, we proceed similarly but use (ii) instead of Lemma \ref{lem:evolp2-1}. We introduce $\tilde{q}^t$, defined as $q^t$, such that $\tilde{q}^t q^t = q^t$ and hence $\Gamma(q^t)=\Gamma(\tilde{q}^t)\Gamma(q^t)$. This yields \begin{align*} & \frac{1}{t} \Big|\Big\langle e^{-\mathrm{i}tH} u, \Big ( \chi(H) \mathrm{d}\Gamma(q^t_i , g^t_i ) \frac{A}{t}\Gamma(q^t) \chi(H) + \mathrm{h.c.} \Big ) e^{-\mathrm{i}tH}u\Big\rangle\Big| \\ & = \frac{1}{t} \Big|\Big\langle e^{-\mathrm{i}tH} u, \Big ( \chi(H) \mathrm{d}\Gamma(q^t_i , g^t_i) (N+1)^{-\frac12}(N+1)^{\frac12}\frac{A}{t}\Gamma(\tilde{q}^t)\Gamma(q^t) \chi(H) + \mathrm{h.c.} \Big ) e^{-\mathrm{i}tH}u\Big\rangle\Big| \\ & \lesssim \frac{1}{t} \Big \| \mathrm{d}\Gamma((g^t_i)^*g^t_i)^{\frac12} \chi(H) e^{-\mathrm{i}tH} u \Big \| \big \| (N+1)^{\frac32} \Gamma(q^t) \chi(H) e^{-\mathrm{i}tH} u \big \| \\ & \lesssim \frac{\alpha(|n|+1)^3}{t} \Big \| \mathrm{d}\Gamma((g^t_i)^*g^t_i)^{\frac12} \chi(H) e^{-\mathrm{i}tH} u \Big \|^2 + \frac{1}{\alpha(|n|+1)^3t} \big \| (N+1)^{\frac32} \Gamma(q^t) \chi(H) e^{-\mathrm{i}tH} u \big \|^2 , \end{align*} where we recall that $n\in\mathbb{Z}$ has been fixed such that $u \in \mathrm{Ran}( \mathds{1}_{\{n\}}(N_{\mathrm{lept}}-N_{\mathrm{neut}}))$. The parameter $\alpha>0$ will be determined later. Summing over $i$, since $N_{\mathrm{lept}}-N_{\mathrm{neut}}$ commutes with $H$, this implies that \begin{align} & \frac{1}{t} \Big|\Big\langle e^{-\mathrm{i}tH} u, \Big ( \chi(H) \mathrm{d}\Gamma(q^t, g^t) \frac{A}{t}\Gamma(q^t) \chi(H) + \mathrm{h.c.} \Big ) e^{-\mathrm{i}tH}u\Big\rangle\Big| \notag \\ & \lesssim \frac{\alpha}{t} \Big \| \mathrm{d}\Gamma((g^t)^*g^t)^{\frac12} \chi(H) e^{-\mathrm{i}tH} (|N_{\mathrm{neut}}-N_{\mathrm{lept}}|+1)^{\frac32} u \Big \|^2 + \frac{1}{\alpha t} \big \| \Gamma(q^t) \chi(H) e^{-\mathrm{i}tH} u \big \|^2 \notag \\ & \qquad + \frac{1}{\alpha t} \big \| (N_{\mathrm{lept}} + N_W +1)^{\frac32} \Gamma(q^t) \chi(H) e^{-\mathrm{i}tH} u \big \|^2 . \label{eq:dd2} \end{align} Remembering that $\tilde \chi \chi = \chi$, the last term of \eqref{eq:dd2} can be decomposed into \begin{align*} & \frac{1}{\alpha t} \big \| (N_{\mathrm{lept}} + N_W +1)^{\frac32} \Gamma(q^t) \chi(H) e^{-\mathrm{i}tH} u \big \|^2 \\ & = \frac{1}{\alpha t} \big \| (N_{\mathrm{lept}} + N_W +1)^{\frac32} \big [ \Gamma(q^t) , \tilde \chi(H) \big ] \chi ( H ) e^{-\mathrm{i}tH} u \big \|^2 + \frac{1}{\alpha t} \big \| (N_{\mathrm{lept}} + N_W +1)^{\frac32} \tilde \chi ( H ) \Gamma(q^t) \chi( H ) e^{-\mathrm{i}tH} u \big \|^2. \end{align*} Using the Helffer-Sj{\"o}strand functional calculus and similar arguments as before, one verifies that \begin{align} \big\|\big[\Gamma(q^t),\chi(H)\big](N_{\mathrm{neut}}+1)^{-1}\big\|=\mathcal{O}(t^{-1}). \label{eq:kha_1} \end{align} By (ii) in Lemma \ref{lemV1NumbestimatesV2}, it is not difficult to deduce from \eqref{eq:kha_1} and the previous equality that \begin{align} & \frac{1}{\alpha t} \big \| (N_{\mathrm{lept}} + N_W +1)^{\frac32} \Gamma(q^t) \chi(H) e^{-\mathrm{i}tH} u \big \|^2 \lesssim \frac{1}{\alpha t^2} \big \| ( N_{\mathrm{neut}}^2 + 1) u \big \|^2 + \frac{1}{\alpha t} \big \| \Gamma(q^t) \chi( H ) e^{-\mathrm{i}tH} u \big \|^2. \label{eq:lart_1} \end{align} Combining \eqref{eq:dd1}, \eqref{eq:dd2} and \eqref{eq:lart_1} gives \begin{align} |R_1(t)| & \lesssim \frac{\alpha}{t} \Big \| \mathrm{d}\Gamma((g^t)^*g^t)^{\frac12} \chi(H) e^{-\mathrm{i}tH} ( |N_{\mathrm{neut}}-N_{\mathrm{lept}}|+1)^{\frac32} u \Big \|^2 + \frac{1}{\alpha t} \big \| \Gamma(q^t) \chi(H) e^{-\mathrm{i}tH} u \big \|^2 \notag \\ &\quad + \mathcal{O}(t^{-2}) \|(N_{\mathrm{neut}}^2 + N_{\mathrm{neut}}+1)u\|^2 \notag \\ &\quad+ \mathcal{O}(t^{-\frac32+\frac{1}{2+\mu}}) \big ( \| \mathrm{d}\Gamma(|p_2|^{-1})^{\frac12} u \| +\| N_{\mathrm{neut}}^{\frac12}u \|+ \|u\| \big ) \big ( \big \| (N_{\mathrm{neut}}+1)^{\frac32} u \big \| + \| u \| \big ). \label{eq:estim_R1} \end{align} Next we consider $R_3(t)$. It follows from \eqref{eq:Mourre_for_propag} that \begin{align*} R_3(t) \ge \frac{1}{t} \Big\langle e^{-\mathrm{i}tH}u, \chi(H) \Gamma(q^t) \big ( \mathrm{c}_0 ( N_{\mathrm{neut}} + \mathds{1} ) - \mathrm{C}_\lambda ( \mathds{1} - \tilde{\chi}^2(H) ) \big ) \Gamma(q^t) \chi(H) e^{-\mathrm{i}tH}u\Big\rangle. \end{align*} Using \eqref{eq:kha_1} together with the previous equation, the facts that $N_{\mathrm{lept}}-N_{\mathrm{neut}}$ commutes with $H$, that $N_W$ and $N_{\mathrm{lept}}$ are relatively $H$-bounded and that $\chi \tilde\chi = \chi$, this gives \begin{align} R_3(t) \ge \frac{\mathrm{c}_0}{t} \Big\langle e^{-\mathrm{i}tH}u, \chi(H) \Gamma(q^t) ( N_{\mathrm{neut}} + \mathds{1} ) \Gamma(q^t) \chi(H) e^{-\mathrm{i}tH}u\Big\rangle + \mathcal{O}(t^{-2})\|(N_{\mathrm{neut}}+1)u\|^2. \label{eq:estim_R3} \end{align} To estimate $R_4(t)$, it suffices to apply \eqref{Step1} together with the fact that $N_W$ and $N_{\mathrm{lept}}$ are relatively $H$-bounded. This yields \begin{align} | R_4(t) | \lesssim \frac{\delta}{t} \Big\langle e^{-\mathrm{i}tH}u, \chi(H) \Gamma(q^t) (N_{\mathrm{neut}}+1 )\Gamma(q^t) \chi(H) e^{-\mathrm{i}tH}u\Big\rangle . \label{eq:estim_R4} \end{align} Putting together \eqref{eq:estim_R2}, \eqref{eq:estim_R1}, \eqref{eq:estim_R3} and \eqref{eq:estim_R4}, we finally arrive at \begin{align*} \partial_t h(t) \ge & \Big ( \frac{\mathrm{c}_0}{t} - \frac{1}{\alpha t} - \frac{ \delta }{ t } \Big ) \Big\langle e^{-\mathrm{i}tH}u, \chi(H) \Gamma(q^t) ( N_{\mathrm{neut}} + \mathds{1} ) \Gamma(q^t) \chi(H) e^{-\mathrm{i}tH}u\Big\rangle \\ & + \frac{\alpha}{t} \Big \| \mathrm{d}\Gamma(g_t^*g_t)^{\frac12} \chi(H) e^{-\mathrm{i}tH} (|N_{\mathrm{neut}}-N_{\mathrm{lept}}|+1)^{\frac32} u \Big \|^2 + \mathcal{O}(t^{-\min(2,1+\mu)})\|(N_{\mathrm{neut}}^2+N_{\mathrm{neut}}+1)u\|^2 \\ & + \mathcal{O}(t^{-\frac32+\frac{1}{2+\mu}}) \big ( \| \mathrm{d}\Gamma(|p_2|^{-1})^{\frac12} u \| + \|u\| \big ) \big ( \big \| (N_{\mathrm{neut}}+1)^{\frac32} u \big \| + \| u \| \big ). \end{align*} Fixing $\alpha$ large enough and $\delta$ small enough and integrating over $t$ from $1$ to $\infty$, we obtain from \eqref{eq:estimh(t)} and (ii) that \begin{align*} & \int_1^\infty \Big\langle e^{-\mathrm{i}tH}u, \chi(H) \Gamma(q^t) ( N_{\mathrm{neut}} + 1 ) \Gamma(q^t) \chi(H) e^{-\mathrm{i}tH}u\Big\rangle \frac{dt}{t} \\ &\lesssim \big\| ( \mathrm{d}\Gamma(|p_2|^{-1}) + N_{\mathrm{neut}} + 1 )^{\frac12} (N_{\mathrm{neut}}+1)^{\frac32} u \big\|^2. \end{align*} This proves (iv) for any $\chi \in \mathrm{C}_0^\infty( ( \lambda-\varepsilon,\lambda+\varepsilon) )$. The extension of the result to any $\chi \in \mathrm{C}_0^\infty( \mathbb{R} \setminus (\sigma_{\mathrm{pp}}(H) \cup \tau ) )$ follows from standard arguments (see, e.g., \cite[Proposition 4.4.7]{DeGe97_01}). \end{proof} As mentioned before, the remainder of this section is devoted to the proof of a different version of propagation estimates, involving a time-dependent modified position operator. In accordance with the notations previously introduced in this section, we set \begin{equation*} x_{t,\rho} := \Big ( \frac{ |p_2| }{ |p_2| + t^{-\rho} } \Big )^{\frac12} x \Big ( \frac{ |p_2| }{ |p_2| + t^{-\rho} } \Big )^{\frac12} , \end{equation*} meaning that $x_{t,\rho,i} = x_i = \mathrm{i}\nabla$ if $i\in\{1,2,3,4,7,8,11,12\}$ (corresponding to the label of a massive particle), and \begin{equation*} x_{t,\rho,i} = \Big ( \frac{ |p_2| }{ |p_2| + t^{-\rho} } \Big )^{\frac12} x_i \Big ( \frac{ |p_2| }{ |p_2| + t^{-\rho} } \Big )^{\frac12} = \Big ( \frac{ |p_2| }{ |p_2| + t^{-\rho} } \Big )^{\frac12} (\mathrm{i}\nabla_{p_2}) \Big ( \frac{ |p_2| }{ |p_2| + t^{-\rho} } \Big )^{\frac12} , \end{equation*} if $i \in \{5,6,9,10,13,14\}$ (corresponding to the label of a neutrino). Likewise, \begin{equation*} \omega_{t,\rho} := \Big ( \frac{ |p_2| }{ |p_2| + t^{-\rho} } \Big )^{\frac12} \omega \Big ( \frac{ |p_2| }{ |p_2| + t^{-\rho} } \Big )^{\frac12} , \end{equation*} where the notation $\omega$ has been introduced in \eqref{eq:not_omega}. Similarly as above, for $R=(R_1,\dots,R_{14})$ and $R'=(R'_1,\dots,R'_{14})$, we have that \begin{equation*} \mathds{1}_{[R,R']}(|x_{t,\rho}|)=\big(\mathds{1}_{[R_1,R'_1]}( | x_{t,\rho,1} | ),\dots,\mathds{1}_{[R_{14},R'_{14}]}( | x_{t,\rho,14} | )\big). \end{equation*} \begin{Th}\label{th:propag_massless_2} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ vanish and consider the Hamiltonian $H = H_0 + g ( H_I^{(1)} + H_I^{(2)} )$ with $H_I^{(1)}$ and $H_I^{(2)}$ given by \eqref{interaction_term}. Assume that \begin{equation*} G \in L^2, \quad a_{(i),\cdot} G \in L^2, \quad |p_3|^{-1} a_{(i),\cdot} G \in L^2, \quad i = 1,2,3 , \end{equation*} and that \begin{equation*} G \in \mathbb{H}^{1+\mu} \text{ for some } \mu > 0. \end{equation*} \begin{itemize} \item[(i)] Let $\rho>0$ be such that $(1+\mu)^{-1} < \rho < 1$ and let $c>\rho^{-1}$. Let $\chi \in \mathrm{C}_0^{\infty}(\mathbb{R})$, $R=(R_1,\dots,R_{14})$ and $R'=(R'_1,\dots,R'_{14})$ be such that $R'_{i}>R_i>c$. There exists $\mathrm{C} > 0$ such that, for all $u \in \mathscr{D}(N_{\mathrm{neut}}^{\frac12})$, \begin{align*} & \int^{\infty}_{1} \Big \| \mathrm{d} \Gamma \Big ( \mathds{1}_{[R,R']} \Big ( \frac{ | x_{t,\rho} | }{ t } \Big ) \Big)^{\frac{1}{2}}\chi(H) e^{-\mathrm{i}tH} u \Big \|^2 \frac{dt}{t} \leq \mathrm{C} \big ( \| N_{\mathrm{neut}}^{\frac12} u \|^2 + \|u\|^{2} \big ). \end{align*} \item[(ii)] Let $\rho>0$ be such that $(1+\mu)^{-1} < \rho < 1$. Let $0<\mathrm{v}_0<\mathrm{v}_1$ and $\chi \in \mathrm{C}_0^{\infty}(\mathbb{R})$. There exists $\mathrm{C}>0$ such that, for all $u \in \mathscr{D}(N_{\mathrm{neut}}^{\frac12})$, \begin{align*} & \int^{\infty}_{1} \Big \| \mathrm{d}\Gamma \Big( \Big\langle \big( \frac{x_{t,\rho}}{t} - \nabla \omega_{t,\rho} \big ) , \mathds{1}_{ [ \mathrm{v}_0 , \mathrm{v}_1 ] }( \frac{x_{t,\rho}}{t} ) \big( \frac{x_{t,\rho}}{t} - \nabla \omega_{t,\rho} \big ) \Big\rangle \Big)^{\frac12} \chi(H)e^{-\mathrm{i} t H} u \Big\|^{2} \frac{dt}{t} \\ & \le \mathrm{C} \big ( \| N_{\mathrm{neut}}^{\frac12} u \|^2 + \|u\|^{2} \big ). \end{align*} \item[(iii)] Let $\rho>0$ be such that $(1+\mu)^{-1} < \rho < 1$. Let $0< \mathrm{v}_0<\mathrm{v}_1$, $J \in \mathrm{C}_0^{\infty}( \{ x \in \mathbb{R}^3 , \mathrm{v}_0 < |x| < \mathrm{v}_1\})$ and $\chi \in \mathrm{C}_0^{\infty}(\mathbb{R})$. There exists $\mathrm{C}>0$ such that, for all $\ell \in \{ 1,2,3\} $ and $u \in \mathscr{D}(N_{\mathrm{neut}}^{\frac12})$, \begin{align*} & \int^{\infty}_{1} \Big \| \mathrm{d}\Gamma \Big( \Big| J\Big(\frac{x_{t,\rho}}{t}\Big) \Big( \frac{x_{t,\rho}^{(\ell)}}{t} - \partial_{(\ell)} \omega_{t,\rho} \Big) + \mathrm{h.c }\Big|\Big)^{\frac{1}{2}}\chi(H)e^{-\mathrm{i} t H} u \Big\|^{2} \frac{dt}{t} \le \mathrm{C} \big ( \| N_{\mathrm{neut}}^{\frac12} u \|^2 + \|u\|^{2} \big ). \end{align*} \item[(iv)] There exists $g_0>0$ such that, for all $|g|\le g_0$, the following holds: let $\rho>0$ be such that $(1+\mu)^{-1} < \rho < 1$. Let $\chi \in \mathrm{C}_{0}^{\infty}(\mathbb{R})$ be supported in $\mathbb{R}\backslash(\tau \cup \sigma_{\mathrm{pp}}(H))$. There exist $\delta>0$ and $\mathrm{C} > 0$ such that, for all $u \in \mathscr{D}( ( \mathrm{d}\Gamma(|p_2|^{-1}) + N_{\mathrm{neut}} + 1 ) (N_{\mathrm{neut}}+1)^{\frac32} )$, \begin{equation*} \int^{\infty}_{1} \Big\| \Gamma \Big ( \mathds{1}_{[0,\delta]} \Big(\frac{|x_{t,\rho}|}{t}\Big)\Big) \chi(H) e^{-\mathrm{i} t H} u \Big\|^2 \frac{dt}{t} \le \mathrm{C} \big\| ( \mathrm{d}\Gamma(|p_2|^{-1}) + N_{\mathrm{neut}} + 1 ) (N_{\mathrm{neut}}+1)^{\frac32} u \big\|^2. \end{equation*} \end{itemize} \end{Th} \begin{proof} (i) The proof is similar to that of Theorem \ref{th:propag_massless} (i), with the following differences. Let $F_i: \mathbb{R} \to \mathbb{R}$ be a bounded non-decreasing function supported in $(1,\infty)$. Instead of the function $f_i$ in \eqref{eq:def_fj}, we consider \begin{align*} \tilde f_i(t) := \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \mathrm{d}\Gamma_{i} \Big( F_i \Big(\frac{ | x_{t,\rho,i} | }{ct}\Big) \Big) e^{ - \mathrm{i} t H } \chi(H) u \Big \rangle , \end{align*} with $c>1$. Obviously, in the massive case ($i\in\{1,2,3,4,7,8,11,12\}$) the proof is the same than in Theorem \ref{th:propag_massless}. Let $i \in \{5,6,9,10,13,14\}$. In the same way as for $f_i$, since $F_i$ is bounded, we have that \begin{equation} \tilde f_i(t) \lesssim \| N_{\mathrm{neut}}^{\frac12} u \|^2 + \| u \|^2. \label{eq:new_propag1} \end{equation} Write $F_i ( r ) = \tilde F_i( r^2 )$. Note that \begin{align*} x_{t,\rho,i}^2 = \sum_{\ell=1}^3 \Big ( \frac{ |p_2| }{ |p_2| + t^{-\rho} } \Big )^{\frac12} x_i^{(\ell)} \Big ( \frac{ |p_2| }{ |p_2| + t^{-\rho} } \Big ) x_i^{(\ell)} \Big ( \frac{ |p_2| }{ |p_2| + t^{-\rho} } \Big )^{\frac12}. \end{align*} We compute \begin{align} & \mathbf{d}_0 \tilde F_i \Big ( \frac{ x_{t,\rho,i}^2 }{ c^2 t^2 } \Big ) \le - 2 \frac{1-\rho}{t} \tilde F_i' \Big ( \frac{ x_{t,\rho,i}^2 }{ c^2 t^2 } \Big ) - 2 \frac{ \rho - c^{-1} }{ t } \Big ( \frac{ |p_2| }{ |p_2| + t^{-\rho} } \Big )^{\frac12} \tilde F_i' \Big ( \frac{ x_{t,\rho}^2 }{ c^2 t^2 } \Big ) \Big ( \frac{ |p_2| }{ |p_2| + t^{-\rho} } \Big )^{\frac12} + \mathcal{O}( t^{-2+\rho} ) . \label{eq:new_propag2} \end{align} Details of the estimate \eqref{eq:new_propag2} are provided in Appendix \ref{app:technical}. The first two terms are non-positive since $\rho < 1$ and $c > \rho^{-1}$. Moreover, the commutators \begin{equation*} \Big [ H_I^{(j)}(G) , \mathrm{d}\Gamma_{i} \Big ( F_i \Big(\frac{ | x_{t,\rho,i} | }{ ct } \Big) \Big) \Big ], \quad j=1,2, \end{equation*} are given by expressions similar to \eqref{commut_11}--\eqref{commut_21}, with the operator $F_i (\frac{ | x_{t,\rho,i} | }{ ct })$ instead of $a_i$. Since $F_i$ vanishes near $0$ and $G \in \mathbb{H}^{1+\mu}$, one can verify using interpolation that $G$ belongs to the domain of $| x_{t,\rho,i} |^{1+\mu}$, and hence \begin{align*} \Big\|F_i\Big(\frac{ | x_{t,\rho,i} | }{ ct }\Big)G\Big\|_2=\mathcal{O}(t^{-1-\mu}). \end{align*} Therefore, in the same way as for Equation \eqref{eq:propagestim4}, Lemma \ref{lemSCW_1} yields \begin{align} & \Big \langle e^{ - \mathrm{i} t H } \chi ( H ) u , \Big [ H_I^{(1)}(G) + H_I^{(2)}(G) , \mathrm{i} \mathrm{d}\Gamma_{j} \Big( F_i \Big( \frac{ | x_{t,\rho,i} | }{ ct }\Big) \Big) \Big ] e^{ - \mathrm{i} t H } \chi(H) u \Big \rangle = \mathcal{O}( t^{-1-\mu} ) \| u \|^2. \label{eq:new_propag3} \end{align} Using \eqref{eq:new_propag1}, \eqref{eq:new_propag2} and \eqref{eq:new_propag3}, one can conclude that (i) holds by arguing in the same way as in the proof of Theorem \ref{th:propag_massless} (i). \vspace{0,2cm} (iv) Again, the proof resembles that of Theorem \ref{th:propag_massless} (iv), we focus on the differences. We consider $\chi \in \mathrm{C}_0^\infty( ( \lambda-\varepsilon,\lambda+\varepsilon) )$, $\delta > 0 $ and $q_i$, $i \in \{1, \dots, 14\}$ as in the proof of Theorem \ref{th:propag_massless}. Let $u \in \mathrm{Ran}( \mathds{1}_{\{n\}}(N_{\mathrm{lept}}-N_{\mathrm{neut}}))$ for some $n \in \mathbb{Z}$. Let $\tilde{q}^t=(q_1(\frac{ x_{t,\rho,1} }{t}),\dots,q_{14}(\frac{ x_{t,\rho,14} }{t}))$ and \begin{equation*} \tilde h(t):=\Big\langle e^{-\mathrm{i}tH}\chi(H)u,\Gamma(\tilde{q}^t)\frac{A}{t}\Gamma(\tilde{q}^t)e^{-\mathrm{i}tH}\chi(H)u\Big\rangle. \end{equation*} To verify that $\tilde h(t)$ is uniformly bounded, we modify \eqref{Step1} as follows. We can decompose \begin{equation*} A=\sum_{i=1}^{14} \mathrm{d}\Gamma_i( a_i ). \end{equation*} Clearly the massive cases ($i\in\{1,2,3,4,7,8,11,12\}$) can be handled as in \eqref{Step1}. For $i \in \{5,6,9,10,13,14\}$, we recall that \begin{equation*} a_i = \frac{\mathrm{i}}{2} \Big ( \frac{p_2}{|p_2|} \cdot \nabla + \nabla \cdot \frac{p_2}{|p_2|} \Big ) , \end{equation*} and write \begin{align} \Big\langle e^{-\mathrm{i}tH}\chi(H)u,\Gamma(\tilde{q}^t) \frac{\mathrm{d}\Gamma_i(a_i)}{t} \Gamma(\tilde{q}^t) e^{-\mathrm{i}tH}\chi(H)u\Big\rangle & = \Big\langle e^{-\mathrm{i}tH}\chi(H)u,\Gamma(\tilde{q}^t) \Big [ \frac{\mathrm{d}\Gamma_i(a_i)}{t} , \Gamma(\tilde{q}^t) \Big ] e^{-\mathrm{i}tH}\chi(H)u\Big\rangle \notag \\ &\quad + \Big\langle e^{-\mathrm{i}tH}\chi(H)u,\Gamma(\tilde{q}^t)^2 \frac{\mathrm{d}\Gamma_i(a_i)}{t} e^{-\mathrm{i}tH}\chi(H)u\Big\rangle. \label{eq:ddd0} \end{align} We estimate each term separately. First, we compute, using \cite[Lemma 5.2]{BoFaSi12_01}, \begin{equation*} \Big [ \frac{ a_i }{ t } , q_i \Big ( \frac{ x_{t,\rho,i} }{ t } \Big ) \Big ] = \mathcal{O}( t^{-1} ) |p_2|^{-1}. \end{equation*} This implies that \begin{align} \Big | \Big\langle e^{-\mathrm{i}tH}\chi(H)u,\Gamma(\tilde{q}^t) \Big [ \frac{\mathrm{d}\Gamma_i(a_i)}{t} , \Gamma(\tilde{q}^t) \Big ] e^{-\mathrm{i}tH}\chi(H)u\Big\rangle \Big | &\lesssim t^{-1} \| u \| \big \| \mathrm{d}\Gamma( |p_2|^{-1} ) e^{-\mathrm{i}tH}\chi(H)u \big \| \notag \\ & \lesssim \| u \| \big ( \| u\|+ \|N_{\mathrm{neut}} u\|+\big\| \mathrm{d}\Gamma( |p_2|^{-1} )u \big \| \big ) , \label{eq:ddd1} \end{align} where we used Lemma \ref{lem:evolp2-1} in the second inequality. Next we have that \begin{align}\label{eq:rewrite_ai} a_i &= \sum_{\ell=1}^3 x_{t,\rho,i}^{(\ell)}\big ( 1 + \mathcal{O}(t^{-\rho} ) |p_2|^{-1} \big ) + \mathcal{O}( 1 ) |p_2|^{-1} , \end{align} see Appendix \ref{app:technical}. Using that $|x_{t,\rho,i}|\le\delta t$ on the support of $q^t_i$, we deduce from the previous equation that \begin{align} & \Big | \Big\langle e^{-\mathrm{i}tH}\chi(H)u,\Gamma( \tilde{q}^t )^2 \frac{\mathrm{d}\Gamma_i(a_i)}{t} e^{-\mathrm{i}tH}\chi(H)u\Big\rangle \Big | \notag \\ &\lesssim \delta \| u \| \big ( \big \| N_{\mathrm{neut}} e^{-\mathrm{i}tH}\chi(H)u \big \| + \mathcal{O}(t^{-\rho}) \big \| \mathrm{d}\Gamma( |p_2|^{-1} ) e^{-\mathrm{i}tH}\chi(H)u \big \| \big ) + \mathcal{O}(t^{-1}) \| u \| \big \| \mathrm{d}\Gamma( |p_2|^{-1} ) e^{-\mathrm{i}tH}\chi(H)u \big \| \notag \\ &\lesssim \| u \| \big ( \big \| N_{\mathrm{neut}} u \big \| + \big \| \mathrm{d}\Gamma( |p_2|^{-1} ) u \big \| + \|u \| \big ) , \label{eq:ddd2} \end{align} where we used Lemma \ref{lem:evolp2-1} and the fact that $\rho > (1+\mu)^{-1}$ in the second inequality. We deduce from \eqref{eq:ddd1} and \eqref{eq:ddd2} that \eqref{eq:ddd0}, and therefore $\tilde h(t)$, are uniformly bounded in $t$. Next, we can decompose the derivative $\partial_t \tilde h ( t ) = \tilde R_1(t) + \tilde R_2(t) + \tilde R_3(t) + \tilde R_4(t)$ analogously as $h(t)$ in \eqref{eq:R1-R4}. The term $\tilde R_2(t)$ can be estimated in the same way as $R_2(t)$, using estimates similar to \eqref{eq:ddd1} and \eqref{eq:ddd2} instead of \eqref{Step1}. This yields \begin{equation} |\tilde R_2(t)|\lesssim \mathcal{O}(t^{-1-\mu})\|u\|\big(\|(N_{\mathrm{neut}} u\| + \big \| \mathrm{d}\Gamma( |p_2|^{-1} ) u \big\| + \|u\| \big). \label{eq:estim_tilde_R2} \end{equation} The estimate for $\tilde R_3(t)$ is identical to that of $R_3(t)$, yielding \begin{align} \tilde R_3(t) \ge \frac{\mathrm{c}_0}{t} \Big\langle e^{-\mathrm{i}tH}u, \chi(H) \Gamma(\tilde{q}^t) ( N_{\mathrm{neut}} + \mathds{1} ) \Gamma(\tilde{q}^t) \chi(H) e^{-\mathrm{i}tH}u\Big\rangle + \mathcal{O}(t^{-2})\|(N_{\mathrm{neut}}+1)u\|^2. \label{eq:estim_tilde_R3} \end{align} To estimate $\tilde R_4(t)$, we proceed as in \eqref{eq:ddd1}--\eqref{eq:ddd2}. This gives \begin{align} | \tilde R_4(t) | = \frac{1}{t} | \tilde h ( t ) | & \lesssim \frac{\delta}{t} \Big\langle e^{-\mathrm{i}tH}u, \chi(H) \Gamma(\tilde{q}^t) (N_{\mathrm{neut}}+1 )\Gamma(\tilde{q}^t) \chi(H) e^{-\mathrm{i}tH}u\Big\rangle \notag \\ &\quad + t^{-1-\rho} \| u \| \big \| \mathrm{d}\Gamma( |p_2|^{-1} ) e^{-\mathrm{i}tH}\chi(H)u \big \| \notag \\ & \lesssim \frac{\delta}{t} \Big\langle e^{-\mathrm{i}tH}u, \chi(H) \Gamma(\tilde{q}^t) (N_{\mathrm{neut}}+1 )\Gamma(\tilde{q}^t) \chi(H) e^{-\mathrm{i}tH}u\Big\rangle \notag \\ &\quad + t^{-1-\rho+\alpha(1+\mu)^{-1}} \| u \| \big(\big\|\mathrm{d}\Gamma( |p_2|^{-1} ) u\big\|+\|u\|\big) , \label{eq:estim_tilde_R4} \end{align} the second inequality being a consequence of Lemma \ref{lem:evolp2-1}. Next we consider $\tilde R_1(t)$. As in the proof of Theorem \ref{th:propag_massless}, we compute \begin{align*} \mathbf{d}_0 \tilde{q}_i^t(x_{t,\rho,i}) = -\frac{1}{2 t} \Big \langle \frac{x_{t,\rho,i}}{t} - \nabla\omega_{t,\rho,i} , \nabla \tilde{q}_t \big( \frac{x_{t,\rho,i}}{t} \big) \Big \rangle + \mathrm{h.c.} + \mathcal{O}(t^{-2+\rho}) , \end{align*} and set $\tilde g^t_i := -\frac{1}{2} \langle \frac{x_{t,\rho,i}}{t} - \nabla\omega_{t,\rho,i} , \nabla \tilde q_j \big( \frac{x_{t,\rho,i}}{t} \big) \rangle + \mathrm{h.c.}$. Let also $\tilde r^t_i$ be the remainder term in the previous equality. We have that \begin{align*} & \Big|\Big\langle e^{-\mathrm{i}tH} u, \Big ( \chi(H) \mathrm{d}\Gamma( \tilde{q}^t_i , \tilde{r}^t_i) \frac{\mathrm{d}\Gamma_i(a_i)}{t}\Gamma(\tilde{q}^t) \chi(H) + \mathrm{h.c.} \Big ) e^{-\mathrm{i}tH}u\Big\rangle\Big| \\ & \lesssim \mathcal{O}(t^{-2+\rho}) \big \| N_{\mathrm{neut}} \chi(H) e^{-\mathrm{i}tH} u \big \| \Big \| \frac{\mathrm{d}\Gamma_i(a_i)}{t}\Gamma(\tilde{q}^t) \chi(H) e^{-\mathrm{i}tH}u \Big \| \\ & \lesssim \mathcal{O}(t^{-2+\rho}) \big \| N_{\mathrm{neut}} \chi(H) e^{-\mathrm{i}tH} u \big \| \Big ( \big \| N_{\mathrm{neut}} \chi(H) e^{-\mathrm{i}tH}u \big \| + \mathcal{O}(t^{-\rho}) \big \| \mathrm{d}\Gamma( |p_2|^{-1} ) \chi(H) e^{-\mathrm{i}tH}u \big \| \Big ) , \end{align*} where we used \eqref{eq:ddd1}--\eqref{eq:ddd2} in the second inequality. By Lemma \ref{lem:evolp2-1}, and since $N_{\mathrm{neut}}-N_{\mathrm{lept}}$ commutes with $H$ and $N_{\mathrm{lept}}$ is relatively $H$-bounded, we obtain that \begin{align*} & \Big|\Big\langle e^{-\mathrm{i}tH} u, \Big ( \chi(H) \mathrm{d}\Gamma_i( \tilde{q}^t_i , \tilde{r}^t_i ) \frac{\mathrm{d}\Gamma_i(a_i)}{t}\Gamma(\tilde{q}^t) \chi(H) + \mathrm{h.c.} \Big ) e^{-\mathrm{i}tH}u\Big\rangle\Big| \\ & \lesssim \mathcal{O}(t^{-2+\rho}) \big ( \big \| N_{\mathrm{neut}} u \big \| + \| u \| \big ) \big ( \big \| N_{\mathrm{neut}} u \big \| + \big \| \mathrm{d}\Gamma(|p_2|^{-1})u\big\|+\|u\|\big). \end{align*} To estimate the term corresponding to $\frac1t\mathrm{d}\Gamma(\tilde{q}^t_i , \tilde{g}^t_i)$, we write \begin{align*} & \frac{1}{t} \Big|\Big\langle e^{-\mathrm{i}tH} u, \Big ( \chi(H) \mathrm{d}\Gamma(\tilde{q}^t_i , \tilde{g}^t_i ) \frac{\mathrm{d}\Gamma_i(a_i)}{t}\Gamma(\tilde{q}^t) \chi(H) + \mathrm{h.c.} \Big ) e^{-\mathrm{i}tH}u\Big\rangle\Big| \\ & = \frac{1}{t} \Big|\Big\langle e^{-\mathrm{i}tH} u, \Big ( \chi(H) \mathrm{d}\Gamma(\tilde{q}^t_i , \tilde{g}^t_i ) (N_{\mathrm{neut}}+1)^{-\frac12}(N_{\mathrm{neut}}+1)^{\frac12}\frac{\mathrm{d}\Gamma_i(a_i)}{t}\Gamma(\tilde{q}^t) \chi(H) + \mathrm{h.c.} \Big ) e^{-\mathrm{i}tH}u\Big\rangle\Big| \\ & \lesssim \frac{1}{t} \Big \| \mathrm{d}\Gamma((\tilde{g}^t_i)^*\tilde{g}^t_i)^{\frac12} \chi(H) e^{-\mathrm{i}tH} u \Big \| \Big ( \big \| ( N_{\mathrm{neut}} + 1 )^{\frac32} \Gamma(\tilde{q}^t) \chi(H) e^{-\mathrm{i}tH}u \big \| \\ & \qquad \qquad \qquad \qquad \qquad \qquad \qquad\qquad + \mathcal{O}(t^{-\rho}) \big \| ( N_{\mathrm{neut}} + 1 )^{\frac12} \mathrm{d}\Gamma( |p_2|^{-1} ) \chi(H) e^{-\mathrm{i}tH}u \big \| \Big ) , \end{align*} where we used again \eqref{eq:ddd1}--\eqref{eq:ddd2}. We expand the last expression and estimate the two terms separately. The first one is estimated exactly as in the proof of Theorem \ref{th:propag_massless}, yielding \begin{align*} & \frac{1}{t} \Big \| \mathrm{d}\Gamma((\tilde{g}^t_i)^*\tilde{g}^t_i)^{\frac12} \chi(H) e^{-\mathrm{i}tH} u \Big \| \big \| ( N_{\mathrm{neut}} + 1 )^{\frac32}\Gamma(\tilde{q}^t) \chi(H) e^{-\mathrm{i}tH}u \big \| \\ & \lesssim \frac{\alpha}{t} \Big \| \mathrm{d}\Gamma((\tilde{g}^t)^*\tilde{g}^t)^{\frac12} \chi(H) e^{-\mathrm{i}tH} (|N_{\mathrm{neut}}-N_{\mathrm{lept}}|+1)^{\frac32} u \Big \|^2 + \frac{1}{\alpha t} \big \| \Gamma(\tilde{q}^t) \chi(H) e^{-\mathrm{i}tH} u \big \|^2 + \frac{1}{\alpha t^2} \big \| ( N_{\mathrm{neut}}^2 + 1) u \big \|^2, \end{align*} with $\alpha > 0$. For the second term, we have that \begin{align*} & \frac{1}{t^{1+\rho}} \Big \| \mathrm{d}\Gamma((\tilde{g}^t_i)^*\tilde{g}^t_i)^{\frac12} \chi(H) e^{-\mathrm{i}tH} u \Big \| \big \| ( N_{\mathrm{neut}} + 1 )^{\frac12} \mathrm{d}\Gamma( |p_2|^{-1} ) \chi(H) e^{-\mathrm{i}tH}u \big \| \\ & \lesssim \frac{1}{t^{1+\rho}} \Big \| \mathrm{d}\Gamma((\tilde{g}^t_i)^*\tilde{g}^t_i)^{\frac12} \chi(H) e^{-\mathrm{i}tH} u \Big \|^2 + \frac{1}{t^{1+\rho}} \big \| ( N_{\mathrm{neut}} + 1 )^{\frac12} \mathrm{d}\Gamma( |p_2|^{-1} ) \chi(H) e^{-\mathrm{i}tH}u \big \|^2, \end{align*} and, since $N_{\mathrm{lept}}$ is relatively $H$-bounded, \begin{align*} &\big \| ( N_{\mathrm{neut}} + 1 )^{\frac12} \mathrm{d}\Gamma( |p_2|^{-1} ) \chi(H) e^{-\mathrm{i}tH}u \big \|^2 \\ &= \big \| \mathrm{d}\Gamma( |p_2|^{-1} ) \chi(H) e^{-\mathrm{i}tH} ( | N_{\mathrm{neut}} - N_{\mathrm{lept}} | + 1 )^{\frac12}u \big \|^2 + \big \| N_{\mathrm{lept}}^{\frac12} \mathrm{d}\Gamma( |p_2|^{-1} ) \chi(H) e^{-\mathrm{i}tH}u \big \|^2 \\ &\le \big \| \mathrm{d}\Gamma( |p_2|^{-1} ) \chi(H) e^{-\mathrm{i}tH} ( | N_{\mathrm{neut}} - N_{\mathrm{lept}} | + 1 )^{\frac12}u \big \|^2 + \big \| (c_1H+c_2)^{\frac12} \mathrm{d}\Gamma( |p_2|^{-1} ) \chi(H) e^{-\mathrm{i}tH}u \big \|^2 , \end{align*} where $c_1$ and $c_2$ are real numbers. Since one easily verifies that $[H,\mathrm{d}\Gamma(|p_2|^{-1})]$ is relatively $H$-bounded, Lemma \ref{lem:evolp2-1} implies that \begin{align*} & \big \| ( N_{\mathrm{neut}} + 1 )^{\frac12} \mathrm{d}\Gamma( |p_2|^{-1} ) \chi(H) e^{-\mathrm{i}tH}u \big \|^2 \lesssim t^{(1+\mu)^{-1}} \big \| ( \mathrm{d}\Gamma( |p_2|^{-1} ) + 1 ) ( N_{\mathrm{neut}} + 1 )^{\frac12} u \big \|^2 . \end{align*} Putting together the previous estimates, we obtain that \begin{align} & |\tilde R_1(t)| \lesssim \frac{\alpha+1}{t} \Big \| \mathrm{d}\Gamma(\tilde{g}_t^*\tilde{g}_t)^{\frac12} \chi(H) e^{-\mathrm{i}tH} (|N_{\mathrm{neut}}-N_{\mathrm{lept}}|+1)^{\frac32} u \Big \|^2 + \frac{1}{\alpha t} \big \| \Gamma(\tilde{q}^t) \chi(H) e^{-\mathrm{i}tH} u \big \|^2 \notag \\ &\quad + \mathcal{O}(t^{-2+\rho}) \big ( \big \| N_{\mathrm{neut}} u \big \| + \| u \| \big ) \big ( \big \| N_{\mathrm{neut}} u \big \| + \big \| \mathrm{d}\Gamma(|p_2|^{-1})u\big\|+\|u\|\big) \notag \\ &\quad + \mathcal{O}(t^{-1-\rho+(1+\mu)^{-1}} ) \big \| ( \mathrm{d}\Gamma( |p_2|^{-1} ) + 1 ) ( N_{\mathrm{neut}} + 1 )^{\frac12} u \big \|^2 . \label{eq:estim_tilde_R1} \end{align} From \eqref{eq:estim_tilde_R2}--\eqref{eq:estim_tilde_R1}, we conclude the proof in the same way as for Theorem \ref{th:propag_massless} (iv). \end{proof} \section{Asymptotic completeness}\label{sec:AC} In this section, we prove Theorem \ref{thm:main}. We begin by recalling the definitions and basic properties of the asymptotic spaces and of the wave operators in Subsection \ref{subsec:wave}. Subsection \ref{subsec:inverse_wave} is devoted to the proof of an important ingredient of the proof of Theorem \ref{thm:main}, namely the existence of inverse wave operators. Finally, in Subsection \ref{subsec:AC}, we establish asymptotic completeness of the wave operators. \subsection{The asymptotic space and the wave operators}\label{subsec:wave} Most of the results of this section are straightforward adaptations of corresponding results established in \cite{DeGe99_01,Am04_01} for Pauli-Fierz Hamiltonians. Therefore we do not give the details of the proofs but refer the reader to \cite{DeGe99_01,Am04_01}. The only exception is Theorem \ref{th:unitary} where we establish unitarity of the wave operators in the case where the masses of the neutrinos vanish. Indeed, when neutrinos are supposed to be massless, Theorem \ref{th:unitary} cannot be proven as in \cite{DeGe99_01,Am04_01}. We rely instead on an elegant argument due to \cite{DeGe00_01}. Recall that the asymptotic creation and annihilation operators for bosons are formally defined by \eqref{eq:def_asympt_a}--\eqref{eq:def_a}, and by analogous formulas for leptons and neutrinos. See Appendix \ref{sec:asympt_a} for precise definitions and basic properties. In particular, the fermionic asymptotic creation or annihilation operators $b^{\pm,\sharp}_{l,\epsilon}(h)$ (for leptons) and $c^{\pm,\sharp}_{l,\epsilon}(h)$ (for neutrinos), with $h \in \mathfrak{h}_2$, are bounded. The bosonic operators $a^{\pm,\sharp}_{\epsilon}(h)$, with $h \in \mathfrak{h}_1$, are closed but unbounded. The space of asymptotic vacua is defined by \begin{equation*} \mathscr{K}^{\pm} = \left\{u \in \mathscr{D} ( N_W^{\frac12} ) , \, \forall h_1 \in \mathfrak{h}_2, h_2,h_3 \in \mathfrak{h}_1, l , \epsilon, a^{\pm}_\epsilon(h_1) u = b^{\pm}_{l,\epsilon}(h_2) u = c^{\pm}_{l,\epsilon}(h_3) u = 0 \right\}. \end{equation*} The asymptotic space is \begin{equation*} \mathscr{H}^{\pm} = \mathscr{K}^{\pm} \otimes \mathscr{H}. \end{equation*} The following proposition can be proven in the same way as \cite[Proposition 5.5]{DeGe99_01}. \begin{Prop}\label{prop:basic_prop_omega} Consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term} and suppose that \begin{equation*} G \in \mathbb{H}^{1+\mu} \text{ for some } \mu > 0. \end{equation*} Then \begin{itemize} \item[(i)] $\mathscr{K}^{\pm}$ is closed and $H$-invariant. \item[(ii)] For all $n \in \mathbb{N}$ and $h_1, \dots, h_n$, $\mathscr{K}^{\pm}$ is contained in the domain of $d^{\pm,*}(h_1)\dots d^{\pm,*}(h_n)$, where $d^{\pm,*}(h_i)$ stands for any of the operators $a^{\pm,*}_\epsilon(h_i)$, with $h_i \in \mathfrak{h}_1$, or $b^{\pm,*}_{l,\epsilon}(h_i)$ or $c^{\pm,*}_{l,\epsilon}(h_i)$, with $h_i \in \mathfrak{h}_2$. \item[(iii)] $\mathscr{H}_{\mathrm{pp}}(H) \subset \mathscr{K}^{\pm}$. \end{itemize} \end{Prop} Recall that the wave operators $\Omega^\pm:\mathscr{H}^\pm\to\mathscr{H}$ are defined by \eqref{eq:def_Omega_1}--\eqref{eq:def_Omega_2} and that $H^{\mathrm{ext}}=H\otimes\mathds{1}+\mathds{1}\otimes H_0$. The following properties are standard consequences of the definitions. \begin{Prop}\label{prop:intert} Consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term} and suppose that \begin{equation*} G \in \mathbb{H}^{1+\mu} \text{ for some } \mu > 0. \end{equation*} Then $\Omega^\pm$ are isometric and we have that \begin{equation*} H \Omega^\pm = \Omega^\pm H^{\mathrm{ext}}. \end{equation*} Moreover, \begin{equation*} d^{\pm,\sharp}(h) \Omega^\pm = \Omega^\pm (\mathds{1} \otimes d^{\sharp}(h)), \end{equation*} where $d^{\pm,\sharp}(h)$ stands for any of the operators $a^{\pm,\sharp}_\epsilon(h)$, with $h \in \mathfrak{h}_1$, or $b^{\pm,\sharp}_{l,\epsilon}(h)$ or $c^{\pm,\sharp}_{l,\epsilon}(h)$, with $h \in \mathfrak{h}_2$. \end{Prop} If the masses of the neutrinos are supposed to be positive, then one can show that the $\Omega^\pm$ are unitary by adapting an argument of \cite{Ho69_02} (see also \cite[Theorem 5.6]{DeGe99_01}). In the massless case, this argument fails, but one can follow the approach of \cite{DeGe00_01}. We give a sketch of the proof. \begin{Th}\label{th:unitary} Suppose that the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ are positive and consider the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. Suppose that \begin{equation*} G \in \mathbb{H}^{1+\mu} \text{ for some } \mu > 0. \end{equation*} Then $\Omega^{\pm}$ are unitary maps form $\mathscr{H}^\pm$ to $\mathscr{H}$. The same holds if the masses of the neutrinos vanish and if one considers instead the Hamiltonian $H = H_0 + g ( H_I^{(1)} + H_I^{(2)} )$ with $H_I^{(1)}$ and $H_I^{(2)}$ given by \eqref{interaction_term}. \end{Th} \begin{proof} As mentioned above, the proof in the massive case is a straightforward adaptation of arguments used in \cite{Ho69_02}. We consider the case where the masses of the neutrinos vanish. By the previous proposition, $\Omega^\pm$ are isometric. It remains to verify that they are onto. Let $\mathfrak{f}_1 \subset \mathfrak{h}_1, \mathfrak{f}_2 \subset \mathfrak{h}_2$ be two subspaces of finite dimensions. For $u \in \mathscr{H}$, let \begin{equation*} n^\pm_{\mathfrak{f}_1, \mathfrak{f}_2}(u) = \sum_{\epsilon=\pm}\sum^{\mathrm{dim} \mathfrak{f}_1}_{i=1} \|a^\pm_\epsilon(h_i)u\|^2 + \sum_{\epsilon=\pm,l=1,2,3}\sum^{\mathrm{dim} \mathfrak{f}_2}_{j=1} \big( \|b^\pm_{l,\epsilon}(g_j)u\|^2 + \|c^\pm_{j,\epsilon}(g_k)u\|^2 \big), \end{equation*} where $\{h_i\}$ and $\{g_j\}$ are orthonormal bases of $\mathfrak{f}_1$ and $\mathfrak{f}_2$, respectively. Note that if $u \notin \mathscr{D}(a^\pm_{l,\epsilon}(h_{i}))$ for some $i$, then $n^\pm_{\mathfrak{f}_1, \mathfrak{f}_2}(u) = \infty$. Clearly, \begin{align*} n^\pm_{\mathfrak{f}_1, \mathfrak{f}_2}(u) & = \lim_{t \to \infty} \sum_{\epsilon=\pm}\sum^{\mathrm{dim} \mathfrak{f}_1}_{i=1} \|a_{\epsilon}(h_{i,\pm t}) e^{\mp \mathrm{i} t H}u\|^2 + \sum_{\epsilon=\pm,l=1,2,3}\sum^{\mathrm{dim} \mathfrak{f}_2}_{j=1} \big( \|b_{l,\epsilon}(g_{j,\pm t}) e^{\mp \mathrm{i} t H}u\|^2 + \|c_{l,\epsilon}(g_{j,\pm t}) e^{\mp \mathrm{i} t H}u\|^2 \big) \\ & \le \big\langle e^{\mp\mathrm{i}tH}u , N e^{ \mp \mathrm{i} t H } u \big\rangle. \end{align*} Decomposing $N = N_W + 2 N_{\mathrm{lept}} + ( N_{\mathrm{neut}} - N_{\mathrm{lept}} )$ and using that $N_W$ and $N_{\mathrm{lept}}$ are relatively $H$-bounded and that $N_{\mathrm{neut}} - N_{\mathrm{lept}}$ commutes with $H$, we deduce that \begin{equation} n^\pm_{\mathfrak{f}_1, \mathfrak{f}_2}(u) \lesssim \langle u,|H|u\rangle + \langle u , N_{\mathrm{neut}} u \rangle + \| u \|^2. \label{eq:npm} \end{equation} Now, as in \cite[Theorem 4.3]{DeGe00_01}, one can verify that \begin{equation} \mathrm{Ran} \, \Omega^\pm = \overline{\mathscr{D}(n^\pm)} , \label{eq:npm_2} \end{equation} where \begin{equation*} n^\pm(u) = \sup_{\mathfrak{f}_1 \subset \mathfrak{h}_1,\mathfrak{f}_2 \subset \mathfrak{h}_2} n^\pm_{\mathfrak{f}_1, \mathfrak{f}_2}(u). \end{equation*} and $\mathscr{D}(n^\pm) = \{ u \in \mathscr{H} , n^\pm(u) < \infty \}$. In the previous equation, the supremum is taken over all finite dimensional subspaces $\mathfrak{f}_1$, $\mathfrak{f}_2$. From \eqref{eq:npm} we deduce that $\mathscr{D}(n^\pm)$ contains the dense subset $\mathscr{D}( |H|^{\frac12} ) \cap \mathscr{D}( N_{\mathrm{neut}}^{\frac12})$. By \eqref{eq:npm_2}, this shows that $\Omega^\pm$ are onto. \end{proof} In the remainder of this subsection we introduce the ``extended wave operators'', defined as in \cite{DeGe99_01}, and state some of their properties. Recall that $\mathscr{H}^{\mathrm{ext}} = \mathscr{H} \otimes \mathscr{H}$ and $H^{\mathrm{ext}} =H \otimes \mathds{1} + \mathds{1} \otimes H_0$. We set \begin{equation*} \mathscr{D} (\Omega^{\mathrm{ext},\pm}) = \bigotimes^{\infty}_{n=0} \mathscr{D}((|H|^{\frac{n}{2}})\otimes \bigoplus_{n_1 + \dots + n_{14} = n} \otimes^{n_1}_{s} \mathfrak{h}_2 \otimes \otimes^{n_2}_{s} \mathfrak{h}_2 \otimes \otimes^{n_3}_{a} \mathfrak{h}_1 \otimes \dots \otimes \otimes^{n_{14}}_{a} \mathfrak{h}_1, \end{equation*} and \begin{align} & \Omega^{\mathrm{ext},\pm} : \mathscr{D}(\Omega^{\mathrm{ext},\pm}) \to \mathscr{H} \notag \\ & \Omega^{\mathrm{ext},\pm} \psi \otimes d^*(h_1)\dots d^*(h_n) \Omega = d^{\pm,*}(h_1)\dots d^{\pm,*}(h_n)\psi , \label{eq:defomegaext} \end{align} where, as above, $d^{\pm,*}(h_i)$ stands for any of the operators $a^{\pm,*}_\epsilon(h_i)$, with $h_i \in \mathfrak{h}_1$, or $b^{\pm,*}_{l,\epsilon}(h_i)$ or $c^{\pm,*}_{l,\epsilon}(h_i)$, with $h_i \in \mathfrak{h}_2$, and likewise for $d^*(h_i)$. The definition \eqref{eq:defomegaext} extends to any vector in $\mathscr{D}(\Omega^{\mathrm{ext},+})$ by linearity. The fact that $\Omega^{\mathrm{ext},\pm}$ are well-defined follows from the properties of the asymptotic creation operators (see Appendix \ref{sec:asympt_a}). It follows directly from the definitions of $\Omega^\pm$ and $\Omega^{\mathrm{ext},\pm}$ that \begin{align*} \Omega^{\mathrm{ext}, \pm} |_{\mathscr{H}^{\pm}} = \Omega^{\pm} . \end{align*} Moreover, in the same way as in \cite[Theorem 5.7]{DeGe99_01}, it is not difficult to verify that, for all $u \in \mathscr{D}(\Omega^{\mathrm{ext},\pm})$, \begin{equation*} \underset{t \to \pm \infty}{\slim} e^{\mathrm{i} t H} I e^{- \mathrm{i} t H^{\mathrm{ext}}}u = \Omega^{\mathrm{ext},\pm}u . \end{equation*} where the scattering identification operator $I$ is defined in \eqref{SIO}. \subsection{The geometric inverse wave operators}\label{subsec:inverse_wave} In this section we establish the existence of two asymptotic observables using the propagation estimates of Section \ref{sec:propag}. Compared to similar results proven in \cite{DeGe99_01,FrGrSc02_01,Am04_01}, the main difficulty we encounter comes from the fact that the propagation observables of Section \ref{sec:propag} only hold for a dense set of states, and for suitable norms. For this reason, the results of this section are not straightforward modifications of previous papers. We begin with the following important proposition. \begin{Prop} \label{Thvelocity} Suppose that the masses of the neutrinos vanish and consider the Hamiltonian $H = H_0 + g ( H_I^{(1)} + H_I^{(2)} )$ with $H_I^{(1)}$ and $H_I^{(2)}$ given by \eqref{interaction_term}. Suppose that \begin{equation*} G \in L^2, \quad a_{(i),\cdot} G \in L^2, \quad |p_3|^{-1} a_{(i),\cdot} G \in L^2, \quad i = 1,2,3 , \end{equation*} and that \begin{equation*} G \in \mathbb{H}^{1+\mu} \text{ for some } \mu > 0. \end{equation*} Let $\delta > 0 $ and $q_i$, $i \in \{1, \dots, 14\}$ be functions in $\mathrm{C}_0^{\infty} (\{ x \in \mathbb{R}^3 , |x| \le 2 \delta \})$ such that $0 \le q_i \le 1$, $q_i=1$ on $\{ x \in \mathbb{R}^3 , |x| \leq \delta \}$ and let $\tilde{q}^t=(q_1(\frac{x_{t,\rho}}{t}),\dots,q_{14}(\frac{x_{t,\rho}}{t}))$. The following limits exist \begin{equation*} \Gamma^{\pm}(q) := \underset{t\to \pm \infty}{\slim} e^{\mathrm{i} t H} \Gamma(\tilde q^t) e^{-\mathrm{i} t H}. \end{equation*} Moreover, for all $\chi \in \mathrm{C}_{0}^{\infty}(\mathbb{R})$ supported in $\mathbb{R}\backslash(\tau \cup \sigma_{\mathrm{pp}}(H))$, there exists $\delta>0$ such that \begin{equation*} \Gamma^{\pm}(q) \chi( H ) = 0. \end{equation*} The same holds if the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ are positive and if one considers the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. \end{Prop} \begin{proof} We consider the more difficult case of $H=H_0 + g(H_I^{(1)}+H_I^{(2)})$ with the masses of the neutrinos equal to $0$. The proof can easily be adapted in the case of $H=H_0 + g ( H_I^{(1)}+H_I^{(2)}+H_I^{(3)}+H_I^{(4)} )$, if the masses of the neutrinos are positive. It suffices to prove the existence of \begin{equation*} \lim_{t\to \pm \infty} e^{\pm \mathrm{i} t H} \Gamma(\tilde{q}^t) e^{\mp \mathrm{i} t H} u, \end{equation*} for $u$ in a dense subset of $\mathscr{H}$. We consider \begin{equation*} \mathscr{E}:=\big\{ u \in \mathscr{H} , \exists \chi \in \mathrm{C}_0^{\infty}(\mathbb{R}) , n \in \mathbb{N} , u = \chi(H) \mathds{1}_{[-n,n]}(N_{\mathrm{neut}}-N_{\mathrm{lept}})u \big\}. \end{equation*} Let $u \in\mathscr{E}$ and let $\tilde \chi \in \mathrm{C}_0^\infty( \mathbb{R} )$ be such that $\tilde \chi \chi = \chi$. As in the proof of Theorem \ref{th:propag_massless_2} (iv), the Helffer-Sj{\"o}strand functional shows that \begin{align*} \big\|\big[\Gamma(\tilde{q}^t),\tilde{\chi}(H)\big](N_{\mathrm{neut}}+1)^{-1}\big\|=\mathcal{O}(t^{-1}). \end{align*} Since $N_{\mathrm{lept}}-N_{\mathrm{neut}}$ commutes with $H$ and $\Gamma(\tilde{q}^t)$ and since $N_{\mathrm{lept}}$ is relatively $H$-bounded, we deduce that \begin{align*} & e^{\pm \mathrm{i} t H} \Gamma(\tilde{q}^t) e^{\mp \mathrm{i} t H} \chi(H) \mathds{1}_{[-n,n]}(N_{\mathrm{neut}}-N_{\mathrm{lept}})u \\ &= \tilde \chi( H ) e^{\pm \mathrm{i} t H} \Gamma(\tilde{q}^t) e^{\mp \mathrm{i} t H} \chi(H) \mathds{1}_{[-n,n]}(N_{\mathrm{neut}}-N_{\mathrm{lept}})u + \mathcal{O}(t^{-1}) \\ &= \mathds{1}_{[-n,n]}(N_{\mathrm{neut}}-N_{\mathrm{lept}}) \tilde \chi( H ) e^{\pm \mathrm{i} t H} \Gamma(\tilde{q}^t) e^{\mp \mathrm{i} t H} \chi(H) \mathds{1}_{[-n,n]}(N_{\mathrm{neut}}-N_{\mathrm{lept}})u + \mathcal{O}(t^{-1}). \end{align*} To shorten notations, let $\chi_{(n)}(H) := \chi(H) \mathds{1}_{[-n,n]}(N_{\mathrm{neut}}-N_{\mathrm{lept}})$, $\tilde\chi_{(n)}(H) := \tilde\chi(H) \mathds{1}_{[-n,n]}(N_{\mathrm{neut}}-N_{\mathrm{lept}})$. By the previous equality, it now suffices to prove the existence of \begin{equation*} \lim_{t\to \pm \infty} \tilde \chi_{(n)}(H) e^{\pm \mathrm{i} t H} \Gamma(\tilde{q}^t) e^{\mp \mathrm{i} t H} \chi_{(n)}(H) u. \end{equation*} Set $W(t) := \tilde \chi_{(n)}(H) e^{\pm \mathrm{i} t H} \Gamma(\tilde{q}^t) e^{\mp \mathrm{i} t H} \chi_{(n)}(H)$ and write, for $t'>t\ge1$, \begin{align} \big\|W(t')u-W(t)u\big\|&=\Big\|\int_{t}^{t'}\partial_sW(s)uds\Big\| \le \sup_{v\in\mathscr{H},\|v\|=1} \int_{t}^{t'} | \langle v , \partial_s W(s) u \rangle | ds . \label{eq:W(t)-} \end{align} We compute \begin{align} \langle v , \partial_t W(t) u \rangle = & \langle v , \partial_t \tilde\chi_{(n)}(H) e^{ \pm \mathrm{i} t H }\Gamma(\tilde{q}^t) e^{ \mp \mathrm{i} t H } \chi_{(n)}(H) u \rangle \notag \\ & = \pm \langle v , \tilde\chi_{(n)}(H) e^{ \pm \mathrm{i} t H } \big( \mathbf{D}_0 \Gamma(\tilde{q}^t) + \mathrm{i} g [H_I^{(1)}+H_I^{(2)} , \Gamma(\tilde{q}^t) ] \big) e^{\mp \mathrm{i} t H } \chi_{(n)}(H) u \rangle \notag \\ & = \pm \langle v , \tilde\chi_{(n)}(H) e^{ \pm \mathrm{i} t H } \big( \mathrm{d}\Gamma(\tilde{q}^t, \mathbf{d}_0 \tilde{q}^t) + \mathrm{i} g [H_I^{(1)}+H_I^{(2)}, \Gamma(\tilde{q}^t) ] \big) e^{\mp \mathrm{i} t H } \chi_{(n)}(H) u \rangle. \label{eq:gamma+_1} \end{align} We will show that the right-hand-side is integrable in $t$ on $[1,\infty)$. We invoke arguments closely related to those used in the proof of Theorem \ref{th:propag_massless_2} (iv). First, the assumption that $G\in\mathbb{H}^{1+\mu}$, the commutation relations of Appendix \ref{sec:standard_def} and Lemma \ref{lemSCW_1} imply that \begin{align*} \big \| [H_I^{(1)} (G) + H_I^{(2)} (G) , \Gamma(\tilde{q}^t)] (N_{\mathrm{lept}}+N_W)^{-1} \big \| = \mathcal{O}(t^{-1-\mu}) . \end{align*} Since $N_W$ and $N_{\mathrm{lept}}$ are relatively $H$-bounded, this yields \begin{align*} \big\| [H_I^{(1)}+H_I^{(2)}, \Gamma(\tilde{q}^t) ] \chi(H) \big\| \lesssim t^{-1-\mu}. \end{align*} Now we consider the term involving $\mathrm{d}\Gamma(\tilde{q}^t, \mathbf{d}_0 \tilde{q}^t)$ in \eqref{eq:gamma+_1}. As in the proof of Theorem \ref{th:propag_massless_2} (iv), we have that \begin{align*} \mathbf{d}_0 \tilde{q}_i^t(x_{t,\rho,i}) = -\frac{1}{2 t} \Big \langle \frac{x_{t,\rho,i}}{t} - \nabla\omega_{t,\rho,i} , \nabla q_i \big( \frac{x_{t,\rho,i}}{t} \big) \Big \rangle + \mathrm{h.c.} + \mathcal{O}(t^{-2}) , \end{align*} if $i\in\{1,2,3,4,7,8,11,12\}$ (corresponding to the label of a massive particle), and \begin{align*} \mathbf{d}_0 \tilde{q}_i^t(x_{t,\rho,i}) = -\frac{1}{2 t} \Big \langle \frac{x_{t,\rho,i}}{t} - \nabla\omega_{t,\rho,i} , \nabla q_i \big( \frac{x_{t,\rho,i}}{t} \big) \Big \rangle + \mathrm{h.c.} + \mathcal{O}(t^{-2+\rho}), \end{align*} if $i \in \{5,6,9,10,13,14\}$ (corresponding to the label of a neutrino). We treat the second case, namely $i \in \{5,6,9,10,13,14\}$, the case of $i\in\{1,2,3,4,7,8,11,12\}$ being easier. Let $\tilde{g}^t_i := -\frac{1}{2} \langle \frac{x_{t,\rho,i}}{t} - \nabla\omega_{t,\rho,i}(k) , \nabla q_i \big( \frac{x_{t,\rho,i}}{t} \big) \rangle + \mathrm{h.c.}$ and let $\tilde{r}^t_i = \mathbf{d}_0 \tilde{q}_i^t( x_{t,\rho,i}) - \frac{1}{t} \tilde{g}^t_i = \mathcal{O}(t^{-2+\rho})$. For the term corresponding to $\tilde{r}^t_i$, we have that \begin{align*} & \big\| e^{ \pm \mathrm{i} t H } \tilde\chi_{(n)}(H) \mathrm{d} \Gamma(\tilde{q}^t_i , \tilde{r}^t_i ) e^{ \mp \mathrm{i} t H } \chi_{(n)}(H) u \big\| \lesssim \mathcal{O}(t^{-2+\rho}) \big\| N e^{- \mathrm{i} t H } \chi_{(n)}(H) u \big\| = \mathcal{O}(t^{-2+\rho}). \end{align*} The equality comes from the facts that $N_{\mathrm{lept}}-N_{\mathrm{neut}}$ commutes with $H$, $N_W$ and $N_{\mathrm{lept}}$ are relatively $H$-bounded and that $(N+1) \tilde\chi_{(n)}(H)$ is bounded. To estimate the term corresponding to $\frac1t\mathrm{d}\Gamma(\tilde{q}^t_i , \tilde{g}^t_i)$, we use \eqref{eq:lau_1}, yielding \begin{align*} & \frac1t \big|\big\langle e^{ \mp \mathrm{i} t H } \tilde\chi_{(n)}(H) v , \mathrm{d} \Gamma(\tilde{q}^t_i , \tilde{g}^t_i) e^{ \mp \mathrm{i} t H } \chi_{(n)}(H) u \big\rangle\big| \\ & \le \frac1t \big\| \mathrm{d}\Gamma( | \tilde{g}^t_i | )^{\frac12} e^{ \mp \mathrm{i} t H } \tilde\chi_{(n)}(H) v \big \| \big \| \mathrm{d}\Gamma( | \tilde{g}^t_i | )^{\frac12} e^{ \mp \mathrm{i} t H } \chi_{(n)}(H) u \big\| . \end{align*} By (iii) of Theorem \ref{th:propag_massless_2}, \begin{align*} &\int_{1}^\infty \frac1t \big\| \mathrm{d}\Gamma( | \tilde{g}^t_i | )^{\frac12} e^{ \mp \mathrm{i} t H } \tilde\chi_{(n)}(H) v \big \| \big \| \mathrm{d}\Gamma( | \tilde{g}^t_i | )^{\frac12} e^{ \mp \mathrm{i} t H } \chi_{(n)}(H) u \big\| dt \lesssim \|u\|\|v\|. \end{align*} From \eqref{eq:W(t)-} and the previous computations, we easily deduce that for any $\varepsilon >0$, \begin{equation*} \big\|W(t')u-W(t)u\big\| \le \varepsilon, \end{equation*} for $t$ and $t'$ large enough. This proves that the limits $\Gamma^\pm(\tilde{q})$ exist. The fact that $\Gamma^\pm(\tilde{q})\chi(H)=0$ for all $\chi \in \mathrm{C}_{0}^{\infty}(\mathbb{R})$ supported in $\mathbb{R}\backslash(\tau \cup \sigma_{\mathrm{pp}}(H))$ is a consequence of Theorem \ref{th:propag_massless_2} (iv). Indeed, Theorem \ref{th:propag_massless_2} (iv) shows that $\Gamma^\pm(\tilde{q})\chi(H)u=0$ for all $u$ in a dense subset of $\mathscr{H}$. Since $\Gamma^\pm(\tilde{q})\chi(H)$ is bounded, the statement follows. \end{proof} We introduce the following notations that will be used in the proof of the next theorem: \begin{align*} & \check{\mathbf{d}}^{(i)}_{0l} b(t) = \frac{\partial b}{\partial t}(t) + \mathrm{i} \big( \omega^{(i)}_l(p_i) \oplus \omega^{(i)}_l(p_i) b(t) - b(t) \omega^{(i)}_l(p_i)\big), \quad i=1,2 , \\ & \check{\mathbf{d}}^{(3)}_{0} b(t) = \frac{\partial b}{\partial t}(t) + \mathrm{i} \big( \omega^{(3)}(p_3) \oplus \omega^{(3)}(p_3) b(t) - b(t) \omega^{(3)}(p_3)\big), \end{align*} if $b(t)$ is a family of operators from $\mathfrak{h}_1$ to $\mathfrak{h}_1\oplus\mathfrak{h}_1$, if $i=1,2$, or from $\mathfrak{h}_2$ to $\mathfrak{h}_2\oplus\mathfrak{h}_2$ if $j=3$. Likewise we set \begin{align*} \check{\mathbf{D}}_0 B(t) & = \frac{\partial B}{\partial t}(t) + \mathrm{i} ( H_0\otimes \mathds{1} + \mathds{1} \otimes H_0) B(t) - \mathrm{i} B(t) H_0 , \end{align*} for any family of operators $B(t) : \mathscr{H} \to \mathscr{H}^{\mathrm{ext}}$. Note that if $B(t) = (b_1(t) , \dots , b_{14}(t) )$ then, as functions of $t$, \begin{align*} \check{\mathbf{D}}_0 \mathrm{d}\Gamma(B) =: \mathrm{d}\Gamma( \check{\mathbf{d}}_0 b ) := \mathrm{d}\Gamma( & \check{\mathbf{d}}_{0,1} b_1 , \dots , \check{\mathbf{d}}_{0,14} b_{14} ) \\ := \mathrm{d}\Gamma\big( & \check{\mathbf{d}}^{(3)}_{0} b_1 , \check{\mathbf{d}}^{(3)}_{0} b_2 , \check{\mathbf{d}}^{(1)}_{01} b_3 , \check{\mathbf{d}}^{(1)}_{01} b_4 , \check{\mathbf{d}}^{(2)}_{01} b_5 , \check{\mathbf{d}}^{(2)}_{01} b_6 , \check{\mathbf{d}}^{(1)}_{02} b_7, \\ &\check{\mathbf{d}}^{(1)}_{02} b_8 , \check{\mathbf{d}}^{(2)}_{02} b_9 , \check{\mathbf{d}}^{(2)}_{02} b_{10} , \check{\mathbf{d}}^{(1)}_{03} b_{11} ,\check{\mathbf{d}}^{(1)}_{03} b_{12} , \check{\mathbf{d}}^{(2)}_{01} b_{13} , \check{\mathbf{d}}^{(2)}_{01} b_{14} \big). \end{align*} The main result of this subsection is stated in the following theorem. It shows the existence of inverse wave operators. \begin{Th}\label{th:inverse_wave} Suppose that the masses of the neutrinos vanish and consider the Hamiltonian $H = H_0 + g ( H_I^{(1)} + H_I^{(2)} )$ with $H_I^{(1)}$ and $H_I^{(2)}$ given by \eqref{interaction_term}. Suppose that \begin{equation*} G \in L^2, \quad a_{(i),\cdot} G \in L^2, \quad |p_3|^{-1} a_{(i),\cdot} G \in L^2, \quad i = 1,2,3 , \end{equation*} and that \begin{equation*} G \in \mathbb{H}^{1+\mu} \text{ for some } \mu > 0. \end{equation*} Let $\delta > 0 $ and $j_{0,i}$, $i \in \{1, \dots, 14\}$ be functions in $\mathrm{C}_0^{\infty} (\{ x \in \mathbb{R}^3 , |x| \le 2 \delta \})$ such that $0 \le j_{0,i} \le 1$, $j_{0,i}=1$ on $\{ x \in \mathbb{R}^3 , |x| \leq \delta \}$ and let $j_{\infty,i} = 1 - j_{0,i}$, $j_i = ( j_{0,i} , j_{\infty,i} )$. Let $\tilde{J}^t = (j_1(\frac{x_{t,\rho,1}}{t}),\dots,j_{14}(\frac{x_{t,\rho,14}}{t}))$. \begin{itemize} \item[(i)] The following limits exist \begin{align*} W^\pm(J) := \underset{t \to +\infty}{\slim} e^{\pm \mathrm{i} t H^{\mathrm{ext}}} \check{\Gamma} \big( \tilde{J}^t \big) e^{\mp \mathrm{i} t H}. \end{align*} \item[(ii)] For all $\chi \in \mathrm{C}_0^\infty( \mathbb{R} )$, we have that \begin{align*} W^\pm(J) \chi(H) = \chi(H^{\mathrm{ext}}) W^{\pm}(J) . \end{align*} \item[(iii)] Let $q = (q_1,\dots,q_{14})$ be such that $q_i j_{i,0} = j_{i,0}$. Then \begin{align*} \big ( \Gamma^{\pm}(q) \otimes \mathds{1} \big ) W^{\pm}(J)=W^{\pm}(J). \end{align*} \item[(iv)] For all $\chi \in \mathrm{C}_0^\infty( \mathbb{R} )$, we have that \begin{align*} \Omega^{\mathrm{ext},\pm} \chi(H^{\mathrm{ext}}) W^{\pm}(J) = \chi(H) . \end{align*} \end{itemize} The same holds if the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ are positive and if one considers the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. \end{Th} \begin{proof} (i) As in the proof of Proposition \ref{Thvelocity}, it suffices to prove the existence of \begin{equation*} \lim_{t\to \pm \infty} e^{\pm \mathrm{i} t H^{\mathrm{ext}}} \check{\Gamma}(\tilde{J}^t) e^{\mp \mathrm{i} t H} u, \end{equation*} for $u$ in a dense subset of $\mathscr{H}$. We consider again \begin{equation*} \mathscr{E}=\big\{ u \in \mathscr{H} , \exists \chi \in \mathrm{C}_0^{\infty}(\mathbb{R}) , n \in \mathbb{N} , u = \chi(H) \mathds{1}_{[-n,n]}(N_{\mathrm{neut}}-N_{\mathrm{lept}})u \big\} , \end{equation*} and fix $u \in\mathscr{E}$. Let $\tilde \chi \in \mathrm{C}_0^\infty( \mathbb{R} )$ be such that $\tilde \chi \chi = \chi$. In the same way as in Lemma \ref{numberestimates3}, one verifies that \begin{align*} \big\| \big ( \check{\Gamma}(\tilde{J}^t) \tilde{\chi}(H) - \chi(H^{\mathrm{ext}}) \check{\Gamma}(\tilde{J}^t) \big ) (N_{\mathrm{neut}}+1)^{-1}\big\|=\mathcal{O}(t^{-1}). \end{align*} Using that $N_{\mathrm{lept}}-N_{\mathrm{neut}}$ commutes with $H$, $\check{\Gamma}(\tilde{J}^t) \mathds{1}_{[-n,n]}(N_{\mathrm{neut}}-N_{\mathrm{lept}}) = \mathds{1}_{[-n,n]}(N_{\mathrm{neut}}^{\mathrm{ext}}-N_{\mathrm{lept}}^{\mathrm{ext}}) \check{\Gamma}(\tilde{J}^t)$, and that $N_{\mathrm{lept}}$ is relatively $H$-bounded, we deduce that \begin{align*} & e^{\pm \mathrm{i} t H^{\mathrm{ext}}} \check{\Gamma}(\tilde{J}^t) e^{\mp \mathrm{i} t H} \chi(H) \mathds{1}_{[-n,n]}(N_{\mathrm{neut}}-N_{\mathrm{lept}})u \\ &= \tilde \chi( H^{\mathrm{ext}} ) e^{\pm \mathrm{i} t H^{\mathrm{ext}} } \check{\Gamma}(\tilde{J}^t) e^{\mp \mathrm{i} t H} \chi(H) \mathds{1}_{[-n,n]}(N_{\mathrm{neut}}-N_{\mathrm{lept}})u + \mathcal{O}(t^{-1}) \\ &= \mathds{1}_{[-n,n]}(N_{\mathrm{neut}}^{\mathrm{ext}}-N_{\mathrm{lept}}^{\mathrm{ext}}) \tilde \chi( H^{\mathrm{ext}} ) e^{\pm \mathrm{i} t H^{\mathrm{ext}}} \check{\Gamma}(\tilde{J}^t) e^{\mp \mathrm{i} t H} \chi(H) \mathds{1}_{[-n,n]}(N_{\mathrm{neut}}-N_{\mathrm{lept}})u + \mathcal{O}(t^{-1}). \end{align*} Similarly as in the proof of Proposition \ref{Thvelocity}, to shorten notation, we set $\chi_{(n)}(H) := \chi(H) \mathds{1}_{[-n,n]}(N_{\mathrm{neut}}-N_{\mathrm{lept}})$ and $\tilde\chi_{(n)}(H^{\mathrm{ext}}) := \tilde\chi(H^{\mathrm{ext}}) \mathds{1}_{[-n,n]}(N_{\mathrm{neut}}^{\mathrm{ext}}-N_{\mathrm{lept}}^{\mathrm{ext}})$. By the previous equality, it now suffices to prove the existence of \begin{equation*} \lim_{t\to \pm \infty} \tilde \chi_{(n)}(H^{\mathrm{ext}}) e^{\pm \mathrm{i} t H^{\mathrm{ext}}} \check{\Gamma}(\tilde{J}^t) e^{\mp \mathrm{i} t H} \chi_{(n)}(H) u. \end{equation*} Set $\check{W}(t) := \tilde \chi_{(n)}(H^{\mathrm{ext}}) e^{\pm \mathrm{i} t H^{\mathrm{ext}}} \check{\Gamma}(\tilde{J}^t) e^{\mp \mathrm{i} t H} \chi_{(n)}(H)$ and write, for $t'>t\ge1$, \begin{align} \big\|\check{W}(t')u-\check{W}(t)u\big\|&=\Big\|\int_{t}^{t'}\partial_s\check{W}(s)uds\Big\| \le \sup_{v\in\mathscr{H}^{\mathrm{ext}},\|v\|=1} \int_{t}^{t'} | \langle v , \partial_s W(s) u \rangle | ds . \label{eq:checkW(t)-} \end{align} We compute \begin{align} &\langle v , \partial_t \check{W}(t) u \rangle = \langle v , \partial_t \tilde\chi_{(n)}(H^{\mathrm{ext}}) e^{ \pm \mathrm{i} t H^{\mathrm{ext}} }\check{\Gamma}(\tilde{J}^t) e^{ \mp \mathrm{i} t H } \chi_{(n)}(H) u \rangle \notag \\ & = \pm \langle v , \tilde\chi_{(n)}(H^{\mathrm{ext}}) e^{ \pm \mathrm{i} t H^{\mathrm{ext}} } \big\{ \check{\mathbf{D}}_0 \check{\Gamma}(\tilde{J}^t) + \mathrm{i} g \big( (( H_I^{(1)}+H_I^{(2)} ) \otimes \mathds{1} ) \check{\Gamma}(\tilde{J}^t) - \check{\Gamma}(\tilde{J}^t) ( H_I^{(1)}+H_I^{(2)} ) \big) \big\} e^{\mp \mathrm{i} t H } \chi_{(n)}(H) u \rangle \notag \\ & = \pm \langle v , \tilde\chi_{(n)}(H^{\mathrm{ext}}) e^{ \pm \mathrm{i} t H^{\mathrm{ext}} } \big\{ \mathrm{d}\check{\Gamma}(\tilde{J}^t, \check{\mathbf{d}}_0 \tilde{J}^t) \notag \\ &\qquad\qquad\qquad\qquad\qquad\qquad + \mathrm{i} g \big( (( H_I^{(1)}+H_I^{(2)} ) \otimes \mathds{1} ) \check{\Gamma}(\tilde{J}^t) - \check{\Gamma}(\tilde{J}^t) ( H_I^{(1)}+H_I^{(2)} ) \big) \big\} e^{\mp \mathrm{i} t H } \chi_{(n)}(H) u \rangle. \notag \end{align} As in the proof of Lemma \ref{numberestimates3}, we have that \begin{equation*} \tilde\chi_{(n)}(H^{\mathrm{ext}}) \big( (( H_I^{(1)}+H_I^{(2)} ) \otimes \mathds{1} ) \check{\Gamma}(\tilde{J}^t) - \check{\Gamma}(\tilde{J}^t) ( H_I^{(1)}+H_I^{(2)} ) \big) \chi_{(n)}(H) = \mathcal{O}(t^{-1-\mu}). \end{equation*} Moreover, similarly as in the proof of Proposition \ref{Thvelocity}, we decompose \begin{align*} & \check{\mathbf{d}}_0 \tilde{J}^t = \frac{1}{t} \tilde{G}^t + \tilde{R}^t, \quad \tilde{G}^t = (\tilde{g}_0^t, \tilde{g}^t_{\infty}), \quad \tilde{g}_{\sharp}^t = - \frac{1}{2} \Big( \Big( \frac{x_{t,\rho}}{t} - \nabla \omega_{t,\rho} \Big) \nabla \tilde{J}_{\sharp} \Big( \frac{x_{t,\rho}}{t}\Big) + \mathrm{h.c.} \Big), \quad \tilde{R}^t = \mathcal{O}(t^{-2+\rho}), \end{align*} and, for all $i \in \{1,\dots,14\}$, we have that \begin{align*} & \big\| e^{ \pm \mathrm{i} t H^{\mathrm{ext}} } \tilde\chi_{(n)}(H^{\mathrm{ext}}) \mathrm{d} \check{\Gamma}(\tilde{J}^t_i, \tilde{R}^t_i) e^{ \mp \mathrm{i} t H } \chi_{(n)}(H) u \big\| \lesssim \mathcal{O}(t^{-2+\rho}) \big\| N e^{\mp \mathrm{i} t H } \chi_{(n)}(H) u \big\| = \mathcal{O}(t^{-2+\rho}). \end{align*} The term corresponding to $\frac1t\mathrm{d}\check{\Gamma}(\tilde{J}^t_i, \tilde{G}^t_i)$ is estimated as \begin{align*} & \frac1t \big|\big\langle e^{ \mp \mathrm{i} t H^{\mathrm{ext}} } \tilde\chi_{(n)}(H^{\mathrm{ext}}) v , \mathrm{d} \check{\Gamma}(\tilde{J}^t_i , \tilde{G}^t_i) e^{ \mp \mathrm{i} t H } \chi_{(n)}(H) u \big\rangle\big| \\ & \le \frac1t \big\| ( \mathrm{d}\Gamma( | \tilde{g}^t_{0,i} | )^{\frac12} \otimes \mathds{1} ) e^{ \mp \mathrm{i} t H^{\mathrm{ext}} } \tilde\chi_{(n)}(H^{\mathrm{ext}}) v \big \| \big \| \mathrm{d}\Gamma( | \tilde{g}^t_{0,i} | )^{\frac12} e^{ \mp \mathrm{i} t H } \chi_{(n)}(H) u \big\| \\ & \quad + \frac1t \big\| ( \mathds{1} \otimes \mathrm{d}\Gamma( | \tilde{g}^t_{\infty,i} | )^{\frac12} ) e^{ \mp \mathrm{i} t H^{\mathrm{ext}} } \tilde\chi_{(n)}(H^{\mathrm{ext}}) v \big \| \big \| \mathrm{d}\Gamma( | \tilde{g}^t_{\infty,i} | )^{\frac12} e^{ \mp \mathrm{i} t H } \chi_{(n)}(H) u \big\| . \end{align*} By (iii) of Theorem \ref{th:propag_massless_2}, \begin{align*} &\int_{1}^\infty \frac1t \big\| ( \mathrm{d}\Gamma( | \tilde{g}^t_{0,i} | )^{\frac12} \otimes \mathds{1} ) e^{ \mp \mathrm{i} t H^{\mathrm{ext}} } \tilde\chi_{(n)}(H^{\mathrm{ext}}) v \big \| \big \| \mathrm{d}\Gamma( | \tilde{g}^t_{0,i} | )^{\frac12} e^{ \mp \mathrm{i} t H } \chi_{(n)}(H) u \big\| dt \lesssim \|u\|\|v\|, \end{align*} and likewise for the second term in the right-hand-side of the previous inequality. Eq. \eqref{eq:checkW(t)-} and the previous estimates imply that, for any $\varepsilon >0$, \begin{equation*} \big\|W(t')u-W(t)u\big\| \le \varepsilon, \end{equation*} for $t$ and $t'$ large enough, which proves that the limits $W^\pm(\tilde{J})$ exist. \vspace{0,2cm} (ii) This is a standard intertwining property. \vspace{0,2cm} (iii) It suffices to write \begin{align*} \big ( \Gamma^{\pm}(q) \otimes \mathds{1} \big ) W^{\pm}(J) u &= \big ( e^{ \pm \mathrm{i} t H} \Gamma(q^t) e^{\mp\mathrm{i} t H} \otimes \mathds{1} \big ) e^{\pm \mathrm{i} t H^{\mathrm{ext}}} \check{\Gamma} \big( \tilde{J}^t \big) e^{\mp \mathrm{i} t H} u + o(1) \\ &= \big ( e^{\pm\mathrm{i} t H^{\mathrm{ext}}} ( \Gamma(q^t) \otimes \mathds{1} \big ) \check{\Gamma} \big( \tilde{J}^t \big) e^{\mp \mathrm{i} t H} u + o(1) \\ &= e^{\pm\mathrm{i} t H^{\mathrm{ext}}} \check{\Gamma} \big( \tilde{J}^t \big) e^{\mp \mathrm{i} t H} u + o(1) \\ &= W^{\pm}(J) u +o(1), \end{align*} where we used that $( \Gamma(q^t) \otimes \mathds{1} \big ) \check{\Gamma} \big( \tilde{J}^t \big) = \check{\Gamma} \big( \tilde{J}^t \big)$ because $q_i j_{i,0} = j_{i,0}$ in the third equality. \vspace{0,2cm} (iv) This is again standard intertwining property. \end{proof} \subsection{Asymptotic completeness}\label{subsec:AC} We are now ready to conclude the proof of Theorem \ref{thm:main}. We begin by showing that the pure point spectral subspace of $H$, $\mathscr{H}_{\mathrm{pp}}(H)$, and the spaces of asymptotic vacua $\mathscr{K}^{\pm}$ coincide. Note that our reasoning process is slightly different from that of \cite{DeGe99_01,Am04_01}. \begin{Th}\label{th:AC} Suppose that the masses of the neutrinos vanish and consider the Hamiltonian $H = H_0 + g ( H_I^{(1)} + H_I^{(2)} )$ with $H_I^{(1)}$ and $H_I^{(2)}$ given by \eqref{interaction_term}. Suppose that \begin{equation*} G \in L^2, \quad a_{(i),\cdot} G \in L^2, \quad |p_3|^{-1} a_{(i),\cdot} G \in L^2, \quad i = 1,2,3 , \end{equation*} and that \begin{equation*} G \in \mathbb{H}^{1+\mu} \text{ for some } \mu > 0. \end{equation*} There exists $g_0>0$ such that, for all $|g| \le g_0$, \begin{equation*} \mathscr{H}_{\mathrm{pp}}(H) = \mathscr{K}^{\pm}. \end{equation*} The same holds, for all $g\in\mathbb{R}$, if the masses of the neutrinos $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$ are positive and if one considers the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}. \end{Th} \begin{proof} By Proposition \ref{prop:basic_prop_omega}, we know that \begin{equation*} \mathscr{H}_{\mathrm{pp}}(H) \subset \mathscr{K}^{\pm}. \end{equation*} Since in addition $\mathscr{H}_{\mathrm{pp}}(H)$ and $\mathscr{K}^{\pm}$ are closed, it remains to establish that $\mathscr{H}_{\mathrm{pp}}(H)^\perp \subset (\mathscr{K}^{\pm})^\perp$. In turn, since $\sigma_{\mathrm{pp}}(H)$ can only accumulate at the closed countable set $\tau$, it suffices to prove that $\mathrm{Ran}( \chi( H ) ) \subset (\mathscr{K}^{\pm})^\perp$ for all $\chi \in \mathrm{C}_0^{\infty}(\mathbb{R} \backslash ( \tau \cup \sigma_{\mathrm{pp}}(H) ) )$. Let $\chi \in \mathrm{C}_0^{\infty}(\mathbb{R} \backslash ( \tau \cup \sigma_{\mathrm{pp}}(H) ) )$ and let $u = \chi(H) v$. Let $J$ be defined as in the statement of Theorem \ref{th:inverse_wave}. By Theorem \ref{th:inverse_wave} (iv), we have that \begin{align*} \chi(H) = \Omega^{\mathrm{ext},\pm} \chi(H^{\mathrm{ext}}) W^{\pm}(J) = \Omega^{\mathrm{ext},\pm} ( \mathds{1} \otimes \Pi_\Omega ) \chi(H^{\mathrm{ext}}) W^{\pm}(J) + \Omega^{\mathrm{ext},\pm} ( \mathds{1} \otimes P_\Omega^\perp ) \chi(H^{\mathrm{ext}}) W^{\pm}(J) , \end{align*} where $\Pi_\Omega$ denotes the projection onto the Fock vacuum and $\Pi_\Omega^\perp$ the projection onto its orthogonal complement. We claim that the first term vanishes. Indeed, we can write \begin{align*} \Omega^{\mathrm{ext},\pm} ( \mathds{1} \otimes \Pi_\Omega ) \chi(H^{\mathrm{ext}}) W^{\pm}(J) & = \Omega^{\mathrm{ext},\pm} ( \chi(H) \otimes \Pi_\Omega ) W^{\pm}(J)\\ & = \Omega^{\mathrm{ext},\pm} ( \chi(H) \otimes \Pi_\Omega ) (\Gamma^\pm( q ) \otimes \mathds{1} ) W^{\pm}(J) , \end{align*} by Theorem \ref{th:inverse_wave} (iii), where $q$ is as in the statement of that result. Since $\chi( H )\Gamma^\pm( q )=0$ by Proposition \ref{Thvelocity}, we see that this term indeed vanishes. Hence we have proven that \begin{equation*} \chi( H ) v = \Omega^{\mathrm{ext},\pm} ( \mathds{1} \otimes \Pi_\Omega^\perp ) \chi(H^{\mathrm{ext}}) W^{\pm}(J) v. \end{equation*} Since $\Omega^{\mathrm{ext},\pm} ( \mathds{1} \otimes d^*(h) ) = d^{\pm,*}(h) \Omega^{\mathrm{ext},\pm}$ for any kind of creation operator $d^*(h)$, the last equality clearly shows that $\chi( H ) v \in ( \mathscr{K}^{\pm} )^\perp$. This concludes the proof of the theorem. \end{proof} Finally, as a consequence of Theorem \ref{th:AC}, we deduce that $H-E$ and $H_0$ are unitary equivalent if the conditions on $G$ are strengthened. \begin{Cor} Suppose that the masses of the neutrinos vanish and consider the Hamiltonian $H = H_0 + g ( H_I^{(1)} + H_I^{(2)} )$ with $H_I^{(1)}$ and $H_I^{(2)}$ given by \eqref{interaction_term}. Under the conditions of Theorem \ref{th:AC}, and assuming in addition that \begin{equation} b_{(i),\cdot} G \in L^2, \quad i = 1,2,3, \quad b_{(i),\cdot} b_{(i'),\cdot} G \in L^2 , \quad i,i'=1,2,3 , \label{eq:cond_bibi'} \end{equation} the operators $H-E$ and $H_0$ are unitarily equivalent. If the masses of the neutrinos are positive and if one considers the Hamiltonian \eqref{total_hamiltonian_g} with $H_I$ given by \eqref{interaction_term}, then, under the conditions of Theorem \ref{th:AC} and assuming in addition that \eqref{eq:cond_bibi'} holds, there exists $g_0>0$, which does not depend on $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$, such that, for all $|g|\le g_0$, $H-E$ and $H_0$ are unitarily equivalent. \end{Cor} \begin{proof} It follows from Proposition \ref{prop:intert}, Theorem \ref{th:unitary} and Theorem \ref{th:AC} that $H$ is unitary equivalent to $H |_{ \mathscr{H}_{\mathrm{pp}}(H) } \otimes \mathds{1} + \mathds{1} \otimes H_0$. In the case where neutrinos are massive, Theorems \ref{Thess2} and \ref{MourreII} imply that, for $|g|\le g_0$, $\mathscr{H}_{\mathrm{pp}}(H) = \{ E \}$ and $E$ is a simple eigenvalue of $H$. This shows that $H-E$ and $H_0$ are unitarily equivalent. Moreover, by Theorem \ref{MourreII}, $g_0$ can be chosen independently of the values of $m_{\nu_e}$, $m_{\nu_{\mu}}$, $m_{\nu_{\tau}}$. The same holds in the massless case, by Theorems \ref{Thess} and \ref{MourreII_1}. This proves the corollary. \end{proof}

2684b4afcbec5066128838783eaa58fb6a898276

\section{Introduction} Classical general relativity has passed all the experimental and observational tests so far. Nevertheless, its inadequateness to \emph{fully} describe our universe has been proven by the celebrated singularity theorems already in the 1970s~\cite{penrose1969,penrose1979} and recently generalized to low-regularity geometries \cite{KunzingerSteinbauerStojkovicVickers:2015,KunzingerSteinbauerVickers:2015,GrafGrantKunzingerSteinbauer:2017}. According to them, curvature singularities appearing in many exact solutions, such as the Schwarzschild black-hole or Friedmann--Lema\^{\i}tre--Robertson--Walker cosmological models, are not specific consequences of their highly symmetric nature, but are quite natural in generic solutions. In such ultra high-curvature regions near the Big Bang or deep inside black holes, the classical Einstein's field equations are no more valid and quantum effects should dominate. Therefore, quantum theory of gravity is required to describe the whole universe in a seamless manner, which is one of the most important unsolved problems in modern physics. Such a theory of quantum gravity governs the wave function of the spacetime which encodes all the possible topological configurations of the classical solutions. Classically, the Einstein equations determine the spacetime locally, while the topological degrees of freedom are reflected in the domains of coordinates. Indeed, one can construct globally different geometries from locally identical spacetimes by different ways of identification. In four and higher dimensions, the number of independent components of the Riemann tensor is larger than those of the Ricci tensor~\cite{wald}, which allows different solutions in vacuum distinct from Minkowski (such as the famous Schwarzschild or Kerr black hole). One of the difficulties of quantum gravity is the fact that both \emph{local dynamical} degrees of freedom and \emph{global topological} degrees of freedom come into play simultaneously. In this regard, 3-dimensional spacetimes are drastically different because the numbers of the independent components of the Riemann tensor and the Ricci tensor are equal. Hence, the Ricci tensor governed by the Einstein equations completely determines the spacetime curvature locally. In other words, there is \emph{no dynamical degree of freedom in three spacetime dimensions}. Consequently, the general vacuum solution of Einstein's equations is the maximally symmetric spacetime, namely Minkowski, de~Sitter (dS), or anti-de~Sitter (AdS) for zero, positive, or negative cosmological constant $\Lambda$, respectively. For this reason of local triviality, 3-dimensional general relativity has attracted much attention as a classroom to study how to handle the global and topological degrees of freedom in the context of quantum gravity~\cite{carlip}. Many explicit exact classical solutions have been recently summarized and classified in a comprehensive monograph \cite{GarciaDiaz:2017}. In particular, in the presence of a negative $\Lambda$, the classical theory admits a black-hole configuration with horizons given by nontrivial identifications, even though the spacetime is locally AdS~\cite{btz1,btz2}. This so-called BTZ black hole shares similar properties to the standard Kerr--AdS black hole in four dimensions, and has been intensively investigated to reveal the quantum aspects of black holes. The BTZ black hole has been generalized to the presence of the Maxwell field~\cite{chargedbtz} and with a null dust fluid (pure radiation)~\cite{husain1994}, which represent a charged black hole and formation of a black hole, respectively. However, it is still not clear if these are all possible solutions in such non-vacuum systems. To obtain more general classical solutions may bring us many insights regarding quantum gravity in the presence of matter. The topic of the present paper is along this direction of research. Indeed, we present here \emph{the most general solution of the Einstein equations with a cosmological constant $\Lambda$ and some kind of matter fields in three dimensions, under only a few weak assumptions}. Specifically, in section~\ref{sec_geom}, we prove that (virtually) all ${D=3}$ Lorentzian geometries belong either to the family of Robinson--Trautman spacetimes or to the family of Kundt spacetimes. Appendix~A contains the Christoffel symbols and all components of the Riemann and Ricci tensors for such a general metric. In appendices~B and~C, we systematically integrate the Einstein field equations for both the disjunct Robinson--Trautman and Kundt families, respectively. The matter content we consider in section~\ref{Einstein} is the mixture of an aligned pure radiation\footnote{The general solution with a timelike dust fluid was obtained in~\cite{BarrowShawTsagas:2006}.} and a gyratonic field (null matter with internal spin/helicity). The most general solutions are summarized and described in sections \ref{sec:discussionRT} and \ref{sec:discussionKundt}, respectively. These are 3-dimensional counterparts of the classic Robinson--Trautman solutions \cite{RobTra60,RobTra62} and the Kundt solutions \cite{Kundt:1961} in four dimensions. In ${D=4}$, they form two huge classes of exact solutions containing, as special cases, many spacetimes, in particular various highly symmetric backgrounds, black holes, accelerated sources, exact gravitational waves, and their combinations. (See \cite{Stephani:2003,GriffithsPodolsky:2009} for a review.) Interestingly, in higher dimensions ${D>4}$, the Robinson--Trautman solutions are more constrained \cite{PodOrt06,OrtPodZof08} so that they basically reduce to highly symmetric and well-known solutions such as the generalizations of the Schwarzschild, Reissner--Nordstr\"om, and Vaidya solutions (however, for $p$-form fields, there are some interesting nontrivial solutions \cite{OrtaggioPodolskyZofka:2015}). The large complementary Kundt class of solutions in higher dimensions was presented in \cite{PodolskyZofka:2009}. As we will show here, the Robinson--Trautman solutions in ${D=3}$ are {\it less} restrictive, in contrast to ${D \ge 4}$. And, surprisingly, we can write down the general solution of the Einstein equations in a closed form starting from only weak assumptions. \section{All geometries in 2+1 dimensions} Let us consider a general 3-dimensional Lorentzian spacetime $({\cal M},g)$ with the metric signature ${(-++)}$. For the curvature tensors, we shall adopt the convention ${[\nabla _a ,\nabla_b]\,V^c ={R^c}_{dab}V^d}$ and ${R_{ab}={R^c}_{acb}}$, and we choose the units such that ${c=1=G}$. \subsection{Uniqueness of Robinson--Trautman and Kundt metrics} \label{sec_geom} Consider an arbitrary null vector field $\mathbf{k}$. At any point we can set up a triad ${\mathbf{e}_{I}\equiv\left\{\mathbf{k},\,\mathbf{l},\,\mathbf{m}\right\}}$ of two null vectors, namely ${\mathbf{k}\cdot\mathbf{k}=0=\mathbf{l}\cdot\mathbf{l}}$, and one spatial vector $\mathbf{m}$ orthogonal to both $\mathbf{k}$ and $\mathbf{l}$ (${\mathbf{k}\cdot\mathbf{m}=0=\mathbf{l}\cdot\mathbf{m}}$), normalized as \begin{equation} \mathbf{k}\cdot\mathbf{l}=-1 \,, \qquad \mathbf{m}\cdot\mathbf{m}=1 \,. \label{triad} \end{equation} With respect to this triad, the covariant derivative of the vector field $\mathbf{k}$ can be decomposed as~\cite{OrtaggioPravdaPravdova:2013} \begin{equation} k_{a;b}=K_{ll}\,k_{a}k_{b}+K_{lk}\,k_{a}l_{b}-K_{lm}\,k_{a}m_{b} -K_{ml}\,m_{a}k_{b}-K_{mk}\,m_{a}l_{b}+K_{mm}\,m_{a}m_{b} \,, \label{CovDerDecomp} \end{equation} where the scalar coefficients $K_{I\!J}$ are defined by ${K_{I\!J}\equiv k_{a;b}\,e^a_{I}e^b_{\!J}}$. Notice that the terms in (\ref{CovDerDecomp}) with ${K_{kk}}$, ${K_{kl}}$, and ${K_{km}}$ are zero due to the constant (null) norm of the vector ${\mathbf{k}}$ implying ${k_{a;b}\,k^a=0}$. To characterize properties of the null congruence generated by $\mathbf{k}$ we employ the 2+1 analogy of classic \emph{optical scalars} (twist, shear, expansion) encoded in the \emph{optical matrix} $\rho$ defined as \begin{equation} \rho \equiv K_{mm}= k_{a;b}\,m^a m^b \,. \end{equation} (See \cite{OrtaggioPravdaPravdova:2013} for more details.) Recall that these optical scalars have an invariant meaning (in the sense that they are independent of the choice of the vectors $\mathbf{l}$ and $\mathbf{m}$ of the triad ${\mathbf{e}_{I}}$) only if the null vector field $\mathbf{k}$ is \emph{geodesic}, that is ${k_{a;b}\,k^b\propto k_a}$. In such a case, the decomposition (\ref{CovDerDecomp}) further simplifies since \begin{equation} K_{mk}\equiv k_{a;b}\,m^ak^b=0 \,. \label{geodeticity} \end{equation} Now, as a \emph{unique property of spacetimes in} ${D=3}$, the optical matrix ${\rho}$ is simply of rank ${1\times 1}$, i.e., it \emph{degenerates to a single scalar function}. Consequently, both its antisymmetric part (twist) and symmetric traceless part (shear) trivially vanish. \emph{Any} 2+1 geometry is thus necessarily \emph{twist-free} and \emph{shear-free}, with respect to the geodesic field $\mathbf{k}$. The only possibly non-trivial optical scalar is \emph{expansion} ${\Theta\equiv\rho}$. By these arguments, we have proved the following uniqueness theorem:\footnote{This observation was already made in 2010 by Chow, Pope and Sezgin~\cite{ChowPopeSezgin:2009}.} \vspace{5mm} \noindent \textbf{Theorem~1:} The only possible 3-dimensional Lorentzian geometries admitting a geodesic null vector field are either \emph{expanding} geometries of the \emph{Robinson--Trautman type} (with ${\Theta\not=0}$) or \emph{non-expanding} geometries of the \emph{Kundt type} (with ${\Theta=0}$). They are necessarily twist-free and shear-free. \vspace{5mm} Interestingly, for 2+1 geometries, the single assumption on \emph{geodesicity}\footnote{Even this weak assumption seems to be superficial. See Remark at the end of this section.} of the null congruence generated by the field~$\mathbf{k}$ is a necessary and \emph{sufficient} condition for such a congruence to be (locally) \emph{orthogonal to a family of null hypersurfaces} ${u=\hbox{const.}}$ To prove this statement, we employ the Frobenius's theorem in the form \begin{equation} k_{[a;b}k_{c]}=0 \qquad \Leftrightarrow \qquad k_a \propto u_{,a}\,. \label{FrobeniusTheorem} \end{equation} (See Eq.~(2.45) in \cite{Stephani:2003}.) In terms of the decomposition (\ref{CovDerDecomp}), the left-hand side of the theorem (\ref{FrobeniusTheorem}) reduces to ${K_{mk}\,m_{[a}l_{b}k_{c]}=0}$. This implies ${K_{mk}=0}$, that is geodesicity because (\ref{CovDerDecomp}) gives \begin{equation} k_{a;b}\,k^b=-K_{lk}\,k_a+K_{mk}\,m_a=-K_{lk}\,k_a \end{equation} due to (\ref{geodeticity}). To summarize: \vspace{5mm} \noindent \textbf{Theorem~2:} A null vector field in any 3-dimensional Lorentzian geometry is \emph{geodesic} if, and only if, it is \emph{hypersurface-orthogonal}. \vspace{5mm} Let us emphasize that this statement is \emph{not} true in dimensions ${D>3}$ because antisymmetric part of the optical matrix (twist) is then also involved~\cite{Stephani:2003,OrtaggioPravdaPravdova:2013}. \vspace{5mm} \noindent \textbf{Remark:} A geodesic null vector field~$\mathbf{k}$ \emph{can (locally) be constructed in any ``reasonable'' ${D=3}$ Lorentzian geometry}. Indeed, it is possible to take a 2-dimensional spatial surface $\Sigma$, at fixed initial time, and an arbitrary null vector at its single point ${P\in\Sigma}$. This vector can be Lie-dragged along arbitrary spatial coordinates of the surface, forming thus a vector field on $\Sigma$. Taking all these null vectors on $\Sigma$ as ``initial data'', one generates a congruence of null geodesics (due to standard existence theorem for ordinary differential equations yielding geodesics). Their tangent vectors everywhere in the spacetime domain form a geodesic null vector field. \vspace{5mm} \subsection{Canonical coordinates} These two theorems combined together allow us to construct convenient \emph{canonical coordinates} for all ${D=3}$ spacetimes admitting a geodesic null vector field $\mathbf{k}$. Due to Theorem~2, these spacetimes are always foliated by a family of null hypersurfaces ${u=\hbox{const.}}$ (such that ${k_a \propto u_{,a}}$), and we naturally choose the coordinate $u$ to label them smoothly. Without loss of generality we may also assume that the congruence of null geodesics generated by $\mathbf{k}$ is affinely parametrized by a parameter~$r$, and ${k_a = -u_{,a}}$. We choose the affine parameter $r$ along the congruence as the second coordinate on the manifold. The remaining spatial coordinate, denoted as~$x$, then span the 1-dimensional ``transverse'' subspace with constant $u$ and $r$. On any null hypersurface given by fixed $u$, the coordinate $x$ is chosen to be constant along each null geodesic generated by $\mathbf{k}$, i.e., ${\mathbf{k}=\mathbf{\partial}_r}$ with ${k^a=g^{ab}k_b= -g^{ab}u_{,b}= -g^{au}}$. In such canonical coordinates $(r,\,u,\,x)$, the specific contravariant metric components $g^{ab}$ thus must have the form ${g^{uu}=0}$, ${g^{xu}=0}$, and ${g^{ru}=-1}$. The only non-vanishing contravariant metric components are $g^{xx}$, ${g^{rx}}$, ${g^{rr}}$, and ${g^{ru}=-1}$. By inverting the matrix $g^{ab}$ we obtain the following nontrivial metric components: \begin{equation} g_{xx}=1/g^{xx}\,, \qquad g_{ur}=-1\,, \qquad g_{ux}= g_{xx}\,g^{rx} \,, \qquad g_{uu}= -g^{rr}+g_{xx}\,{(g^{rx})}^2 \,. \label{CovariantMetricComp} \end{equation} Consequently, we have proved the following: \vspace{5mm} \noindent \textbf{Theorem~3:} All 3-dimensional Lorentzian geometries admitting a geodesic null vector field~$\mathbf{k}$ can be parametrized by \emph{canonical coordinates} ${\{r,\,u,\,x\}}$ in which the metric takes the form \begin{equation} \dd s^2 = g_{xx}(r,u,x)\, \dd x^2+2\,g_{ux}(r,u,x)\, \dd u\, \dd x -2\,\dd u\,\dd r+g_{uu}(r,u,x)\, \dd u^2 \,, \label{general nontwist} \end{equation} and in which ${\mathbf{k}=\mathbf{\partial}_r\,}$ holds. \vspace{5mm} In these canonical coordinates, we can now explicitly express the null triad frame ${\left\{\mathbf{k},\,\mathbf{l},\,\mathbf{m}\right\}}$ introduced in Eq.~(\ref{triad}). The most natural choice is \begin{equation} \mathbf{k}=\partial_r\,, \qquad \mathbf{l}={\textstyle \frac{1}{2}}\,g_{uu}\,\partial_r+\partial_u \,, \qquad \mathbf{m}=m^x\big(g_{ux}\,\partial_r+\partial_x\big) \,, \label{triadexpl} \end{equation} where ${m^x=1/\sqrt{g_{xx}}}$ to satisfy the normalization condition ${\mathbf{m}\cdot\mathbf{m}=1}$. It is noted that $\mathbf{l}$ and $\mathbf{m}$ are not geodesics in general. A direct calculation for the metric (\ref{general nontwist}) immediately shows that the covariant derivative of $\mathbf{k}$ is ${k_{a;b}=\Gamma^{u}_{ab}=\frac{1}{2}g_{ab,r}}$. In particular, ${k_{r;b}=0=k_{a;r}}$ holds. Explicit form of the \emph{expansion scalar} thus becomes \begin{equation} \Theta=\rho=k_{x;x}\,m^x m^x = \frac{g_{xx,r}}{2\,g_{xx}} \,, \label{OptMat} \end{equation} implying the relation \begin{equation} g_{xx,r}=2\Theta\, g_{xx} \,. \label{shearfree condition} \end{equation} The Christoffel symbols and all coordinate components of the Riemann and Ricci curvature tensors for the general metric (\ref{general nontwist}) are calculated, using Eq.~(\ref{shearfree condition}), in Appendix~A. Moreover, the expression (\ref{shearfree condition}) can be integrated as \begin{equation} {\textstyle g_{xx}=\mathrm{R}^2(r,u,x)\,h_{xx}(u,x) \,, \qquad \hbox{where} \qquad \mathrm{R}=\exp\big(\int\Theta(r,u,x)\,\dd r\big)} \,. \label{IntShearFreeCond} \end{equation} The generic case ${\Theta\neq 0}$ gives the expanding \emph{Robinson--Trautman class}. When the expansion vanishes, ${\Theta=0}$, this effectively reduces to ${\mathrm{R}=1}$ so that the spatial metric ${g_{xx}(u,x)}$ is independent of the coordinate $r$. It yields exactly the \emph{Kundt class} of non-expanding, twist-free and shear-free geometries \cite{Stephani:2003,GriffithsPodolsky:2009,PodolskyZofka:2009,OrtaggioPravdaPravdova:2013}. These results can be summarized as: \vspace{5mm} \noindent \textbf{Theorem~4:} All geometries of Theorem~3 split into \emph{two distinct classes}. The first one is the \emph{Robinson--Trautman} class of metrics for which the geodesic null vector field~${\mathbf{k}=\mathbf{\partial}_r\,}$ is expanding, ${\Theta\not=0}$, while the second one is the non-expanding \emph{Kundt} class, ${\Theta=0}$. For the Kundt class, the metric coefficient $g_{xx}(u,x)$ in the canonical coordinates (\ref{general nontwist}) is independent of the affine parameter~$r$, while for the Robinson--Trautman subclass, ${g_{xx}(r,u,x)}$ is given by Eq.~(\ref{IntShearFreeCond}) in terms of~$\Theta$. \subsection{Einstein's field equations: $\Lambda$-vacuum, pure radiation, or gyratons} \label{Einstein} Having thus identified all 3-dimensional Lorentzian geometries (admitting a geodesic null vector field), we can now apply the field equations. Einstein's equations for the metric $g_{ab}$ have the form ${R_{ab}-\frac{1}{2}R\,g_{ab}+\Lambda\, g_{ab}=8\pi\, T_{ab}}$, where a non-vanishing \emph{cosmological constant} $\Lambda$ is admitted, and an arbitrary matter field is given by its energy momentum-tensor $T_{ab}$ with the trace ${T\equiv g^{ab}\,T_{ab}}$. By substituting their trace ${R=2(3\Lambda-8\pi\, T)}$ we obtain \begin{equation} R_{ab}= {\textstyle 2\Lambda\,g_{ab}+8\pi\big(T_{ab}-T\,g_{ab}\big)} \,. \label{EinstinEq} \end{equation} Our main aim here is to completely integrate the field equations (\ref{EinstinEq}) for the most general metric (\ref{general nontwist}), in both subcases of expanding Robinson--Trautman spacetimes and non-expanding Kundt spacetimes. We will assume a \emph{null radiation matter field aligned} with the geodesic null congruence generated by ${\mathbf{k}=\mathbf{\partial}_r\,}$. In addition, we will also admit a \emph{gyratonic matter}, which is a generalization of the null radiation field to include a spin of the null source \cite{Bonnor:1970b,FrolovFursaev:2005,KrtousPodolskyZelnikovKadlecova:2012}. Specifically, we will assume that the only non-vanishing components of the energy-momen\-tum tensor~$T_{ab}$ are ${T_{uu}}$, corresponding to the classical null radiation component, and ${T_{ux}}$, which encodes the inner gyratonic angular momentum. These can be functions of all coordinates: \begin{equation} T_{uu}(r,u,x) \,,\qquad T_{ux}(r,u,x)\,. \label{TuuTux} \end{equation} We immediately observe from Eq.~(\ref{general nontwist}) that the trace of such an energy-momentum tensor vanishes, \begin{equation} T=0 \quad \Rightarrow \quad R=6\Lambda \,. \label {T=0} \end{equation} Moreover, the \emph{conservation of energy-momentum} ${T_{ab;c}\,g^{bc}=0}$ (which follows from the simplified Bianchi identities ${{R_{ab}}^{;b}=0}$) yields the constraints \begin{eqnarray} T_{ax;x}\,g^{xx}+(T_{ar;x}+T_{ax;r})\,g^{rx}+T_{ar;r}\,g^{rr}-T_{ar;u}-T_{au;r}=0\,. \label{Bianchi1} \end{eqnarray} Using the Christoffel symbols (\ref{ChristoffelBegin})--(\ref{ChristoffelEnd}), the non-vanishing terms ${T_{ab;c}}$ entering Eq.~(\ref{Bianchi1}) are \begin{eqnarray} T_{xx;x} \rovno -2\Theta\, g_{xx}\,T_{ux}\,,\\ T_{ux;x} \rovno T_{ux,x}+{\textstyle(\Theta\, g^{rx}g_{xx} -\frac{1}{2}g_{ux,r}-\frac{1}{2}g^{xx}g_{xx,x})}\,T_{ux}-\Theta\, g_{xx}\,T_{uu}\,,\\ T_{ur;x} \rovno -\Theta\,T_{ux}\,,\\ T_{ux;r} \rovno T_{ux,r}-\Theta\,T_{ux}\,,\\ T_{ur;u} \rovno -{\textstyle\frac{1}{2}}g^{xx}g_{ux,r}\,T_{ux}\,,\\ T_{uu;r} \rovno T_{uu,r}-g^{xx}g_{ux,r}\,T_{ux}\,. \end{eqnarray} For ${a=x}$ and ${a=u}$, respectively, we thus obtain the following two constraints \begin{eqnarray} && T_{ux,r}+\Theta\,T_{ux}=0 \,, \label{EqTup} \\ && {\textstyle T_{uu,r}+\Theta\, T_{uu}=g^{xx}\big[\,T_{ux,x}+\big(g_{ux,r}-2\Theta\, g_{ux}-\frac{1}{2}g^{xx}g_{xx,x}\big)T_{ux}\big] } \label{EqTuu} \,, \end{eqnarray} while the constraint for ${a=r}$ is identically satisfied. Recall that such an energy-momentum tensor represents a generic (aligned) \emph{pure radiation (null) gyratonic matter field}. Of course, by setting ${T_{ux}=0}$ we recover the standard \emph{pure radiation (null dust) without gyratons}, and for ${T_{uu}=0=T_{ux}}$ \emph{vacuum} Einstein's equations are obtained. Explicit integration of the Einstein field equations (\ref{EinstinEq}) in 2+1 dimensions with the matter field (\ref{TuuTux}) constrained by Eqs.~(\ref{EqTup}) and (\ref{EqTuu}) has to be done separately for the Robinson--Trautman class ${(\Theta\neq 0)}$ and for the Kundt class ${(\Theta=0)}$ of spacetimes. Such a step-by-step integration is performed in Appendices~B and C, respectively. In subsequent sections~\ref{sec:discussionRT} and \ref{sec:discussionKundt}, we present the summary of the results, together with their physical discussions. \subsection{Energy conditions} In the spacetime~(\ref{general nontwist}), a set of three vectors $({\bf k},{\bf l},{\bf m})$ given by Eq.~(\ref{triadexpl}) forms a pseudo-orthonormal basis \begin{equation} {\bar E}^a_{(i)}\equiv({\bar E}^a_{(0)},{\bar E}^a_{(1)},{\bar E}^a_{(2)})=({\bf k},{\bf l},{\bf m})\,, \label{psude-bases} \end{equation} which satisfies \begin{equation} {\bar E}^a_{(i)}{\bar E}_{(j)a}={\bar \eta}_{(i)(j)}=\left( \begin{array}{ccc} 0 & -1 & 0 \\ -1 & 0 & 0 \\ 0 & 0 & 1 \\ \end{array} \right), \end{equation} where ${\bar \eta}_{(i)(j)}$ is the metric in the local Lorentz frame. The spacetime metric $g_{ab}$ is then given by \begin{equation} g_{ab}={\bar \eta}_{(i)(j)}\,{\bar E}^{(i)}_{a}{\bar E}^{(j)}_{b}=-k_al_b-l_ak_b+m_am_b\,. \end{equation} From ${\bf k}$ and ${\bf l}$, we construct unit timelike and spacelike vectors ${\bf u}$ and ${\bf s}$ such that \begin{align} {\bf u}\equiv &{\textstyle \frac{1}{\sqrt{2}}({\bf k}+{\bf l})=\frac{1}{\sqrt{2}}\bigl[(1+\frac{1}{2}\,g_{uu})\,\partial_r+\partial_u\bigl]}\,,\\ {\bf s}\equiv & {\textstyle\frac{1}{\sqrt{2}}({\bf k}-{\bf l})=\frac{1}{\sqrt{2}}\bigl[(1-\frac{1}{2}\,g_{uu})\,\partial_r-\partial_u\bigl]}\,, \end{align} which satisfy ${\bf u}\cdot {\bf u}=-1$, ${\bf s}\cdot{\bf s}=1$, ${\bf u}\cdot{\bf s}=0$, and ${\bf m}\cdot {\bf u}={\bf m}\cdot {\bf s}=0$. A set of three vectors $({\bf u},{\bf s},{\bf m})$ forms an orthonormal basis in the spacetime (\ref{general nontwist}): \begin{equation} {E}^a_{(i)}\equiv ({E}^a_{(0)},{E}^a_{(1)},{E}^a_{(2)})=({\bf u},{\bf s},{\bf m})\,.\label{normal-bases} \end{equation} These basis vectors satisfy \begin{equation} {E}^a_{(i)}{E}_{(j)a}={\eta}_{(i)(j)}=\left( \begin{array}{ccc} -1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \\ \end{array} \right) \end{equation} and the metric $g_{ab}$ is given by \begin{equation} g_{ab}={\eta}_{(i)(j)}{E}^{(i)}_{a}{E}^{(j)}_{b}=-u_au_b+s_as_b+m_am_b \,. \end{equation} The components of the energy-momentum tensor $T_{ab}$ in the orthonormal frame (\ref{normal-bases}) are given~by \begin{equation} T^{(i)(j)}=\eta^{(i)(k)}\eta^{(j)(l)}E^a_{(k)}E^b_{(l)}\,T_{ab} =\left( \begin{array}{ccc} \sigma+\nu & \nu & \zeta \\ \nu & -\sigma+\nu & \zeta \\ \zeta & \zeta & -\sigma \\ \end{array} \right)\,, \label{T(a)(b)} \end{equation} where \begin{equation} \textstyle{ \sigma=0\,,\qquad \nu=\frac12\, T_{uu}\,,\qquad \zeta=-\frac{1}{\sqrt{2}}\,m^x\,T_{ux} }\,. \end{equation} The expression (\ref{T(a)(b)}) with non-zero $\zeta$ is the 3-dimensional version of the type III energy-momentum tensor in the Hawking--Ellis classification~\cite{Hawking:1973uf,mmv2017}, and reduces to a special case of the type II if $\zeta\equiv 0$ holds. The type III energy-momentum tensor violates the null energy condition and therefore the weak and dominant energy conditions as well~\cite{energyconditions}. In the case of ${\zeta\equiv 0}$ (that is for ${T_{ux}=0}$), as shown in~\cite{energyconditions}, the standard energy conditions are equivalent to \begin{itemize} \item \emph{Null} energy condition (NEC): ${\nu\ge 0}$\,, \item \emph{Weak} energy condition (WEC): ${\nu\ge 0}$ and ${\sigma\ge 0}$\,, \item \emph{Strong} energy condition (SEC): ${\nu\ge 0}$ and ${\sigma\le 0}$\,, \item \emph{Dominant} energy condition (DEC): ${\nu\ge 0}$ and ${\sigma\ge 0}$\,. \end{itemize} Thus, while gyratons ($T_{ux}\ne 0$) violate NEC (and hence WEC and DEC as well), a pure radiation ($T_{ux}\equiv 0$) respects all the standard energy conditions if and only if $T_{uu}\ge 0$. \section{All Robinson--Trautman solutions and their properties} \label{sec:discussionRT} The complete integration of Einstein's field equations for a general 3-dimensional Robinson--Trautman metric in vacuum, with a cosmological constant $\Lambda$, and possibly a pure radiation field ${T_{uu}}$ and gyratons ${T_{ux}}$ is performed in appendix~B. Equations~(\ref{ExplTuu}) and (\ref{ExplTup}) show that the matter field necessarily takes the form \begin{eqnarray} && T_{uu} = \frac{\mathcal{N}}{r}-\frac{P(P\mathcal{J})_{,x}}{r^2}+\frac{fP^2\mathcal{J}}{r^3} \,, \label{RTTuu}\\ && T_{ux} = \frac{\mathcal{J}}{r} \,, \label{RTTup} \end{eqnarray} where ${\mathcal{N}(u,x)}$ and ${\mathcal{J}(u,x)}$ are functions determining the (density of) energy and angular momentum. Notice that ${T_{uu}\to0}$ and ${T_{ux}\to0}$ as ${r\to \infty}$, i.e., asymptotically the solutions ``become vacuum''. Indeed, it can be shown that ${R_{abcd}\to \Lambda\,(g_{ac}\,g_{bd}-g_{ad}\,g_{bc})}$ in the limit of ${r\to \infty}$ with constant $u$ and $x$, namely the spacetime is asymptotically maximally symmetric (Minkowski, dS, or AdS), at least locally. The corresponding generic metric can be written in the form (\ref{RTmetric}), that is \begin{eqnarray} \dd s^2 \rovno \frac{r^2}{P^2}\, \dd x^2+2\,(e\,r^2+f\,)\,\dd u \dd x -2\,\dd u\dd r \nonumber \\ && +\Big(-a +2\big[ P(Pe)_{,x}+(\ln P)_{,u} \big]\,r +(\Lambda+P^2e^2)\,r^2\Big)\, \dd u^2 \,. \label{RTmetricsummary} \end{eqnarray} The metric functions $P(u,x)$, $e(u,x)$, $f(u,x)$ and $a(u,x)$ are constrained just by two field equations (\ref{fieleq_ux}) and (\ref{RTEq2}): \begin{eqnarray} a_{,x} \rovno c f - 2 f_{,u}-16\pi\, \mathcal{J} \,, \label{RTE1}\\ a_{,u}\rovno ac+\triangle c+2(\Lambda+P^2e^2)P(Pf)_{,x} +3P^2f(P^2e^2)_{,x} \nonumber\\ && -2P^2f\,e_{,u} -P^2e(4 f_{,u}-cf+48\pi\, \mathcal{J}) + 16\pi\,\mathcal{N} \,, \label{RTE2} \end{eqnarray} where ${\triangle c \equiv P(Pc_{,x})_{,x}}$ is the transverse-space Laplace operator applied on the function $c$, defined by $c \equiv 2\big[P(Pe)_{,x}+(\ln P)_{,u}\big]$. (See Eqs.~(\ref{Laplace}) and (\ref{idno3}).) Generically, by prescribing an \emph{arbitrary gyratonic function} $\mathcal{J}$ (as well as \emph{any} metric functions ${P, e, f}$) we can always integrate (\ref{RTE1}) to obtain ${a(u,x)}$. Subsequently, its partial derivative $a_{,u}$ (and other given functions) uniquely determines the pure radiation energy profile $\mathcal{N}$ via the field equation (\ref{RTE2}). It is remarkable that in ${D=3}$ the function $f(u,x)$ in the metric (\ref{RTmetricsummary}) \emph{remains arbitrary} and, in general, \emph{non-vanishing}. This is an \emph{entirely new feature which does not occur in dimensions} ${D\ge4}$. Indeed, it was demonstrated in \cite{PodOrt06,OrtPodZof08,OrtaggioPodolskyZofka:2015} that for the Robinson--Trautman class of spacetimes in four and any higher dimensions necessarily ${f_i=0}$ for all ${(D-2)}$ spatial components. In this sense, the ${D=3}$ case is richer than the ${D\ge4}$ cases. \subsection{Gauge freedom} \label{gaugeRT} The most general form of the Robinson--Trautman metric (\ref{RTmetricsummary}) can be simplified using the gauge freedom. In the following, we consider the coordinate transformations ${(u,x)\to (u,x')}$ such that ${x=x(u,x')}$. Then, we have \begin{eqnarray} &&\dd x=\frac{\partial x}{\partial u}\dd u+\frac{\partial x}{\partial x'}\dd x'\,. \end{eqnarray} Since this equation gives \begin{eqnarray} &&\dd x=\frac{\partial x}{\partial u}\dd u+\frac{\partial x}{\partial x'}\biggl(\frac{\partial x'}{\partial u}\dd u+\frac{\partial x'}{\partial x}\dd x\biggl)=\biggl(\frac{\partial x}{\partial u}+\frac{\partial x}{\partial x'}\frac{\partial x'}{\partial u}\biggl)\dd u+\frac{\partial x}{\partial x'}\frac{\partial x'}{\partial x}\dd x\, , \end{eqnarray} we have the following relations: \begin{eqnarray} \frac{\partial x'}{\partial u}=-\frac{\partial x}{\partial u}\biggl(\frac{\partial x}{\partial x'}\biggl)^{-1},\qquad \frac{\partial x'}{\partial x}=\biggl(\frac{\partial x}{\partial x'}\biggl)^{-1}\, , \end{eqnarray} with which we obtain \begin{eqnarray} &&\frac{\partial}{\partial u}\biggl|_{(u,x)}=\biggl(\frac{\partial}{\partial u}+\frac{\partial x'}{\partial u}\frac{\partial}{\partial x'}\biggl)\biggl|_{(u,x')}=\biggl[\frac{\partial}{\partial u}-\frac{\partial x}{\partial u}\biggl(\frac{\partial x}{\partial x'}\biggl)^{-1}\frac{\partial}{\partial x'}\biggl]\biggl|_{(u,x')}\,,\\ &&\frac{\partial}{\partial x}\biggl|_{(u,x)}=\frac{\partial x'}{\partial x}\frac{\partial}{\partial x'}\biggl|_{(u,x')}=\biggl(\frac{\partial x}{\partial x'}\biggl)^{-1}\frac{\partial}{\partial x'}\biggl|_{(u,x')}\,. \end{eqnarray} Consequently, the Robinson--Trautman metric (\ref{RTmetricsummary}) is transformed to be \begin{eqnarray} \dd s^2 \rovno \frac{r^2}{P^2}\biggl(\frac{\partial x}{\partial x'}\biggl)^2\dd {x'}^2+2\frac{\partial x}{\partial x'}\biggl(\frac{r^2}{P^2}\frac{\partial x}{\partial u}+e\,r^2+f\biggl)\dd u\dd x' -2\,\dd u\dd r \nonumber \\ && +\Bigg[-a +2\biggl[ P\biggl(\frac{\partial x}{\partial x'}\biggl)^{-1}(Pe)_{,x'}+(\ln P)_{,u}-\frac{\partial x}{\partial u}\biggl(\frac{\partial x}{\partial x'}\biggl)^{-1}(\ln P)_{,x'} \biggl]\,r \nonumber \\ &&\qquad +(\Lambda+P^2e^2)\,r^2+\frac{r^2}{P^2}\biggl(\frac{\partial x}{\partial u}\biggl)^2+2\,(e\,r^2+f\,)\frac{\partial x}{\partial u}\Bigg]\, \dd u^2 \,, \end{eqnarray} where we have assumed $\frac{\partial x}{\partial x'}\ne 0$. Indeed, there are two natural gauge choices for the Robinson--Trautman metric (\ref{RTmetricsummary}): \begin{itemize} \item \textbf{Setting ${e=0}$} The function $e(u,x)$, introduced in the off-diagonal metric component ${g_{ux}=e\,r^2+f}$, \emph{can locally be removed}\footnote{But it may be \emph{convenient to retain} the function $e$ non-trivial because it may directly encode some physical or geometrical property, e.g. the gyratonic aspect of the matter (or as in the Garc\'{\i}a D\'{\i}az--Pleba\'nski 1981 gauge form \cite{GarciaPlebanski:1981,BicakPodolsky:1999a} of the ${D=4}$ Robinson--Trautman metric, where it directly encodes amplitudes of gravitational waves).} by a gauge transformation ${x \to x'}$ such that \begin{equation} x = -{\textstyle\int} P^2(u,x')\,e(u,x')\,\dd u \,. \label{gauge of e} \end{equation} (In the degenerate case when ${\frac{\partial x}{\partial x'}=0}$, namely, if both $P$ and $e$ are independent of $x$, we remove $e$ by ${x= -\int P^2(u)e(u)\,\dd u - \lambda\,x'}$, where $\lambda$ is a suitable real parameter \cite{PodOrt06}.) After this coordinate transformation and redefinitions such that ${P/\frac{\partial x}{\partial x'}\to P}$, ${f\frac{\partial x}{\partial x'}\to f}$, and ${a+2P^2ef \to a}$, the metric (\ref{RTmetricsummary}) reduces to (without a prime on the new coordinate $x'$) \begin{equation} \dd s^2 = \frac{r^2}{P^2}\, \dd x^2+2\,f\,\dd u \dd x -2\,\dd u\dd r +\Big[-a + 2\,(\ln P)_{,u}\,r + \Lambda\,r^2\,\Big]\, \dd u^2 \,, \label{RTmet} \end{equation} and the remaining two field equations (\ref{RTE1}), (\ref{RTE2}) are simplified considerably as \begin{eqnarray} a_{,x} \rovno 2\,\big[ f\,(\ln P)_{,u} - f_{,u} \big] -16\pi\, \mathcal{J} \,, \label{RTEq1eje0}\\ a_{,u}\rovno 2\,a(\ln P)_{,u}+2\,\triangle (\ln P)_{,u}+2\Lambda\,P(Pf)_{,x} + 16\pi\,\mathcal{N} \,, \label{RTEq2eje0} \end{eqnarray} where \begin{equation} \triangle (\ln P)_{,u} \equiv P[P(\ln P)_{,ux}]_{,x}\,. \label{Laplace2} \end{equation} Prescribing \emph{arbitrary} functions $P(u,x)$, $f(u,x)$ and $\mathcal{J}(u,x)$, we can thus integrate equation (\ref{RTEq1eje0}) to obtain $a(u,x)$. The second equation (\ref{RTEq2eje0}) then determines the corresponding unique form of the pure radiation profile $\mathcal{N}(u,x)$. \item \textbf{Setting ${P=1}$} Alternatively, it is possible to employ the gauge freedom to simplify the metric function $P$ to ${P=1}$, at the expense of keeping the off-diagonal metric function $e(u,x)$. This is achieved by the transformation ${x \to x'}$ given by \begin{equation} x = {\textstyle\int} P(u,x')\,\dd x' \,, \label{gauge of P} \end{equation} and redefinitions ${P^{-1}\frac{\partial x}{\partial u}+Pe\to e}$, ${Pf\to f}$, and ${a-2\frac{\partial x}{\partial u}f\to a}$. This leads to the metric \begin{equation} \dd s^2 = r^2\, \dd x^2+2\,(e\,r^2+f\,)\,\dd u \dd x -2\,\dd u\dd r +\Big[-a +2\,e_{,x}\,r +(\Lambda+e^2)\,r^2\Big]\, \dd u^2 \,, \label{RTmetP=1} \end{equation} for which the field equations (\ref{RTE1}), (\ref{RTE2}) are \begin{eqnarray} a_{,x} \rovno c f - 2 f_{,u}-16\pi\, \mathcal{J} \,, \label{RTE1Pje1}\\ a_{,u}\rovno ac+c_{,xx}+2(\Lambda+e^2)f_{,x} +3f(e^2)_{,x} -2f\,e_{,u} -e(4 f_{,u}-cf+48\pi\, \mathcal{J}) + 16\pi\,\mathcal{N} \,, \label{RTE2Pje1} \end{eqnarray} where ${c \equiv 2e_{,x}}$. These equations look more complicated than Eqs.~(\ref{RTEq1eje0}) and (\ref{RTEq2eje0}). \end{itemize} In any case, this is a very large family of explicit exact solutions to ${D=3}$ Einstein's gravity. In particular, as a unique feature, it admits the non-trivial off-diagonal metric function $f(u,x)$ which gives an additional degree of freedom. \subsection{Special subcase ${f=0}$} In this special subcase, the Robinson--Trautman metric (in the gauge ${e=0}$ corresponding to Eq.~(\ref{RTmet})) is further simplified to \begin{equation} \dd s^2 = \frac{r^2}{P^2}\, \dd x^2 -2\,\dd u\dd r +\Big[-a + 2\,(\ln P)_{,u}\,r + \Lambda\,r^2\,\Big]\, \dd u^2 \,, \label{RTmet_f=0} \end{equation} with $a(u,x)$ such that \begin{eqnarray} a_{,x} \rovno -16\pi\, \mathcal{J} \,, \label{RTEq1eje0_f=0}\\ a_{,u} - 2\,a(\ln P)_{,u}-2\,\triangle (\ln P)_{,u} \rovno \ \ 16\pi\,\mathcal{N} \,. \label{RTEq2eje0_f=0} \end{eqnarray} Notice that these field equations are \emph{fully independent of the cosmological constant} $\Lambda$. We are free to prescribe arbitrary metric functions $P(u,x)$ and ${a(u,x)}$. The corresponding matter contents $\mathcal{J} $ and $\mathcal{N}$ are then evaluated by Eqs.~(\ref{RTEq1eje0_f=0}) and (\ref{RTEq2eje0_f=0}), respectively. Interestingly, due to Eq.~(\ref{RTEq1eje0_f=0}), the key metric function $a(u)$ \emph{does not depend on the spatial coordinate $x$ if, and only if}, ${\mathcal{J}=0}$, i.e., \emph{in the absence of gyratons}. The remaining field equation (\ref{RTEq2eje0_f=0}) couples the metric functions $a$ and $(\ln P)_{,u}$ to the matter profile $\mathcal{N}(u,x)$. There also exists a special gyratonic solution given by ${P=1}$. In such a case, the metric has a very simple form \begin{equation} \dd s^2 = r^2\, \dd x^2 -2\,\dd u\dd r +\big[-a + \Lambda\,r^2\,\big]\, \dd u^2 \,, \label{RTmet_f=0_P=1} \end{equation} where \begin{eqnarray} a_{,x} = -16\pi\,\mathcal{J} \,, \qquad a_{,u} = 16\pi\,\mathcal{N} \,. \label{RTEq2eje0_f=0_P=1} \end{eqnarray} These two field equations are consistent if, and only if, ${\mathcal{N}_{,x}=-\mathcal{J}_{,u}}$. In fact, this solution can be interpreted (replacing $x$ by $\phi$ with a compact range) as a \emph{Vaidya-type object} emitting a null radiation gyratonic matter, i.e., a gyratonic generalization of the solution given by Eqs.~(\ref{Vaidya}) and (\ref{RTEVaidya}). \subsection{Arbitrarily moving object with pure radiation (${\mathcal{J}=0}$)} \label{sec_moving object} It is interesting to observe that the Robinson--Trautman line element (\ref{RTmet_f=0}) with only pure null radiation matter field ${T_{uu}=\mathcal{N}\,r^{-1}}$, where $\mathcal{N}(u,x)$ is given by expression (\ref{RTEq2eje0_f=0}), is an \emph{exact solution of Einstein's equations representing an arbitrarily moving object}. It accelerates due to an \emph{anisotropic emission} of the null matter. In fact, it is the 3-dimensional version of the so-called \emph{Kinnerley photon rocket} moving in Minkowski, de~Sitter, or anti--de~Sitter universe~\cite{Kinnersley:1969,Bonnor:1994,DainMoreschiGleiser:1996,Podolsky:2008,GursesSarioglu:2002,Podolsky:2011,NewmanUnti:1963}. To describe and physically interpret this family of spacetimes, it is better to introduce \emph{angular coordinate} ${\phi\in[0,2\pi)}$ instead of ${x\in {\rm R}}$ (i.e., to compactify the transverse space to a circle~$S^1$). Moreover, for convenience we relabel $a(u)$ to ${-\mu(u)}$ in the metric (\ref{RTmet_f=0}), \begin{equation} \dd s^2 = \frac{r^2}{P^2}\, \dd \phi^2 -2\,\dd u\dd r +\Big[\,\mu(u) + 2\,(\ln P)_{,u}\,r + \Lambda\,r^2\,\Big]\, \dd u^2 \,, \label{photon_rocket} \end{equation} and we take the key function $P(u, \phi)$ to be of the following special form \begin{equation} P(u, \phi) = \dot{z}^0(u)-\dot{z}^1(u)\cos\phi-\dot{z}^2(u)\sin\phi \,, \label{p_for_photon_rocket} \end{equation} where ${\dot{z}^0(u), \dot{z}^1(u), \dot{z}^2(u)}$ are \emph{arbitrary functions of the coordinate}~$u$ that satisfy a normalization condition ${(\dot{z}^0)^2-(\dot{z}^1)^2-(\dot{z}^2)^2=1}$ and a dot denotes derivative with respect to $u$. Such a constraint is always satisfied if we employ a parametrization ${\dot{z}^0\equiv \sqrt{1+v^2}}$, ${\dot{z}^1\equiv v\sin\psi}$, ${\dot{z}^2\equiv v\cos\psi}$, that is \begin{equation} P(u, \phi) = \sqrt{1+v^2(u)}-v(u)\sin\psi(u)\cos\phi-v(u)\cos\psi(u)\sin\phi \, . \label{p_for_photon_rocket_conv} \end{equation} \begin{figure}[htbp] \begin{center} \includegraphics[width=0.6\linewidth]{fig1} \caption{\label{fig1} The coordinate system (\ref{photon_rocket}) is adapted to an arbitrarily moving test particle in a space with Minkowski coordinates ${Z^0, Z^1, Z^2}$. At any event ${\cal P}$ of its worldline ${Z^\alpha=z^\alpha(u)}$, with velocity {\boldmath$u$}, a future null cone is labeled by $u$ and it is assigned the value of the corresponding proper time. The coordinate $r$ is an affine parameter along null geodesics which connect ${\cal P}$ with any event ${\cal Q}$ on the cone. The position of~${\cal Q}$ on the circle $u=\hbox{const.}$, $r=\hbox{const.}$ is specified by the angular coordinate $\phi$. The moving particle is always located at ${r=0}$. } \end{center} \end{figure} In fact, ${\{\dot{z}^0(u), \dot{z}^1(u), \dot{z}^2(u)\}}$ are \emph{components of the 3-velocity} {\boldmath$u$} of an object whose \emph{mass} is proportional to the function $\mu(u)$, located at the (singular) origin ${r=0}$ of the comoving coordinates~\cite{NewmanUnti:1963,Podolsky:2011}. This is illustrated in Fig.~\ref{fig1}. The function $v(u)$ in Eq.~(\ref{p_for_photon_rocket_conv}) thus denotes the \emph{speed} of the object in two spatial directions, while the angle $\psi(u)$ determines the actual \emph{direction} of the spatial velocity at an instant~$u$. Since these three functions can be prescribed arbitrarily, the metric (\ref{photon_rocket}) with the function (\ref{p_for_photon_rocket}) or (\ref{p_for_photon_rocket_conv}) is an \emph{exact solution for the gravitational field of an arbitrarily moving massive object}. Notice that, compared to the previously studied cases in ${D\ge 4}$ \cite{Podolsky:2008,Podolsky:2011,NewmanUnti:1963}, there is no Gaussian/Ricci curvature term ${K}$ in the component $g_{uu}$ of the metric (\ref{photon_rocket}) because the 1-dimensional transverse space spanned by $\phi$ is always flat and thus ${K=0}$. The single field equation (\ref{RTEq2eje0_f=0}) determines the corresponding $\mathcal{N}$ as \begin{eqnarray} \mathcal{N}(u,\phi) \rovno {\textstyle\frac{1}{16\pi}}\Big[-\mu_{,u}+2\mu\,(\ln P)_{,u}-2\,\triangle (\ln P)_{,u} \Big]\nonumber\\ \rovno {\textstyle\frac{1}{8\pi}}\Big[-{\textstyle\frac{1}{2}}\mu_{,u}+\big(\,\mu-(P_{,\phi})^2 \big)\,(\ln P)_{,u}-P P_{,u\phi\phi}+P_{,u} P_{,\phi\phi}+P_{,u\phi} P_{,\phi} \Big]\,, \label{NKinnersley} \end{eqnarray} where we have also employed the relation (\ref{Laplace2}). Interestingly, for the specific function (\ref{p_for_photon_rocket_conv}) the term ${\triangle (\ln P)_{,u}}$ is just equal to ${-(\ln P)_{,u}}$, so that the null matter field is simply given by \begin{equation} \mathcal{N}(u,\phi) = {\textstyle\frac{1}{8\pi}}\Big[-{\textstyle\frac{1}{2}}\mu_{,u}+(\mu+1)(\ln P)_{,u} \Big]\,. \label{NKinnersleySimple} \end{equation} Notice that the motion of the object with ${\mathcal{N}>0}$ is related to its \emph{decreasing mass function} $\mu(u)$. As an interesting special subcase, we get an explicit solution representing \emph{radial straight flight} of the photon rocket. It is obtained, e.g., by setting ${\psi(u)=\frac{\pi}{2}}$, so that ${\dot{z}^1(u)=v(u)}$ and ${\dot{z}^2(u)=0}$, and hence ${P = \sqrt{1+v^2(u)}-v(u) \cos\phi }$. In fact, it is convenient to express the speed function as \begin{equation} v(u) = \dot{z}^1(u) = \sinh\big({\textstyle\int}\alpha(u)\,\dd u\big)\,, \label{acceleration function} \end{equation} where the function $\alpha(u)$ determines the instantaneous \emph{acceleration} of the rocket. The key function then reads \begin{equation} P = \cosh\big({\textstyle\int}\alpha\,\dd u\big)-\sinh\big({\textstyle\int}\alpha\,\dd u\big)\, \cos\phi\,. \label{transfX-P} \end{equation} Now we perform the transformation \begin{equation} \sin\varphi =\frac{\epsilon\sin\phi}{P(u,\phi)}\,, \label{transfX} \end{equation} where $\epsilon=\pm1$. Equations~(\ref{transfX-P}) and (\ref{transfX}) imply \begin{equation} \cos\varphi=P^{-1}[\sinh \big({\textstyle\int}\alpha\,\dd u\big)-\cosh \big({\textstyle\int}\alpha\,\dd u\big) \cos\phi]\,, \label{transfX-cos} \end{equation} by using a remaining degree of freedom of transformation $\varphi\to \varphi+\pi$ because the sign of $\epsilon$ in Eq.~(\ref{transfX}) is arbitrary. Equations~(\ref{transfX-P}), (\ref{transfX}), and (\ref{transfX-cos}) also imply interesting relations ${(\ln P)_{,u} = \alpha(u)\,\cos\varphi}$, ${\epsilon\cot(\phi/2)=\tan(\varphi/2)\,\exp\big({\textstyle\int}\alpha(u)\,\dd u\big)}$, and ${\dd \varphi+\alpha(u)\sin\varphi\,\dd u=-\epsilon\:\dd \phi/P}$, so that finally the metric (\ref{photon_rocket}) is put into the following form: \begin{equation} \dd s^2 = r^2\big[\,\dd \varphi+\alpha(u)\sin\varphi\,\dd u\,\big]^2 -2\,\dd u\dd r +\big[\,\mu(u) + 2\,r\,\alpha(u)\cos\varphi + \Lambda\,r^2\,\big]\, \dd u^2 \,. \label{photon_rocket 1 direction} \end{equation} This is the ${D=3}$ version of the classic Kinnersley metric \cite{Kinnersley:1969, Bonnor:1994} with a cosmological constant \cite{Podolsky:2008, Podolsky:2011} and Eq.~(19.54) in \cite{GriffithsPodolsky:2009}. Gravitational field generated by a photon rocket \emph{uniformly accelerating} in a single spatial direction is obtained by simply assuming ${\alpha=\hbox{const.}}$ in the metric (\ref{photon_rocket 1 direction}). \subsection{Standing Vaidya-type object with pure radiation (${\mathcal{J}=0}$)} \label{sec_standing object} As an important special subcase of the family of solutions given by Eqs.~(\ref{photon_rocket}) and (\ref{p_for_photon_rocket}), we immediately obtain an exact gravitational field of a \emph{non-moving object}. Of course, this has a trivial spatial velocity ${v=0}$, equivalent to ${\dot{z}^1(u)=0=\dot{z}^2(u)}$, and the corresponding metric function (\ref{p_for_photon_rocket_conv}) is simply ${P=1}$. The metric (\ref{photon_rocket}) then reduces~to \begin{equation} \dd s^2 = r^2\, \dd \phi^2 -2\,\dd u\dd r +\big[\,\mu(u) + \Lambda\,r^2\,\big]\, \dd u^2 \,, \label{Vaidya} \end{equation} and it is circularly symmetric. The corresponding null matter field is given by \begin{equation} \mathcal{N}(u) = - {\textstyle\frac{1}{16\pi}}\,\mu_{,u}\,, \label{RTEVaidya} \end{equation} and it is also independent of $\phi$. In other words, the radiation from the source is \emph{isotropic} and thus causes no acceleration. The same result is obtain by setting ${\alpha=0}$ in the metric (\ref{photon_rocket 1 direction}). \subsection{Vacuum solutions: Minkowski, de~Sitter, anti--de~Sitter (${\mathcal{N}=0=\mathcal{J}}$)} \label{sec_MAdS} It is well-known that in ${D=3}$, all \emph{vacuum} solutions are just maximally symmetric background spacetimes of constant curvature, that is Minkowski, de~Sitter, anti--de~Sitter universe, according to the sign of the cosmological constant $\Lambda$. In the context of the family of metrics (\ref{RTmet})--(\ref{RTEq2eje0}), this occurs if, and only if, ${\mathcal{N}=0=\mathcal{J}}$. In such a case we need to satisfy a \emph{single integrability condition} for the function $a(x,u)$, namely \begin{equation} a_{,xu}=a_{,ux}\,. \label{integrabilit} \end{equation} The left-hand side is obtained by taking the $u$-derivative of Eq.~(\ref{RTEq1eje0}), while the right-hand side is obtained from the $x$-derivative of Eq.~(\ref{RTEq2eje0}), yielding the following constraint: \begin{eqnarray} a\,(\ln P)_{,ux} \!\!\!&=&\!\!\! f\,(\ln P)_{,uu} - 2f\,[(\ln P)_{,u}]^2 +3 f_{,u}\,(\ln P)_{,u}-f_{,uu} \nonumber\\ && -[\triangle (\ln P)_{,u}]_{,x}-\Lambda\, [P(Pf)_{,x}]_{,x} \,. \label{integrability} \end{eqnarray} There are two distinct subcases to be considered: \begin{itemize} \item \textbf{The case} ${(\ln P)_{,ux} \not=0}$ allows us to divide both parts of equation (\ref{integrability}) by the term ${(\ln P)_{,ux}}$. We thus obtain an explicit expression for the function $a(u,x)$, uniquely determined by the remaining free metric functions $P(u,x)$ and $f(u,x)$. \item \textbf{The case} ${(\ln P)_{,ux} =0 \Leftrightarrow}$ the function $P$ is \emph{separable} in the form ${P(u,x)=U(u)X(x)}$. In such a case, the function $a(u,x)$ is given by equation (\ref{RTEq1eje0}), while the functions $P(u,x)$ and $f(u,x)$ must satisfy the constraint (\ref{integrability}) with vanishing left-hand side. Interestingly, such a condition can be rewritten as \begin{equation} \big[\,U^{-1} (U^{-1}f)_{,u}\big]_{,u}= -\Lambda\, \big[X(Xf)_{,x}\big]_{,x} \,. \label{constraintPf} \end{equation} Assuming the function $f$ is \emph{also separable}, this equation can be explicitly solved by separation of variables $u$ and $x$. Moreover, when ${\Lambda=0}$, a general solution of Eq.~(\ref{constraintPf}) is ${f(u,x)=U(u)\,\big[\alpha(x)\int U(u)\,\dd u+\beta(x)\big]}$, where $\alpha(x)$ and $\beta(x)$ are arbitrary functions. \end{itemize} The vacuum Robinson--Trautman metric is \emph{simplified for} ${f=0}$, i.e., for the ${\mathcal{J}=0=\mathcal{N}}$ subcase of the metric (\ref{RTmet_f=0}). In this case, the two metric functions $a(u)$ and ${P(x,u)}$ are related by the single field equation (\ref{RTEq2eje0_f=0}), namely \begin{equation} \big[\triangle + a \big](\ln P)_{,u} = {\textstyle \frac{1}{2}}a_{,u}\,. \label{f=0vacuum} \end{equation} Prescribing an arbitrary $a(u)$, we obtain $P(x,u)$ by solving this differential equation. \emph{Assuming separability}, ${P(u,x)=U(u)X(x)}$, the only possibility is ${U=\sqrt{a(u)}}$, and (by redefining $x$) the corresponding metric takes the form \begin{equation} \dd s^2 = (r^2/a)\, \dd x^2 -2\,\dd u\dd r +\big[-a + (\ln a)_{,u}\,r + \Lambda\,r^2\,\big]\, \dd u^2 \,. \label{RTmet_f=0vacuum} \end{equation} The transformation \begin{equation} R=r/\sqrt{a(u)}\,,\qquad U={\textstyle\int}\sqrt{a(u)}\,\dd u\,, \qquad \phi=x\,, \label{transfRTmet_f=0vacuum} \end{equation} puts it into the ``Eddington form'' of maximally symmetric spaces \begin{equation} \dd s^2 = R^2\, \dd \phi^2 -2\,\dd U\dd R +\big(-1 + \Lambda\,R^2\,\big)\, \dd U^2 \,, \label{MinkAdS} \end{equation} which is actually the metric (\ref{Vaidya}) with ${\mu=-1}$. The following canonical form of Minkowski and (anti-)de~Sitter space \begin{equation} \dd s^2 = R^2\, \dd \phi^2 + (1-\Lambda\,R^2)^{-1}\, \dd R^2 - (1-\Lambda\,R^2)\, \dd T^2 \,, \label{MinkAdScanon} \end{equation} is obtained by expressing the null coordinate as ${U=T-\int(1-\Lambda\,R^2)^{-1}\dd R}$. \section{All Kundt solutions and their properties} \label{sec:discussionKundt} In appendix~C, we completely integrate Einstein's field equations for a general 3-dimensional Kundt line element in vacuum, with a cosmological constant $\Lambda$, and possibly a null radiation matter field $T_{uu}$ with a gyratonic component $T_{ux}$. These matter components have a generic form (\ref{KExplTup}) and~(\ref{KExplTuu}): \begin{eqnarray} && T_{uu} = \mathcal{N}+P^2\Big[\mathcal{J}_{,x} + \mathcal{J}\Big(f+ \frac{P_{,x}}{P}\Big) \Big]\,r \,, \label{KundtTuu}\\ && T_{ux} = \mathcal{J} \,. \label{KundtTup} \end{eqnarray} The metric can be written in the form (\ref{Kundtmetric}) which, in general, contains five metric functions $P(u,x)$, $e(u,x)$, $f(u,x)$, $a(u,x)$, and $b(u,x)$. This large family of spacetimes splits into \emph{two distinct subclasses}, namely: \begin{itemize} \item \textbf{The case} ${f=0}$ necessarily requires ${\Lambda=0}$ and ${F=0}$, so that such a general metric reads \begin{eqnarray} \dd s^2 \rovno \frac{1}{P^2}\, \dd x^2+2\,e\,\dd u \dd x -2\,\dd u\dd r +\big(a+b\,r\big)\, \dd u^2 \,, \label{Kundtmetric_f=0} \end{eqnarray} where $a$, $b$, $e$, and $P$ remain \emph{arbitrary functions} of $u$ and $x$. This represents a class of exact solutions of Einstein's equations with a gyrating null matter characterized by the functions \begin{eqnarray} \mathcal{J}(u,x) \rovno -{\textstyle\frac{1}{16\pi}}\,b_{,x}\,, \label{Jf=0sumar}\\ \mathcal{N}(u,x) \rovno {\textstyle\frac{1}{8\pi}}\,P^2 \Big[-{\textstyle\frac{1}{2}}\Big(a_{,xx}+a_{,x}\frac{P_{,x}}{P}\Big) -{\textstyle\frac{1}{2}}b\Big(e_{,x}+\frac{P_{,x}}{P}\,e+ \frac{P_{,u}}{P^3}\Big) \nonumber\\ &&\hspace{14.0mm} +e_{,ux}+ \frac{P_{,x}}{P}\,e_{,u} + \frac{P_{,uu}}{P^3}-2\frac{P_{,u}^2}{P^4}\Big]+2P^2 e\,\mathcal{J} \,. \label{Nf=0sumar} \end{eqnarray} (See Eqs.~(\ref{JKundtf=0}) and (\ref{RuuKundt2f=0}).) They explicitly relate the matter functions to the metric functions of the gravitational field. While the gyratonic part $\mathcal{J}$ is directly determined by $b_{,x}$, the null matter energy profile $\mathcal{N}$ is given primarily by the spatial derivatives of $a$. Whenever $b(u)$ is independent of $x$, the gyrating component vanishes (${\mathcal{J}=0}$) and the spacetimes contains only pure radiation. When ${\mathcal{J}=0=\mathcal{N}}$, the metric (\ref{Kundtmetric_f=0}) is a \emph{vacuum solution}. In fact, it represents Minkowski spacetime. For ${P=1}$, ${e=b=0}$, and ${a=0}$, this is directly the usual form of flat spacetime ${\dd s^2 = \dd x^2-2\,\dd u\dd r}$ in the double-null coordinates (as well as for any $a$ linear in $x$). \item \textbf{The case} ${f\not=0}$ is the most general form of the Kundt metrics with the matter content studied here, \begin{eqnarray} \dd s^2 \rovno \frac{1}{P^2}\, \dd x^2+2\,(e+f\,r)\,\dd u \dd x -2\,\dd u\dd r +\Big[a+b\,r +\big(\Lambda+{\textstyle\frac{1}{4}} \F \big)\,r^2 \Big]\, \dd u^2 \,, \label{Kundtmetric_fnot=0} \end{eqnarray} where ${\F \equiv P^2f^2}$ satisfies Eq.~(\ref{F}), namely \begin{equation} \F_{,x}+ (\F+4\Lambda)f=0 \,. \label{Fsummary} \end{equation} Its general solution can be obtained by treating $F(u,x)$ as an arbitrary generating function, from which we obtain ${f=-\F_{,x}/(\F+4\Lambda)}$ and ${P^2=\F/f^2=\F(\F+4\Lambda)^2/(\F_{,x})^2}$. Alternatively, we prescribe any $f(x,u)$, which generates ${\F=-4\Lambda+g(u)\,\exp[-\int\! f(x,u)\,\dd x]}$ and ${P^2=\F/f^2}$. In this case ${f\not=0}$, the gyratonic matter field is characterized by Eqs.~(\ref{Rux2}) and (\ref{RuuKundt2fnot=0}), namely \begin{eqnarray} \mathcal{J}(u,x) \rovno {\textstyle\frac{1}{16\pi} \Big[\, f_{,u}-b_{,x} -\frac{1}{2}(F+4\Lambda)e-f\,(\ln P)_{,u}\,\Big]}\,, \label{Jfnot=0sumar}\\ \mathcal{N}(u,x) \rovno {\textstyle\frac{1}{8\pi}}\,P^2 \Big[-{\textstyle\frac{1}{2}}a_{,xx}+{\textstyle\frac{1}{2}}a_{,x}\Big(f-\frac{P_{,x}}{P}\Big) -{\textstyle\frac{1}{4}}a\Big(f^2+\frac{4\Lambda}{P^2}\Big) -{\textstyle\frac{1}{2}}b\Big(e_{,x}+\frac{P_{,x}}{P}\,e+ \frac{P_{,u}}{P^3}\Big) \nonumber\\ &&\hspace{2.0mm} +fe\frac{P_{,u}}{P}+e_{,ux}+ \frac{P_{,x}}{P}\,e_{,u} +(\Lambda+{\textstyle\frac{1}{4}}P^2f^2) e^2 + \frac{P_{,uu}}{P^3}-2\frac{P_{,u}^2}{P^4}\Big] +2P^2 e\,\mathcal{J}\,. \label{NKundt2fnot=0sum} \end{eqnarray} The function $\mathcal{J}$ can be prescribed arbitrarily by choosing, e.g., (the spatial derivative of) any free function $b(u,x)$, while the function $\mathcal{N}$ can be prescribed by the free function $a(u,x)$. Moreover, the general expressions (\ref{Jfnot=0sumar}) and (\ref{NKundt2fnot=0sum}) reduce to (\ref{Jf=0sumar}) and (\ref{Nf=0sumar}) when ${f=0}$ (in which case necessarily ${\Lambda=0}$). The metric (\ref{Kundtmetric_fnot=0}) is a vacuum solution if, and only if, ${\mathcal{J}=0=\mathcal{N}}$. It is Minkowski, de~Sitter, or anti-de~Sitter space, according to the sign of the cosmological constant $\Lambda$. \end{itemize} In fact, as already identified in subsection 6.2 of~\cite{PodolskyZofka:2009} for general Kundt geometries in an arbitrary dimension, \emph{these are two distinct invariant subclasses} defined by the property ${f=0}$ and ${f \not=0}$ since the scalar \begin{equation} f^if_i\equiv g^{ij}f_if_j=g^{xx}f_xf_x=P^2f^2=\F \label{invariantF} \end{equation} is gauge invariant. Moreover, this distinction is related to the (non)vanishing of the Ricci rotation (NP) coefficient ${\tau\equiv k_{a;b}\,m^al^b}$. The null triad is given by Eq.~(\ref{triadexpl}), for which ${\tau=\frac{1}{2}Pf}$. Therefore, for ${f\not=0}$, we obtain ${\tau\not=0}$ which means that the privileged null congruence generated by ${\mathbf{k}=\mathbf{\partial}_r}$ has its \emph{internal spatial rotation}. This effect is absent if, and only if, ${f=0}$. Such an invariant classification is analogous to the division into the class of {\it pp\,}-waves and the class of Kundt waves within the Kundt geometry in the standard ${D=4}$. (See section~18.3 of \cite{GriffithsPodolsky:2009}.) \subsection{Gauge freedom} \label{Kundtgauge} Further simplification is achieved by employing the gauge freedom. As in section~\ref{gaugeRT}, we consider the coordinate transformations ${(u,x)\to (u,x')}$ defined by ${x=x(u,x')}$. Then, the Kundt metric (\ref{Kundtmetric_fnot=0}) is transformed to \begin{eqnarray} \dd s^2 \rovno \frac{1}{P^2}\biggl(\frac{\partial x}{\partial x'}\biggl)^2\dd x'^2+2\frac{\partial x}{\partial x'}\biggl[\frac{1}{P^2}\frac{\partial x}{\partial u}+\,(e+f\,r)\biggl]\,\dd u \dd x' \nonumber \\ &&-2\,\dd u\dd r +\Bigg[a+b\,r +\big(\Lambda+{\textstyle\frac{1}{4}} \F \big)\,r^2+\frac{1}{P^2}\biggl(\frac{\partial x}{\partial u}\biggl)^2+2\,(e+f\,r)\frac{\partial x}{\partial u} \Bigg]\, \dd u^2\,. \end{eqnarray} Two natural possibilities are: \begin{itemize} \item \textbf{Setting ${e=0}$} The transformation (\ref{gauge of e}), if we redefine the functions such that ${P/\frac{\partial x}{\partial x'}\to P}$, ${\frac{\partial x}{\partial x'}f\to f}$, ${a-P^2e^2\to a}$, and ${b-2P^2ef\to b}$, locally removes the function $e$ from the metric (\ref{Kundtmetric_fnot=0}), and then the resulting Kundt metric (without a prime on the new coordinate $x'$) is given by \begin{eqnarray} \dd s^2 \rovno \frac{1}{P^2}\, \dd x^2+2\,fr\,\dd u \dd x -2\,\dd u\dd r +\Big[a+b\,r +\big(\Lambda+{\textstyle\frac{1}{4}} \F \big)\,r^2 \Big]\, \dd u^2 \,, \label{Kundtmetric_fnot=0e=0} \end{eqnarray} with \begin{eqnarray} \mathcal{J}(u,x) \rovno {\textstyle\frac{1}{16\pi} \Big[\, f_{,u}-b_{,x} -f\,(\ln P)_{,u}\,\Big]}\,, \label{Jfnot=0sumare=0}\\ \mathcal{N}(u,x) \rovno {\textstyle\frac{1}{8\pi}}\,P^2 \Big[\!\!-{\textstyle\frac{1}{2}}a_{,xx}+{\textstyle\frac{1}{2}}a_{,x}\Big(f-\frac{P_{,x}}{P}\Big) \nonumber \\ &&\qquad\ -{\textstyle\frac{1}{4}}a\Big(f^2+\frac{4\Lambda}{P^2}\Big)-{\textstyle\frac{1}{2}}b\frac{P_{,u}}{P^3} + \frac{P_{,uu}}{P^3}-2\frac{P_{,u}^2}{P^4}\Big]\,. \label{NKundt2fnot=0sume=0} \end{eqnarray} \item \textbf{Setting ${P=1}$} Alternatively, the transformation (\ref{gauge of P}) with redefinition of the functions ${e, f, a, b}$ such that ${P^{-1}\frac{\partial x}{\partial u}+Pe\to e}$, ${Pf\to f}$, ${a+\frac{1}{P^2}(\frac{\partial x}{\partial u})^2+2e\frac{\partial x}{\partial u}\to a}$, and ${b+2f\frac{\partial x}{\partial u}\to b}$, puts the Kundt metric (\ref{Kundtmetric_fnot=0}) into the form with ${P=1}$, namely \begin{eqnarray} \dd s^2 \rovno \dd x^2+2\,(e+f\,r)\,\dd u \dd x -2\,\dd u\dd r +\Big[a+b\,r +\big(\Lambda+{\textstyle\frac{1}{4}} f^2 \big)\,r^2 \Big]\, \dd u^2 \,, \label{Kundtmetric_fnot=0P=1} \end{eqnarray} for which the field equations (\ref{Jfnot=0sumar}) and (\ref{NKundt2fnot=0sum}) are \begin{eqnarray} \mathcal{J}(u,x) \rovno {\textstyle\frac{1}{16\pi} \Big[\, f_{,u}-b_{,x} -\frac{1}{2}(f^2+4\Lambda)e\,\Big]}\,, \label{Jfnot=0sumarp=1}\\ \mathcal{N}(u,x) \rovno {\textstyle\frac{1}{8\pi}} \Big[\!\!-{\textstyle\frac{1}{2}}a_{,xx}+{\textstyle\frac{1}{2}}a_{,x}\,f +{\textstyle\frac{1}{4}} (e^2-a)(f^2+4\Lambda) -{\textstyle\frac{1}{2}}b \, e_{,x} +e_{,ux} \Big] +2 e\,\mathcal{J}\,. \label{NKundt2fnot=0sumP=1} \end{eqnarray} \end{itemize} \subsection{Important subclasses} \label{Kundtsubclasses} Finally, let us mention three important subclasses of the general Kundt metric: \begin{itemize} \item \textbf{pp-waves} These are defined geometrically as admitting a \emph{covariantly constant null vector} field ${\mathbf{k}}$ (and thus are sometimes denoted as CCNV spacetimes). Here we have ${\mathbf{k}=\mathbf{\partial}_r\,}$ and, since ${k_{\alpha;\beta}=\frac12g_{\alpha\beta,r}}$, it follows that all the functions in the metric (\ref{Kundtmetric_fnot=0}) must be independent of the coordinate $r$. Therefore, it takes the Brinkmann form \begin{eqnarray} \dd s^2 \rovno \frac{1}{P^2}\, \dd x^2+2\,e\,\dd u \dd x -2\,\dd u\dd r + a\, \dd u^2 \,, \label{pp-metric} \end{eqnarray} which is Eq.~(\ref{Kundtmetric_f=0}) with ${b=0}$, and necessarily ${\Lambda=0}$. From Eq.~(\ref{Jf=0sumar}) it immediately follows that ${\mathcal{J}=0}$, i.e., there are \emph{no gyratonic pp-waves in Einstein's theory in three dimensions}. Only the standard pure radiation is allowed. Using the gauge ${e=0}$, corresponding to the metric (\ref{Kundtmetric_fnot=0e=0}), this is given by the following single remaining field equation: \begin{equation} \mathcal{N}(u,x) = {\textstyle\frac{1}{8\pi}}\, \Big[-{\textstyle\frac{1}{2}}(P^2 a_{,xx}+PP_{,x} a_{,x}) + \frac{P_{,uu}}{P} -2\frac{P_{,u}^2}{P^2} \Big]\,. \label{ppe=0} \end{equation} \item \textbf{VSI spacetimes} VSI spacetimes are Lorentzian geometries with \emph{vanishing scalar curvature invariants of all orders}. A review and list of references can be found in~\cite{OrtaggioPravdaPravdova:2013}. It was found that all such spacetimes belong to the Kundt class and can be written in 2+1 dimensions in the following canonical form: \begin{equation} \dd s^2=\dd x^2+2\,(e+ f\,r )\,\dd x\dd u-2\,\dd u\,\dd r+(a + b\,r + c\,r^2)\,\dd u^2 \,, \label{metricVSI} \end{equation} where ${a, b, c, e,f}$ are functions of $u$ and $x$. This is a particular subcase of the metric (\ref{Kundtmetric_fnot=0}) in which \begin{equation} P=1\,,\qquad c\equiv\Lambda+{\textstyle\frac{1}{4}} \F=\Lambda+{\textstyle\frac{1}{4}} f^2 \,. \label{flatspace} \end{equation} Again, two subclasses can now be distinguished, namely the case ${f=0}$ and the case ${f \not=0}$. (For more details, see \cite{ColeyFusterHervikPelavas:2006, PodolskyZofka:2009}.) The special subclass of VSI spacetimes with ${f=0}$ and ${b=0=c}$ represent pp-waves (\ref{pp-metric}) with a trivial spatial metric part. In view of Eq.~(\ref{ppe=0}), applying the gauge ${e=0}$, it is a pure-radiation spacetime \begin{equation} \dd s^2=\dd x^2-2\,\dd u\,\dd r+a(x,u)\,\dd u^2 \label{metricVSIe=0=f} \end{equation} with \begin{equation} \mathcal{N}(u,x) = -{\textstyle\frac{1}{16\pi}}\, a_{,xx}\,. \label{VSIppe=0} \end{equation} \item \textbf{CSI spacetimes} Moreover, in \cite{ColeyHervikPelavas:2008}, Lorentzian manifolds in $D=3$ were studied for which all polynomial scalar invariants constructed from the Riemann tensor and its covariant derivatives are \emph{constant}. These CSI geometries are either locally homogeneous or locally Kundt. They thus also belong to the metrics investigated in this contribution. Using our notation, the Kundt CSI geometries of \cite{ColeyHervikPelavas:2008} correspond to the line element \begin{equation} \dd s^2=\dd x^2+2\,(e+ f\,r )\,\dd x\dd u-2\,\dd u\,\dd r+\Big[a + b\,r + \big(\sigma+{\textstyle \frac{1}{4}f^2}\big)\,r^2\Big]\,\dd u^2 \,, \label{metricCSI} \end{equation} with the metric function $f(u,x)$ constrained by \begin{equation} f_{,x}+{\textstyle \frac{1}{2}}f^2=-s \,, \qquad (2\sigma-s)f=-2\alpha \,, \end{equation} where ${\alpha}$, ${\sigma}$, and ${s}$ are constants. Comparing this geometric construction with the explicit solution of Einstein's equations (\ref{Kundtmetric_fnot=0}) and (\ref{Fsummary}), in the gauge (\ref{Kundtmetric_fnot=0P=1}), we identify ${\sigma=\Lambda}$ and we find ${s=2\Lambda}$, that is ${\alpha=0}$. This brings us to the Segre types ${\{3\}}$, ${\{(21)\}}$, and ${\{(1, 11)\}}$ of \cite{ColeyHervikPelavas:2008}. \end{itemize} \newpage \section{Summary and future prospects} In the present paper, we have obtained the general solution of the Einstein equations in 2+1 dimensions in the presence of a cosmological constant for a pure radiation (null dust) or gyratons under a few weak assumptions. Our result is summarized as follows: \vspace{5mm} \noindent \textbf{Theorem~5:} If the spacetime admits a geodesic null vector field, the general local solution of the 3-dimensional Einstein equations with an aligned pure radiation or gyratons in the presence of a cosmological constant consists of the Robinson--Trautman class (\ref{RTmetricsummary}) with the non-zero components of the energy-momentum tensor (\ref{RTTuu}) and (\ref{RTTup}) satisfying Eqs.~(\ref{RTE1}) and (\ref{RTE2}), and the Kundt class (\ref{Kundtmetric_fnot=0}) with the non-zero components of the energy-momentum tensor (\ref{KundtTuu}) and (\ref{KundtTup}) satisfying Eqs.~(\ref{Jfnot=0sumar}), (\ref{NKundt2fnot=0sum}), and (\ref{Fsummary}). There are two invariantly distinct subclasses, namely ${f=0}$ and ${f\ne0}$. \vspace{5mm} As stated above, the general solution in three dimensions is divided into the Robinson--Trautman class and the Kundt class. A characteristic feature of the Robinson--Trautman class is the fact that the function $f(u,x)$ can be non-zero, in spite that it has to be vanishing in four and higher dimensions. This surprising fact shows a richer structure of the 3-dimensional solution than its higher-dimensional counterparts. The spacetime with vanishing $f$ would share the same causal structure and physical interpretation with the higher-dimensional counterparts. Actually, in the case of ${f=0}$, we have constructed the 3-dimensional version of the Kinnersley photon rocket in the absence of gyratons. We leave it for future investigations to clarify a physical interpretation of the solution with non-vanishing $f$. This new functional degree of freedom $f(u,x)$ exists in the Kundt class as well. It would be interesting to compare all these explicit Kundt-type solutions in Einstein's theory with the analogous Kundt solutions of topologically massive gravity (a 3-dimensional theory with a cosmological constant, extending the Einstein gravity to contain higher derivative terms), as given in \cite{ChowPopeSezgin:2009} and summarized in Chapters~18 and~19 of \cite{GarciaDiaz:2017}. Also, comparison to higher-dimensional Kundt spacetimes in the Einstein theory in any dimension ${D\ge4}$, investigated in \cite{PodolskyZofka:2009}, could reveal specific properties of these 2+1 spacetimes. Our general solution could be used to provide answers to open problems in classical general relativity in the framework of 2+1 dimensions. An example of such problems is the cosmic censorship hypothesis~\cite{penrose1969,penrose1979} in the presence of a negative cosmological constant. In order to answer this fundamental question, we need to clarify whether the formation of a naked singularity is generic or not within our solution which describes gravitational collapse of a pure radiation or gyratons from regular initial data. Another possible application of our solution is for canonical quantum gravity. As shown in the ADM formalism of the Einstein equations, general relativity can be treated as a constrained dynamical system. There are two possible ways of canonical quantization for such systems: the Dirac quantization and the reduced phase-space quantization. (See~\cite{carlip} for a textbook of quantum gravity in 2+1 dimensions.) In the latter approach, one first solves the classical constraint equations, namely the Hamiltonian constraint and momentum constraint. Then, we put the solutions of these constraint equations back to the action, perform the spatial integration, and finally quantize the system on the constraint surface. This is a natural way to quantize constrained dynamical systems, however, it is an extremely difficult task to solve the classical constraint equations with full generality. In the present paper, we have succeeded to derive the general solution of the Einstein equations thanks to the very convenient canonical coordinate system adopted. Such an appropriate coordinate system could give us a hint to solve constraint equations and perform a complete reduced phase-space quantization of 3-dimensional gravity. If this is possible in the presence of a negative cosmological constant, in particular, the result would bring us a full description of the quantum states of 3-dimensional black holes. \section*{Acknowledgments} H.~M.~thanks the Institute of Theoretical Physics at the Charles University for a kind hospitality in Prague where this work was initiated, and J.~P.~thanks Hokkai-Gakuen University for a kind hospitality in Sapporo, where this work was completed. This article was supported by the Czech Science Foundation grant No.~GA\v{C}R 17-01625S.

b64554fb231c1fd5565ab4ec35eb55da1cebf67f

\section{\label{sec:Intro}Introduction} Ferronematics, i.e., suspensions of anisotropic ferromagnetic particles dispersed in a nematic liquid crystal (NLC), attract both theoretical \cite{2006_Zadorozhnii, 2007_Zadorozhnii, 2014_Brand, 2014_Zakhlevnykh, 2015_Zakhlevnykh, 2016_Boychuk, 2016_Zakhlevnykh, 2017_Zakhlevnykh} and experimental \cite{2011_Podoliak, 2013_Mertelj, 2014_Mertelj, 2015_Hess, 2017_Mertelj, 2018_Maximean} interest due to their ability to exhibit fluidity due to the solvent as well as macroscopic magnetization due to the colloidal inclusions. The anisotropic nature of the solvent implies broken rotational symmetry as compared to a simple isotropic liquid. The interaction of the anisotropic ferromagnetic colloids with the solvent depends on the orientation of the former with respect to the nematic director of the latter. As a result, the individual magnetic moments of the colloids become effectively trapped around the two possible orientations of the nematic order. Therefore, suitably prepared samples can exhibit a macroscopically ferromagnetic phase. The phase behavior of this complex system follows from its free energy density. The authors of Ref.~\cite{2013_Mertelj} proposed a phenomenological expression thereof which is formulated in terms of the magnetization $\mathbf{M}$ and the nematic director $\mathbf{n}$. A similar expression was derived analytically starting from a microscopic description of the system \cite{2018_Zarubin}: \begin{equation} \label{eq:FN} f(\mathbf{M}, \mathbf{n})= \frac{a}{2}|\mathbf{M}|^2 - \frac{1}{2}\gamma\mu_0(\mathbf{M}\cdot\mathbf{n})^2 - \mathbf{M}\cdot\mathbf{B} \end{equation} where $\mu_0=4\pi\times 10^{-7}$ N/A$^2$ is the permeability of vacuum, $a>0$ is a constant which depends on properties of both the nematic medium and the colloids (for an explicit form see Ref.\ \cite{2018_Zarubin}), $\gamma \geq 0$ measures the coupling between the magnetization and the nematic director, and $\mathbf{B}=B\mathbf{e}_x$ is the external magnetic field. Both $a$ and $\gamma$ are functions of the microscopic coupling constant $c:=WR/K$ where $W$ is the anchoring strength measuring the interaction energy of the NLC per surface area of a single colloid, $R$ is the radius of a colloidal particle modeled as a thin disc (i.e., a disc whose thickness is much smaller than its radius such that the interaction of the rim with the NLC medium can be disregarded), and $K$ is the elastic constant of the NLC within the one-elastic-constant approximation. (The value of $K=3.5\times10^{-12}$ N, corresponding to the twist elastic constant of 5CB \cite{2014_Mertelj}, is used throughout the current study if not specified otherwise.) One of the interesting results of the experiments reported in Ref.~\cite{2013_Mertelj} was the observation of a complex response of the ferronematic slab to a uniform external magnetic field, which depends on the initial state of the sample: If, on one hand, the sample was prepared by quenching the NLC solvent from the isotropic into the nematic phase in the absence of an external magnetic field, the colloids formed \textit{various} magnetic domains within which $\mathbf{M}||\mathbf{n}$. If, on the other hand, the NLC solvent was quenched in the presence of a uniform magnetic field, a \textit{single} domain formed with the entire sample being magnetized in one direction with $\mathbf{M}||\mathbf{n}$. Applying thereafter a uniform external magnetic field opposite to the direction of the magnetization of the single-domain sample yielded a complex, optically observable response of a nonuniform director field. Here we focus on the case of single-domain samples. So far such samples have been thoroughly investigated theoretically in the situation in which the external magnetic field is applied \textit{perpendicular} to the initial magnetization of the sample \cite{2006_Zadorozhnii,2007_Zadorozhnii,2015_Zakhlevnykh}. Also the dynamics of such a configuration was investigated experimentally \cite{2017_Potisk,2018_Potisk,2018_Sebastian}. A \textit{thresholdless} distortion of the nematic was observed. Moreover, the authors of Refs.\ \cite{2006_Zadorozhnii,2007_Zadorozhnii,2015_Zakhlevnykh} considered infinitely strong anchoring at the walls and external magnetic fields up to magnitudes strong enough to directly interact with the magnetically anisotropic NLC molecules. Our aim is to investigate in detail the behavior of a monodomain sample exposed to a uniform external magnetic field which is applied in the direction \textit{antiparallel} to the initial magnetization. Moreover, we consider only magnetic fields of small ($\leq 25$ mT) amplitudes, such that the direct magnetic field influence on the NLC can be neglected. It was observed experimentally \cite{2013_Mertelj, 2014_Mertelj} that, like for common ferromagnets, ferronematics exhibit hysteresis in the magnetic properties as a function of the external field. Moreover, the critical field, i.e., the magnetic field strength at which the ferronematic becomes distorted (for a more precise definition see Sec.~\ref{subsec:I} below), is another feature of the ferronematic sample. We obtain the hysteresis curves numerically by using a conjugate-gradient technique in order to minimize an appropriate free energy functional of the ferronematic in slab geometry. From the hysteresis curves one can infer the value of the critical magnetic field as function of the parameters of the model and compare them with the expressions derived in Ref.~\cite{2013_Mertelj}. In our previous study \cite{2018_Zarubin} we derived the dependence of the coupling parameter $\gamma$ on the microscopic coupling $c$ which in turn depends on the particle size. Having obtained the critical field as a function of $\gamma$ allows us to relate it to the size of the colloids and therefore one can potentially tune the value of the critical field by tuning the mean of the size distribution of the particles used. It turns out that the switching process of the considered ferronematic slab from one phase to the other takes place according to one of two possible scenarios which we shall discuss. In the first scenario, regions nucleate near the system walls in which the magnetization is flipped, whereas the nematic director is kept in place by the walls. In the second scenario, the nematic director follows the magnetization, i.e., it makes a full rotation by 180$^\circ$, everywhere throughout the sample. Finally, we propose a novel technique which can be used, e.g., in data storage devices. It is based on magnetic fields which control the magneto-optical properties of ferronematic cells, and thus allows one to switch between magnetized and demagnetized states by applying a uniform magnetic field of suitable orientation. The paper is organized as follows. In Sec.\ \ref{sec:Theory} we introduce the free energy functional in order to describe the system and the numerical method to minimize it. Section \ref{subsec:I} contains the description of the first of the two switching mechanism as well as the results concerning the critical magnetic field and its dependence on the parameters of the model. In Sec.\ \ref{subsec:II} we present the second switching mechanism and provide a map which relates the parameters of the model to the character of the switching. In Sec.\ \ref{subsec:TwoWalls} we report that a combination of the two mechanisms leads to a sample the magnetization of which can be reversibly switched off by using the external magnetic field. The role of the phenomenon of segregation is discussed in Sec.~\ref{sec:Segregation}. In Sec.\ \ref{sec:Discussion} we conclude by discussing the main results. \section{\label{sec:Theory}Numerical model} \begin{figure}[t!] \centering \includegraphics[scale=1]{./slab.eps} \caption{Sketch of a ferronematic in a slab of width $D$. The $x$-direction corresponds to the (lateral) easy direction of the liquid crystalline medium. Due to lateral translational invariance, all profiles depend only on the normal coordinate $z$. The nematic director $\mathbf{n}(z)$, the magnetization $\mathbf{M}(z)$, and the external magnetic field $\mathbf{B}$ are parallel to the $x$-$z$-plane and their directions with respect to the positive $x$-direction are described by the angles $\varphi(z)$, $\theta(z)$, and $\psi$, respectively.} \label{fig:Slab} \end{figure} We consider the experimental setup studied in Ref.~\cite{2013_Mertelj}. It consists of a ferronematic confined by two parallel and planar walls at a distance $D$ with $\mathbf{e}_x$ as the so-called easy axis at both surfaces, which imposes a parallel orientation of the nematic director there (see Fig.~\ref{fig:Slab}). We assume that the sample was prepared in the presence of a homogeneous external magnetic field $\mathbf{B}=B\mathbf{e}_x$ in the direction parallel to the easy axis $\mathbf{e}_x$ of the walls (i.e., $B>0$), thus producing a single domain of the ferromagnetic phase. In the following the effect of applying an external magnetic field in the direction opposite to the one used during this preparation (i.e., $B<0$) is investigated numerically. The system is described by two spatially varying fields: the director field $\mathbf{n}(\mathbf{r})$ and the magnetization field $\mathbf{M}(\mathbf{r})$. We assume the absolute value of the magnetization is spatially constant, $|\mathbf{M}(\mathbf{r})|=\text{const}=m\rho_\text{iso}$ (where $m$ is the absolute value of the magnetic moment of a single colloid, here taken to be $3\times 10^{-18}$~A~m$^2$ \cite{2013_Mertelj}, and $\rho_\text{iso}$ is the number density of the colloids dispersed in the isotropic phase of the liquid crystal during the preparation of the sample (see Refs.\ \cite{2013_Mertelj} and \cite{2018_Zarubin}), i.e., segregation effects are assumed to be small \cite{footnote1}. (For a discussion of the possible influence of segregation see Sec.\ \ref{sec:Segregation}.) Due to translational invariance in the lateral $x$-$y$-plane all physical quantities depend on the normal coordinate $z$ only. We consider that both $\mathbf{n}(z)$ and $\mathbf{M}(z)/(m\rho_\text{iso})$ are parallel to the $x$-$z$-plane \cite{footnote2}, so that they can be described by the angles $\varphi(z)$ and $\theta(z)$, respectively (see Fig.~\ref{fig:Slab}). The initial configuration is given by the uniform profiles $\varphi(z)=0$ and $\theta(z)=0$, which corresponds to an unstable state when a uniform magnetic field $\mathbf{B}$ is applied in the direction $\psi=\pi$ (see Fig.~\ref{fig:Slab}). In terms of the profiles $\varphi$ and $\theta$ the free energy functional of the system is given by \begin{equation} \label{eq:SlabF} \frac{1}{S}\beta\mathcal{F}[\varphi,\theta] = \beta F_\text{ferr}[\varphi,\theta] + \beta F_\text{elas}[\varphi] + \beta F_\text{surf}[\varphi], \end{equation} where $S$ is the surface area of one of the glass plates, $\beta:=1/(k_\text{B}T)$, \begin{equation} \label{eq:Fferr} \beta F_\text{ferr}[\varphi,\theta]= \int_0^D\!\!\ensuremath{\mathop{}\!\mathrm{d}} z\ \beta f(\mathbf{M}(z),\mathbf{n}(z)) \end{equation} with the free energy density $f$ given by Eq.~\eqref{eq:FN}, which is the contribution due to the \textit{ferr}onematic, \begin{equation} \label{eq:NLCEl} \beta F_\text{elas}[\varphi]= \frac{1}{2}\beta K\int_0^D\!\!\ensuremath{\mathop{}\!\mathrm{d}} z\ \left( \frac{\ensuremath{\mathop{}\!\mathrm{d}}\varphi(z)}{\ensuremath{\mathop{}\!\mathrm{d}} z}\right)^2, \end{equation} is the contribution due to the elastic distortions of the liquid crystal, and \begin{equation} \label{eq:WallC} \beta F_\text{surf}[\varphi]= -\frac{1}{2}\beta W_\text{wall}\left(\cos(\varphi(0))^2+\cos(\varphi(D))^2\right) \end{equation} is the contribution due to the coupling of the liquid crystal to the glass plates. The equilibrium profiles $\varphi(z)$ and $\theta(z)$ correspond to the minimum of the free energy in Eq.~\eqref{eq:SlabF}, which has been determined numerically by using the Fletcher-Reeves-Polak-Ribiere general function minimization algorithm \cite{2002_Press}. The absolute value $|\mathbf{M}|$ of the magnetization is assumed to have a constant value $m\rho_\text{iso}$ and is taken to be independent of the external field $\mathbf{B}$ throughout Sec.\ \ref{sec:Results}. For the discussion of the problem in the case of a spatially varying $|\mathbf{M}|$ see Sec.\ \ref{sec:Segregation}. We have used the following parameter values: $K=3.5\times 10^{-12}$~N, $\rho_\text{iso}=1.5\times 10^{19}$~m$^{-3}$, $m=3\times 10^{-18}$~A~m$^2$, and $T=300$~K. These values are consistent with experimental data reported in Refs.\ \cite{2013_Mertelj,2014_Mertelj}. The thickness $D$ of the slab is taken to be 20~$\mu$m throughout Sec.\ \ref{sec:Results}. Thicker slabs are considered in Sec.\ \ref{sec:Segregation} \cite{footnote7,1995_Burylov}. \section{\label{sec:Results}Results} \subsection{\label{subsec:I}Switching mechanism I and the critical field} The experiments in Refs.~\cite{2013_Mertelj, 2014_Mertelj} demonstrate that, upon applying a uniform external magnetic field to the setup described in Sec.~\ref{sec:Theory} (see also Fig.~\ref{fig:Slab}), there is a nonvanishing critical magnetic field strength $\mathbf{B}_\text{cr}=B_\text{cr}\mathbf{e}_x$ such, that for $B<B_\text{cr}<0$ ($B>B_\text{cr}>0$) in the case of an initial magnetization pointing along the positive (negative) $x$-direction, elastic distortions of the liquid crystal matrix occur. The occurrence of such a critical magnetic field strength $B_\text{cr}$ can be explained qualitatively in terms of a diverging relaxation time of the fluctuations of the nematic director field $\mathbf{n}$ (or $\varphi$) (see Ref.~\cite{2013_Mertelj}). Here we aim at exploring the dependence of $B_\text{cr}$ on the coupling constant $\gamma$ and the wall anchoring strength $W_\text{wall}$ \cite{footnote5}. Moreover, we are also interested in the intermediate metastable states preceding the switches of the magnetization field $\mathbf{M}$ to the ground state parallel to the external field $\mathbf{B}$. \begin{figure}[t] \centering \includegraphics[scale=1]{./hyst.eps} \caption{Hysteresis of the spatially averaged magnetization $\mathcal{M}$ along the $x$-direction (see Eq.~\eqref{eq:PrettyM}) of the ferronematic slab as function of the $x$-component $B$ of the external magnetic field $\mathbf{B}$. Negative values of $\mathcal{M}$ or $B$ correspond to orientations in the negative $x$-direction (see Fig.~\ref{fig:Slab}). For initially saturated samples with $\mathcal{M}=1$ (i.e., magnetized in the positive $x$-direction) there is a nonvanishing critical magnetic field $B_\text{cr}<0$ (indicated in the plot) such that for $B\in[B_\text{cr},0]$ the magnetization $\mathcal{M}$ does not respond to the external field. Upon increasing the field in the negative $x$-direction (i.e., for $B<B_\text{cr}$, left branch of the loop) the system evolves through a series of qualitatively distinct metastable states (red circles) corresponding to the profiles displayed in Fig.~\ref{fig:MetaS} and eventually reaches saturation along the negative $x$-direction (i.e., $\mathcal{M}=-1$). Gradually lowering the magnitude $|B|$ of the magnetic field does not influence the magnetization of the sample (i.e., for $B<0$, see the part of the loop along $\mathcal{M}=-1$). After $B=0$ is crossed, the situation is identical to the one described above up to a change of sign of $B$ and $\mathcal{M}$ (right branch of the loop). Dotted vertical lines separate regions of qualitatively different metastable states. Note that state (1) corresponds to $B_\text{cr}<B<0$. $W_\text{wall}=3.1\times 10^{-5}$~J/m$^2$, $\gamma=240$, $K=3.5\times 10^{-12}$~N, $\rho_\text{iso}=1.5\times 10^{19}$~m$^{-3}$, $m=3\times 10^{-18}$~A~m$^2$, and $T=300$~K} \label{fig:Hyst} \end{figure} \begin{figure*}[t!] \centering \includegraphics[scale=0.85]{./profiles.eps} \caption{Qualitatively distinct metastable states of the ferronematic along the hysteresis curve in Fig.~\ref{fig:Hyst} in terms of the profiles $\varphi(z)$ of the nematic director field (black lines) and $\theta(z)$ of the local magnetization (red lines). State (1) (see the numbered states in Fig.\ \ref{fig:Hyst}): For magnetic fields $\mathrm{B}$ along the $x$-direction with components $B > B_\text{cr}$ the saturated profiles $\mathcal{M}=1$ are unperturbed by $B$. State (2): For magnetic field components $B < B_\text{cr}$ both profiles deviate significantly from the saturated ones. Note that the profiles in panel (1) correspond to a saddle point of the free energy so that spontaneous symmetry breaking can occur from state (1) to state (2) (i.e., the magnetization and the director start to rotate in either clockwise or counterclockwise direction, see Fig.\ \ref{fig:Slab}). The equally probable profiles which correspond to the same projection $\mathcal{M}$ and thus correspond to the same points (1)-(4) on the hysteresis loop (see Fig.\ \ref{fig:Hyst}), are obtained by the map $(\varphi,\theta)\mapsto(-\varphi,-\theta)$ for the profiles (a)-(d). State (3): Upon further increasing the magnetic field strength in the negative $x$-direction, layers form near the walls where the magnetization interpolates between the direction along ($\theta$ large) and opposite ($\theta$ small) to the magnetic field. The coupling of the magnetization and the director causes $\theta>\pi$ within the layers. State (4): For even stronger magnetic fields the entire slab (besides thin layers near the walls, which require yet higher fields to switch) is magnetized along the direction of the magnetic field, i.e., $\theta=\pi$. The values of the system parameters are the same as in Fig.\ \ref{fig:Hyst}.} \label{fig:MetaS} \end{figure*} Figure \ref{fig:Hyst} shows the hysteresis curve of the spatially averaged projection of the magnetization $\mathcal{M}$ onto the $x$-axis, \begin{equation} \label{eq:PrettyM} \mathcal{M}:=\frac{1}{D}\int_0^D\!\!\ensuremath{\mathop{}\!\mathrm{d}} z\,\,\cos\theta (z), \end{equation} as function of the component of the external magnetic field $\mathbf{B}$ along the $x$-axis for the particular choice of the coupling constant $\gamma=240$ (which corresponds to a value of the microscopic coupling constant $c\approx 0.035$ \cite{2018_Zarubin}) and of the wall anchoring strength $W_\text{wall}=3.1\times 10^{-5}$ J/m$^2$; this choice of the parameters is reasonable in the context of available experimental data (see Refs.\ \cite{2013_Mertelj}, \cite{2014_Mertelj}, and \cite{2018_Zarubin}). In order to investigate the switching process of the magnetization in the ferromagnetic phase as function of the external magnetic field, Fig.~\ref{fig:MetaS} displays the orientation profiles $\varphi(z)$ and $\theta(z)$ for a series of intermediate metastable states corresponding to the hysteresis loop in Fig.~\ref{fig:Hyst}. For initially saturated samples with $\mathcal{M}=1$, the magnetization does not change significantly in the presence of $x$-components of the magnetic field $B > B_\text{cr}$, whereas for $B < B_\text{cr}$ there is a noticable deviation of the $x$-component of the spatially averaged magnetization $\mathcal{M}$ from the initial saturation value (see Fig.~\ref{fig:Hyst}). This defines a critical magnetic field strength $B_\text{cr} < 0$. For $B > B_\text{cr}$ both the magnetization and the nematic director field profiles, i.e., $\theta(z)$ and $\varphi(z)$, de facto do not deviate from the saturated ones (see Fig.~\ref{fig:MetaS}(a)). While the magnetization tends to align with the external magnetic field, due to the interaction described by the coupling constant $\gamma$ it is dragging the nematic director field along. At $B=B_\text{cr}$ the metastable state corresponding to the unperturbed nematic director becomes unfavorable compared to the metastable state corresponding to the perturbation induced in the interior of the slab (see Fig.~\ref{fig:MetaS}(b)). We note that the saturated sample with $\mathcal{M}=1$ in a magnetic field in the negative $x$-direction with $B=B_\text{cr}$ corresponds to a saddle point of the free energy so that spontaneous symmetry breaking induced by fluctuations leads to perturbations of the magnetization orientation profile $\theta(z)$ with either $\theta(z) > 0$ or $\theta(z)<0$. In the following we focus only on the first case, while the second, conjugated one, follows from changing signs. It is the perturbed nematic director field which manifests itself as a brightening of the sample when viewed with crossed polarizers as in the experiment reported in Ref.~\cite{2013_Mertelj}, and it occurs only due to the coupling of the magnetization field $\mathbf{M}$ to the nematic director field $\mathbf{n}$. The external magnetic field imposes a torque onto the magnetization field which in turn leads to a torque onto the nematic director field. The latter is opposing the torque generated by the walls of the cell and which is transmitted due to the elasticity of the NLC (Eq.~\eqref{eq:NLCEl}). Upon increasing the external magnetic field the variations inside the slab become more and more pronounced for both the magnetization and the nematic director field. However, in the case of soft anchoring \cite{footnote6} (distinct from the case of infinitely strong anchoring, see Ref.~\cite{1970_Brochard}) at the surface of the colloid, the angle between the magnetization and the nematic director is nonzero for $B < B_\text{cr}$ (see Figs.~\ref{fig:Slab} and \ref{fig:MetaS}), i.e., $\theta\neq\varphi$. \begin{figure}[t!] \centering \includegraphics[scale=1]{./vectors.eps} \caption{Explicit magnetization field (red arrows) and nematic director field (black rods) of the pretransitional metastable state (3) (see Figs.~\ref{fig:Hyst} and \ref{fig:MetaS}(c)) of the ferronematic in between two glass walls (hatched regions). While in the interior of the slab (white region) the magnetization field (red arrows) is, to a large extent, aligned with the external magnetic field in the negative $x$-direction, close to the walls (grey regions) its orientation interpolates between the configurations being parallel and antiparallel to the magnetic field. This is due to the coupling to the nematic director field (black rods), which is aligned along the easy direction at the walls. Upon further increasing the strength of the external magnetic field in the negative $x$-direction ($B < B_\text{cr} < 0$), the grey regions widen and eventually produce an almost uniformly magnetized sample (see Fig.~\ref{fig:MetaS}(d)). The pale red regions denote thin layers very close to the walls which switch last.} \label{fig:Nucleation} \end{figure} Before reaching the magnetic phase with the sample being magnetized along the field in negative $x$-direction, the system passes through the metastable state (3) in Fig.~\ref{fig:Hyst}, in which the magnetization profile $\theta(z)$ has a peculiar form (see Figs.~\ref{fig:MetaS}(c) and \ref{fig:Nucleation}). In this metastable state the magnetization in the interior of the sample is aligned along the magnetic field. Within certain transition regions close to the walls the orientation of the magnetization interpolates between the direction along the magnetic field and the opposite direction. These transition regions occur because the magnetization is coupled to the nematic director field, which is aligned along the easy axis ($\varphi =0$) at the walls. The width of these transition regions grows upon increasing the external magnetic field strength so that eventually the minimum of the free energy given by Eq.~\eqref{eq:SlabF} corresponds to the magnetization being oriented parallel to the external magnetic field in the entire slab ($\theta=\pi$). Concerning the transition regions following observations can be made: (i) Due to the soft coupling between the colloids and the nematic director field of the NLC the ground state, in which the entire sample is magnetized along the external field, is attained by means of ``switching'' the magnetization locally, i.e., by inverting the direction of the magnetization (and thus of the orientation of the magnetic colloids) without the simultaneous rotation of the local nematic director field. (ii) Layers of the incipient ferronematic phase are nucleated in the regions close to the walls due to the interplay between elastic and magnetic torques and because the coupling energy is invariant with respect to an inversion of the magnetization. \begin{figure}[t!] \centering \includegraphics[scale=1]{./phase.eps} \caption{The dependence of the critical magnetic field strength $B_\text{cr}$ on the coupling constant $\gamma$ and on the anchoring strength $W_\text{wall}$ at the walls. The color code denotes values of $B_\text{cr}$ measured in units of mT. The thin black curves are contour lines; from left to right $B_\text{cr}=1.2,1.3,1.4,1.5,1.6,1.65,1.7,1.75$ mT. The thick black lines correspond to the cuts in Fig.~\ref{fig:Cuts}(a) and (b) for $\gamma\approx 195$ and $W_\text{wall}=4\times 10^{-5}$ J/m$^2$, respectively (the values of the remaining parameters are the same as in Fig.\ \ref{fig:Hyst}).} \label{fig:Phase} \end{figure} Naturally the question arises concerning the dependences of the critical magnetic field strength $B_\text{cr}$ on the coupling constant $\gamma$ and on the wall anchoring $W_\text{wall}$. Here we define $B_\text{cr}$ as the magnetic field strength at which the spatially averaged magnetization $\mathcal{M}$ equals $0.97$; note that $\mathcal{M}=1$ in the saturated state. This definition differs from the one used in Refs.~\cite{2013_Mertelj, 2014_Mertelj}, where $B_\text{cr}$ is defined as the magnetic field strength at which the relaxation time of thermal fluctuations of the direction of $\mathbf{n}$ diverges. Here we do not consider dynamic processes, instead we propose the above alternative definition of $B_\text{cr}$. Obviously, the choice of $0.97$ for the threshold value contains some degree of arbitrariness. However, as can be inferred from the steep slope of the hysteresis loop close to state (1) in Fig.~\ref{fig:Hyst}, no significant changes are expected to occur by choosing different threshold values not too much less than unity. Figure \ref{fig:Phase} shows the dependence of $B_\text{cr}$ on the coupling constant $\gamma$ and the wall anchoring strength $W_\text{wall}$. One can infer from Fig.~\ref{fig:Phase} that for fixed $W_\text{wall}$ the critical field $B_\text{cr}$ increases upon increasing $\gamma$. Indeed, for a given value of $W_\text{wall}$, the magnetization field is aligned with the nematic director field, the rotation of which is opposed by the torque imposed by the walls. The system sustains the alignment for increasing external magnetic field strengths which in their turn are due to an increasing strength of the coupling $\gamma$ between the magnetization and the nematic director field. One can also infer from Fig.~\ref{fig:Phase} that the critical field $B_\text{cr}$ depends rather weakly on the wall anchoring strength $W_\text{wall}$: Within the considered range of the anchoring strengths, $W_\text{wall}\in[0.5\times 10^{-5},10\times 10^{-5}]\mathrm{J/m^2}$, for fixed $\gamma$ the critical field strength $B_\text{cr}$ varies by $\approx 0.1\text{ mT}$. For large values of $W_\text{wall}$ the critical field reaches a plateau (see Fig.\ \ref{fig:Cuts} (a)) and it becomes independent of the wall anchoring $W_\text{wall}$. \begin{figure}[t!] \centering \includegraphics[scale=1]{./cuts.eps} \caption{Dependence of the critical magnetic field strength $B_\text{cr}$ (a) on the wall anchoring strength $W_\text{wall}$ (along the thick vertical black line in Fig.~\ref{fig:Phase}, i.e., for $\gamma\approx 195$), (b) on the coupling constant $\gamma$ (present approach and Eq.\ \eqref{eq:BCr} along the thick horizontal black line in Fig.~\ref{fig:Phase}, i.e., for $W_\text{wall}=4\times 10^{-5}$ J/m$^2$), and (c) on the microscopic coupling constant $c$ (see see below Eq.\ \eqref{eq:FN}, $W_\text{wall}=4\times 10^{-5}$ J/m$^2$). The values of the remaining parameters are the same as in Fig.\ \ref{fig:Hyst}.} \label{fig:Cuts} \end{figure} Figure~\ref{fig:Cuts} illustrates these trends of the critical magnetic field strength $B_\text{cr}$ via cuts in Fig.~\ref{fig:Phase} for $\gamma\approx 195$ (see Fig.~\ref{fig:Cuts}(a)) and $W_\text{wall}=4\times 10^{-5}$ J/m$^2$ (see Fig.~\ref{fig:Cuts}(b)). Moreover, Fig.~\ref{fig:Cuts}(c) displays the dependence of $B_\text{cr}$ on the microscopic coupling constant $c$ (see below Eq.\ \eqref{eq:FN}). Finally, Fig.~\ref{fig:Cuts}(b) compares the values of the critical magnetic field $B_\text{cr}$ as defined in the present approach (circles) with the corresponding expression given in Ref.~\cite{2013_Mertelj}, \begin{equation} \label{eq:BCr} B^\text{\cite{2013_Mertelj}}_\text{cr}= \frac{\pi^2 \gamma \mu_0 K M_\text{s}}{\pi^2 K+\gamma\mu_0 M_\text{s}^2D^2}, \end{equation} where $M_\text{s}:=m\rho_\text{iso}$ is the magnetization of the saturated sample, obtained in the limit $W_\text{wall}\to\infty$. (Here, we consider the particular NLC used for the experiments in Refs.\ \cite{2013_Mertelj,2014_Mertelj} and therefore a fixed value of the elastic constant $K$. Although the variation of the expression given here as function of the elastic constant of the NLC is interesting, we leave this issue for future work due to the highly non-trivial occurrences of $K$.) Remarkably, $B_\text{cr} \approx B^\text{\cite{2013_Mertelj}}_\text{cr}$ appears to hold although the two definitions of the critical magnetic field strength differ and although $B^\text{\cite{2013_Mertelj}}_\text{cr}$ in Eq.~\eqref{eq:BCr} does not take the dependence on $W_\text{wall}$ into account. \subsection{\label{subsec:II}Switching mechanism II} In Sec.~\ref{subsec:I} we revealed a mechanism of switching the sample magnetization in the case of large values of the wall anchoring strength $W_\text{wall}$. Here we show another possible mechanism which corresponds, however, to small values of $W_\text{wall}$. For suitable combinations of $\gamma$ and $W_\text{wall}$ the magnetization field is able to drag the nematic director field along, thereby inducing a large change of the angle $\varphi$ compared with the initial configuration. For small values of $W_\text{wall}$ the anchoring at the wall is so weak, that the nematic director field at the surface of the walls is able to deviate from the direction of the easy axis and to rotate with the magnetic field due to the coupling between the magnetization and the nematic director. \begin{figure*}[t!] \centering \includegraphics[scale=0.85]{./profiles_II.eps} \caption{Final two stages of switching mechanism II for the magnetization in an initially oppositely oriented external magnetic field for $\gamma=240$ and $W_\text{wall}=0.1\times 10^{-5}$~J/m$^2$ (the values of the remaining parameters are the same as in Fig.\ \ref{fig:Hyst}). It is characterized by a nematic director field with significant elastic distortions throughout the sample. Even at the surface rotations of the nematic director with respect to the easy axis occur due to the weak anchoring at the walls (note $\varphi\neq 0$ and $\theta\neq 0$ in the left panel). For sufficiently strong magnetic fields both the magnetization field and the nematic director rotate by the angle $\pi$ (see right panel). The panels are denoted by ``(3)'' and ``(4)'' in order to make the comparison easier with the corresponding panels in Fig.~\ref{fig:MetaS} describing switching mechanism I. Note that $\theta\neq\varphi$ even directly at the walls.} \label{fig:MetaSII} \end{figure*} We have performed calculations analogous to those described in Sec.~\ref{subsec:I} but for small values of the anchoring strength $W_\text{wall}$ at the sample walls. It turns out that for values $W_\text{wall} <0.5\times 10^{-5}$~J/m$^2$ there are corresponding values of the coupling constant $\gamma$ which produce a switching mechanism which is qualitatively different from the one described in the previous section. In this mechanism the early stages of the switching are similar to those described in the previous section (see Fig.~\ref{fig:MetaS}(1) and (2)). However, the subsequent stage, as displayed in Fig.~\ref{fig:MetaSII}(3), is qualitatively different in the sense that the system does not separate in distinct spatial regions with different orientations of the magnetization (compare with Fig.~\ref{fig:MetaS}(3)). Obviously, the change in character of the switching mechanism is directly related to the anchoring at the walls being too weak to prevent the liquid crystal from rotating along with the magnetization field. This weakness is revealed also by nonzero angles $\varphi$ and $\theta$ at the walls. \begin{figure}[t!] \centering \includegraphics[scale=1]{./switch.eps} \caption{Regions of dominance for the two switching mechanisms between two ferromagnetic phases in terms of the coupling constant $\gamma$ and the wall anchoring strength $W_\text{wall}$, which is the same for both walls. (The values of the remaining parameters are the same as in Fig.\ \ref{fig:Hyst}.) The region denoted as "I" (full red circles) corresponds to switching mechanism I in which the magnetization leaves the nematic director field behind (see Sec.~\ref{subsec:I}). The region denoted as "II" (full blue circles) corresponds to switching mechanism II in which the nematic director field is weakly coupled to the sample walls and therefore is able to follow the magnetization.} \label{fig:Switch} \end{figure} At such low anchoring strengths the torque imposed on the liquid crystal director by the walls cannot compete with the drag imposed by the rotating magnetization field and it is energetically more favorable for the director at the walls to flip its orientation. The regions of dominance for the two switching mechanisms are depicted in Fig.~\ref{fig:Switch}, where a map spanned by the coordinates ($\gamma$, $W_\text{wall}=W_\text{wall}^{(1)}=W_\text{wall}^{(2)}$) marks region "I" (corresponding to switching mechanism I, for which the director field returns back to its initial configuration upon increasing the external magnetic field and reaching the saturation of the magnetization in the direction of the field) and region "II" (corresponding to switching mechanism II, for which the director field follows the magnetization). Within our numerical approach switching mechanism II is observed only for values of the wall anchoring $W_\text{wall}$ which are significantly smaller than the one estimated from the experiment \cite{2014_Mertelj}, i.e., $W^\text{exp}_\text{wall}\approx (3.40\pm 0.11)\times10^{-5}$~J/m$^2$. Therefore, we expect switching mechanism I to be the one realized experimentally. The segregation of colloids might play an important role. Segregation amounts to a redistribution of the colloids dispersed in the liquid crystal. This effect is caused by the opportunity to lower the free energy of the magnetic colloids in an external magnetic field by migrating away from regions in which the liquid crystal prevents alignment along the external field. Thereby energy is gained by accomplishing alignment at the expense of the entropic contribution due to denser packing which is proportional to $\rho\log\rho$, where $\rho$ is the local number density of the colloids (see Ref.~\cite{2018_Zarubin}). Although segregation is neglected in Sec.\ \ref{sec:Results}, we nonetheless do not expect segregation to influence our results qualitatively (see Sec.\ \ref{sec:Segregation}). On the other hand, quantitative changes are conceivable, i.e., the map in Fig.~\ref{fig:Switch} might be affected. \subsection{\label{subsec:TwoWalls}Confining walls with different anchoring strengths} In this section we study a combination of switching mechanisms I and II described in Secs.~\ref{subsec:I} and \ref{subsec:II}, respectively, by considering a strong anchoring strength at one wall and a weak one at the other, sharing the same easy axis. Thus the system is described by three parameters (assuming $a$, $T$, $K$, $D$, $m$, and $\rho_\text{iso}$ to be fixed): (i) coupling constant $\gamma$, (ii) the anchoring strength at one of the walls, and (iii) the ratio of the anchoring strengths at the two walls. Note that introducing different but still uniform and parallel anchorings at the walls renders the system still effectively one-dimensional along the $z$-direction and hence from a numerical point of view its complexity does not change. Adding a third parameter (i.e., the ratio of the anchoring strengths at the two walls $w:=W_\text{wall}^{(1)}/W_\text{wall}^{(2)}$) introduces a third dimension to the map considered in Fig.~\ref{fig:Switch}. The cut of this three-dimensional map along $w=1$ produces the two-dimensional map shown in Fig.~\ref{fig:Switch}. While the two-dimensional map in Fig.~\ref{fig:Switch} exhibits only two switching regions (I and II), in the three-dimensional parameter space the situation can be more involved. It is reasonable to expect that if both $W_\text{wall}^{(1)}$ and $W_\text{wall}^{(2)}$ become infinitely strong, the magnetization of the sample switches according to mechanism I. If, on the other hand, $W_\text{wall}^{(1)},W_\text{wall}^{(2)}\to 0$ one can expect that the magnetization of the sample switches according to mechanism II. However, pairs $(W_\text{wall}^{(1)},W_\text{wall}^{(2)})$ can exist such that the magnetization in the vicinity of one wall would switch according to mechanism I and the magnetization in the vicinity of the other wall would switch according to mechanism II. Thus, the three-dimensional parameter space consists of \textit{three} regions: dominance of mechanism I, II, and their combination. \begin{figure}[t!] \centering \includegraphics[scale=1]{./two_walls.eps} \caption{Orientation profiles $\varphi(z)$ (nematic director, black line) and $\theta(z)$ (magnetization, red line) for $B \leq -12$ mT (with the sample initially magnetized along the positive $x$-direction). There is a gradual variation of the nematic director from one wall to the other. The parameters are chosen as $\gamma=240$, $W^{(1)}_\text{wall}=10^{-5}$~J/m$^2$ (see $z=0$), and $W^{(2)}_\text{wall}=0.1\times10^{-5}$~J/m$^2$ (see $z=D$), $K=9\times 10^{-12}$ N \cite{footnote3}; the values of the remaining parameters are the same as in Fig.\ \ref{fig:Hyst}.} \label{fig:TwoWalls} \end{figure} The line separating the two regions in Fig.~\ref{fig:Switch} becomes a two-dimensional manifold in the three-dimensional parameter space ($W_\text{wall}^{(1)},w,\gamma$). In the vicinity of the plane $w=1$ this manifold, which is separating the regions ``I'' and ``II'', is considered to be perpendicular to the plane $w=1$ and only the two regions ``I'' and ``II'' occur. Therefore, if one would like to find a point ($W_\text{wall}^{(1)},w,\gamma$) that belongs to the region, which corresponds to the combination of the two switching mechanisms, it is necessary to pick the value of $w$ significantly different from 1. To this end, for fixed $\gamma$ it seems to be natural to take as an estimate the anchoring at one of the walls from region ``I'' in Fig.~\ref{fig:Switch} and the anchoring at the other wall from region ``II'' in Fig.~\ref{fig:Switch}. It has turned out that for $D=20$~$\mu$m (for a discussion concerning larger values of $D$ see Sec.\ \ref{sec:Segregation}) the combination $W_\text{wall}^{(1)}=10^{-5}$~J/m$^2$ and $W_\text{wall}^{(2)}=0.1\times 10^{-5}$~J/m$^2$ (i.e., $w=10$) yields profiles $\varphi(z)$ (nematic director) and $\theta(z)$ (magnetization) which consist of one part due to switching mechanism I and another part due to switching mechanism II. Figure \ref{fig:TwoWalls} shows the actual profiles (i.e., for magnetic field strengths $B \leq -12$ mT and for the initial magnetization pointing into the positive $x$-direction) of the magnetization and of the nematic director field for $\gamma=240$, $W^{(1)}_\text{wall}=10^{-5}$ J/m$^2$, and $W^{(2)}_\text{wall}=0.1\times10^{-5}$ J/m$^2$, where superscript $(1)$ denotes the wall at $z=0$ and superscript $(2)$ denotes the wall at $z=D$. Both the orientation field $\varphi(z)$ of the nematic director and the orientational field $\theta(z)$ of the magnetization have a nontrivial form. The nematic director field profile exhibits a smooth rotation by an angle of $\pi$ from one wall to the other. According to Fig.\ \ref{fig:TwoWalls}, the wall with the strong anchoring at $z/D=0$ is able to align the nematic director along the easy axis there ($\varphi=0$), while the magnetization field is switching to the negative $x$-direction parallel to the external field (compare Sec.~\ref{subsec:I}). On the other side, the weak anchoring at $z/D=1$ in Fig.~\ref{fig:TwoWalls} allows the nematic director there to follow the magnetization ($\theta=\pi$ implies $\varphi=\pi$; compare Sec.~\ref{subsec:II}). This provides a situation in which $\varphi=0$ at one wall and $\varphi=\pi$ at the other. The elastic contribution in Eq.~\eqref{eq:NLCEl} ensures that no singularities occur in the interior of the slab so that there is a smooth crossover between the two boundary values. \begin{figure}[t!] \centering \includegraphics[scale=1]{./two_walls_vectors.eps} \caption{Vector fields corresponding to Fig.~\ref{fig:TwoWalls}. The notation is the same as in Fig.~\ref{fig:Nucleation}. Note the distortion of the nematic director field (black rods) in the interior and the interface between two magnetization (red arrows) domains at $z\approx 0.55D$ due to the combination of the switching mechanisms I and II. Basically throughout the whole sample the magnetization has reached the switched state in negative $x$-direction. $\gamma=240$, $W^{(1)}_\text{wall}=10^{-5}$~J/m$^2$ (see $z=0$), and $W^{(2)}_\text{wall}=0.1\times10^{-5}$~J/m$^2$ (see $z=D$), $K=9\times 10^{-12}$ N \cite{footnote3} (the values of the remaining parameters are the same as in Fig.\ \ref{fig:Hyst}).} \label{fig:TwoWallsVectors} \end{figure} In the middle ($z\approx 0.55D$) of the sample the magnetization field exhibits an interface between two halves of the slab (see Fig.~\ref{fig:TwoWallsVectors}). The orientation of the magnetization within the two halves differs only in how the magnetization approaches the value $\theta=\pi$ in the vicinity of the center of the slab. This behavior of the magnetization field profile is caused by the necessity to be compatible with the nematic director profile in the center region. This means that the rotation of the nematic director in the interior of the slab forces the magnetization direction to reach its value $\theta=\pi$ at $z\approx D/2$ either from $\theta>\pi$ at $z\lesssim D/2$ or from $\theta<\pi$ at $z\gtrsim D/2$ (see the red curve in Fig.\ \ref{fig:TwoWalls}). In the situation of Fig.~\ref{fig:TwoWalls}, upon switching off the external magnetic field, we found that the system relaxes into a state with a uniform nematic director field and two domains with the magnetization pointing into opposite directions (see Fig.~\ref{fig:Storage}(c)) \cite{footnote3}. The position of the interface between these two domains depends on the position of the interface plane formed while the magnetic field was still on \cite{footnote4}. Application of the external magnetic field to the two-domain configuration opens up two possibilities: (i) If the external magnetic field is applied in the same direction as the field used to create the two-domain sample, the resulting state is identical to the one in Fig.~\ref{fig:TwoWalls}. (ii) If, on the other hand, the magnetic field is applied opposite to the direction of the magnetic field used to create the two-domain sample, one of the domains (i.e., the one the magnetization of which is opposite to the external field) switches. This yields a uniform sample both in terms of the nematic director and the magnetization field, thus returning the system to its initial state. These steps are summarized in Fig.~\ref{fig:Storage}. Note that the states shown in Figs.~\ref{fig:Storage} (a) and \ref{fig:Storage} (d) are identical. Also note that the state depicted in Fig.~\ref{fig:Storage} (a) exhibits saturated magnetization $\mathcal{M}=1$ (Eq.\ \eqref{eq:PrettyM}) whereas the state depicted on Fig.~\ref{fig:Storage} (c) exhibits $\mathcal{M}\ll 1$. Since one can restore the initial state (see Fig.~\ref{fig:Storage} (a)) from the two-domain state (see Fig.~\ref{fig:Storage}(c)) by applying an external magnetic field $B>0$, one is able to cycle through three states (see Figs.~\ref{fig:Storage} (a), (b), and (c)). Accordingly, this ferronematic cell with two walls of different anchoring strength can be put in either of two states (i.e., magnetized or demagnetized) by using an external magnetic field of suitable direction. This opens up the possibility, e.g., to use an array of such cells for storage of binary information with a "bit" being represented by the state of the cell (magnetized/demagnetized) or as a spatially resolving magnetic field detector with memory function. \begin{figure*}[t!] \centering \includegraphics[scale=1]{./storage.eps} \caption{Orientational profiles $\varphi(z)$ (nematic director, black line) and $\theta(z)$ (magnetization, red line) in a ferronematic cell with walls of different anchoring strength, as discussed in Sec.~\ref{subsec:TwoWalls}. The initial state (a) is a uniformly magnetized ferronematic slab, i.e., the director field ($\varphi=0$, black line) and the magnetization field ($\theta=0$, red line) are uniform. Upon application of an external magnetic field $\mathbf{B}=B\mathbf{e_x}$ in the direction opposite to the initial magnetization direction (i.e., $B<0$) transfers the sample into the disturbed state (b) (compare Fig.~\ref{fig:TwoWalls}). After suddenly switching off the external field (i.e., for $B=0$) the system relaxes into the state (c) in which the nematic director is uniform; that part of the sample, which is close to the wall with strong anchoring, retains its magnetization direction $\theta\approx \pi$ whereas the magnetization near the wall with weak anchoring follows the relaxation of the nematic director and attains $\theta=0$. Thus having two halves of the sample being magnetized in opposite directions yields zero overall sample magnetization, i.e., $\mathcal{M}\ll 1$. This configuration offers two options: (i) The application of an external magnetic field in the direction of the initial magnetization (i.e., $B>0$) returns the sample to the initial, uniform state (d)=(a) with magnetization $\mathcal{M}=1$. (ii) The application of an external magnetic field in the direction opposite to the initial magnetization (i.e., $B<0$) produces the disturbed state (e)=(b). The fact that the states (d) and (a) are identical allows one to cycle through the states (a), (b), and (c) by applying the external magnetic field $\mathbf{B}$ in suitable directions. The width of the interface between two domains in (c) is not larger than the numerical grid discretization, i.e., less than $D/100$. The values of the parameters are the same as in Fig.\ \ref{fig:TwoWalls}.} \label{fig:Storage} \end{figure*} \begin{figure}[t!] \centering \includegraphics[scale=1]{./seg.eps} \caption{Results of the numerical minimization of the functional in Eq.~\eqref{eq:SlabF} with segregation effects included (see Eqs.~\eqref{eq:Split} and \eqref{eq:FSeg}). The sample thickness is $D=120\,\mathrm{\mu m}$, the wall anchoring is $W_\text{wall}=3.4\times10^{-5}$~J/m$^2$ for both walls, the coupling constant is $\gamma=240$, the external magnetic field is $B=4$~mT, and $a\approx 3.1\times 10^{-4}$~N/A$^2$. (The values of the remaining parameters are the same as in Fig.\ \ref{fig:Hyst}.) Due to the large thickness $D$ of the cell the switching mechanism I is now combined with the switching mechanism II even in the case that both walls provide strong anchoring. The particles are expelled (see the blue solid line) from the regions with strong gradients of the nematic director field $\varphi$ (black solid line). The depletion layers (i.e., the minima of the blue solid line) are separated from the walls due to the soft anchoring between the magnetization and the nematic director field. This differs from the situation described in Ref.~\cite{1970_Brochard} in which the depletion layer is located in close vicinity of the wall.} \label{fig:Segregation} \end{figure} \section{\label{sec:Segregation}Segregation effects} It was pointed out by Brochard and de Gennes \cite{1970_Brochard} that anisotropic magnetic colloids tend to move away from regions of the NLC where distortions of the director field prevent them from minimizing their free energy in the external magnetic field. The segregation parameter defined as (see Ref.~\cite{1970_Brochard}) \begin{equation} s:=\beta mB \end{equation} is greater than unity already for $B\gtrsim 1.4$ mT and therefore one can expect segregation to occur for external fields stronger than $1.4$ mT. In the following we investigate the impact of segregation on the switching mechanisms I and II. The present theoretical approach (see Eqs.~\eqref{eq:FN} and \eqref{eq:SlabF}) includes the possibility of segregation to occur through the dependence of the magnetization field $\mathbf{M}$ on the spatial coordinate $z$. In particular, we are interested in spatial inhomogeneities of the \textit{absolute value} $|\mathbf{M}(z)|$ of the magnetization vector. It is convenient to introduce the dimensionless quantity $\tau (z) := |\mathbf{M}(z)|/(m\rho_\text{iso})$. So far all our results have been obtained in the limit $\tau(z) = \text{const} =1$. If $\tau (z) \neq 1$, that part of the free energy density, which depends on it (see Eq.~\eqref{eq:FN} and Fig.\ \ref{fig:Slab}), is given by \begin{equation} \begin{aligned} \label{eq:FNWithS} \beta &f(\theta(z), \varphi(z), \tau(z))=\\ &\beta (m\rho_\text{iso})^2\tau(z)^2 \Big(\frac{a}{2}-\frac{1}{2}\gamma\mu_0 \cos(\theta(z)-\varphi(z))^2\Big)\\ &-\beta m\rho_\text{iso}B\tau(z)\cos(\theta(z)-\psi). \end{aligned} \end{equation} Concerning the segregation effects the value of $a$ matters. According to Ref.\ \cite{2018_Zarubin} the value of $\gamma=240$ implies $c\approx 0.035$ and therefore $a\approx 10\times k_\text{B}T/(m^2\rho_\text{iso})\approx 3.1\times 10^{-4}$ N/A$^2$ (concerning the definition of $a$ in terms of $c$ see Ref.\ \cite{2018_Zarubin}). Since the sample always contains a fixed number of magnetic colloids (i.e., neglecting aggregation) the field $\tau(z)$ is subject to the constraint (see Appendix \ref{sec:Appendix}) \begin{equation} \label{eq:TauCons} \frac{1}{D}\int_0^D\!\!\ensuremath{\mathop{}\!\mathrm{d}} z\,\,\tau(z)=1. \end{equation} It is convenient to consider deviations $\delta\tau(z)$ from the homogeneous case, i.e., \begin{equation} \tau(z)=1+\delta\tau(z), \label{eq:deltataudef} \end{equation} which allows one to rewrite the free energy density in Eq.~\eqref{eq:FNWithS} as the sum of the free energy density evaluated for $\tau(z) = \text{const} =1$ and the contribution due to segregation: \begin{align} \label{eq:Split} \beta &f(\theta(z), \varphi(z), \tau(z))=\nonumber\\ &\beta f(\theta(z),\varphi(z), \tau(z)=1)+\beta f_\text{seg}(\theta(z), \varphi(z), \tau(z)), \end{align} where $f_\text{seg}(\theta(z), \varphi(z), \tau(z))$ is defined as \begin{equation} \label{eq:FSeg} \begin{aligned} \beta &f_\text{seg}(\theta(z), \varphi(z), \tau(z)):=\\ &\beta (m\rho_\text{iso})^2\Big(a-\gamma\mu_0 \cos(\theta(z)-\varphi(z))^2\\ &-\beta m\rho_\text{iso}B\cos(\theta(z)-\psi)\Big)\delta\tau(z)\\ &+\beta (m\rho_\text{iso})^2 \Big(\frac{a}{2}-\frac{1}{2}\gamma\mu_0 \cos(\theta(z)-\varphi(z))^2\Big)\delta\tau(z)^2. \end{aligned} \end{equation} The constraint in Eq.~\eqref{eq:TauCons} turns into \begin{equation} \label{eq:DTauCons} \int_0^D\!\!\ensuremath{\mathop{}\!\mathrm{d}} z\,\,\delta\tau(z)=0. \end{equation} We express $\delta\tau(z)$ in terms of a Fourier series: \begin{equation} \delta\tau(z)=\frac{a_0}{2}+\sum_{n=1}^\infty \Big[a_n\cos\Big(\frac{2\pi nz}{D}\Big)+b_n\sin\Big(\frac{2\pi nz}{D}\Big)\Big]. \end{equation} Equation~\eqref{eq:DTauCons} implies $a_0=0$. The functional in Eq.~\eqref{eq:SlabF} is minimized with respect to the fields $\theta(z)$ and $\varphi(z)$ and the coefficients $a_n$ and $b_n$, $n\in \{1,2,...,N\}$. The number of coefficients $N$ to be taken into account has to be chosen. It is reasonable to set the minimum wavelength in the Fourier series to be larger than the colloid diameter $d\sim 100$ nm. Therefore, $N$ has to be smaller than $N_\text{max}=[D/d\,]$ where $[x]$ denotes the integer part of $x$. For a slab of thickness $D=20$~$\mu$m one has $N_\text{max}=200$. It turned out that for slab thicknesses $D<60\,\mathrm{\mu m}$ the equilibrium profile $\tau(z)$ obtained from Eq.~\eqref{eq:FNWithS} takes negative values, i.e., $\delta\tau(z)<-1$, which contradicts its physical meaning $\tau(z)\sim|\mathbf{M}(z)| \geq0$. This behavior is related to the absence of contributions in $\tau(z)$ beyond quadratic order. However, for slab thicknesses $D\geq 60$ $\mu$m the algorithm does provide the profiles $\theta(z)$ and $\varphi(z)$ together with a physically reasonable segregation profile $\delta\tau(z)$. Figure \ref{fig:Segregation} shows the calculated profiles for $D= 120$~$\mu$m, equal walls with strong anchoring $W_\text{wall}=3.4\times10^{-5}$~J/m$^2$, coupling constant $\gamma=240$, and external magnetic field $B=4$~mT. It is evident that switching mechanism I is observed even in the presence of segregation effects. The density of magnetic colloids is largely reduced (with $\delta\tau(z)$ close to -1) in the regions of nonzero gradient of the director profile. An important difference to the case of infinitely strong coupling of the colloids to the liquid crystal (see Ref.~\cite{1970_Brochard}) is the \textit{depletion layer} being shifted away from the walls towards the interior of the sample. The functional form of the profiles $\varphi(z)$ and $\theta(z)$ obtained for asymmetric pairs of walls with strong and weak anchoring (see Fig.~\ref{fig:TwoWalls} in Sec.~\ref{subsec:TwoWalls}) is found at both walls for sufficiently thick slabs. We have performed a series of calculations for different slab thicknesses $D$ in order to determine for which thickness the pure switching mechanism I turns into a combination of mechanisms I and II (see Fig.~\ref{fig:Segregation}); we have found $D=95\pm 5$~$\mu$m. This observation opens the possibility to manipulate the sample magnetization in a manner similar to that described in Sec.~\ref{subsec:TwoWalls} but without the need to use a second, weakly anchored wall. However, the state of nonzero net magnetization is not necessarily saturated, but it might exhibit $\mathcal{M} < 1$. In order to have nonetheless a state with the magnetization $\mathcal{M}\approx 1$, one would need to adjust the system parameters (e.g., wall anchoring, elastic constant of the NLC etc.) such that each region of switched magnetization in the vicinity of the walls (see Fig.~\ref{fig:Segregation}) takes up $\approx 25\%$ of the slab thickness. We have also performed a calculation for the case of two walls with equally weak anchoring. We found that segregation did not have a qualitative impact in that case either. Therefore we expect that the results of Secs.\ \ref{subsec:I} and \ref{subsec:II} are not affected qualitatively by segregation. \section{\label{sec:Discussion}Summary and conclusions} In this analysis we have studied theoretically a ferronematic confined between two planar, parallel walls which impose an easy axis on the NLC director field. Inspired by the experimental studies reported in Refs. \cite{2013_Mertelj, 2014_Mertelj}, the system is subjected to an external magnetic field. The ferronematic is an anisotropic polar fluid and thus the system is characterized by the relative directions of the NLC director, the easy axes due to the walls, the magnetization, and the external magnetic field. We have considered the situation in which the ferronematic is initially prepared with a uniform magnetization along the easy axis of the NLC. Subsequently an external magnetic field is applied in the direction opposite to the magnetization. This choice of the geometry reduces the theoretical description to an effectively one-dimensional one. The experiments reported in Refs.~\cite{2013_Mertelj, 2014_Mertelj} showed that for such a setup there exists a critical external magnetic field $B_\text{cr}>0$ such that for magnetic field strengths $|\mathbf{B}|<B_\text{cr}$ the sample remains unperturbed. The authors of Ref.~\cite{2013_Mertelj} also provided the expression in Eq.~\eqref{eq:BCr} for the critical magnetic field strength in terms of the coupling $\gamma$ between the magnetization and the nematic director field. This critical magnetic field strength, which increases upon increasing $\gamma$, has been determined as that magnetic field strength for which the relaxation rate of long-wavelength fluctuations of the nematic director field vanish. Here we study the system by numerical minimization of the corresponding free energy functional in Eq.~\eqref{eq:SlabF}. The numerical minimization is performed by using the Fletcher-Reeves-Polak-Ribiere general function minimization algorithm. It is obvious, that the global minimum of the free energy functional in Eq.~\eqref{eq:SlabF} before the external magnetic field has been applied is the initial state of the ferronematic being uniformly magnetized along the easy axis. Once the magnetic field is applied in the direction opposite to the initial magnetization, this state becomes only a local (metastable) minimum. The new global minimum is the ferronematic magnetized in the direction of the field. By means of a conjugate gradient algorithm one is able to search for the local minimum of the free energy and therefore to identify metastable states of the system. This is particularly useful in the present context, because this way one can investigate possible intermediate orientation profiles between the initial, now metastable, state with a uniform magnetization in the positive $x$-direction, i.e., opposite to the magnetic field $\mathbf{B}=B\mathbf{e}_x$ pointing into the negative $x$-direction, i.e., $B<0$, and the final stable state with the magnetization in the direction along the magnetic field, i.e., in negative $x$-direction ($B<0$). Figure \ref{fig:Hyst} shows the dependence of the dimensionless magnetization $\mathcal{M}$ (Eq.~\eqref{eq:PrettyM}) of the metastable state described above on the strength of the external magnetic field for particular values of the coupling constant $\gamma$ and of the anchoring strength $W_\text{wall}$ at a wall. One observes hysteresis of the magnetization for which a critical magnetic field strength $B_\text{cr}$ can be identified as the one for which significant deviations from the saturation magnetization $\mathcal{M}=1$ occur. One can distinguish several, qualitatively different, intermediate states (red circles in Fig.~\ref{fig:Hyst}). First, for a magnetic field $\mathbf{B}=B\mathbf{e}_x$, with component $B$ in the direction of the initial magnetization ($B>0$), the sample remains practically unperturbed for $B > B_\text{cr}$ ($B_\text{cr}<0$), i.e., for the magnetic field either along the initial magnetization ($B>0$) or sufficiently weak in the direction opposite to the initial magnetization ($B<0$) (see Fig.~\ref{fig:MetaS}(1)). Upon decreasing the component $B$ of the magnetic field $\mathbf{B}=B\mathbf{e}_x$ further (i.e., making it less positive or more negative), the torque imposed on the NLC by the walls is no longer able to keep the ferronematic in the initial unperturbed state and thus the profiles become perturbed (see Fig.~\ref{fig:MetaS}(2)). These states correspond to the brightening of the sample when viewed via crossed polarizers \cite{2013_Mertelj}. If one decreases the magnetic field component $B$ even further (i.e., making $B$ even more negative), one encounters the interesting metastable state shown in Fig.~\ref{fig:MetaS}(3). In this state, near each wall a layer is formed within which the nematic director is close to the easy axis and the magnetization has inverted its direction, pointing along the external magnetic field (i.e., in negative $x$-direction). This flipping of the magnetization is energetically favorable for a sufficiently large strength $|\mathbf{B}|=|B|$ of the magnetic field pointing in the direction opposite to the initial magnetization (i.e., pointing into the negative $x$-direction), because the contribution to the free energy (Eq.~\eqref{eq:FN}) of the coupling between the magnetization and the nematic director is invariant upon inversion of the magnetization, $\mathbf{M} \mapsto -\mathbf{M}$, but the contribution of the coupling between the magnetization and the external magnetic field is not. When the external magnetic field becomes even stronger (i.e., $B$ becomes even more negative and $|B|$ even larger), the regions of flipped magnetization expand into the interior of the system, eventually giving rise to the whole sample (except for thin layers in the very vicinity of the walls) being magnetized along the magnetic field. The qualitatively different scenario, which we refer to as switching mechanism II, occurs if the wall anchoring $W_\text{wall}$ is too weak to prevent the nematic director field from following the rotating magnetization. In accordance to Figs.\ \ref{fig:MetaS}(3) and \ref{fig:MetaS}(4), in this scenario the final stages of the switching are not realized. Figure \ref{fig:MetaSII} illustrates how the final stages (3) and (4) of the switching mechanism II are realized, according to which the magnetization \textit{and} the director field rotate in parallel. Figure \ref{fig:Switch} shows whether certain combinations of wall anchoring strengths $W_\text{wall}^{(1)}=W_\text{wall}^{(2)}=W_\text{wall}$ as well as of the coupling constant $\gamma$ lead to switching mechanism I or II. According to the map in Fig.\ \ref{fig:Switch} the switching mechanism I is the dominant one in the experiments described in Refs.\ \cite{2013_Mertelj,2014_Mertelj} ($W_\text{wall}^\text{exp}\approx 3.4\times 10^{-5}$~J/m$^2$). The dependences of the critical magnetic field strength $B_\text{cr}$ on the coupling constant $\gamma$ and on the (equal) wall anchoring strength $W_\text{wall}$ are presented in Figs.~\ref{fig:Phase} and \ref{fig:Cuts}. On one hand, $B_\text{cr}$ increases as function of $W_\text{wall}$ and, on the other hand, it also increases as function of $\gamma$, which is consistent with the results of Ref.~\cite{2013_Mertelj}. In Fig.~\ref{fig:Cuts}(b) a comparison of the critical magnetic field strength $B_\text{cr}$ as defined here with that introduced in Ref.~\cite{2013_Mertelj} (see Eq.~\eqref{eq:BCr}) shows good agreement, although the two expressions involve different properties of the ferronematic. Within a recently developed theory of ferronematics \cite{2018_Zarubin} one can relate the coupling coefficient $\gamma$ to the microscopic coupling $c$ (see below Eq.\ \eqref{eq:FN}) which depends on the size of the colloids in the suspension. Figure~\ref{fig:Cuts}(c) shows the dependence of $B_\text{cr}$ on the microscopic coupling $c$ for a particular value of the wall anchoring strength $W_\text{wall}$. This allows one to vary the critical magnetic field by tuning the mean value of the size distribution of the colloids participating in the ferronematic. Combining two walls with different anchoring strengths allows one to design a sample such that its switching mechanism is a superposition of type I and type II. The resulting nematic director and magnetization field profiles (see Figs.~\ref{fig:TwoWalls} and \ref{fig:TwoWallsVectors}) are obtained by applying an external magnetic field. In turn, switching off this field divides the sample into two domains with opposite magnetizations (see Fig.~\ref{fig:Storage} (c)), rendering a sample with zero net magnetization. The initial state (i.e., the magnetized slab) can be restored by applying an external magnetic field of suitable direction to the two-domain sample. This cycle can be repeated arbitrarily, thus facilitating the switching between two states (magnetized/demagnetized slab) by using a uniform magnetic field only. This opens application perspectives such as storage of information and magnetic fields detection. Similar controllable magnetic slabs can be constructed by using two walls with equally strong anchoring for samples of larger thickness ($D\geq 95$ $\mu$m for $\gamma=240$, $K=3.5\times 10^{-12}$ N, and $W_\text{wall}=3.4\times 10^{-5}$ J/m$^2$). We have found segregation to be \textit{quantitatively} different from the case of walls with infinitely strong anchoring and of infinitely strong coupling of the colloids to the NLC (see Ref.\ \cite{1970_Brochard}). However, segregation effects do not affect the switching mechanisms qualitatively. \section*{Conflicts of interest} There are no conflicts to declare.

62627f74502f35ccf19f3ad1028ce1d9df641415

\section{Introduction} Although the idea of automated translation can in principle be traced back to as long as the 17th century with Ren\'{e} Descartes proposal of an ``universal language'' \cite{dascal1982universal}, machine translation has only existed as a technological field since the 1950s, with a pioneering memorandum by Warren Weaver \cite{locke1955machine,weaver1955translation} discussing the possibility of employing digital computers to perform automated translation. The now famous Georgetown-IBM experiment followed not long after, providing the first experimental demonstration of the prospects of automating translation by the means of successfully converting more than sixty Russian sentences into English \cite{gordin2015scientific}. Early systems improved upon the results of the Georgetown-IBM experiment by exploiting Noam Chomsky's theory of generative linguistics, and the field experienced a sense of optimism about the prospects of fully automating natural language translation. As is customary with artificial intelligence, the initial optimistic stage was followed by an extended period of strong disillusionment with the field, of which the catalyst was the influential 1966 ALPAC (Automatic Language Processing Advisory Committee) report( \cite{hutchins1986machine}. Such research was then disfavoured in the United States, making a re-entrance in the 1970s before the 1980s surge in statistical methods for machine translation \cite{koehn2009statistical,Moses2007}. Statistical and example-based machine translation have been on the rise ever since \cite{Bahdanau2014,carl2003recent,Firat2017}, with highly successful applications such as Google Translate (recently ported to a neural translation technology \cite{wu2016google}) amounting to over $200$ million users daily. In spite of the recent commercial success of automated translation tools (or perhaps stemming directly from it), machine translation has amounted a significant deal of criticism. Noted philosopher and founding father of generative linguistics Noam Chomsky has argued that the achievements of machine translation, while successes in a particular sense, are \emph{not successes in the sense that science has ever been interested in}: they merely provide effective ways, according to Chomsky, of approximating unanalyzed data \cite{Chomsky2011,norvig2017chomsky}. Chomsky argues that the faith of the MT community in statistical methods is absurd by analogy with a standard scientific field such as physics \cite{Chomsky2011}: \begin{quotation} \textsl{I mean actually you could do physics this way, instead of studying things like balls rolling down frictionless planes, which can't happen in nature, if you took a ton of video tapes of what's happening outside my office window, let's say, you know, leaves flying and various things, and you did an extensive analysis of them, you would get some kind of prediction of what's likely to happen next, certainly way better than anybody in the physics department could do. Well that's a notion of success which is I think novel, I don't know of anything like it in the history of science}. \end{quotation} Leading AI researcher and Google's Director of Research Peter Norvig responds to these arguments by suggesting that even standard physical theories such as the Newtonian model of gravitation are, in a sense, \emph{trained} \cite{norvig2017chomsky}: \begin{quotation} \textsl{As another example, consider the Newtonian model of gravitational attraction, which says that the force between two objects of mass $m_1$ and $m_2$ a distance $r$ apart is given by} \begin{equation*} F = G m_1 m_2 / r^2 \end{equation*} \textsl{where $G$ is the universal gravitational constant. This is a trained model because the gravitational constant G is determined by statistical inference over the results of a series of experiments that contain stochastic experimental error. It is also a deterministic (non-probabilistic) model because it states an exact functional relationship. I believe that Chomsky has no objection to this kind of statistical model. Rather, he seems to reserve his criticism for statistical models like Shannon's that have quadrillions of parameters, not just one or two.} \end{quotation} Chomsky and Norvig's debate \cite{norvig2017chomsky} is a microcosm of the two leading standpoints about the future of science in the face of increasingly sophisticated statistical models. Are we, as Chomsky seems to argue, jeopardizing science by relying on statistical tools to perform predictions instead of perfecting traditional science models, or are these tools, as Norvig argues, components of the scientific standard since its conception? Currently there are no satisfactory resolutions to this conundrum, but perhaps statistical models pose an even greater and more urgent threat to our society. On a 2014 article, Londa Schiebinger suggested that scientific research fails to take gender issues into account, arguing that the phenomenon of male defaults on new technologies such as Google Translate provides a window into this asymmetry \cite{schiebinger2014scientific}. Since then, recent worrisome results in machine learning have somewhat supported Schiebinger's view. Not only Google photos' statistical image labeling algorithm has been found to classify dark-skinned people as gorillas \cite{garcia2016racist} and purportedly intelligent programs have been suggested to be negatively biased against black prisoners when predicting criminal behavior \cite{angwin2016machine} but the machine learning revolution has also indirectly revived heated debates about the controversial field of physiognomy, with proposals of AI systems capable of identifying the sexual orientation of an individual through its facial characteristics \cite{wang2017deep}. Similar concerns are growing at an unprecedented rate in the media, with reports of Apple's Iphone X face unlock feature failing to differentiate between two different Asian people \cite{womanunlockphone2017} and automatic soap dispensers which reportedly do not recognize black hands \cite{racistsoapdispenser2017}. \emph{Machine bias}, the phenomenon by which trained statistical models unbeknownst to their creators grow to reflect controversial societal asymmetries, is growing into a pressing concern for the modern times, invites us to ask ourselves whether there are limits to our dependence on these techniques -- and more importantly, whether some of these limits have already been traversed. In the wave of algorithmic bias, some have argued for the creation of some kind of agency in the likes of the Food and Drug Administration, with the sole purpose of regulating algorithmic discrimination \cite{kirkpatrick2016battling}. With this in mind, we propose a quantitative analysis of the phenomenon of gender bias in machine translation. We illustrate how this can be done by simply exploiting Google Translate to map sentences from a gender neutral language into English. As Figure \ref{fig:screenshot-gtranslate-hungarian} exemplifies, this approach produces results consistent with the hypothesis that sentences about stereotypical gender roles are translated accordingly with high probability: \emph{nurse} and \emph{baker} are translated with female pronouns while \emph{engineer} and \emph{CEO} are translated with male ones. \begin{figure}[h] \centering \fbox{\includegraphics[width=\linewidth]{pictures/screenshot-gtranslate-hungarian}} \caption{Translating sentences from a gender neutral language such as Hungarian to English provides a glimpse into the phenomenon of gender bias in machine translation. This screenshot from Google Translate shows how occupations from traditionally male-dominated fields \cite{WB2014} such as scholar, engineer and CEO are interpreted as male, while occupations such as nurse, baker and wedding organizer are interpreted as female.} \label{fig:screenshot-gtranslate-hungarian} \end{figure} \section{Motivation} As of 2018, Google Translate is one of the largest publicly available machine translation tools in existence, amounting 200 million users daily\cite{Gtranslate200daily2017}. Initially relying on United Nations and European Parliament transcripts to gather data, since 2014 Google Translate has inputed content from its users through the Translate Community initiative\cite{TranslateCommunity}. Recently however there has been a growing concern about gender asymmetries in the translation mechanism, with some heralding it as ``sexist'' \cite{AlgorithmGtranslateSexist2018}. This concern has to at least some extent a scientific backup: A recent study has shown that word embeddings are particularly prone to yielding gender stereotypes\cite{bolukbasi2016man}. Fortunately, the researchers propose a relatively simple \emph{debiasing} algorithm with promising results: they were able to cut the proportion of stereotypical analogies from $19\%$ to $6\%$ without any significant compromise in the performance of the word embedding technique. They are not alone: there is a growing effort to systematically discover and resolve issues of algorithmic bias in black-box algorithms\cite{hajian2016algorithmic}. The success of these results suggest that a similar technique could be used to remove gender bias from Google Translate outputs, should it exist. This paper intends to investigate whether it does. We are optimistic that our research endeavors can be used to argue that there is a positive payoff in redesigning modern statistical translation tools. \section{Assumptions and Preliminaries} In this paper we assume that a statistical translation tool should reflect at most the inequality existent in society -- it is only logical that a translation tool will poll from examples that society produced and, as such, will inevitably retain some of that bias. It has been argued that one's language affects one's knowledge and cognition about the world \cite{kay1984sapir}, and this leads to the discussion that languages that distinguish between female and male genders grammatically may enforce a bias in the person's perception of the world, with some studies corroborating this, as shown in \cite{boroditsky2003sex}, as well some relating this with sexism \cite{thompson2014linguistic} and gender inequalities \cite{santacreu2013female}. With this in mind, one can argue that a move towards gender neutrality in language and communication should be striven as a means to promote improved gender equality. Thus, in languages where gender neutrality can be achieved -- such as English -- it would be a valid aim to create translation tools that keep the gender-neutrality of texts translated into such a language, instead of defaulting to male or female variants. We will thus assume throughout this paper that although the distribution of translated gender pronouns may deviate from 50:50, it should not deviate to the extent of misrepresenting the demographics of job positions. That is to say we shall assume that Google Translate incorporates a negative gender bias if the frequency of male defaults overestimates the (possibly unequal) distribution of male employees per female employee in a given occupation. \section{Materials and Methods} We shall assume and then show that the phenomenon of gender bias in machine translation can be assessed by mapping sentences constructed in gender neutral languages to English by the means of an automated translation tool. Specifically, we can translate sentences such as the Hungarian ``ő egy ápolónő'', where ``ápolónő'' translates to ``nurse'' and ``ő'' is a gender-neutral pronoun meaning either he, she or it, to English, yielding in this example the result ``she's a nurse'' on Google Translate. As Figure \ref{fig:screenshot-gtranslate-hungarian} clearly shows, the same template yields a male pronoun when ``nurse'' is replaced by ``engineer''. The same basic template can be ported to all other gender neutral languages, as depicted in Table \ref{tab:templates}. Given the success of Google Translate, which amounts to 200 million users daily, we have chosen to exploit its API to obtain the desired thermometer of gender bias. Also, in order to solidify our results, we have decided to work with a fair amount of gender neutral languages, forming a list of these with help from the World Atlas of Language Structures (WALS) \cite{wals} and other sources. Table \ref{tab:gender-neutral-languages} compiles all languages we chose to use, with additional columns informing whether they (1) exhibit a gender markers in the sentence and (2) are supported by Google Translate. However, we stumbled on some difficulties which led to some of those langauges being removed, which will be explained in . There is a prohibitively large class of nouns and adjectives that could in principle be substituted into our templates. To simplify our dataset, we have decided to focus our work on job positions -- which, we believe, are an interesting window into the nature of gender bias --, and were able to obtain a comprehensive list of professional occupations from the Bureau of Labor Statistics' detailed occupations table \cite{BLS2017}, from the United States Department of Labor. The values inside, however, had to be expanded since each line contained multiple occupations and sometimes very specific ones. Fortunately this table also provided a percentage of women participation in the jobs shown, for those that had more than 50 thousand workers. We filtered some of these because they were too generic ( ``Computer occupations, all other'', and others) or because they had gender specific words for the profession (``host/hostess'', ``waiter/waitress''). We then separated the curated jobs into broader categories (Artistic, Corporate, Theatre, etc.) as shown in Table \ref{tab:occupations}. Finally, Table \ref{tab:occupations-examples} shows thirty examples of randomly selected occupations from our dataset. For the occupations that had less than 50 thousand workers, and thus no data about the participation of women, we assumed that its women participation was that of its upper category. Finally, as complementary evidence we have decided to include a small subset of 21 adjectives in our study. All adjectives were obtained from the top one thousand most frequent words in this category as featured in the Corpus of Contemporary American English (COCA) https://corpus.byu.edu/coca/, but it was necessary to manually curate them because a substantial fraction of these adjectives cannot be applied to human subjects. Also because the sentiment associated with each adjective is not as easily accessible as for example the occupation category of each job position, we performed a manual selection of a subset of such words which we believe to be meaningful to this study. These words are presented in Table \ref{tab:adjectives}. We made all code and data used to generate and compile the results presented in the following sections publicly available in the following Github repository: https://github.com/marceloprates/Gender-Bias. Note however that because the Google Translate algorithm can change, unfortunately we cannot guarantee full reproducibility of our results. All experiments reported here were conducted on April 2018. \begin{table}[H] \centering \begin{small} \begin{tabular}{|c|m{1.5cm}|m{1.5cm}|m{2.2cm}|m{1.5cm}|} \hline Language Family & Language & Phrases have male/female markers & Tested \\ \hline \hline Austronesian & Malay & \xmark & \checkmark\\ \hline \multirow{3}{*}{Uralic} & Estonian & \xmark & \checkmark\\ & Finnish & \xmark & \checkmark\\ & Hungarian & \xmark & \checkmark\\ \hline \multirow{5}{*}{Indo-European} & Armenian & \xmark & \checkmark\\ & Bengali & O & \checkmark\\ & English & \checkmark & \xmark \\ & {\color{red}Persian} & \xmark & \checkmark\\ & {\color{red}Nepali} & O & \checkmark\\ \hline Japonic & Japanese & \xmark & \checkmark\\ \hline Koreanic & {\color{red}Korean} & \checkmark& \xmark\\ \hline Turkic & Turkish & \xmark & \checkmark\\ \hline \multirow{2}{*}{Niger-Congo} & Yoruba & \xmark & \checkmark\\ & Swahili & \xmark & \checkmark\\ \hline Isolate & Basque & \xmark & \checkmark\\ \hline Sino-Tibetan & Chinese & O & \checkmark\\ \hline \end{tabular} \end{small} \caption{Gender neutral languages supported by Google Translate. Languages are grouped according to language families and classified according to whether they enforce any kind of mandatory gender (male/female) demarcation on simple phrases (\checkmark: yes, \xmark:~never, O:~some). For the purposes of this work, we have decided to work only with languages lacking such demarcation. Languages colored in red have been omitted for other reasons. See Section} \ref{ssec:rationale_for_language_exceptions} for further explanation. \label{tab:gender-neutral-languages} \end{table} \begin{table}[H] \centering \begin{small} \centering \begin{tabular}{|m{3.5cm}|m{2.0cm}|m{2.5cm}|m{2.5cm}|} \hline Category & Group & \# Occupations & Female Participation \\ \hline \hline Education, training, and library & Education & $22$ & $73.0\%$ \\ \hline Business and financial operations & Corporate & $46$ & $54.0\%$ \\ \hline Office and administrative support & Service & $87$ & $72.2\%$ \\ \hline Healthcare support & Healthcare & $16$ & $87.1\%$ \\ \hline Management & Corporate & $46$ & $39.8\%$ \\ \hline Installation, maintenance, and repair & Service & $91$ & $4.0\%$ \\ \hline Healthcare practitioners and technical & Healthcare & $43$ & $75.0\%$ \\ \hline Community and social service & Service & $14$ & $66.1\%$ \\ \hline Sales and related & Corporate & $28$ & $49.1\%$ \\ \hline Production & Production & $264$ & $28.9\%$ \\ \hline Architecture and engineering & STEM & $29$ & $16.2\%$ \\ \hline Life, physical, and social science & STEM & $34$ & $47.4\%$ \\ \hline Transportation and material moving & Service & $70$ & $17.3\%$ \\ \hline Arts, design, entertainment, sports, and media & Arts / Entertainment & $37$ & $46.9\%$ \\ \hline Legal & Legal & $7$ & $52.8\%$ \\ \hline Protective Service & Service & $28$ & $22.3\%$ \\ \hline Food preparation and serving related & Service & $17$ & $53.8\%$ \\ \hline Farming, fishing, and forestry & Farming / Fishing / Forestry & $13$ & $23.4\%$ \\ \hline Computer and mathematical & STEM & $16$ & $25.5\%$ \\ \hline Personal care and service & Service & $33$ & $76.1\%$ \\ \hline Construction and extraction & Construction / Extraction & $68$ & $3.0\%$ \\ \hline Building and grounds cleaning and maintenance & Service & $10$ & $40.7\%$ \\ \hline \hline Total & - & $1019$ & $41.3\%$ \\ \hline \end{tabular} \end{small} \caption{Selected occupations obtained from the U.S. Bureau of Labor Statistics https://www.bls.gov/cps/cpsaat11.htm, grouped by category. We obtained a total of $1019$ occupations from $22$ distinct categories. We have further grouped them into broader groups (or \emph{super-categories}) to ease analysis and visualization.} \label{tab:occupations} \end{table} \begin{table}[H] \centering \small{ \begin{tabular}{|c|c|c|} \hline Language & Occupation sentence template & Adjective sentence template \\ \hline \hline Malay & dia adalah $\langle occupation \rangle$ & dia $\langle adjective \rangle$ \\ \hline Estonian & ta on $\langle occupation \rangle$ & ta on $\langle adjective \rangle$ \\ \hline Finnish & hän on $\langle occupation \rangle$ & hän on $\langle adjective \rangle$ \\ \hline Hungarian & ő egy $\langle occupation \rangle$ & ő $\langle adjective \rangle$ \\ \hline Armenian & na $\langle occupation \rangle$ e & na $\langle adjective \rangle$ e \\ \hline \multirow{6}{*}{Bengali} & Ē ēkajana $\langle occupation \rangle$ & Ē $\langle adjective \rangle$\\ & Yini ēkajana $\langle occupation \rangle$ & Yini $\langle adjective \rangle$\\ & Ō ēkajana $\langle occupation \rangle$ & Ō $\langle adjective \rangle$\\ & Uni ēkajana $\langle occupation \rangle$ & Uni $\langle adjective \rangle$\\ & Sē ēkajana $\langle occupation \rangle$ & Sē $\langle adjective \rangle$\\ & Tini ēkajana $\langle occupation \rangle$ & Tini $\langle adjective \rangle$\\ \hline Japanese & \begin{CJK*}{UTF8}{min}あの人は\end{CJK*} $\langle occupation \rangle$ \begin{CJK*}{UTF8}{min}です\end{CJK*} & \begin{CJK*}{UTF8}{min}あの人は\end{CJK*} $\langle adjective \rangle$ \begin{CJK*}{UTF8}{min}です\end{CJK*}\\ \hline Turkish & o bir $\langle occupation \rangle$ & o $\langle adjective \rangle$ \\ \hline Yoruba & o jẹ $\langle occupation \rangle$ & o jẹ $\langle adjective \rangle$ \\ \hline Basque & $\langle occupation \rangle$ bat da & $\langle adjective \rangle$ da \\ \hline Swahili & yeye ni $\langle occupation \rangle$ & yeye ni $\langle adjective \rangle$ \\ \hline Chinese & ta shi $\langle occupation \rangle$ & ta hen $\langle adjective \rangle$ \\ \hline \end{tabular} } \caption{Templates used to infer gender biases in the translation of job occupations and adjectives to the English language.} \label{tab:templates} \end{table} \begin{table}[H] \centering \small{ \begin{tabular}{|c|c|c|c|c|} \hline Insurance sales agent & Editor & Rancher \\ \hline Ticket taker & Pile-driver operator & Tool maker \\ \hline Jeweler & Judicial law clerk & Auditing clerk \\ \hline Physician & Embalmer & Door-to-door salesperson \\ \hline Packer & Bookkeeping clerk & Community health worker \\ \hline Sales worker & Floor finisher & Social science technician \\ \hline Probation officer & Paper goods machine setter & Heating installer \\ \hline Animal breeder & Instructor & Teacher assistant \\ \hline Statistical assistant & Shipping clerk & Trapper \\ \hline Pharmacy aide & Sewing machine operator & Service unit operator \\ \hline \end{tabular} } \caption{A randomly selected example subset of thirty occupations obtained from our dataset with a total of $1019$ different occupations.} \label{tab:occupations-examples} \end{table} \begin{table}[H] \centering \small{ \begin{tabular}{|c|c|c|} \hline Happy & Sad & Right \\ \hline Wrong & Afraid & Brave \\ \hline Smart & Dumb & Proud \\ \hline Strong & Polite & Cruel \\ \hline Desirable & Loving & Sympathetic \\ \hline Modest & Successful & Guilty \\ \hline Innocent & Mature & Shy \\ \hline \end{tabular} } \caption{Curated list of 21 adjectives obtained from the top one thousand most frequent words in this category in the Corpus of Contemporary American English (COCA)} https://corpus.byu.edu/coca/. \label{tab:adjectives} \end{table} \subsection{Rationale for language exceptions} \label{ssec:rationale_for_language_exceptions} While it is possible to construct gender neutral sentences in two of the languages omitted in our experiments (namely Korean and Nepali), we have chosen to omit them for the following reasons: \begin{enumerate} \item We faced technical difficulties to form templates and automatically translate sentences with the right-to-left, top-to-bottom nature of the script and, as such, we have decided not to include it in our experiments. \item Due to Nepali having a rather complex grammar, with possible male/female gender demarcations on the phrases and due to none of the authors being fluent or able to reach someone fluent in the language, we were not confident enough in our ability to produce the required templates. Bengali was almost discarded under the same rationale, but we have decided to keep it because of our sentence template for Bengali has a simple grammatical structure which does not require any kind of inflection. \item One can construct gender neutral phrases in Korean by omitting the gender pronoun; in fact, this is the default procedure. However, the expressiveness of this omission depends on the context of the sentence being clear, which is not possible in the way we frame phrases. \end{enumerate} \section{Distribution of translated gender pronouns per occupation category} A sensible way to group translation data is to coalesce occupations in the same category and collect statistics among languages about how prominent male defaults are in each field. What we have found is that Google Translate does indeed translate sentences with male pronouns with greater probability than it does either with female or gender-neutral pronouns, in general. Furthermore, this bias is seemingly aggravated for fields suggested to be troubled by male stereotypes, such as life and physical sciences, architecture, engineering, computer science and mathematics \cite{moss2015can}. Table \ref{tab:gender-by-category} summarizes these data, and Table \ref{tab:gender-by-category-grouped} summarizes it even further by coalescing occupation categories into broader groups to ease interpretation. For instance, STEM (Science, Technology, Engineering and Mathematics) fields are grouped into a single category, which helps us compare the large asymmetry between gender pronouns in these fields ($72\%$ of male defaults) to that of more evenly distributed fields such as healthcare ($50\%$). \begin{table}[H] \footnotesize\setlength{\tabcolsep}{2.5pt} \centering \small{ \begin{tabular}{|c|c|c|c|} \hline Category & Female ($\%$) & Male ($\%$) & Neutral ($\%$) \\ \hline \hline Office and administrative support & 11.015 & 58.812 & 16.954 \\ \hline Architecture and engineering & 2.299 & 72.701 & 10.92 \\ \hline Farming, fishing, and forestry & 12.179 & 62.179 & 14.744 \\ \hline Management & 11.232 & 66.667 & 12.681 \\ \hline Community and social service & 20.238 & 62.5 & 10.119 \\ \hline Healthcare support & 25.0 & 43.75 & 17.188 \\ \hline Sales and related & 8.929 & 62.202 & 16.964 \\ \hline Installation, maintenance, and repair & 5.22 & 58.333 & 17.125 \\ \hline Transportation and material moving & 8.81 & 62.976 & 17.5 \\ \hline Legal & 11.905 & 72.619 & 10.714 \\ \hline Business and financial operations & 7.065 & 67.935 & 15.58 \\ \hline Life, physical, and social science & 5.882 & 73.284 & 10.049 \\ \hline Arts, design, entertainment, sports, and media & 10.36 & 67.342 & 11.486 \\ \hline Education, training, and library & 23.485 & 53.03 & 9.091 \\ \hline Building and grounds cleaning and maintenance & 12.5 & 68.333 & 11.667 \\ \hline Personal care and service & 18.939 & 49.747 & 18.434 \\ \hline Healthcare practitioners and technical & 22.674 & 51.744 & 15.116 \\ \hline Production & 14.331 & 51.199 & 18.245 \\ \hline Computer and mathematical & 4.167 & 66.146 & 14.062 \\ \hline Construction and extraction & 8.578 & 61.887 & 17.525 \\ \hline Protective service & 8.631 & 65.179 & 12.5 \\ \hline Food preparation and serving related & 21.078 & 58.333 & 17.647 \\ \hline \hline Total & 11.76 & 58.93 & 15.939 \\ \hline \end{tabular} } \caption{Percentage of female, male and neutral gender pronouns obtained for each BLS occupation category, averaged over all occupations in said category and tested languages detailed in Table} \ref{tab:gender-neutral-languages}. Note that rows do not in general add up to $100\%$, as there is a fair amount of translated sentences for which we cannot obtain a gender pronoun. \label{tab:gender-by-category} \end{table} \begin{table}[H] \centering \small{ \begin{tabular}{|c|c|c|c|} \hline Category & Female ($\%$) & Male ($\%$) & Neutral ($\%$) \\ \hline \hline Service & 10.5 & 59.548 & 16.476 \\ \hline STEM & 4.219 & 71.624 & 11.181 \\ \hline Farming / Fishing / Forestry & 12.179 & 62.179 & 14.744 \\ \hline Corporate & 9.167 & 66.042 & 14.861 \\ \hline Healthcare & 23.305 & 49.576 & 15.537 \\ \hline Legal & 11.905 & 72.619 & 10.714 \\ \hline Arts / Entertainment & 10.36 & 67.342 & 11.486 \\ \hline Education & 23.485 & 53.03 & 9.091 \\ \hline Production & 14.331 & 51.199 & 18.245 \\ \hline Construction / Extraction & 8.578 & 61.887 & 17.525 \\ \hline \hline Total & 11.76 & 58.93 & 15.939 \\ \hline \end{tabular} } \caption{Percentage of female, male and neutral gender pronouns obtained for each of the merged occupation category, averaged over all occupations in said category and tested languages detailed in Table} \ref{tab:gender-neutral-languages}. Note that rows do not in general add up to $100\%$, as there is a fair amount of translated sentences for which we cannot obtain a gender pronoun. \label{tab:gender-by-category-grouped} \end{table} \definecolor{beige}{HTML}{D3CA39} Plotting histograms for the number of gender pronouns per occupation category sheds further light on how female, male and gender-neutral pronouns are differently distributed. The histogram in Figure~\ref{fig:histogram-female-grouped} suggests that the number of female pronouns is inversely distributed -- which is mirrored in the data for gender-neutral pronouns in Figure~\ref{fig:histogram-neutral-grouped} --, while the same data for male pronouns (shown in Figure~\ref{fig:histogram-male-grouped}) suggests a skew normal distribution. Furthermore we can see both on Figures \ref{fig:histogram-female-grouped} and \ref{fig:histogram-male-grouped} how STEM fields (labeled in \colorbox{beige}{beige exhibit predominantly male defaults -- amounting predominantly near $X = 0$ in the female histogram although much to the right in the male histogram. These values contrast with BLS' report of gender participation, which will be discussed in more detail in Section \ref{sec:comparison-women-participation}. \begin{figure}[H] \centering \includegraphics[width=10cm]{pictures/histograms/all-categories-stacked-Female-grouped} \caption{The data for the number of translated female pronouns per merged occupation category totaled among languages suggests and inverse distribution. STEM fields are nearly exclusively concentrated at $X = 0$, while more evenly distributed in fields such as production and healthcare (See Table}~\ref{tab:gender-by-category-grouped}) extends to higher values. \label{fig:histogram-female-grouped} \end{figure} \begin{figure}[H] \centering \includegraphics[width=10cm]{pictures/histograms/all-categories-stacked-Male-grouped} \caption{In contrast to Figure}\ref{fig:histogram-female-grouped} male pronouns are seemingly skew normally distributed, with a peak at $X = 6$. One can see how STEM fields concentrate mainly to the right ($X \geq 6$). \label{fig:histogram-male-grouped} \end{figure} \begin{figure}[H] \centering \includegraphics[width=10cm]{pictures/histograms/all-categories-stacked-Neutral-grouped} \caption{The scarcity of gender-neutral pronouns is manifest in their histogram. Once again, STEM fields are predominantly concentrated at $X = 0$.} \label{fig:histogram-neutral-grouped} \end{figure} We can also visualize male, female, and gender neutral histograms side by side, in which context is useful to compare the dissimilar distributions of translated STEM and Healthcare occupations (Figures \ref{fig:histogram-dodged-STEM} and \ref{fig:histogram-dodged-Healthcare} respectively). The number of translated female pronouns among languages is not normally distributed for any of the individual categories in Table \ref{tab:occupations}, but Healthcare is in many ways the most balanced category, which can be seen in comparison with STEM -- in which male defaults are second to most prominent. \begin{figure}[H] \centering \includegraphics[width=10cm]{pictures/histograms/categories/gender-dodged-STEM} \caption{Histograms for the distribution of the number of translated female, male and gender neutral pronouns totaled among languages are plotted side by side for job occupations in the STEM (Science, Technology, Engineering and Mathematics) field, in which male defaults are the second-to-most prominent (after Legal).} \label{fig:histogram-dodged-STEM} \end{figure} \begin{figure}[H] \centering \includegraphics[width=10cm]{pictures/histograms/categories/gender-dodged-Healthcare} \caption{Histograms for the distribution of the number of translated female, male and gender neutral pronouns totaled among languages are plotted side by side for job occupations in the Healthcare field, in which male defaults are least prominent.} \label{fig:histogram-dodged-Healthcare} \end{figure} The bar plots in Figure \ref{fig:gender-by-category} help us visualize how much of the distribution of each occupation category is composed of female, male and gender-neutral pronouns. In this context, STEM fields, which show a predominance of male defaults, are contrasted with Healthcare and educations, which show a larger proportion of female pronouns. \begin{figure}[H] \centering \includegraphics[width=10cm]{pictures/barplot-gender-by-category} \caption{Bar plots show how much of the distribution of translated gender pronouns for each occupation category (grouped as in Table \ref{tab:gender-by-category-grouped}) is composed of female, male and neutral terms. Legal and STEM fields exhibit a predominance of male defaults and contrast with Healthcare and Education, with a larger proportion of female and neutral pronouns. Note that in general the bars do not add up to $100\%$, as there is a fair amount of translated sentences for which we cannot obtain a gender pronoun. Categories are sorted with respect to the proportions of male, female and neutral translated pronouns respectively}. \label{fig:gender-by-category} \end{figure} Although computing our statistics over the set of all languages has practical value, this may erase subtleties characteristic to each individual idiom. In this context, it is also important to visualize how each language translates job occupations in each category. The heatmaps in Figures \ref{fig:heatmap-female}, \ref{fig:heatmap-male} and \ref{fig:heatmap-neutral} show the translation probabilities into female, male and neutral pronouns, respectively, for each pair of language and category (blue is $0\%$ and red is $100\%$). Both axes are sorted in these Figures, which helps us visualize both languages and categories in an spectrum of increasing male/female/neutral translation tendencies. In agreement with suggested stereotypes, \cite{moss2015can} STEM fields are second only to Legal ones in the prominence of male defaults. These two are followed by Arts \& Entertainment and Corporate, in this order, while Healthcare, Production and Education lie on the opposite end of the spectrum. \begin{figure}[H] \centering \includegraphics[width=10cm]{pictures/heatmap-languages-categories-Female} \caption{Heatmap for the translation probability into female pronouns for each pair of language and occupation category. Probabilities range from $0\%$ (blue) to $100\%$ (red), and both axes are sorted in such a way that higher probabilities concentrate on the bottom right corner.} \label{fig:heatmap-female} \end{figure} \begin{figure}[H] \centering \includegraphics[width=10cm]{pictures/heatmap-languages-categories-Male} \caption{Heatmap for the translation probability into male pronouns for each pair of language and occupation category. Probabilities range from $0\%$ (blue) to $100\%$ (red), and both axes are sorted in such a way that higher probabilities concentrate on the bottom right corner.} \label{fig:heatmap-male} \end{figure} \begin{figure}[H] \centering \includegraphics[width=10cm]{pictures/heatmap-languages-categories-Neutral} \caption{Heatmap for the translation probability into gender neutral pronouns for each pair of language and occupation category. Probabilities range from $0\%$ (blue) to $100\%$ (red), and both axes are sorted in such a way that higher probabilities concentrate on the bottom right corner.} \label{fig:heatmap-neutral} \end{figure} Our analysis is not truly complete without tests for statistical significant differences in the translation tendencies among female, male and gender neutral pronouns. We want to know for which languages and categories does Google Translate translate sentences with significantly more male than female, or male than neutral, or neutral than female, pronouns. We ran one-sided t-tests to assess this question for each pair of language and category and also totaled among either languages or categories. The corresponding p-values are presented in Tables \ref{tab:pvalues-MF}, \ref{tab:pvalues-MN}, \ref{tab:pvalues-NF} respectively. Language-Category pairs for which the null hypothesis was not rejected for a confidence level of $\alpha = .005$ are highlighted in blue. It is important to note that when the null hypothesis is accepted, we cannot discard the possibility of the complementary null hypothesis being rejected. For example, neither male nor female pronouns are significantly more common for Healthcare positions in the Estonian language, but female pronouns \emph{are} significantly more common for the same category in Finnish and Hungarian. Because of this, Language-Category pairs for which the complementary null hypothesis is rejected are painted in a darker shade of blue (see Table \ref{tab:pvalues-MF} for the three examples cited above. Although there is a noticeable level of variation among languages and categories, the null hypothesis that male pronouns are not significantly more frequent than female ones was consistently rejected for all languages and all categories examined. The same is true for the null hypothesis that male pronouns are not significantly more frequent than gender neutral pronouns, with the one exception of the Basque language (which exhibits a rather strong tendency towards neutral pronouns). The null hypothesis that neutral pronouns are not significantly more frequent than female ones is accepted with much more frequency, namely for the languages Malay, Estonian, Finnish, Hungarian, Armenian and for the categories Farming \& Fishing \& Forestry, Healthcare, Legal, Arts \& Entertainment, Education. In all three cases, the null hypothesis corresponding to the aggregate for all languages and categories is rejected. We can learn from this, in summary, that Google Translate translates male pronouns more frequently than both female and gender neutral ones, either in general for Language-Category pairs or consistently among languages and among categories (with the notable exception of the Basque idiom). \begin{table}[H] \centering \footnotesize\setlength{\tabcolsep}{2.5pt} \small \begin{tabular}{|m{1.75cm}|cccccccccccc|c|} \hline & Mal. & Est. & Fin. & Hun. & Arm. & Ben. & Jap. & Tur. & Yor. & Bas. & Swa. & Chi. & Total \\ \hline Service & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline STEM & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline Farming Fishing Forestry & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.603$ & \cellcolor{blue!25}$.786$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$*$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline Corporate & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline Healthcare & $<\alpha$ & \cellcolor{blue!25}$.938$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$.999$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline Legal & $<\alpha$ & \cellcolor{blue!25}$.368$ & \cellcolor{blue!25}$.632$ & \cellcolor{blue!25}$.368$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.086$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline Arts Entertainment & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.08$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline Education & $<\alpha$ & \cellcolor{blue!25}$.808$ & \cellcolor{blue!25}$.333$ & \cellcolor{blue!25}$.263$ & \cellcolor{blue!25}$.588$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.417$ & $<\alpha$ & \cellcolor{blue!25}$.052$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline Production & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.5$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.159$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline Construction Extraction & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.16$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline Total & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline \end{tabular} \caption{Computed p-values relative to the null hypothesis that the number of translated male pronouns is not significantly greater than that of female pronouns, organized for each language and each occupation category. Cells corresponding to the acceptance of the null hypothesis are marked in blue, and within those cells, those corresponding to cases in which the complementary null hypothesis (that the number of female pronouns is not significantly greater than that of male pronouns) was rejected are marked with a darker shade of the same color. A significance level of $\alpha =.05$ was adopted. Asterisks indicate cases in which all pronouns are translated with gender neutral pronouns.} \label{tab:pvalues-MF} \end{table} \begin{table}[H] \centering \footnotesize\setlength{\tabcolsep}{2.5pt} \small \begin{tabular}{|m{1.75cm}|cccccccccccc|c|} \hline & Mal. & Est. & Fin. & Hun. & Arm. & Ben. & Jap. & Tur. & Yor. & Bas. & Swa. & Chi. & Total \\ \hline Service & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!45}$1.0$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline STEM & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!45}$.984$ & $<\alpha$ & \cellcolor{blue!25}$.07$ & $<\alpha$ \\ \hline Farming Fishing Forestry & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.135$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.068$ & \cellcolor{blue!45}$1.0$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline Corporate & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!45}$1.0$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline Healthcare & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.39$ & \cellcolor{blue!45}$1.0$ & $<\alpha$ & \cellcolor{blue!25}$.088$ & $<\alpha$ \\ \hline Legal & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.145$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.771$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline Arts Entertainment & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.07$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!45}$1.0$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline Education & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.093$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.5$ & $<\alpha$ & \cellcolor{blue!25}$.068$ & $<\alpha$ \\ \hline Production & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.412$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline Construction Extraction & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.92$ & \cellcolor{blue!45}$1.0$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline Total & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!45}$1.0$ & $<\alpha$ & $<\alpha$ & $<\alpha$ \\ \hline \end{tabular} \caption{Computed p-values relative to the null hypothesis that the number of translated male pronouns is not significantly greater than that of gender neutral pronouns, organized for each language and each occupation category. Cells corresponding to the acceptance of the null hypothesis are marked in blue, and within those cells, those corresponding to cases in which the complementary null hypothesis (that the number of gender neutral pronouns is not significantly greater than that of male pronouns) was rejected are marked with a darker shade of the same color. A significance level of $\alpha =.05$ was adopted. Asterisks indicate cases in which all pronouns are translated with gender neutral pronouns.} \label{tab:pvalues-MN} \end{table} \begin{table}[H] \centering \footnotesize\setlength{\tabcolsep}{2.5pt} \small \begin{tabular}{|m{1.75cm}|cccccccccccc|c|} \hline & Mal. & Est. & Fin. & Hun. & Arm. & Ben. & Jap. & Tur. & Yor. & Bas. & Swa. & Chi. & Total \\ \hline Service & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$.981$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!45}$1.0$ & $<\alpha$ & $<\alpha$ \\ \hline STEM & \cellcolor{blue!25}$.84$ & \cellcolor{blue!45}$.978$ & \cellcolor{blue!45}$.998$ & \cellcolor{blue!45}$.993$ & \cellcolor{blue!25}$.84$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.079$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.84$ & $<\alpha$ & $<\alpha$ \\ \hline Farming Fishing Forestry & \cellcolor{blue!25}$*$ & \cellcolor{blue!25}$*$ & \cellcolor{blue!45}$.999$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!25}$*$ & \cellcolor{blue!25}$.167$ & \cellcolor{blue!25}$.169$ & \cellcolor{blue!25}$.292$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$*$ & \cellcolor{blue!25}$.083$ & \cellcolor{blue!25}$.147$ \\ \hline Corporate & \cellcolor{blue!25}$*$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$.996$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!45}$.977$ & $<\alpha$ & $<\alpha$ \\ \hline Healthcare & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!25}$.086$ & $<\alpha$ & \cellcolor{blue!25}$.87$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!45}$1.0$ & $<\alpha$ & \cellcolor{blue!45}$.977$ \\ \hline Legal & \cellcolor{blue!25}$*$ & \cellcolor{blue!45}$.961$ & \cellcolor{blue!45}$.985$ & \cellcolor{blue!45}$.961$ & \cellcolor{blue!25}$*$ & $<\alpha$ & \cellcolor{blue!25}$.086$ & \cellcolor{blue!25}$*$ & \cellcolor{blue!25}$.178$ & $<\alpha$ & \cellcolor{blue!25}$*$ & \cellcolor{blue!25}$*$ & \cellcolor{blue!25}$.072$ \\ \hline Arts Entertainment & \cellcolor{blue!25}$.92$ & \cellcolor{blue!45}$.994$ & \cellcolor{blue!45}$.999$ & \cellcolor{blue!45}$.998$ & \cellcolor{blue!45}$.998$ & \cellcolor{blue!25}$.067$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!25}$.92$ & \cellcolor{blue!25}$.162$ & \cellcolor{blue!25}$.097$ \\ \hline Education & \cellcolor{blue!25}$*$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$.999$ & \cellcolor{blue!45}$.999$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!25}$.058$ & $<\alpha$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!25}$.164$ & \cellcolor{blue!25}$.052$ & \cellcolor{blue!45}$.995$ & \cellcolor{blue!25}$.052$ & \cellcolor{blue!45}$.992$ \\ \hline Production & \cellcolor{blue!45}$.996$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!25}$.113$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!45}$1.0$ & $<\alpha$ & $<\alpha$ \\ \hline Construction Extraction & \cellcolor{blue!25}$.84$ & \cellcolor{blue!45}$.996$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!25}$*$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!45}$1.0$ & $<\alpha$ & $<\alpha$ \\ \hline Total & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & \cellcolor{blue!45}$1.0$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & $<\alpha$ & \cellcolor{blue!45}$1.0$ & $<\alpha$ & $<\alpha$ \\ \hline \end{tabular} \caption{Computed p-values relative to the null hypothesis that the number of translated gender neutral pronouns is not significantly greater than that of female pronouns, organized for each language and each occupation category. Cells corresponding to the acceptance of the null hypothesis are marked in blue, and within those cells, those corresponding to cases in which the complementary null hypothesis (that the number of female pronouns is not significantly greater than that of gender neutral pronouns) was rejected are marked with a darker shade of the same color. A significance level of $\alpha =.05$ was adopted. Asterisks indicate cases in which all pronouns are translated with gender neutral pronouns.} \label{tab:pvalues-NF} \end{table} \section{Distribution of translated gender pronouns per language} We have taken the care of experimenting with a fair amount of different gender neutral languages. Because of that, another sensible way of coalescing our data is by language groups, as shown in Table \ref{tab:gender-by-language}. This can help us visualize the effect of different cultures in the genesis -- or lack thereof -- of gender bias. Nevertheless, the barplots in Figure \ref{fig:gender-by-language} are perhaps most useful to identifying the difficulty of extracting a gender pronoun when translating from certain languages. Basque is a good example of this difficulty, although the quality of Bengali, Yoruba, Chinese and Turkish translations are also compromised. \begin{table}[H] \small{ \centering \begin{tabular}{|c|c|c|c|} \hline Language & Female ($\%$) & Male ($\%$) & Neutral ($\%$) \\ \hline \hline Malay & 3.827 & 88.420 & 0.000 \\ \hline Estonian & 17.370 & 72.228 & 0.491 \\ \hline Finnish & 34.446 & 56.624 & 0.000 \\ \hline Hungarian & 34.347 & 58.292 & 0.000 \\ \hline Armenian & 10.010 & 82.041 & 0.687 \\ \hline Bengali & 16.765 & 37.782 & 25.63 \\ \hline Japanese & 0.000 & 66.928 & 24.436 \\ \hline Turkish & 2.748 & 62.807 & 18.744 \\ \hline Yoruba & 1.178 & 48.184 & 38.371 \\ \hline Basque & 0.393 & 5.496 & 58.587 \\ \hline Swahili & 14.033 & 76.644 & 0.000 \\ \hline Chinese & 5.986 & 51.717 & 24.338 \\ \hline \end{tabular} \caption{Percentage of female, male and neutral gender pronouns obtained for each language, averaged over all occupations detailed in Table} \ref{tab:gender-neutral-languages}. Note that rows do not in general add up to $100\%$, as there is a fair amount of translated sentences for which we cannot obtain a gender pronoun. \label{tab:gender-by-language} } \end{table} \begin{figure}[H] \centering \includegraphics[width=10cm]{pictures/barplot-gender-by-language} \caption{The distribution of pronominal genders per language also suggests a tendency towards male defaults, with female pronouns reaching as low as $0.196\%$ and $1.865\%$ for Japanese and Chinese respectively. Once again not all bars add up to $100\%$ , as there is a fair amount of translated sentences for which we cannot obtain a gender pronoun, particularly in Basque. Among all tested languages, Basque was the only one to yield more gender neutral than male pronouns, with Bengali and Yoruba following after in this order. Languages are sorted with respect to the proportions of male, female and neutral translated pronouns respectively.} \label{fig:gender-by-language} \end{figure} \section{Distribution of translated gender pronouns for varied adjectives} We queried the 1000 most frequently used adjectives in English, as classified in the COCA corpus [https://corpus.byu.edu/coca/], but since not all of them were readily applicable to the sentence template we used, we filtered the N adjectives that would fit the templates and made sense for describing a human being. The list of adjectives extracted from the corpus is available on the Github repository: https://github.com/marceloprates/Gender-Bias. Apart from occupations, which we have exhaustively examined by collecting labor data from the U.S. Bureau of Labor Statistics, we have also selected a small subset of adjectives from the Corpus of Contemporary American English (COCA) https://corpus.byu.edu/coca/, in an attempt to provide preliminary evidence that the phenomenon of gender bias may extend beyond the professional context examined in this paper. Because a large number of adjectives are not applicable to human subjects, we manually curated a reasonable subset of such words. The template used for adjectives is similar to that used for occupations, and is provided again for reference in Table \ref{tab:templates}. Once again the data points towards male defaults, but some variation can be observed throughout different adjectives. Sentences containing the words \emph{Shy}, \emph{Attractive}, \emph{Happy}, \emph{Kind} and \emph{Ashamed} are predominantly female translated (\emph{Attractive} is translated as female and gender-neutral in equal parts), while \emph{Arrogant}, \emph{Cruel} and \emph{Guilty} are disproportionately translated with male pronouns (\emph{Guilty} is in fact never translated with female or neutral pronouns). \begin{table}[H] \small{ \centering \begin{tabular}{|c|c|c|c|} \hline Adjective & Female ($\%$) & Male ($\%$) & Neutral ($\%$) \\ \hline \hline Happy & 36.364 & 27.273 & 18.182 \\ \hline Sad & 18.182 & 36.364 & 18.182 \\ \hline Right & 0.000 & 63.636 & 27.273 \\ \hline Wrong & 0.000 & 54.545 & 36.364 \\ \hline Afraid & 9.091 & 54.545 & 0.000 \\ \hline Brave & 9.091 & 63.636 & 18.182 \\ \hline Smart & 18.182 & 45.455 & 18.182 \\ \hline Dumb & 18.182 & 36.364 & 18.182 \\ \hline Proud & 9.091 & 72.727 & 9.091 \\ \hline Strong & 9.091 & 54.545 & 18.182 \\ \hline Polite & 18.182 & 45.455 & 18.182 \\ \hline Cruel & 9.091 & 63.636 & 18.182 \\ \hline Desirable & 9.091 & 36.364 & 45.455 \\ \hline Loving & 18.182 & 45.455 & 27.273 \\ \hline Sympathetic & 18.182 & 45.455 & 18.182 \\ \hline Modest & 18.182 & 45.455 & 27.273 \\ \hline Successful & 9.091 & 54.545 & 27.273 \\ \hline Guilty & 0.000 & 72.727 & 0.000 \\ \hline Innocent & 9.091 & 54.545 & 9.091 \\ \hline Mature & 36.364 & 36.364 & 9.091 \\ \hline Shy & 36.364 & 27.273 & 27.273 \\ \hline \hline Total & 30.3 & 98.1 & 41.7 \\ \hline \end{tabular} \caption{Number of female, male and neutral pronominal genders in the translated sentences for each selected adjective.} \label{tab:gender-by-adjective} } \end{table} \begin{figure}[H] \centering \includegraphics[width=10cm]{pictures/barplot-adjectives} \caption{The distribution of pronominal genders for each word in Table} \ref{tab:adjectives} shows how stereotypical gender roles can play a part on the automatic translation of simple adjectives. One can see that adjectives such as \emph{Shy} and \emph{Desirable}, \emph{Sad} and \emph{Dumb} amass at the female side of the spectrum, contrasting with \emph{Proud}, \emph{Guilty}, \emph{Cruel} and \emph{Brave} which are almost exclusively translated with male pronouns. \label{fig:barplot-adjectives} \end{figure} \section{Comparison with women participation data across job positions}\label{sec:comparison-women-participation} A sensible objection to the conclusions we draw from our study is that the perceived gender bias in Google Translate results stems from the fact that possibly female participation in some job positions is itself low. We must account for the possibility that the statistics of gender pronouns in Google Translate outputs merely reflects the demographics of male-dominated fields (male-dominated fields can be considered those that have less than 25\% of women participation\cite{WB2014}, according to the U.S. Department of Labor Women's Bureau). In this context, the argument in favor of a critical revision of statistic translation algorithms weakens considerably, and possibly shifts the blame away from these tools. The U.S. Bureau of Labor Statistics data summarized in Table \ref{tab:occupations} contains statistics about the percentage of women participation in each occupation category. This data is also available for each individual occupation, which allows us to compute the frequency of women participation for each 12-quantile. We carried the same computation in the context of frequencies of translated female pronouns, and the resulting histograms are plotted side-by-side in Figure \ref{fig:histogram-compare-gt-real}. The data shows us that Google Translate outputs fail to follow the real-world distribution of female workers across a comprehensive set of job positions. The distribution of translated female pronouns is consistently inversely distributed, with female pronouns accumulating in the first 12-quantile. By contrast, BLS data shows that female participation peaks in the fourth 12-quantile and remains significant throughout the next ones. \begin{figure}[H] \centering \includegraphics[width=10cm]{pictures/histogram-compare-gt-real} \caption{Women participation ($\%$) data obtained from the U.S. Bureau of Labor Statistics allows us to assess whether the Google Translate bias towards male defaults is at least to some extent explained by small frequencies of female workers in some job positions. Our data does not make a very good case for that hypothesis: the total frequency of translated female pronouns (in blue) for each 12-quantile does not seem to respond to the higher proportion of female workers (in yellow) in the last quantiles.} \label{fig:histogram-compare-gt-real} \end{figure} Averaged over occupations and languages, sentences are translated with female pronouns $11.76\%$ of the time. In contrast, the gender participation frequency for female workers averaged over all occupations in the BLS report yields a consistently larger figure of $35.94\%$. The variance reported for the translation results is also lower, at $\approx 0.028$ in contrast with the report's $\approx 0.067$. We ran an one-sided t-test to evaluate the null hypothesis that the female participation frequency is not significantly greater then the GT female pronoun frequency for the same job positions, obtaining a p-value $p \approx 6.2 10^{-94}$ vastly inferior to our confidence level of $\alpha = 0.005$ and thus rejecting H0 and concluding that Google Translate's female translation frequencies sub-estimates female participation frequencies in US job positions. As a result, it is not possible to understand this asymmetry as a reflection of workplace demographics, and the prominence of male defaults in Google Translate is, we believe, yet lacking a clear justification. \section{Discussion} At the time of the writing up this paper, Google Translate offered only one official translation for each input word, along with a list of synonyms. In this context, all experiments reported here offer an analysis of a ``screenshot'' of that tool as of August 2018, the moment they were carried out. A preprint version of this paper was posted the in well-known Cornell University-based \url{arXiv.org} open repository on September 6, 2018. The manuscript soon enjoyed a significant amount of media coverage, featuring on \emph{The Register} \cite{clauburn2018boffins}, \emph{Datanews} \cite{bellens2018sexiste}, \emph{t3n} \cite{rixecker2018sexistische}, among others, and more recently on \emph{Slator} \cite{Dino2019hesaid} and \emph{Jornal do Comercio} \cite{knebel2019nosrobos}. On December 6, 2018 the company's policy changed, and a statement was released detailing their efforts to reduce gender bias on Google Translate, which included a new feature presenting the user with a feminine as well as a masculine official translation (Figure \ref{fig:gender-before-after}). According to the company, this decision is part of a broader goal of promoting fairness and reducing biases in machine learning. They also acknowledged the technical reasons behind gender bias in their model, stating that: \begin{quote} \textsl{Google Translate learns from hundreds of millions of already-translated examples from the web. Historically, it has provided only one translation for a query, even if the translation could have either a feminine or masculine form. So when the model produced one translation, it inadvertently replicated gender biases that already existed. For example: it would skew masculine for words like ``strong'' or ``doctor,'' and feminine for other words, like ``nurse'' or ``beautiful.''} \end{quote} \begin{figure}[h] \centering \includegraphics[width=\linewidth]{pictures/gender-before-after.png} \caption{Comparison between the GUI of Google Translate before (left) and after (right) the introduction of the new feature intended to promote gender fairness in translation. The results described in this paper relate to the older version.} \label{fig:gender-before-after} \end{figure} Their statement is very similar to the conclusions drawn on this paper, as is their motivation for redesigning the tool. As authors, we are incredibly happy to see our vision and beliefs align with those of Google in such a short timespan from the initial publishing of our work, although the company's statement does not cite any study or report in particular and thus we cannot know for sure whether this paper had an effect on their decision or not. Regardless of whether their decision was monocratic, guided by public opinion or based on published research, we understand it as an important first step on an ongoing fight against algorithmic bias, and we praise the Google Translate team for their efforts. Google Translate's new feminine and masculine forms for translated sentences exemplifies how, as this paper also suggests, machine learning translation tools can be \emph{debiased}, dropping the need for resorting to a balanced training set. However, it should be noted that important as it is, GT's new feature is still a first step. It does not address all of the shortcomings described in this paper, and the limited language coverage means that many users will still experience gender biased translation results. Furthermore, the system does not yet have support for non-binary results, which may exclude part of their user base. In addition, one should note that further evidence is mounting about the kind of bias examined in this paper: it is becoming clear that this is a statistical phenomenon independent from any proprietary tool. In this context, the research carried out in \cite{bolukbasi2016man} presents a very convincing argument for the sensitivity of word embeddings to gender bias in the training dataset. This suggests that machine translation engineers should be especially aware of their training data when designing a system. It is not feasible to train these models on unbiased texts, as they are probably scarce. What must be done instead is to engineer solutions to remove bias from the system after an initial training, which seems to be the goal of Google Translate's recent efforts. Fortunately, as \cite{bolukbasi2016man} also show, debiasing can be implemented with relatively low effort and modest resources. The technology to promote social justice on machine translation in particular and machine learning in general is often already available. The most significant effort which must be taken in this context is to promote social awareness on these issues so that society can be invited into the conversation. \section{Conclusions} In this paper, we have provided evidence that statistical translation tools such as Google Translate can exhibit gender biases and a strong tendency towards male defaults. Although implicit, these biases possibly stem from the real world data which is used to train them, and in this context possibly provide a window into the way our society talks (and writes) about women in the workplace. In this paper, we suggest that and test the hypothesis that statistical translation tools can be probed to yield insights about stereotypical gender roles in our society -- or at least in their training data. By translating professional-related sentences such as ``He/She is an engineer'' from gender neutral languages such as Hungarian and Chinese into English, we were able to collect statistics about the asymmetry between female and male pronominal genders in the translation outputs. Our results show that male defaults are not only prominent, but exaggerated in fields suggested to be troubled with gender stereotypes, such as STEM (Science, Technology, Engineering and Mathematics) occupations. And because Google Translate typically uses English as a \emph{lingua franca} to translate between other languages (e.g. Chinese $\rightarrow$ English $\rightarrow$ Portuguese) \cite{LanguageSupportNMT2018,boitet2010mt}, our findings possibly extend to translations between gender neutral languages and non-gender neutral languages (apart from English) in general, although we have not tested this hypothesis. Our results seem to suggest that this phenomenon extends beyond the scope of the workplace, with the proportion of female pronouns varying significantly according to adjectives used to describe a person. Adjectives such as \emph{Shy} and \emph{Desirable} are translated with a larger proportion of female pronouns, while \emph{Guilty} and \emph{Cruel} are almost exclusively translated with male ones. Different languages also seemingly have a significant impact in machine gender bias, with Hungarian exhibiting a better equilibrium between male and female pronouns than, for instance, Chinese. Some languages such as Yoruba and Basque were found to translate sentences with gender neutral pronouns very often, although this is the exception rather than the rule and Basque also exhibits a high frequency of phrases for which we could not automatically extract a gender pronoun. In order to strengthen our results, we ran pronominal gender translation statistics against the U.S. Bureau of Labor Statistics data on the frequency of women participation for each job position. Although Google Translate exhibits male defaults, this phenomenon may merely reflect the unequal distribution of male and female workers in some job positions. To test this hypothesis, we compared the distribution of female workers with the frequency of female translations, finding no correlation between said variables. Our data shows that Google Translate outputs fail to reflect the real-world distribution of female workers, under-estimating the expected frequency. That is to say that even if we do not expect a 50:50 distribution of translated gender pronouns, Google Translate exhibits male defaults in a greater frequency that job occupation data alone would suggest. The prominence of male defaults in Google Translate is therefore to the best of our knowledge yet lacking a clear justification. We think this work sheds new light on a pressing ethical difficulty arising from modern statistical machine translation, and hope that it will lead to discussions about the role of AI engineers on minimizing potential harmful effects of the current concerns about machine bias. We are optimistic that unbiased results can be obtained with relatively little effort and marginal cost to the performance of current methods, to which current \emph{debiasing} algorithms in the scientific literature are a testament. \section{Acknowledgments} This study was financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - Brasil (CAPES) - Finance Code 001 and the Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq). This is a pre-print of an article published in Neural Computing and Applications. \vskip 0.2in

4951652e36b8e9fd94355f63d8c424aafaadd3f5

\section{Introduction} \label{sec:Introduction} One of the most studied and central problems in graph mining is the identification of cohesive subgraphs. This problem has been raised in several contexts, from social network analysis \cite{DBLP:journals/cn/KumarRRT99} to finding functional motifs in biological networks \cite{bioinfo2006}. Different definitions of cohesive graphs have been proposed and applied in literature. One of the most remarkable example is clique, and finding a maximum size clique is a well-known and studied problem in theoretical computer science \cite{DBLP:conf/coco/Karp72}. Most of the definitions of cohesive subgraph lead to NP-hard problems, in some cases even hard to approximate. For example, finding a clique of maximum size in a graph $G=(V,E)$ is an NP-hard problem \cite{DBLP:conf/coco/Karp72} and it is even hard to approximate within factor $O(|V|^{1 - \varepsilon})$, for each $\varepsilon > 0$ \cite{DBLP:journals/toc/Zuckerman07}. A definition of dense subgraph that leads to a polynomial-time algorithm is that of average-degree density. For this problem, called {\sf Densest Subgraph}, Goldberg gave an elegant polynomial-time algorithm \cite{Goldberg:1984:FMD:894477}. Furthermore, a linear-time greedy algorithm that achieves an approximation factor of $\frac{1}{2}$ for {\sf Densest Subgraph} has been given in \cite{DBLP:conf/swat/AsahiroITT96,DBLP:conf/approx/Charikar00}. The {\sf Densest Subgraph} problem aims at finding a single subgraph, but in many applications it is of interest finding a collection of dense subgraphs of a given graph. More precisely, it is interesting to compute a collection of subgraphs having maximum density in a given graph. A recent approach proposed in \cite{DBLP:journals/datamine/GalbrunGT16} asks for a collection of top $k$ densest, possibly overlapping, subgraphs (denoted as {\sf Top-k-Overlapping Densest Subgraphs}{}), since in many real-world cases dense subgraphs are related to non disjoint communities. As pointed out in \cite{DBLP:journals/im/LeskovecLDM09,DBLP:journals/datamine/GalbrunGT16}, for example hubs are vertices that may be part of several communities and hence of several densest subgraphs, thus motivating the quest for overlapping distinct subgraphs. {\sf Top-k-Overlapping Densest Subgraphs}{}, proposed in \cite{DBLP:journals/datamine/GalbrunGT16}, addresses this problem by looking for a collection of $k$ subgraphs that maximize an objective function that takes into account both the density of the subgraphs and the distance between the subgraphs of the solution, thus allowing an overlap between the subgraphs which depends on a parameter $\lambda$. When $\lambda$ is small, then the density plays a dominant role in the objective function, so the output subgraphs can share a significant part of vertices. On the other hand, if $\lambda$ is large, then the subgraphs will share few or no vertices, so the subgraphs may be disjoint. An approach similar to {\sf Top-k-Overlapping Densest Subgraphs}{} was proposed in \cite{DBLP:conf/wsdm/BalalauBCGS15}, where the goal is to find a set of $k$ subgraphs of maximum density, such that the maximum pairwise Jaccard coefficient of the subgraphs is bounded. A dynamic variant of the problem, whose goal is finding a set of $k$ disjoint subgraphs, has been recently considered in \cite{DBLP:conf/cikm/NasirGMG17}. {\sf Top-k-Overlapping Densest Subgraphs}{} has been shown to be approximable within factor $\frac{1}{10}$ \cite{DBLP:journals/datamine/GalbrunGT16}, while its computational complexity has been left open \cite{DBLP:journals/datamine/GalbrunGT16}. In this paper, we present algorithmic and complexity results for {\sf Top-k-Overlapping Densest Subgraphs}{} when $k$ is less than the number of vertices in the graph. This last assumption (required in Section \ref{sec:Approximation}) is reasonable, for example notice that in the experimental results presented in \cite{DBLP:journals/datamine/GalbrunGT16} $k$ is equal to $20$, even for graphs having thousands or millions of vertices. Concerning the approximation of the problem, we provide in Section \ref{sec:Approximation} a $\frac{2}{3}$-approximation algorithm when $k$ is a constant, and we present a $\frac{1}{2}$-approximation algorithm when $k$ is smaller than the size of the vertex set. From the computational complexity point of view, we show in Section \ref{sec:complexity} that {\sf Top-k Overlapping Densest Subgraphs} is NP-hard even if $k=3$ (that is we ask for three densest subgraphs), when $\lambda = 3|V|^3$, for an input graph $G=(V,E)$. We conclude the paper in Section \ref{sec:conclusion} with some open problems. Some of the proofs and the pseudocode of some algorithms are omitted due to page limit. \section{Definitions} \label{sec:Def} In this section, we present some definitions that will be useful in the rest of the paper. Moreover, we provide the formal definition of the problem we are interested in. All the graphs we consider in this paper are undirected. Given a graph $G=(V,E)$, and a set $V' \subseteq V$, we denote by $G[V']$ the \emph{subgraph} of $G$ induced by $V'$, formally $G[V']=(V',E')$, where $E'$ is defined as follows: $ E'= \{ \{ u,v \}: \{ u,v \} \in E \wedge u,v \in V' \} $. If $G[V']$ is a subgraph of $G[V'']$, with $V' \subseteq V'' \subseteq V$, then $G[V'']$ is a \emph{supergraph} of $G[V']$. $G[V']$ is a proper subgraph of $G[V'']$, if $V' \subset V'' \subseteq V$; in this case $G[V'']$ is a proper supergraph of $G[V']$. A subgraph $G[V']$ of $G$ is a \emph{singleton}, if $|V'|=1$. Given a subset $U \subseteq V$, we denote by $E(U)$ the set of edges of $G$ having both endpoints in $U$. Moreover, we denote by $E(V_1,V_2)$, with $V_1 \subseteq V'$, $V_2 \subseteq V'$ and $V_1 \cap V_2 = \emptyset$, the set of edges having exactly one endpoint in $V_1$ and exactly one endpoint in $V_2$, formally $ E(V_1,V_2)=\{\{u,v\}: u\in V_1, v \in V_2 \} $. Two subgraphs $G[V_1]$ and $G[V_2]$ of a graph $G=(V,E)$ are called \emph{distinct} when $V_1 \neq V_2$. Next, we present the definition of \emph{crossing subgraphs}, which is fundamental in Section~\ref{subsec:approxNotConstant}. \begin{definition} Given a graph $G=(V,E)$, let $G[V_1]$ and $G[V_2]$ be two subgraphs of $G=(V,E)$. $G[V_1]$ and $G[V_2]$ are \emph{crossing} when $V_1 \cap V_2 \neq \emptyset$, $V_1 \setminus V_2 \neq \emptyset$ and $V_2 \setminus V_1 \neq \emptyset$ (notice that $V_1 \nsubseteq V_2$ and $V_2 \nsubseteq V_1$). \end{definition} Consider the example of Fig. \ref{fig:DenseSubgraph}. The two subgraphs induced by $\{ v_5,v_6, v_7, v_8, v_9, v_{10} \}$ and $\{ v_1, v_2, v_3, v_4, v_5 \}$ are crossing, while the two subgraphs induced by $\{ v_5,v_6, v_7, v_8, v_9, v_{10} \}$ and $\{ v_5,v_6, v_7, v_8, v_9 \}$ are not crossing. Now, we present the definition of density of a subgraph. \begin{definition} Given a graph $G=(V,E)$ and a subgraph $G[V']=(V',E')$, with $V' \subseteq V$, the density of $G[V']$, denoted by $dens(G[V'])$, is defined as $ dens(G[V'])=\frac{|E'|}{|V'|} $. \end{definition} A \emph{densest subgraph} of a graph $G=(V,E)$ is a subgraph $G[U]$, with $U \subseteq V$, that maximizes $dens(G[U])$, among the subgraphs of $G$. In the example of Fig. \ref{fig:DenseSubgraph} the subgraph induced by $\{ v_5, v_6, v_7, v_8, v_9, v_{10} \}$ is the densest subgraph and has density $\frac{11}{6}$. Given a graph $G=(V,E)$ and a collection of subgraphs $\mathcal{W}= \{G[W_1], \dots, G[W_k] \}$ where each $G[W_i]$ is a subgraph of $G$, that is $W_i \subseteq V$, with $1 \leq i \leq k$, then the density of $\mathcal{W}$, denoted by $dens(\mathcal{W})$, is defined as $ dens(\mathcal{W}) = \sum_{i=1}^{k} dens (G[W_i]) $. The goal of the problem we are interested in is to find a collection of $k$, with $1 \leq k < |V|$, possibly overlapping subgraphs having high density. However, allowing overlap leads to a solution that may contain $k$ copies of the same subgraph. To address such an issue, in \cite{DBLP:journals/datamine/GalbrunGT16} a distance function between subgraphs of the collection is included in the objective function (to be maximized). We present here the distance function between two subgraphs presented in \cite{DBLP:journals/datamine/GalbrunGT16}. \begin{definition} Given a graph $G=(V,E)$ and two subgraphs $G[U]$, $G[Z]$, with $U,Z \subseteq V$, define the distance function $d: 2^{G[V]} \times 2^{G[V]} \rightarrow \mathbb{R_{+}}$ between $G[U]$ and $G[Z]$ as follows: \begin{equation*} d(G[U],G[Z]) = \left\{ \begin{array}{ll} 2-\frac{|U \cap Z|^2}{|U||Z|} & \text{if } U \neq Z,\\ 0 & \text{else}. \end{array} \right. \end{equation*} \end{definition} Notice that $d(G[U],G[Z]) \leq 2$, for each $U,Z \subseteq V$. Now, we are able to define the problem we are interested in. \begin{problem}{\sf Top-k-Overlapping Densest Subgraphs}{} \\ \noindent \textbf{Input:} a graph $G=(V,E)$, a parameter $\lambda > 0$. \\ \textbf{Output:} a set $\mathcal{W} = \{ G[W_1], \dots , G[W_k] \}$ of $k$ subgraphs, with $1 \leq k <|V|$ and $W_i \subseteq V$, $1 \leq i \leq k$, that maximizes the following value \[ r(\mathcal{W})= dens(\mathcal{W})+ \lambda \sum_{i=1}^{k-1} \sum_{j=i+1}^k d(G[W_i],G[W_j]) \] \end{problem} We assume in what follows that $|V| > 5$ (it is required in the proof of Lemma \ref{lem:secondPartCorrect}). Notice that, when $|V| \leq 5$, {\sf Top-k-Overlapping Densest Subgraphs}{} can be solved optimally in constant time. \begin{figure} \centering \includegraphics[scale=.45]{DenseSubgraph.pdf} \caption{A graph and a solution $\mathcal{W}$ of {\sf Top-k-Overlapping Densest Subgraphs}{}, for $k=3$, consisting of the three subgraphs included in boxes. } \label{fig:DenseSubgraph} \end{figure} \subsection{Goldberg's Algorithm and Extended Goldberg's Algorithm} Goldberg's Algorithm \cite{Goldberg:1984:FMD:894477} computes in polynomial time an optimal solution for the \textsf{Densest-Subgraph} problem that, given as input a graph $G=(V,E)$, asks for a subgraph $G[V']$ in $G$ having maximum density. Goldberg's Algorithm reduces $\ensuremath{\mathsf{Densest \mhyphen Subgraph}}$ to the problem of computing a minimum cut in a weighted auxiliary graph computed by adding two vertices $s$ (the sink) and $t$ (the target) to $V$, where both $s$ and $t$ are connected to every vertex of $V$. The time complexity of Goldberg's Algorithm is $O(|V||E|\log(|V|^2/|E|))$ applying parametric flow algorithm \cite{DBLP:journals/siamcomp/GalloGT89}. Given a graph $G=(V,E)$ and a subgraph $G[V']$, with $V' \subseteq V$, we denote by \textit{Dense-Subgraph}$(G[V'])$ a densest subgraph in $G[V']$, which can be computed with Goldberg's Algorithm. \begin{comment} Formally, $G'=(V',E',w)$ is defined as follows: \[ V'= V \cup \{s,t\} \] \[ E'=E \cup \{ \{ s,v\} \{t,v \} : v \in V \} \] The weight function is defined as follows. For each edge $\{s,v\} \in E'$, with $v \in V \cap V'$, we have \[ w(\{s,v\})= d_G(v) \] For each edge $\{u,v\} \in E \cap E'$, with $u,v \in V \cap V'$, we have \[ w(\{u,v\})= 1 \] For each edge $\{v,t\} \in E'$, with $v \in V \cap V'$, we have \[ w(\{u,v\})= z \] where $z$, with $0 < z \leq \frac{|V||V|-1}{|V|}$. The algorithm computes whether there exists a densest subgraph of density $z$ by checking if there exists a minimum cut in graph $G'$ of value \[ \sum_{v \in V\setminus S} d(v) + |e(S,V\setminus S)|+ c|S| \leq 2|E| \] where $S \subseteq V' \setminus \{s,t\}$. We denote by $G[S]$, with $S \subseteq V$, be the subgraph returned by Goldberg's Algorithm. \end{comment} In this paper, we consider also a modification of Goldberg's Algorithm given in \cite{Zou2013}. We refer to this algorithm as the Extended Goldberg's Algorithm. Extended Goldberg's Algorithm \cite{Zou2013} addresses a constrained variant of $\ensuremath{\mathsf{Densest \mhyphen Subgraph}}$, where some vertices are forced to be in a densest subgraph, that is we want to compute a densest subgraph $G[V']$ constrained to the fact that a set $S \subseteq V'$. We denote by $Dense\mhyphen Subgraph(G[V'],C(S))$ a densest subgraph of $G[V']$ that is forced to contain $S$, where $S$ is called the \emph{constrained set} of $Dense\mhyphen Subgraph(G[V'],C(S))$. Notice that \textit{Dense-Subgraph}$(G[V'],C(S))$ can be computed with the Extended Goldberg's Algorithm in time $O(|V||E|\log(|V|^2/|E|))$ \cite{DBLP:journals/siamcomp/GalloGT89,Zou2013}. \begin{comment} Let $G_c$ be the graph obtained by this modification. We show next that a minimum cut of this graph corresponds to a densest subgraph of $G$. \begin{lemma} \label{lem:modifiedGoldberg} There exists a densest subgraph $G[S]$ of $G=(V,E)$, with $S \subseteq V$, if and only if there exists a minimum cut in $G_c$ of value \[ \sum_{v \in V\setminus S} d(v) + |e(S,V\setminus S)|+ c|S| \leq 2|E| \] \end{lemma} \begin{proof} The result follows from the analysis of Goldberg's Algorithm that show that there exists a densest subgraph $G[S]$ of $G=(V,E)$, with $S \subseteq V$, if and only if there exists a minimum cut in $G_c$ of value \[ \sum_{v \in V\setminus S} d(v) + |e(S,V\setminus S)|+ c|S| \leq 2|E| \] and from the observation that It follows that edges $\{ s,v \}$ and $\{ s,u \}$ are not cut, hence $u,v \in S$. \end{proof} \end{comment} \section{Approximating {\sf Top-k-Overlapping Densest Subgraphs}{}} \label{sec:Approximation} In this section, we present a $\frac{2}{3}$-approximation algorithm for {\sf Top-k-Overlapping Densest Subgraphs}{} when $k$ is a constant and a $\frac{1}{2}$-approximation algorithm when $k$ is not a constant. First, the two approximation algorithms compute a densest subgraph of $G$, denoted by $G[W_1]$. Then, they iteratively compute a solution for an intermediate problem, called {\sf Densest-Distinct-Subgraph}. When $k$ is constant we are able to solve the $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ problem in polynomial time, while for general $k$ we are able to provide a $\frac{1}{2}$-approximation algorithm for it. First, we introduce the $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ problem, then we present the two approximation algorithms and the analysis of their approximation factors. \begin{problem}\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}} \\ \noindent \textbf{Input:} a graph $G=(V,E)$ and a set $\mathcal{W} = \{ G[W_1], \dots , G[W_t] \}$, with $1 \leq t \leq k-1$, of subgraphs of $G$.\\ \textbf{Output:} a subgraph $G[Z]$ of $G$ such that $Z \neq W_i$, for each $1 \leq i \leq t$, and $dens(G[Z])$ is maximum. \end{problem} Notice that $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ is not identical to compute a densest subgraph of $G$, as we need to ensure that the returned subgraph $G[Z]$ is distinct from any subgraph in $\mathcal{W}$. Moreover, notice that we assume that $|\mathcal{W}| \leq k-1$, since if $|\mathcal{W}| = k$ we already have $k$ subgraphs in our solution of {\sf Top-k-Overlapping Densest Subgraphs}{}. \subsection{Approximation for Constant $k$} \label{subsec:ApproxConstant} First, we show that $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ is polynomial-time solvable when $k$ is a constant, then we show how to obtain an approximation algorithm for {\sf Top-k-Overlapping Densest Subgraphs}{} by iteratively solving $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ and by combining this solution with one that consists of $k$ singletons. \subsubsection{A Polynomial-Time Algorithm for $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$\\} We start by proving a property of solutions of $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$. \begin{lemma} \label{lem:DenseK} Consider a graph $G=(V,E)$ and a set $\mathcal{W} = \{ G[W_1], \dots , G[W_t]\}$, $1 \leq t \leq k-1$, of subgraphs of $G$. Given a subgraph $G[Z]$ distinct from the subgraphs in $\mathcal{W}$, there exist at most $t$ vertices $u_1, \dots , u_{t}$, with $u_i \in W_i$, $1 \leq i \leq t$, that can be partitioned into two sets $U_1$, $U_2$ such that $Z \supseteq U_1$, $Z \cap U_2 = \emptyset$ and there is no $G[W_j]$ in $\mathcal{W}$, with $1 \leq j \leq t$, such that $W_j \supseteq U_1$ and $W_j \cap U_2 = \emptyset$. \end{lemma} \begin{proof} Consider $G[Z]$ and a subgraph $G[W_j]$, $1 \leq j \leq t$. Since $G[Z]$ is distinct from $G[W_j]$, it follows that there exists a vertex $u_j \in V$ such that $u_j \in Z \setminus W_j$ (in this case $u_j \in U_1$) or $u_j \in W_j \setminus Z$ (in this case $u_j \in U_2$), otherwise $Z$ is identical to $W_j$. By construction, the two sets $U_1$ and $U_2$ satisfy the lemma. \end{proof} Next, based on Lemma \ref{lem:DenseK}, we show how to compute an optimal solution of \textsf{Dense-Distinct-Subgraph}, when $k$ is a constant. Algorithm \ref{algo:approx3} iterates over each subset $\{ u_1,\dots, u_t \}$ of at most $t$ vertices (recall that $|\mathcal{W}|=t$) and over the subsets $U_1, U_2 \subseteq \{ u_1,\dots, u_t \}$, such that $U_1 \uplus U_2 = \{ u_1,\dots, u_t \}$. Algorithm \ref{algo:approx3} computes a densest subgraph $G[Z]$ of $G$, with constrained set $U_1$ and with $Z \cap U_2 = \emptyset$, such that there is no subgraph of $\mathcal{W}$ that contains $U_1$ and whose set of vertices is disjoint from $U_2$. Algorithm \ref{algo:approx3} applies the Extended Goldberg's algorithm on the subgraph $G[ V \setminus U_2]$, with constrained set $C(U_1)$. We prove the correctness of Algorithm \ref{algo:approx3} in the next theorem. \begin{theorem} \label{teo:OptDenseK} Let $G[Z]$ be the solution returned by Algorithm \ref{algo:approx3}. Then, an optimal solution $G[Z']$ of $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ over instance $(G, \mathcal{W})$ has density at most $dens(G[Z])$. \end{theorem} We recall that a densest subgraph constrained to a given set can be computed in time $O(|V||E| \log(\frac{|V|^2}{|E|}))$ with the Extended Goldberg's Algorithm \cite{DBLP:journals/siamcomp/GalloGT89,Zou2013}. It follows that Algorithm~\ref{algo:approx3} returns an optimal solution of $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ in time $O(|V|^{2k+2}|E| \log(\frac{|V|^2}{|E|}))$. \subsubsection{A $\frac{2}{3}$-Approximation Algorithm when $k$ is a Constant\\} We show that, by solving the $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ problem optimally, we achieve a $\frac{2}{3}$ approximation ratio for {\sf Top-k-Overlapping Densest Subgraphs}{}. The approximation algorithm returns the solution of maximum value between the solution returned by Algorithm \ref{algo:approxDef} and a solution consisting of $k$ singletons. First, we consider the solution returned by Algorithm \ref{algo:approxDef}. At each step, Algorithm \ref{algo:approxDef} computes an optimal solution of $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ in time $O(|V|^{2k+2}|E| \log(\frac{|V|^2}{|E|}))$ and the output subgraph is added to the solution. Since $k$ is a constant, hence the number of iterations of Algorithm \ref{algo:approxDef} is a constant, the overall time complexity of Algorithm \ref{algo:approxDef} is $O(|V|^{2k+2}|E| \log(\frac{|V|^2}{|E|}))$. \begin{algorithm}[H] \KwData{a graph $G$ } \KwResult{ a set $\mathcal{W} = \{ G[W_1], \dots , G[W_k] \}$ of subgraphs of $G$} $\mathcal{W} \leftarrow \{ G[W_1] \}$ /* $G[W_1]$ is a densest subgraph of $G$ */\; \For{$i \leftarrow 2$ \KwTo $k$}{ Compute an optimal solution $G[Z]$ of $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ with input $(G,\mathcal{W})$ /* Applying Algorithm \ref{algo:approx3} */\; $\mathcal{W} \leftarrow \mathcal{W} \cup \{ G[Z] \}$ } Return($\mathcal{W}$)\; \caption{Algorithm that returns an approximated solution of {\sf Top-k-Overlapping Densest Subgraphs}{}} \label{algo:approxDef} \end{algorithm} Consider the solution $\mathcal{W}= \{ G[W_1], \dots, G[W_k] \}$ returned by Algorithm \ref{algo:approxDef}, we prove a result on the distance between two subgraphs of $\mathcal{W}$. \begin{lemma} \label{lem:ApproxDistance} Let $\mathcal{W}= \{ G[W_1], \dots, G[W_k] \}$ be a set of subgraphs returned by Algorithm~\ref{algo:approxDef}. Then, for each $G[W_i],G[W_j] \in \mathcal{W}$, with $1 \leq i \leq k$, $1 \leq j \leq k$ and $i \neq j$, it holds $d(G[W_i],G[W_j]) > 1$. \end{lemma} \begin{proof} By the definition of distance $d$, since $G[W_i]$ and $G[W_j]$, with $1 \leq i \leq k$, $1 \leq j \leq k$ and $i \neq j$, are distinct subgraphs of $G$, it follows that $ d(G[W_i],G[W_j]) = 2- \frac{|W_i \cap W_j|^2}{|W_i||W_j|} $. Since $\frac{|W_i \cap W_j|^2}{|W_i||W_j|} \leq 1$, it follows that $d(G[W_i],G[W_j]) \geq 1$. \end{proof} Now, we prove a bound on the value $r(\mathcal{W})$ of a solution $\mathcal{W}$ returned by Algorithm \ref{algo:approxDef}. \begin{lemma} \label{lem:ApproxRkconstant} Let $\mathcal{W}= \{ G[W_1], \dots, G[W_k] \}$ be a set of subgraphs returned by Algorithm~\ref{algo:approxDef} and let $\mathcal{W}^o= \{ G[W^o_1], \dots, G[W^o_k] \}$ be an optimal solution of {\sf Top-k-Overlapping Densest Subgraphs}{} over instance $G$. Then, for each $\lambda > 0$, \[ dens(\mathcal{W}) \geq dens(\mathcal{W}^o) \text{ and } \lambda \sum_{i=1}^{k-1} \sum_{j=i+1}^k d(W_i,W_j) \geq \frac{1}{2} \lambda \sum_{i=1}^{k-1} \sum_{j=i+1}^k d(W_i^o,W_j^o). \] \end{lemma} \begin{proof} The second inequality follows from Lemma \ref{lem:ApproxDistance} and from the fact that $ d(G[U],G[Z]) \leq 2$, for each $U,Z \subseteq V$. We prove the first inequality of the lemma by induction on $k$. Let $G[W_i]$, with $2 \leq i \leq k$, be the subgraph added to $\mathcal{W}$ by the $i$-th iteration of Algorithm \ref{algo:approxDef}. By construction, $dens(G[W_1])$ $\geq$ $dens(G[W_2]) \geq $ $\dots$ $\geq dens(G[W_k])$. Moreover, assume w.l.o.g. that $dens(G[W^o_1]) \geq$ $dens(G[W^o_2]) \geq $ $\dots$ $ \geq dens(G[W^o_k])$. When $k=1$, by construction of Algorithm \ref{algo:approxDef}, $G[W_1]$ is a densest subgraph of $G$, it follows that $dens(G[W_1]) \geq dens(G[W^o_1])$. Assume that the lemma holds for $ k -1$, we prove that it holds for $k$. First, notice that $ \sum_{i=1}^k dens(G[W_i]) = \sum_{i=1}^{k-1} dens(G[W_i]) + dens(G[W_k]) $. By induction hypothesis \[ \sum_{i=1}^{k-1} dens(G[W_i]) \geq \sum_{i=1}^{k-1} dens(G[W^o_i]). \] Notice that $G[W_k]$ is an optimal solution of $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ on instance $(G, $ $\{ G[W_1]$, $G[W_2],$ $\dots, G[W_{k-1}] \})$. By the pigeon principle at least one of $G[W^o_1]$, $G[W^o_2]$, $\dots$, $G[W^o_k]$ does not belong to the set $\{ G[W_1], G[W_2],$ $\dots,$ $G[W_{k-1}]\}$ of subgraphs, hence, by the optimality of $G[W_k]$, $dens(G[W_k]) \geq dens(G[W^o_p])$, for some $p$ with $1 \leq p \leq k$, and $dens(G[W^o_p]) \geq dens(G[W^o_k])$. Now, \[ \sum_{i=1}^k dens(G[W_i]) = \sum_{i=1}^{k-1} dens(G[W_i]) + dens(G[W_k]) \geq \] \[ \sum_{i=1}^{k-1} dens(G[W^o_i]) + dens(G[W^o_k]) \geq \sum_{i=1}^{k} dens(G[W^o_i] \] thus concluding the proof. \end{proof} Consider Algorithm $A_T$ that, given an instance $G$ of {\sf Top-k-Overlapping Densest Subgraphs}{}, returns a solution $\mathcal{W}'=\{ G[W'_1], \dots, G[W'_k] \}$ consisting of $k$ distinct singletons. Notice that, since each $G[W'_i]$, with $1 \leq i \leq k$, is a singleton, it follows that $dens(\mathcal{W}')=0$. Moreover, since the subgraphs in $\mathcal{W}'$ are pairwise disjoint, we have $d(G[W'_i],G[W'_j])=2$, for each $G[W'_i]$, $G[W'_j] \in \mathcal{W}'$ with $1 \leq i,j \leq k$ and $i \neq j$. We can prove now that the maximum between $r(\mathcal{W})$ (where $\mathcal{W}$ is the solution returned by Algorithm \ref{algo:approxDef}) and $r(\mathcal{W}')$ (where $\mathcal{W}'$ is the solution returned by Algorithm $A_T$) is at least $\frac{2}{3}$ of the value of an optimal solution of {\sf Top-k-Overlapping Densest Subgraphs}{}. \begin{theorem} \label{teo:finalApprox} Let $\mathcal{W}= \{ G[W_1], \dots, G[W_k] \}$ be the solution returned by Algorithm \ref{algo:approxDef} and let $\mathcal{W}'= \{ G[W'_1], \dots, G[W'_k] \}$ be the solution returned by Algorithm $A_T$. Let $\mathcal{W}^o= \{ G[W^o_1], \dots,$ $G[W^o_k] \}$ be an optimal solution of {\sf Top-k-Overlapping Densest Subgraphs}{} over instance $G$. Then $ \max(r(\mathcal{W}), r(\mathcal{W}')) \geq \frac{2}{3}~r(\mathcal{W}^o). $ \end{theorem} \subsection{Approximation When $k$ is not a Constant} \label{subsec:approxNotConstant} Now, we show that {\sf Top-k-Overlapping Densest Subgraphs}{} can be approximated within factor $\frac{1}{2}$ when $k$ is not a constant. The approximation algorithm (Algorithm \ref{algo:approxkNotConstant}), consists of two phases. In the first phase, while $\mathcal{W}$ does not contain crossing subgraphs (see Property~\ref{prop:Stop}), Algorithm~\ref{algo:approxkNotConstant} adds to $\mathcal{W}$ a subgraph which is an optimal solution of $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$. When Property \ref{prop:Stop} holds, Phase 2 of Algorithm \ref{algo:approxkNotConstant} completes $\mathcal{W}$, so that $\mathcal{W}$ contains $k$ subgraphs. \begin{algorithm} \KwData{a graph $G$ } \KwResult{$\mathcal{W} = \{ G[W_1], \dots , G[W_k] \}$ of subgraphs of $G$} $\mathcal{W} \leftarrow \{ G[W_1] \}$ /* $G[W_1]$ is a densest subgraph of $G$ */\; \textbf{Phase 1}\; \While{$|\mathcal{W}| < k$ and Property \ref{prop:Stop} does not hold} { Compute an optimal solution $G[Z]$ of $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ with input $(G,\mathcal{W})$ /* Applying Algorithm \ref{algo:approx4} (described later) */\; $\mathcal{W} \leftarrow \mathcal{W} \cup \{ G[Z] \}$\; } \textbf{Phase 2} (Only if $|\mathcal{W}|<k$)\; $W_{i,j} \leftarrow W_i \cap W_j$, with $W_i$ and $W_j$ two crossing subgraphs in $\mathcal{W}$\; \If{$|W_{i,j}| \leq 3$} { Complete $\mathcal{W}$ by adding the densest distinct subgraphs (not already in $\mathcal{W}$) induced by $W_i \cup \{ v\}$, with $v \in V \setminus W_i$, and by $W_j \cup \{ u \}$, with $u \in V \setminus W_j$; } \If{$|W_{i,j}| \geq 4$} { Complete $\mathcal{W}$ by adding the densest distinct subgraphs (not already in $\mathcal{W}$) induced by $W_i \cup \{ v\}$, with $v \in V \setminus W_i$, by $W_j \cup \{ u \}$, with $u \in V \setminus W_j$, and by $W_j \setminus \{ w\}$, with $w \in W_{i,j}$; } Return($\mathcal{W}$)\; \caption{Returns an approximated solution of {\sf Top-k-Overlapping Densest Subgraphs}{}} \label{algo:approxkNotConstant} \end{algorithm} First, we define formally the property on which Algorithm \ref{algo:approxkNotConstant} is based. \begin{property} \label{prop:Stop} Given a collection $\mathcal{W}$ of $t$ subgraphs, with $2 \leq t \leq k-1$, there exist two crossing subgraphs $G[W_i]$ and $G[W_j]$ in $\mathcal{W}$, with $1 \leq i \leq t$, $1 \leq j \leq t$ and $i \neq j$. \end{property} \subsubsection{Analysis of Phase 1} We show that, while $\mathcal{W}$ does not satisfy Property \ref{prop:Stop}, $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ can be solved optimally in polynomial time. First, we prove a property of a solution of the $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ problem when Property \ref{prop:Stop} does not hold. \begin{lemma} \label{lem:PrelAprrox} Consider a graph $G=(V,E)$ and a set $\mathcal{W} = \{ G[W_1], \dots , G[W_t]\}$, $1 \leq t \leq k-1$, of subgraphs of $G$ that does not satisfy Property \ref{prop:Stop}. Given a subgraph $G[Z]$ distinct from the subgraphs in $\mathcal{W}$, there exist at most three vertices $u_1, u_2, u_3 \in V$ such that $ u_1, u_2 \in Z $, $u_3 \notin Z$ and there is no $G[W_j]$ in $\mathcal{W}$, $1 \leq j \leq t$, with $u_1, u_2 \in W_j$ and $u_3 \notin W_j$. \end{lemma} \begin{proof} Consider a subgraph $G[Z]$ distinct from the subgraphs in $\mathcal{W}$ and a vertex $u_1 \in Z$. Notice that, for each subgraph in $\mathcal{W}$ that does not contain $u_1$, the lemma holds. Now, we consider the set $\mathcal{W}'$ of subgraphs in $\mathcal{W}$ that contain $u_1$. If $\mathcal{W}' = \emptyset$, then the lemma holds, since there is no subgraph in $\mathcal{W}$ that contains $u_1$. Consider the pair $(\mathcal{W}', \subseteq)$ where $\subseteq$ is the subgraph inclusion relation \footnote{Given $A,B \subseteq V$, $G[A] \subseteq G[B]$ if and only if $A \subseteq B$}. $(\mathcal{W}', \subseteq)$ is a well-ordered set \footnote{We recall that a well-ordered set is a pair ($S, \leq$), where $S$ is a set and $\leq$ is a binary relation on $S$ such that (1) Relation $\leq$ satisfies the following properties: reflexivety, antisymmetry, transitivity and comparability; (2) every non-empty subset of $S$ has a least element based on relation $\leq$.}. Clearly, $\subseteq$ is reflexive, antysimmetric and transitive on $\mathcal{W}'$. We show that is comparable, that is, given $G[W_x], G[W_y] \in \mathcal{W'}$ with $W_x \neq W_y$, either $W_x \subset W_y$ or $W_y \subset W_x$. Indeed, consider two subgraphs $G[W_x], G[W_y] \in \mathcal{W}'$, such that neither $W_x \subset W_y$ nor $W_y \subset W_x$. It follows that they are crossing subgraphs, since they both contain $u_1$, contradicting the hypothesis that Property \ref{prop:Stop} does not hold. Since $\mathcal{W}'$ is a finite set, it follows that $(\mathcal{W}', \subseteq)$ is a well-ordered set. Consider now the set $\mathcal{W}'_C$ of subgraphs in $\mathcal{W}'$ that are subgraphs of $G[Z]$ and notice that $(\mathcal{W}', \subseteq)$ and $(\mathcal{W}'_C, \subseteq)$ are well-ordered sets. Let $G[W_v]$ be the largest subgraph in $\mathcal{W}'_C$. Since $G[W_v]$ is a subgraph of $G[Z]$, there exists a vertex $u_2 \in Z \setminus W_v$. Since $(\mathcal{W}'_C, \subseteq)$ is a well-ordered set, each subgraph in $\mathcal{W}'_C \setminus \{ G[W_v] \}$ is a subgraph of $G[W_v]$, thus each subgraph in $\mathcal{W}'_C$ does not contain $u_2$. Consider now the set $\mathcal{W}'_N$ of subgraphs in $\mathcal{W}'$ which are not subgraphs of $G[Z]$. Notice that $(\mathcal{W}'_N, \subseteq)$ is a well-ordered set and let $G[W_y]$ be the graph of minimum cardinality in $\mathcal{W}'_N$. It follows that there exists a vertex $u_3 \in W_y \setminus Z$, and notice that, since $(\mathcal{W}'_N, \subseteq)$ is a well-ordered set, $u_3$ belongs to each subgraph in $\mathcal{W}'_N$. Since we have shown that there exists a vertex $u_2 \in Z$ that does not belong to any subgraph of $\mathcal{W}'_C$ and there exists a vertex $u_3 \notin Z$ that belongs to each subgraph of $\mathcal{W}'_N$, the lemma follows. \end{proof} Algorithm \ref{algo:approx4} computes an optimal solution $G[Z]$ of $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ when Property 1 does not hold. Algorithm \ref{algo:approx4} is a modified variant of Algorithm \ref{algo:approx3} (see Section \ref{subsec:ApproxConstant}), which considers a set of at most three vertices $u_1,u_2, u_3$, with $U_1 = \{ u_1,u_2\}$ and $U_2=\{ u_3\}$. Based on Lemma \ref{lem:PrelAprrox}, we can prove the following result. \begin{theorem} \label{teo:OptPhase1} Let $G[Z]$ be the solution returned by Algorithm \ref{algo:approx4}. Then, an optimal solution of $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ over instance $(G,\mathcal{W})$ when Property \ref{prop:Stop} does not hold has density at most $dens(G[Z])$. \end{theorem} Notice that Algorithm \ref{algo:approx4} returns an optimal solution of $\ensuremath{\mathsf{Densest\mhyphen Distinct\mhyphen Subgraph}}$ in Case 3 of Lemma \ref{lem:PrelAprrox} in time $O(|V|^{4}|E| \log(\frac{|V|^2}{|E|}))$, since it applies the Extended Goldberg's Algorithm of complexity $O(|V||E| \log(\frac{|V|^2}{|E|}))$ \cite{Zou2013}, for each set of three vertices in $V$. \paragraph{\textbf{Analysis of Phase 2}\\ } Assuming that Property \ref{prop:Stop} holds and $|\mathcal{W}|=t < k$, we consider Phase 2 of Algorithm \ref{algo:approxkNotConstant}. Given two crossing subgraphs $G[W_i]$ and $G[W_j]$ of $\mathcal{W}$, with $1 \leq i \leq t$, $1 \leq j \leq t$ and $i \neq j$, define $W_{i,j} = W_i \cap W_j$, Algorithm \ref{algo:approxkNotConstant} adds $h = k-t$ subgraphs to $\mathcal{W}$ until $|\mathcal{W}|=k$, as follows. If $|W_{i,j}| \leq 3$, then Phase 2 of Algorithm \ref{algo:approxkNotConstant} adds the $h$ densest distinct subgraphs (not already in $\mathcal{W}$) induced by $W_i \cup \{ v\}$, for some $v \in V \setminus W_i$, and by $W_j \cup \{ u \}$, for some $u \in V \setminus W_j$. If $|W_{i,j}| \geq 4$, then Phase 2 of Algorithm \ref{algo:approxkNotConstant} adds the $h$ densest distinct subgraphs (not already in $\mathcal{W}$) induced by $W_i \cup \{ v\}$, for some $v \in V \setminus W_i$, by $W_j \cup \{ u \}$, for some $u \in V \setminus W_j$, and by $W_j \setminus \{ w\}$, for some $w \in W_{i,j}$. Next, we show that, after Phase 2 of Algorithm \ref{algo:approxkNotConstant}, $|\mathcal{W}|=k$ and each subgraph added by Phase 2 has density at least $\frac{1}{2} dens(G[W_j])$, where $G[W_j]$ is a subgraph added to $\mathcal{W}$ in Phase~1. \begin{lemma} \label{lem:secondPartCorrect} $|\mathcal{W}|=k$ after the execution of Phase 2 of Algorithm \ref{algo:approxkNotConstant}. \end{lemma} \begin{lemma} \label{lem:secondPart} Let $G[W']$ be a subgraph added to $\mathcal{W}$ by Phase 2 of Algorithm \ref{algo:approxkNotConstant}. Then, $dens(G[W']) \geq \frac{1}{2} dens(G[W_j])$, with $G[W_j]$ a subgraph added to $\mathcal{W}$ by Phase 1 of Algorithm~\ref{algo:approxkNotConstant}. \end{lemma} Phase 2 of Algorithm \ref{algo:approxkNotConstant} requires time $O(k^2 |V|)$, since we have to compare each subgraph to be added to $\mathcal{W}$ with the subgraphs already in $\mathcal{W}$ and each of this comparison requires time $O(k|V|)$. Each iteration of Phase 1 of Algorithm~\ref{algo:approxkNotConstant} requires time $O(|V|^{4}|E| \log(\frac{|V|^2}{|E|}))$, hence the overall complexity of Algorithm~\ref{algo:approxkNotConstant} is $O(|V|^{5}|E| \log(\frac{|V|^2}{|E|}))$, since Phase 1 is iterated at most $k \leq |V|-1$ times. Now, thanks to Lemma \ref{lem:secondPart}, we are able to prove that the density of the solution returned by Algorithm \ref{algo:approxkNotConstant} is at least half the density of an optimal solution of {\sf Top-k-Overlapping Densest Subgraphs}{}. \begin{lemma} \label{lem:GeneralApproxLemma} Let $\mathcal{W}= \{ G[W_1], \dots, G[W_k] \}$ be the solution returned by Algorithm \ref{algo:approxkNotConstant} and let $\mathcal{W}^o= \{ G[W^o_1], \dots, G[W^o_k] \}$ be an optimal solution of {\sf Top-k-Overlapping Densest Subgraphs}{} over instance $G$. Then $\sum_{i=1}^{k} dens(G[W_i]) \geq \frac{1}{2} \sum_{i=1}^{k} dens(G[W^o_i]). $ \end{lemma} We can conclude the analysis of the approximation factor with the following result. \begin{theorem} \label{teo:finalApproxGeneralRes} Let $\mathcal{W}= \{ G[W_1], \dots, G[W_k] \}$ be the solution returned by Algorithm \ref{algo:approxkNotConstant} and let $\mathcal{W}^o= \{ G[W^o_1], \dots, G[W^o_k] \}$ be an optimal solution of {\sf Top-k-Overlapping Densest Subgraphs}{} over instance $G$. Then $ r(\mathcal{W}) \geq \frac{1}{2} r(\mathcal{W}^o) $. \end{theorem} \begin{proof} First, by Lemma \ref{lem:GeneralApproxLemma}, it holds $dens(\mathcal{W}) \geq \frac{1}{2} dens(\mathcal{W}^o)$. Similarly to the proof Lemma \ref{lem:ApproxDistance}, since the subgraphs in $\{ G[W_1], \dots, G[W_k] \}$ are all distinct, it holds $d(G[W_i],G[W_j]) \geq 1$, for each $i$, $j$ with $1 \leq i \leq k$, $1 \leq j \leq k$ and $i \neq j$, and by definition $d(G[W^o_i],G[W^o_j]) \leq 2$, thus $ \lambda \sum_{i=1}^{k-1} \sum_{j=i+1}^k d(G[W_i],G[W_j]) \geq \frac{1}{2} \lambda \sum_{i=1}^{k-1} \sum_{j=i+1}^k d(G[W_i^o],G[W_j^o]) $. We can conclude that $ r(\mathcal{W}) \geq \frac{1}{2} r(\mathcal{W}^o).$ \end{proof} \section{Complexity of {\sf Top-k-Overlapping Densest Subgraphs}{}} \label{sec:complexity} In this section, we consider the computational complexity of {\sf Top-k-Overlapping Densest Subgraphs}{} and we show that the problem is NP-hard even if $k=3$. We denote this restriction of the problem by \ensuremath{\mathsf{Top \mhyphen 3 \mhyphen Overlapping\mhyphen %Subgraphs}}{}. We prove the result by giving a reduction from {\sf 3-Clique Partition}, which is NP-complete \cite{DBLP:conf/coco/Karp72}. We recall that {\sf 3-Clique Partition}, given an input graph $G_P=(V_P,E_P)$, asks for a partition of $V_P$ into $V_{P,1}$, $V_{P,2}$, $V_{P,3}$ such that $V_P = V_{P,1} \uplus V_{P,2} \uplus V_{P,3}$ and each $G[V_{P,i}]$, with $1 \leq i \leq 3$, is a clique. Given an instance $G_P=(V_P,E_P)$ of {\sf 3-Clique Partition}, the input graph $G=(V,E)$ of \ensuremath{\mathsf{Top \mhyphen 3 \mhyphen Overlapping\mhyphen %Subgraphs}}{} is identical to $G_P=(V_P,E_P)$. Define $\lambda = 3|V|^3$. In order to define a reduction from {\sf 3-Clique Partition} to \ensuremath{\mathsf{Top \mhyphen 3 \mhyphen Overlapping\mhyphen %Subgraphs}}{}, we show the following results. \begin{lemma} \label{lem:3Case1} Let $G_P=(V_P,E_P)$ be a graph instance of {\sf 3-Clique Partition} and let $G=(V,E)$ be the corresponding graph instance of \ensuremath{\mathsf{Top \mhyphen 3 \mhyphen Overlapping\mhyphen %Subgraphs}}{}. Given three cliques $G_P[V_{P,1}]$, $G_P[V_{P,2}]$, $G_P[V_{P,3}]$ such that $V_{P,1}$, $V_{P,2}$, $V_{P,3}$ partition $V_P$, we can compute in polynomial time a set $\mathcal{W}= \{G[V_1], G[V_2], G[V_3] \}$ such that $r(\mathcal{W}) \geq \frac{|V| -3}{2} + 18|V|^3$. \end{lemma} \begin{lemma} \label{lem:3Case2} Let $G_P=(V_P,E_P)$ be a graph instance of {\sf 3-Clique Partition} and let $G=(V,E)$ be the corresponding graph instance of \ensuremath{\mathsf{Top \mhyphen 3 \mhyphen Overlapping\mhyphen %Subgraphs}}{}. Given a solution $\mathcal{W}= \{G[V_1], G[V_2], G[V_3] \}$ of \ensuremath{\mathsf{Top \mhyphen 3 \mhyphen Overlapping\mhyphen %Subgraphs}}{} on instance $G$, with $r(\mathcal{W}) \geq \frac{|V| -3}{2}+ 18|V|^3$, we can compute in polynomial time three cliques $G_P[V_{P,1}]$, $G_P[V_{P,2}]$, $G_P[V_{P,3}]$ of $G_P$ such that $V_{P,1}$, $V_{P,2}$, $V_{P,3}$ partition $V_P$. \end{lemma} We can conclude that \ensuremath{\mathsf{Top \mhyphen 3 \mhyphen Overlapping\mhyphen %Subgraphs}}{} is NP-hard. \begin{theorem} \ensuremath{\mathsf{Top \mhyphen 3 \mhyphen Overlapping\mhyphen %Subgraphs}}{} is NP-hard. \end{theorem} \begin{proof} From Lemma \ref{lem:3Case1} and Lemma \ref{lem:3Case2}, it follows that we have described a polynomial-time reduction from {\sf 3-Clique Partition} to \ensuremath{\mathsf{Top \mhyphen 3 \mhyphen Overlapping\mhyphen %Subgraphs}}{}. Since {\sf 3-Clique Partition} is NP-complete \cite{DBLP:conf/coco/Karp72}, it follows that also \ensuremath{\mathsf{Top \mhyphen 3 \mhyphen Overlapping\mhyphen %Subgraphs}}{} is NP-hard. \end{proof} \section{Conclusion} \label{sec:conclusion} We have shown that {\sf Top-k-Overlapping Densest Subgraphs}{} is NP-hard when $k=3$ and we have given two approximation algorithms of factor $\frac{2}{3}$ and $\frac{1}{2}$, when $k$ is a constant and when $k$ is smaller than the number of vertices in the graph, respectively. For future works, it would be interesting to further investigate the approximability of {\sf Top-k-Overlapping Densest Subgraphs}{}, possibly improving the approximation factor or improving the time complexity of our approximation algorithms. A second interesting open problem is the computational complexity of {\sf Top-k-Overlapping Densest Subgraphs}{}, in particular when $\lambda$ is a constant. Another open problem of theoretical interest is the computational complexity of {\sf Top-k-Overlapping Densest Subgraphs}{} when $k=2$. \bibliographystyle{splncs03}

a9dd46befbe377c2fc4cc55a66bf0d1d30243d3b

\section{Introduction} \section{Introduction} \section{Introduction} Recent works in the state of the art in RDF data management have shown that extraction and exploitation of the implicit schema of the data can be beneficial in both storage and SPARQL query performance \cite{pham2015deriving}\cite{pham2016exploiting}\cite{meimaris2017extended}\cite{montoya2017odyssey}. In order to organize on disk, index and query triples efficiently, these trends heavily rely on two structural components of an RDF dataset, namely (i) the notion of \textit{characteristic sets }(CS), i.e., different property sets that characterize subject nodes, and (ii) the join links between CSs. For the latter, in our previous work, we introduced \textit{Extended Characteristic Sets} (ECS)\cite{meimaris2017extended}, which are typed links between CSs that exist only when there are object-subject joins between their triples, and we showed how RDF data management can rely extensively on CSs and ECSs for both storage and indexing, yielding significant performance benefits in heavy SPARQL workloads. However, this approach failed to address schema heterogeneity in loosely-structured datasets, as this implied a large number of CSs and ECSs (e.g., Geonames contains 851 CSs and 12136 CS links), and thus, skewed data distributions that impose large overheads in the extraction, storage and disk-based retrieval\cite{pham2015deriving}\cite{meimaris2017extended}. In this paper, we exploit the hierarchical relationships between CSs, as captured by subsumption of their respective property sets, in order to merge related CSs. We follow a relational implementation approach by storing all triples corresponding to a set of merged CSs into a separate relational table and by executing queries through a SPARQL to SQL transformation. Although, alternative storage technologies can be considered (key-value, graph stores,etc), we have selected well-established technologies and database systems for the implementation of our approach, in order to take advantage of existing data indexing and query processing techniques that have been proven to scale efficiently in large and complex datasets. To this end, we present a novel system, named \textit{raxonDB}, that exploits these hierarchies in order to merge together hierarchically related CSs and decrease the number of CSs and the links between them, resulting in a more compact schema with better data distribution. The resulting system, built on top of PostgreSQL, provides significant performance improvements in both storage and query performance of RDF data. In short, our contributions are as follows: \begin{itemize} \item We introduce a novel CS merging algorithm that takes advantage of CS hierarchies, \item we implement \textit{raxonDB}, an RDF engine built on top of a relational backbone that takes advantage of this merging for both storing and query processing, \item we perform an experimental evaluation that indicates significant performance improvements for various parameter configurations. \end{itemize} \section{Distance-Based Reordering} \subsection{Query Representation} Let $T$ be the set of all triple patterns in an incoming query $q$. Each triple pattern $t_i \in T$ consists of three nodes, either bound (URIs, blank node IDs or literals) or unbound (variable), i.e., the subject, predicate and object of $t_i$. Furthermore, let $M$ represent the set of all unique nodes in the subject and object positions in $T$, both bound and unbound. We map each triple pattern $t_i$ from $T$ to a $\left\vert{M}\right\vert$-dimensional vector space, where each vector attribute is an element from $M$, i.e., a subject or object node from the query. In this preliminary approach, we assume a solution space of left-deep trees, where the original problem of finding the best order of triple evaluation is factorial. This is because for a query $q$ consisting of $\left\vert{T}\right\vert$ triple patterns, there are $\left\vert{T}\right\vert!$ different permutations of triples. The factorial problem, commonly solved by dynamic programming algorithms, guarantees that the solution will be optimal with respect to the chosen cost model. Unfortunately, recent studies have suggested that even industrial-level query optimizers do not yield satisfactory results when they are heavily dependent on the cost model \cite{leis2015good}, especially when cost model errors propagate through multiple joins. In our approach, we map triple patterns from a query into a multidimensional space, and decide the join order based on the spatial correlations of the patterns, i.e., by comparing the pair-wise distances of the vectors that represent the triple patterns in the multiple dimensions, and ranking the patterns based on this comparison. The multidimensional space is built based on cardinality statistics that are held in a separate index. These pre-computed statistics are commonly found in RDF stores, and include the cardinality of triples with a bound subject, property, and object, as well as the size of the whole dataset. Specifically, we assume that each triple pattern becomes a $m$-dimensional vector, and the value of an attribute $m_i$ for a given triple pattern $t = (s,p,o)$, is given by the following function: \[ f(t,m_i) = \begin{cases} card(t),& \text{if } m_i \in (s,o)\\ 0, & \text{otherwise} \end{cases} \] Where $card(t)$ is the cardinality of the triple pattern. This is calculated using the following rules: \begin{itemize} \item{ $card(t) = card(p)$, if $p$ is bound and $s$,$o$ are unbound,} \item{ $card(t) = max(1,\frac{card(p)}{\left\vert{S}\right\vert})$, if $p$ and $o$ are bound, and $s$ is unbound,} \item{ $card(t) = 1$, if $p$ and $s$ are bound, and $o$ is either bound or unbound,} \item{ $card(t) = \left\vert{D}\right\vert$, for all other cases} \end{itemize} The cardinality for a bound property $card(p)$ is simply given as the number of triples with $p$ as the predicate. $\left\vert{D}\right\vert$ and $\left\vert{S}\right\vert$ are the number of triples in the dataset $D$, and the number of distinct subject nodes $S$ respectively. Notice that we do not represent predicate attributes, which means that the query space will be built entirely based on the subject and object nodes of all triple patterns in the query. Instead, the information that comes from the predicates in the triple patterns becomes quantized in the values of the attributes, as is given by the aforementioned cardinality estimation rules. Especially for the \textit{rdf:type} property, we assume the existence of an aggregate index that holds the exact cardinalities of particular class types. Therefore, in this special case, the second rule is altered to reflect the exact enumeration of subjects with a specific type. For instance, in our running example, the cardinality of $t_1$ is 99k, as there exist 99k subjects with \textit{rdf:type} \textit{ub:Student}. Function $f(t,m)$ can be used to calculate the values of a $\left\vert{T}\right\vert\times\left\vert{M}\right\vert$ matrix, denoted as $Q_m$, where each row represents a triple pattern in $T$, and each column represents an attribute in $M$, or simply a node in the query pattern. By applying a distance function and processing the pair-wise distances between rows in $Q_m$, the search space becomes quadratic with respect to $\left\vert{T}\right\vert$, and the comparisons will be of $O(\left\vert{T}\right\vert^2)$ complexity, rather than the original factorial one. Moreover, as we do not evaluate self-distances and permuations of the same pair (e.g., $[t_1,t_2]$ and $[t_2,t_1]$, the actual number of comparisons will be $\binom{\left\vert{T}\right\vert}{2}$, or $\left\vert{T}\right\vert \times (\left\vert{T}\right\vert -1)/2$. Even though this reduction in cost comes with the loss of guarantee of optimality, our experiments show that this approach tends to work well for all types of queries, even with large numbers of triple patterns. The matrix $Q_m$ for the query of Listing \ref{lst:sparql}, is shown in Table 1 for the synthetic dataset LUBM10 with 1.5m triples. To construct this for dataset $D$, we apply the estimation rules based on the per-computed statistics. For instance, there are 99k triples with \textit{rdf:type} as a property and \textit{ub:Student} as an object, which can be seen in columns 1 and 2 for $t_1$. Similarly, there are 106k triples with \textit{ub:emailAddress} as a predicate, which is encoded in $t_5$. Application of a distance function will generate a $\left\vert{5}\right\vert\times\left\vert{5}\right\vert$ distance matrix of pair-wise distances between the triple patterns $t_1-t_5$. \subsection{Plan Generation} When $Q_m$ is constructed and the distance matrix is calculated, we build sub-plans for the query, based on the ascending ranking of the pair-wise distances of the triple patterns in $T$. The algorithm to build sub-plans can be seen in Algorithm \ref{alg:generateSubPlans}. Specifically, we sort pair-wise distances in ascending order and create an empty queue for sub-plans (Lines 1-3). Then, we iterate through each pair of sorted triple patterns $t_a, t_b$ (Line 4). If there exists a sub-plan $p_i$ that contains only $t_a$ or only $t_b$, we append the non-contained triple pattern to $p_i$ (Lines 5-10). If none of $t_a, t_b$ are contained in a sub-plan, we create a new sub-plan and add $t_a, t_b$, then push the new sub-plan to the existing queue. The order in which we add these two depends on the cost of each triple pattern (Lines 11-15). In case both patterns $t_a$ and $t_b$ exist in sub-plans, we continue to the next pair. After we have created a series of (ordered) sub-plans, we iterate through consecutive sub-plans and attempt to reorder their patterns in order to better capture join relationships that occur between these. More accurately, if there are more than one sub-plans found, then we try to rearrange the triples that share common variables between adjacent sub-plans, so that the joined triple patterns between the two sub-plans are closely located. This can be seen in Algorithm \ref{alg:reorderSubPlans}. Specifically, if there is only one sub-plan, then this is returned as the final plan (Lines 1-3). Else, the algorithm iterates through the $subPlans$ queue with two pointers (Line 4), and prioritizes each pair of sub-plans (Line 5). The $prioritizePair$ function checks for joins between two consecutive sub-plans (Line 12), i.e. it checks whether the last triple (tail) in $p_i$ is joined with the first triple (head) in $p_{i+1}$. If they are joined, it returns the pair as is (Lines 13-15). If not, it finds the first triple in $p_{i+1}$ that can be joined with the tail of $p_i$ (if one exists), pushes it at the tail of $p_i$ and recursively checks the same pair (Lines 17-21). Finally, the $prioritizePair$ function returns an array of two plans, the first of which is inserted in the final plan, while the second is used as the first sub-plan of the next iteration, as it holds the updated ordering. This means that the triple patterns that are ranked lower in $p_{i+1}$ will be bubbled-up and re-ranked in order to reflect the join relationship between $p_i$ and $p_{i+1}$. \begin{algorithm}[ht] \caption{\emph{generateSubPlans}}\label{alg:generateSubPlans} \textbf{Input: $map$:} A matrix of triple patterns from the query as rows, and query nodes as columns\\ \textbf{Output: $subPlans$:} An ordered list of sub-sets of query triple patterns, that constitute subplans. \begin{algorithmic}[1] \State $subPlans \leftarrow new Queue()$ \State $distances \leftarrow distanceMatrix(map)$ \State sort($distances$) \For{\textbf{each} $pair$ $t_i, t_j \in distances \mid t_i \neq t_j$} \If {$\exists p_i \in subPlans \mid t_a \in p_i$ AND $t_b \not\in p_i$} \State append $t_b$ to $p_i$ \State continue \ElsIf {$\exists p_i \in subPlans \mid t_a \not\in p_i$ AND $t_b \in p_i$} \State append $t_a$ to $p_i$ \State continue \ElsIf {$\not\exists p_i \in subPlans \mid t_a \in p_i$ OR $t_b \in p_i$} \State $p \leftarrow new List()$ \State append minCost($t_a, t_b)$ to $p$ \State append maxCost($t_a, t_b)$ to $p$ \State $subPlans$.push($p$) \Else \State continue \\ \EndIf \EndFor \Return $subPlans$ \end{algorithmic} \end{algorithm} \begin{algorithm}[ht] \caption{\emph{reorderSubPlans}}\label{alg:reorderSubPlans} \textbf{Input: $subPlans(p_0,\dots,p_{n-1})$:} A queue of sub-plans \\ \textbf{Output: $finalPlan$:} An ordered list of triple patterns, to be executed by the query engine. \begin{algorithmic}[1] \If {$subPlans$.size $== 1$} \State $finalPlan$.push($p_0$) \State \Return $finalPlan$ \EndIf \For{\textbf{each} $p_i, p_{i+1} \in subPlans$} \State $nextPair \leftarrow $ prioritizePair($p_i, p_{i+1}$) \State $finalPlan$.push($nextPair$[0]) \State $p_{i+1} \leftarrow nextPair$[1] \If {$p_{i+1}$.isEmpty} \State $subPlans.remove(p_{i+1})$ \EndIf \EndFor \Return $finalPlan$ \Procedure{prioritizePair}{$p_i, p_{i+1}$} \State $newPair \leftarrow array[2]$ \If {$p_i.tail$ is joined with $p_{i+1}.head$} \State $newPair$[0]$\leftarrow p_i$ \State $newPair$[1]$\leftarrow p_{i+1}$ \State \Return $newPair$ \Else \For{\textbf{each} $t_k, \in p_{i+1}$} \If {$p_i.tail$ is joined with $t_k$} \State $p_i.push(t_k)$ \State $p_{i+1}.remove(t_k)$ \State $newPair \leftarrow $prioritizePair($p_i, p_{i+1}$) \EndIf \EndFor \State $newPair$[0]$\leftarrow p_i$ \State $newPair$[1]$\leftarrow p_{i+1}$ \Return $newPair$ \EndIf \EndProcedure \end{algorithmic} \end{algorithm} \subsection{Experiments} We implemented a proof-of-concept version of the query planner in Jena ARQ, and conducted experiments on Jena TDB\footnote{https://jena.apache.org/documentation/tdb/}, comparing against several statistical reordering approaches, namely Stocker et al's PFJ and ONS \cite{stocker2008sparql}, Kalayci et al's Ant System \cite{kalayci2015ant} as well as Jena TDB's Fixed (JF) and Weighted (JW) optimizers, and the reordering performed by the open source edition of Openlink Virtuoso 7.2. As triple pattern ordering is a problem that is orthogonal to the low-level implementation specifics and design choices (e.g., join implementations, indexing), we use Jena TDB as a common testbed for all approaches to strictly assess and compare the effect of triple pattern orderings on query processing. We used LUBM to generate a synthetic dataset of 15m triples, and measured execution times for 29 queries on the LUBM dataset. The queries consist of 14 original queries provided by LUBM, and an additional 15 queries of increasingly complex shapes and sizes, used in \cite{kalayci2015ant}. All code and queries are available online\footnote{https://github.com/mmeimaris/sparqlDistancePlanner}. For each approach, we extract an ordering, and feed that directly to the Jena query processor. Table 2 summarizes the percentages of best plans for all queries. These are computed by measuring, for each approach, how many plans were the fastest\footnote{We measure execution times by executing each query 10 times and reporting the mean.} in relation to the other approaches, for all queries. Because of the fact that the fastest plan can be reached by more than one method, we differentiate between plans based on the generated orderings, and then we compare their execution times. As can be seen, our method outperforms all other methods for queries of different patterns, namely the original 14 queries, and the 15 additional queries of star, chain, chain-star, and cyclic patterns, achieving the best plan 90\% of the time, with Virtuoso coming second with 66\% of its orderings being the best. In order to assess how our planner performs with respect to different query types, we use 15 extended queries originally constructed in \cite{kalayci2015ant}. These represent different query types, and are thus classified into star, chain, cyclic, and chain-star query types, of varying triple pattern sizes. Their original intention was to provide more complex query structures of up to 14 triple patterns for LUBM dataset, because the original queries are limited to 1-6 triple patterns. We measure execution times for each query type separately, and report these in Tables 3, 4, 5, and 6, for star, chain, cyclic, and chain-star queries respectively. For each query, the best ordering is marked with a bold execution time. While we do not show the actual triple orders that were derived by each method, multiple bold values in the same row indicate that the same order has been achieved by more than one approach. Note that small differences in execution times for the same orderings can be attributed to lags imposed by I/O, available CPU resources and other external factors. For computing the distance matrix from $Q_m$, we used a simple Euclidean distance function that measuring distances between triple patterns in the m-dimensional space. Overall, the results are encouraging for further pursuing this direction. In fact, in 90\% of the queries, our method generated the fastest plan in relation to the other approaches, of whom the best (VIRT) achieved a rate of 66\%, and the rest achieved less than 60\%. \begin{table} \centering \caption{$Q_m$ matrix for reference query.} \begin{tabular}{|c|c c c c c c|} \hline & ?X & Student & ?Y & Dpt & Univ0 & ?Z\\ \hline $t_1$ & 99k & 99k & 0 & 0 & 0 & 0\\ $t_2$ & 0 & 0 & 189 & 189 &0 & 0 \\ $t_3$ & 106k & 0 & 106k & 0 & 0 & 0 \\ $t_4$ & 0 & 0 & 239 & 0 & 239 & 0 \\ $t_5$ & 106k & 0 & 0 & 0 & 0 & 106k \\ \hline \end{tabular} \end{table} \begin{table} \scriptsize \setlength{\tabcolsep}{8pt} \renewcommand{\arraystretch}{1.8} \caption{Percentage of plans that are best compared to other methods for all queries.} \begin{tabular}{|c c c c c c c |} \hline JW & JF & ONS & PFJ & ANT & VIRT & OUR\\ \hline 59\% & 48\% & 28\% & 59\% & 31\% & 66\% & 90\% \\ \hline \end{tabular} \end{table} \begin{table}[ht] \scriptsize \setlength{\tabcolsep}{6pt} \renewcommand{\arraystretch}{1.8} \caption{Query execution times (seconds) for the star queries.} \begin{tabular}{|c|c c c c c c c|} \hline & JW & JF & ONS & PFJ & ANT & VIRT & OUR\\ \hline Q1 & 1.01 & 1.14 & 1.07& \textbf{0.61} & 1.05 & 0.63 & \textbf{0.57} \\ Q2 & 0.91 & 0.92 & 4.71 & 0.76 & 0.79 & 0.74 & \textbf{0.67} \\ Q3 & 0.23 & 0.27 & 0.23 & 0.15 & 3.62 & 0.21 & \textbf{0.13} \\ Q4 & 0.17 & 0.08 & 0.07 & 0.07 & 0.42 & \textbf{0.05} & \textbf{0.03} \\ \hline \end{tabular} \end{table} \begin{table}[ht] \scriptsize \setlength{\tabcolsep}{6pt} \renewcommand{\arraystretch}{1.8} \caption{Query execution times (seconds) for the chain queries.} \begin{tabular}{|c|c c c c c c c |} \hline & JW & JF & ONS & PFJ & ANT & VIRT & OUR\\ \hline Q1 & 0.30 & 0.33 & 0.37 & \textbf{0.14} & 1.09 & 0.39 & \textbf{0.14} \\ Q2 & \textbf{0.49} & \textbf{0.56} & \textbf{0.52} & 0.77 & \textbf{0.50} & \textbf{0.48} & \textbf{0.48} \\ Q3 & 322 & 323 & 131 & 292 & 139 & \textbf{55} & \textbf{53} \\ Q4 & \textbf{54} & \textbf{53} & \textbf{53} & 163 & 209 & 182 & \textbf{49} \\ \hline \end{tabular} \end{table} \begin{table}[ht] \scriptsize \setlength{\tabcolsep}{8pt} \renewcommand{\arraystretch}{1.8} \caption{Query execution times (seconds) for the cyclic queries.} \begin{tabular}{|c|c c c c c c c |} \hline & JW & JF & ONS & PFJ & ANT & VIRT & OUR\\ \hline Q1 & 0.05 & 0.07 & 0.05 & \textbf{0.04} & \textbf{0.04} & \textbf{0.04} & \textbf{0.04}\\ Q2 & 0.25 & 0.24 & 0.19 & \textbf{0.02} & 0.20 & 0.18 & \textbf{0.02} \\ Q3 & \textbf{0.06} & \textbf{0.06} & 0.31 & 2.37 & 0.74 & \textbf{0.04} & 4.50 \\ Q4 & 1.06 & 0.94 & 1.30 & 0.64 & 0.81 & \textbf{0.04} & 0.52 \\ \hline \end{tabular} \end{table} \begin{table} \scriptsize \setlength{\tabcolsep}{8pt} \renewcommand{\arraystretch}{1.8} \caption{Query execution times (seconds) for the chain-star queries.} \begin{tabular}{|c|c c c c c c c |} \hline & JW & JF & ONS & PFJ & ANT & VIRT & OUR\\ \hline Q1 & \textbf{4.16} & \textbf{4.16} & \textbf{4.15} & 7.39 & \textbf{4.15} & 5.15 & \textbf{4.15} \\ Q2 & 5.95 & 5.92 & 5.93 & 5.83 & 6.67 & \textbf{0.49} & \textbf{0.42} \\ Q3 & 6.11 & 4.98 & 4.75 & 214.92 & 6.37 & \textbf{2.35} & \textbf{2.31} \\ \hline \end{tabular} \end{table} \section{Related Work} RDF data management systems generally follow three storage schemes, namely \textit{triples tables}, \textit{property tables}, and \textit{vertical partitioning}. A triples table has three columns, representing the subject, predicate and object (SPO) of an RDF triple. This technique replicates data in different orderings in order to facilitate sort-merge joins. RDF-3X \cite{neumann2010RDF} and Hexastore \cite{weiss2008hexastore} build tables on all six permutations of SPO. Built on a relational backbone, Virtuoso \cite{erling2010virtuoso} uses a 4-column table for quads, and a combination of full and partial indexes. These methods work well for queries with small numbers of joins, however, they degrade with increasing sizes, unbound variables and joins. \textit{Property Tables} places data in tables with columns corresponding to properties of the dataset, where each table identifies a specific resource type. Each row identifies a subject node and holds the value of each property. This technique has been implemented experimentally in Jena \cite{wilkinson2006jena} and DB2RDF \cite{bornea2013building}, and shows promising results when resource types and their properties are well-defined. However, this causes extra space overhead for null values in cases of sparse properties \cite{abadi2007scalable}. Also, it raises performance issues when handling complex queries with many joins, as the amounts of intermediate results increase \cite{janik2005brahms}. Vertical partitioning segments data in two-column tables. Each table corresponds to a property, and each row to a subject node \cite{abadi2007scalable}. This provides great performance for queries with bound objects, but suffers when the table sizes have large variations in size \cite{sidirourgos2008column}. TripleBit \cite{yuan2013triplebit} broadly falls under vertical partitioning. In TripleBit, the data is vertically partitioned in chunks per predicate. While this reduces replication, it suffers from the same problems as property tables. It does not consider the inherent schema of the triples in order to speed up the evaluation of complex query patterns. In distributed settings, a growing body of literature exists, with systems such as Sempala \cite{schatzle2014sempala}, H2RDF \cite{papailiou2014h} and S2RDF \cite{schatzle2016s2rdf}. However, these are based on parallelization of centralized indexing and query evaluation schemes. For these reasons, latest state of the art approaches rely on implicit schema detection in order to derive a hidden schema from RDF data and index/store triples based on this schema. Furthremore, due to the tabular structure that tends to implicitly underly RDF data, recent works have been implemented in relational backbones. In our previous work \cite{meimaris2017extended}, we defined \textit{Extended Characteristic Sets (ECSs)} as typed links betwen CSs, and we showed how ECSs can be used to index triples and greatly improve query performance. In \cite{pham2015deriving}, the authors identify and merge CSs, similar to our approach, into what they call an \textit{emergent schema}. However, their main focus is to extract a human-readable schema with appropriate relation labelling. They do not use hierarchical information of CSs, rather they use semantics to drive the merging process. In \cite{pham2016exploiting} it is shown how this \textit{emergent schema} approach can assist query performance, however, the approach is limited by the constraints of human-readable schema discovery. In our work, query performance, indexing and storage optimization are the main aims of the merging process, and thus we are not concerned about providing human-readable schema information or any form of schema exploration. In \cite{montoya2017odyssey}, the authors use CSs and ECSs in order to assist cost estimation for federated queries, while in \cite{gubichev2014exploiting}, the authors use CSs in order to provide better triple reordering plans. To the best of our knowledge, this is the first work to exploit hierarchical CS relations in order to merge CSs and improve query performance. \section{Hierarchical CS Merging} \subsection{Preliminaries} The RDF model does not generally enforce structural rules in the representation of triples; within the same dataset there can be largely diverse sets of predicates emitting from nodes of the same semantic type \cite{meimaris2017extended,pham2015deriving,neumann2011characteristic}. \textit{Characteristic Sets} (CS)\cite{neumann2011characteristic} capture this diversity by representing implied node types based on the set of properties they emit. Formally, given a collection of triples $D$, and a node $s$, the characteristic set $cs(s)$ of $s$ is $cs(s) = \{p \mid \exists o: (s, p, o) \in D\}$. The set of properties of a CS $cs_i$ is denoted with $P_i$. Furthermore, in a given dataset, each CS represents a set of records identified by a subject node, and all of the values of the subject node (i.e., objects) for the predicates in $P_i$. We denote the set of all records of $cs_i$ as $r_i$, while $cs_i$ is represented by a relational table $c_i$ that is defined by these two elements, i.e., $c_i = (P_i, r_i)$. The tuples in $c_i$ are of the form $(s,p_{i,1},\dots, p_{i,k})$, where $s$ is the identifier column (e.g. URI) of a subject node and $p_{i,1}, p_{i,2}, \dots ,p_{i,k}$ are the values, i.e. object nodes, of the properties in $P_i$ for $s$. In the context of this paper, with the term \textit{Characteristic Set} we will refer collectively to the properties and records of a CS, i.e., its relational table, rather than just the set of properties proposed in the original definition, for the sake of simplicity. Within a given dataset, CSs often exhibit hierarchical relationships, as a result of the overlaps in their comprising sets of properties. For example, consider two CSs, $c_1, c_2$, describing human beings, with $P_{1} = \{type, name\}$ and $P_{2} = \{type, name, marriedTo\}$. It can be seen that $P_{1} \subset P_{2}$ and thus $c_1$ is a parent of $c_2$. This relationship entails an overlap of properties that define the CSs, and can be exploited in order to provide a means to merge common CSs based on the specialization or generalization of the node types they describe. In what follows, we formally define the notions of CS \textit{subsumption}, \textit{hierarchy} and \textit{ancestral sub-graphs}. \textit{Definition 1}. \textbf{\textit{(CS Subsumption)}}. Given two CSs, $c_i$ and $c_j$, and their property sets $P_i$ and $P_j$, then $c_i$ subsumes $c_j$, or $c_i \succ c_j$, when the property set of $c_i$ is a proper subset of the property set of $c_j$, or $P_i \subset P_j$. This subsumption forms parent-child relationships between CSs. CS subsumption relationships can be seen in Figure \ref{fig:combinations}(a) as directed edges between nodes. The set of all parent-child relationships defines a CS hierarchy as defined in the following. \textit{Definition 2}. \textbf{\textit{(CS Hierarchy and Inferred Hierarchy)}}. CS subsumption creates a partial ordering that essentially defines a \textit{hierarchy} such that when $c_i \succ c_{j}$, then $c_{i}$ is a parent of $c_{j}$. Formally, a CS hierarchy is a graph lattice $L = (V, E)$ where $V \in C$ and $E \in (V \times V)$. A directed edge between two CS nodes $c_1, c_2$ exists in $L$, when $c_1 \succ c_2 $ and there exists no other $c_i$ such that $c_1 \succ c_i \succ c_2$. An example CS hierarchy can be seen in Figure \ref{fig:combinations}(a). Given a hierarchy $L$, we denote the \textit{hierarchical closure} of $L$ with $L_{c}$, so that $L_{c}$ extends $L$ to contain inferred edges between hierarchically related nodes that are not consecutive, e.g. a node and its grandchildren. An example inferred hierarchy can be seen in Figure \ref{fig:combinations}(c) for a sub-graph of the graph in Figure \ref{fig:combinations}(a), with the inferred relationships in dashed lines. In the remainder of this paper, we refer to $L_{c}$ as the \textit{inferred hierarchy} of $L$ \textit{Definition 3}. \textbf{\textit{(CS Ancestral Sub-graphs)}}. Given an inferred hierarchy $L_{c} = (V, E)$, a CS $c_{base}$ and set of CSs $c_1, \dots, c_k$, then $a = (V^{'}, E^{'})$ is an ancestral sub-graph with $c_{base}$ as the lowermost child when $\forall i \in [1..k]$, it holds that $c_{i} \succ c_{base}$, and $(c_{i}, c_{base}) \in E^{'}$. This means that any sub-graph with $c_{base}$ as a sink node will be an ancestral sub-graph of $c_{base}$. Thus, it holds that $a \subset L_c$. For instance, in Figure \ref{fig:combinations}(c), nodes $c_7, c_4, c_2$ form an ancestral sub-graph with $c_7$ as the base CS. Similarly, nodes $c_6, c_4, c_2$ and $c_6, c_5, c_2$ form ancestral sub-graphs with $c_6$ as base CS. Logically, we map each CS to a relational table, so that for a CS $c_i$ we create a relational table $t_i = (s, p_{i,1}, p_{i,2}, \dots, p_{i,k})$, where $s$ is the id of the subject and $p_{i,1}\dots, p_{i,k}$ are the properties that belong to $P_i$, and then we use the CS hierarchy in order to merge the nodes of an ancestral sub-graph with $c_i$ as base into a single table. Specifically, we exploit the property set overlap in order to merge together smaller parent CSs with larger child CSs, in order to minimize the effect of NULL values that will appear for properties in smaller CSs that do no exist in the larger CSs. Thus, $c_{base}$ will be the most specialized CS in its ancestral sub-graph. For this reason, we define a merge operator, $hier\_merge$, as follows. \textit{Definition 4}. \textbf{\textit{(Hierarchical CS Merge)}}. Given an ancestral sub-graph $a = (V^{'}, E^{'})$, where $V^{'} = \{c_1 = (P_1, r_1), c_2 = (P_2, r_2),\dots, c_k=(P_k, r_k)\}$ as defined above, then a hierarchical merge of $a$ is given as follows: $hier\_merge(a) = c_{a}$, where $c_a = (P_1, r_a)$. Here, $P_1$ is the most specialized property set in $a$, as $c_1$ does not have any children in $a$, while $r_a = \bigcup\limits_{i=1}^{k}r_i^{'}$ is the UNION of the records of all CSs in $V^{'}$, where $r_i^{'}$ is the projection of $r_i$ on $P_1$. This means that $r_i^{'}$ will contain NULL values for all the non-shared properties of $P_1$ and $P_i$, i.e., $P_1 \setminus P_i$. In essence, $hier\_merge$ is an \textit{edge contraction} operator that merges all nodes of an ancestral sub-graph into one, while removing the edges that connect them. For instance, assume that $V^{'} = \{c_0 = (P_0, r_0), c_1 = (P_1, r_1), c_2 = (P_2, r_2)\}$ is the set of vertices of an ancestral sub-graph with three CSs, with $P_0 = \{p_a,p_b\}$, $P_1 = \{p_a, p_b, p_c\}$ and $P_2 = \{p_a, p_b, p_c, p_d\}$. Thus, $c_0 \succ c_1 \succ c_2$. Hierarchical merging can be seen in Figure \ref{fig:table_merging}. \textit{Definition 5}. \textbf{\textit{(Merge Graph)}}. Given an inferred CS hierarchy $L_{c} = (V,E)$, a merge graph is a graph $L^{'}=(V^{'}, E^{'})$ that consists of a set of $n$ ancestral sub-graphs, and has the following properties: (i) $L^{'}$ contains all nodes in $L$ such that $V^{'} \equiv V$, i.e., it covers all CSs in the input dataset, (ii) $L^{'}$ contains a subset of the edges in $L$ such that $E^{'} \subset E$, (iii) each node is contained in exactly one ancestral sub-graph $a_i$, (iv) all ancestral sub-graphs are pair-wise disconnected, i.e., there exist no edges between the nodes of different ancestral sub-graphs. Thus, each ancestral sub-graph can be contracted into one node unambiguously, using the $hier\_merge$ operator. Also, the total number of relational tables will be equal to the number of ancestral sub-graphs in the merge graph. \begin{figure}[t] \centering \includegraphics[width=1\textwidth]{example_graph.png} \caption{(a) A CS hierarchy graph with dense nodes colored in deep purple, (b) the connected components derived by cutting off descendants from dense nodes, (c) a connected component with dashed lines representing inferred hierarchical relationships, (d) all possible assignments of dense nodes to non-dense nodes.}\label{fig:combinations} \end{figure} \textbf{Problem Formulation. }Given an inferred CS hierarchy $L_{c} = (V, E)$, the problem is to find a merge graph $L^{'} = (V, E^{'})$ in the form of a set of disconnected ancestral sub-graphs, that provides an optimal way to merge CS nodes. In other words, the problem is to find the best set of ancestral sub-graphs from an inferred hierarchy $L_{c}$ that minimize an objective cost function $cost(x)$, or more formally: \begin{align} L^{'} = argmin_{x \subset L_c}cost(x) \end{align} This formulation entails several problems. First, the notion of cost depends on possibly subjective factors, such as the query workload, the storage technology, the input dataset and so on. There is no universal cost model that can be deployed in order to assess the effectiveness of a merge graph. Moreover, neither the number of ancestral sub-graphs, nor the set of sub-graph roots is known as part of the input. A CS hierarchy of $n$ nodes can potentially create $2^n$ sub-graphs, while the number of possible sub-graph roots is also exponential with the respect to the hierarchy size. Thus, given an arbitrary cost function, this is a problem of \textit{non-uniform graph partitioning} on the inferred hierarchy $L_c$, which is known to be NP-Hard. That is, even with a deployed cost model, it is still an exponential problem to enumerate all possible sub-graphs and find the one with the minimum cost. For these reasons, we approach the problem by deploying a set of rules and heuristics that find a good merge graph efficiently and offer improved storage and query performance, as will be shown in the experiments. \subsection{CS Retrieval and Merging} The primary focus of this work is to improve the efficiency of the storage and query capabilities of relational RDF engines by exploiting the implicit schema of the data in the form of CSs. However, CS merging results in several problems that need to be addressed in this context. These are discussed in what follows. First, the problem of selecting ancestral sub-graphs is a computationally hard one, as mentioned earlier. For this reason, we rely on a simple heuristic in order to seed the process and provide an initial set of ancestral sub-graph \textit{sink nodes}, that will form the bases of the final merged tables, as defined in Definition 3. For this, we identify \textit{dense} CS nodes in the hierarchy (i.e, with large cardinalities) and use these nodes as the bases of the ancestral sub-graphs. While node density can be defined in many different ways, in the context of this work we define a $c_i$ to be dense, if its cardinality is larger than a linear function of the maximum cardinality of CSs in $D$, i.e., a function $d: N \rightarrow R$, with $d(c_i) = m \times |r_{max}|$. Here, $m \in \left[0,1\right]$ is called the \textit{density factor}, and $r_{max}$ is the cardinality of the largest CS in $D$. This means that, by definition, if $m = 0$, no CSs will be merged, because all CSs will be considered dense and thus each CS will define its own ancestral sub-graph, while if $m = 1$, all no ancestral sub-graph will be defined, and all CSs will be merged to one large table, as no CS has a cardinality larger than that of the largest CS. With a given $m$, the problem is reduced to finding the optimal ancestral sub-graph for each given dense node. Second, merging tables results in the introduction of NULL values for the non-shared columns, which can degrade performance. Specifically, merging CSs with different attribute sets can result in large numbers of NULL values in the resulting table. Given a parent CS $c_1 = (P_1, r_1)$ and a child CS $c_2 = (P_2, r_2)$ with $\left\vert{P_1}\right\vert < \left\vert{P_2}\right\vert$ and $\left\vert{r_1}\right\vert >> \left\vert{r_2}\right\vert$, the resulting $\left\vert{P_2 \setminus P_1}\right\vert \times \left\vert{r_1}\right\vert$ NULL cells will be significantly large compared to the total number of $r_1+r_2$ records, thus potentially causing poor storage and querying performance\cite{pham2015deriving}. For this reason, CS merging must be performed in a way that will minimize the presence of NULL values. The following function captures the NULL-value effect of the merge of two CSs $c_i = (P_i, r_i), c_j = (P_j,r_j)$ with $c_i \succ c_j$: \begin{align} r_{null}(c_i,c_j) = \frac{|P_j \setminus P_i| \times |r_i|}{(|r_j|)} \end{align} Intuitively, $r_{null}$ represents the ratio of null values to the cardinality of the base CS in the merge. The numerator of the fraction represents the total number of cell values that will be null, as the product of the number of non-shared properties and the cardinality of the parent CS. The denominator represents the cardinality of the base CS. Hence, the base CS must be a descendant (i.e., with more properties) in order to minimize the presence of NULLs. In order to assess an ancestral sub-graph, we use a generalized version of $r_{null}$ that captures the NULL value effect on the whole sub-graph: \begin{align} r^{g}_{null}(g)|_{c_{d}} = \frac{\sum_{i=1}^{|g|}{|P_{d} \setminus P_i| \times |r_i|}}{|r_d|+\sum_{i=1}^{|g|}(|r_i|)} \end{align} Here, $c_d = (P_d, r_d)$ is the dense root of sub-graph $g$. However, merging a parent to a dense child changes the structure of the input graph, as the cardinality of the dense node is increased. To accommodate this, we define a cost function that works on the graph level, as follows: \begin{align} cost(g) = \sum_{i=1}^{n}{r^{g}_{null}(g_i)|_{c_{di}}} \end{align} where $n$ is the number of dense nodes, $c_{di}$ is a dense node and $g_i$ is the ancestral sub-graph with $c_{di}$ as the base node. Given this cost model and a pre-defined set of dense nodes, our exact algorithm will find the optimal sub-graph for each dense node. An inferred hierarchy graph can be converted to a set of connected components that are derived by removing the outgoing edges from dense nodes, since we are not interested in merging children to parents, but only parents to children. An example of this can be seen in Figure \ref{fig:combinations}(b). For each component, we can compute $cost(g)$ as the sum of the costs of these components. The main idea is to identify all connected components in the CS graph, iterate through these components, enumerate all sub-graphs within the components that start from the given set of dense nodes, and select the optimal partitioning for each component. The algorithm can be seen in Algorithm \ref{alg:optimalMerge}. The algorithm works by first identifying all connected components of the inferred hierarchy (Line 2). Identifying connected components is trivially done using standard DFS traversal, and is not shown in the Algorithm. Then, we iterate each component (Line 3), and for each component, we generate all possible sub-graphs. Then, we calculate the cost of each sub-graph (Line 7) and if it is smaller than the current minimum, the minimum cost and best sub-graph are updated (Lines 8-9). Finally, we add the best sub-graph to the final list (Line 11) and move to the next component. \begin{figure}[t] \centering \includegraphics[width=0.4\textwidth]{table_merging_3.png} \caption{Merging the tables of $c_0$, $c_1$ and $c_2$.}\label{fig:table_merging} \end{figure} To generate the sub-graphs, we do not need to do an exhaustive generation of $2^n$ combinations, but we can rely on the observation that each non-dense node must be merged to exactly one dense node. Therefore, sub-graph generation is reduced to finding all possible assignments of dense nodes to the non-dense nodes. An example of this can be seen in figure \ref{fig:combinations}. In the figure, nodes $c_2, c_4, c_5$ are non-dense, while nodes $c_6, c_7, c_8$ are dense. All possible and meaningful sub-graphs are enumerated in the table at the right of the figure, where we assign a dense node to each of the non-dense nodes. An assignment is only possible if there exists a parent-child relationship between a non-dense node and a dense node, even if it is an inferred one (e.g. $c_2$ is an inferred parent of $c_7$). Hence, the problem of sub-graph generation becomes one of generating combinations from different lists, by selecting one element from each list. The number of lists is equal to the number of non-dense nodes, and the elements of each list are the dense nodes that are related to the non-dense node. \textbf{Complexity Analysis. } Assuming that a connected component $g$ has $k$ non-dense nodes and $d$ dense nodes, and each non-dense node $k_i$ is related to $e(k_i)$ dense nodes, then the number of sub-graphs that need to be enumerated are $\prod_{i=1}^{k}{e(k_i)}$. In the example of figure \ref{fig:combinations}, the total number of sub-graphs is $e(c_2) \times e(c_4) \times e(c_5) = 3 \times 2 \times 1 = 6$. In the worst case all $k$ nodes are parents of all $d$ nodes. Then, the number of total sub-graphs is $k^d$, which makes the asymptotic time complexity of the algorithm $O(k^d)$. \begin{algorithm} \KwData{An inferred hierarchy lattice $L_c$ as a adjacency list , and a set of dense CSs $D$} \KwResult{A set of optimal ancestral sub-graphs} init $finalList$\; $connectedComponents \gets findConnectedComponents(L_c)$\; \For {\textbf{each} $connectedComponent$}{ init $min \gets MAX\_VALUE$\; init $bestSubgraph$ \; \While {$next \gets connectedComponent.generateNextSubgraph()$} { \uIf{$cost(next) < min$}{ $min \gets cost(next)$\; $bestSubgraph \gets next$\; } } $finalList.add(bestSubgraph)$\; } return $finalList$\; \caption{\emph{optimalMerge}}\label{alg:optimalMerge} \end{algorithm} \subsection{Greedy Approximation} For very small $d$ (e.g. $d < 4$), the asymptotic complexity of $O(k^d)$ is acceptable. However, in real-world cases, the number of connected components can be small, making $d$ large. For this reason, we introduce a heuristic algorithm for approximating the problem, that does not need to enumerate all possible combinations, but instead relies on a greedy objective function that attempts to find the local minimum with respect to our defined cost model for each non-dense node. Note that it lies beyond the scope of this work to compute the degree of approximation to the optimal solution, however, in our experiments, the heuristic solution is shown to provide significant performance gains. The main idea behind the algorithm is to iterate the non-dense nodes, and for each non-dense node, calculate the $r_{null}$ function and find the dense node that minimizes this function for the given non-dense node. Then, the cardinalities will be recomputed and the next non-dense node will be examined. The algorithm can be seen in Algorithm \ref{alg:greedyMerge}. In the beginning, the algorithm initiates a hash table, $mergeMap$, with an empty list for each dense node (Lines 1-4). Then, the algorithm iterates all non-dense nodes (Line 5), and for each dense node, it calculates the cost $r_{null}$ of merging it to each of its connected dense nodes (Lines 5-13), keeping the current minimum cost and dense node. In the end, the current non-dense node is added to the list of the dense node that minimizes $r_{null}$ (Line 14). Notice that we do not need to split the hierarchy into connected components in order for $greedyMerge$ to work. \textbf{Complexity Analysis. }Given $k$ non-dense nodes and $d$ dense nodes, where each non-dense node $k_i$ is related to $e(k_i)$ dense nodes, the $greedyMerge$ algorithm needs $\sum_{i=1}^{k}{e(k_i)}$ iterations, because we need to iterate all $e(k_i)$ nodes for each $k_i$. In the worst case, every $k_i$ is related to all $d$ dense nodes, requiring $kd$ iterations. Assuming a constant cost for the computation of $r_{null}$, then the asymptotic complexity of the greedy algorithm is $O(kd)$, which is a significant performance when compared to the exponential complexity of $optimalMerge$. \begin{algorithm} \KwData{A hash table $p$ mapping non-dense CSs to their dense descendants, a set of dense CSs $D$, and a set of non-dense CSs $K$} \KwResult{A hash table mapping dense CSs to sets of non-dense CSs to be merged} init $mergeMap$\; \For {\textbf{each} $d \in D$}{ $mergeMap.put(d, new List())$\; } \For {\textbf{each} $k \in K$}{ $min \gets MAX\_VALUE$\; init $bestDense$\; \For {\textbf{each} $d_k \in p.get(k)$}{ $cost \gets r_{null}(k, d_k)$\; \uIf{$cost < min$}{ $min \gets cost$\; $bestDense \gets d_k$\; } } $mergeMap.get(bestDense).add(k)$\; } return $mergeMap$\; \caption{\emph{greedyMerge}}\label{alg:greedyMerge} \end{algorithm} Obviously, this process does not necessarily cover all CSs of the input dataset. The percentage of the dataset that is covered by this process is called \textit{dense CS coverage}. The remainder of the CSs that are not contained by any merge path are aggregated into one large table containing all of their predicates. If the total coverage of the merging process is large, then this large table does not impose a heavy overhead in query performance, as will be shown in the experiments. Finally, we load the data in the corresponding tables. \begin{figure}[ht] \centering \includegraphics[width=1\textwidth]{greedy_merging.png} \caption{An example of greedy merging. Dense nodes are coloured in deep purple. At each step, the non-dense node under examination is coloured with green, while the edge that minimizes $r_{null}$ can be seen in bold.}\label{fig:path_pruning} \end{figure} \subsection{Implementation} We implemented \textit{raxonDB} as a storage and querying engine that supports hierarchical CS merging, and can be deployed on top of standard RDBMS solutions. Specifically, we used PostgreSQL 9.6, but \textit{raxonDB} can be adapted for other relational databases as well. The architecture of \textit{raxonDB} can be seen in Figure \ref{fig:arch}. \textbf{CS Retrieval and Merging. }The processes of retrieving and merging CSs take place during the loading stage of an incoming RDF dataset. CS retrieval is a trivial procedure that requires scanning the whole dataset and storing the unique sets of properties that are emitted from the subject nodes in the incoming triples, and is adopted from our previous work in \cite{meimaris2017extended} where it is described in detail. After retrieving the CSs, the main idea is to compute the inferred CS hierarchy and apply one of the described merging algorithms. Finally, each set of merged CSs is stored in a relational table. In each table, the first column represents the subject identifier, while the rest of the columns represent the union of the property sets of the merged CSs. For multi-valued properties, we use PostgreSQL's array data type in order to avoid duplication of the rows. \textbf{Indexing. }We deploy several indexes in \textit{raxonDB}. First off, we index the subject id for each row. We also build foreign-key indexes on object-subject links between rows in different CSs, i.e., when a value of a property in one CS is the subject id of another CS. Next, we use standard B+tree for indexing single-valued property columns, while we use PostgreSQL's GIN indexes, which apply to array datatypes for indexing multi-valued properties. This enables fast access to CS chain queries, i.e., queries that apply successive joins for object-subject relationships. Furthermore, we store these links on the schema level as well, i.e., we keep an index of CS pairs that are linked with at least one object-subject pair of records. These links are called Extended Characteristic Sets (ECSs) and are based on our previous work in \cite{meimaris2017extended}. With the ECS index, we can quickly filter out CSs that are guaranteed not to be related, i.e., no joins exist between them, even if they are individually matched in a chain of query CSs. Other metadata and indexes include the property sets of CSs, and which properties can contain multiple values in the same CS. \begin{figure}[t] \centering \includegraphics[scale=0.35]{arch3.png} \caption{Architecture of \textit{raxonDB}.}\label{fig:arch} \end{figure} \textbf{Query Processing. }Processing SPARQL queries on top of merged CSs entails (i) parsing the queries, (ii) retrieving the query CSs, (iii) identifying the joins between them, and (iv) mapping them to merged tables in the database. Steps (i)-(iii) are inherited from our previous work in \cite{meimaris2017extended}. For (iv), a query CS can match with more than one table in the database. For instance, consider a query containing a chain of three CSs, $q_1 \bowtie q_2 \bowtie q_3$, joined sequentially with object-subject joins. Each query CS $q_i$ matches with all tables whose property sets are supersets of the property set of $q_i$. Thus, each join in the initial query creates a set of \textit{permutations} of table joins that need to be evaluated. For instance, assume that $q_1$ matches with $c_1, c_2$, while $q_2$ matches with $c_3$ and $q_3$ matches with $c_4, c_5$. Furthermore, by looking up the ECS index, we derived that the links $\left[c_1,c_3\right]$, $\left[c_2,c_3\right]$, $\left[c_3,c_4\right]$ and $\left[c_3,c_5\right]$ are all valid, i.e., they correspond to candidate joins in the data. Then, $\left[c_1, c_3, c_4\right]$, $\left[c_1, c_3, c_5\right]$, $\left[c_2, c_3, c_4\right]$ and $\left[c_2, c_3, c_5\right]$ are all valid table permutations that must be processed. Two strategies can be employed here. The first is to join the UNIONs of the matching tables for each $q_i$, and the other is to process each permutation of tables separately and append the results. Given the filtering performed by the ECS indexing approach, where we can pre-filter CSs based on the relationships between them, the UNION would impose significant overhead and eliminate the advantage of ECS indexing. Therefore, we have implemented the second approach, that is, process a separate query for each permutation. Finally, due to the existence of NULL values in the merged tables, we must add explicit IS NOT NULL restrictions for all the properties that are contained in each matched CS and are not part of any other restriction or filter in the original query. \section{Experimental Evaluation} We implemented \textit{raxonDB} on top of PostgreSQL\footnote{The code and queries are available in https://github.com/mmeimaris/raxonDB}. We did not extend our previous native RDF implementation of \textit{axonDB} \cite{meimaris2017extended}, because given the underlying relational schema of the CS tables, we decided to rely on a well-established relational engine for both the planning and the execution of queries, instead of re-implementing it. As the focus of this paper is to improve RDF storage and querying efficiency in relational settings, we rely on existing mechanisms within PostgreSQL for I/O operations, physical storage and query planning. In this set of experiments, we report results of implementing $hier\_merge$ with the greedy approximation algorithm, as experimenting with the optimal algorithm failed to finish the merging process even in datasets with small numbers of CSs. \textbf{Datasets. }For this set of experiments, we used two synthetic datasets, namely \textit{LUBM2000} ($\approx$300m triples), and WatDiv ($\approx$100m triples), as well as two real-world datasets, namely \textit{Geonames} ($\approx$170m triples) and \textit{Reactome} ($\approx$15m triples). LUBM \cite{guo2005lubm} is a customizable generator of synthetic data that describes academic information about universities, departments, faculty, and so on. Similarly, WatDiv\cite{alucc2014diversified} is a customizable generator with more options for the production and distribution of triples to classes. \textit{Reactome}\footnote{http://www.ebi.ac.uk/rdf/services/reactome} is a biological dataset that describes biological pathways, and \textit{Geonames}\footnote{http://www.geonames.org/ontology/documentation.html} is a widely used ontology of geographical entities with varying properties. Geonames maintains a rich graph structure as there is a heavy usage of hierarchical area features on a multitude of levels. \textbf{Loading. }In order to assess the effect of hierarchical merging in the loading phase, we performed a series of experiments using all four datasets. For this experiment, we measure the size on disk, the loading time, the final number of merged tables, as well as the number of ECSs (joins between merged tables) and the percentage of triple coverage by CSs included in the merging process, for varying values of the density factor $m \in [0,1]$. The results are summarized in Table \ref{table:loading}. As can be seen, the number of CS, and consequently tables, is greatly reduced with increasing values of $m$. As the number of CSs is reduced, the expected number of joins between CSs is also reduced, which can be seen in the column that measures ECSs. Consequently, the number of tables can be decreased significantly without trading off large amounts of coverage by dense CSs, i.e. large tables with many null values. Loading time tends to be slightly greater as the number of CSs decreases, and thus the number of merges increases, the only exception being WatDiv, where loading time is actually decreased. This is a side-effect of the excessive number of tables ($=5667$) in the simple case which imposes large overheads for the persistence of the tables on disk and the generation of indexes and statistics for each one. \begin{figure*}[t] \def\tabularxcolumn#1{m{#1}} \begin{tabularx}{\linewidth}{@{}cXX@{}} \begin{tabular}{ccc} \subfloat[Execution time (seconds) for LUBM]{\includegraphics[width=0.31\textwidth]{queries_lubm_merge_4.png}} & \subfloat[Execution time (seconds) for Geonames]{\includegraphics[width=0.3\textwidth]{queries_geonames_merge_5.png}} & \subfloat[Execution time (seconds) for Reactome]{\includegraphics[width=0.31\textwidth]{queries_reactome_merge_4.png}} \\ \end{tabular} \end{tabularx} \caption{Query execution times in milliseconds}\label{fig:exp_fig_1} \end{figure*} \begin{figure*}[t] \def\tabularxcolumn#1{m{#1}} \begin{tabularx}{\linewidth}{@{}cXX@{}} \begin{tabular}{ccc} \subfloat[\# of CS permutations for LUBM]{\includegraphics[width=0.31\textwidth]{permutations_lubm_merge_2.png}} & \subfloat[\# of CS permutations for Geonames]{\includegraphics[width=0.31\textwidth]{permutations_geonames_merge_2.png}} & \subfloat[\# of CS permutations for Reactome]{\includegraphics[width=0.31\textwidth]{permutations_reactome_merge_2.png}} \\ \end{tabular} \end{tabularx} \caption{\# of CS permutations for increasing m}\label{fig:exp_fig_2} \end{figure*} \begin{table} \small \centering \caption{Loading experiments for all datasets} \resizebox{\columnwidth}{!}{ \begin{tabular}{|c|c c c c c |} \hline \textbf{Dataset} & \textbf{Size (MB)} & \textbf{Time} & \textbf{\# Tables (CSs)} & \textbf{\# of ECSs} & \textbf{Dense CS}\\ & & & & & Coverage\\ \hline Reactome Simple & 781 & 3min & 112 & 346 & 100\% \\ Reactome (m=0.05) & 675 & 4min & 35 & 252 & 97\% \\ Reactome (m=0.25) & 865 & 4min & 14 & 73 & 77\% \\ \hline Geonames Simple & 4991 & 69min & 851 & 12136 & 100\% \\ Geonames (m=0.0025) & 4999 & 70min & 82 & 2455 & 97\% \\ Geonames (m=0.05) & 5093 & 91min & 19 & 76 & 87\% \\ Geonames (m=0.1) & 5104 & 92min & 6 & 28 & 83\% \\ \hline LUBM Simple & 591 & 3min & 14 & 68 & 100\% \\ LUBM (m=0.25) & 610 & 3min & 6 & 21 & 90\% \\ LUBM (m=0.5) & 620 & 3min & 3 & 6 & 58\% \\ \hline WatDiv Simple & 4910 & 97min & 5667 & 802 & 100\% \\ WatDiv (m=0.01) & 5094 & 75min & 67 & 99 & 77\% \\ WatDiv (m=0.1) & 5250 & 75min & 25 & 23 & 63\% \\ WatDiv (m=0.5) & 5250 & 77min & 16 & 19 & 55\% \\ \hline \end{tabular} } \label{table:loading} \end{table} \textbf{Query Performance. } In order to assess the effect of the density factor parameter $m$ during query processing, we perform a series of experiments on LUBM, Reactome and Geonames. For the workload, we used the sets of queries from \cite{meimaris2017extended}. We employ two metrics, namely \textit{execution time} and \textit{number of table permutations}. The results can be seen in Figures \ref{fig:exp_fig_1} and \ref{fig:exp_fig_2}. As can be seen, hierarchical CS merging can help speed up query performance significantly as long as the dense coverage remains high. For example, in all datasets, query performance degrades dramatically when $m=1$, in which case the merging process cannot find any dense CSs. In this case, all rows are added to one large table, which makes the database only contain one table with many NULL cells. These findings are consistent across all three datasets and require further future work in order to identify the optimal value for $m$. In order to assess the performance of \textit{raxonDB} and establish that no overhead is imposed by the relational backbone, we performed a series of queries on LUBM2000, Geonames and Reactome, assuming the best merging of CSs is employed as captured by $m$ with respect to our previous findings. We also compared the query performance with rdf-3x, Virtuoso 7.1, TripleBit and the emergent schema approach described in \cite{pham2016exploiting}. The results can be seen in Figure \ref{fig:comp_with_old} and indicate that \textit{raxonDB} provides equal or better performance from the original \textit{axonDB} implementation, as well as the rest of the systems, including the emergent schema approach, which is the only direct competitor for merging CSs. Especially for queries with large intermediate results and low selectivity that correspond to a few CSs and ECSs (e.g. LUBM Q5 and Q6, Geonames Q5 and Q6) several of the other approaches fail to answer fast and in some cases time out. \begin{figure*}[t] \def\tabularxcolumn#1{m{#1}} \begin{tabularx}{\linewidth}{@{}cXX@{}} \begin{tabular}{ccc} \subfloat[Execution time (seconds) for LUBM2000]{\includegraphics[width=0.33\textwidth]{comp_with_old_2.png}} & \subfloat[Execution time (seconds) for Geonames]{\includegraphics[width=0.33\textwidth]{comp_with_old_geonames.png}} & \subfloat[Execution time (seconds) for Reactome]{\includegraphics[width=0.33\textwidth]{comp_with_old_reactome.png}} \\ \end{tabular} \end{tabularx} \caption{Query execution times in milliseconds for different RDF engines}\label{fig:comp_with_old} \end{figure*} \section{Conclusions and Future Work} In this paper, we tackled the problem of merging characteristic sets based on their hierarchical relationships. As future work, we will study computation of the optimal value for $m$, taking into consideration workload characteristics as well as a more refined cost model for the ancestral paths. Furthermore, we will study application of these findings in a distributed architecture, in order to further scale the capabilities of \textit{raxonDB}. \section{Introduction} Recent works in the state of the art in RDF data management have shown that extraction and exploitation of the implicit schema of the data can be beneficial in both storage and SPARQL query performance \cite{pham2015deriving}\cite{pham2016exploiting}\cite{meimaris2017extended}\cite{montoya2017odyssey}. In order to organize on disk, index and query triples efficiently, these trends heavily rely on two structural components of an RDF dataset, namely (i) the notion of \textit{characteristic sets }(CS), i.e., different property sets that characterize subject nodes, and (ii) the join links between CSs. For the latter, in our previous work, we introduced \textit{Extended Characteristic Sets} (ECS)\cite{meimaris2017extended}, which are typed links between CSs that exist only when there are object-subject joins between their triples, and we showed how RDF data management can rely extensively on CSs and ECSs for both storage and indexing, yielding significant performance benefits in heavy SPARQL workloads. However, this approach failed to address schema heterogeneity in loosely-structured datasets, as this implied a large number of CSs and ECSs (e.g., Geonames contains 851 CSs and 12136 CS links), and thus, skewed data distributions that impose large overheads in the extraction, storage and disk-based retrieval\cite{pham2015deriving}\cite{meimaris2017extended}. In this paper, we exploit the hierarchical relationships between CSs, as captured by subsumption of their respective property sets, in order to merge related CSs. We follow a relational implementation approach by storing all triples corresponding to a set of merged CSs into a separate relational table and by executing queries through a SPARQL to SQL transformation. Although, alternative storage technologies can be considered (key-value, graph stores,etc), we have selected well-established technologies and database systems for the implementation of our approach, in order to take advantage of existing data indexing and query processing techniques that have been proven to scale efficiently in large and complex datasets. To this end, we present a novel system, named \textit{raxonDB}, that exploits these hierarchies in order to merge together hierarchically related CSs and decrease the number of CSs and the links between them, resulting in a more compact schema with better data distribution. The resulting system, built on top of PostgreSQL, provides significant performance improvements in both storage and query performance of RDF data. In short, our contributions are as follows: \begin{itemize} \item We introduce a novel CS merging algorithm that takes advantage of CS hierarchies, \item we implement \textit{raxonDB}, an RDF engine built on top of a relational backbone that takes advantage of this merging for both storing and query processing, \item we perform an experimental evaluation that indicates significant performance improvements for various parameter configurations. \end{itemize} \section{Distance-Based Reordering} \subsection{Query Representation} Let $T$ be the set of all triple patterns in an incoming query $q$. Each triple pattern $t_i \in T$ consists of three nodes, either bound (URIs, blank node IDs or literals) or unbound (variable), i.e., the subject, predicate and object of $t_i$. Furthermore, let $M$ represent the set of all unique nodes in the subject and object positions in $T$, both bound and unbound. We map each triple pattern $t_i$ from $T$ to a $\left\vert{M}\right\vert$-dimensional vector space, where each vector attribute is an element from $M$, i.e., a subject or object node from the query. In this preliminary approach, we assume a solution space of left-deep trees, where the original problem of finding the best order of triple evaluation is factorial. This is because for a query $q$ consisting of $\left\vert{T}\right\vert$ triple patterns, there are $\left\vert{T}\right\vert!$ different permutations of triples. The factorial problem, commonly solved by dynamic programming algorithms, guarantees that the solution will be optimal with respect to the chosen cost model. Unfortunately, recent studies have suggested that even industrial-level query optimizers do not yield satisfactory results when they are heavily dependent on the cost model \cite{leis2015good}, especially when cost model errors propagate through multiple joins. In our approach, we map triple patterns from a query into a multidimensional space, and decide the join order based on the spatial correlations of the patterns, i.e., by comparing the pair-wise distances of the vectors that represent the triple patterns in the multiple dimensions, and ranking the patterns based on this comparison. The multidimensional space is built based on cardinality statistics that are held in a separate index. These pre-computed statistics are commonly found in RDF stores, and include the cardinality of triples with a bound subject, property, and object, as well as the size of the whole dataset. Specifically, we assume that each triple pattern becomes a $m$-dimensional vector, and the value of an attribute $m_i$ for a given triple pattern $t = (s,p,o)$, is given by the following function: \[ f(t,m_i) = \begin{cases} card(t),& \text{if } m_i \in (s,o)\\ 0, & \text{otherwise} \end{cases} \] Where $card(t)$ is the cardinality of the triple pattern. This is calculated using the following rules: \begin{itemize} \item{ $card(t) = card(p)$, if $p$ is bound and $s$,$o$ are unbound,} \item{ $card(t) = max(1,\frac{card(p)}{\left\vert{S}\right\vert})$, if $p$ and $o$ are bound, and $s$ is unbound,} \item{ $card(t) = 1$, if $p$ and $s$ are bound, and $o$ is either bound or unbound,} \item{ $card(t) = \left\vert{D}\right\vert$, for all other cases} \end{itemize} The cardinality for a bound property $card(p)$ is simply given as the number of triples with $p$ as the predicate. $\left\vert{D}\right\vert$ and $\left\vert{S}\right\vert$ are the number of triples in the dataset $D$, and the number of distinct subject nodes $S$ respectively. Notice that we do not represent predicate attributes, which means that the query space will be built entirely based on the subject and object nodes of all triple patterns in the query. Instead, the information that comes from the predicates in the triple patterns becomes quantized in the values of the attributes, as is given by the aforementioned cardinality estimation rules. Especially for the \textit{rdf:type} property, we assume the existence of an aggregate index that holds the exact cardinalities of particular class types. Therefore, in this special case, the second rule is altered to reflect the exact enumeration of subjects with a specific type. For instance, in our running example, the cardinality of $t_1$ is 99k, as there exist 99k subjects with \textit{rdf:type} \textit{ub:Student}. Function $f(t,m)$ can be used to calculate the values of a $\left\vert{T}\right\vert\times\left\vert{M}\right\vert$ matrix, denoted as $Q_m$, where each row represents a triple pattern in $T$, and each column represents an attribute in $M$, or simply a node in the query pattern. By applying a distance function and processing the pair-wise distances between rows in $Q_m$, the search space becomes quadratic with respect to $\left\vert{T}\right\vert$, and the comparisons will be of $O(\left\vert{T}\right\vert^2)$ complexity, rather than the original factorial one. Moreover, as we do not evaluate self-distances and permuations of the same pair (e.g., $[t_1,t_2]$ and $[t_2,t_1]$, the actual number of comparisons will be $\binom{\left\vert{T}\right\vert}{2}$, or $\left\vert{T}\right\vert \times (\left\vert{T}\right\vert -1)/2$. Even though this reduction in cost comes with the loss of guarantee of optimality, our experiments show that this approach tends to work well for all types of queries, even with large numbers of triple patterns. The matrix $Q_m$ for the query of Listing \ref{lst:sparql}, is shown in Table 1 for the synthetic dataset LUBM10 with 1.5m triples. To construct this for dataset $D$, we apply the estimation rules based on the per-computed statistics. For instance, there are 99k triples with \textit{rdf:type} as a property and \textit{ub:Student} as an object, which can be seen in columns 1 and 2 for $t_1$. Similarly, there are 106k triples with \textit{ub:emailAddress} as a predicate, which is encoded in $t_5$. Application of a distance function will generate a $\left\vert{5}\right\vert\times\left\vert{5}\right\vert$ distance matrix of pair-wise distances between the triple patterns $t_1-t_5$. \subsection{Plan Generation} When $Q_m$ is constructed and the distance matrix is calculated, we build sub-plans for the query, based on the ascending ranking of the pair-wise distances of the triple patterns in $T$. The algorithm to build sub-plans can be seen in Algorithm \ref{alg:generateSubPlans}. Specifically, we sort pair-wise distances in ascending order and create an empty queue for sub-plans (Lines 1-3). Then, we iterate through each pair of sorted triple patterns $t_a, t_b$ (Line 4). If there exists a sub-plan $p_i$ that contains only $t_a$ or only $t_b$, we append the non-contained triple pattern to $p_i$ (Lines 5-10). If none of $t_a, t_b$ are contained in a sub-plan, we create a new sub-plan and add $t_a, t_b$, then push the new sub-plan to the existing queue. The order in which we add these two depends on the cost of each triple pattern (Lines 11-15). In case both patterns $t_a$ and $t_b$ exist in sub-plans, we continue to the next pair. After we have created a series of (ordered) sub-plans, we iterate through consecutive sub-plans and attempt to reorder their patterns in order to better capture join relationships that occur between these. More accurately, if there are more than one sub-plans found, then we try to rearrange the triples that share common variables between adjacent sub-plans, so that the joined triple patterns between the two sub-plans are closely located. This can be seen in Algorithm \ref{alg:reorderSubPlans}. Specifically, if there is only one sub-plan, then this is returned as the final plan (Lines 1-3). Else, the algorithm iterates through the $subPlans$ queue with two pointers (Line 4), and prioritizes each pair of sub-plans (Line 5). The $prioritizePair$ function checks for joins between two consecutive sub-plans (Line 12), i.e. it checks whether the last triple (tail) in $p_i$ is joined with the first triple (head) in $p_{i+1}$. If they are joined, it returns the pair as is (Lines 13-15). If not, it finds the first triple in $p_{i+1}$ that can be joined with the tail of $p_i$ (if one exists), pushes it at the tail of $p_i$ and recursively checks the same pair (Lines 17-21). Finally, the $prioritizePair$ function returns an array of two plans, the first of which is inserted in the final plan, while the second is used as the first sub-plan of the next iteration, as it holds the updated ordering. This means that the triple patterns that are ranked lower in $p_{i+1}$ will be bubbled-up and re-ranked in order to reflect the join relationship between $p_i$ and $p_{i+1}$. \begin{algorithm}[ht] \caption{\emph{generateSubPlans}}\label{alg:generateSubPlans} \textbf{Input: $map$:} A matrix of triple patterns from the query as rows, and query nodes as columns\\ \textbf{Output: $subPlans$:} An ordered list of sub-sets of query triple patterns, that constitute subplans. \begin{algorithmic}[1] \State $subPlans \leftarrow new Queue()$ \State $distances \leftarrow distanceMatrix(map)$ \State sort($distances$) \For{\textbf{each} $pair$ $t_i, t_j \in distances \mid t_i \neq t_j$} \If {$\exists p_i \in subPlans \mid t_a \in p_i$ AND $t_b \not\in p_i$} \State append $t_b$ to $p_i$ \State continue \ElsIf {$\exists p_i \in subPlans \mid t_a \not\in p_i$ AND $t_b \in p_i$} \State append $t_a$ to $p_i$ \State continue \ElsIf {$\not\exists p_i \in subPlans \mid t_a \in p_i$ OR $t_b \in p_i$} \State $p \leftarrow new List()$ \State append minCost($t_a, t_b)$ to $p$ \State append maxCost($t_a, t_b)$ to $p$ \State $subPlans$.push($p$) \Else \State continue \\ \EndIf \EndFor \Return $subPlans$ \end{algorithmic} \end{algorithm} \begin{algorithm}[ht] \caption{\emph{reorderSubPlans}}\label{alg:reorderSubPlans} \textbf{Input: $subPlans(p_0,\dots,p_{n-1})$:} A queue of sub-plans \\ \textbf{Output: $finalPlan$:} An ordered list of triple patterns, to be executed by the query engine. \begin{algorithmic}[1] \If {$subPlans$.size $== 1$} \State $finalPlan$.push($p_0$) \State \Return $finalPlan$ \EndIf \For{\textbf{each} $p_i, p_{i+1} \in subPlans$} \State $nextPair \leftarrow $ prioritizePair($p_i, p_{i+1}$) \State $finalPlan$.push($nextPair$[0]) \State $p_{i+1} \leftarrow nextPair$[1] \If {$p_{i+1}$.isEmpty} \State $subPlans.remove(p_{i+1})$ \EndIf \EndFor \Return $finalPlan$ \Procedure{prioritizePair}{$p_i, p_{i+1}$} \State $newPair \leftarrow array[2]$ \If {$p_i.tail$ is joined with $p_{i+1}.head$} \State $newPair$[0]$\leftarrow p_i$ \State $newPair$[1]$\leftarrow p_{i+1}$ \State \Return $newPair$ \Else \For{\textbf{each} $t_k, \in p_{i+1}$} \If {$p_i.tail$ is joined with $t_k$} \State $p_i.push(t_k)$ \State $p_{i+1}.remove(t_k)$ \State $newPair \leftarrow $prioritizePair($p_i, p_{i+1}$) \EndIf \EndFor \State $newPair$[0]$\leftarrow p_i$ \State $newPair$[1]$\leftarrow p_{i+1}$ \Return $newPair$ \EndIf \EndProcedure \end{algorithmic} \end{algorithm} \subsection{Experiments} We implemented a proof-of-concept version of the query planner in Jena ARQ, and conducted experiments on Jena TDB\footnote{https://jena.apache.org/documentation/tdb/}, comparing against several statistical reordering approaches, namely Stocker et al's PFJ and ONS \cite{stocker2008sparql}, Kalayci et al's Ant System \cite{kalayci2015ant} as well as Jena TDB's Fixed (JF) and Weighted (JW) optimizers, and the reordering performed by the open source edition of Openlink Virtuoso 7.2. As triple pattern ordering is a problem that is orthogonal to the low-level implementation specifics and design choices (e.g., join implementations, indexing), we use Jena TDB as a common testbed for all approaches to strictly assess and compare the effect of triple pattern orderings on query processing. We used LUBM to generate a synthetic dataset of 15m triples, and measured execution times for 29 queries on the LUBM dataset. The queries consist of 14 original queries provided by LUBM, and an additional 15 queries of increasingly complex shapes and sizes, used in \cite{kalayci2015ant}. All code and queries are available online\footnote{https://github.com/mmeimaris/sparqlDistancePlanner}. For each approach, we extract an ordering, and feed that directly to the Jena query processor. Table 2 summarizes the percentages of best plans for all queries. These are computed by measuring, for each approach, how many plans were the fastest\footnote{We measure execution times by executing each query 10 times and reporting the mean.} in relation to the other approaches, for all queries. Because of the fact that the fastest plan can be reached by more than one method, we differentiate between plans based on the generated orderings, and then we compare their execution times. As can be seen, our method outperforms all other methods for queries of different patterns, namely the original 14 queries, and the 15 additional queries of star, chain, chain-star, and cyclic patterns, achieving the best plan 90\% of the time, with Virtuoso coming second with 66\% of its orderings being the best. In order to assess how our planner performs with respect to different query types, we use 15 extended queries originally constructed in \cite{kalayci2015ant}. These represent different query types, and are thus classified into star, chain, cyclic, and chain-star query types, of varying triple pattern sizes. Their original intention was to provide more complex query structures of up to 14 triple patterns for LUBM dataset, because the original queries are limited to 1-6 triple patterns. We measure execution times for each query type separately, and report these in Tables 3, 4, 5, and 6, for star, chain, cyclic, and chain-star queries respectively. For each query, the best ordering is marked with a bold execution time. While we do not show the actual triple orders that were derived by each method, multiple bold values in the same row indicate that the same order has been achieved by more than one approach. Note that small differences in execution times for the same orderings can be attributed to lags imposed by I/O, available CPU resources and other external factors. For computing the distance matrix from $Q_m$, we used a simple Euclidean distance function that measuring distances between triple patterns in the m-dimensional space. Overall, the results are encouraging for further pursuing this direction. In fact, in 90\% of the queries, our method generated the fastest plan in relation to the other approaches, of whom the best (VIRT) achieved a rate of 66\%, and the rest achieved less than 60\%. \begin{table} \centering \caption{$Q_m$ matrix for reference query.} \begin{tabular}{|c|c c c c c c|} \hline & ?X & Student & ?Y & Dpt & Univ0 & ?Z\\ \hline $t_1$ & 99k & 99k & 0 & 0 & 0 & 0\\ $t_2$ & 0 & 0 & 189 & 189 &0 & 0 \\ $t_3$ & 106k & 0 & 106k & 0 & 0 & 0 \\ $t_4$ & 0 & 0 & 239 & 0 & 239 & 0 \\ $t_5$ & 106k & 0 & 0 & 0 & 0 & 106k \\ \hline \end{tabular} \end{table} \begin{table} \scriptsize \setlength{\tabcolsep}{8pt} \renewcommand{\arraystretch}{1.8} \caption{Percentage of plans that are best compared to other methods for all queries.} \begin{tabular}{|c c c c c c c |} \hline JW & JF & ONS & PFJ & ANT & VIRT & OUR\\ \hline 59\% & 48\% & 28\% & 59\% & 31\% & 66\% & 90\% \\ \hline \end{tabular} \end{table} \begin{table}[ht] \scriptsize \setlength{\tabcolsep}{6pt} \renewcommand{\arraystretch}{1.8} \caption{Query execution times (seconds) for the star queries.} \begin{tabular}{|c|c c c c c c c|} \hline & JW & JF & ONS & PFJ & ANT & VIRT & OUR\\ \hline Q1 & 1.01 & 1.14 & 1.07& \textbf{0.61} & 1.05 & 0.63 & \textbf{0.57} \\ Q2 & 0.91 & 0.92 & 4.71 & 0.76 & 0.79 & 0.74 & \textbf{0.67} \\ Q3 & 0.23 & 0.27 & 0.23 & 0.15 & 3.62 & 0.21 & \textbf{0.13} \\ Q4 & 0.17 & 0.08 & 0.07 & 0.07 & 0.42 & \textbf{0.05} & \textbf{0.03} \\ \hline \end{tabular} \end{table} \begin{table}[ht] \scriptsize \setlength{\tabcolsep}{6pt} \renewcommand{\arraystretch}{1.8} \caption{Query execution times (seconds) for the chain queries.} \begin{tabular}{|c|c c c c c c c |} \hline & JW & JF & ONS & PFJ & ANT & VIRT & OUR\\ \hline Q1 & 0.30 & 0.33 & 0.37 & \textbf{0.14} & 1.09 & 0.39 & \textbf{0.14} \\ Q2 & \textbf{0.49} & \textbf{0.56} & \textbf{0.52} & 0.77 & \textbf{0.50} & \textbf{0.48} & \textbf{0.48} \\ Q3 & 322 & 323 & 131 & 292 & 139 & \textbf{55} & \textbf{53} \\ Q4 & \textbf{54} & \textbf{53} & \textbf{53} & 163 & 209 & 182 & \textbf{49} \\ \hline \end{tabular} \end{table} \begin{table}[ht] \scriptsize \setlength{\tabcolsep}{8pt} \renewcommand{\arraystretch}{1.8} \caption{Query execution times (seconds) for the cyclic queries.} \begin{tabular}{|c|c c c c c c c |} \hline & JW & JF & ONS & PFJ & ANT & VIRT & OUR\\ \hline Q1 & 0.05 & 0.07 & 0.05 & \textbf{0.04} & \textbf{0.04} & \textbf{0.04} & \textbf{0.04}\\ Q2 & 0.25 & 0.24 & 0.19 & \textbf{0.02} & 0.20 & 0.18 & \textbf{0.02} \\ Q3 & \textbf{0.06} & \textbf{0.06} & 0.31 & 2.37 & 0.74 & \textbf{0.04} & 4.50 \\ Q4 & 1.06 & 0.94 & 1.30 & 0.64 & 0.81 & \textbf{0.04} & 0.52 \\ \hline \end{tabular} \end{table} \begin{table} \scriptsize \setlength{\tabcolsep}{8pt} \renewcommand{\arraystretch}{1.8} \caption{Query execution times (seconds) for the chain-star queries.} \begin{tabular}{|c|c c c c c c c |} \hline & JW & JF & ONS & PFJ & ANT & VIRT & OUR\\ \hline Q1 & \textbf{4.16} & \textbf{4.16} & \textbf{4.15} & 7.39 & \textbf{4.15} & 5.15 & \textbf{4.15} \\ Q2 & 5.95 & 5.92 & 5.93 & 5.83 & 6.67 & \textbf{0.49} & \textbf{0.42} \\ Q3 & 6.11 & 4.98 & 4.75 & 214.92 & 6.37 & \textbf{2.35} & \textbf{2.31} \\ \hline \end{tabular} \end{table} \section{Related Work} RDF data management systems generally follow three storage schemes, namely \textit{triples tables}, \textit{property tables}, and \textit{vertical partitioning}. A triples table has three columns, representing the subject, predicate and object (SPO) of an RDF triple. This technique replicates data in different orderings in order to facilitate sort-merge joins. RDF-3X \cite{neumann2010RDF} and Hexastore \cite{weiss2008hexastore} build tables on all six permutations of SPO. Built on a relational backbone, Virtuoso \cite{erling2010virtuoso} uses a 4-column table for quads, and a combination of full and partial indexes. These methods work well for queries with small numbers of joins, however, they degrade with increasing sizes, unbound variables and joins. \textit{Property Tables} places data in tables with columns corresponding to properties of the dataset, where each table identifies a specific resource type. Each row identifies a subject node and holds the value of each property. This technique has been implemented experimentally in Jena \cite{wilkinson2006jena} and DB2RDF \cite{bornea2013building}, and shows promising results when resource types and their properties are well-defined. However, this causes extra space overhead for null values in cases of sparse properties \cite{abadi2007scalable}. Also, it raises performance issues when handling complex queries with many joins, as the amounts of intermediate results increase \cite{janik2005brahms}. Vertical partitioning segments data in two-column tables. Each table corresponds to a property, and each row to a subject node \cite{abadi2007scalable}. This provides great performance for queries with bound objects, but suffers when the table sizes have large variations in size \cite{sidirourgos2008column}. TripleBit \cite{yuan2013triplebit} broadly falls under vertical partitioning. In TripleBit, the data is vertically partitioned in chunks per predicate. While this reduces replication, it suffers from the same problems as property tables. It does not consider the inherent schema of the triples in order to speed up the evaluation of complex query patterns. In distributed settings, a growing body of literature exists, with systems such as Sempala \cite{schatzle2014sempala}, H2RDF \cite{papailiou2014h} and S2RDF \cite{schatzle2016s2rdf}. However, these are based on parallelization of centralized indexing and query evaluation schemes. For these reasons, latest state of the art approaches rely on implicit schema detection in order to derive a hidden schema from RDF data and index/store triples based on this schema. Furthremore, due to the tabular structure that tends to implicitly underly RDF data, recent works have been implemented in relational backbones. In our previous work \cite{meimaris2017extended}, we defined \textit{Extended Characteristic Sets (ECSs)} as typed links betwen CSs, and we showed how ECSs can be used to index triples and greatly improve query performance. In \cite{pham2015deriving}, the authors identify and merge CSs, similar to our approach, into what they call an \textit{emergent schema}. However, their main focus is to extract a human-readable schema with appropriate relation labelling. They do not use hierarchical information of CSs, rather they use semantics to drive the merging process. In \cite{pham2016exploiting} it is shown how this \textit{emergent schema} approach can assist query performance, however, the approach is limited by the constraints of human-readable schema discovery. In our work, query performance, indexing and storage optimization are the main aims of the merging process, and thus we are not concerned about providing human-readable schema information or any form of schema exploration. In \cite{montoya2017odyssey}, the authors use CSs and ECSs in order to assist cost estimation for federated queries, while in \cite{gubichev2014exploiting}, the authors use CSs in order to provide better triple reordering plans. To the best of our knowledge, this is the first work to exploit hierarchical CS relations in order to merge CSs and improve query performance. \section{Hierarchical CS Merging} \subsection{Preliminaries} The RDF model does not generally enforce structural rules in the representation of triples; within the same dataset there can be largely diverse sets of predicates emitting from nodes of the same semantic type \cite{meimaris2017extended,pham2015deriving,neumann2011characteristic}. \textit{Characteristic Sets} (CS)\cite{neumann2011characteristic} capture this diversity by representing implied node types based on the set of properties they emit. Formally, given a collection of triples $D$, and a node $s$, the characteristic set $cs(s)$ of $s$ is $cs(s) = \{p \mid \exists o: (s, p, o) \in D\}$. The set of properties of a CS $cs_i$ is denoted with $P_i$. Furthermore, in a given dataset, each CS represents a set of records identified by a subject node, and all of the values of the subject node (i.e., objects) for the predicates in $P_i$. We denote the set of all records of $cs_i$ as $r_i$, while $cs_i$ is represented by a relational table $c_i$ that is defined by these two elements, i.e., $c_i = (P_i, r_i)$. The tuples in $c_i$ are of the form $(s,p_{i,1},\dots, p_{i,k})$, where $s$ is the identifier column (e.g. URI) of a subject node and $p_{i,1}, p_{i,2}, \dots ,p_{i,k}$ are the values, i.e. object nodes, of the properties in $P_i$ for $s$. In the context of this paper, with the term \textit{Characteristic Set} we will refer collectively to the properties and records of a CS, i.e., its relational table, rather than just the set of properties proposed in the original definition, for the sake of simplicity. Within a given dataset, CSs often exhibit hierarchical relationships, as a result of the overlaps in their comprising sets of properties. For example, consider two CSs, $c_1, c_2$, describing human beings, with $P_{1} = \{type, name\}$ and $P_{2} = \{type, name, marriedTo\}$. It can be seen that $P_{1} \subset P_{2}$ and thus $c_1$ is a parent of $c_2$. This relationship entails an overlap of properties that define the CSs, and can be exploited in order to provide a means to merge common CSs based on the specialization or generalization of the node types they describe. In what follows, we formally define the notions of CS \textit{subsumption}, \textit{hierarchy} and \textit{ancestral sub-graphs}. \textit{Definition 1}. \textbf{\textit{(CS Subsumption)}}. Given two CSs, $c_i$ and $c_j$, and their property sets $P_i$ and $P_j$, then $c_i$ subsumes $c_j$, or $c_i \succ c_j$, when the property set of $c_i$ is a proper subset of the property set of $c_j$, or $P_i \subset P_j$. This subsumption forms parent-child relationships between CSs. CS subsumption relationships can be seen in Figure \ref{fig:combinations}(a) as directed edges between nodes. The set of all parent-child relationships defines a CS hierarchy as defined in the following. \textit{Definition 2}. \textbf{\textit{(CS Hierarchy and Inferred Hierarchy)}}. CS subsumption creates a partial ordering that essentially defines a \textit{hierarchy} such that when $c_i \succ c_{j}$, then $c_{i}$ is a parent of $c_{j}$. Formally, a CS hierarchy is a graph lattice $L = (V, E)$ where $V \in C$ and $E \in (V \times V)$. A directed edge between two CS nodes $c_1, c_2$ exists in $L$, when $c_1 \succ c_2 $ and there exists no other $c_i$ such that $c_1 \succ c_i \succ c_2$. An example CS hierarchy can be seen in Figure \ref{fig:combinations}(a). Given a hierarchy $L$, we denote the \textit{hierarchical closure} of $L$ with $L_{c}$, so that $L_{c}$ extends $L$ to contain inferred edges between hierarchically related nodes that are not consecutive, e.g. a node and its grandchildren. An example inferred hierarchy can be seen in Figure \ref{fig:combinations}(c) for a sub-graph of the graph in Figure \ref{fig:combinations}(a), with the inferred relationships in dashed lines. In the remainder of this paper, we refer to $L_{c}$ as the \textit{inferred hierarchy} of $L$ \textit{Definition 3}. \textbf{\textit{(CS Ancestral Sub-graphs)}}. Given an inferred hierarchy $L_{c} = (V, E)$, a CS $c_{base}$ and set of CSs $c_1, \dots, c_k$, then $a = (V^{'}, E^{'})$ is an ancestral sub-graph with $c_{base}$ as the lowermost child when $\forall i \in [1..k]$, it holds that $c_{i} \succ c_{base}$, and $(c_{i}, c_{base}) \in E^{'}$. This means that any sub-graph with $c_{base}$ as a sink node will be an ancestral sub-graph of $c_{base}$. Thus, it holds that $a \subset L_c$. For instance, in Figure \ref{fig:combinations}(c), nodes $c_7, c_4, c_2$ form an ancestral sub-graph with $c_7$ as the base CS. Similarly, nodes $c_6, c_4, c_2$ and $c_6, c_5, c_2$ form ancestral sub-graphs with $c_6$ as base CS. Logically, we map each CS to a relational table, so that for a CS $c_i$ we create a relational table $t_i = (s, p_{i,1}, p_{i,2}, \dots, p_{i,k})$, where $s$ is the id of the subject and $p_{i,1}\dots, p_{i,k}$ are the properties that belong to $P_i$, and then we use the CS hierarchy in order to merge the nodes of an ancestral sub-graph with $c_i$ as base into a single table. Specifically, we exploit the property set overlap in order to merge together smaller parent CSs with larger child CSs, in order to minimize the effect of NULL values that will appear for properties in smaller CSs that do no exist in the larger CSs. Thus, $c_{base}$ will be the most specialized CS in its ancestral sub-graph. For this reason, we define a merge operator, $hier\_merge$, as follows. \textit{Definition 4}. \textbf{\textit{(Hierarchical CS Merge)}}. Given an ancestral sub-graph $a = (V^{'}, E^{'})$, where $V^{'} = \{c_1 = (P_1, r_1), c_2 = (P_2, r_2),\dots, c_k=(P_k, r_k)\}$ as defined above, then a hierarchical merge of $a$ is given as follows: $hier\_merge(a) = c_{a}$, where $c_a = (P_1, r_a)$. Here, $P_1$ is the most specialized property set in $a$, as $c_1$ does not have any children in $a$, while $r_a = \bigcup\limits_{i=1}^{k}r_i^{'}$ is the UNION of the records of all CSs in $V^{'}$, where $r_i^{'}$ is the projection of $r_i$ on $P_1$. This means that $r_i^{'}$ will contain NULL values for all the non-shared properties of $P_1$ and $P_i$, i.e., $P_1 \setminus P_i$. In essence, $hier\_merge$ is an \textit{edge contraction} operator that merges all nodes of an ancestral sub-graph into one, while removing the edges that connect them. For instance, assume that $V^{'} = \{c_0 = (P_0, r_0), c_1 = (P_1, r_1), c_2 = (P_2, r_2)\}$ is the set of vertices of an ancestral sub-graph with three CSs, with $P_0 = \{p_a,p_b\}$, $P_1 = \{p_a, p_b, p_c\}$ and $P_2 = \{p_a, p_b, p_c, p_d\}$. Thus, $c_0 \succ c_1 \succ c_2$. Hierarchical merging can be seen in Figure \ref{fig:table_merging}. \textit{Definition 5}. \textbf{\textit{(Merge Graph)}}. Given an inferred CS hierarchy $L_{c} = (V,E)$, a merge graph is a graph $L^{'}=(V^{'}, E^{'})$ that consists of a set of $n$ ancestral sub-graphs, and has the following properties: (i) $L^{'}$ contains all nodes in $L$ such that $V^{'} \equiv V$, i.e., it covers all CSs in the input dataset, (ii) $L^{'}$ contains a subset of the edges in $L$ such that $E^{'} \subset E$, (iii) each node is contained in exactly one ancestral sub-graph $a_i$, (iv) all ancestral sub-graphs are pair-wise disconnected, i.e., there exist no edges between the nodes of different ancestral sub-graphs. Thus, each ancestral sub-graph can be contracted into one node unambiguously, using the $hier\_merge$ operator. Also, the total number of relational tables will be equal to the number of ancestral sub-graphs in the merge graph. \begin{figure}[t] \centering \includegraphics[width=1\textwidth]{example_graph.png} \caption{(a) A CS hierarchy graph with dense nodes colored in deep purple, (b) the connected components derived by cutting off descendants from dense nodes, (c) a connected component with dashed lines representing inferred hierarchical relationships, (d) all possible assignments of dense nodes to non-dense nodes.}\label{fig:combinations} \end{figure} \textbf{Problem Formulation. }Given an inferred CS hierarchy $L_{c} = (V, E)$, the problem is to find a merge graph $L^{'} = (V, E^{'})$ in the form of a set of disconnected ancestral sub-graphs, that provides an optimal way to merge CS nodes. In other words, the problem is to find the best set of ancestral sub-graphs from an inferred hierarchy $L_{c}$ that minimize an objective cost function $cost(x)$, or more formally: \begin{align} L^{'} = argmin_{x \subset L_c}cost(x) \end{align} This formulation entails several problems. First, the notion of cost depends on possibly subjective factors, such as the query workload, the storage technology, the input dataset and so on. There is no universal cost model that can be deployed in order to assess the effectiveness of a merge graph. Moreover, neither the number of ancestral sub-graphs, nor the set of sub-graph roots is known as part of the input. A CS hierarchy of $n$ nodes can potentially create $2^n$ sub-graphs, while the number of possible sub-graph roots is also exponential with the respect to the hierarchy size. Thus, given an arbitrary cost function, this is a problem of \textit{non-uniform graph partitioning} on the inferred hierarchy $L_c$, which is known to be NP-Hard. That is, even with a deployed cost model, it is still an exponential problem to enumerate all possible sub-graphs and find the one with the minimum cost. For these reasons, we approach the problem by deploying a set of rules and heuristics that find a good merge graph efficiently and offer improved storage and query performance, as will be shown in the experiments. \subsection{CS Retrieval and Merging} The primary focus of this work is to improve the efficiency of the storage and query capabilities of relational RDF engines by exploiting the implicit schema of the data in the form of CSs. However, CS merging results in several problems that need to be addressed in this context. These are discussed in what follows. First, the problem of selecting ancestral sub-graphs is a computationally hard one, as mentioned earlier. For this reason, we rely on a simple heuristic in order to seed the process and provide an initial set of ancestral sub-graph \textit{sink nodes}, that will form the bases of the final merged tables, as defined in Definition 3. For this, we identify \textit{dense} CS nodes in the hierarchy (i.e, with large cardinalities) and use these nodes as the bases of the ancestral sub-graphs. While node density can be defined in many different ways, in the context of this work we define a $c_i$ to be dense, if its cardinality is larger than a linear function of the maximum cardinality of CSs in $D$, i.e., a function $d: N \rightarrow R$, with $d(c_i) = m \times |r_{max}|$. Here, $m \in \left[0,1\right]$ is called the \textit{density factor}, and $r_{max}$ is the cardinality of the largest CS in $D$. This means that, by definition, if $m = 0$, no CSs will be merged, because all CSs will be considered dense and thus each CS will define its own ancestral sub-graph, while if $m = 1$, all no ancestral sub-graph will be defined, and all CSs will be merged to one large table, as no CS has a cardinality larger than that of the largest CS. With a given $m$, the problem is reduced to finding the optimal ancestral sub-graph for each given dense node. Second, merging tables results in the introduction of NULL values for the non-shared columns, which can degrade performance. Specifically, merging CSs with different attribute sets can result in large numbers of NULL values in the resulting table. Given a parent CS $c_1 = (P_1, r_1)$ and a child CS $c_2 = (P_2, r_2)$ with $\left\vert{P_1}\right\vert < \left\vert{P_2}\right\vert$ and $\left\vert{r_1}\right\vert >> \left\vert{r_2}\right\vert$, the resulting $\left\vert{P_2 \setminus P_1}\right\vert \times \left\vert{r_1}\right\vert$ NULL cells will be significantly large compared to the total number of $r_1+r_2$ records, thus potentially causing poor storage and querying performance\cite{pham2015deriving}. For this reason, CS merging must be performed in a way that will minimize the presence of NULL values. The following function captures the NULL-value effect of the merge of two CSs $c_i = (P_i, r_i), c_j = (P_j,r_j)$ with $c_i \succ c_j$: \begin{align} r_{null}(c_i,c_j) = \frac{|P_j \setminus P_i| \times |r_i|}{(|r_j|)} \end{align} Intuitively, $r_{null}$ represents the ratio of null values to the cardinality of the base CS in the merge. The numerator of the fraction represents the total number of cell values that will be null, as the product of the number of non-shared properties and the cardinality of the parent CS. The denominator represents the cardinality of the base CS. Hence, the base CS must be a descendant (i.e., with more properties) in order to minimize the presence of NULLs. In order to assess an ancestral sub-graph, we use a generalized version of $r_{null}$ that captures the NULL value effect on the whole sub-graph: \begin{align} r^{g}_{null}(g)|_{c_{d}} = \frac{\sum_{i=1}^{|g|}{|P_{d} \setminus P_i| \times |r_i|}}{|r_d|+\sum_{i=1}^{|g|}(|r_i|)} \end{align} Here, $c_d = (P_d, r_d)$ is the dense root of sub-graph $g$. However, merging a parent to a dense child changes the structure of the input graph, as the cardinality of the dense node is increased. To accommodate this, we define a cost function that works on the graph level, as follows: \begin{align} cost(g) = \sum_{i=1}^{n}{r^{g}_{null}(g_i)|_{c_{di}}} \end{align} where $n$ is the number of dense nodes, $c_{di}$ is a dense node and $g_i$ is the ancestral sub-graph with $c_{di}$ as the base node. Given this cost model and a pre-defined set of dense nodes, our exact algorithm will find the optimal sub-graph for each dense node. An inferred hierarchy graph can be converted to a set of connected components that are derived by removing the outgoing edges from dense nodes, since we are not interested in merging children to parents, but only parents to children. An example of this can be seen in Figure \ref{fig:combinations}(b). For each component, we can compute $cost(g)$ as the sum of the costs of these components. The main idea is to identify all connected components in the CS graph, iterate through these components, enumerate all sub-graphs within the components that start from the given set of dense nodes, and select the optimal partitioning for each component. The algorithm can be seen in Algorithm \ref{alg:optimalMerge}. The algorithm works by first identifying all connected components of the inferred hierarchy (Line 2). Identifying connected components is trivially done using standard DFS traversal, and is not shown in the Algorithm. Then, we iterate each component (Line 3), and for each component, we generate all possible sub-graphs. Then, we calculate the cost of each sub-graph (Line 7) and if it is smaller than the current minimum, the minimum cost and best sub-graph are updated (Lines 8-9). Finally, we add the best sub-graph to the final list (Line 11) and move to the next component. \begin{figure}[t] \centering \includegraphics[width=0.4\textwidth]{table_merging_3.png} \caption{Merging the tables of $c_0$, $c_1$ and $c_2$.}\label{fig:table_merging} \end{figure} To generate the sub-graphs, we do not need to do an exhaustive generation of $2^n$ combinations, but we can rely on the observation that each non-dense node must be merged to exactly one dense node. Therefore, sub-graph generation is reduced to finding all possible assignments of dense nodes to the non-dense nodes. An example of this can be seen in figure \ref{fig:combinations}. In the figure, nodes $c_2, c_4, c_5$ are non-dense, while nodes $c_6, c_7, c_8$ are dense. All possible and meaningful sub-graphs are enumerated in the table at the right of the figure, where we assign a dense node to each of the non-dense nodes. An assignment is only possible if there exists a parent-child relationship between a non-dense node and a dense node, even if it is an inferred one (e.g. $c_2$ is an inferred parent of $c_7$). Hence, the problem of sub-graph generation becomes one of generating combinations from different lists, by selecting one element from each list. The number of lists is equal to the number of non-dense nodes, and the elements of each list are the dense nodes that are related to the non-dense node. \textbf{Complexity Analysis. } Assuming that a connected component $g$ has $k$ non-dense nodes and $d$ dense nodes, and each non-dense node $k_i$ is related to $e(k_i)$ dense nodes, then the number of sub-graphs that need to be enumerated are $\prod_{i=1}^{k}{e(k_i)}$. In the example of figure \ref{fig:combinations}, the total number of sub-graphs is $e(c_2) \times e(c_4) \times e(c_5) = 3 \times 2 \times 1 = 6$. In the worst case all $k$ nodes are parents of all $d$ nodes. Then, the number of total sub-graphs is $k^d$, which makes the asymptotic time complexity of the algorithm $O(k^d)$. \begin{algorithm} \KwData{An inferred hierarchy lattice $L_c$ as a adjacency list , and a set of dense CSs $D$} \KwResult{A set of optimal ancestral sub-graphs} init $finalList$\; $connectedComponents \gets findConnectedComponents(L_c)$\; \For {\textbf{each} $connectedComponent$}{ init $min \gets MAX\_VALUE$\; init $bestSubgraph$ \; \While {$next \gets connectedComponent.generateNextSubgraph()$} { \uIf{$cost(next) < min$}{ $min \gets cost(next)$\; $bestSubgraph \gets next$\; } } $finalList.add(bestSubgraph)$\; } return $finalList$\; \caption{\emph{optimalMerge}}\label{alg:optimalMerge} \end{algorithm} \subsection{Greedy Approximation} For very small $d$ (e.g. $d < 4$), the asymptotic complexity of $O(k^d)$ is acceptable. However, in real-world cases, the number of connected components can be small, making $d$ large. For this reason, we introduce a heuristic algorithm for approximating the problem, that does not need to enumerate all possible combinations, but instead relies on a greedy objective function that attempts to find the local minimum with respect to our defined cost model for each non-dense node. Note that it lies beyond the scope of this work to compute the degree of approximation to the optimal solution, however, in our experiments, the heuristic solution is shown to provide significant performance gains. The main idea behind the algorithm is to iterate the non-dense nodes, and for each non-dense node, calculate the $r_{null}$ function and find the dense node that minimizes this function for the given non-dense node. Then, the cardinalities will be recomputed and the next non-dense node will be examined. The algorithm can be seen in Algorithm \ref{alg:greedyMerge}. In the beginning, the algorithm initiates a hash table, $mergeMap$, with an empty list for each dense node (Lines 1-4). Then, the algorithm iterates all non-dense nodes (Line 5), and for each dense node, it calculates the cost $r_{null}$ of merging it to each of its connected dense nodes (Lines 5-13), keeping the current minimum cost and dense node. In the end, the current non-dense node is added to the list of the dense node that minimizes $r_{null}$ (Line 14). Notice that we do not need to split the hierarchy into connected components in order for $greedyMerge$ to work. \textbf{Complexity Analysis. }Given $k$ non-dense nodes and $d$ dense nodes, where each non-dense node $k_i$ is related to $e(k_i)$ dense nodes, the $greedyMerge$ algorithm needs $\sum_{i=1}^{k}{e(k_i)}$ iterations, because we need to iterate all $e(k_i)$ nodes for each $k_i$. In the worst case, every $k_i$ is related to all $d$ dense nodes, requiring $kd$ iterations. Assuming a constant cost for the computation of $r_{null}$, then the asymptotic complexity of the greedy algorithm is $O(kd)$, which is a significant performance when compared to the exponential complexity of $optimalMerge$. \begin{algorithm} \KwData{A hash table $p$ mapping non-dense CSs to their dense descendants, a set of dense CSs $D$, and a set of non-dense CSs $K$} \KwResult{A hash table mapping dense CSs to sets of non-dense CSs to be merged} init $mergeMap$\; \For {\textbf{each} $d \in D$}{ $mergeMap.put(d, new List())$\; } \For {\textbf{each} $k \in K$}{ $min \gets MAX\_VALUE$\; init $bestDense$\; \For {\textbf{each} $d_k \in p.get(k)$}{ $cost \gets r_{null}(k, d_k)$\; \uIf{$cost < min$}{ $min \gets cost$\; $bestDense \gets d_k$\; } } $mergeMap.get(bestDense).add(k)$\; } return $mergeMap$\; \caption{\emph{greedyMerge}}\label{alg:greedyMerge} \end{algorithm} Obviously, this process does not necessarily cover all CSs of the input dataset. The percentage of the dataset that is covered by this process is called \textit{dense CS coverage}. The remainder of the CSs that are not contained by any merge path are aggregated into one large table containing all of their predicates. If the total coverage of the merging process is large, then this large table does not impose a heavy overhead in query performance, as will be shown in the experiments. Finally, we load the data in the corresponding tables. \begin{figure}[ht] \centering \includegraphics[width=1\textwidth]{greedy_merging.png} \caption{An example of greedy merging. Dense nodes are coloured in deep purple. At each step, the non-dense node under examination is coloured with green, while the edge that minimizes $r_{null}$ can be seen in bold.}\label{fig:path_pruning} \end{figure} \subsection{Implementation} We implemented \textit{raxonDB} as a storage and querying engine that supports hierarchical CS merging, and can be deployed on top of standard RDBMS solutions. Specifically, we used PostgreSQL 9.6, but \textit{raxonDB} can be adapted for other relational databases as well. The architecture of \textit{raxonDB} can be seen in Figure \ref{fig:arch}. \textbf{CS Retrieval and Merging. }The processes of retrieving and merging CSs take place during the loading stage of an incoming RDF dataset. CS retrieval is a trivial procedure that requires scanning the whole dataset and storing the unique sets of properties that are emitted from the subject nodes in the incoming triples, and is adopted from our previous work in \cite{meimaris2017extended} where it is described in detail. After retrieving the CSs, the main idea is to compute the inferred CS hierarchy and apply one of the described merging algorithms. Finally, each set of merged CSs is stored in a relational table. In each table, the first column represents the subject identifier, while the rest of the columns represent the union of the property sets of the merged CSs. For multi-valued properties, we use PostgreSQL's array data type in order to avoid duplication of the rows. \textbf{Indexing. }We deploy several indexes in \textit{raxonDB}. First off, we index the subject id for each row. We also build foreign-key indexes on object-subject links between rows in different CSs, i.e., when a value of a property in one CS is the subject id of another CS. Next, we use standard B+tree for indexing single-valued property columns, while we use PostgreSQL's GIN indexes, which apply to array datatypes for indexing multi-valued properties. This enables fast access to CS chain queries, i.e., queries that apply successive joins for object-subject relationships. Furthermore, we store these links on the schema level as well, i.e., we keep an index of CS pairs that are linked with at least one object-subject pair of records. These links are called Extended Characteristic Sets (ECSs) and are based on our previous work in \cite{meimaris2017extended}. With the ECS index, we can quickly filter out CSs that are guaranteed not to be related, i.e., no joins exist between them, even if they are individually matched in a chain of query CSs. Other metadata and indexes include the property sets of CSs, and which properties can contain multiple values in the same CS. \begin{figure}[t] \centering \includegraphics[scale=0.35]{arch3.png} \caption{Architecture of \textit{raxonDB}.}\label{fig:arch} \end{figure} \textbf{Query Processing. }Processing SPARQL queries on top of merged CSs entails (i) parsing the queries, (ii) retrieving the query CSs, (iii) identifying the joins between them, and (iv) mapping them to merged tables in the database. Steps (i)-(iii) are inherited from our previous work in \cite{meimaris2017extended}. For (iv), a query CS can match with more than one table in the database. For instance, consider a query containing a chain of three CSs, $q_1 \bowtie q_2 \bowtie q_3$, joined sequentially with object-subject joins. Each query CS $q_i$ matches with all tables whose property sets are supersets of the property set of $q_i$. Thus, each join in the initial query creates a set of \textit{permutations} of table joins that need to be evaluated. For instance, assume that $q_1$ matches with $c_1, c_2$, while $q_2$ matches with $c_3$ and $q_3$ matches with $c_4, c_5$. Furthermore, by looking up the ECS index, we derived that the links $\left[c_1,c_3\right]$, $\left[c_2,c_3\right]$, $\left[c_3,c_4\right]$ and $\left[c_3,c_5\right]$ are all valid, i.e., they correspond to candidate joins in the data. Then, $\left[c_1, c_3, c_4\right]$, $\left[c_1, c_3, c_5\right]$, $\left[c_2, c_3, c_4\right]$ and $\left[c_2, c_3, c_5\right]$ are all valid table permutations that must be processed. Two strategies can be employed here. The first is to join the UNIONs of the matching tables for each $q_i$, and the other is to process each permutation of tables separately and append the results. Given the filtering performed by the ECS indexing approach, where we can pre-filter CSs based on the relationships between them, the UNION would impose significant overhead and eliminate the advantage of ECS indexing. Therefore, we have implemented the second approach, that is, process a separate query for each permutation. Finally, due to the existence of NULL values in the merged tables, we must add explicit IS NOT NULL restrictions for all the properties that are contained in each matched CS and are not part of any other restriction or filter in the original query. \section{Experimental Evaluation} We implemented \textit{raxonDB} on top of PostgreSQL\footnote{The code and queries are available in https://github.com/mmeimaris/raxonDB}. We did not extend our previous native RDF implementation of \textit{axonDB} \cite{meimaris2017extended}, because given the underlying relational schema of the CS tables, we decided to rely on a well-established relational engine for both the planning and the execution of queries, instead of re-implementing it. As the focus of this paper is to improve RDF storage and querying efficiency in relational settings, we rely on existing mechanisms within PostgreSQL for I/O operations, physical storage and query planning. In this set of experiments, we report results of implementing $hier\_merge$ with the greedy approximation algorithm, as experimenting with the optimal algorithm failed to finish the merging process even in datasets with small numbers of CSs. \textbf{Datasets. }For this set of experiments, we used two synthetic datasets, namely \textit{LUBM2000} ($\approx$300m triples), and WatDiv ($\approx$100m triples), as well as two real-world datasets, namely \textit{Geonames} ($\approx$170m triples) and \textit{Reactome} ($\approx$15m triples). LUBM \cite{guo2005lubm} is a customizable generator of synthetic data that describes academic information about universities, departments, faculty, and so on. Similarly, WatDiv\cite{alucc2014diversified} is a customizable generator with more options for the production and distribution of triples to classes. \textit{Reactome}\footnote{http://www.ebi.ac.uk/rdf/services/reactome} is a biological dataset that describes biological pathways, and \textit{Geonames}\footnote{http://www.geonames.org/ontology/documentation.html} is a widely used ontology of geographical entities with varying properties. Geonames maintains a rich graph structure as there is a heavy usage of hierarchical area features on a multitude of levels. \textbf{Loading. }In order to assess the effect of hierarchical merging in the loading phase, we performed a series of experiments using all four datasets. For this experiment, we measure the size on disk, the loading time, the final number of merged tables, as well as the number of ECSs (joins between merged tables) and the percentage of triple coverage by CSs included in the merging process, for varying values of the density factor $m \in [0,1]$. The results are summarized in Table \ref{table:loading}. As can be seen, the number of CS, and consequently tables, is greatly reduced with increasing values of $m$. As the number of CSs is reduced, the expected number of joins between CSs is also reduced, which can be seen in the column that measures ECSs. Consequently, the number of tables can be decreased significantly without trading off large amounts of coverage by dense CSs, i.e. large tables with many null values. Loading time tends to be slightly greater as the number of CSs decreases, and thus the number of merges increases, the only exception being WatDiv, where loading time is actually decreased. This is a side-effect of the excessive number of tables ($=5667$) in the simple case which imposes large overheads for the persistence of the tables on disk and the generation of indexes and statistics for each one. \begin{figure*}[t] \def\tabularxcolumn#1{m{#1}} \begin{tabularx}{\linewidth}{@{}cXX@{}} \begin{tabular}{ccc} \subfloat[Execution time (seconds) for LUBM]{\includegraphics[width=0.31\textwidth]{queries_lubm_merge_4.png}} & \subfloat[Execution time (seconds) for Geonames]{\includegraphics[width=0.3\textwidth]{queries_geonames_merge_5.png}} & \subfloat[Execution time (seconds) for Reactome]{\includegraphics[width=0.31\textwidth]{queries_reactome_merge_4.png}} \\ \end{tabular} \end{tabularx} \caption{Query execution times in milliseconds}\label{fig:exp_fig_1} \end{figure*} \begin{figure*}[t] \def\tabularxcolumn#1{m{#1}} \begin{tabularx}{\linewidth}{@{}cXX@{}} \begin{tabular}{ccc} \subfloat[\# of CS permutations for LUBM]{\includegraphics[width=0.31\textwidth]{permutations_lubm_merge_2.png}} & \subfloat[\# of CS permutations for Geonames]{\includegraphics[width=0.31\textwidth]{permutations_geonames_merge_2.png}} & \subfloat[\# of CS permutations for Reactome]{\includegraphics[width=0.31\textwidth]{permutations_reactome_merge_2.png}} \\ \end{tabular} \end{tabularx} \caption{\# of CS permutations for increasing m}\label{fig:exp_fig_2} \end{figure*} \begin{table} \small \centering \caption{Loading experiments for all datasets} \resizebox{\columnwidth}{!}{ \begin{tabular}{|c|c c c c c |} \hline \textbf{Dataset} & \textbf{Size (MB)} & \textbf{Time} & \textbf{\# Tables (CSs)} & \textbf{\# of ECSs} & \textbf{Dense CS}\\ & & & & & Coverage\\ \hline Reactome Simple & 781 & 3min & 112 & 346 & 100\% \\ Reactome (m=0.05) & 675 & 4min & 35 & 252 & 97\% \\ Reactome (m=0.25) & 865 & 4min & 14 & 73 & 77\% \\ \hline Geonames Simple & 4991 & 69min & 851 & 12136 & 100\% \\ Geonames (m=0.0025) & 4999 & 70min & 82 & 2455 & 97\% \\ Geonames (m=0.05) & 5093 & 91min & 19 & 76 & 87\% \\ Geonames (m=0.1) & 5104 & 92min & 6 & 28 & 83\% \\ \hline LUBM Simple & 591 & 3min & 14 & 68 & 100\% \\ LUBM (m=0.25) & 610 & 3min & 6 & 21 & 90\% \\ LUBM (m=0.5) & 620 & 3min & 3 & 6 & 58\% \\ \hline WatDiv Simple & 4910 & 97min & 5667 & 802 & 100\% \\ WatDiv (m=0.01) & 5094 & 75min & 67 & 99 & 77\% \\ WatDiv (m=0.1) & 5250 & 75min & 25 & 23 & 63\% \\ WatDiv (m=0.5) & 5250 & 77min & 16 & 19 & 55\% \\ \hline \end{tabular} } \label{table:loading} \end{table} \textbf{Query Performance. } In order to assess the effect of the density factor parameter $m$ during query processing, we perform a series of experiments on LUBM, Reactome and Geonames. For the workload, we used the sets of queries from \cite{meimaris2017extended}. We employ two metrics, namely \textit{execution time} and \textit{number of table permutations}. The results can be seen in Figures \ref{fig:exp_fig_1} and \ref{fig:exp_fig_2}. As can be seen, hierarchical CS merging can help speed up query performance significantly as long as the dense coverage remains high. For example, in all datasets, query performance degrades dramatically when $m=1$, in which case the merging process cannot find any dense CSs. In this case, all rows are added to one large table, which makes the database only contain one table with many NULL cells. These findings are consistent across all three datasets and require further future work in order to identify the optimal value for $m$. In order to assess the performance of \textit{raxonDB} and establish that no overhead is imposed by the relational backbone, we performed a series of queries on LUBM2000, Geonames and Reactome, assuming the best merging of CSs is employed as captured by $m$ with respect to our previous findings. We also compared the query performance with rdf-3x, Virtuoso 7.1, TripleBit and the emergent schema approach described in \cite{pham2016exploiting}. The results can be seen in Figure \ref{fig:comp_with_old} and indicate that \textit{raxonDB} provides equal or better performance from the original \textit{axonDB} implementation, as well as the rest of the systems, including the emergent schema approach, which is the only direct competitor for merging CSs. Especially for queries with large intermediate results and low selectivity that correspond to a few CSs and ECSs (e.g. LUBM Q5 and Q6, Geonames Q5 and Q6) several of the other approaches fail to answer fast and in some cases time out. \begin{figure*}[t] \def\tabularxcolumn#1{m{#1}} \begin{tabularx}{\linewidth}{@{}cXX@{}} \begin{tabular}{ccc} \subfloat[Execution time (seconds) for LUBM2000]{\includegraphics[width=0.33\textwidth]{comp_with_old_2.png}} & \subfloat[Execution time (seconds) for Geonames]{\includegraphics[width=0.33\textwidth]{comp_with_old_geonames.png}} & \subfloat[Execution time (seconds) for Reactome]{\includegraphics[width=0.33\textwidth]{comp_with_old_reactome.png}} \\ \end{tabular} \end{tabularx} \caption{Query execution times in milliseconds for different RDF engines}\label{fig:comp_with_old} \end{figure*} \section{Conclusions and Future Work} In this paper, we tackled the problem of merging characteristic sets based on their hierarchical relationships. As future work, we will study computation of the optimal value for $m$, taking into consideration workload characteristics as well as a more refined cost model for the ancestral paths. Furthermore, we will study application of these findings in a distributed architecture, in order to further scale the capabilities of \textit{raxonDB}.

35d0192c90b1fd4c31e1d501dc91f553fd1e4953